tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
docs: rcu: convert some articles from html to ReST
There are 4 RCU articles that are written on html format.
The way they are, they can't be part of the Linux Kernel
documentation body nor share the styles and pdf output.
So, convert them to ReST format.
This way, make htmldocs and make pdfdocs will produce a
documentation output that will be like the original ones, but
will be part of the Linux Kernel documentation body.
Part of the conversion was done with the help of pandoc, but
the result had some broken things that had to be manually
fixed.
Following are manual changes Mauro made when doing the automatic conversion:
Quoting from: https://lore.kernel.org/rcu/20190726154550.5eeae294@coco.lan/
> > At least the pandoc's version I used here has a bug: its conversion
> > from html to ReST on those files only start after a <body> tag - or
> > when the first quiz table starts. I only discovered that adding a
> > <body> at the beginning of the file solve this book at the last
> > conversions.
> >
> > So, for most html->ReST conversions, I manually converted the first
> > part of the document, basically stripping html paragraph tags and
> > by replacing highlights by the ReST syntax.
> >
> > Also, all the quiz tables seem to assume some javascript macro or
> > css style that would be hiding the answer part until the mouse moves
> > to it. Such macro/css was not there at the kernel tree. So, the quiz
> > answers have the same color as the background, making them invisible.
> > Even if we had such macro/css, this is not portable for pdf/LaTeX output
> > (and I'm not sure if this would work with ePub).
> >
> > So, I ended by manually doing the table conversion.
> >
> > Finally, I double-checked if the conversions ended ok, addressing any
> > issues that might have heppened.
> >
> > So, after both automatic conversion and manual fixes, I opened both the
> > html files produced by Sphinx and the original ones and compared them
> > line per line (except for the indexes, as Sphinx produces them
> > automatically), in order to see if all information from the original
> > files will be there on a format close to what we have on other ReST
> > files, fixing any pending issues if any.
Signed-off-by: Mauro Carvalho Chehab <mchehab+samsung@kernel.org>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
authored by
Mauro Carvalho Chehab
and committed by
Paul E. McKenney
ccc9971e
c07e6f36
+4978
-6104
-1391
Documentation/RCU/Design/Data-Structures/Data-Structures.html
-1391
Documentation/RCU/Design/Data-Structures/Data-Structures.html
···
1
-
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN"
2
-
"http://www.w3.org/TR/html4/loose.dtd">
3
-
<html>
4
-
<head><title>A Tour Through TREE_RCU's Data Structures [LWN.net]</title>
5
-
<meta HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=iso-8859-1">
6
-
7
-
<p>December 18, 2016</p>
8
-
<p>This article was contributed by Paul E. McKenney</p>
9
-
10
-
<h3>Introduction</h3>
11
-
12
-
This document describes RCU's major data structures and their relationship
13
-
to each other.
14
-
15
-
<ol>
16
-
<li> <a href="#Data-Structure Relationships">
17
-
Data-Structure Relationships</a>
18
-
<li> <a href="#The rcu_state Structure">
19
-
The <tt>rcu_state</tt> Structure</a>
20
-
<li> <a href="#The rcu_node Structure">
21
-
The <tt>rcu_node</tt> Structure</a>
22
-
<li> <a href="#The rcu_segcblist Structure">
23
-
The <tt>rcu_segcblist</tt> Structure</a>
24
-
<li> <a href="#The rcu_data Structure">
25
-
The <tt>rcu_data</tt> Structure</a>
26
-
<li> <a href="#The rcu_head Structure">
27
-
The <tt>rcu_head</tt> Structure</a>
28
-
<li> <a href="#RCU-Specific Fields in the task_struct Structure">
29
-
RCU-Specific Fields in the <tt>task_struct</tt> Structure</a>
30
-
<li> <a href="#Accessor Functions">
31
-
Accessor Functions</a>
32
-
</ol>
33
-
34
-
<h3><a name="Data-Structure Relationships">Data-Structure Relationships</a></h3>
35
-
36
-
<p>RCU is for all intents and purposes a large state machine, and its
37
-
data structures maintain the state in such a way as to allow RCU readers
38
-
to execute extremely quickly, while also processing the RCU grace periods
39
-
requested by updaters in an efficient and extremely scalable fashion.
40
-
The efficiency and scalability of RCU updaters is provided primarily
41
-
by a combining tree, as shown below:
42
-
43
-
</p><p><img src="BigTreeClassicRCU.svg" alt="BigTreeClassicRCU.svg" width="30%">
44
-
45
-
</p><p>This diagram shows an enclosing <tt>rcu_state</tt> structure
46
-
containing a tree of <tt>rcu_node</tt> structures.
47
-
Each leaf node of the <tt>rcu_node</tt> tree has up to 16
48
-
<tt>rcu_data</tt> structures associated with it, so that there
49
-
are <tt>NR_CPUS</tt> number of <tt>rcu_data</tt> structures,
50
-
one for each possible CPU.
51
-
This structure is adjusted at boot time, if needed, to handle the
52
-
common case where <tt>nr_cpu_ids</tt> is much less than
53
-
<tt>NR_CPUs</tt>.
54
-
For example, a number of Linux distributions set <tt>NR_CPUs=4096</tt>,
55
-
which results in a three-level <tt>rcu_node</tt> tree.
56
-
If the actual hardware has only 16 CPUs, RCU will adjust itself
57
-
at boot time, resulting in an <tt>rcu_node</tt> tree with only a single node.
58
-
59
-
</p><p>The purpose of this combining tree is to allow per-CPU events
60
-
such as quiescent states, dyntick-idle transitions,
61
-
and CPU hotplug operations to be processed efficiently
62
-
and scalably.
63
-
Quiescent states are recorded by the per-CPU <tt>rcu_data</tt> structures,
64
-
and other events are recorded by the leaf-level <tt>rcu_node</tt>
65
-
structures.
66
-
All of these events are combined at each level of the tree until finally
67
-
grace periods are completed at the tree's root <tt>rcu_node</tt>
68
-
structure.
69
-
A grace period can be completed at the root once every CPU
70
-
(or, in the case of <tt>CONFIG_PREEMPT_RCU</tt>, task)
71
-
has passed through a quiescent state.
72
-
Once a grace period has completed, record of that fact is propagated
73
-
back down the tree.
74
-
75
-
</p><p>As can be seen from the diagram, on a 64-bit system
76
-
a two-level tree with 64 leaves can accommodate 1,024 CPUs, with a fanout
77
-
of 64 at the root and a fanout of 16 at the leaves.
78
-
79
-
<table>
80
-
<tr><th> </th></tr>
81
-
<tr><th align="left">Quick Quiz:</th></tr>
82
-
<tr><td>
83
-
Why isn't the fanout at the leaves also 64?
84
-
</td></tr>
85
-
<tr><th align="left">Answer:</th></tr>
86
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
87
-
Because there are more types of events that affect the leaf-level
88
-
<tt>rcu_node</tt> structures than further up the tree.
89
-
Therefore, if the leaf <tt>rcu_node</tt> structures have fanout of
90
-
64, the contention on these structures' <tt>->structures</tt>
91
-
becomes excessive.
92
-
Experimentation on a wide variety of systems has shown that a fanout
93
-
of 16 works well for the leaves of the <tt>rcu_node</tt> tree.
94
-
</font>
95
-
96
-
<p><font color="ffffff">Of course, further experience with
97
-
systems having hundreds or thousands of CPUs may demonstrate
98
-
that the fanout for the non-leaf <tt>rcu_node</tt> structures
99
-
must also be reduced.
100
-
Such reduction can be easily carried out when and if it proves
101
-
necessary.
102
-
In the meantime, if you are using such a system and running into
103
-
contention problems on the non-leaf <tt>rcu_node</tt> structures,
104
-
you may use the <tt>CONFIG_RCU_FANOUT</tt> kernel configuration
105
-
parameter to reduce the non-leaf fanout as needed.
106
-
</font>
107
-
108
-
<p><font color="ffffff">Kernels built for systems with
109
-
strong NUMA characteristics might also need to adjust
110
-
<tt>CONFIG_RCU_FANOUT</tt> so that the domains of the
111
-
<tt>rcu_node</tt> structures align with hardware boundaries.
112
-
However, there has thus far been no need for this.
113
-
</font></td></tr>
114
-
<tr><td> </td></tr>
115
-
</table>
116
-
117
-
<p>If your system has more than 1,024 CPUs (or more than 512 CPUs on
118
-
a 32-bit system), then RCU will automatically add more levels to the
119
-
tree.
120
-
For example, if you are crazy enough to build a 64-bit system with 65,536
121
-
CPUs, RCU would configure the <tt>rcu_node</tt> tree as follows:
122
-
123
-
</p><p><img src="HugeTreeClassicRCU.svg" alt="HugeTreeClassicRCU.svg" width="50%">
124
-
125
-
</p><p>RCU currently permits up to a four-level tree, which on a 64-bit system
126
-
accommodates up to 4,194,304 CPUs, though only a mere 524,288 CPUs for
127
-
32-bit systems.
128
-
On the other hand, you can set both <tt>CONFIG_RCU_FANOUT</tt> and
129
-
<tt>CONFIG_RCU_FANOUT_LEAF</tt> to be as small as 2, which would result
130
-
in a 16-CPU test using a 4-level tree.
131
-
This can be useful for testing large-system capabilities on small test
132
-
machines.
133
-
134
-
</p><p>This multi-level combining tree allows us to get most of the
135
-
performance and scalability
136
-
benefits of partitioning, even though RCU grace-period detection is
137
-
inherently a global operation.
138
-
The trick here is that only the last CPU to report a quiescent state
139
-
into a given <tt>rcu_node</tt> structure need advance to the <tt>rcu_node</tt>
140
-
structure at the next level up the tree.
141
-
This means that at the leaf-level <tt>rcu_node</tt> structure, only
142
-
one access out of sixteen will progress up the tree.
143
-
For the internal <tt>rcu_node</tt> structures, the situation is even
144
-
more extreme: Only one access out of sixty-four will progress up
145
-
the tree.
146
-
Because the vast majority of the CPUs do not progress up the tree,
147
-
the lock contention remains roughly constant up the tree.
148
-
No matter how many CPUs there are in the system, at most 64 quiescent-state
149
-
reports per grace period will progress all the way to the root
150
-
<tt>rcu_node</tt> structure, thus ensuring that the lock contention
151
-
on that root <tt>rcu_node</tt> structure remains acceptably low.
152
-
153
-
</p><p>In effect, the combining tree acts like a big shock absorber,
154
-
keeping lock contention under control at all tree levels regardless
155
-
of the level of loading on the system.
156
-
157
-
</p><p>RCU updaters wait for normal grace periods by registering
158
-
RCU callbacks, either directly via <tt>call_rcu()</tt>
159
-
or indirectly via <tt>synchronize_rcu()</tt> and friends.
160
-
RCU callbacks are represented by <tt>rcu_head</tt> structures,
161
-
which are queued on <tt>rcu_data</tt> structures while they are
162
-
waiting for a grace period to elapse, as shown in the following figure:
163
-
164
-
</p><p><img src="BigTreePreemptRCUBHdyntickCB.svg" alt="BigTreePreemptRCUBHdyntickCB.svg" width="40%">
165
-
166
-
</p><p>This figure shows how <tt>TREE_RCU</tt>'s and
167
-
<tt>PREEMPT_RCU</tt>'s major data structures are related.
168
-
Lesser data structures will be introduced with the algorithms that
169
-
make use of them.
170
-
171
-
</p><p>Note that each of the data structures in the above figure has
172
-
its own synchronization:
173
-
174
-
<p><ol>
175
-
<li> Each <tt>rcu_state</tt> structures has a lock and a mutex,
176
-
and some fields are protected by the corresponding root
177
-
<tt>rcu_node</tt> structure's lock.
178
-
<li> Each <tt>rcu_node</tt> structure has a spinlock.
179
-
<li> The fields in <tt>rcu_data</tt> are private to the corresponding
180
-
CPU, although a few can be read and written by other CPUs.
181
-
</ol>
182
-
183
-
<p>It is important to note that different data structures can have
184
-
very different ideas about the state of RCU at any given time.
185
-
For but one example, awareness of the start or end of a given RCU
186
-
grace period propagates slowly through the data structures.
187
-
This slow propagation is absolutely necessary for RCU to have good
188
-
read-side performance.
189
-
If this balkanized implementation seems foreign to you, one useful
190
-
trick is to consider each instance of these data structures to be
191
-
a different person, each having the usual slightly different
192
-
view of reality.
193
-
194
-
</p><p>The general role of each of these data structures is as
195
-
follows:
196
-
197
-
</p><ol>
198
-
<li> <tt>rcu_state</tt>:
199
-
This structure forms the interconnection between the
200
-
<tt>rcu_node</tt> and <tt>rcu_data</tt> structures,
201
-
tracks grace periods, serves as short-term repository
202
-
for callbacks orphaned by CPU-hotplug events,
203
-
maintains <tt>rcu_barrier()</tt> state,
204
-
tracks expedited grace-period state,
205
-
and maintains state used to force quiescent states when
206
-
grace periods extend too long,
207
-
<li> <tt>rcu_node</tt>: This structure forms the combining
208
-
tree that propagates quiescent-state
209
-
information from the leaves to the root, and also propagates
210
-
grace-period information from the root to the leaves.
211
-
It provides local copies of the grace-period state in order
212
-
to allow this information to be accessed in a synchronized
213
-
manner without suffering the scalability limitations that
214
-
would otherwise be imposed by global locking.
215
-
In <tt>CONFIG_PREEMPT_RCU</tt> kernels, it manages the lists
216
-
of tasks that have blocked while in their current
217
-
RCU read-side critical section.
218
-
In <tt>CONFIG_PREEMPT_RCU</tt> with
219
-
<tt>CONFIG_RCU_BOOST</tt>, it manages the
220
-
per-<tt>rcu_node</tt> priority-boosting
221
-
kernel threads (kthreads) and state.
222
-
Finally, it records CPU-hotplug state in order to determine
223
-
which CPUs should be ignored during a given grace period.
224
-
<li> <tt>rcu_data</tt>: This per-CPU structure is the
225
-
focus of quiescent-state detection and RCU callback queuing.
226
-
It also tracks its relationship to the corresponding leaf
227
-
<tt>rcu_node</tt> structure to allow more-efficient
228
-
propagation of quiescent states up the <tt>rcu_node</tt>
229
-
combining tree.
230
-
Like the <tt>rcu_node</tt> structure, it provides a local
231
-
copy of the grace-period information to allow for-free
232
-
synchronized
233
-
access to this information from the corresponding CPU.
234
-
Finally, this structure records past dyntick-idle state
235
-
for the corresponding CPU and also tracks statistics.
236
-
<li> <tt>rcu_head</tt>:
237
-
This structure represents RCU callbacks, and is the
238
-
only structure allocated and managed by RCU users.
239
-
The <tt>rcu_head</tt> structure is normally embedded
240
-
within the RCU-protected data structure.
241
-
</ol>
242
-
243
-
<p>If all you wanted from this article was a general notion of how
244
-
RCU's data structures are related, you are done.
245
-
Otherwise, each of the following sections give more details on
246
-
the <tt>rcu_state</tt>, <tt>rcu_node</tt> and <tt>rcu_data</tt> data
247
-
structures.
248
-
249
-
<h3><a name="The rcu_state Structure">
250
-
The <tt>rcu_state</tt> Structure</a></h3>
251
-
252
-
<p>The <tt>rcu_state</tt> structure is the base structure that
253
-
represents the state of RCU in the system.
254
-
This structure forms the interconnection between the
255
-
<tt>rcu_node</tt> and <tt>rcu_data</tt> structures,
256
-
tracks grace periods, contains the lock used to
257
-
synchronize with CPU-hotplug events,
258
-
and maintains state used to force quiescent states when
259
-
grace periods extend too long,
260
-
261
-
</p><p>A few of the <tt>rcu_state</tt> structure's fields are discussed,
262
-
singly and in groups, in the following sections.
263
-
The more specialized fields are covered in the discussion of their
264
-
use.
265
-
266
-
<h5>Relationship to rcu_node and rcu_data Structures</h5>
267
-
268
-
This portion of the <tt>rcu_state</tt> structure is declared
269
-
as follows:
270
-
271
-
<pre>
272
-
1 struct rcu_node node[NUM_RCU_NODES];
273
-
2 struct rcu_node *level[NUM_RCU_LVLS + 1];
274
-
3 struct rcu_data __percpu *rda;
275
-
</pre>
276
-
277
-
<table>
278
-
<tr><th> </th></tr>
279
-
<tr><th align="left">Quick Quiz:</th></tr>
280
-
<tr><td>
281
-
Wait a minute!
282
-
You said that the <tt>rcu_node</tt> structures formed a tree,
283
-
but they are declared as a flat array!
284
-
What gives?
285
-
</td></tr>
286
-
<tr><th align="left">Answer:</th></tr>
287
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
288
-
The tree is laid out in the array.
289
-
The first node In the array is the head, the next set of nodes in the
290
-
array are children of the head node, and so on until the last set of
291
-
nodes in the array are the leaves.
292
-
</font>
293
-
294
-
<p><font color="ffffff">See the following diagrams to see how
295
-
this works.
296
-
</font></td></tr>
297
-
<tr><td> </td></tr>
298
-
</table>
299
-
300
-
<p>The <tt>rcu_node</tt> tree is embedded into the
301
-
<tt>->node[]</tt> array as shown in the following figure:
302
-
303
-
</p><p><img src="TreeMapping.svg" alt="TreeMapping.svg" width="40%">
304
-
305
-
</p><p>One interesting consequence of this mapping is that a
306
-
breadth-first traversal of the tree is implemented as a simple
307
-
linear scan of the array, which is in fact what the
308
-
<tt>rcu_for_each_node_breadth_first()</tt> macro does.
309
-
This macro is used at the beginning and ends of grace periods.
310
-
311
-
</p><p>Each entry of the <tt>->level</tt> array references
312
-
the first <tt>rcu_node</tt> structure on the corresponding level
313
-
of the tree, for example, as shown below:
314
-
315
-
</p><p><img src="TreeMappingLevel.svg" alt="TreeMappingLevel.svg" width="40%">
316
-
317
-
</p><p>The zero<sup>th</sup> element of the array references the root
318
-
<tt>rcu_node</tt> structure, the first element references the
319
-
first child of the root <tt>rcu_node</tt>, and finally the second
320
-
element references the first leaf <tt>rcu_node</tt> structure.
321
-
322
-
</p><p>For whatever it is worth, if you draw the tree to be tree-shaped
323
-
rather than array-shaped, it is easy to draw a planar representation:
324
-
325
-
</p><p><img src="TreeLevel.svg" alt="TreeLevel.svg" width="60%">
326
-
327
-
</p><p>Finally, the <tt>->rda</tt> field references a per-CPU
328
-
pointer to the corresponding CPU's <tt>rcu_data</tt> structure.
329
-
330
-
</p><p>All of these fields are constant once initialization is complete,
331
-
and therefore need no protection.
332
-
333
-
<h5>Grace-Period Tracking</h5>
334
-
335
-
<p>This portion of the <tt>rcu_state</tt> structure is declared
336
-
as follows:
337
-
338
-
<pre>
339
-
1 unsigned long gp_seq;
340
-
</pre>
341
-
342
-
<p>RCU grace periods are numbered, and
343
-
the <tt>->gp_seq</tt> field contains the current grace-period
344
-
sequence number.
345
-
The bottom two bits are the state of the current grace period,
346
-
which can be zero for not yet started or one for in progress.
347
-
In other words, if the bottom two bits of <tt>->gp_seq</tt> are
348
-
zero, then RCU is idle.
349
-
Any other value in the bottom two bits indicates that something is broken.
350
-
This field is protected by the root <tt>rcu_node</tt> structure's
351
-
<tt>->lock</tt> field.
352
-
353
-
</p><p>There are <tt>->gp_seq</tt> fields
354
-
in the <tt>rcu_node</tt> and <tt>rcu_data</tt> structures
355
-
as well.
356
-
The fields in the <tt>rcu_state</tt> structure represent the
357
-
most current value, and those of the other structures are compared
358
-
in order to detect the beginnings and ends of grace periods in a distributed
359
-
fashion.
360
-
The values flow from <tt>rcu_state</tt> to <tt>rcu_node</tt>
361
-
(down the tree from the root to the leaves) to <tt>rcu_data</tt>.
362
-
363
-
<h5>Miscellaneous</h5>
364
-
365
-
<p>This portion of the <tt>rcu_state</tt> structure is declared
366
-
as follows:
367
-
368
-
<pre>
369
-
1 unsigned long gp_max;
370
-
2 char abbr;
371
-
3 char *name;
372
-
</pre>
373
-
374
-
<p>The <tt>->gp_max</tt> field tracks the duration of the longest
375
-
grace period in jiffies.
376
-
It is protected by the root <tt>rcu_node</tt>'s <tt>->lock</tt>.
377
-
378
-
<p>The <tt>->name</tt> and <tt>->abbr</tt> fields distinguish
379
-
between preemptible RCU (“rcu_preempt” and “p”)
380
-
and non-preemptible RCU (“rcu_sched” and “s”).
381
-
These fields are used for diagnostic and tracing purposes.
382
-
383
-
<h3><a name="The rcu_node Structure">
384
-
The <tt>rcu_node</tt> Structure</a></h3>
385
-
386
-
<p>The <tt>rcu_node</tt> structures form the combining
387
-
tree that propagates quiescent-state
388
-
information from the leaves to the root and also that propagates
389
-
grace-period information from the root down to the leaves.
390
-
They provides local copies of the grace-period state in order
391
-
to allow this information to be accessed in a synchronized
392
-
manner without suffering the scalability limitations that
393
-
would otherwise be imposed by global locking.
394
-
In <tt>CONFIG_PREEMPT_RCU</tt> kernels, they manage the lists
395
-
of tasks that have blocked while in their current
396
-
RCU read-side critical section.
397
-
In <tt>CONFIG_PREEMPT_RCU</tt> with
398
-
<tt>CONFIG_RCU_BOOST</tt>, they manage the
399
-
per-<tt>rcu_node</tt> priority-boosting
400
-
kernel threads (kthreads) and state.
401
-
Finally, they record CPU-hotplug state in order to determine
402
-
which CPUs should be ignored during a given grace period.
403
-
404
-
</p><p>The <tt>rcu_node</tt> structure's fields are discussed,
405
-
singly and in groups, in the following sections.
406
-
407
-
<h5>Connection to Combining Tree</h5>
408
-
409
-
<p>This portion of the <tt>rcu_node</tt> structure is declared
410
-
as follows:
411
-
412
-
<pre>
413
-
1 struct rcu_node *parent;
414
-
2 u8 level;
415
-
3 u8 grpnum;
416
-
4 unsigned long grpmask;
417
-
5 int grplo;
418
-
6 int grphi;
419
-
</pre>
420
-
421
-
<p>The <tt>->parent</tt> pointer references the <tt>rcu_node</tt>
422
-
one level up in the tree, and is <tt>NULL</tt> for the root
423
-
<tt>rcu_node</tt>.
424
-
The RCU implementation makes heavy use of this field to push quiescent
425
-
states up the tree.
426
-
The <tt>->level</tt> field gives the level in the tree, with
427
-
the root being at level zero, its children at level one, and so on.
428
-
The <tt>->grpnum</tt> field gives this node's position within
429
-
the children of its parent, so this number can range between 0 and 31
430
-
on 32-bit systems and between 0 and 63 on 64-bit systems.
431
-
The <tt>->level</tt> and <tt>->grpnum</tt> fields are
432
-
used only during initialization and for tracing.
433
-
The <tt>->grpmask</tt> field is the bitmask counterpart of
434
-
<tt>->grpnum</tt>, and therefore always has exactly one bit set.
435
-
This mask is used to clear the bit corresponding to this <tt>rcu_node</tt>
436
-
structure in its parent's bitmasks, which are described later.
437
-
Finally, the <tt>->grplo</tt> and <tt>->grphi</tt> fields
438
-
contain the lowest and highest numbered CPU served by this
439
-
<tt>rcu_node</tt> structure, respectively.
440
-
441
-
</p><p>All of these fields are constant, and thus do not require any
442
-
synchronization.
443
-
444
-
<h5>Synchronization</h5>
445
-
446
-
<p>This field of the <tt>rcu_node</tt> structure is declared
447
-
as follows:
448
-
449
-
<pre>
450
-
1 raw_spinlock_t lock;
451
-
</pre>
452
-
453
-
<p>This field is used to protect the remaining fields in this structure,
454
-
unless otherwise stated.
455
-
That said, all of the fields in this structure can be accessed without
456
-
locking for tracing purposes.
457
-
Yes, this can result in confusing traces, but better some tracing confusion
458
-
than to be heisenbugged out of existence.
459
-
460
-
<h5>Grace-Period Tracking</h5>
461
-
462
-
<p>This portion of the <tt>rcu_node</tt> structure is declared
463
-
as follows:
464
-
465
-
<pre>
466
-
1 unsigned long gp_seq;
467
-
2 unsigned long gp_seq_needed;
468
-
</pre>
469
-
470
-
<p>The <tt>rcu_node</tt> structures' <tt>->gp_seq</tt> fields are
471
-
the counterparts of the field of the same name in the <tt>rcu_state</tt>
472
-
structure.
473
-
They each may lag up to one step behind their <tt>rcu_state</tt>
474
-
counterpart.
475
-
If the bottom two bits of a given <tt>rcu_node</tt> structure's
476
-
<tt>->gp_seq</tt> field is zero, then this <tt>rcu_node</tt>
477
-
structure believes that RCU is idle.
478
-
</p><p>The <tt>>gp_seq</tt> field of each <tt>rcu_node</tt>
479
-
structure is updated at the beginning and the end
480
-
of each grace period.
481
-
482
-
<p>The <tt>->gp_seq_needed</tt> fields record the
483
-
furthest-in-the-future grace period request seen by the corresponding
484
-
<tt>rcu_node</tt> structure. The request is considered fulfilled when
485
-
the value of the <tt>->gp_seq</tt> field equals or exceeds that of
486
-
the <tt>->gp_seq_needed</tt> field.
487
-
488
-
<table>
489
-
<tr><th> </th></tr>
490
-
<tr><th align="left">Quick Quiz:</th></tr>
491
-
<tr><td>
492
-
Suppose that this <tt>rcu_node</tt> structure doesn't see
493
-
a request for a very long time.
494
-
Won't wrapping of the <tt>->gp_seq</tt> field cause
495
-
problems?
496
-
</td></tr>
497
-
<tr><th align="left">Answer:</th></tr>
498
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
499
-
No, because if the <tt>->gp_seq_needed</tt> field lags behind the
500
-
<tt>->gp_seq</tt> field, the <tt>->gp_seq_needed</tt> field
501
-
will be updated at the end of the grace period.
502
-
Modulo-arithmetic comparisons therefore will always get the
503
-
correct answer, even with wrapping.
504
-
</font></td></tr>
505
-
<tr><td> </td></tr>
506
-
</table>
507
-
508
-
<h5>Quiescent-State Tracking</h5>
509
-
510
-
<p>These fields manage the propagation of quiescent states up the
511
-
combining tree.
512
-
513
-
</p><p>This portion of the <tt>rcu_node</tt> structure has fields
514
-
as follows:
515
-
516
-
<pre>
517
-
1 unsigned long qsmask;
518
-
2 unsigned long expmask;
519
-
3 unsigned long qsmaskinit;
520
-
4 unsigned long expmaskinit;
521
-
</pre>
522
-
523
-
<p>The <tt>->qsmask</tt> field tracks which of this
524
-
<tt>rcu_node</tt> structure's children still need to report
525
-
quiescent states for the current normal grace period.
526
-
Such children will have a value of 1 in their corresponding bit.
527
-
Note that the leaf <tt>rcu_node</tt> structures should be
528
-
thought of as having <tt>rcu_data</tt> structures as their
529
-
children.
530
-
Similarly, the <tt>->expmask</tt> field tracks which
531
-
of this <tt>rcu_node</tt> structure's children still need to report
532
-
quiescent states for the current expedited grace period.
533
-
An expedited grace period has
534
-
the same conceptual properties as a normal grace period, but the
535
-
expedited implementation accepts extreme CPU overhead to obtain
536
-
much lower grace-period latency, for example, consuming a few
537
-
tens of microseconds worth of CPU time to reduce grace-period
538
-
duration from milliseconds to tens of microseconds.
539
-
The <tt>->qsmaskinit</tt> field tracks which of this
540
-
<tt>rcu_node</tt> structure's children cover for at least
541
-
one online CPU.
542
-
This mask is used to initialize <tt>->qsmask</tt>,
543
-
and <tt>->expmaskinit</tt> is used to initialize
544
-
<tt>->expmask</tt> and the beginning of the
545
-
normal and expedited grace periods, respectively.
546
-
547
-
<table>
548
-
<tr><th> </th></tr>
549
-
<tr><th align="left">Quick Quiz:</th></tr>
550
-
<tr><td>
551
-
Why are these bitmasks protected by locking?
552
-
Come on, haven't you heard of atomic instructions???
553
-
</td></tr>
554
-
<tr><th align="left">Answer:</th></tr>
555
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
556
-
Lockless grace-period computation! Such a tantalizing possibility!
557
-
</font>
558
-
559
-
<p><font color="ffffff">But consider the following sequence of events:
560
-
</font>
561
-
562
-
<ol>
563
-
<li> <font color="ffffff">CPU 0 has been in dyntick-idle
564
-
mode for quite some time.
565
-
When it wakes up, it notices that the current RCU
566
-
grace period needs it to report in, so it sets a
567
-
flag where the scheduling clock interrupt will find it.
568
-
</font><p>
569
-
<li> <font color="ffffff">Meanwhile, CPU 1 is running
570
-
<tt>force_quiescent_state()</tt>,
571
-
and notices that CPU 0 has been in dyntick idle mode,
572
-
which qualifies as an extended quiescent state.
573
-
</font><p>
574
-
<li> <font color="ffffff">CPU 0's scheduling clock
575
-
interrupt fires in the
576
-
middle of an RCU read-side critical section, and notices
577
-
that the RCU core needs something, so commences RCU softirq
578
-
processing.
579
-
</font>
580
-
<p>
581
-
<li> <font color="ffffff">CPU 0's softirq handler
582
-
executes and is just about ready
583
-
to report its quiescent state up the <tt>rcu_node</tt>
584
-
tree.
585
-
</font><p>
586
-
<li> <font color="ffffff">But CPU 1 beats it to the punch,
587
-
completing the current
588
-
grace period and starting a new one.
589
-
</font><p>
590
-
<li> <font color="ffffff">CPU 0 now reports its quiescent
591
-
state for the wrong
592
-
grace period.
593
-
That grace period might now end before the RCU read-side
594
-
critical section.
595
-
If that happens, disaster will ensue.
596
-
</font>
597
-
</ol>
598
-
599
-
<p><font color="ffffff">So the locking is absolutely required in
600
-
order to coordinate clearing of the bits with updating of the
601
-
grace-period sequence number in <tt>->gp_seq</tt>.
602
-
</font></td></tr>
603
-
<tr><td> </td></tr>
604
-
</table>
605
-
606
-
<h5>Blocked-Task Management</h5>
607
-
608
-
<p><tt>PREEMPT_RCU</tt> allows tasks to be preempted in the
609
-
midst of their RCU read-side critical sections, and these tasks
610
-
must be tracked explicitly.
611
-
The details of exactly why and how they are tracked will be covered
612
-
in a separate article on RCU read-side processing.
613
-
For now, it is enough to know that the <tt>rcu_node</tt>
614
-
structure tracks them.
615
-
616
-
<pre>
617
-
1 struct list_head blkd_tasks;
618
-
2 struct list_head *gp_tasks;
619
-
3 struct list_head *exp_tasks;
620
-
4 bool wait_blkd_tasks;
621
-
</pre>
622
-
623
-
<p>The <tt>->blkd_tasks</tt> field is a list header for
624
-
the list of blocked and preempted tasks.
625
-
As tasks undergo context switches within RCU read-side critical
626
-
sections, their <tt>task_struct</tt> structures are enqueued
627
-
(via the <tt>task_struct</tt>'s <tt>->rcu_node_entry</tt>
628
-
field) onto the head of the <tt>->blkd_tasks</tt> list for the
629
-
leaf <tt>rcu_node</tt> structure corresponding to the CPU
630
-
on which the outgoing context switch executed.
631
-
As these tasks later exit their RCU read-side critical sections,
632
-
they remove themselves from the list.
633
-
This list is therefore in reverse time order, so that if one of the tasks
634
-
is blocking the current grace period, all subsequent tasks must
635
-
also be blocking that same grace period.
636
-
Therefore, a single pointer into this list suffices to track
637
-
all tasks blocking a given grace period.
638
-
That pointer is stored in <tt>->gp_tasks</tt> for normal
639
-
grace periods and in <tt>->exp_tasks</tt> for expedited
640
-
grace periods.
641
-
These last two fields are <tt>NULL</tt> if either there is
642
-
no grace period in flight or if there are no blocked tasks
643
-
preventing that grace period from completing.
644
-
If either of these two pointers is referencing a task that
645
-
removes itself from the <tt>->blkd_tasks</tt> list,
646
-
then that task must advance the pointer to the next task on
647
-
the list, or set the pointer to <tt>NULL</tt> if there
648
-
are no subsequent tasks on the list.
649
-
650
-
</p><p>For example, suppose that tasks T1, T2, and T3 are
651
-
all hard-affinitied to the largest-numbered CPU in the system.
652
-
Then if task T1 blocked in an RCU read-side
653
-
critical section, then an expedited grace period started,
654
-
then task T2 blocked in an RCU read-side critical section,
655
-
then a normal grace period started, and finally task 3 blocked
656
-
in an RCU read-side critical section, then the state of the
657
-
last leaf <tt>rcu_node</tt> structure's blocked-task list
658
-
would be as shown below:
659
-
660
-
</p><p><img src="blkd_task.svg" alt="blkd_task.svg" width="60%">
661
-
662
-
</p><p>Task T1 is blocking both grace periods, task T2 is
663
-
blocking only the normal grace period, and task T3 is blocking
664
-
neither grace period.
665
-
Note that these tasks will not remove themselves from this list
666
-
immediately upon resuming execution.
667
-
They will instead remain on the list until they execute the outermost
668
-
<tt>rcu_read_unlock()</tt> that ends their RCU read-side critical
669
-
section.
670
-
671
-
<p>
672
-
The <tt>->wait_blkd_tasks</tt> field indicates whether or not
673
-
the current grace period is waiting on a blocked task.
674
-
675
-
<h5>Sizing the <tt>rcu_node</tt> Array</h5>
676
-
677
-
<p>The <tt>rcu_node</tt> array is sized via a series of
678
-
C-preprocessor expressions as follows:
679
-
680
-
<pre>
681
-
1 #ifdef CONFIG_RCU_FANOUT
682
-
2 #define RCU_FANOUT CONFIG_RCU_FANOUT
683
-
3 #else
684
-
4 # ifdef CONFIG_64BIT
685
-
5 # define RCU_FANOUT 64
686
-
6 # else
687
-
7 # define RCU_FANOUT 32
688
-
8 # endif
689
-
9 #endif
690
-
10
691
-
11 #ifdef CONFIG_RCU_FANOUT_LEAF
692
-
12 #define RCU_FANOUT_LEAF CONFIG_RCU_FANOUT_LEAF
693
-
13 #else
694
-
14 # ifdef CONFIG_64BIT
695
-
15 # define RCU_FANOUT_LEAF 64
696
-
16 # else
697
-
17 # define RCU_FANOUT_LEAF 32
698
-
18 # endif
699
-
19 #endif
700
-
20
701
-
21 #define RCU_FANOUT_1 (RCU_FANOUT_LEAF)
702
-
22 #define RCU_FANOUT_2 (RCU_FANOUT_1 * RCU_FANOUT)
703
-
23 #define RCU_FANOUT_3 (RCU_FANOUT_2 * RCU_FANOUT)
704
-
24 #define RCU_FANOUT_4 (RCU_FANOUT_3 * RCU_FANOUT)
705
-
25
706
-
26 #if NR_CPUS <= RCU_FANOUT_1
707
-
27 # define RCU_NUM_LVLS 1
708
-
28 # define NUM_RCU_LVL_0 1
709
-
29 # define NUM_RCU_NODES NUM_RCU_LVL_0
710
-
30 # define NUM_RCU_LVL_INIT { NUM_RCU_LVL_0 }
711
-
31 # define RCU_NODE_NAME_INIT { "rcu_node_0" }
712
-
32 # define RCU_FQS_NAME_INIT { "rcu_node_fqs_0" }
713
-
33 # define RCU_EXP_NAME_INIT { "rcu_node_exp_0" }
714
-
34 #elif NR_CPUS <= RCU_FANOUT_2
715
-
35 # define RCU_NUM_LVLS 2
716
-
36 # define NUM_RCU_LVL_0 1
717
-
37 # define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
718
-
38 # define NUM_RCU_NODES (NUM_RCU_LVL_0 + NUM_RCU_LVL_1)
719
-
39 # define NUM_RCU_LVL_INIT { NUM_RCU_LVL_0, NUM_RCU_LVL_1 }
720
-
40 # define RCU_NODE_NAME_INIT { "rcu_node_0", "rcu_node_1" }
721
-
41 # define RCU_FQS_NAME_INIT { "rcu_node_fqs_0", "rcu_node_fqs_1" }
722
-
42 # define RCU_EXP_NAME_INIT { "rcu_node_exp_0", "rcu_node_exp_1" }
723
-
43 #elif NR_CPUS <= RCU_FANOUT_3
724
-
44 # define RCU_NUM_LVLS 3
725
-
45 # define NUM_RCU_LVL_0 1
726
-
46 # define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_2)
727
-
47 # define NUM_RCU_LVL_2 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
728
-
48 # define NUM_RCU_NODES (NUM_RCU_LVL_0 + NUM_RCU_LVL_1 + NUM_RCU_LVL_2)
729
-
49 # define NUM_RCU_LVL_INIT { NUM_RCU_LVL_0, NUM_RCU_LVL_1, NUM_RCU_LVL_2 }
730
-
50 # define RCU_NODE_NAME_INIT { "rcu_node_0", "rcu_node_1", "rcu_node_2" }
731
-
51 # define RCU_FQS_NAME_INIT { "rcu_node_fqs_0", "rcu_node_fqs_1", "rcu_node_fqs_2" }
732
-
52 # define RCU_EXP_NAME_INIT { "rcu_node_exp_0", "rcu_node_exp_1", "rcu_node_exp_2" }
733
-
53 #elif NR_CPUS <= RCU_FANOUT_4
734
-
54 # define RCU_NUM_LVLS 4
735
-
55 # define NUM_RCU_LVL_0 1
736
-
56 # define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_3)
737
-
57 # define NUM_RCU_LVL_2 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_2)
738
-
58 # define NUM_RCU_LVL_3 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
739
-
59 # define NUM_RCU_NODES (NUM_RCU_LVL_0 + NUM_RCU_LVL_1 + NUM_RCU_LVL_2 + NUM_RCU_LVL_3)
740
-
60 # define NUM_RCU_LVL_INIT { NUM_RCU_LVL_0, NUM_RCU_LVL_1, NUM_RCU_LVL_2, NUM_RCU_LVL_3 }
741
-
61 # define RCU_NODE_NAME_INIT { "rcu_node_0", "rcu_node_1", "rcu_node_2", "rcu_node_3" }
742
-
62 # define RCU_FQS_NAME_INIT { "rcu_node_fqs_0", "rcu_node_fqs_1", "rcu_node_fqs_2", "rcu_node_fqs_3" }
743
-
63 # define RCU_EXP_NAME_INIT { "rcu_node_exp_0", "rcu_node_exp_1", "rcu_node_exp_2", "rcu_node_exp_3" }
744
-
64 #else
745
-
65 # error "CONFIG_RCU_FANOUT insufficient for NR_CPUS"
746
-
66 #endif
747
-
</pre>
748
-
749
-
<p>The maximum number of levels in the <tt>rcu_node</tt> structure
750
-
is currently limited to four, as specified by lines 21-24
751
-
and the structure of the subsequent “if” statement.
752
-
For 32-bit systems, this allows 16*32*32*32=524,288 CPUs, which
753
-
should be sufficient for the next few years at least.
754
-
For 64-bit systems, 16*64*64*64=4,194,304 CPUs is allowed, which
755
-
should see us through the next decade or so.
756
-
This four-level tree also allows kernels built with
757
-
<tt>CONFIG_RCU_FANOUT=8</tt> to support up to 4096 CPUs,
758
-
which might be useful in very large systems having eight CPUs per
759
-
socket (but please note that no one has yet shown any measurable
760
-
performance degradation due to misaligned socket and <tt>rcu_node</tt>
761
-
boundaries).
762
-
In addition, building kernels with a full four levels of <tt>rcu_node</tt>
763
-
tree permits better testing of RCU's combining-tree code.
764
-
765
-
</p><p>The <tt>RCU_FANOUT</tt> symbol controls how many children
766
-
are permitted at each non-leaf level of the <tt>rcu_node</tt> tree.
767
-
If the <tt>CONFIG_RCU_FANOUT</tt> Kconfig option is not specified,
768
-
it is set based on the word size of the system, which is also
769
-
the Kconfig default.
770
-
771
-
</p><p>The <tt>RCU_FANOUT_LEAF</tt> symbol controls how many CPUs are
772
-
handled by each leaf <tt>rcu_node</tt> structure.
773
-
Experience has shown that allowing a given leaf <tt>rcu_node</tt>
774
-
structure to handle 64 CPUs, as permitted by the number of bits in
775
-
the <tt>->qsmask</tt> field on a 64-bit system, results in
776
-
excessive contention for the leaf <tt>rcu_node</tt> structures'
777
-
<tt>->lock</tt> fields.
778
-
The number of CPUs per leaf <tt>rcu_node</tt> structure is therefore
779
-
limited to 16 given the default value of <tt>CONFIG_RCU_FANOUT_LEAF</tt>.
780
-
If <tt>CONFIG_RCU_FANOUT_LEAF</tt> is unspecified, the value
781
-
selected is based on the word size of the system, just as for
782
-
<tt>CONFIG_RCU_FANOUT</tt>.
783
-
Lines 11-19 perform this computation.
784
-
785
-
</p><p>Lines 21-24 compute the maximum number of CPUs supported by
786
-
a single-level (which contains a single <tt>rcu_node</tt> structure),
787
-
two-level, three-level, and four-level <tt>rcu_node</tt> tree,
788
-
respectively, given the fanout specified by <tt>RCU_FANOUT</tt>
789
-
and <tt>RCU_FANOUT_LEAF</tt>.
790
-
These numbers of CPUs are retained in the
791
-
<tt>RCU_FANOUT_1</tt>,
792
-
<tt>RCU_FANOUT_2</tt>,
793
-
<tt>RCU_FANOUT_3</tt>, and
794
-
<tt>RCU_FANOUT_4</tt>
795
-
C-preprocessor variables, respectively.
796
-
797
-
</p><p>These variables are used to control the C-preprocessor <tt>#if</tt>
798
-
statement spanning lines 26-66 that computes the number of
799
-
<tt>rcu_node</tt> structures required for each level of the tree,
800
-
as well as the number of levels required.
801
-
The number of levels is placed in the <tt>NUM_RCU_LVLS</tt>
802
-
C-preprocessor variable by lines 27, 35, 44, and 54.
803
-
The number of <tt>rcu_node</tt> structures for the topmost level
804
-
of the tree is always exactly one, and this value is unconditionally
805
-
placed into <tt>NUM_RCU_LVL_0</tt> by lines 28, 36, 45, and 55.
806
-
The rest of the levels (if any) of the <tt>rcu_node</tt> tree
807
-
are computed by dividing the maximum number of CPUs by the
808
-
fanout supported by the number of levels from the current level down,
809
-
rounding up. This computation is performed by lines 37,
810
-
46-47, and 56-58.
811
-
Lines 31-33, 40-42, 50-52, and 62-63 create initializers
812
-
for lockdep lock-class names.
813
-
Finally, lines 64-66 produce an error if the maximum number of
814
-
CPUs is too large for the specified fanout.
815
-
816
-
<h3><a name="The rcu_segcblist Structure">
817
-
The <tt>rcu_segcblist</tt> Structure</a></h3>
818
-
819
-
The <tt>rcu_segcblist</tt> structure maintains a segmented list of
820
-
callbacks as follows:
821
-
822
-
<pre>
823
-
1 #define RCU_DONE_TAIL 0
824
-
2 #define RCU_WAIT_TAIL 1
825
-
3 #define RCU_NEXT_READY_TAIL 2
826
-
4 #define RCU_NEXT_TAIL 3
827
-
5 #define RCU_CBLIST_NSEGS 4
828
-
6
829
-
7 struct rcu_segcblist {
830
-
8 struct rcu_head *head;
831
-
9 struct rcu_head **tails[RCU_CBLIST_NSEGS];
832
-
10 unsigned long gp_seq[RCU_CBLIST_NSEGS];
833
-
11 long len;
834
-
12 long len_lazy;
835
-
13 };
836
-
</pre>
837
-
838
-
<p>
839
-
The segments are as follows:
840
-
841
-
<ol>
842
-
<li> <tt>RCU_DONE_TAIL</tt>: Callbacks whose grace periods have elapsed.
843
-
These callbacks are ready to be invoked.
844
-
<li> <tt>RCU_WAIT_TAIL</tt>: Callbacks that are waiting for the
845
-
current grace period.
846
-
Note that different CPUs can have different ideas about which
847
-
grace period is current, hence the <tt>->gp_seq</tt> field.
848
-
<li> <tt>RCU_NEXT_READY_TAIL</tt>: Callbacks waiting for the next
849
-
grace period to start.
850
-
<li> <tt>RCU_NEXT_TAIL</tt>: Callbacks that have not yet been
851
-
associated with a grace period.
852
-
</ol>
853
-
854
-
<p>
855
-
The <tt>->head</tt> pointer references the first callback or
856
-
is <tt>NULL</tt> if the list contains no callbacks (which is
857
-
<i>not</i> the same as being empty).
858
-
Each element of the <tt>->tails[]</tt> array references the
859
-
<tt>->next</tt> pointer of the last callback in the corresponding
860
-
segment of the list, or the list's <tt>->head</tt> pointer if
861
-
that segment and all previous segments are empty.
862
-
If the corresponding segment is empty but some previous segment is
863
-
not empty, then the array element is identical to its predecessor.
864
-
Older callbacks are closer to the head of the list, and new callbacks
865
-
are added at the tail.
866
-
This relationship between the <tt>->head</tt> pointer, the
867
-
<tt>->tails[]</tt> array, and the callbacks is shown in this
868
-
diagram:
869
-
870
-
</p><p><img src="nxtlist.svg" alt="nxtlist.svg" width="40%">
871
-
872
-
</p><p>In this figure, the <tt>->head</tt> pointer references the
873
-
first
874
-
RCU callback in the list.
875
-
The <tt>->tails[RCU_DONE_TAIL]</tt> array element references
876
-
the <tt>->head</tt> pointer itself, indicating that none
877
-
of the callbacks is ready to invoke.
878
-
The <tt>->tails[RCU_WAIT_TAIL]</tt> array element references callback
879
-
CB 2's <tt>->next</tt> pointer, which indicates that
880
-
CB 1 and CB 2 are both waiting on the current grace period,
881
-
give or take possible disagreements about exactly which grace period
882
-
is the current one.
883
-
The <tt>->tails[RCU_NEXT_READY_TAIL]</tt> array element
884
-
references the same RCU callback that <tt>->tails[RCU_WAIT_TAIL]</tt>
885
-
does, which indicates that there are no callbacks waiting on the next
886
-
RCU grace period.
887
-
The <tt>->tails[RCU_NEXT_TAIL]</tt> array element references
888
-
CB 4's <tt>->next</tt> pointer, indicating that all the
889
-
remaining RCU callbacks have not yet been assigned to an RCU grace
890
-
period.
891
-
Note that the <tt>->tails[RCU_NEXT_TAIL]</tt> array element
892
-
always references the last RCU callback's <tt>->next</tt> pointer
893
-
unless the callback list is empty, in which case it references
894
-
the <tt>->head</tt> pointer.
895
-
896
-
<p>
897
-
There is one additional important special case for the
898
-
<tt>->tails[RCU_NEXT_TAIL]</tt> array element: It can be <tt>NULL</tt>
899
-
when this list is <i>disabled</i>.
900
-
Lists are disabled when the corresponding CPU is offline or when
901
-
the corresponding CPU's callbacks are offloaded to a kthread,
902
-
both of which are described elsewhere.
903
-
904
-
</p><p>CPUs advance their callbacks from the
905
-
<tt>RCU_NEXT_TAIL</tt> to the <tt>RCU_NEXT_READY_TAIL</tt> to the
906
-
<tt>RCU_WAIT_TAIL</tt> to the <tt>RCU_DONE_TAIL</tt> list segments
907
-
as grace periods advance.
908
-
909
-
</p><p>The <tt>->gp_seq[]</tt> array records grace-period
910
-
numbers corresponding to the list segments.
911
-
This is what allows different CPUs to have different ideas as to
912
-
which is the current grace period while still avoiding premature
913
-
invocation of their callbacks.
914
-
In particular, this allows CPUs that go idle for extended periods
915
-
to determine which of their callbacks are ready to be invoked after
916
-
reawakening.
917
-
918
-
</p><p>The <tt>->len</tt> counter contains the number of
919
-
callbacks in <tt>->head</tt>, and the
920
-
<tt>->len_lazy</tt> contains the number of those callbacks that
921
-
are known to only free memory, and whose invocation can therefore
922
-
be safely deferred.
923
-
924
-
<p><b>Important note</b>: It is the <tt>->len</tt> field that
925
-
determines whether or not there are callbacks associated with
926
-
this <tt>rcu_segcblist</tt> structure, <i>not</i> the <tt>->head</tt>
927
-
pointer.
928
-
The reason for this is that all the ready-to-invoke callbacks
929
-
(that is, those in the <tt>RCU_DONE_TAIL</tt> segment) are extracted
930
-
all at once at callback-invocation time (<tt>rcu_do_batch</tt>), due
931
-
to which <tt>->head</tt> may be set to NULL if there are no not-done
932
-
callbacks remaining in the <tt>rcu_segcblist</tt>.
933
-
If callback invocation must be postponed, for example, because a
934
-
high-priority process just woke up on this CPU, then the remaining
935
-
callbacks are placed back on the <tt>RCU_DONE_TAIL</tt> segment and
936
-
<tt>->head</tt> once again points to the start of the segment.
937
-
In short, the head field can briefly be <tt>NULL</tt> even though the
938
-
CPU has callbacks present the entire time.
939
-
Therefore, it is not appropriate to test the <tt>->head</tt> pointer
940
-
for <tt>NULL</tt>.
941
-
942
-
<p>In contrast, the <tt>->len</tt> and <tt>->len_lazy</tt> counts
943
-
are adjusted only after the corresponding callbacks have been invoked.
944
-
This means that the <tt>->len</tt> count is zero only if
945
-
the <tt>rcu_segcblist</tt> structure really is devoid of callbacks.
946
-
Of course, off-CPU sampling of the <tt>->len</tt> count requires
947
-
careful use of appropriate synchronization, for example, memory barriers.
948
-
This synchronization can be a bit subtle, particularly in the case
949
-
of <tt>rcu_barrier()</tt>.
950
-
951
-
<h3><a name="The rcu_data Structure">
952
-
The <tt>rcu_data</tt> Structure</a></h3>
953
-
954
-
<p>The <tt>rcu_data</tt> maintains the per-CPU state for the RCU subsystem.
955
-
The fields in this structure may be accessed only from the corresponding
956
-
CPU (and from tracing) unless otherwise stated.
957
-
This structure is the
958
-
focus of quiescent-state detection and RCU callback queuing.
959
-
It also tracks its relationship to the corresponding leaf
960
-
<tt>rcu_node</tt> structure to allow more-efficient
961
-
propagation of quiescent states up the <tt>rcu_node</tt>
962
-
combining tree.
963
-
Like the <tt>rcu_node</tt> structure, it provides a local
964
-
copy of the grace-period information to allow for-free
965
-
synchronized
966
-
access to this information from the corresponding CPU.
967
-
Finally, this structure records past dyntick-idle state
968
-
for the corresponding CPU and also tracks statistics.
969
-
970
-
</p><p>The <tt>rcu_data</tt> structure's fields are discussed,
971
-
singly and in groups, in the following sections.
972
-
973
-
<h5>Connection to Other Data Structures</h5>
974
-
975
-
<p>This portion of the <tt>rcu_data</tt> structure is declared
976
-
as follows:
977
-
978
-
<pre>
979
-
1 int cpu;
980
-
2 struct rcu_node *mynode;
981
-
3 unsigned long grpmask;
982
-
4 bool beenonline;
983
-
</pre>
984
-
985
-
<p>The <tt>->cpu</tt> field contains the number of the
986
-
corresponding CPU and the <tt>->mynode</tt> field references the
987
-
corresponding <tt>rcu_node</tt> structure.
988
-
The <tt>->mynode</tt> is used to propagate quiescent states
989
-
up the combining tree.
990
-
These two fields are constant and therefore do not require synchronization.
991
-
992
-
<p>The <tt>->grpmask</tt> field indicates the bit in
993
-
the <tt>->mynode->qsmask</tt> corresponding to this
994
-
<tt>rcu_data</tt> structure, and is also used when propagating
995
-
quiescent states.
996
-
The <tt>->beenonline</tt> flag is set whenever the corresponding
997
-
CPU comes online, which means that the debugfs tracing need not dump
998
-
out any <tt>rcu_data</tt> structure for which this flag is not set.
999
-
1000
-
<h5>Quiescent-State and Grace-Period Tracking</h5>
1001
-
1002
-
<p>This portion of the <tt>rcu_data</tt> structure is declared
1003
-
as follows:
1004
-
1005
-
<pre>
1006
-
1 unsigned long gp_seq;
1007
-
2 unsigned long gp_seq_needed;
1008
-
3 bool cpu_no_qs;
1009
-
4 bool core_needs_qs;
1010
-
5 bool gpwrap;
1011
-
</pre>
1012
-
1013
-
<p>The <tt>->gp_seq</tt> field is the counterpart of the field of the same
1014
-
name in the <tt>rcu_state</tt> and <tt>rcu_node</tt> structures. The
1015
-
<tt>->gp_seq_needed</tt> field is the counterpart of the field of the same
1016
-
name in the rcu_node</tt> structure.
1017
-
They may each lag up to one behind their <tt>rcu_node</tt>
1018
-
counterparts, but in <tt>CONFIG_NO_HZ_IDLE</tt> and
1019
-
<tt>CONFIG_NO_HZ_FULL</tt> kernels can lag
1020
-
arbitrarily far behind for CPUs in dyntick-idle mode (but these counters
1021
-
will catch up upon exit from dyntick-idle mode).
1022
-
If the lower two bits of a given <tt>rcu_data</tt> structure's
1023
-
<tt>->gp_seq</tt> are zero, then this <tt>rcu_data</tt>
1024
-
structure believes that RCU is idle.
1025
-
1026
-
<table>
1027
-
<tr><th> </th></tr>
1028
-
<tr><th align="left">Quick Quiz:</th></tr>
1029
-
<tr><td>
1030
-
All this replication of the grace period numbers can only cause
1031
-
massive confusion.
1032
-
Why not just keep a global sequence number and be done with it???
1033
-
</td></tr>
1034
-
<tr><th align="left">Answer:</th></tr>
1035
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
1036
-
Because if there was only a single global sequence
1037
-
numbers, there would need to be a single global lock to allow
1038
-
safely accessing and updating it.
1039
-
And if we are not going to have a single global lock, we need
1040
-
to carefully manage the numbers on a per-node basis.
1041
-
Recall from the answer to a previous Quick Quiz that the consequences
1042
-
of applying a previously sampled quiescent state to the wrong
1043
-
grace period are quite severe.
1044
-
</font></td></tr>
1045
-
<tr><td> </td></tr>
1046
-
</table>
1047
-
1048
-
<p>The <tt>->cpu_no_qs</tt> flag indicates that the
1049
-
CPU has not yet passed through a quiescent state,
1050
-
while the <tt>->core_needs_qs</tt> flag indicates that the
1051
-
RCU core needs a quiescent state from the corresponding CPU.
1052
-
The <tt>->gpwrap</tt> field indicates that the corresponding
1053
-
CPU has remained idle for so long that the
1054
-
<tt>gp_seq</tt> counter is in danger of overflow, which
1055
-
will cause the CPU to disregard the values of its counters on
1056
-
its next exit from idle.
1057
-
1058
-
<h5>RCU Callback Handling</h5>
1059
-
1060
-
<p>In the absence of CPU-hotplug events, RCU callbacks are invoked by
1061
-
the same CPU that registered them.
1062
-
This is strictly a cache-locality optimization: callbacks can and
1063
-
do get invoked on CPUs other than the one that registered them.
1064
-
After all, if the CPU that registered a given callback has gone
1065
-
offline before the callback can be invoked, there really is no other
1066
-
choice.
1067
-
1068
-
</p><p>This portion of the <tt>rcu_data</tt> structure is declared
1069
-
as follows:
1070
-
1071
-
<pre>
1072
-
1 struct rcu_segcblist cblist;
1073
-
2 long qlen_last_fqs_check;
1074
-
3 unsigned long n_cbs_invoked;
1075
-
4 unsigned long n_nocbs_invoked;
1076
-
5 unsigned long n_cbs_orphaned;
1077
-
6 unsigned long n_cbs_adopted;
1078
-
7 unsigned long n_force_qs_snap;
1079
-
8 long blimit;
1080
-
</pre>
1081
-
1082
-
<p>The <tt>->cblist</tt> structure is the segmented callback list
1083
-
described earlier.
1084
-
The CPU advances the callbacks in its <tt>rcu_data</tt> structure
1085
-
whenever it notices that another RCU grace period has completed.
1086
-
The CPU detects the completion of an RCU grace period by noticing
1087
-
that the value of its <tt>rcu_data</tt> structure's
1088
-
<tt>->gp_seq</tt> field differs from that of its leaf
1089
-
<tt>rcu_node</tt> structure.
1090
-
Recall that each <tt>rcu_node</tt> structure's
1091
-
<tt>->gp_seq</tt> field is updated at the beginnings and ends of each
1092
-
grace period.
1093
-
1094
-
<p>
1095
-
The <tt>->qlen_last_fqs_check</tt> and
1096
-
<tt>->n_force_qs_snap</tt> coordinate the forcing of quiescent
1097
-
states from <tt>call_rcu()</tt> and friends when callback
1098
-
lists grow excessively long.
1099
-
1100
-
</p><p>The <tt>->n_cbs_invoked</tt>,
1101
-
<tt>->n_cbs_orphaned</tt>, and <tt>->n_cbs_adopted</tt>
1102
-
fields count the number of callbacks invoked,
1103
-
sent to other CPUs when this CPU goes offline,
1104
-
and received from other CPUs when those other CPUs go offline.
1105
-
The <tt>->n_nocbs_invoked</tt> is used when the CPU's callbacks
1106
-
are offloaded to a kthread.
1107
-
1108
-
<p>
1109
-
Finally, the <tt>->blimit</tt> counter is the maximum number of
1110
-
RCU callbacks that may be invoked at a given time.
1111
-
1112
-
<h5>Dyntick-Idle Handling</h5>
1113
-
1114
-
<p>This portion of the <tt>rcu_data</tt> structure is declared
1115
-
as follows:
1116
-
1117
-
<pre>
1118
-
1 int dynticks_snap;
1119
-
2 unsigned long dynticks_fqs;
1120
-
</pre>
1121
-
1122
-
The <tt>->dynticks_snap</tt> field is used to take a snapshot
1123
-
of the corresponding CPU's dyntick-idle state when forcing
1124
-
quiescent states, and is therefore accessed from other CPUs.
1125
-
Finally, the <tt>->dynticks_fqs</tt> field is used to
1126
-
count the number of times this CPU is determined to be in
1127
-
dyntick-idle state, and is used for tracing and debugging purposes.
1128
-
1129
-
<p>
1130
-
This portion of the rcu_data structure is declared as follows:
1131
-
1132
-
<pre>
1133
-
1 long dynticks_nesting;
1134
-
2 long dynticks_nmi_nesting;
1135
-
3 atomic_t dynticks;
1136
-
4 bool rcu_need_heavy_qs;
1137
-
5 bool rcu_urgent_qs;
1138
-
</pre>
1139
-
1140
-
<p>These fields in the rcu_data structure maintain the per-CPU dyntick-idle
1141
-
state for the corresponding CPU.
1142
-
The fields may be accessed only from the corresponding CPU (and from tracing)
1143
-
unless otherwise stated.
1144
-
1145
-
<p>The <tt>->dynticks_nesting</tt> field counts the
1146
-
nesting depth of process execution, so that in normal circumstances
1147
-
this counter has value zero or one.
1148
-
NMIs, irqs, and tracers are counted by the <tt>->dynticks_nmi_nesting</tt>
1149
-
field.
1150
-
Because NMIs cannot be masked, changes to this variable have to be
1151
-
undertaken carefully using an algorithm provided by Andy Lutomirski.
1152
-
The initial transition from idle adds one, and nested transitions
1153
-
add two, so that a nesting level of five is represented by a
1154
-
<tt>->dynticks_nmi_nesting</tt> value of nine.
1155
-
This counter can therefore be thought of as counting the number
1156
-
of reasons why this CPU cannot be permitted to enter dyntick-idle
1157
-
mode, aside from process-level transitions.
1158
-
1159
-
<p>However, it turns out that when running in non-idle kernel context,
1160
-
the Linux kernel is fully capable of entering interrupt handlers that
1161
-
never exit and perhaps also vice versa.
1162
-
Therefore, whenever the <tt>->dynticks_nesting</tt> field is
1163
-
incremented up from zero, the <tt>->dynticks_nmi_nesting</tt> field
1164
-
is set to a large positive number, and whenever the
1165
-
<tt>->dynticks_nesting</tt> field is decremented down to zero,
1166
-
the the <tt>->dynticks_nmi_nesting</tt> field is set to zero.
1167
-
Assuming that the number of misnested interrupts is not sufficient
1168
-
to overflow the counter, this approach corrects the
1169
-
<tt>->dynticks_nmi_nesting</tt> field every time the corresponding
1170
-
CPU enters the idle loop from process context.
1171
-
1172
-
</p><p>The <tt>->dynticks</tt> field counts the corresponding
1173
-
CPU's transitions to and from either dyntick-idle or user mode, so
1174
-
that this counter has an even value when the CPU is in dyntick-idle
1175
-
mode or user mode and an odd value otherwise. The transitions to/from
1176
-
user mode need to be counted for user mode adaptive-ticks support
1177
-
(see timers/NO_HZ.txt).
1178
-
1179
-
</p><p>The <tt>->rcu_need_heavy_qs</tt> field is used
1180
-
to record the fact that the RCU core code would really like to
1181
-
see a quiescent state from the corresponding CPU, so much so that
1182
-
it is willing to call for heavy-weight dyntick-counter operations.
1183
-
This flag is checked by RCU's context-switch and <tt>cond_resched()</tt>
1184
-
code, which provide a momentary idle sojourn in response.
1185
-
1186
-
</p><p>Finally, the <tt>->rcu_urgent_qs</tt> field is used to record
1187
-
the fact that the RCU core code would really like to see a quiescent state from
1188
-
the corresponding CPU, with the various other fields indicating just how badly
1189
-
RCU wants this quiescent state.
1190
-
This flag is checked by RCU's context-switch path
1191
-
(<tt>rcu_note_context_switch</tt>) and the cond_resched code.
1192
-
1193
-
<table>
1194
-
<tr><th> </th></tr>
1195
-
<tr><th align="left">Quick Quiz:</th></tr>
1196
-
<tr><td>
1197
-
Why not simply combine the <tt>->dynticks_nesting</tt>
1198
-
and <tt>->dynticks_nmi_nesting</tt> counters into a
1199
-
single counter that just counts the number of reasons that
1200
-
the corresponding CPU is non-idle?
1201
-
</td></tr>
1202
-
<tr><th align="left">Answer:</th></tr>
1203
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
1204
-
Because this would fail in the presence of interrupts whose
1205
-
handlers never return and of handlers that manage to return
1206
-
from a made-up interrupt.
1207
-
</font></td></tr>
1208
-
<tr><td> </td></tr>
1209
-
</table>
1210
-
1211
-
<p>Additional fields are present for some special-purpose
1212
-
builds, and are discussed separately.
1213
-
1214
-
<h3><a name="The rcu_head Structure">
1215
-
The <tt>rcu_head</tt> Structure</a></h3>
1216
-
1217
-
<p>Each <tt>rcu_head</tt> structure represents an RCU callback.
1218
-
These structures are normally embedded within RCU-protected data
1219
-
structures whose algorithms use asynchronous grace periods.
1220
-
In contrast, when using algorithms that block waiting for RCU grace periods,
1221
-
RCU users need not provide <tt>rcu_head</tt> structures.
1222
-
1223
-
</p><p>The <tt>rcu_head</tt> structure has fields as follows:
1224
-
1225
-
<pre>
1226
-
1 struct rcu_head *next;
1227
-
2 void (*func)(struct rcu_head *head);
1228
-
</pre>
1229
-
1230
-
<p>The <tt>->next</tt> field is used
1231
-
to link the <tt>rcu_head</tt> structures together in the
1232
-
lists within the <tt>rcu_data</tt> structures.
1233
-
The <tt>->func</tt> field is a pointer to the function
1234
-
to be called when the callback is ready to be invoked, and
1235
-
this function is passed a pointer to the <tt>rcu_head</tt>
1236
-
structure.
1237
-
However, <tt>kfree_rcu()</tt> uses the <tt>->func</tt>
1238
-
field to record the offset of the <tt>rcu_head</tt>
1239
-
structure within the enclosing RCU-protected data structure.
1240
-
1241
-
</p><p>Both of these fields are used internally by RCU.
1242
-
From the viewpoint of RCU users, this structure is an
1243
-
opaque “cookie”.
1244
-
1245
-
<table>
1246
-
<tr><th> </th></tr>
1247
-
<tr><th align="left">Quick Quiz:</th></tr>
1248
-
<tr><td>
1249
-
Given that the callback function <tt>->func</tt>
1250
-
is passed a pointer to the <tt>rcu_head</tt> structure,
1251
-
how is that function supposed to find the beginning of the
1252
-
enclosing RCU-protected data structure?
1253
-
</td></tr>
1254
-
<tr><th align="left">Answer:</th></tr>
1255
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
1256
-
In actual practice, there is a separate callback function per
1257
-
type of RCU-protected data structure.
1258
-
The callback function can therefore use the <tt>container_of()</tt>
1259
-
macro in the Linux kernel (or other pointer-manipulation facilities
1260
-
in other software environments) to find the beginning of the
1261
-
enclosing structure.
1262
-
</font></td></tr>
1263
-
<tr><td> </td></tr>
1264
-
</table>
1265
-
1266
-
<h3><a name="RCU-Specific Fields in the task_struct Structure">
1267
-
RCU-Specific Fields in the <tt>task_struct</tt> Structure</a></h3>
1268
-
1269
-
<p>The <tt>CONFIG_PREEMPT_RCU</tt> implementation uses some
1270
-
additional fields in the <tt>task_struct</tt> structure:
1271
-
1272
-
<pre>
1273
-
1 #ifdef CONFIG_PREEMPT_RCU
1274
-
2 int rcu_read_lock_nesting;
1275
-
3 union rcu_special rcu_read_unlock_special;
1276
-
4 struct list_head rcu_node_entry;
1277
-
5 struct rcu_node *rcu_blocked_node;
1278
-
6 #endif /* #ifdef CONFIG_PREEMPT_RCU */
1279
-
7 #ifdef CONFIG_TASKS_RCU
1280
-
8 unsigned long rcu_tasks_nvcsw;
1281
-
9 bool rcu_tasks_holdout;
1282
-
10 struct list_head rcu_tasks_holdout_list;
1283
-
11 int rcu_tasks_idle_cpu;
1284
-
12 #endif /* #ifdef CONFIG_TASKS_RCU */
1285
-
</pre>
1286
-
1287
-
<p>The <tt>->rcu_read_lock_nesting</tt> field records the
1288
-
nesting level for RCU read-side critical sections, and
1289
-
the <tt>->rcu_read_unlock_special</tt> field is a bitmask
1290
-
that records special conditions that require <tt>rcu_read_unlock()</tt>
1291
-
to do additional work.
1292
-
The <tt>->rcu_node_entry</tt> field is used to form lists of
1293
-
tasks that have blocked within preemptible-RCU read-side critical
1294
-
sections and the <tt>->rcu_blocked_node</tt> field references
1295
-
the <tt>rcu_node</tt> structure whose list this task is a member of,
1296
-
or <tt>NULL</tt> if it is not blocked within a preemptible-RCU
1297
-
read-side critical section.
1298
-
1299
-
<p>The <tt>->rcu_tasks_nvcsw</tt> field tracks the number of
1300
-
voluntary context switches that this task had undergone at the
1301
-
beginning of the current tasks-RCU grace period,
1302
-
<tt>->rcu_tasks_holdout</tt> is set if the current tasks-RCU
1303
-
grace period is waiting on this task, <tt>->rcu_tasks_holdout_list</tt>
1304
-
is a list element enqueuing this task on the holdout list,
1305
-
and <tt>->rcu_tasks_idle_cpu</tt> tracks which CPU this
1306
-
idle task is running, but only if the task is currently running,
1307
-
that is, if the CPU is currently idle.
1308
-
1309
-
<h3><a name="Accessor Functions">
1310
-
Accessor Functions</a></h3>
1311
-
1312
-
<p>The following listing shows the
1313
-
<tt>rcu_get_root()</tt>, <tt>rcu_for_each_node_breadth_first</tt> and
1314
-
<tt>rcu_for_each_leaf_node()</tt> function and macros:
1315
-
1316
-
<pre>
1317
-
1 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
1318
-
2 {
1319
-
3 return &rsp->node[0];
1320
-
4 }
1321
-
5
1322
-
6 #define rcu_for_each_node_breadth_first(rsp, rnp) \
1323
-
7 for ((rnp) = &(rsp)->node[0]; \
1324
-
8 (rnp) < &(rsp)->node[NUM_RCU_NODES]; (rnp)++)
1325
-
9
1326
-
10 #define rcu_for_each_leaf_node(rsp, rnp) \
1327
-
11 for ((rnp) = (rsp)->level[NUM_RCU_LVLS - 1]; \
1328
-
12 (rnp) < &(rsp)->node[NUM_RCU_NODES]; (rnp)++)
1329
-
</pre>
1330
-
1331
-
<p>The <tt>rcu_get_root()</tt> simply returns a pointer to the
1332
-
first element of the specified <tt>rcu_state</tt> structure's
1333
-
<tt>->node[]</tt> array, which is the root <tt>rcu_node</tt>
1334
-
structure.
1335
-
1336
-
</p><p>As noted earlier, the <tt>rcu_for_each_node_breadth_first()</tt>
1337
-
macro takes advantage of the layout of the <tt>rcu_node</tt>
1338
-
structures in the <tt>rcu_state</tt> structure's
1339
-
<tt>->node[]</tt> array, performing a breadth-first traversal by
1340
-
simply traversing the array in order.
1341
-
Similarly, the <tt>rcu_for_each_leaf_node()</tt> macro traverses only
1342
-
the last part of the array, thus traversing only the leaf
1343
-
<tt>rcu_node</tt> structures.
1344
-
1345
-
<table>
1346
-
<tr><th> </th></tr>
1347
-
<tr><th align="left">Quick Quiz:</th></tr>
1348
-
<tr><td>
1349
-
What does
1350
-
<tt>rcu_for_each_leaf_node()</tt> do if the <tt>rcu_node</tt> tree
1351
-
contains only a single node?
1352
-
</td></tr>
1353
-
<tr><th align="left">Answer:</th></tr>
1354
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
1355
-
In the single-node case,
1356
-
<tt>rcu_for_each_leaf_node()</tt> traverses the single node.
1357
-
</font></td></tr>
1358
-
<tr><td> </td></tr>
1359
-
</table>
1360
-
1361
-
<h3><a name="Summary">
1362
-
Summary</a></h3>
1363
-
1364
-
So the state of RCU is represented by an <tt>rcu_state</tt> structure,
1365
-
which contains a combining tree of <tt>rcu_node</tt> and
1366
-
<tt>rcu_data</tt> structures.
1367
-
Finally, in <tt>CONFIG_NO_HZ_IDLE</tt> kernels, each CPU's dyntick-idle
1368
-
state is tracked by dynticks-related fields in the <tt>rcu_data</tt> structure.
1369
-
1370
-
If you made it this far, you are well prepared to read the code
1371
-
walkthroughs in the other articles in this series.
1372
-
1373
-
<h3><a name="Acknowledgments">
1374
-
Acknowledgments</a></h3>
1375
-
1376
-
I owe thanks to Cyrill Gorcunov, Mathieu Desnoyers, Dhaval Giani, Paul
1377
-
Turner, Abhishek Srivastava, Matt Kowalczyk, and Serge Hallyn
1378
-
for helping me get this document into a more human-readable state.
1379
-
1380
-
<h3><a name="Legal Statement">
1381
-
Legal Statement</a></h3>
1382
-
1383
-
<p>This work represents the view of the author and does not necessarily
1384
-
represent the view of IBM.
1385
-
1386
-
</p><p>Linux is a registered trademark of Linus Torvalds.
1387
-
1388
-
</p><p>Other company, product, and service names may be trademarks or
1389
-
service marks of others.
1390
-
1391
-
</body></html>
+1163
Documentation/RCU/Design/Data-Structures/Data-Structures.rst
+1163
Documentation/RCU/Design/Data-Structures/Data-Structures.rst
···
1
+
===================================================
2
+
A Tour Through TREE_RCU's Data Structures [LWN.net]
3
+
===================================================
4
+
5
+
December 18, 2016
6
+
7
+
This article was contributed by Paul E. McKenney
8
+
9
+
Introduction
10
+
============
11
+
12
+
This document describes RCU's major data structures and their relationship
13
+
to each other.
14
+
15
+
Data-Structure Relationships
16
+
============================
17
+
18
+
RCU is for all intents and purposes a large state machine, and its
19
+
data structures maintain the state in such a way as to allow RCU readers
20
+
to execute extremely quickly, while also processing the RCU grace periods
21
+
requested by updaters in an efficient and extremely scalable fashion.
22
+
The efficiency and scalability of RCU updaters is provided primarily
23
+
by a combining tree, as shown below:
24
+
25
+
.. kernel-figure:: BigTreeClassicRCU.svg
26
+
27
+
This diagram shows an enclosing ``rcu_state`` structure containing a tree
28
+
of ``rcu_node`` structures. Each leaf node of the ``rcu_node`` tree has up
29
+
to 16 ``rcu_data`` structures associated with it, so that there are
30
+
``NR_CPUS`` number of ``rcu_data`` structures, one for each possible CPU.
31
+
This structure is adjusted at boot time, if needed, to handle the common
32
+
case where ``nr_cpu_ids`` is much less than ``NR_CPUs``.
33
+
For example, a number of Linux distributions set ``NR_CPUs=4096``,
34
+
which results in a three-level ``rcu_node`` tree.
35
+
If the actual hardware has only 16 CPUs, RCU will adjust itself
36
+
at boot time, resulting in an ``rcu_node`` tree with only a single node.
37
+
38
+
The purpose of this combining tree is to allow per-CPU events
39
+
such as quiescent states, dyntick-idle transitions,
40
+
and CPU hotplug operations to be processed efficiently
41
+
and scalably.
42
+
Quiescent states are recorded by the per-CPU ``rcu_data`` structures,
43
+
and other events are recorded by the leaf-level ``rcu_node``
44
+
structures.
45
+
All of these events are combined at each level of the tree until finally
46
+
grace periods are completed at the tree's root ``rcu_node``
47
+
structure.
48
+
A grace period can be completed at the root once every CPU
49
+
(or, in the case of ``CONFIG_PREEMPT_RCU``, task)
50
+
has passed through a quiescent state.
51
+
Once a grace period has completed, record of that fact is propagated
52
+
back down the tree.
53
+
54
+
As can be seen from the diagram, on a 64-bit system
55
+
a two-level tree with 64 leaves can accommodate 1,024 CPUs, with a fanout
56
+
of 64 at the root and a fanout of 16 at the leaves.
57
+
58
+
+-----------------------------------------------------------------------+
59
+
| **Quick Quiz**: |
60
+
+-----------------------------------------------------------------------+
61
+
| Why isn't the fanout at the leaves also 64? |
62
+
+-----------------------------------------------------------------------+
63
+
| **Answer**: |
64
+
+-----------------------------------------------------------------------+
65
+
| Because there are more types of events that affect the leaf-level |
66
+
| ``rcu_node`` structures than further up the tree. Therefore, if the |
67
+
| leaf ``rcu_node`` structures have fanout of 64, the contention on |
68
+
| these structures' ``->structures`` becomes excessive. Experimentation |
69
+
| on a wide variety of systems has shown that a fanout of 16 works well |
70
+
| for the leaves of the ``rcu_node`` tree. |
71
+
| |
72
+
| Of course, further experience with systems having hundreds or |
73
+
| thousands of CPUs may demonstrate that the fanout for the non-leaf |
74
+
| ``rcu_node`` structures must also be reduced. Such reduction can be |
75
+
| easily carried out when and if it proves necessary. In the meantime, |
76
+
| if you are using such a system and running into contention problems |
77
+
| on the non-leaf ``rcu_node`` structures, you may use the |
78
+
| ``CONFIG_RCU_FANOUT`` kernel configuration parameter to reduce the |
79
+
| non-leaf fanout as needed. |
80
+
| |
81
+
| Kernels built for systems with strong NUMA characteristics might |
82
+
| also need to adjust ``CONFIG_RCU_FANOUT`` so that the domains of |
83
+
| the ``rcu_node`` structures align with hardware boundaries. |
84
+
| However, there has thus far been no need for this. |
85
+
+-----------------------------------------------------------------------+
86
+
87
+
If your system has more than 1,024 CPUs (or more than 512 CPUs on a
88
+
32-bit system), then RCU will automatically add more levels to the tree.
89
+
For example, if you are crazy enough to build a 64-bit system with
90
+
65,536 CPUs, RCU would configure the ``rcu_node`` tree as follows:
91
+
92
+
.. kernel-figure:: HugeTreeClassicRCU.svg
93
+
94
+
RCU currently permits up to a four-level tree, which on a 64-bit system
95
+
accommodates up to 4,194,304 CPUs, though only a mere 524,288 CPUs for
96
+
32-bit systems. On the other hand, you can set both
97
+
``CONFIG_RCU_FANOUT`` and ``CONFIG_RCU_FANOUT_LEAF`` to be as small as
98
+
2, which would result in a 16-CPU test using a 4-level tree. This can be
99
+
useful for testing large-system capabilities on small test machines.
100
+
101
+
This multi-level combining tree allows us to get most of the performance
102
+
and scalability benefits of partitioning, even though RCU grace-period
103
+
detection is inherently a global operation. The trick here is that only
104
+
the last CPU to report a quiescent state into a given ``rcu_node``
105
+
structure need advance to the ``rcu_node`` structure at the next level
106
+
up the tree. This means that at the leaf-level ``rcu_node`` structure,
107
+
only one access out of sixteen will progress up the tree. For the
108
+
internal ``rcu_node`` structures, the situation is even more extreme:
109
+
Only one access out of sixty-four will progress up the tree. Because the
110
+
vast majority of the CPUs do not progress up the tree, the lock
111
+
contention remains roughly constant up the tree. No matter how many CPUs
112
+
there are in the system, at most 64 quiescent-state reports per grace
113
+
period will progress all the way to the root ``rcu_node`` structure,
114
+
thus ensuring that the lock contention on that root ``rcu_node``
115
+
structure remains acceptably low.
116
+
117
+
In effect, the combining tree acts like a big shock absorber, keeping
118
+
lock contention under control at all tree levels regardless of the level
119
+
of loading on the system.
120
+
121
+
RCU updaters wait for normal grace periods by registering RCU callbacks,
122
+
either directly via ``call_rcu()`` or indirectly via
123
+
``synchronize_rcu()`` and friends. RCU callbacks are represented by
124
+
``rcu_head`` structures, which are queued on ``rcu_data`` structures
125
+
while they are waiting for a grace period to elapse, as shown in the
126
+
following figure:
127
+
128
+
.. kernel-figure:: BigTreePreemptRCUBHdyntickCB.svg
129
+
130
+
This figure shows how ``TREE_RCU``'s and ``PREEMPT_RCU``'s major data
131
+
structures are related. Lesser data structures will be introduced with
132
+
the algorithms that make use of them.
133
+
134
+
Note that each of the data structures in the above figure has its own
135
+
synchronization:
136
+
137
+
#. Each ``rcu_state`` structures has a lock and a mutex, and some fields
138
+
are protected by the corresponding root ``rcu_node`` structure's lock.
139
+
#. Each ``rcu_node`` structure has a spinlock.
140
+
#. The fields in ``rcu_data`` are private to the corresponding CPU,
141
+
although a few can be read and written by other CPUs.
142
+
143
+
It is important to note that different data structures can have very
144
+
different ideas about the state of RCU at any given time. For but one
145
+
example, awareness of the start or end of a given RCU grace period
146
+
propagates slowly through the data structures. This slow propagation is
147
+
absolutely necessary for RCU to have good read-side performance. If this
148
+
balkanized implementation seems foreign to you, one useful trick is to
149
+
consider each instance of these data structures to be a different
150
+
person, each having the usual slightly different view of reality.
151
+
152
+
The general role of each of these data structures is as follows:
153
+
154
+
#. ``rcu_state``: This structure forms the interconnection between the
155
+
``rcu_node`` and ``rcu_data`` structures, tracks grace periods,
156
+
serves as short-term repository for callbacks orphaned by CPU-hotplug
157
+
events, maintains ``rcu_barrier()`` state, tracks expedited
158
+
grace-period state, and maintains state used to force quiescent
159
+
states when grace periods extend too long,
160
+
#. ``rcu_node``: This structure forms the combining tree that propagates
161
+
quiescent-state information from the leaves to the root, and also
162
+
propagates grace-period information from the root to the leaves. It
163
+
provides local copies of the grace-period state in order to allow
164
+
this information to be accessed in a synchronized manner without
165
+
suffering the scalability limitations that would otherwise be imposed
166
+
by global locking. In ``CONFIG_PREEMPT_RCU`` kernels, it manages the
167
+
lists of tasks that have blocked while in their current RCU read-side
168
+
critical section. In ``CONFIG_PREEMPT_RCU`` with
169
+
``CONFIG_RCU_BOOST``, it manages the per-\ ``rcu_node``
170
+
priority-boosting kernel threads (kthreads) and state. Finally, it
171
+
records CPU-hotplug state in order to determine which CPUs should be
172
+
ignored during a given grace period.
173
+
#. ``rcu_data``: This per-CPU structure is the focus of quiescent-state
174
+
detection and RCU callback queuing. It also tracks its relationship
175
+
to the corresponding leaf ``rcu_node`` structure to allow
176
+
more-efficient propagation of quiescent states up the ``rcu_node``
177
+
combining tree. Like the ``rcu_node`` structure, it provides a local
178
+
copy of the grace-period information to allow for-free synchronized
179
+
access to this information from the corresponding CPU. Finally, this
180
+
structure records past dyntick-idle state for the corresponding CPU
181
+
and also tracks statistics.
182
+
#. ``rcu_head``: This structure represents RCU callbacks, and is the
183
+
only structure allocated and managed by RCU users. The ``rcu_head``
184
+
structure is normally embedded within the RCU-protected data
185
+
structure.
186
+
187
+
If all you wanted from this article was a general notion of how RCU's
188
+
data structures are related, you are done. Otherwise, each of the
189
+
following sections give more details on the ``rcu_state``, ``rcu_node``
190
+
and ``rcu_data`` data structures.
191
+
192
+
The ``rcu_state`` Structure
193
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~
194
+
195
+
The ``rcu_state`` structure is the base structure that represents the
196
+
state of RCU in the system. This structure forms the interconnection
197
+
between the ``rcu_node`` and ``rcu_data`` structures, tracks grace
198
+
periods, contains the lock used to synchronize with CPU-hotplug events,
199
+
and maintains state used to force quiescent states when grace periods
200
+
extend too long,
201
+
202
+
A few of the ``rcu_state`` structure's fields are discussed, singly and
203
+
in groups, in the following sections. The more specialized fields are
204
+
covered in the discussion of their use.
205
+
206
+
Relationship to rcu_node and rcu_data Structures
207
+
''''''''''''''''''''''''''''''''''''''''''''''''
208
+
209
+
This portion of the ``rcu_state`` structure is declared as follows:
210
+
211
+
::
212
+
213
+
1 struct rcu_node node[NUM_RCU_NODES];
214
+
2 struct rcu_node *level[NUM_RCU_LVLS + 1];
215
+
3 struct rcu_data __percpu *rda;
216
+
217
+
+-----------------------------------------------------------------------+
218
+
| **Quick Quiz**: |
219
+
+-----------------------------------------------------------------------+
220
+
| Wait a minute! You said that the ``rcu_node`` structures formed a |
221
+
| tree, but they are declared as a flat array! What gives? |
222
+
+-----------------------------------------------------------------------+
223
+
| **Answer**: |
224
+
+-----------------------------------------------------------------------+
225
+
| The tree is laid out in the array. The first node In the array is the |
226
+
| head, the next set of nodes in the array are children of the head |
227
+
| node, and so on until the last set of nodes in the array are the |
228
+
| leaves. |
229
+
| See the following diagrams to see how this works. |
230
+
+-----------------------------------------------------------------------+
231
+
232
+
The ``rcu_node`` tree is embedded into the ``->node[]`` array as shown
233
+
in the following figure:
234
+
235
+
.. kernel-figure:: TreeMapping.svg
236
+
237
+
One interesting consequence of this mapping is that a breadth-first
238
+
traversal of the tree is implemented as a simple linear scan of the
239
+
array, which is in fact what the ``rcu_for_each_node_breadth_first()``
240
+
macro does. This macro is used at the beginning and ends of grace
241
+
periods.
242
+
243
+
Each entry of the ``->level`` array references the first ``rcu_node``
244
+
structure on the corresponding level of the tree, for example, as shown
245
+
below:
246
+
247
+
.. kernel-figure:: TreeMappingLevel.svg
248
+
249
+
The zero\ :sup:`th` element of the array references the root
250
+
``rcu_node`` structure, the first element references the first child of
251
+
the root ``rcu_node``, and finally the second element references the
252
+
first leaf ``rcu_node`` structure.
253
+
254
+
For whatever it is worth, if you draw the tree to be tree-shaped rather
255
+
than array-shaped, it is easy to draw a planar representation:
256
+
257
+
.. kernel-figure:: TreeLevel.svg
258
+
259
+
Finally, the ``->rda`` field references a per-CPU pointer to the
260
+
corresponding CPU's ``rcu_data`` structure.
261
+
262
+
All of these fields are constant once initialization is complete, and
263
+
therefore need no protection.
264
+
265
+
Grace-Period Tracking
266
+
'''''''''''''''''''''
267
+
268
+
This portion of the ``rcu_state`` structure is declared as follows:
269
+
270
+
::
271
+
272
+
1 unsigned long gp_seq;
273
+
274
+
RCU grace periods are numbered, and the ``->gp_seq`` field contains the
275
+
current grace-period sequence number. The bottom two bits are the state
276
+
of the current grace period, which can be zero for not yet started or
277
+
one for in progress. In other words, if the bottom two bits of
278
+
``->gp_seq`` are zero, then RCU is idle. Any other value in the bottom
279
+
two bits indicates that something is broken. This field is protected by
280
+
the root ``rcu_node`` structure's ``->lock`` field.
281
+
282
+
There are ``->gp_seq`` fields in the ``rcu_node`` and ``rcu_data``
283
+
structures as well. The fields in the ``rcu_state`` structure represent
284
+
the most current value, and those of the other structures are compared
285
+
in order to detect the beginnings and ends of grace periods in a
286
+
distributed fashion. The values flow from ``rcu_state`` to ``rcu_node``
287
+
(down the tree from the root to the leaves) to ``rcu_data``.
288
+
289
+
Miscellaneous
290
+
'''''''''''''
291
+
292
+
This portion of the ``rcu_state`` structure is declared as follows:
293
+
294
+
::
295
+
296
+
1 unsigned long gp_max;
297
+
2 char abbr;
298
+
3 char *name;
299
+
300
+
The ``->gp_max`` field tracks the duration of the longest grace period
301
+
in jiffies. It is protected by the root ``rcu_node``'s ``->lock``.
302
+
303
+
The ``->name`` and ``->abbr`` fields distinguish between preemptible RCU
304
+
(“rcu_preempt” and “p”) and non-preemptible RCU (“rcu_sched” and “s”).
305
+
These fields are used for diagnostic and tracing purposes.
306
+
307
+
The ``rcu_node`` Structure
308
+
~~~~~~~~~~~~~~~~~~~~~~~~~~
309
+
310
+
The ``rcu_node`` structures form the combining tree that propagates
311
+
quiescent-state information from the leaves to the root and also that
312
+
propagates grace-period information from the root down to the leaves.
313
+
They provides local copies of the grace-period state in order to allow
314
+
this information to be accessed in a synchronized manner without
315
+
suffering the scalability limitations that would otherwise be imposed by
316
+
global locking. In ``CONFIG_PREEMPT_RCU`` kernels, they manage the lists
317
+
of tasks that have blocked while in their current RCU read-side critical
318
+
section. In ``CONFIG_PREEMPT_RCU`` with ``CONFIG_RCU_BOOST``, they
319
+
manage the per-\ ``rcu_node`` priority-boosting kernel threads
320
+
(kthreads) and state. Finally, they record CPU-hotplug state in order to
321
+
determine which CPUs should be ignored during a given grace period.
322
+
323
+
The ``rcu_node`` structure's fields are discussed, singly and in groups,
324
+
in the following sections.
325
+
326
+
Connection to Combining Tree
327
+
''''''''''''''''''''''''''''
328
+
329
+
This portion of the ``rcu_node`` structure is declared as follows:
330
+
331
+
::
332
+
333
+
1 struct rcu_node *parent;
334
+
2 u8 level;
335
+
3 u8 grpnum;
336
+
4 unsigned long grpmask;
337
+
5 int grplo;
338
+
6 int grphi;
339
+
340
+
The ``->parent`` pointer references the ``rcu_node`` one level up in the
341
+
tree, and is ``NULL`` for the root ``rcu_node``. The RCU implementation
342
+
makes heavy use of this field to push quiescent states up the tree. The
343
+
``->level`` field gives the level in the tree, with the root being at
344
+
level zero, its children at level one, and so on. The ``->grpnum`` field
345
+
gives this node's position within the children of its parent, so this
346
+
number can range between 0 and 31 on 32-bit systems and between 0 and 63
347
+
on 64-bit systems. The ``->level`` and ``->grpnum`` fields are used only
348
+
during initialization and for tracing. The ``->grpmask`` field is the
349
+
bitmask counterpart of ``->grpnum``, and therefore always has exactly
350
+
one bit set. This mask is used to clear the bit corresponding to this
351
+
``rcu_node`` structure in its parent's bitmasks, which are described
352
+
later. Finally, the ``->grplo`` and ``->grphi`` fields contain the
353
+
lowest and highest numbered CPU served by this ``rcu_node`` structure,
354
+
respectively.
355
+
356
+
All of these fields are constant, and thus do not require any
357
+
synchronization.
358
+
359
+
Synchronization
360
+
'''''''''''''''
361
+
362
+
This field of the ``rcu_node`` structure is declared as follows:
363
+
364
+
::
365
+
366
+
1 raw_spinlock_t lock;
367
+
368
+
This field is used to protect the remaining fields in this structure,
369
+
unless otherwise stated. That said, all of the fields in this structure
370
+
can be accessed without locking for tracing purposes. Yes, this can
371
+
result in confusing traces, but better some tracing confusion than to be
372
+
heisenbugged out of existence.
373
+
374
+
.. _grace-period-tracking-1:
375
+
376
+
Grace-Period Tracking
377
+
'''''''''''''''''''''
378
+
379
+
This portion of the ``rcu_node`` structure is declared as follows:
380
+
381
+
::
382
+
383
+
1 unsigned long gp_seq;
384
+
2 unsigned long gp_seq_needed;
385
+
386
+
The ``rcu_node`` structures' ``->gp_seq`` fields are the counterparts of
387
+
the field of the same name in the ``rcu_state`` structure. They each may
388
+
lag up to one step behind their ``rcu_state`` counterpart. If the bottom
389
+
two bits of a given ``rcu_node`` structure's ``->gp_seq`` field is zero,
390
+
then this ``rcu_node`` structure believes that RCU is idle.
391
+
392
+
The ``>gp_seq`` field of each ``rcu_node`` structure is updated at the
393
+
beginning and the end of each grace period.
394
+
395
+
The ``->gp_seq_needed`` fields record the furthest-in-the-future grace
396
+
period request seen by the corresponding ``rcu_node`` structure. The
397
+
request is considered fulfilled when the value of the ``->gp_seq`` field
398
+
equals or exceeds that of the ``->gp_seq_needed`` field.
399
+
400
+
+-----------------------------------------------------------------------+
401
+
| **Quick Quiz**: |
402
+
+-----------------------------------------------------------------------+
403
+
| Suppose that this ``rcu_node`` structure doesn't see a request for a |
404
+
| very long time. Won't wrapping of the ``->gp_seq`` field cause |
405
+
| problems? |
406
+
+-----------------------------------------------------------------------+
407
+
| **Answer**: |
408
+
+-----------------------------------------------------------------------+
409
+
| No, because if the ``->gp_seq_needed`` field lags behind the |
410
+
| ``->gp_seq`` field, the ``->gp_seq_needed`` field will be updated at |
411
+
| the end of the grace period. Modulo-arithmetic comparisons therefore |
412
+
| will always get the correct answer, even with wrapping. |
413
+
+-----------------------------------------------------------------------+
414
+
415
+
Quiescent-State Tracking
416
+
''''''''''''''''''''''''
417
+
418
+
These fields manage the propagation of quiescent states up the combining
419
+
tree.
420
+
421
+
This portion of the ``rcu_node`` structure has fields as follows:
422
+
423
+
::
424
+
425
+
1 unsigned long qsmask;
426
+
2 unsigned long expmask;
427
+
3 unsigned long qsmaskinit;
428
+
4 unsigned long expmaskinit;
429
+
430
+
The ``->qsmask`` field tracks which of this ``rcu_node`` structure's
431
+
children still need to report quiescent states for the current normal
432
+
grace period. Such children will have a value of 1 in their
433
+
corresponding bit. Note that the leaf ``rcu_node`` structures should be
434
+
thought of as having ``rcu_data`` structures as their children.
435
+
Similarly, the ``->expmask`` field tracks which of this ``rcu_node``
436
+
structure's children still need to report quiescent states for the
437
+
current expedited grace period. An expedited grace period has the same
438
+
conceptual properties as a normal grace period, but the expedited
439
+
implementation accepts extreme CPU overhead to obtain much lower
440
+
grace-period latency, for example, consuming a few tens of microseconds
441
+
worth of CPU time to reduce grace-period duration from milliseconds to
442
+
tens of microseconds. The ``->qsmaskinit`` field tracks which of this
443
+
``rcu_node`` structure's children cover for at least one online CPU.
444
+
This mask is used to initialize ``->qsmask``, and ``->expmaskinit`` is
445
+
used to initialize ``->expmask`` and the beginning of the normal and
446
+
expedited grace periods, respectively.
447
+
448
+
+-----------------------------------------------------------------------+
449
+
| **Quick Quiz**: |
450
+
+-----------------------------------------------------------------------+
451
+
| Why are these bitmasks protected by locking? Come on, haven't you |
452
+
| heard of atomic instructions??? |
453
+
+-----------------------------------------------------------------------+
454
+
| **Answer**: |
455
+
+-----------------------------------------------------------------------+
456
+
| Lockless grace-period computation! Such a tantalizing possibility! |
457
+
| But consider the following sequence of events: |
458
+
| |
459
+
| #. CPU 0 has been in dyntick-idle mode for quite some time. When it |
460
+
| wakes up, it notices that the current RCU grace period needs it to |
461
+
| report in, so it sets a flag where the scheduling clock interrupt |
462
+
| will find it. |
463
+
| #. Meanwhile, CPU 1 is running ``force_quiescent_state()``, and |
464
+
| notices that CPU 0 has been in dyntick idle mode, which qualifies |
465
+
| as an extended quiescent state. |
466
+
| #. CPU 0's scheduling clock interrupt fires in the middle of an RCU |
467
+
| read-side critical section, and notices that the RCU core needs |
468
+
| something, so commences RCU softirq processing. |
469
+
| #. CPU 0's softirq handler executes and is just about ready to report |
470
+
| its quiescent state up the ``rcu_node`` tree. |
471
+
| #. But CPU 1 beats it to the punch, completing the current grace |
472
+
| period and starting a new one. |
473
+
| #. CPU 0 now reports its quiescent state for the wrong grace period. |
474
+
| That grace period might now end before the RCU read-side critical |
475
+
| section. If that happens, disaster will ensue. |
476
+
| |
477
+
| So the locking is absolutely required in order to coordinate clearing |
478
+
| of the bits with updating of the grace-period sequence number in |
479
+
| ``->gp_seq``. |
480
+
+-----------------------------------------------------------------------+
481
+
482
+
Blocked-Task Management
483
+
'''''''''''''''''''''''
484
+
485
+
``PREEMPT_RCU`` allows tasks to be preempted in the midst of their RCU
486
+
read-side critical sections, and these tasks must be tracked explicitly.
487
+
The details of exactly why and how they are tracked will be covered in a
488
+
separate article on RCU read-side processing. For now, it is enough to
489
+
know that the ``rcu_node`` structure tracks them.
490
+
491
+
::
492
+
493
+
1 struct list_head blkd_tasks;
494
+
2 struct list_head *gp_tasks;
495
+
3 struct list_head *exp_tasks;
496
+
4 bool wait_blkd_tasks;
497
+
498
+
The ``->blkd_tasks`` field is a list header for the list of blocked and
499
+
preempted tasks. As tasks undergo context switches within RCU read-side
500
+
critical sections, their ``task_struct`` structures are enqueued (via
501
+
the ``task_struct``'s ``->rcu_node_entry`` field) onto the head of the
502
+
``->blkd_tasks`` list for the leaf ``rcu_node`` structure corresponding
503
+
to the CPU on which the outgoing context switch executed. As these tasks
504
+
later exit their RCU read-side critical sections, they remove themselves
505
+
from the list. This list is therefore in reverse time order, so that if
506
+
one of the tasks is blocking the current grace period, all subsequent
507
+
tasks must also be blocking that same grace period. Therefore, a single
508
+
pointer into this list suffices to track all tasks blocking a given
509
+
grace period. That pointer is stored in ``->gp_tasks`` for normal grace
510
+
periods and in ``->exp_tasks`` for expedited grace periods. These last
511
+
two fields are ``NULL`` if either there is no grace period in flight or
512
+
if there are no blocked tasks preventing that grace period from
513
+
completing. If either of these two pointers is referencing a task that
514
+
removes itself from the ``->blkd_tasks`` list, then that task must
515
+
advance the pointer to the next task on the list, or set the pointer to
516
+
``NULL`` if there are no subsequent tasks on the list.
517
+
518
+
For example, suppose that tasks T1, T2, and T3 are all hard-affinitied
519
+
to the largest-numbered CPU in the system. Then if task T1 blocked in an
520
+
RCU read-side critical section, then an expedited grace period started,
521
+
then task T2 blocked in an RCU read-side critical section, then a normal
522
+
grace period started, and finally task 3 blocked in an RCU read-side
523
+
critical section, then the state of the last leaf ``rcu_node``
524
+
structure's blocked-task list would be as shown below:
525
+
526
+
.. kernel-figure:: blkd_task.svg
527
+
528
+
Task T1 is blocking both grace periods, task T2 is blocking only the
529
+
normal grace period, and task T3 is blocking neither grace period. Note
530
+
that these tasks will not remove themselves from this list immediately
531
+
upon resuming execution. They will instead remain on the list until they
532
+
execute the outermost ``rcu_read_unlock()`` that ends their RCU
533
+
read-side critical section.
534
+
535
+
The ``->wait_blkd_tasks`` field indicates whether or not the current
536
+
grace period is waiting on a blocked task.
537
+
538
+
Sizing the ``rcu_node`` Array
539
+
'''''''''''''''''''''''''''''
540
+
541
+
The ``rcu_node`` array is sized via a series of C-preprocessor
542
+
expressions as follows:
543
+
544
+
::
545
+
546
+
1 #ifdef CONFIG_RCU_FANOUT
547
+
2 #define RCU_FANOUT CONFIG_RCU_FANOUT
548
+
3 #else
549
+
4 # ifdef CONFIG_64BIT
550
+
5 # define RCU_FANOUT 64
551
+
6 # else
552
+
7 # define RCU_FANOUT 32
553
+
8 # endif
554
+
9 #endif
555
+
10
556
+
11 #ifdef CONFIG_RCU_FANOUT_LEAF
557
+
12 #define RCU_FANOUT_LEAF CONFIG_RCU_FANOUT_LEAF
558
+
13 #else
559
+
14 # ifdef CONFIG_64BIT
560
+
15 # define RCU_FANOUT_LEAF 64
561
+
16 # else
562
+
17 # define RCU_FANOUT_LEAF 32
563
+
18 # endif
564
+
19 #endif
565
+
20
566
+
21 #define RCU_FANOUT_1 (RCU_FANOUT_LEAF)
567
+
22 #define RCU_FANOUT_2 (RCU_FANOUT_1 * RCU_FANOUT)
568
+
23 #define RCU_FANOUT_3 (RCU_FANOUT_2 * RCU_FANOUT)
569
+
24 #define RCU_FANOUT_4 (RCU_FANOUT_3 * RCU_FANOUT)
570
+
25
571
+
26 #if NR_CPUS <= RCU_FANOUT_1
572
+
27 # define RCU_NUM_LVLS 1
573
+
28 # define NUM_RCU_LVL_0 1
574
+
29 # define NUM_RCU_NODES NUM_RCU_LVL_0
575
+
30 # define NUM_RCU_LVL_INIT { NUM_RCU_LVL_0 }
576
+
31 # define RCU_NODE_NAME_INIT { "rcu_node_0" }
577
+
32 # define RCU_FQS_NAME_INIT { "rcu_node_fqs_0" }
578
+
33 # define RCU_EXP_NAME_INIT { "rcu_node_exp_0" }
579
+
34 #elif NR_CPUS <= RCU_FANOUT_2
580
+
35 # define RCU_NUM_LVLS 2
581
+
36 # define NUM_RCU_LVL_0 1
582
+
37 # define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
583
+
38 # define NUM_RCU_NODES (NUM_RCU_LVL_0 + NUM_RCU_LVL_1)
584
+
39 # define NUM_RCU_LVL_INIT { NUM_RCU_LVL_0, NUM_RCU_LVL_1 }
585
+
40 # define RCU_NODE_NAME_INIT { "rcu_node_0", "rcu_node_1" }
586
+
41 # define RCU_FQS_NAME_INIT { "rcu_node_fqs_0", "rcu_node_fqs_1" }
587
+
42 # define RCU_EXP_NAME_INIT { "rcu_node_exp_0", "rcu_node_exp_1" }
588
+
43 #elif NR_CPUS <= RCU_FANOUT_3
589
+
44 # define RCU_NUM_LVLS 3
590
+
45 # define NUM_RCU_LVL_0 1
591
+
46 # define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_2)
592
+
47 # define NUM_RCU_LVL_2 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
593
+
48 # define NUM_RCU_NODES (NUM_RCU_LVL_0 + NUM_RCU_LVL_1 + NUM_RCU_LVL_2)
594
+
49 # define NUM_RCU_LVL_INIT { NUM_RCU_LVL_0, NUM_RCU_LVL_1, NUM_RCU_LVL_2 }
595
+
50 # define RCU_NODE_NAME_INIT { "rcu_node_0", "rcu_node_1", "rcu_node_2" }
596
+
51 # define RCU_FQS_NAME_INIT { "rcu_node_fqs_0", "rcu_node_fqs_1", "rcu_node_fqs_2" }
597
+
52 # define RCU_EXP_NAME_INIT { "rcu_node_exp_0", "rcu_node_exp_1", "rcu_node_exp_2" }
598
+
53 #elif NR_CPUS <= RCU_FANOUT_4
599
+
54 # define RCU_NUM_LVLS 4
600
+
55 # define NUM_RCU_LVL_0 1
601
+
56 # define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_3)
602
+
57 # define NUM_RCU_LVL_2 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_2)
603
+
58 # define NUM_RCU_LVL_3 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
604
+
59 # define NUM_RCU_NODES (NUM_RCU_LVL_0 + NUM_RCU_LVL_1 + NUM_RCU_LVL_2 + NUM_RCU_LVL_3)
605
+
60 # define NUM_RCU_LVL_INIT { NUM_RCU_LVL_0, NUM_RCU_LVL_1, NUM_RCU_LVL_2, NUM_RCU_LVL_3 }
606
+
61 # define RCU_NODE_NAME_INIT { "rcu_node_0", "rcu_node_1", "rcu_node_2", "rcu_node_3" }
607
+
62 # define RCU_FQS_NAME_INIT { "rcu_node_fqs_0", "rcu_node_fqs_1", "rcu_node_fqs_2", "rcu_node_fqs_3" }
608
+
63 # define RCU_EXP_NAME_INIT { "rcu_node_exp_0", "rcu_node_exp_1", "rcu_node_exp_2", "rcu_node_exp_3" }
609
+
64 #else
610
+
65 # error "CONFIG_RCU_FANOUT insufficient for NR_CPUS"
611
+
66 #endif
612
+
613
+
The maximum number of levels in the ``rcu_node`` structure is currently
614
+
limited to four, as specified by lines 21-24 and the structure of the
615
+
subsequent “if” statement. For 32-bit systems, this allows
616
+
16*32*32*32=524,288 CPUs, which should be sufficient for the next few
617
+
years at least. For 64-bit systems, 16*64*64*64=4,194,304 CPUs is
618
+
allowed, which should see us through the next decade or so. This
619
+
four-level tree also allows kernels built with ``CONFIG_RCU_FANOUT=8``
620
+
to support up to 4096 CPUs, which might be useful in very large systems
621
+
having eight CPUs per socket (but please note that no one has yet shown
622
+
any measurable performance degradation due to misaligned socket and
623
+
``rcu_node`` boundaries). In addition, building kernels with a full four
624
+
levels of ``rcu_node`` tree permits better testing of RCU's
625
+
combining-tree code.
626
+
627
+
The ``RCU_FANOUT`` symbol controls how many children are permitted at
628
+
each non-leaf level of the ``rcu_node`` tree. If the
629
+
``CONFIG_RCU_FANOUT`` Kconfig option is not specified, it is set based
630
+
on the word size of the system, which is also the Kconfig default.
631
+
632
+
The ``RCU_FANOUT_LEAF`` symbol controls how many CPUs are handled by
633
+
each leaf ``rcu_node`` structure. Experience has shown that allowing a
634
+
given leaf ``rcu_node`` structure to handle 64 CPUs, as permitted by the
635
+
number of bits in the ``->qsmask`` field on a 64-bit system, results in
636
+
excessive contention for the leaf ``rcu_node`` structures' ``->lock``
637
+
fields. The number of CPUs per leaf ``rcu_node`` structure is therefore
638
+
limited to 16 given the default value of ``CONFIG_RCU_FANOUT_LEAF``. If
639
+
``CONFIG_RCU_FANOUT_LEAF`` is unspecified, the value selected is based
640
+
on the word size of the system, just as for ``CONFIG_RCU_FANOUT``.
641
+
Lines 11-19 perform this computation.
642
+
643
+
Lines 21-24 compute the maximum number of CPUs supported by a
644
+
single-level (which contains a single ``rcu_node`` structure),
645
+
two-level, three-level, and four-level ``rcu_node`` tree, respectively,
646
+
given the fanout specified by ``RCU_FANOUT`` and ``RCU_FANOUT_LEAF``.
647
+
These numbers of CPUs are retained in the ``RCU_FANOUT_1``,
648
+
``RCU_FANOUT_2``, ``RCU_FANOUT_3``, and ``RCU_FANOUT_4`` C-preprocessor
649
+
variables, respectively.
650
+
651
+
These variables are used to control the C-preprocessor ``#if`` statement
652
+
spanning lines 26-66 that computes the number of ``rcu_node`` structures
653
+
required for each level of the tree, as well as the number of levels
654
+
required. The number of levels is placed in the ``NUM_RCU_LVLS``
655
+
C-preprocessor variable by lines 27, 35, 44, and 54. The number of
656
+
``rcu_node`` structures for the topmost level of the tree is always
657
+
exactly one, and this value is unconditionally placed into
658
+
``NUM_RCU_LVL_0`` by lines 28, 36, 45, and 55. The rest of the levels
659
+
(if any) of the ``rcu_node`` tree are computed by dividing the maximum
660
+
number of CPUs by the fanout supported by the number of levels from the
661
+
current level down, rounding up. This computation is performed by
662
+
lines 37, 46-47, and 56-58. Lines 31-33, 40-42, 50-52, and 62-63 create
663
+
initializers for lockdep lock-class names. Finally, lines 64-66 produce
664
+
an error if the maximum number of CPUs is too large for the specified
665
+
fanout.
666
+
667
+
The ``rcu_segcblist`` Structure
668
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
669
+
670
+
The ``rcu_segcblist`` structure maintains a segmented list of callbacks
671
+
as follows:
672
+
673
+
::
674
+
675
+
1 #define RCU_DONE_TAIL 0
676
+
2 #define RCU_WAIT_TAIL 1
677
+
3 #define RCU_NEXT_READY_TAIL 2
678
+
4 #define RCU_NEXT_TAIL 3
679
+
5 #define RCU_CBLIST_NSEGS 4
680
+
6
681
+
7 struct rcu_segcblist {
682
+
8 struct rcu_head *head;
683
+
9 struct rcu_head **tails[RCU_CBLIST_NSEGS];
684
+
10 unsigned long gp_seq[RCU_CBLIST_NSEGS];
685
+
11 long len;
686
+
12 long len_lazy;
687
+
13 };
688
+
689
+
The segments are as follows:
690
+
691
+
#. ``RCU_DONE_TAIL``: Callbacks whose grace periods have elapsed. These
692
+
callbacks are ready to be invoked.
693
+
#. ``RCU_WAIT_TAIL``: Callbacks that are waiting for the current grace
694
+
period. Note that different CPUs can have different ideas about which
695
+
grace period is current, hence the ``->gp_seq`` field.
696
+
#. ``RCU_NEXT_READY_TAIL``: Callbacks waiting for the next grace period
697
+
to start.
698
+
#. ``RCU_NEXT_TAIL``: Callbacks that have not yet been associated with a
699
+
grace period.
700
+
701
+
The ``->head`` pointer references the first callback or is ``NULL`` if
702
+
the list contains no callbacks (which is *not* the same as being empty).
703
+
Each element of the ``->tails[]`` array references the ``->next``
704
+
pointer of the last callback in the corresponding segment of the list,
705
+
or the list's ``->head`` pointer if that segment and all previous
706
+
segments are empty. If the corresponding segment is empty but some
707
+
previous segment is not empty, then the array element is identical to
708
+
its predecessor. Older callbacks are closer to the head of the list, and
709
+
new callbacks are added at the tail. This relationship between the
710
+
``->head`` pointer, the ``->tails[]`` array, and the callbacks is shown
711
+
in this diagram:
712
+
713
+
.. kernel-figure:: nxtlist.svg
714
+
715
+
In this figure, the ``->head`` pointer references the first RCU callback
716
+
in the list. The ``->tails[RCU_DONE_TAIL]`` array element references the
717
+
``->head`` pointer itself, indicating that none of the callbacks is
718
+
ready to invoke. The ``->tails[RCU_WAIT_TAIL]`` array element references
719
+
callback CB 2's ``->next`` pointer, which indicates that CB 1 and CB 2
720
+
are both waiting on the current grace period, give or take possible
721
+
disagreements about exactly which grace period is the current one. The
722
+
``->tails[RCU_NEXT_READY_TAIL]`` array element references the same RCU
723
+
callback that ``->tails[RCU_WAIT_TAIL]`` does, which indicates that
724
+
there are no callbacks waiting on the next RCU grace period. The
725
+
``->tails[RCU_NEXT_TAIL]`` array element references CB 4's ``->next``
726
+
pointer, indicating that all the remaining RCU callbacks have not yet
727
+
been assigned to an RCU grace period. Note that the
728
+
``->tails[RCU_NEXT_TAIL]`` array element always references the last RCU
729
+
callback's ``->next`` pointer unless the callback list is empty, in
730
+
which case it references the ``->head`` pointer.
731
+
732
+
There is one additional important special case for the
733
+
``->tails[RCU_NEXT_TAIL]`` array element: It can be ``NULL`` when this
734
+
list is *disabled*. Lists are disabled when the corresponding CPU is
735
+
offline or when the corresponding CPU's callbacks are offloaded to a
736
+
kthread, both of which are described elsewhere.
737
+
738
+
CPUs advance their callbacks from the ``RCU_NEXT_TAIL`` to the
739
+
``RCU_NEXT_READY_TAIL`` to the ``RCU_WAIT_TAIL`` to the
740
+
``RCU_DONE_TAIL`` list segments as grace periods advance.
741
+
742
+
The ``->gp_seq[]`` array records grace-period numbers corresponding to
743
+
the list segments. This is what allows different CPUs to have different
744
+
ideas as to which is the current grace period while still avoiding
745
+
premature invocation of their callbacks. In particular, this allows CPUs
746
+
that go idle for extended periods to determine which of their callbacks
747
+
are ready to be invoked after reawakening.
748
+
749
+
The ``->len`` counter contains the number of callbacks in ``->head``,
750
+
and the ``->len_lazy`` contains the number of those callbacks that are
751
+
known to only free memory, and whose invocation can therefore be safely
752
+
deferred.
753
+
754
+
.. important::
755
+
756
+
It is the ``->len`` field that determines whether or
757
+
not there are callbacks associated with this ``rcu_segcblist``
758
+
structure, *not* the ``->head`` pointer. The reason for this is that all
759
+
the ready-to-invoke callbacks (that is, those in the ``RCU_DONE_TAIL``
760
+
segment) are extracted all at once at callback-invocation time
761
+
(``rcu_do_batch``), due to which ``->head`` may be set to NULL if there
762
+
are no not-done callbacks remaining in the ``rcu_segcblist``. If
763
+
callback invocation must be postponed, for example, because a
764
+
high-priority process just woke up on this CPU, then the remaining
765
+
callbacks are placed back on the ``RCU_DONE_TAIL`` segment and
766
+
``->head`` once again points to the start of the segment. In short, the
767
+
head field can briefly be ``NULL`` even though the CPU has callbacks
768
+
present the entire time. Therefore, it is not appropriate to test the
769
+
``->head`` pointer for ``NULL``.
770
+
771
+
In contrast, the ``->len`` and ``->len_lazy`` counts are adjusted only
772
+
after the corresponding callbacks have been invoked. This means that the
773
+
``->len`` count is zero only if the ``rcu_segcblist`` structure really
774
+
is devoid of callbacks. Of course, off-CPU sampling of the ``->len``
775
+
count requires careful use of appropriate synchronization, for example,
776
+
memory barriers. This synchronization can be a bit subtle, particularly
777
+
in the case of ``rcu_barrier()``.
778
+
779
+
The ``rcu_data`` Structure
780
+
~~~~~~~~~~~~~~~~~~~~~~~~~~
781
+
782
+
The ``rcu_data`` maintains the per-CPU state for the RCU subsystem. The
783
+
fields in this structure may be accessed only from the corresponding CPU
784
+
(and from tracing) unless otherwise stated. This structure is the focus
785
+
of quiescent-state detection and RCU callback queuing. It also tracks
786
+
its relationship to the corresponding leaf ``rcu_node`` structure to
787
+
allow more-efficient propagation of quiescent states up the ``rcu_node``
788
+
combining tree. Like the ``rcu_node`` structure, it provides a local
789
+
copy of the grace-period information to allow for-free synchronized
790
+
access to this information from the corresponding CPU. Finally, this
791
+
structure records past dyntick-idle state for the corresponding CPU and
792
+
also tracks statistics.
793
+
794
+
The ``rcu_data`` structure's fields are discussed, singly and in groups,
795
+
in the following sections.
796
+
797
+
Connection to Other Data Structures
798
+
'''''''''''''''''''''''''''''''''''
799
+
800
+
This portion of the ``rcu_data`` structure is declared as follows:
801
+
802
+
::
803
+
804
+
1 int cpu;
805
+
2 struct rcu_node *mynode;
806
+
3 unsigned long grpmask;
807
+
4 bool beenonline;
808
+
809
+
The ``->cpu`` field contains the number of the corresponding CPU and the
810
+
``->mynode`` field references the corresponding ``rcu_node`` structure.
811
+
The ``->mynode`` is used to propagate quiescent states up the combining
812
+
tree. These two fields are constant and therefore do not require
813
+
synchronization.
814
+
815
+
The ``->grpmask`` field indicates the bit in the ``->mynode->qsmask``
816
+
corresponding to this ``rcu_data`` structure, and is also used when
817
+
propagating quiescent states. The ``->beenonline`` flag is set whenever
818
+
the corresponding CPU comes online, which means that the debugfs tracing
819
+
need not dump out any ``rcu_data`` structure for which this flag is not
820
+
set.
821
+
822
+
Quiescent-State and Grace-Period Tracking
823
+
'''''''''''''''''''''''''''''''''''''''''
824
+
825
+
This portion of the ``rcu_data`` structure is declared as follows:
826
+
827
+
::
828
+
829
+
1 unsigned long gp_seq;
830
+
2 unsigned long gp_seq_needed;
831
+
3 bool cpu_no_qs;
832
+
4 bool core_needs_qs;
833
+
5 bool gpwrap;
834
+
835
+
The ``->gp_seq`` field is the counterpart of the field of the same name
836
+
in the ``rcu_state`` and ``rcu_node`` structures. The
837
+
``->gp_seq_needed`` field is the counterpart of the field of the same
838
+
name in the rcu_node structure. They may each lag up to one behind their
839
+
``rcu_node`` counterparts, but in ``CONFIG_NO_HZ_IDLE`` and
840
+
``CONFIG_NO_HZ_FULL`` kernels can lag arbitrarily far behind for CPUs in
841
+
dyntick-idle mode (but these counters will catch up upon exit from
842
+
dyntick-idle mode). If the lower two bits of a given ``rcu_data``
843
+
structure's ``->gp_seq`` are zero, then this ``rcu_data`` structure
844
+
believes that RCU is idle.
845
+
846
+
+-----------------------------------------------------------------------+
847
+
| **Quick Quiz**: |
848
+
+-----------------------------------------------------------------------+
849
+
| All this replication of the grace period numbers can only cause |
850
+
| massive confusion. Why not just keep a global sequence number and be |
851
+
| done with it??? |
852
+
+-----------------------------------------------------------------------+
853
+
| **Answer**: |
854
+
+-----------------------------------------------------------------------+
855
+
| Because if there was only a single global sequence numbers, there |
856
+
| would need to be a single global lock to allow safely accessing and |
857
+
| updating it. And if we are not going to have a single global lock, we |
858
+
| need to carefully manage the numbers on a per-node basis. Recall from |
859
+
| the answer to a previous Quick Quiz that the consequences of applying |
860
+
| a previously sampled quiescent state to the wrong grace period are |
861
+
| quite severe. |
862
+
+-----------------------------------------------------------------------+
863
+
864
+
The ``->cpu_no_qs`` flag indicates that the CPU has not yet passed
865
+
through a quiescent state, while the ``->core_needs_qs`` flag indicates
866
+
that the RCU core needs a quiescent state from the corresponding CPU.
867
+
The ``->gpwrap`` field indicates that the corresponding CPU has remained
868
+
idle for so long that the ``gp_seq`` counter is in danger of overflow,
869
+
which will cause the CPU to disregard the values of its counters on its
870
+
next exit from idle.
871
+
872
+
RCU Callback Handling
873
+
'''''''''''''''''''''
874
+
875
+
In the absence of CPU-hotplug events, RCU callbacks are invoked by the
876
+
same CPU that registered them. This is strictly a cache-locality
877
+
optimization: callbacks can and do get invoked on CPUs other than the
878
+
one that registered them. After all, if the CPU that registered a given
879
+
callback has gone offline before the callback can be invoked, there
880
+
really is no other choice.
881
+
882
+
This portion of the ``rcu_data`` structure is declared as follows:
883
+
884
+
::
885
+
886
+
1 struct rcu_segcblist cblist;
887
+
2 long qlen_last_fqs_check;
888
+
3 unsigned long n_cbs_invoked;
889
+
4 unsigned long n_nocbs_invoked;
890
+
5 unsigned long n_cbs_orphaned;
891
+
6 unsigned long n_cbs_adopted;
892
+
7 unsigned long n_force_qs_snap;
893
+
8 long blimit;
894
+
895
+
The ``->cblist`` structure is the segmented callback list described
896
+
earlier. The CPU advances the callbacks in its ``rcu_data`` structure
897
+
whenever it notices that another RCU grace period has completed. The CPU
898
+
detects the completion of an RCU grace period by noticing that the value
899
+
of its ``rcu_data`` structure's ``->gp_seq`` field differs from that of
900
+
its leaf ``rcu_node`` structure. Recall that each ``rcu_node``
901
+
structure's ``->gp_seq`` field is updated at the beginnings and ends of
902
+
each grace period.
903
+
904
+
The ``->qlen_last_fqs_check`` and ``->n_force_qs_snap`` coordinate the
905
+
forcing of quiescent states from ``call_rcu()`` and friends when
906
+
callback lists grow excessively long.
907
+
908
+
The ``->n_cbs_invoked``, ``->n_cbs_orphaned``, and ``->n_cbs_adopted``
909
+
fields count the number of callbacks invoked, sent to other CPUs when
910
+
this CPU goes offline, and received from other CPUs when those other
911
+
CPUs go offline. The ``->n_nocbs_invoked`` is used when the CPU's
912
+
callbacks are offloaded to a kthread.
913
+
914
+
Finally, the ``->blimit`` counter is the maximum number of RCU callbacks
915
+
that may be invoked at a given time.
916
+
917
+
Dyntick-Idle Handling
918
+
'''''''''''''''''''''
919
+
920
+
This portion of the ``rcu_data`` structure is declared as follows:
921
+
922
+
::
923
+
924
+
1 int dynticks_snap;
925
+
2 unsigned long dynticks_fqs;
926
+
927
+
The ``->dynticks_snap`` field is used to take a snapshot of the
928
+
corresponding CPU's dyntick-idle state when forcing quiescent states,
929
+
and is therefore accessed from other CPUs. Finally, the
930
+
``->dynticks_fqs`` field is used to count the number of times this CPU
931
+
is determined to be in dyntick-idle state, and is used for tracing and
932
+
debugging purposes.
933
+
934
+
This portion of the rcu_data structure is declared as follows:
935
+
936
+
::
937
+
938
+
1 long dynticks_nesting;
939
+
2 long dynticks_nmi_nesting;
940
+
3 atomic_t dynticks;
941
+
4 bool rcu_need_heavy_qs;
942
+
5 bool rcu_urgent_qs;
943
+
944
+
These fields in the rcu_data structure maintain the per-CPU dyntick-idle
945
+
state for the corresponding CPU. The fields may be accessed only from
946
+
the corresponding CPU (and from tracing) unless otherwise stated.
947
+
948
+
The ``->dynticks_nesting`` field counts the nesting depth of process
949
+
execution, so that in normal circumstances this counter has value zero
950
+
or one. NMIs, irqs, and tracers are counted by the
951
+
``->dynticks_nmi_nesting`` field. Because NMIs cannot be masked, changes
952
+
to this variable have to be undertaken carefully using an algorithm
953
+
provided by Andy Lutomirski. The initial transition from idle adds one,
954
+
and nested transitions add two, so that a nesting level of five is
955
+
represented by a ``->dynticks_nmi_nesting`` value of nine. This counter
956
+
can therefore be thought of as counting the number of reasons why this
957
+
CPU cannot be permitted to enter dyntick-idle mode, aside from
958
+
process-level transitions.
959
+
960
+
However, it turns out that when running in non-idle kernel context, the
961
+
Linux kernel is fully capable of entering interrupt handlers that never
962
+
exit and perhaps also vice versa. Therefore, whenever the
963
+
``->dynticks_nesting`` field is incremented up from zero, the
964
+
``->dynticks_nmi_nesting`` field is set to a large positive number, and
965
+
whenever the ``->dynticks_nesting`` field is decremented down to zero,
966
+
the the ``->dynticks_nmi_nesting`` field is set to zero. Assuming that
967
+
the number of misnested interrupts is not sufficient to overflow the
968
+
counter, this approach corrects the ``->dynticks_nmi_nesting`` field
969
+
every time the corresponding CPU enters the idle loop from process
970
+
context.
971
+
972
+
The ``->dynticks`` field counts the corresponding CPU's transitions to
973
+
and from either dyntick-idle or user mode, so that this counter has an
974
+
even value when the CPU is in dyntick-idle mode or user mode and an odd
975
+
value otherwise. The transitions to/from user mode need to be counted
976
+
for user mode adaptive-ticks support (see timers/NO_HZ.txt).
977
+
978
+
The ``->rcu_need_heavy_qs`` field is used to record the fact that the
979
+
RCU core code would really like to see a quiescent state from the
980
+
corresponding CPU, so much so that it is willing to call for
981
+
heavy-weight dyntick-counter operations. This flag is checked by RCU's
982
+
context-switch and ``cond_resched()`` code, which provide a momentary
983
+
idle sojourn in response.
984
+
985
+
Finally, the ``->rcu_urgent_qs`` field is used to record the fact that
986
+
the RCU core code would really like to see a quiescent state from the
987
+
corresponding CPU, with the various other fields indicating just how
988
+
badly RCU wants this quiescent state. This flag is checked by RCU's
989
+
context-switch path (``rcu_note_context_switch``) and the cond_resched
990
+
code.
991
+
992
+
+-----------------------------------------------------------------------+
993
+
| **Quick Quiz**: |
994
+
+-----------------------------------------------------------------------+
995
+
| Why not simply combine the ``->dynticks_nesting`` and |
996
+
| ``->dynticks_nmi_nesting`` counters into a single counter that just |
997
+
| counts the number of reasons that the corresponding CPU is non-idle? |
998
+
+-----------------------------------------------------------------------+
999
+
| **Answer**: |
1000
+
+-----------------------------------------------------------------------+
1001
+
| Because this would fail in the presence of interrupts whose handlers |
1002
+
| never return and of handlers that manage to return from a made-up |
1003
+
| interrupt. |
1004
+
+-----------------------------------------------------------------------+
1005
+
1006
+
Additional fields are present for some special-purpose builds, and are
1007
+
discussed separately.
1008
+
1009
+
The ``rcu_head`` Structure
1010
+
~~~~~~~~~~~~~~~~~~~~~~~~~~
1011
+
1012
+
Each ``rcu_head`` structure represents an RCU callback. These structures
1013
+
are normally embedded within RCU-protected data structures whose
1014
+
algorithms use asynchronous grace periods. In contrast, when using
1015
+
algorithms that block waiting for RCU grace periods, RCU users need not
1016
+
provide ``rcu_head`` structures.
1017
+
1018
+
The ``rcu_head`` structure has fields as follows:
1019
+
1020
+
::
1021
+
1022
+
1 struct rcu_head *next;
1023
+
2 void (*func)(struct rcu_head *head);
1024
+
1025
+
The ``->next`` field is used to link the ``rcu_head`` structures
1026
+
together in the lists within the ``rcu_data`` structures. The ``->func``
1027
+
field is a pointer to the function to be called when the callback is
1028
+
ready to be invoked, and this function is passed a pointer to the
1029
+
``rcu_head`` structure. However, ``kfree_rcu()`` uses the ``->func``
1030
+
field to record the offset of the ``rcu_head`` structure within the
1031
+
enclosing RCU-protected data structure.
1032
+
1033
+
Both of these fields are used internally by RCU. From the viewpoint of
1034
+
RCU users, this structure is an opaque “cookie”.
1035
+
1036
+
+-----------------------------------------------------------------------+
1037
+
| **Quick Quiz**: |
1038
+
+-----------------------------------------------------------------------+
1039
+
| Given that the callback function ``->func`` is passed a pointer to |
1040
+
| the ``rcu_head`` structure, how is that function supposed to find the |
1041
+
| beginning of the enclosing RCU-protected data structure? |
1042
+
+-----------------------------------------------------------------------+
1043
+
| **Answer**: |
1044
+
+-----------------------------------------------------------------------+
1045
+
| In actual practice, there is a separate callback function per type of |
1046
+
| RCU-protected data structure. The callback function can therefore use |
1047
+
| the ``container_of()`` macro in the Linux kernel (or other |
1048
+
| pointer-manipulation facilities in other software environments) to |
1049
+
| find the beginning of the enclosing structure. |
1050
+
+-----------------------------------------------------------------------+
1051
+
1052
+
RCU-Specific Fields in the ``task_struct`` Structure
1053
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1054
+
1055
+
The ``CONFIG_PREEMPT_RCU`` implementation uses some additional fields in
1056
+
the ``task_struct`` structure:
1057
+
1058
+
::
1059
+
1060
+
1 #ifdef CONFIG_PREEMPT_RCU
1061
+
2 int rcu_read_lock_nesting;
1062
+
3 union rcu_special rcu_read_unlock_special;
1063
+
4 struct list_head rcu_node_entry;
1064
+
5 struct rcu_node *rcu_blocked_node;
1065
+
6 #endif /* #ifdef CONFIG_PREEMPT_RCU */
1066
+
7 #ifdef CONFIG_TASKS_RCU
1067
+
8 unsigned long rcu_tasks_nvcsw;
1068
+
9 bool rcu_tasks_holdout;
1069
+
10 struct list_head rcu_tasks_holdout_list;
1070
+
11 int rcu_tasks_idle_cpu;
1071
+
12 #endif /* #ifdef CONFIG_TASKS_RCU */
1072
+
1073
+
The ``->rcu_read_lock_nesting`` field records the nesting level for RCU
1074
+
read-side critical sections, and the ``->rcu_read_unlock_special`` field
1075
+
is a bitmask that records special conditions that require
1076
+
``rcu_read_unlock()`` to do additional work. The ``->rcu_node_entry``
1077
+
field is used to form lists of tasks that have blocked within
1078
+
preemptible-RCU read-side critical sections and the
1079
+
``->rcu_blocked_node`` field references the ``rcu_node`` structure whose
1080
+
list this task is a member of, or ``NULL`` if it is not blocked within a
1081
+
preemptible-RCU read-side critical section.
1082
+
1083
+
The ``->rcu_tasks_nvcsw`` field tracks the number of voluntary context
1084
+
switches that this task had undergone at the beginning of the current
1085
+
tasks-RCU grace period, ``->rcu_tasks_holdout`` is set if the current
1086
+
tasks-RCU grace period is waiting on this task,
1087
+
``->rcu_tasks_holdout_list`` is a list element enqueuing this task on
1088
+
the holdout list, and ``->rcu_tasks_idle_cpu`` tracks which CPU this
1089
+
idle task is running, but only if the task is currently running, that
1090
+
is, if the CPU is currently idle.
1091
+
1092
+
Accessor Functions
1093
+
~~~~~~~~~~~~~~~~~~
1094
+
1095
+
The following listing shows the ``rcu_get_root()``,
1096
+
``rcu_for_each_node_breadth_first`` and ``rcu_for_each_leaf_node()``
1097
+
function and macros:
1098
+
1099
+
::
1100
+
1101
+
1 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
1102
+
2 {
1103
+
3 return &rsp->node[0];
1104
+
4 }
1105
+
5
1106
+
6 #define rcu_for_each_node_breadth_first(rsp, rnp) \
1107
+
7 for ((rnp) = &(rsp)->node[0]; \
1108
+
8 (rnp) < &(rsp)->node[NUM_RCU_NODES]; (rnp)++)
1109
+
9
1110
+
10 #define rcu_for_each_leaf_node(rsp, rnp) \
1111
+
11 for ((rnp) = (rsp)->level[NUM_RCU_LVLS - 1]; \
1112
+
12 (rnp) < &(rsp)->node[NUM_RCU_NODES]; (rnp)++)
1113
+
1114
+
The ``rcu_get_root()`` simply returns a pointer to the first element of
1115
+
the specified ``rcu_state`` structure's ``->node[]`` array, which is the
1116
+
root ``rcu_node`` structure.
1117
+
1118
+
As noted earlier, the ``rcu_for_each_node_breadth_first()`` macro takes
1119
+
advantage of the layout of the ``rcu_node`` structures in the
1120
+
``rcu_state`` structure's ``->node[]`` array, performing a breadth-first
1121
+
traversal by simply traversing the array in order. Similarly, the
1122
+
``rcu_for_each_leaf_node()`` macro traverses only the last part of the
1123
+
array, thus traversing only the leaf ``rcu_node`` structures.
1124
+
1125
+
+-----------------------------------------------------------------------+
1126
+
| **Quick Quiz**: |
1127
+
+-----------------------------------------------------------------------+
1128
+
| What does ``rcu_for_each_leaf_node()`` do if the ``rcu_node`` tree |
1129
+
| contains only a single node? |
1130
+
+-----------------------------------------------------------------------+
1131
+
| **Answer**: |
1132
+
+-----------------------------------------------------------------------+
1133
+
| In the single-node case, ``rcu_for_each_leaf_node()`` traverses the |
1134
+
| single node. |
1135
+
+-----------------------------------------------------------------------+
1136
+
1137
+
Summary
1138
+
~~~~~~~
1139
+
1140
+
So the state of RCU is represented by an ``rcu_state`` structure, which
1141
+
contains a combining tree of ``rcu_node`` and ``rcu_data`` structures.
1142
+
Finally, in ``CONFIG_NO_HZ_IDLE`` kernels, each CPU's dyntick-idle state
1143
+
is tracked by dynticks-related fields in the ``rcu_data`` structure. If
1144
+
you made it this far, you are well prepared to read the code
1145
+
walkthroughs in the other articles in this series.
1146
+
1147
+
Acknowledgments
1148
+
~~~~~~~~~~~~~~~
1149
+
1150
+
I owe thanks to Cyrill Gorcunov, Mathieu Desnoyers, Dhaval Giani, Paul
1151
+
Turner, Abhishek Srivastava, Matt Kowalczyk, and Serge Hallyn for
1152
+
helping me get this document into a more human-readable state.
1153
+
1154
+
Legal Statement
1155
+
~~~~~~~~~~~~~~~
1156
+
1157
+
This work represents the view of the author and does not necessarily
1158
+
represent the view of IBM.
1159
+
1160
+
Linux is a registered trademark of Linus Torvalds.
1161
+
1162
+
Other company, product, and service names may be trademarks or service
1163
+
marks of others.
-668
Documentation/RCU/Design/Expedited-Grace-Periods/Expedited-Grace-Periods.html
-668
Documentation/RCU/Design/Expedited-Grace-Periods/Expedited-Grace-Periods.html
···
1
-
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN"
2
-
"http://www.w3.org/TR/html4/loose.dtd">
3
-
<html>
4
-
<head><title>A Tour Through TREE_RCU's Expedited Grace Periods</title>
5
-
<meta HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=iso-8859-1">
6
-
7
-
<h2>Introduction</h2>
8
-
9
-
This document describes RCU's expedited grace periods.
10
-
Unlike RCU's normal grace periods, which accept long latencies to attain
11
-
high efficiency and minimal disturbance, expedited grace periods accept
12
-
lower efficiency and significant disturbance to attain shorter latencies.
13
-
14
-
<p>
15
-
There are two flavors of RCU (RCU-preempt and RCU-sched), with an earlier
16
-
third RCU-bh flavor having been implemented in terms of the other two.
17
-
Each of the two implementations is covered in its own section.
18
-
19
-
<ol>
20
-
<li> <a href="#Expedited Grace Period Design">
21
-
Expedited Grace Period Design</a>
22
-
<li> <a href="#RCU-preempt Expedited Grace Periods">
23
-
RCU-preempt Expedited Grace Periods</a>
24
-
<li> <a href="#RCU-sched Expedited Grace Periods">
25
-
RCU-sched Expedited Grace Periods</a>
26
-
<li> <a href="#Expedited Grace Period and CPU Hotplug">
27
-
Expedited Grace Period and CPU Hotplug</a>
28
-
<li> <a href="#Expedited Grace Period Refinements">
29
-
Expedited Grace Period Refinements</a>
30
-
</ol>
31
-
32
-
<h2><a name="Expedited Grace Period Design">
33
-
Expedited Grace Period Design</a></h2>
34
-
35
-
<p>
36
-
The expedited RCU grace periods cannot be accused of being subtle,
37
-
given that they for all intents and purposes hammer every CPU that
38
-
has not yet provided a quiescent state for the current expedited
39
-
grace period.
40
-
The one saving grace is that the hammer has grown a bit smaller
41
-
over time: The old call to <tt>try_stop_cpus()</tt> has been
42
-
replaced with a set of calls to <tt>smp_call_function_single()</tt>,
43
-
each of which results in an IPI to the target CPU.
44
-
The corresponding handler function checks the CPU's state, motivating
45
-
a faster quiescent state where possible, and triggering a report
46
-
of that quiescent state.
47
-
As always for RCU, once everything has spent some time in a quiescent
48
-
state, the expedited grace period has completed.
49
-
50
-
<p>
51
-
The details of the <tt>smp_call_function_single()</tt> handler's
52
-
operation depend on the RCU flavor, as described in the following
53
-
sections.
54
-
55
-
<h2><a name="RCU-preempt Expedited Grace Periods">
56
-
RCU-preempt Expedited Grace Periods</a></h2>
57
-
58
-
<p>
59
-
<tt>CONFIG_PREEMPT=y</tt> kernels implement RCU-preempt.
60
-
The overall flow of the handling of a given CPU by an RCU-preempt
61
-
expedited grace period is shown in the following diagram:
62
-
63
-
<p><img src="ExpRCUFlow.svg" alt="ExpRCUFlow.svg" width="55%">
64
-
65
-
<p>
66
-
The solid arrows denote direct action, for example, a function call.
67
-
The dotted arrows denote indirect action, for example, an IPI
68
-
or a state that is reached after some time.
69
-
70
-
<p>
71
-
If a given CPU is offline or idle, <tt>synchronize_rcu_expedited()</tt>
72
-
will ignore it because idle and offline CPUs are already residing
73
-
in quiescent states.
74
-
Otherwise, the expedited grace period will use
75
-
<tt>smp_call_function_single()</tt> to send the CPU an IPI, which
76
-
is handled by <tt>rcu_exp_handler()</tt>.
77
-
78
-
<p>
79
-
However, because this is preemptible RCU, <tt>rcu_exp_handler()</tt>
80
-
can check to see if the CPU is currently running in an RCU read-side
81
-
critical section.
82
-
If not, the handler can immediately report a quiescent state.
83
-
Otherwise, it sets flags so that the outermost <tt>rcu_read_unlock()</tt>
84
-
invocation will provide the needed quiescent-state report.
85
-
This flag-setting avoids the previous forced preemption of all
86
-
CPUs that might have RCU read-side critical sections.
87
-
In addition, this flag-setting is done so as to avoid increasing
88
-
the overhead of the common-case fastpath through the scheduler.
89
-
90
-
<p>
91
-
Again because this is preemptible RCU, an RCU read-side critical section
92
-
can be preempted.
93
-
When that happens, RCU will enqueue the task, which will the continue to
94
-
block the current expedited grace period until it resumes and finds its
95
-
outermost <tt>rcu_read_unlock()</tt>.
96
-
The CPU will report a quiescent state just after enqueuing the task because
97
-
the CPU is no longer blocking the grace period.
98
-
It is instead the preempted task doing the blocking.
99
-
The list of blocked tasks is managed by <tt>rcu_preempt_ctxt_queue()</tt>,
100
-
which is called from <tt>rcu_preempt_note_context_switch()</tt>, which
101
-
in turn is called from <tt>rcu_note_context_switch()</tt>, which in
102
-
turn is called from the scheduler.
103
-
104
-
<table>
105
-
<tr><th> </th></tr>
106
-
<tr><th align="left">Quick Quiz:</th></tr>
107
-
<tr><td>
108
-
Why not just have the expedited grace period check the
109
-
state of all the CPUs?
110
-
After all, that would avoid all those real-time-unfriendly IPIs.
111
-
</td></tr>
112
-
<tr><th align="left">Answer:</th></tr>
113
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
114
-
Because we want the RCU read-side critical sections to run fast,
115
-
which means no memory barriers.
116
-
Therefore, it is not possible to safely check the state from some
117
-
other CPU.
118
-
And even if it was possible to safely check the state, it would
119
-
still be necessary to IPI the CPU to safely interact with the
120
-
upcoming <tt>rcu_read_unlock()</tt> invocation, which means that
121
-
the remote state testing would not help the worst-case
122
-
latency that real-time applications care about.
123
-
124
-
<p><font color="ffffff">One way to prevent your real-time
125
-
application from getting hit with these IPIs is to
126
-
build your kernel with <tt>CONFIG_NO_HZ_FULL=y</tt>.
127
-
RCU would then perceive the CPU running your application
128
-
as being idle, and it would be able to safely detect that
129
-
state without needing to IPI the CPU.
130
-
</font></td></tr>
131
-
<tr><td> </td></tr>
132
-
</table>
133
-
134
-
<p>
135
-
Please note that this is just the overall flow:
136
-
Additional complications can arise due to races with CPUs going idle
137
-
or offline, among other things.
138
-
139
-
<h2><a name="RCU-sched Expedited Grace Periods">
140
-
RCU-sched Expedited Grace Periods</a></h2>
141
-
142
-
<p>
143
-
<tt>CONFIG_PREEMPT=n</tt> kernels implement RCU-sched.
144
-
The overall flow of the handling of a given CPU by an RCU-sched
145
-
expedited grace period is shown in the following diagram:
146
-
147
-
<p><img src="ExpSchedFlow.svg" alt="ExpSchedFlow.svg" width="55%">
148
-
149
-
<p>
150
-
As with RCU-preempt, RCU-sched's
151
-
<tt>synchronize_rcu_expedited()</tt> ignores offline and
152
-
idle CPUs, again because they are in remotely detectable
153
-
quiescent states.
154
-
However, because the
155
-
<tt>rcu_read_lock_sched()</tt> and <tt>rcu_read_unlock_sched()</tt>
156
-
leave no trace of their invocation, in general it is not possible to tell
157
-
whether or not the current CPU is in an RCU read-side critical section.
158
-
The best that RCU-sched's <tt>rcu_exp_handler()</tt> can do is to check
159
-
for idle, on the off-chance that the CPU went idle while the IPI
160
-
was in flight.
161
-
If the CPU is idle, then <tt>rcu_exp_handler()</tt> reports
162
-
the quiescent state.
163
-
164
-
<p> Otherwise, the handler forces a future context switch by setting the
165
-
NEED_RESCHED flag of the current task's thread flag and the CPU preempt
166
-
counter.
167
-
At the time of the context switch, the CPU reports the quiescent state.
168
-
Should the CPU go offline first, it will report the quiescent state
169
-
at that time.
170
-
171
-
<h2><a name="Expedited Grace Period and CPU Hotplug">
172
-
Expedited Grace Period and CPU Hotplug</a></h2>
173
-
174
-
<p>
175
-
The expedited nature of expedited grace periods require a much tighter
176
-
interaction with CPU hotplug operations than is required for normal
177
-
grace periods.
178
-
In addition, attempting to IPI offline CPUs will result in splats, but
179
-
failing to IPI online CPUs can result in too-short grace periods.
180
-
Neither option is acceptable in production kernels.
181
-
182
-
<p>
183
-
The interaction between expedited grace periods and CPU hotplug operations
184
-
is carried out at several levels:
185
-
186
-
<ol>
187
-
<li> The number of CPUs that have ever been online is tracked
188
-
by the <tt>rcu_state</tt> structure's <tt>->ncpus</tt>
189
-
field.
190
-
The <tt>rcu_state</tt> structure's <tt>->ncpus_snap</tt>
191
-
field tracks the number of CPUs that have ever been online
192
-
at the beginning of an RCU expedited grace period.
193
-
Note that this number never decreases, at least in the absence
194
-
of a time machine.
195
-
<li> The identities of the CPUs that have ever been online is
196
-
tracked by the <tt>rcu_node</tt> structure's
197
-
<tt>->expmaskinitnext</tt> field.
198
-
The <tt>rcu_node</tt> structure's <tt>->expmaskinit</tt>
199
-
field tracks the identities of the CPUs that were online
200
-
at least once at the beginning of the most recent RCU
201
-
expedited grace period.
202
-
The <tt>rcu_state</tt> structure's <tt>->ncpus</tt> and
203
-
<tt>->ncpus_snap</tt> fields are used to detect when
204
-
new CPUs have come online for the first time, that is,
205
-
when the <tt>rcu_node</tt> structure's <tt>->expmaskinitnext</tt>
206
-
field has changed since the beginning of the last RCU
207
-
expedited grace period, which triggers an update of each
208
-
<tt>rcu_node</tt> structure's <tt>->expmaskinit</tt>
209
-
field from its <tt>->expmaskinitnext</tt> field.
210
-
<li> Each <tt>rcu_node</tt> structure's <tt>->expmaskinit</tt>
211
-
field is used to initialize that structure's
212
-
<tt>->expmask</tt> at the beginning of each RCU
213
-
expedited grace period.
214
-
This means that only those CPUs that have been online at least
215
-
once will be considered for a given grace period.
216
-
<li> Any CPU that goes offline will clear its bit in its leaf
217
-
<tt>rcu_node</tt> structure's <tt>->qsmaskinitnext</tt>
218
-
field, so any CPU with that bit clear can safely be ignored.
219
-
However, it is possible for a CPU coming online or going offline
220
-
to have this bit set for some time while <tt>cpu_online</tt>
221
-
returns <tt>false</tt>.
222
-
<li> For each non-idle CPU that RCU believes is currently online, the grace
223
-
period invokes <tt>smp_call_function_single()</tt>.
224
-
If this succeeds, the CPU was fully online.
225
-
Failure indicates that the CPU is in the process of coming online
226
-
or going offline, in which case it is necessary to wait for a
227
-
short time period and try again.
228
-
The purpose of this wait (or series of waits, as the case may be)
229
-
is to permit a concurrent CPU-hotplug operation to complete.
230
-
<li> In the case of RCU-sched, one of the last acts of an outgoing CPU
231
-
is to invoke <tt>rcu_report_dead()</tt>, which
232
-
reports a quiescent state for that CPU.
233
-
However, this is likely paranoia-induced redundancy. <!-- @@@ -->
234
-
</ol>
235
-
236
-
<table>
237
-
<tr><th> </th></tr>
238
-
<tr><th align="left">Quick Quiz:</th></tr>
239
-
<tr><td>
240
-
Why all the dancing around with multiple counters and masks
241
-
tracking CPUs that were once online?
242
-
Why not just have a single set of masks tracking the currently
243
-
online CPUs and be done with it?
244
-
</td></tr>
245
-
<tr><th align="left">Answer:</th></tr>
246
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
247
-
Maintaining single set of masks tracking the online CPUs <i>sounds</i>
248
-
easier, at least until you try working out all the race conditions
249
-
between grace-period initialization and CPU-hotplug operations.
250
-
For example, suppose initialization is progressing down the
251
-
tree while a CPU-offline operation is progressing up the tree.
252
-
This situation can result in bits set at the top of the tree
253
-
that have no counterparts at the bottom of the tree.
254
-
Those bits will never be cleared, which will result in
255
-
grace-period hangs.
256
-
In short, that way lies madness, to say nothing of a great many
257
-
bugs, hangs, and deadlocks.
258
-
259
-
<p><font color="ffffff">
260
-
In contrast, the current multi-mask multi-counter scheme ensures
261
-
that grace-period initialization will always see consistent masks
262
-
up and down the tree, which brings significant simplifications
263
-
over the single-mask method.
264
-
265
-
<p><font color="ffffff">
266
-
This is an instance of
267
-
<a href="http://www.cs.columbia.edu/~library/TR-repository/reports/reports-1992/cucs-039-92.ps.gz"><font color="ffffff">
268
-
deferring work in order to avoid synchronization</a>.
269
-
Lazily recording CPU-hotplug events at the beginning of the next
270
-
grace period greatly simplifies maintenance of the CPU-tracking
271
-
bitmasks in the <tt>rcu_node</tt> tree.
272
-
</font></td></tr>
273
-
<tr><td> </td></tr>
274
-
</table>
275
-
276
-
<h2><a name="Expedited Grace Period Refinements">
277
-
Expedited Grace Period Refinements</a></h2>
278
-
279
-
<ol>
280
-
<li> <a href="#Idle-CPU Checks">Idle-CPU checks</a>.
281
-
<li> <a href="#Batching via Sequence Counter">
282
-
Batching via sequence counter</a>.
283
-
<li> <a href="#Funnel Locking and Wait/Wakeup">
284
-
Funnel locking and wait/wakeup</a>.
285
-
<li> <a href="#Use of Workqueues">Use of Workqueues</a>.
286
-
<li> <a href="#Stall Warnings">Stall warnings</a>.
287
-
<li> <a href="#Mid-Boot Operation">Mid-boot operation</a>.
288
-
</ol>
289
-
290
-
<h3><a name="Idle-CPU Checks">Idle-CPU Checks</a></h3>
291
-
292
-
<p>
293
-
Each expedited grace period checks for idle CPUs when initially forming
294
-
the mask of CPUs to be IPIed and again just before IPIing a CPU
295
-
(both checks are carried out by <tt>sync_rcu_exp_select_cpus()</tt>).
296
-
If the CPU is idle at any time between those two times, the CPU will
297
-
not be IPIed.
298
-
Instead, the task pushing the grace period forward will include the
299
-
idle CPUs in the mask passed to <tt>rcu_report_exp_cpu_mult()</tt>.
300
-
301
-
<p>
302
-
For RCU-sched, there is an additional check:
303
-
If the IPI has interrupted the idle loop, then
304
-
<tt>rcu_exp_handler()</tt> invokes <tt>rcu_report_exp_rdp()</tt>
305
-
to report the corresponding quiescent state.
306
-
307
-
<p>
308
-
For RCU-preempt, there is no specific check for idle in the
309
-
IPI handler (<tt>rcu_exp_handler()</tt>), but because
310
-
RCU read-side critical sections are not permitted within the
311
-
idle loop, if <tt>rcu_exp_handler()</tt> sees that the CPU is within
312
-
RCU read-side critical section, the CPU cannot possibly be idle.
313
-
Otherwise, <tt>rcu_exp_handler()</tt> invokes
314
-
<tt>rcu_report_exp_rdp()</tt> to report the corresponding quiescent
315
-
state, regardless of whether or not that quiescent state was due to
316
-
the CPU being idle.
317
-
318
-
<p>
319
-
In summary, RCU expedited grace periods check for idle when building
320
-
the bitmask of CPUs that must be IPIed, just before sending each IPI,
321
-
and (either explicitly or implicitly) within the IPI handler.
322
-
323
-
<h3><a name="Batching via Sequence Counter">
324
-
Batching via Sequence Counter</a></h3>
325
-
326
-
<p>
327
-
If each grace-period request was carried out separately, expedited
328
-
grace periods would have abysmal scalability and
329
-
problematic high-load characteristics.
330
-
Because each grace-period operation can serve an unlimited number of
331
-
updates, it is important to <i>batch</i> requests, so that a single
332
-
expedited grace-period operation will cover all requests in the
333
-
corresponding batch.
334
-
335
-
<p>
336
-
This batching is controlled by a sequence counter named
337
-
<tt>->expedited_sequence</tt> in the <tt>rcu_state</tt> structure.
338
-
This counter has an odd value when there is an expedited grace period
339
-
in progress and an even value otherwise, so that dividing the counter
340
-
value by two gives the number of completed grace periods.
341
-
During any given update request, the counter must transition from
342
-
even to odd and then back to even, thus indicating that a grace
343
-
period has elapsed.
344
-
Therefore, if the initial value of the counter is <tt>s</tt>,
345
-
the updater must wait until the counter reaches at least the
346
-
value <tt>(s+3)&~0x1</tt>.
347
-
This counter is managed by the following access functions:
348
-
349
-
<ol>
350
-
<li> <tt>rcu_exp_gp_seq_start()</tt>, which marks the start of
351
-
an expedited grace period.
352
-
<li> <tt>rcu_exp_gp_seq_end()</tt>, which marks the end of an
353
-
expedited grace period.
354
-
<li> <tt>rcu_exp_gp_seq_snap()</tt>, which obtains a snapshot of
355
-
the counter.
356
-
<li> <tt>rcu_exp_gp_seq_done()</tt>, which returns <tt>true</tt>
357
-
if a full expedited grace period has elapsed since the
358
-
corresponding call to <tt>rcu_exp_gp_seq_snap()</tt>.
359
-
</ol>
360
-
361
-
<p>
362
-
Again, only one request in a given batch need actually carry out
363
-
a grace-period operation, which means there must be an efficient
364
-
way to identify which of many concurrent reqeusts will initiate
365
-
the grace period, and that there be an efficient way for the
366
-
remaining requests to wait for that grace period to complete.
367
-
However, that is the topic of the next section.
368
-
369
-
<h3><a name="Funnel Locking and Wait/Wakeup">
370
-
Funnel Locking and Wait/Wakeup</a></h3>
371
-
372
-
<p>
373
-
The natural way to sort out which of a batch of updaters will initiate
374
-
the expedited grace period is to use the <tt>rcu_node</tt> combining
375
-
tree, as implemented by the <tt>exp_funnel_lock()</tt> function.
376
-
The first updater corresponding to a given grace period arriving
377
-
at a given <tt>rcu_node</tt> structure records its desired grace-period
378
-
sequence number in the <tt>->exp_seq_rq</tt> field and moves up
379
-
to the next level in the tree.
380
-
Otherwise, if the <tt>->exp_seq_rq</tt> field already contains
381
-
the sequence number for the desired grace period or some later one,
382
-
the updater blocks on one of four wait queues in the
383
-
<tt>->exp_wq[]</tt> array, using the second-from-bottom
384
-
and third-from bottom bits as an index.
385
-
An <tt>->exp_lock</tt> field in the <tt>rcu_node</tt> structure
386
-
synchronizes access to these fields.
387
-
388
-
<p>
389
-
An empty <tt>rcu_node</tt> tree is shown in the following diagram,
390
-
with the white cells representing the <tt>->exp_seq_rq</tt> field
391
-
and the red cells representing the elements of the
392
-
<tt>->exp_wq[]</tt> array.
393
-
394
-
<p><img src="Funnel0.svg" alt="Funnel0.svg" width="75%">
395
-
396
-
<p>
397
-
The next diagram shows the situation after the arrival of Task A
398
-
and Task B at the leftmost and rightmost leaf <tt>rcu_node</tt>
399
-
structures, respectively.
400
-
The current value of the <tt>rcu_state</tt> structure's
401
-
<tt>->expedited_sequence</tt> field is zero, so adding three and
402
-
clearing the bottom bit results in the value two, which both tasks
403
-
record in the <tt>->exp_seq_rq</tt> field of their respective
404
-
<tt>rcu_node</tt> structures:
405
-
406
-
<p><img src="Funnel1.svg" alt="Funnel1.svg" width="75%">
407
-
408
-
<p>
409
-
Each of Tasks A and B will move up to the root
410
-
<tt>rcu_node</tt> structure.
411
-
Suppose that Task A wins, recording its desired grace-period sequence
412
-
number and resulting in the state shown below:
413
-
414
-
<p><img src="Funnel2.svg" alt="Funnel2.svg" width="75%">
415
-
416
-
<p>
417
-
Task A now advances to initiate a new grace period, while Task B
418
-
moves up to the root <tt>rcu_node</tt> structure, and, seeing that
419
-
its desired sequence number is already recorded, blocks on
420
-
<tt>->exp_wq[1]</tt>.
421
-
422
-
<table>
423
-
<tr><th> </th></tr>
424
-
<tr><th align="left">Quick Quiz:</th></tr>
425
-
<tr><td>
426
-
Why <tt>->exp_wq[1]</tt>?
427
-
Given that the value of these tasks' desired sequence number is
428
-
two, so shouldn't they instead block on <tt>->exp_wq[2]</tt>?
429
-
</td></tr>
430
-
<tr><th align="left">Answer:</th></tr>
431
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
432
-
No.
433
-
434
-
<p><font color="ffffff">
435
-
Recall that the bottom bit of the desired sequence number indicates
436
-
whether or not a grace period is currently in progress.
437
-
It is therefore necessary to shift the sequence number right one
438
-
bit position to obtain the number of the grace period.
439
-
This results in <tt>->exp_wq[1]</tt>.
440
-
</font></td></tr>
441
-
<tr><td> </td></tr>
442
-
</table>
443
-
444
-
<p>
445
-
If Tasks C and D also arrive at this point, they will compute the
446
-
same desired grace-period sequence number, and see that both leaf
447
-
<tt>rcu_node</tt> structures already have that value recorded.
448
-
They will therefore block on their respective <tt>rcu_node</tt>
449
-
structures' <tt>->exp_wq[1]</tt> fields, as shown below:
450
-
451
-
<p><img src="Funnel3.svg" alt="Funnel3.svg" width="75%">
452
-
453
-
<p>
454
-
Task A now acquires the <tt>rcu_state</tt> structure's
455
-
<tt>->exp_mutex</tt> and initiates the grace period, which
456
-
increments <tt>->expedited_sequence</tt>.
457
-
Therefore, if Tasks E and F arrive, they will compute
458
-
a desired sequence number of 4 and will record this value as
459
-
shown below:
460
-
461
-
<p><img src="Funnel4.svg" alt="Funnel4.svg" width="75%">
462
-
463
-
<p>
464
-
Tasks E and F will propagate up the <tt>rcu_node</tt>
465
-
combining tree, with Task F blocking on the root <tt>rcu_node</tt>
466
-
structure and Task E wait for Task A to finish so that
467
-
it can start the next grace period.
468
-
The resulting state is as shown below:
469
-
470
-
<p><img src="Funnel5.svg" alt="Funnel5.svg" width="75%">
471
-
472
-
<p>
473
-
Once the grace period completes, Task A
474
-
starts waking up the tasks waiting for this grace period to complete,
475
-
increments the <tt>->expedited_sequence</tt>,
476
-
acquires the <tt>->exp_wake_mutex</tt> and then releases the
477
-
<tt>->exp_mutex</tt>.
478
-
This results in the following state:
479
-
480
-
<p><img src="Funnel6.svg" alt="Funnel6.svg" width="75%">
481
-
482
-
<p>
483
-
Task E can then acquire <tt>->exp_mutex</tt> and increment
484
-
<tt>->expedited_sequence</tt> to the value three.
485
-
If new tasks G and H arrive and moves up the combining tree at the
486
-
same time, the state will be as follows:
487
-
488
-
<p><img src="Funnel7.svg" alt="Funnel7.svg" width="75%">
489
-
490
-
<p>
491
-
Note that three of the root <tt>rcu_node</tt> structure's
492
-
waitqueues are now occupied.
493
-
However, at some point, Task A will wake up the
494
-
tasks blocked on the <tt>->exp_wq</tt> waitqueues, resulting
495
-
in the following state:
496
-
497
-
<p><img src="Funnel8.svg" alt="Funnel8.svg" width="75%">
498
-
499
-
<p>
500
-
Execution will continue with Tasks E and H completing
501
-
their grace periods and carrying out their wakeups.
502
-
503
-
<table>
504
-
<tr><th> </th></tr>
505
-
<tr><th align="left">Quick Quiz:</th></tr>
506
-
<tr><td>
507
-
What happens if Task A takes so long to do its wakeups
508
-
that Task E's grace period completes?
509
-
</td></tr>
510
-
<tr><th align="left">Answer:</th></tr>
511
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
512
-
Then Task E will block on the <tt>->exp_wake_mutex</tt>,
513
-
which will also prevent it from releasing <tt>->exp_mutex</tt>,
514
-
which in turn will prevent the next grace period from starting.
515
-
This last is important in preventing overflow of the
516
-
<tt>->exp_wq[]</tt> array.
517
-
</font></td></tr>
518
-
<tr><td> </td></tr>
519
-
</table>
520
-
521
-
<h3><a name="Use of Workqueues">Use of Workqueues</a></h3>
522
-
523
-
<p>
524
-
In earlier implementations, the task requesting the expedited
525
-
grace period also drove it to completion.
526
-
This straightforward approach had the disadvantage of needing to
527
-
account for POSIX signals sent to user tasks,
528
-
so more recent implemementations use the Linux kernel's
529
-
<a href="https://www.kernel.org/doc/Documentation/core-api/workqueue.rst">workqueues</a>.
530
-
531
-
<p>
532
-
The requesting task still does counter snapshotting and funnel-lock
533
-
processing, but the task reaching the top of the funnel lock
534
-
does a <tt>schedule_work()</tt> (from <tt>_synchronize_rcu_expedited()</tt>
535
-
so that a workqueue kthread does the actual grace-period processing.
536
-
Because workqueue kthreads do not accept POSIX signals, grace-period-wait
537
-
processing need not allow for POSIX signals.
538
-
539
-
In addition, this approach allows wakeups for the previous expedited
540
-
grace period to be overlapped with processing for the next expedited
541
-
grace period.
542
-
Because there are only four sets of waitqueues, it is necessary to
543
-
ensure that the previous grace period's wakeups complete before the
544
-
next grace period's wakeups start.
545
-
This is handled by having the <tt>->exp_mutex</tt>
546
-
guard expedited grace-period processing and the
547
-
<tt>->exp_wake_mutex</tt> guard wakeups.
548
-
The key point is that the <tt>->exp_mutex</tt> is not released
549
-
until the first wakeup is complete, which means that the
550
-
<tt>->exp_wake_mutex</tt> has already been acquired at that point.
551
-
This approach ensures that the previous grace period's wakeups can
552
-
be carried out while the current grace period is in process, but
553
-
that these wakeups will complete before the next grace period starts.
554
-
This means that only three waitqueues are required, guaranteeing that
555
-
the four that are provided are sufficient.
556
-
557
-
<h3><a name="Stall Warnings">Stall Warnings</a></h3>
558
-
559
-
<p>
560
-
Expediting grace periods does nothing to speed things up when RCU
561
-
readers take too long, and therefore expedited grace periods check
562
-
for stalls just as normal grace periods do.
563
-
564
-
<table>
565
-
<tr><th> </th></tr>
566
-
<tr><th align="left">Quick Quiz:</th></tr>
567
-
<tr><td>
568
-
But why not just let the normal grace-period machinery
569
-
detect the stalls, given that a given reader must block
570
-
both normal and expedited grace periods?
571
-
</td></tr>
572
-
<tr><th align="left">Answer:</th></tr>
573
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
574
-
Because it is quite possible that at a given time there
575
-
is no normal grace period in progress, in which case the
576
-
normal grace period cannot emit a stall warning.
577
-
</font></td></tr>
578
-
<tr><td> </td></tr>
579
-
</table>
580
-
581
-
The <tt>synchronize_sched_expedited_wait()</tt> function loops waiting
582
-
for the expedited grace period to end, but with a timeout set to the
583
-
current RCU CPU stall-warning time.
584
-
If this time is exceeded, any CPUs or <tt>rcu_node</tt> structures
585
-
blocking the current grace period are printed.
586
-
Each stall warning results in another pass through the loop, but the
587
-
second and subsequent passes use longer stall times.
588
-
589
-
<h3><a name="Mid-Boot Operation">Mid-boot operation</a></h3>
590
-
591
-
<p>
592
-
The use of workqueues has the advantage that the expedited
593
-
grace-period code need not worry about POSIX signals.
594
-
Unfortunately, it has the
595
-
corresponding disadvantage that workqueues cannot be used until
596
-
they are initialized, which does not happen until some time after
597
-
the scheduler spawns the first task.
598
-
Given that there are parts of the kernel that really do want to
599
-
execute grace periods during this mid-boot “dead zone”,
600
-
expedited grace periods must do something else during thie time.
601
-
602
-
<p>
603
-
What they do is to fall back to the old practice of requiring that the
604
-
requesting task drive the expedited grace period, as was the case
605
-
before the use of workqueues.
606
-
However, the requesting task is only required to drive the grace period
607
-
during the mid-boot dead zone.
608
-
Before mid-boot, a synchronous grace period is a no-op.
609
-
Some time after mid-boot, workqueues are used.
610
-
611
-
<p>
612
-
Non-expedited non-SRCU synchronous grace periods must also operate
613
-
normally during mid-boot.
614
-
This is handled by causing non-expedited grace periods to take the
615
-
expedited code path during mid-boot.
616
-
617
-
<p>
618
-
The current code assumes that there are no POSIX signals during
619
-
the mid-boot dead zone.
620
-
However, if an overwhelming need for POSIX signals somehow arises,
621
-
appropriate adjustments can be made to the expedited stall-warning code.
622
-
One such adjustment would reinstate the pre-workqueue stall-warning
623
-
checks, but only during the mid-boot dead zone.
624
-
625
-
<p>
626
-
With this refinement, synchronous grace periods can now be used from
627
-
task context pretty much any time during the life of the kernel.
628
-
That is, aside from some points in the suspend, hibernate, or shutdown
629
-
code path.
630
-
631
-
<h3><a name="Summary">
632
-
Summary</a></h3>
633
-
634
-
<p>
635
-
Expedited grace periods use a sequence-number approach to promote
636
-
batching, so that a single grace-period operation can serve numerous
637
-
requests.
638
-
A funnel lock is used to efficiently identify the one task out of
639
-
a concurrent group that will request the grace period.
640
-
All members of the group will block on waitqueues provided in
641
-
the <tt>rcu_node</tt> structure.
642
-
The actual grace-period processing is carried out by a workqueue.
643
-
644
-
<p>
645
-
CPU-hotplug operations are noted lazily in order to prevent the need
646
-
for tight synchronization between expedited grace periods and
647
-
CPU-hotplug operations.
648
-
The dyntick-idle counters are used to avoid sending IPIs to idle CPUs,
649
-
at least in the common case.
650
-
RCU-preempt and RCU-sched use different IPI handlers and different
651
-
code to respond to the state changes carried out by those handlers,
652
-
but otherwise use common code.
653
-
654
-
<p>
655
-
Quiescent states are tracked using the <tt>rcu_node</tt> tree,
656
-
and once all necessary quiescent states have been reported,
657
-
all tasks waiting on this expedited grace period are awakened.
658
-
A pair of mutexes are used to allow one grace period's wakeups
659
-
to proceed concurrently with the next grace period's processing.
660
-
661
-
<p>
662
-
This combination of mechanisms allows expedited grace periods to
663
-
run reasonably efficiently.
664
-
However, for non-time-critical tasks, normal grace periods should be
665
-
used instead because their longer duration permits much higher
666
-
degrees of batching, and thus much lower per-request overheads.
667
-
668
-
</body></html>
+521
Documentation/RCU/Design/Expedited-Grace-Periods/Expedited-Grace-Periods.rst
+521
Documentation/RCU/Design/Expedited-Grace-Periods/Expedited-Grace-Periods.rst
···
1
+
=================================================
2
+
A Tour Through TREE_RCU's Expedited Grace Periods
3
+
=================================================
4
+
5
+
Introduction
6
+
============
7
+
8
+
This document describes RCU's expedited grace periods.
9
+
Unlike RCU's normal grace periods, which accept long latencies to attain
10
+
high efficiency and minimal disturbance, expedited grace periods accept
11
+
lower efficiency and significant disturbance to attain shorter latencies.
12
+
13
+
There are two flavors of RCU (RCU-preempt and RCU-sched), with an earlier
14
+
third RCU-bh flavor having been implemented in terms of the other two.
15
+
Each of the two implementations is covered in its own section.
16
+
17
+
Expedited Grace Period Design
18
+
=============================
19
+
20
+
The expedited RCU grace periods cannot be accused of being subtle,
21
+
given that they for all intents and purposes hammer every CPU that
22
+
has not yet provided a quiescent state for the current expedited
23
+
grace period.
24
+
The one saving grace is that the hammer has grown a bit smaller
25
+
over time: The old call to ``try_stop_cpus()`` has been
26
+
replaced with a set of calls to ``smp_call_function_single()``,
27
+
each of which results in an IPI to the target CPU.
28
+
The corresponding handler function checks the CPU's state, motivating
29
+
a faster quiescent state where possible, and triggering a report
30
+
of that quiescent state.
31
+
As always for RCU, once everything has spent some time in a quiescent
32
+
state, the expedited grace period has completed.
33
+
34
+
The details of the ``smp_call_function_single()`` handler's
35
+
operation depend on the RCU flavor, as described in the following
36
+
sections.
37
+
38
+
RCU-preempt Expedited Grace Periods
39
+
===================================
40
+
41
+
``CONFIG_PREEMPT=y`` kernels implement RCU-preempt.
42
+
The overall flow of the handling of a given CPU by an RCU-preempt
43
+
expedited grace period is shown in the following diagram:
44
+
45
+
.. kernel-figure:: ExpRCUFlow.svg
46
+
47
+
The solid arrows denote direct action, for example, a function call.
48
+
The dotted arrows denote indirect action, for example, an IPI
49
+
or a state that is reached after some time.
50
+
51
+
If a given CPU is offline or idle, ``synchronize_rcu_expedited()``
52
+
will ignore it because idle and offline CPUs are already residing
53
+
in quiescent states.
54
+
Otherwise, the expedited grace period will use
55
+
``smp_call_function_single()`` to send the CPU an IPI, which
56
+
is handled by ``rcu_exp_handler()``.
57
+
58
+
However, because this is preemptible RCU, ``rcu_exp_handler()``
59
+
can check to see if the CPU is currently running in an RCU read-side
60
+
critical section.
61
+
If not, the handler can immediately report a quiescent state.
62
+
Otherwise, it sets flags so that the outermost ``rcu_read_unlock()``
63
+
invocation will provide the needed quiescent-state report.
64
+
This flag-setting avoids the previous forced preemption of all
65
+
CPUs that might have RCU read-side critical sections.
66
+
In addition, this flag-setting is done so as to avoid increasing
67
+
the overhead of the common-case fastpath through the scheduler.
68
+
69
+
Again because this is preemptible RCU, an RCU read-side critical section
70
+
can be preempted.
71
+
When that happens, RCU will enqueue the task, which will the continue to
72
+
block the current expedited grace period until it resumes and finds its
73
+
outermost ``rcu_read_unlock()``.
74
+
The CPU will report a quiescent state just after enqueuing the task because
75
+
the CPU is no longer blocking the grace period.
76
+
It is instead the preempted task doing the blocking.
77
+
The list of blocked tasks is managed by ``rcu_preempt_ctxt_queue()``,
78
+
which is called from ``rcu_preempt_note_context_switch()``, which
79
+
in turn is called from ``rcu_note_context_switch()``, which in
80
+
turn is called from the scheduler.
81
+
82
+
83
+
+-----------------------------------------------------------------------+
84
+
| **Quick Quiz**: |
85
+
+-----------------------------------------------------------------------+
86
+
| Why not just have the expedited grace period check the state of all |
87
+
| the CPUs? After all, that would avoid all those real-time-unfriendly |
88
+
| IPIs. |
89
+
+-----------------------------------------------------------------------+
90
+
| **Answer**: |
91
+
+-----------------------------------------------------------------------+
92
+
| Because we want the RCU read-side critical sections to run fast, |
93
+
| which means no memory barriers. Therefore, it is not possible to |
94
+
| safely check the state from some other CPU. And even if it was |
95
+
| possible to safely check the state, it would still be necessary to |
96
+
| IPI the CPU to safely interact with the upcoming |
97
+
| ``rcu_read_unlock()`` invocation, which means that the remote state |
98
+
| testing would not help the worst-case latency that real-time |
99
+
| applications care about. |
100
+
| |
101
+
| One way to prevent your real-time application from getting hit with |
102
+
| these IPIs is to build your kernel with ``CONFIG_NO_HZ_FULL=y``. RCU |
103
+
| would then perceive the CPU running your application as being idle, |
104
+
| and it would be able to safely detect that state without needing to |
105
+
| IPI the CPU. |
106
+
+-----------------------------------------------------------------------+
107
+
108
+
Please note that this is just the overall flow: Additional complications
109
+
can arise due to races with CPUs going idle or offline, among other
110
+
things.
111
+
112
+
RCU-sched Expedited Grace Periods
113
+
---------------------------------
114
+
115
+
``CONFIG_PREEMPT=n`` kernels implement RCU-sched. The overall flow of
116
+
the handling of a given CPU by an RCU-sched expedited grace period is
117
+
shown in the following diagram:
118
+
119
+
.. kernel-figure:: ExpSchedFlow.svg
120
+
121
+
As with RCU-preempt, RCU-sched's ``synchronize_rcu_expedited()`` ignores
122
+
offline and idle CPUs, again because they are in remotely detectable
123
+
quiescent states. However, because the ``rcu_read_lock_sched()`` and
124
+
``rcu_read_unlock_sched()`` leave no trace of their invocation, in
125
+
general it is not possible to tell whether or not the current CPU is in
126
+
an RCU read-side critical section. The best that RCU-sched's
127
+
``rcu_exp_handler()`` can do is to check for idle, on the off-chance
128
+
that the CPU went idle while the IPI was in flight. If the CPU is idle,
129
+
then ``rcu_exp_handler()`` reports the quiescent state.
130
+
131
+
Otherwise, the handler forces a future context switch by setting the
132
+
NEED_RESCHED flag of the current task's thread flag and the CPU preempt
133
+
counter. At the time of the context switch, the CPU reports the
134
+
quiescent state. Should the CPU go offline first, it will report the
135
+
quiescent state at that time.
136
+
137
+
Expedited Grace Period and CPU Hotplug
138
+
--------------------------------------
139
+
140
+
The expedited nature of expedited grace periods require a much tighter
141
+
interaction with CPU hotplug operations than is required for normal
142
+
grace periods. In addition, attempting to IPI offline CPUs will result
143
+
in splats, but failing to IPI online CPUs can result in too-short grace
144
+
periods. Neither option is acceptable in production kernels.
145
+
146
+
The interaction between expedited grace periods and CPU hotplug
147
+
operations is carried out at several levels:
148
+
149
+
#. The number of CPUs that have ever been online is tracked by the
150
+
``rcu_state`` structure's ``->ncpus`` field. The ``rcu_state``
151
+
structure's ``->ncpus_snap`` field tracks the number of CPUs that
152
+
have ever been online at the beginning of an RCU expedited grace
153
+
period. Note that this number never decreases, at least in the
154
+
absence of a time machine.
155
+
#. The identities of the CPUs that have ever been online is tracked by
156
+
the ``rcu_node`` structure's ``->expmaskinitnext`` field. The
157
+
``rcu_node`` structure's ``->expmaskinit`` field tracks the
158
+
identities of the CPUs that were online at least once at the
159
+
beginning of the most recent RCU expedited grace period. The
160
+
``rcu_state`` structure's ``->ncpus`` and ``->ncpus_snap`` fields are
161
+
used to detect when new CPUs have come online for the first time,
162
+
that is, when the ``rcu_node`` structure's ``->expmaskinitnext``
163
+
field has changed since the beginning of the last RCU expedited grace
164
+
period, which triggers an update of each ``rcu_node`` structure's
165
+
``->expmaskinit`` field from its ``->expmaskinitnext`` field.
166
+
#. Each ``rcu_node`` structure's ``->expmaskinit`` field is used to
167
+
initialize that structure's ``->expmask`` at the beginning of each
168
+
RCU expedited grace period. This means that only those CPUs that have
169
+
been online at least once will be considered for a given grace
170
+
period.
171
+
#. Any CPU that goes offline will clear its bit in its leaf ``rcu_node``
172
+
structure's ``->qsmaskinitnext`` field, so any CPU with that bit
173
+
clear can safely be ignored. However, it is possible for a CPU coming
174
+
online or going offline to have this bit set for some time while
175
+
``cpu_online`` returns ``false``.
176
+
#. For each non-idle CPU that RCU believes is currently online, the
177
+
grace period invokes ``smp_call_function_single()``. If this
178
+
succeeds, the CPU was fully online. Failure indicates that the CPU is
179
+
in the process of coming online or going offline, in which case it is
180
+
necessary to wait for a short time period and try again. The purpose
181
+
of this wait (or series of waits, as the case may be) is to permit a
182
+
concurrent CPU-hotplug operation to complete.
183
+
#. In the case of RCU-sched, one of the last acts of an outgoing CPU is
184
+
to invoke ``rcu_report_dead()``, which reports a quiescent state for
185
+
that CPU. However, this is likely paranoia-induced redundancy.
186
+
187
+
+-----------------------------------------------------------------------+
188
+
| **Quick Quiz**: |
189
+
+-----------------------------------------------------------------------+
190
+
| Why all the dancing around with multiple counters and masks tracking |
191
+
| CPUs that were once online? Why not just have a single set of masks |
192
+
| tracking the currently online CPUs and be done with it? |
193
+
+-----------------------------------------------------------------------+
194
+
| **Answer**: |
195
+
+-----------------------------------------------------------------------+
196
+
| Maintaining single set of masks tracking the online CPUs *sounds* |
197
+
| easier, at least until you try working out all the race conditions |
198
+
| between grace-period initialization and CPU-hotplug operations. For |
199
+
| example, suppose initialization is progressing down the tree while a |
200
+
| CPU-offline operation is progressing up the tree. This situation can |
201
+
| result in bits set at the top of the tree that have no counterparts |
202
+
| at the bottom of the tree. Those bits will never be cleared, which |
203
+
| will result in grace-period hangs. In short, that way lies madness, |
204
+
| to say nothing of a great many bugs, hangs, and deadlocks. |
205
+
| In contrast, the current multi-mask multi-counter scheme ensures that |
206
+
| grace-period initialization will always see consistent masks up and |
207
+
| down the tree, which brings significant simplifications over the |
208
+
| single-mask method. |
209
+
| |
210
+
| This is an instance of `deferring work in order to avoid |
211
+
| synchronization <http://www.cs.columbia.edu/~library/TR-repository/re |
212
+
| ports/reports-1992/cucs-039-92.ps.gz>`__. |
213
+
| Lazily recording CPU-hotplug events at the beginning of the next |
214
+
| grace period greatly simplifies maintenance of the CPU-tracking |
215
+
| bitmasks in the ``rcu_node`` tree. |
216
+
+-----------------------------------------------------------------------+
217
+
218
+
Expedited Grace Period Refinements
219
+
----------------------------------
220
+
221
+
Idle-CPU Checks
222
+
~~~~~~~~~~~~~~~
223
+
224
+
Each expedited grace period checks for idle CPUs when initially forming
225
+
the mask of CPUs to be IPIed and again just before IPIing a CPU (both
226
+
checks are carried out by ``sync_rcu_exp_select_cpus()``). If the CPU is
227
+
idle at any time between those two times, the CPU will not be IPIed.
228
+
Instead, the task pushing the grace period forward will include the idle
229
+
CPUs in the mask passed to ``rcu_report_exp_cpu_mult()``.
230
+
231
+
For RCU-sched, there is an additional check: If the IPI has interrupted
232
+
the idle loop, then ``rcu_exp_handler()`` invokes
233
+
``rcu_report_exp_rdp()`` to report the corresponding quiescent state.
234
+
235
+
For RCU-preempt, there is no specific check for idle in the IPI handler
236
+
(``rcu_exp_handler()``), but because RCU read-side critical sections are
237
+
not permitted within the idle loop, if ``rcu_exp_handler()`` sees that
238
+
the CPU is within RCU read-side critical section, the CPU cannot
239
+
possibly be idle. Otherwise, ``rcu_exp_handler()`` invokes
240
+
``rcu_report_exp_rdp()`` to report the corresponding quiescent state,
241
+
regardless of whether or not that quiescent state was due to the CPU
242
+
being idle.
243
+
244
+
In summary, RCU expedited grace periods check for idle when building the
245
+
bitmask of CPUs that must be IPIed, just before sending each IPI, and
246
+
(either explicitly or implicitly) within the IPI handler.
247
+
248
+
Batching via Sequence Counter
249
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
250
+
251
+
If each grace-period request was carried out separately, expedited grace
252
+
periods would have abysmal scalability and problematic high-load
253
+
characteristics. Because each grace-period operation can serve an
254
+
unlimited number of updates, it is important to *batch* requests, so
255
+
that a single expedited grace-period operation will cover all requests
256
+
in the corresponding batch.
257
+
258
+
This batching is controlled by a sequence counter named
259
+
``->expedited_sequence`` in the ``rcu_state`` structure. This counter
260
+
has an odd value when there is an expedited grace period in progress and
261
+
an even value otherwise, so that dividing the counter value by two gives
262
+
the number of completed grace periods. During any given update request,
263
+
the counter must transition from even to odd and then back to even, thus
264
+
indicating that a grace period has elapsed. Therefore, if the initial
265
+
value of the counter is ``s``, the updater must wait until the counter
266
+
reaches at least the value ``(s+3)&~0x1``. This counter is managed by
267
+
the following access functions:
268
+
269
+
#. ``rcu_exp_gp_seq_start()``, which marks the start of an expedited
270
+
grace period.
271
+
#. ``rcu_exp_gp_seq_end()``, which marks the end of an expedited grace
272
+
period.
273
+
#. ``rcu_exp_gp_seq_snap()``, which obtains a snapshot of the counter.
274
+
#. ``rcu_exp_gp_seq_done()``, which returns ``true`` if a full expedited
275
+
grace period has elapsed since the corresponding call to
276
+
``rcu_exp_gp_seq_snap()``.
277
+
278
+
Again, only one request in a given batch need actually carry out a
279
+
grace-period operation, which means there must be an efficient way to
280
+
identify which of many concurrent reqeusts will initiate the grace
281
+
period, and that there be an efficient way for the remaining requests to
282
+
wait for that grace period to complete. However, that is the topic of
283
+
the next section.
284
+
285
+
Funnel Locking and Wait/Wakeup
286
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
287
+
288
+
The natural way to sort out which of a batch of updaters will initiate
289
+
the expedited grace period is to use the ``rcu_node`` combining tree, as
290
+
implemented by the ``exp_funnel_lock()`` function. The first updater
291
+
corresponding to a given grace period arriving at a given ``rcu_node``
292
+
structure records its desired grace-period sequence number in the
293
+
``->exp_seq_rq`` field and moves up to the next level in the tree.
294
+
Otherwise, if the ``->exp_seq_rq`` field already contains the sequence
295
+
number for the desired grace period or some later one, the updater
296
+
blocks on one of four wait queues in the ``->exp_wq[]`` array, using the
297
+
second-from-bottom and third-from bottom bits as an index. An
298
+
``->exp_lock`` field in the ``rcu_node`` structure synchronizes access
299
+
to these fields.
300
+
301
+
An empty ``rcu_node`` tree is shown in the following diagram, with the
302
+
white cells representing the ``->exp_seq_rq`` field and the red cells
303
+
representing the elements of the ``->exp_wq[]`` array.
304
+
305
+
.. kernel-figure:: Funnel0.svg
306
+
307
+
The next diagram shows the situation after the arrival of Task A and
308
+
Task B at the leftmost and rightmost leaf ``rcu_node`` structures,
309
+
respectively. The current value of the ``rcu_state`` structure's
310
+
``->expedited_sequence`` field is zero, so adding three and clearing the
311
+
bottom bit results in the value two, which both tasks record in the
312
+
``->exp_seq_rq`` field of their respective ``rcu_node`` structures:
313
+
314
+
.. kernel-figure:: Funnel1.svg
315
+
316
+
Each of Tasks A and B will move up to the root ``rcu_node`` structure.
317
+
Suppose that Task A wins, recording its desired grace-period sequence
318
+
number and resulting in the state shown below:
319
+
320
+
.. kernel-figure:: Funnel2.svg
321
+
322
+
Task A now advances to initiate a new grace period, while Task B moves
323
+
up to the root ``rcu_node`` structure, and, seeing that its desired
324
+
sequence number is already recorded, blocks on ``->exp_wq[1]``.
325
+
326
+
+-----------------------------------------------------------------------+
327
+
| **Quick Quiz**: |
328
+
+-----------------------------------------------------------------------+
329
+
| Why ``->exp_wq[1]``? Given that the value of these tasks' desired |
330
+
| sequence number is two, so shouldn't they instead block on |
331
+
| ``->exp_wq[2]``? |
332
+
+-----------------------------------------------------------------------+
333
+
| **Answer**: |
334
+
+-----------------------------------------------------------------------+
335
+
| No. |
336
+
| Recall that the bottom bit of the desired sequence number indicates |
337
+
| whether or not a grace period is currently in progress. It is |
338
+
| therefore necessary to shift the sequence number right one bit |
339
+
| position to obtain the number of the grace period. This results in |
340
+
| ``->exp_wq[1]``. |
341
+
+-----------------------------------------------------------------------+
342
+
343
+
If Tasks C and D also arrive at this point, they will compute the same
344
+
desired grace-period sequence number, and see that both leaf
345
+
``rcu_node`` structures already have that value recorded. They will
346
+
therefore block on their respective ``rcu_node`` structures'
347
+
``->exp_wq[1]`` fields, as shown below:
348
+
349
+
.. kernel-figure:: Funnel3.svg
350
+
351
+
Task A now acquires the ``rcu_state`` structure's ``->exp_mutex`` and
352
+
initiates the grace period, which increments ``->expedited_sequence``.
353
+
Therefore, if Tasks E and F arrive, they will compute a desired sequence
354
+
number of 4 and will record this value as shown below:
355
+
356
+
.. kernel-figure:: Funnel4.svg
357
+
358
+
Tasks E and F will propagate up the ``rcu_node`` combining tree, with
359
+
Task F blocking on the root ``rcu_node`` structure and Task E wait for
360
+
Task A to finish so that it can start the next grace period. The
361
+
resulting state is as shown below:
362
+
363
+
.. kernel-figure:: Funnel5.svg
364
+
365
+
Once the grace period completes, Task A starts waking up the tasks
366
+
waiting for this grace period to complete, increments the
367
+
``->expedited_sequence``, acquires the ``->exp_wake_mutex`` and then
368
+
releases the ``->exp_mutex``. This results in the following state:
369
+
370
+
.. kernel-figure:: Funnel6.svg
371
+
372
+
Task E can then acquire ``->exp_mutex`` and increment
373
+
``->expedited_sequence`` to the value three. If new tasks G and H arrive
374
+
and moves up the combining tree at the same time, the state will be as
375
+
follows:
376
+
377
+
.. kernel-figure:: Funnel7.svg
378
+
379
+
Note that three of the root ``rcu_node`` structure's waitqueues are now
380
+
occupied. However, at some point, Task A will wake up the tasks blocked
381
+
on the ``->exp_wq`` waitqueues, resulting in the following state:
382
+
383
+
.. kernel-figure:: Funnel8.svg
384
+
385
+
Execution will continue with Tasks E and H completing their grace
386
+
periods and carrying out their wakeups.
387
+
388
+
+-----------------------------------------------------------------------+
389
+
| **Quick Quiz**: |
390
+
+-----------------------------------------------------------------------+
391
+
| What happens if Task A takes so long to do its wakeups that Task E's |
392
+
| grace period completes? |
393
+
+-----------------------------------------------------------------------+
394
+
| **Answer**: |
395
+
+-----------------------------------------------------------------------+
396
+
| Then Task E will block on the ``->exp_wake_mutex``, which will also |
397
+
| prevent it from releasing ``->exp_mutex``, which in turn will prevent |
398
+
| the next grace period from starting. This last is important in |
399
+
| preventing overflow of the ``->exp_wq[]`` array. |
400
+
+-----------------------------------------------------------------------+
401
+
402
+
Use of Workqueues
403
+
~~~~~~~~~~~~~~~~~
404
+
405
+
In earlier implementations, the task requesting the expedited grace
406
+
period also drove it to completion. This straightforward approach had
407
+
the disadvantage of needing to account for POSIX signals sent to user
408
+
tasks, so more recent implemementations use the Linux kernel's
409
+
`workqueues <https://www.kernel.org/doc/Documentation/core-api/workqueue.rst>`__.
410
+
411
+
The requesting task still does counter snapshotting and funnel-lock
412
+
processing, but the task reaching the top of the funnel lock does a
413
+
``schedule_work()`` (from ``_synchronize_rcu_expedited()`` so that a
414
+
workqueue kthread does the actual grace-period processing. Because
415
+
workqueue kthreads do not accept POSIX signals, grace-period-wait
416
+
processing need not allow for POSIX signals. In addition, this approach
417
+
allows wakeups for the previous expedited grace period to be overlapped
418
+
with processing for the next expedited grace period. Because there are
419
+
only four sets of waitqueues, it is necessary to ensure that the
420
+
previous grace period's wakeups complete before the next grace period's
421
+
wakeups start. This is handled by having the ``->exp_mutex`` guard
422
+
expedited grace-period processing and the ``->exp_wake_mutex`` guard
423
+
wakeups. The key point is that the ``->exp_mutex`` is not released until
424
+
the first wakeup is complete, which means that the ``->exp_wake_mutex``
425
+
has already been acquired at that point. This approach ensures that the
426
+
previous grace period's wakeups can be carried out while the current
427
+
grace period is in process, but that these wakeups will complete before
428
+
the next grace period starts. This means that only three waitqueues are
429
+
required, guaranteeing that the four that are provided are sufficient.
430
+
431
+
Stall Warnings
432
+
~~~~~~~~~~~~~~
433
+
434
+
Expediting grace periods does nothing to speed things up when RCU
435
+
readers take too long, and therefore expedited grace periods check for
436
+
stalls just as normal grace periods do.
437
+
438
+
+-----------------------------------------------------------------------+
439
+
| **Quick Quiz**: |
440
+
+-----------------------------------------------------------------------+
441
+
| But why not just let the normal grace-period machinery detect the |
442
+
| stalls, given that a given reader must block both normal and |
443
+
| expedited grace periods? |
444
+
+-----------------------------------------------------------------------+
445
+
| **Answer**: |
446
+
+-----------------------------------------------------------------------+
447
+
| Because it is quite possible that at a given time there is no normal |
448
+
| grace period in progress, in which case the normal grace period |
449
+
| cannot emit a stall warning. |
450
+
+-----------------------------------------------------------------------+
451
+
452
+
The ``synchronize_sched_expedited_wait()`` function loops waiting for
453
+
the expedited grace period to end, but with a timeout set to the current
454
+
RCU CPU stall-warning time. If this time is exceeded, any CPUs or
455
+
``rcu_node`` structures blocking the current grace period are printed.
456
+
Each stall warning results in another pass through the loop, but the
457
+
second and subsequent passes use longer stall times.
458
+
459
+
Mid-boot operation
460
+
~~~~~~~~~~~~~~~~~~
461
+
462
+
The use of workqueues has the advantage that the expedited grace-period
463
+
code need not worry about POSIX signals. Unfortunately, it has the
464
+
corresponding disadvantage that workqueues cannot be used until they are
465
+
initialized, which does not happen until some time after the scheduler
466
+
spawns the first task. Given that there are parts of the kernel that
467
+
really do want to execute grace periods during this mid-boot “dead
468
+
zone”, expedited grace periods must do something else during thie time.
469
+
470
+
What they do is to fall back to the old practice of requiring that the
471
+
requesting task drive the expedited grace period, as was the case before
472
+
the use of workqueues. However, the requesting task is only required to
473
+
drive the grace period during the mid-boot dead zone. Before mid-boot, a
474
+
synchronous grace period is a no-op. Some time after mid-boot,
475
+
workqueues are used.
476
+
477
+
Non-expedited non-SRCU synchronous grace periods must also operate
478
+
normally during mid-boot. This is handled by causing non-expedited grace
479
+
periods to take the expedited code path during mid-boot.
480
+
481
+
The current code assumes that there are no POSIX signals during the
482
+
mid-boot dead zone. However, if an overwhelming need for POSIX signals
483
+
somehow arises, appropriate adjustments can be made to the expedited
484
+
stall-warning code. One such adjustment would reinstate the
485
+
pre-workqueue stall-warning checks, but only during the mid-boot dead
486
+
zone.
487
+
488
+
With this refinement, synchronous grace periods can now be used from
489
+
task context pretty much any time during the life of the kernel. That
490
+
is, aside from some points in the suspend, hibernate, or shutdown code
491
+
path.
492
+
493
+
Summary
494
+
~~~~~~~
495
+
496
+
Expedited grace periods use a sequence-number approach to promote
497
+
batching, so that a single grace-period operation can serve numerous
498
+
requests. A funnel lock is used to efficiently identify the one task out
499
+
of a concurrent group that will request the grace period. All members of
500
+
the group will block on waitqueues provided in the ``rcu_node``
501
+
structure. The actual grace-period processing is carried out by a
502
+
workqueue.
503
+
504
+
CPU-hotplug operations are noted lazily in order to prevent the need for
505
+
tight synchronization between expedited grace periods and CPU-hotplug
506
+
operations. The dyntick-idle counters are used to avoid sending IPIs to
507
+
idle CPUs, at least in the common case. RCU-preempt and RCU-sched use
508
+
different IPI handlers and different code to respond to the state
509
+
changes carried out by those handlers, but otherwise use common code.
510
+
511
+
Quiescent states are tracked using the ``rcu_node`` tree, and once all
512
+
necessary quiescent states have been reported, all tasks waiting on this
513
+
expedited grace period are awakened. A pair of mutexes are used to allow
514
+
one grace period's wakeups to proceed concurrently with the next grace
515
+
period's processing.
516
+
517
+
This combination of mechanisms allows expedited grace periods to run
518
+
reasonably efficiently. However, for non-time-critical tasks, normal
519
+
grace periods should be used instead because their longer duration
520
+
permits much higher degrees of batching, and thus much lower per-request
521
+
overheads.
-9
Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Diagram.html
-9
Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Diagram.html
···
1
-
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN"
2
-
"http://www.w3.org/TR/html4/loose.dtd">
3
-
<html>
4
-
<head><title>A Diagram of TREE_RCU's Grace-Period Memory Ordering</title>
5
-
<meta HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=iso-8859-1">
6
-
7
-
<p><img src="TreeRCU-gp.svg" alt="TreeRCU-gp.svg">
8
-
9
-
</body></html>
-704
Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.html
-704
Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.html
···
1
-
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN"
2
-
"http://www.w3.org/TR/html4/loose.dtd">
3
-
<html>
4
-
<head><title>A Tour Through TREE_RCU's Grace-Period Memory Ordering</title>
5
-
<meta HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=iso-8859-1">
6
-
7
-
<p>August 8, 2017</p>
8
-
<p>This article was contributed by Paul E. McKenney</p>
9
-
10
-
<h3>Introduction</h3>
11
-
12
-
<p>This document gives a rough visual overview of how Tree RCU's
13
-
grace-period memory ordering guarantee is provided.
14
-
15
-
<ol>
16
-
<li> <a href="#What Is Tree RCU's Grace Period Memory Ordering Guarantee?">
17
-
What Is Tree RCU's Grace Period Memory Ordering Guarantee?</a>
18
-
<li> <a href="#Tree RCU Grace Period Memory Ordering Building Blocks">
19
-
Tree RCU Grace Period Memory Ordering Building Blocks</a>
20
-
<li> <a href="#Tree RCU Grace Period Memory Ordering Components">
21
-
Tree RCU Grace Period Memory Ordering Components</a>
22
-
<li> <a href="#Putting It All Together">Putting It All Together</a>
23
-
</ol>
24
-
25
-
<h3><a name="What Is Tree RCU's Grace Period Memory Ordering Guarantee?">
26
-
What Is Tree RCU's Grace Period Memory Ordering Guarantee?</a></h3>
27
-
28
-
<p>RCU grace periods provide extremely strong memory-ordering guarantees
29
-
for non-idle non-offline code.
30
-
Any code that happens after the end of a given RCU grace period is guaranteed
31
-
to see the effects of all accesses prior to the beginning of that grace
32
-
period that are within RCU read-side critical sections.
33
-
Similarly, any code that happens before the beginning of a given RCU grace
34
-
period is guaranteed to see the effects of all accesses following the end
35
-
of that grace period that are within RCU read-side critical sections.
36
-
37
-
<p>Note well that RCU-sched read-side critical sections include any region
38
-
of code for which preemption is disabled.
39
-
Given that each individual machine instruction can be thought of as
40
-
an extremely small region of preemption-disabled code, one can think of
41
-
<tt>synchronize_rcu()</tt> as <tt>smp_mb()</tt> on steroids.
42
-
43
-
<p>RCU updaters use this guarantee by splitting their updates into
44
-
two phases, one of which is executed before the grace period and
45
-
the other of which is executed after the grace period.
46
-
In the most common use case, phase one removes an element from
47
-
a linked RCU-protected data structure, and phase two frees that element.
48
-
For this to work, any readers that have witnessed state prior to the
49
-
phase-one update (in the common case, removal) must not witness state
50
-
following the phase-two update (in the common case, freeing).
51
-
52
-
<p>The RCU implementation provides this guarantee using a network
53
-
of lock-based critical sections, memory barriers, and per-CPU
54
-
processing, as is described in the following sections.
55
-
56
-
<h3><a name="Tree RCU Grace Period Memory Ordering Building Blocks">
57
-
Tree RCU Grace Period Memory Ordering Building Blocks</a></h3>
58
-
59
-
<p>The workhorse for RCU's grace-period memory ordering is the
60
-
critical section for the <tt>rcu_node</tt> structure's
61
-
<tt>->lock</tt>.
62
-
These critical sections use helper functions for lock acquisition, including
63
-
<tt>raw_spin_lock_rcu_node()</tt>,
64
-
<tt>raw_spin_lock_irq_rcu_node()</tt>, and
65
-
<tt>raw_spin_lock_irqsave_rcu_node()</tt>.
66
-
Their lock-release counterparts are
67
-
<tt>raw_spin_unlock_rcu_node()</tt>,
68
-
<tt>raw_spin_unlock_irq_rcu_node()</tt>, and
69
-
<tt>raw_spin_unlock_irqrestore_rcu_node()</tt>,
70
-
respectively.
71
-
For completeness, a
72
-
<tt>raw_spin_trylock_rcu_node()</tt>
73
-
is also provided.
74
-
The key point is that the lock-acquisition functions, including
75
-
<tt>raw_spin_trylock_rcu_node()</tt>, all invoke
76
-
<tt>smp_mb__after_unlock_lock()</tt> immediately after successful
77
-
acquisition of the lock.
78
-
79
-
<p>Therefore, for any given <tt>rcu_node</tt> structure, any access
80
-
happening before one of the above lock-release functions will be seen
81
-
by all CPUs as happening before any access happening after a later
82
-
one of the above lock-acquisition functions.
83
-
Furthermore, any access happening before one of the
84
-
above lock-release function on any given CPU will be seen by all
85
-
CPUs as happening before any access happening after a later one
86
-
of the above lock-acquisition functions executing on that same CPU,
87
-
even if the lock-release and lock-acquisition functions are operating
88
-
on different <tt>rcu_node</tt> structures.
89
-
Tree RCU uses these two ordering guarantees to form an ordering
90
-
network among all CPUs that were in any way involved in the grace
91
-
period, including any CPUs that came online or went offline during
92
-
the grace period in question.
93
-
94
-
<p>The following litmus test exhibits the ordering effects of these
95
-
lock-acquisition and lock-release functions:
96
-
97
-
<pre>
98
-
1 int x, y, z;
99
-
2
100
-
3 void task0(void)
101
-
4 {
102
-
5 raw_spin_lock_rcu_node(rnp);
103
-
6 WRITE_ONCE(x, 1);
104
-
7 r1 = READ_ONCE(y);
105
-
8 raw_spin_unlock_rcu_node(rnp);
106
-
9 }
107
-
10
108
-
11 void task1(void)
109
-
12 {
110
-
13 raw_spin_lock_rcu_node(rnp);
111
-
14 WRITE_ONCE(y, 1);
112
-
15 r2 = READ_ONCE(z);
113
-
16 raw_spin_unlock_rcu_node(rnp);
114
-
17 }
115
-
18
116
-
19 void task2(void)
117
-
20 {
118
-
21 WRITE_ONCE(z, 1);
119
-
22 smp_mb();
120
-
23 r3 = READ_ONCE(x);
121
-
24 }
122
-
25
123
-
26 WARN_ON(r1 == 0 && r2 == 0 && r3 == 0);
124
-
</pre>
125
-
126
-
<p>The <tt>WARN_ON()</tt> is evaluated at “the end of time”,
127
-
after all changes have propagated throughout the system.
128
-
Without the <tt>smp_mb__after_unlock_lock()</tt> provided by the
129
-
acquisition functions, this <tt>WARN_ON()</tt> could trigger, for example
130
-
on PowerPC.
131
-
The <tt>smp_mb__after_unlock_lock()</tt> invocations prevent this
132
-
<tt>WARN_ON()</tt> from triggering.
133
-
134
-
<p>This approach must be extended to include idle CPUs, which need
135
-
RCU's grace-period memory ordering guarantee to extend to any
136
-
RCU read-side critical sections preceding and following the current
137
-
idle sojourn.
138
-
This case is handled by calls to the strongly ordered
139
-
<tt>atomic_add_return()</tt> read-modify-write atomic operation that
140
-
is invoked within <tt>rcu_dynticks_eqs_enter()</tt> at idle-entry
141
-
time and within <tt>rcu_dynticks_eqs_exit()</tt> at idle-exit time.
142
-
The grace-period kthread invokes <tt>rcu_dynticks_snap()</tt> and
143
-
<tt>rcu_dynticks_in_eqs_since()</tt> (both of which invoke
144
-
an <tt>atomic_add_return()</tt> of zero) to detect idle CPUs.
145
-
146
-
<table>
147
-
<tr><th> </th></tr>
148
-
<tr><th align="left">Quick Quiz:</th></tr>
149
-
<tr><td>
150
-
But what about CPUs that remain offline for the entire
151
-
grace period?
152
-
</td></tr>
153
-
<tr><th align="left">Answer:</th></tr>
154
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
155
-
Such CPUs will be offline at the beginning of the grace period,
156
-
so the grace period won't expect quiescent states from them.
157
-
Races between grace-period start and CPU-hotplug operations
158
-
are mediated by the CPU's leaf <tt>rcu_node</tt> structure's
159
-
<tt>->lock</tt> as described above.
160
-
</font></td></tr>
161
-
<tr><td> </td></tr>
162
-
</table>
163
-
164
-
<p>The approach must be extended to handle one final case, that
165
-
of waking a task blocked in <tt>synchronize_rcu()</tt>.
166
-
This task might be affinitied to a CPU that is not yet aware that
167
-
the grace period has ended, and thus might not yet be subject to
168
-
the grace period's memory ordering.
169
-
Therefore, there is an <tt>smp_mb()</tt> after the return from
170
-
<tt>wait_for_completion()</tt> in the <tt>synchronize_rcu()</tt>
171
-
code path.
172
-
173
-
<table>
174
-
<tr><th> </th></tr>
175
-
<tr><th align="left">Quick Quiz:</th></tr>
176
-
<tr><td>
177
-
What? Where???
178
-
I don't see any <tt>smp_mb()</tt> after the return from
179
-
<tt>wait_for_completion()</tt>!!!
180
-
</td></tr>
181
-
<tr><th align="left">Answer:</th></tr>
182
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
183
-
That would be because I spotted the need for that
184
-
<tt>smp_mb()</tt> during the creation of this documentation,
185
-
and it is therefore unlikely to hit mainline before v4.14.
186
-
Kudos to Lance Roy, Will Deacon, Peter Zijlstra, and
187
-
Jonathan Cameron for asking questions that sensitized me
188
-
to the rather elaborate sequence of events that demonstrate
189
-
the need for this memory barrier.
190
-
</font></td></tr>
191
-
<tr><td> </td></tr>
192
-
</table>
193
-
194
-
<p>Tree RCU's grace--period memory-ordering guarantees rely most
195
-
heavily on the <tt>rcu_node</tt> structure's <tt>->lock</tt>
196
-
field, so much so that it is necessary to abbreviate this pattern
197
-
in the diagrams in the next section.
198
-
For example, consider the <tt>rcu_prepare_for_idle()</tt> function
199
-
shown below, which is one of several functions that enforce ordering
200
-
of newly arrived RCU callbacks against future grace periods:
201
-
202
-
<pre>
203
-
1 static void rcu_prepare_for_idle(void)
204
-
2 {
205
-
3 bool needwake;
206
-
4 struct rcu_data *rdp;
207
-
5 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
208
-
6 struct rcu_node *rnp;
209
-
7 struct rcu_state *rsp;
210
-
8 int tne;
211
-
9
212
-
10 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) ||
213
-
11 rcu_is_nocb_cpu(smp_processor_id()))
214
-
12 return;
215
-
13 tne = READ_ONCE(tick_nohz_active);
216
-
14 if (tne != rdtp->tick_nohz_enabled_snap) {
217
-
15 if (rcu_cpu_has_callbacks(NULL))
218
-
16 invoke_rcu_core();
219
-
17 rdtp->tick_nohz_enabled_snap = tne;
220
-
18 return;
221
-
19 }
222
-
20 if (!tne)
223
-
21 return;
224
-
22 if (rdtp->all_lazy &&
225
-
23 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
226
-
24 rdtp->all_lazy = false;
227
-
25 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
228
-
26 invoke_rcu_core();
229
-
27 return;
230
-
28 }
231
-
29 if (rdtp->last_accelerate == jiffies)
232
-
30 return;
233
-
31 rdtp->last_accelerate = jiffies;
234
-
32 for_each_rcu_flavor(rsp) {
235
-
33 rdp = this_cpu_ptr(rsp->rda);
236
-
34 if (rcu_segcblist_pend_cbs(&rdp->cblist))
237
-
35 continue;
238
-
36 rnp = rdp->mynode;
239
-
37 raw_spin_lock_rcu_node(rnp);
240
-
38 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
241
-
39 raw_spin_unlock_rcu_node(rnp);
242
-
40 if (needwake)
243
-
41 rcu_gp_kthread_wake(rsp);
244
-
42 }
245
-
43 }
246
-
</pre>
247
-
248
-
<p>But the only part of <tt>rcu_prepare_for_idle()</tt> that really
249
-
matters for this discussion are lines 37–39.
250
-
We will therefore abbreviate this function as follows:
251
-
252
-
</p><p><img src="rcu_node-lock.svg" alt="rcu_node-lock.svg">
253
-
254
-
<p>The box represents the <tt>rcu_node</tt> structure's <tt>->lock</tt>
255
-
critical section, with the double line on top representing the additional
256
-
<tt>smp_mb__after_unlock_lock()</tt>.
257
-
258
-
<h3><a name="Tree RCU Grace Period Memory Ordering Components">
259
-
Tree RCU Grace Period Memory Ordering Components</a></h3>
260
-
261
-
<p>Tree RCU's grace-period memory-ordering guarantee is provided by
262
-
a number of RCU components:
263
-
264
-
<ol>
265
-
<li> <a href="#Callback Registry">Callback Registry</a>
266
-
<li> <a href="#Grace-Period Initialization">Grace-Period Initialization</a>
267
-
<li> <a href="#Self-Reported Quiescent States">
268
-
Self-Reported Quiescent States</a>
269
-
<li> <a href="#Dynamic Tick Interface">Dynamic Tick Interface</a>
270
-
<li> <a href="#CPU-Hotplug Interface">CPU-Hotplug Interface</a>
271
-
<li> <a href="Forcing Quiescent States">Forcing Quiescent States</a>
272
-
<li> <a href="Grace-Period Cleanup">Grace-Period Cleanup</a>
273
-
<li> <a href="Callback Invocation">Callback Invocation</a>
274
-
</ol>
275
-
276
-
<p>Each of the following section looks at the corresponding component
277
-
in detail.
278
-
279
-
<h4><a name="Callback Registry">Callback Registry</a></h4>
280
-
281
-
<p>If RCU's grace-period guarantee is to mean anything at all, any
282
-
access that happens before a given invocation of <tt>call_rcu()</tt>
283
-
must also happen before the corresponding grace period.
284
-
The implementation of this portion of RCU's grace period guarantee
285
-
is shown in the following figure:
286
-
287
-
</p><p><img src="TreeRCU-callback-registry.svg" alt="TreeRCU-callback-registry.svg">
288
-
289
-
<p>Because <tt>call_rcu()</tt> normally acts only on CPU-local state,
290
-
it provides no ordering guarantees, either for itself or for
291
-
phase one of the update (which again will usually be removal of
292
-
an element from an RCU-protected data structure).
293
-
It simply enqueues the <tt>rcu_head</tt> structure on a per-CPU list,
294
-
which cannot become associated with a grace period until a later
295
-
call to <tt>rcu_accelerate_cbs()</tt>, as shown in the diagram above.
296
-
297
-
<p>One set of code paths shown on the left invokes
298
-
<tt>rcu_accelerate_cbs()</tt> via
299
-
<tt>note_gp_changes()</tt>, either directly from <tt>call_rcu()</tt> (if
300
-
the current CPU is inundated with queued <tt>rcu_head</tt> structures)
301
-
or more likely from an <tt>RCU_SOFTIRQ</tt> handler.
302
-
Another code path in the middle is taken only in kernels built with
303
-
<tt>CONFIG_RCU_FAST_NO_HZ=y</tt>, which invokes
304
-
<tt>rcu_accelerate_cbs()</tt> via <tt>rcu_prepare_for_idle()</tt>.
305
-
The final code path on the right is taken only in kernels built with
306
-
<tt>CONFIG_HOTPLUG_CPU=y</tt>, which invokes
307
-
<tt>rcu_accelerate_cbs()</tt> via
308
-
<tt>rcu_advance_cbs()</tt>, <tt>rcu_migrate_callbacks</tt>,
309
-
<tt>rcutree_migrate_callbacks()</tt>, and <tt>takedown_cpu()</tt>,
310
-
which in turn is invoked on a surviving CPU after the outgoing
311
-
CPU has been completely offlined.
312
-
313
-
<p>There are a few other code paths within grace-period processing
314
-
that opportunistically invoke <tt>rcu_accelerate_cbs()</tt>.
315
-
However, either way, all of the CPU's recently queued <tt>rcu_head</tt>
316
-
structures are associated with a future grace-period number under
317
-
the protection of the CPU's lead <tt>rcu_node</tt> structure's
318
-
<tt>->lock</tt>.
319
-
In all cases, there is full ordering against any prior critical section
320
-
for that same <tt>rcu_node</tt> structure's <tt>->lock</tt>, and
321
-
also full ordering against any of the current task's or CPU's prior critical
322
-
sections for any <tt>rcu_node</tt> structure's <tt>->lock</tt>.
323
-
324
-
<p>The next section will show how this ordering ensures that any
325
-
accesses prior to the <tt>call_rcu()</tt> (particularly including phase
326
-
one of the update)
327
-
happen before the start of the corresponding grace period.
328
-
329
-
<table>
330
-
<tr><th> </th></tr>
331
-
<tr><th align="left">Quick Quiz:</th></tr>
332
-
<tr><td>
333
-
But what about <tt>synchronize_rcu()</tt>?
334
-
</td></tr>
335
-
<tr><th align="left">Answer:</th></tr>
336
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
337
-
The <tt>synchronize_rcu()</tt> passes <tt>call_rcu()</tt>
338
-
to <tt>wait_rcu_gp()</tt>, which invokes it.
339
-
So either way, it eventually comes down to <tt>call_rcu()</tt>.
340
-
</font></td></tr>
341
-
<tr><td> </td></tr>
342
-
</table>
343
-
344
-
<h4><a name="Grace-Period Initialization">Grace-Period Initialization</a></h4>
345
-
346
-
<p>Grace-period initialization is carried out by
347
-
the grace-period kernel thread, which makes several passes over the
348
-
<tt>rcu_node</tt> tree within the <tt>rcu_gp_init()</tt> function.
349
-
This means that showing the full flow of ordering through the
350
-
grace-period computation will require duplicating this tree.
351
-
If you find this confusing, please note that the state of the
352
-
<tt>rcu_node</tt> changes over time, just like Heraclitus's river.
353
-
However, to keep the <tt>rcu_node</tt> river tractable, the
354
-
grace-period kernel thread's traversals are presented in multiple
355
-
parts, starting in this section with the various phases of
356
-
grace-period initialization.
357
-
358
-
<p>The first ordering-related grace-period initialization action is to
359
-
advance the <tt>rcu_state</tt> structure's <tt>->gp_seq</tt>
360
-
grace-period-number counter, as shown below:
361
-
362
-
</p><p><img src="TreeRCU-gp-init-1.svg" alt="TreeRCU-gp-init-1.svg" width="75%">
363
-
364
-
<p>The actual increment is carried out using <tt>smp_store_release()</tt>,
365
-
which helps reject false-positive RCU CPU stall detection.
366
-
Note that only the root <tt>rcu_node</tt> structure is touched.
367
-
368
-
<p>The first pass through the <tt>rcu_node</tt> tree updates bitmasks
369
-
based on CPUs having come online or gone offline since the start of
370
-
the previous grace period.
371
-
In the common case where the number of online CPUs for this <tt>rcu_node</tt>
372
-
structure has not transitioned to or from zero,
373
-
this pass will scan only the leaf <tt>rcu_node</tt> structures.
374
-
However, if the number of online CPUs for a given leaf <tt>rcu_node</tt>
375
-
structure has transitioned from zero,
376
-
<tt>rcu_init_new_rnp()</tt> will be invoked for the first incoming CPU.
377
-
Similarly, if the number of online CPUs for a given leaf <tt>rcu_node</tt>
378
-
structure has transitioned to zero,
379
-
<tt>rcu_cleanup_dead_rnp()</tt> will be invoked for the last outgoing CPU.
380
-
The diagram below shows the path of ordering if the leftmost
381
-
<tt>rcu_node</tt> structure onlines its first CPU and if the next
382
-
<tt>rcu_node</tt> structure has no online CPUs
383
-
(or, alternatively if the leftmost <tt>rcu_node</tt> structure offlines
384
-
its last CPU and if the next <tt>rcu_node</tt> structure has no online CPUs).
385
-
386
-
</p><p><img src="TreeRCU-gp-init-2.svg" alt="TreeRCU-gp-init-1.svg" width="75%">
387
-
388
-
<p>The final <tt>rcu_gp_init()</tt> pass through the <tt>rcu_node</tt>
389
-
tree traverses breadth-first, setting each <tt>rcu_node</tt> structure's
390
-
<tt>->gp_seq</tt> field to the newly advanced value from the
391
-
<tt>rcu_state</tt> structure, as shown in the following diagram.
392
-
393
-
</p><p><img src="TreeRCU-gp-init-3.svg" alt="TreeRCU-gp-init-1.svg" width="75%">
394
-
395
-
<p>This change will also cause each CPU's next call to
396
-
<tt>__note_gp_changes()</tt>
397
-
to notice that a new grace period has started, as described in the next
398
-
section.
399
-
But because the grace-period kthread started the grace period at the
400
-
root (with the advancing of the <tt>rcu_state</tt> structure's
401
-
<tt>->gp_seq</tt> field) before setting each leaf <tt>rcu_node</tt>
402
-
structure's <tt>->gp_seq</tt> field, each CPU's observation of
403
-
the start of the grace period will happen after the actual start
404
-
of the grace period.
405
-
406
-
<table>
407
-
<tr><th> </th></tr>
408
-
<tr><th align="left">Quick Quiz:</th></tr>
409
-
<tr><td>
410
-
But what about the CPU that started the grace period?
411
-
Why wouldn't it see the start of the grace period right when
412
-
it started that grace period?
413
-
</td></tr>
414
-
<tr><th align="left">Answer:</th></tr>
415
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
416
-
In some deep philosophical and overly anthromorphized
417
-
sense, yes, the CPU starting the grace period is immediately
418
-
aware of having done so.
419
-
However, if we instead assume that RCU is not self-aware,
420
-
then even the CPU starting the grace period does not really
421
-
become aware of the start of this grace period until its
422
-
first call to <tt>__note_gp_changes()</tt>.
423
-
On the other hand, this CPU potentially gets early notification
424
-
because it invokes <tt>__note_gp_changes()</tt> during its
425
-
last <tt>rcu_gp_init()</tt> pass through its leaf
426
-
<tt>rcu_node</tt> structure.
427
-
</font></td></tr>
428
-
<tr><td> </td></tr>
429
-
</table>
430
-
431
-
<h4><a name="Self-Reported Quiescent States">
432
-
Self-Reported Quiescent States</a></h4>
433
-
434
-
<p>When all entities that might block the grace period have reported
435
-
quiescent states (or as described in a later section, had quiescent
436
-
states reported on their behalf), the grace period can end.
437
-
Online non-idle CPUs report their own quiescent states, as shown
438
-
in the following diagram:
439
-
440
-
</p><p><img src="TreeRCU-qs.svg" alt="TreeRCU-qs.svg" width="75%">
441
-
442
-
<p>This is for the last CPU to report a quiescent state, which signals
443
-
the end of the grace period.
444
-
Earlier quiescent states would push up the <tt>rcu_node</tt> tree
445
-
only until they encountered an <tt>rcu_node</tt> structure that
446
-
is waiting for additional quiescent states.
447
-
However, ordering is nevertheless preserved because some later quiescent
448
-
state will acquire that <tt>rcu_node</tt> structure's <tt>->lock</tt>.
449
-
450
-
<p>Any number of events can lead up to a CPU invoking
451
-
<tt>note_gp_changes</tt> (or alternatively, directly invoking
452
-
<tt>__note_gp_changes()</tt>), at which point that CPU will notice
453
-
the start of a new grace period while holding its leaf
454
-
<tt>rcu_node</tt> lock.
455
-
Therefore, all execution shown in this diagram happens after the
456
-
start of the grace period.
457
-
In addition, this CPU will consider any RCU read-side critical
458
-
section that started before the invocation of <tt>__note_gp_changes()</tt>
459
-
to have started before the grace period, and thus a critical
460
-
section that the grace period must wait on.
461
-
462
-
<table>
463
-
<tr><th> </th></tr>
464
-
<tr><th align="left">Quick Quiz:</th></tr>
465
-
<tr><td>
466
-
But a RCU read-side critical section might have started
467
-
after the beginning of the grace period
468
-
(the advancing of <tt>->gp_seq</tt> from earlier), so why should
469
-
the grace period wait on such a critical section?
470
-
</td></tr>
471
-
<tr><th align="left">Answer:</th></tr>
472
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
473
-
It is indeed not necessary for the grace period to wait on such
474
-
a critical section.
475
-
However, it is permissible to wait on it.
476
-
And it is furthermore important to wait on it, as this
477
-
lazy approach is far more scalable than a “big bang”
478
-
all-at-once grace-period start could possibly be.
479
-
</font></td></tr>
480
-
<tr><td> </td></tr>
481
-
</table>
482
-
483
-
<p>If the CPU does a context switch, a quiescent state will be
484
-
noted by <tt>rcu_node_context_switch()</tt> on the left.
485
-
On the other hand, if the CPU takes a scheduler-clock interrupt
486
-
while executing in usermode, a quiescent state will be noted by
487
-
<tt>rcu_sched_clock_irq()</tt> on the right.
488
-
Either way, the passage through a quiescent state will be noted
489
-
in a per-CPU variable.
490
-
491
-
<p>The next time an <tt>RCU_SOFTIRQ</tt> handler executes on
492
-
this CPU (for example, after the next scheduler-clock
493
-
interrupt), <tt>rcu_core()</tt> will invoke
494
-
<tt>rcu_check_quiescent_state()</tt>, which will notice the
495
-
recorded quiescent state, and invoke
496
-
<tt>rcu_report_qs_rdp()</tt>.
497
-
If <tt>rcu_report_qs_rdp()</tt> verifies that the quiescent state
498
-
really does apply to the current grace period, it invokes
499
-
<tt>rcu_report_rnp()</tt> which traverses up the <tt>rcu_node</tt>
500
-
tree as shown at the bottom of the diagram, clearing bits from
501
-
each <tt>rcu_node</tt> structure's <tt>->qsmask</tt> field,
502
-
and propagating up the tree when the result is zero.
503
-
504
-
<p>Note that traversal passes upwards out of a given <tt>rcu_node</tt>
505
-
structure only if the current CPU is reporting the last quiescent
506
-
state for the subtree headed by that <tt>rcu_node</tt> structure.
507
-
A key point is that if a CPU's traversal stops at a given <tt>rcu_node</tt>
508
-
structure, then there will be a later traversal by another CPU
509
-
(or perhaps the same one) that proceeds upwards
510
-
from that point, and the <tt>rcu_node</tt> <tt>->lock</tt>
511
-
guarantees that the first CPU's quiescent state happens before the
512
-
remainder of the second CPU's traversal.
513
-
Applying this line of thought repeatedly shows that all CPUs'
514
-
quiescent states happen before the last CPU traverses through
515
-
the root <tt>rcu_node</tt> structure, the “last CPU”
516
-
being the one that clears the last bit in the root <tt>rcu_node</tt>
517
-
structure's <tt>->qsmask</tt> field.
518
-
519
-
<h4><a name="Dynamic Tick Interface">Dynamic Tick Interface</a></h4>
520
-
521
-
<p>Due to energy-efficiency considerations, RCU is forbidden from
522
-
disturbing idle CPUs.
523
-
CPUs are therefore required to notify RCU when entering or leaving idle
524
-
state, which they do via fully ordered value-returning atomic operations
525
-
on a per-CPU variable.
526
-
The ordering effects are as shown below:
527
-
528
-
</p><p><img src="TreeRCU-dyntick.svg" alt="TreeRCU-dyntick.svg" width="50%">
529
-
530
-
<p>The RCU grace-period kernel thread samples the per-CPU idleness
531
-
variable while holding the corresponding CPU's leaf <tt>rcu_node</tt>
532
-
structure's <tt>->lock</tt>.
533
-
This means that any RCU read-side critical sections that precede the
534
-
idle period (the oval near the top of the diagram above) will happen
535
-
before the end of the current grace period.
536
-
Similarly, the beginning of the current grace period will happen before
537
-
any RCU read-side critical sections that follow the
538
-
idle period (the oval near the bottom of the diagram above).
539
-
540
-
<p>Plumbing this into the full grace-period execution is described
541
-
<a href="#Forcing Quiescent States">below</a>.
542
-
543
-
<h4><a name="CPU-Hotplug Interface">CPU-Hotplug Interface</a></h4>
544
-
545
-
<p>RCU is also forbidden from disturbing offline CPUs, which might well
546
-
be powered off and removed from the system completely.
547
-
CPUs are therefore required to notify RCU of their comings and goings
548
-
as part of the corresponding CPU hotplug operations.
549
-
The ordering effects are shown below:
550
-
551
-
</p><p><img src="TreeRCU-hotplug.svg" alt="TreeRCU-hotplug.svg" width="50%">
552
-
553
-
<p>Because CPU hotplug operations are much less frequent than idle transitions,
554
-
they are heavier weight, and thus acquire the CPU's leaf <tt>rcu_node</tt>
555
-
structure's <tt>->lock</tt> and update this structure's
556
-
<tt>->qsmaskinitnext</tt>.
557
-
The RCU grace-period kernel thread samples this mask to detect CPUs
558
-
having gone offline since the beginning of this grace period.
559
-
560
-
<p>Plumbing this into the full grace-period execution is described
561
-
<a href="#Forcing Quiescent States">below</a>.
562
-
563
-
<h4><a name="Forcing Quiescent States">Forcing Quiescent States</a></h4>
564
-
565
-
<p>As noted above, idle and offline CPUs cannot report their own
566
-
quiescent states, and therefore the grace-period kernel thread
567
-
must do the reporting on their behalf.
568
-
This process is called “forcing quiescent states”, it is
569
-
repeated every few jiffies, and its ordering effects are shown below:
570
-
571
-
</p><p><img src="TreeRCU-gp-fqs.svg" alt="TreeRCU-gp-fqs.svg" width="100%">
572
-
573
-
<p>Each pass of quiescent state forcing is guaranteed to traverse the
574
-
leaf <tt>rcu_node</tt> structures, and if there are no new quiescent
575
-
states due to recently idled and/or offlined CPUs, then only the
576
-
leaves are traversed.
577
-
However, if there is a newly offlined CPU as illustrated on the left
578
-
or a newly idled CPU as illustrated on the right, the corresponding
579
-
quiescent state will be driven up towards the root.
580
-
As with self-reported quiescent states, the upwards driving stops
581
-
once it reaches an <tt>rcu_node</tt> structure that has quiescent
582
-
states outstanding from other CPUs.
583
-
584
-
<table>
585
-
<tr><th> </th></tr>
586
-
<tr><th align="left">Quick Quiz:</th></tr>
587
-
<tr><td>
588
-
The leftmost drive to root stopped before it reached
589
-
the root <tt>rcu_node</tt> structure, which means that
590
-
there are still CPUs subordinate to that structure on
591
-
which the current grace period is waiting.
592
-
Given that, how is it possible that the rightmost drive
593
-
to root ended the grace period?
594
-
</td></tr>
595
-
<tr><th align="left">Answer:</th></tr>
596
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
597
-
Good analysis!
598
-
It is in fact impossible in the absence of bugs in RCU.
599
-
But this diagram is complex enough as it is, so simplicity
600
-
overrode accuracy.
601
-
You can think of it as poetic license, or you can think of
602
-
it as misdirection that is resolved in the
603
-
<a href="#Putting It All Together">stitched-together diagram</a>.
604
-
</font></td></tr>
605
-
<tr><td> </td></tr>
606
-
</table>
607
-
608
-
<h4><a name="Grace-Period Cleanup">Grace-Period Cleanup</a></h4>
609
-
610
-
<p>Grace-period cleanup first scans the <tt>rcu_node</tt> tree
611
-
breadth-first advancing all the <tt>->gp_seq</tt> fields, then it
612
-
advances the <tt>rcu_state</tt> structure's <tt>->gp_seq</tt> field.
613
-
The ordering effects are shown below:
614
-
615
-
</p><p><img src="TreeRCU-gp-cleanup.svg" alt="TreeRCU-gp-cleanup.svg" width="75%">
616
-
617
-
<p>As indicated by the oval at the bottom of the diagram, once
618
-
grace-period cleanup is complete, the next grace period can begin.
619
-
620
-
<table>
621
-
<tr><th> </th></tr>
622
-
<tr><th align="left">Quick Quiz:</th></tr>
623
-
<tr><td>
624
-
But when precisely does the grace period end?
625
-
</td></tr>
626
-
<tr><th align="left">Answer:</th></tr>
627
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
628
-
There is no useful single point at which the grace period
629
-
can be said to end.
630
-
The earliest reasonable candidate is as soon as the last
631
-
CPU has reported its quiescent state, but it may be some
632
-
milliseconds before RCU becomes aware of this.
633
-
The latest reasonable candidate is once the <tt>rcu_state</tt>
634
-
structure's <tt>->gp_seq</tt> field has been updated,
635
-
but it is quite possible that some CPUs have already completed
636
-
phase two of their updates by that time.
637
-
In short, if you are going to work with RCU, you need to
638
-
learn to embrace uncertainty.
639
-
</font></td></tr>
640
-
<tr><td> </td></tr>
641
-
</table>
642
-
643
-
644
-
<h4><a name="Callback Invocation">Callback Invocation</a></h4>
645
-
646
-
<p>Once a given CPU's leaf <tt>rcu_node</tt> structure's
647
-
<tt>->gp_seq</tt> field has been updated, that CPU can begin
648
-
invoking its RCU callbacks that were waiting for this grace period
649
-
to end.
650
-
These callbacks are identified by <tt>rcu_advance_cbs()</tt>,
651
-
which is usually invoked by <tt>__note_gp_changes()</tt>.
652
-
As shown in the diagram below, this invocation can be triggered by
653
-
the scheduling-clock interrupt (<tt>rcu_sched_clock_irq()</tt> on
654
-
the left) or by idle entry (<tt>rcu_cleanup_after_idle()</tt> on
655
-
the right, but only for kernels build with
656
-
<tt>CONFIG_RCU_FAST_NO_HZ=y</tt>).
657
-
Either way, <tt>RCU_SOFTIRQ</tt> is raised, which results in
658
-
<tt>rcu_do_batch()</tt> invoking the callbacks, which in turn
659
-
allows those callbacks to carry out (either directly or indirectly
660
-
via wakeup) the needed phase-two processing for each update.
661
-
662
-
</p><p><img src="TreeRCU-callback-invocation.svg" alt="TreeRCU-callback-invocation.svg" width="60%">
663
-
664
-
<p>Please note that callback invocation can also be prompted by any
665
-
number of corner-case code paths, for example, when a CPU notes that
666
-
it has excessive numbers of callbacks queued.
667
-
In all cases, the CPU acquires its leaf <tt>rcu_node</tt> structure's
668
-
<tt>->lock</tt> before invoking callbacks, which preserves the
669
-
required ordering against the newly completed grace period.
670
-
671
-
<p>However, if the callback function communicates to other CPUs,
672
-
for example, doing a wakeup, then it is that function's responsibility
673
-
to maintain ordering.
674
-
For example, if the callback function wakes up a task that runs on
675
-
some other CPU, proper ordering must in place in both the callback
676
-
function and the task being awakened.
677
-
To see why this is important, consider the top half of the
678
-
<a href="#Grace-Period Cleanup">grace-period cleanup</a> diagram.
679
-
The callback might be running on a CPU corresponding to the leftmost
680
-
leaf <tt>rcu_node</tt> structure, and awaken a task that is to run on
681
-
a CPU corresponding to the rightmost leaf <tt>rcu_node</tt> structure,
682
-
and the grace-period kernel thread might not yet have reached the
683
-
rightmost leaf.
684
-
In this case, the grace period's memory ordering might not yet have
685
-
reached that CPU, so again the callback function and the awakened
686
-
task must supply proper ordering.
687
-
688
-
<h3><a name="Putting It All Together">Putting It All Together</a></h3>
689
-
690
-
<p>A stitched-together diagram is
691
-
<a href="Tree-RCU-Diagram.html">here</a>.
692
-
693
-
<h3><a name="Legal Statement">
694
-
Legal Statement</a></h3>
695
-
696
-
<p>This work represents the view of the author and does not necessarily
697
-
represent the view of IBM.
698
-
699
-
</p><p>Linux is a registered trademark of Linus Torvalds.
700
-
701
-
</p><p>Other company, product, and service names may be trademarks or
702
-
service marks of others.
703
-
704
-
</body></html>
+625
Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst
+625
Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst
···
1
+
======================================================
2
+
A Tour Through TREE_RCU's Grace-Period Memory Ordering
3
+
======================================================
4
+
5
+
August 8, 2017
6
+
7
+
This article was contributed by Paul E. McKenney
8
+
9
+
Introduction
10
+
============
11
+
12
+
This document gives a rough visual overview of how Tree RCU's
13
+
grace-period memory ordering guarantee is provided.
14
+
15
+
What Is Tree RCU's Grace Period Memory Ordering Guarantee?
16
+
==========================================================
17
+
18
+
RCU grace periods provide extremely strong memory-ordering guarantees
19
+
for non-idle non-offline code.
20
+
Any code that happens after the end of a given RCU grace period is guaranteed
21
+
to see the effects of all accesses prior to the beginning of that grace
22
+
period that are within RCU read-side critical sections.
23
+
Similarly, any code that happens before the beginning of a given RCU grace
24
+
period is guaranteed to see the effects of all accesses following the end
25
+
of that grace period that are within RCU read-side critical sections.
26
+
27
+
Note well that RCU-sched read-side critical sections include any region
28
+
of code for which preemption is disabled.
29
+
Given that each individual machine instruction can be thought of as
30
+
an extremely small region of preemption-disabled code, one can think of
31
+
``synchronize_rcu()`` as ``smp_mb()`` on steroids.
32
+
33
+
RCU updaters use this guarantee by splitting their updates into
34
+
two phases, one of which is executed before the grace period and
35
+
the other of which is executed after the grace period.
36
+
In the most common use case, phase one removes an element from
37
+
a linked RCU-protected data structure, and phase two frees that element.
38
+
For this to work, any readers that have witnessed state prior to the
39
+
phase-one update (in the common case, removal) must not witness state
40
+
following the phase-two update (in the common case, freeing).
41
+
42
+
The RCU implementation provides this guarantee using a network
43
+
of lock-based critical sections, memory barriers, and per-CPU
44
+
processing, as is described in the following sections.
45
+
46
+
Tree RCU Grace Period Memory Ordering Building Blocks
47
+
=====================================================
48
+
49
+
The workhorse for RCU's grace-period memory ordering is the
50
+
critical section for the ``rcu_node`` structure's
51
+
``->lock``. These critical sections use helper functions for lock
52
+
acquisition, including ``raw_spin_lock_rcu_node()``,
53
+
``raw_spin_lock_irq_rcu_node()``, and ``raw_spin_lock_irqsave_rcu_node()``.
54
+
Their lock-release counterparts are ``raw_spin_unlock_rcu_node()``,
55
+
``raw_spin_unlock_irq_rcu_node()``, and
56
+
``raw_spin_unlock_irqrestore_rcu_node()``, respectively.
57
+
For completeness, a ``raw_spin_trylock_rcu_node()`` is also provided.
58
+
The key point is that the lock-acquisition functions, including
59
+
``raw_spin_trylock_rcu_node()``, all invoke ``smp_mb__after_unlock_lock()``
60
+
immediately after successful acquisition of the lock.
61
+
62
+
Therefore, for any given ``rcu_node`` structure, any access
63
+
happening before one of the above lock-release functions will be seen
64
+
by all CPUs as happening before any access happening after a later
65
+
one of the above lock-acquisition functions.
66
+
Furthermore, any access happening before one of the
67
+
above lock-release function on any given CPU will be seen by all
68
+
CPUs as happening before any access happening after a later one
69
+
of the above lock-acquisition functions executing on that same CPU,
70
+
even if the lock-release and lock-acquisition functions are operating
71
+
on different ``rcu_node`` structures.
72
+
Tree RCU uses these two ordering guarantees to form an ordering
73
+
network among all CPUs that were in any way involved in the grace
74
+
period, including any CPUs that came online or went offline during
75
+
the grace period in question.
76
+
77
+
The following litmus test exhibits the ordering effects of these
78
+
lock-acquisition and lock-release functions::
79
+
80
+
1 int x, y, z;
81
+
2
82
+
3 void task0(void)
83
+
4 {
84
+
5 raw_spin_lock_rcu_node(rnp);
85
+
6 WRITE_ONCE(x, 1);
86
+
7 r1 = READ_ONCE(y);
87
+
8 raw_spin_unlock_rcu_node(rnp);
88
+
9 }
89
+
10
90
+
11 void task1(void)
91
+
12 {
92
+
13 raw_spin_lock_rcu_node(rnp);
93
+
14 WRITE_ONCE(y, 1);
94
+
15 r2 = READ_ONCE(z);
95
+
16 raw_spin_unlock_rcu_node(rnp);
96
+
17 }
97
+
18
98
+
19 void task2(void)
99
+
20 {
100
+
21 WRITE_ONCE(z, 1);
101
+
22 smp_mb();
102
+
23 r3 = READ_ONCE(x);
103
+
24 }
104
+
25
105
+
26 WARN_ON(r1 == 0 && r2 == 0 && r3 == 0);
106
+
107
+
The ``WARN_ON()`` is evaluated at “the end of time”,
108
+
after all changes have propagated throughout the system.
109
+
Without the ``smp_mb__after_unlock_lock()`` provided by the
110
+
acquisition functions, this ``WARN_ON()`` could trigger, for example
111
+
on PowerPC.
112
+
The ``smp_mb__after_unlock_lock()`` invocations prevent this
113
+
``WARN_ON()`` from triggering.
114
+
115
+
This approach must be extended to include idle CPUs, which need
116
+
RCU's grace-period memory ordering guarantee to extend to any
117
+
RCU read-side critical sections preceding and following the current
118
+
idle sojourn.
119
+
This case is handled by calls to the strongly ordered
120
+
``atomic_add_return()`` read-modify-write atomic operation that
121
+
is invoked within ``rcu_dynticks_eqs_enter()`` at idle-entry
122
+
time and within ``rcu_dynticks_eqs_exit()`` at idle-exit time.
123
+
The grace-period kthread invokes ``rcu_dynticks_snap()`` and
124
+
``rcu_dynticks_in_eqs_since()`` (both of which invoke
125
+
an ``atomic_add_return()`` of zero) to detect idle CPUs.
126
+
127
+
+-----------------------------------------------------------------------+
128
+
| **Quick Quiz**: |
129
+
+-----------------------------------------------------------------------+
130
+
| But what about CPUs that remain offline for the entire grace period? |
131
+
+-----------------------------------------------------------------------+
132
+
| **Answer**: |
133
+
+-----------------------------------------------------------------------+
134
+
| Such CPUs will be offline at the beginning of the grace period, so |
135
+
| the grace period won't expect quiescent states from them. Races |
136
+
| between grace-period start and CPU-hotplug operations are mediated |
137
+
| by the CPU's leaf ``rcu_node`` structure's ``->lock`` as described |
138
+
| above. |
139
+
+-----------------------------------------------------------------------+
140
+
141
+
The approach must be extended to handle one final case, that of waking a
142
+
task blocked in ``synchronize_rcu()``. This task might be affinitied to
143
+
a CPU that is not yet aware that the grace period has ended, and thus
144
+
might not yet be subject to the grace period's memory ordering.
145
+
Therefore, there is an ``smp_mb()`` after the return from
146
+
``wait_for_completion()`` in the ``synchronize_rcu()`` code path.
147
+
148
+
+-----------------------------------------------------------------------+
149
+
| **Quick Quiz**: |
150
+
+-----------------------------------------------------------------------+
151
+
| What? Where??? I don't see any ``smp_mb()`` after the return from |
152
+
| ``wait_for_completion()``!!! |
153
+
+-----------------------------------------------------------------------+
154
+
| **Answer**: |
155
+
+-----------------------------------------------------------------------+
156
+
| That would be because I spotted the need for that ``smp_mb()`` during |
157
+
| the creation of this documentation, and it is therefore unlikely to |
158
+
| hit mainline before v4.14. Kudos to Lance Roy, Will Deacon, Peter |
159
+
| Zijlstra, and Jonathan Cameron for asking questions that sensitized |
160
+
| me to the rather elaborate sequence of events that demonstrate the |
161
+
| need for this memory barrier. |
162
+
+-----------------------------------------------------------------------+
163
+
164
+
Tree RCU's grace--period memory-ordering guarantees rely most heavily on
165
+
the ``rcu_node`` structure's ``->lock`` field, so much so that it is
166
+
necessary to abbreviate this pattern in the diagrams in the next
167
+
section. For example, consider the ``rcu_prepare_for_idle()`` function
168
+
shown below, which is one of several functions that enforce ordering of
169
+
newly arrived RCU callbacks against future grace periods:
170
+
171
+
::
172
+
173
+
1 static void rcu_prepare_for_idle(void)
174
+
2 {
175
+
3 bool needwake;
176
+
4 struct rcu_data *rdp;
177
+
5 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
178
+
6 struct rcu_node *rnp;
179
+
7 struct rcu_state *rsp;
180
+
8 int tne;
181
+
9
182
+
10 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) ||
183
+
11 rcu_is_nocb_cpu(smp_processor_id()))
184
+
12 return;
185
+
13 tne = READ_ONCE(tick_nohz_active);
186
+
14 if (tne != rdtp->tick_nohz_enabled_snap) {
187
+
15 if (rcu_cpu_has_callbacks(NULL))
188
+
16 invoke_rcu_core();
189
+
17 rdtp->tick_nohz_enabled_snap = tne;
190
+
18 return;
191
+
19 }
192
+
20 if (!tne)
193
+
21 return;
194
+
22 if (rdtp->all_lazy &&
195
+
23 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
196
+
24 rdtp->all_lazy = false;
197
+
25 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
198
+
26 invoke_rcu_core();
199
+
27 return;
200
+
28 }
201
+
29 if (rdtp->last_accelerate == jiffies)
202
+
30 return;
203
+
31 rdtp->last_accelerate = jiffies;
204
+
32 for_each_rcu_flavor(rsp) {
205
+
33 rdp = this_cpu_ptr(rsp->rda);
206
+
34 if (rcu_segcblist_pend_cbs(&rdp->cblist))
207
+
35 continue;
208
+
36 rnp = rdp->mynode;
209
+
37 raw_spin_lock_rcu_node(rnp);
210
+
38 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
211
+
39 raw_spin_unlock_rcu_node(rnp);
212
+
40 if (needwake)
213
+
41 rcu_gp_kthread_wake(rsp);
214
+
42 }
215
+
43 }
216
+
217
+
But the only part of ``rcu_prepare_for_idle()`` that really matters for
218
+
this discussion are lines 37–39. We will therefore abbreviate this
219
+
function as follows:
220
+
221
+
.. kernel-figure:: rcu_node-lock.svg
222
+
223
+
The box represents the ``rcu_node`` structure's ``->lock`` critical
224
+
section, with the double line on top representing the additional
225
+
``smp_mb__after_unlock_lock()``.
226
+
227
+
Tree RCU Grace Period Memory Ordering Components
228
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
229
+
230
+
Tree RCU's grace-period memory-ordering guarantee is provided by a
231
+
number of RCU components:
232
+
233
+
#. `Callback Registry <#Callback%20Registry>`__
234
+
#. `Grace-Period Initialization <#Grace-Period%20Initialization>`__
235
+
#. `Self-Reported Quiescent
236
+
States <#Self-Reported%20Quiescent%20States>`__
237
+
#. `Dynamic Tick Interface <#Dynamic%20Tick%20Interface>`__
238
+
#. `CPU-Hotplug Interface <#CPU-Hotplug%20Interface>`__
239
+
#. `Forcing Quiescent States <Forcing%20Quiescent%20States>`__
240
+
#. `Grace-Period Cleanup <Grace-Period%20Cleanup>`__
241
+
#. `Callback Invocation <Callback%20Invocation>`__
242
+
243
+
Each of the following section looks at the corresponding component in
244
+
detail.
245
+
246
+
Callback Registry
247
+
^^^^^^^^^^^^^^^^^
248
+
249
+
If RCU's grace-period guarantee is to mean anything at all, any access
250
+
that happens before a given invocation of ``call_rcu()`` must also
251
+
happen before the corresponding grace period. The implementation of this
252
+
portion of RCU's grace period guarantee is shown in the following
253
+
figure:
254
+
255
+
.. kernel-figure:: TreeRCU-callback-registry.svg
256
+
257
+
Because ``call_rcu()`` normally acts only on CPU-local state, it
258
+
provides no ordering guarantees, either for itself or for phase one of
259
+
the update (which again will usually be removal of an element from an
260
+
RCU-protected data structure). It simply enqueues the ``rcu_head``
261
+
structure on a per-CPU list, which cannot become associated with a grace
262
+
period until a later call to ``rcu_accelerate_cbs()``, as shown in the
263
+
diagram above.
264
+
265
+
One set of code paths shown on the left invokes ``rcu_accelerate_cbs()``
266
+
via ``note_gp_changes()``, either directly from ``call_rcu()`` (if the
267
+
current CPU is inundated with queued ``rcu_head`` structures) or more
268
+
likely from an ``RCU_SOFTIRQ`` handler. Another code path in the middle
269
+
is taken only in kernels built with ``CONFIG_RCU_FAST_NO_HZ=y``, which
270
+
invokes ``rcu_accelerate_cbs()`` via ``rcu_prepare_for_idle()``. The
271
+
final code path on the right is taken only in kernels built with
272
+
``CONFIG_HOTPLUG_CPU=y``, which invokes ``rcu_accelerate_cbs()`` via
273
+
``rcu_advance_cbs()``, ``rcu_migrate_callbacks``,
274
+
``rcutree_migrate_callbacks()``, and ``takedown_cpu()``, which in turn
275
+
is invoked on a surviving CPU after the outgoing CPU has been completely
276
+
offlined.
277
+
278
+
There are a few other code paths within grace-period processing that
279
+
opportunistically invoke ``rcu_accelerate_cbs()``. However, either way,
280
+
all of the CPU's recently queued ``rcu_head`` structures are associated
281
+
with a future grace-period number under the protection of the CPU's lead
282
+
``rcu_node`` structure's ``->lock``. In all cases, there is full
283
+
ordering against any prior critical section for that same ``rcu_node``
284
+
structure's ``->lock``, and also full ordering against any of the
285
+
current task's or CPU's prior critical sections for any ``rcu_node``
286
+
structure's ``->lock``.
287
+
288
+
The next section will show how this ordering ensures that any accesses
289
+
prior to the ``call_rcu()`` (particularly including phase one of the
290
+
update) happen before the start of the corresponding grace period.
291
+
292
+
+-----------------------------------------------------------------------+
293
+
| **Quick Quiz**: |
294
+
+-----------------------------------------------------------------------+
295
+
| But what about ``synchronize_rcu()``? |
296
+
+-----------------------------------------------------------------------+
297
+
| **Answer**: |
298
+
+-----------------------------------------------------------------------+
299
+
| The ``synchronize_rcu()`` passes ``call_rcu()`` to ``wait_rcu_gp()``, |
300
+
| which invokes it. So either way, it eventually comes down to |
301
+
| ``call_rcu()``. |
302
+
+-----------------------------------------------------------------------+
303
+
304
+
Grace-Period Initialization
305
+
^^^^^^^^^^^^^^^^^^^^^^^^^^^
306
+
307
+
Grace-period initialization is carried out by the grace-period kernel
308
+
thread, which makes several passes over the ``rcu_node`` tree within the
309
+
``rcu_gp_init()`` function. This means that showing the full flow of
310
+
ordering through the grace-period computation will require duplicating
311
+
this tree. If you find this confusing, please note that the state of the
312
+
``rcu_node`` changes over time, just like Heraclitus's river. However,
313
+
to keep the ``rcu_node`` river tractable, the grace-period kernel
314
+
thread's traversals are presented in multiple parts, starting in this
315
+
section with the various phases of grace-period initialization.
316
+
317
+
The first ordering-related grace-period initialization action is to
318
+
advance the ``rcu_state`` structure's ``->gp_seq`` grace-period-number
319
+
counter, as shown below:
320
+
321
+
.. kernel-figure:: TreeRCU-gp-init-1.svg
322
+
323
+
The actual increment is carried out using ``smp_store_release()``, which
324
+
helps reject false-positive RCU CPU stall detection. Note that only the
325
+
root ``rcu_node`` structure is touched.
326
+
327
+
The first pass through the ``rcu_node`` tree updates bitmasks based on
328
+
CPUs having come online or gone offline since the start of the previous
329
+
grace period. In the common case where the number of online CPUs for
330
+
this ``rcu_node`` structure has not transitioned to or from zero, this
331
+
pass will scan only the leaf ``rcu_node`` structures. However, if the
332
+
number of online CPUs for a given leaf ``rcu_node`` structure has
333
+
transitioned from zero, ``rcu_init_new_rnp()`` will be invoked for the
334
+
first incoming CPU. Similarly, if the number of online CPUs for a given
335
+
leaf ``rcu_node`` structure has transitioned to zero,
336
+
``rcu_cleanup_dead_rnp()`` will be invoked for the last outgoing CPU.
337
+
The diagram below shows the path of ordering if the leftmost
338
+
``rcu_node`` structure onlines its first CPU and if the next
339
+
``rcu_node`` structure has no online CPUs (or, alternatively if the
340
+
leftmost ``rcu_node`` structure offlines its last CPU and if the next
341
+
``rcu_node`` structure has no online CPUs).
342
+
343
+
.. kernel-figure:: TreeRCU-gp-init-1.svg
344
+
345
+
The final ``rcu_gp_init()`` pass through the ``rcu_node`` tree traverses
346
+
breadth-first, setting each ``rcu_node`` structure's ``->gp_seq`` field
347
+
to the newly advanced value from the ``rcu_state`` structure, as shown
348
+
in the following diagram.
349
+
350
+
.. kernel-figure:: TreeRCU-gp-init-1.svg
351
+
352
+
This change will also cause each CPU's next call to
353
+
``__note_gp_changes()`` to notice that a new grace period has started,
354
+
as described in the next section. But because the grace-period kthread
355
+
started the grace period at the root (with the advancing of the
356
+
``rcu_state`` structure's ``->gp_seq`` field) before setting each leaf
357
+
``rcu_node`` structure's ``->gp_seq`` field, each CPU's observation of
358
+
the start of the grace period will happen after the actual start of the
359
+
grace period.
360
+
361
+
+-----------------------------------------------------------------------+
362
+
| **Quick Quiz**: |
363
+
+-----------------------------------------------------------------------+
364
+
| But what about the CPU that started the grace period? Why wouldn't it |
365
+
| see the start of the grace period right when it started that grace |
366
+
| period? |
367
+
+-----------------------------------------------------------------------+
368
+
| **Answer**: |
369
+
+-----------------------------------------------------------------------+
370
+
| In some deep philosophical and overly anthromorphized sense, yes, the |
371
+
| CPU starting the grace period is immediately aware of having done so. |
372
+
| However, if we instead assume that RCU is not self-aware, then even |
373
+
| the CPU starting the grace period does not really become aware of the |
374
+
| start of this grace period until its first call to |
375
+
| ``__note_gp_changes()``. On the other hand, this CPU potentially gets |
376
+
| early notification because it invokes ``__note_gp_changes()`` during |
377
+
| its last ``rcu_gp_init()`` pass through its leaf ``rcu_node`` |
378
+
| structure. |
379
+
+-----------------------------------------------------------------------+
380
+
381
+
Self-Reported Quiescent States
382
+
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
383
+
384
+
When all entities that might block the grace period have reported
385
+
quiescent states (or as described in a later section, had quiescent
386
+
states reported on their behalf), the grace period can end. Online
387
+
non-idle CPUs report their own quiescent states, as shown in the
388
+
following diagram:
389
+
390
+
.. kernel-figure:: TreeRCU-qs.svg
391
+
392
+
This is for the last CPU to report a quiescent state, which signals the
393
+
end of the grace period. Earlier quiescent states would push up the
394
+
``rcu_node`` tree only until they encountered an ``rcu_node`` structure
395
+
that is waiting for additional quiescent states. However, ordering is
396
+
nevertheless preserved because some later quiescent state will acquire
397
+
that ``rcu_node`` structure's ``->lock``.
398
+
399
+
Any number of events can lead up to a CPU invoking ``note_gp_changes``
400
+
(or alternatively, directly invoking ``__note_gp_changes()``), at which
401
+
point that CPU will notice the start of a new grace period while holding
402
+
its leaf ``rcu_node`` lock. Therefore, all execution shown in this
403
+
diagram happens after the start of the grace period. In addition, this
404
+
CPU will consider any RCU read-side critical section that started before
405
+
the invocation of ``__note_gp_changes()`` to have started before the
406
+
grace period, and thus a critical section that the grace period must
407
+
wait on.
408
+
409
+
+-----------------------------------------------------------------------+
410
+
| **Quick Quiz**: |
411
+
+-----------------------------------------------------------------------+
412
+
| But a RCU read-side critical section might have started after the |
413
+
| beginning of the grace period (the advancing of ``->gp_seq`` from |
414
+
| earlier), so why should the grace period wait on such a critical |
415
+
| section? |
416
+
+-----------------------------------------------------------------------+
417
+
| **Answer**: |
418
+
+-----------------------------------------------------------------------+
419
+
| It is indeed not necessary for the grace period to wait on such a |
420
+
| critical section. However, it is permissible to wait on it. And it is |
421
+
| furthermore important to wait on it, as this lazy approach is far |
422
+
| more scalable than a “big bang” all-at-once grace-period start could |
423
+
| possibly be. |
424
+
+-----------------------------------------------------------------------+
425
+
426
+
If the CPU does a context switch, a quiescent state will be noted by
427
+
``rcu_node_context_switch()`` on the left. On the other hand, if the CPU
428
+
takes a scheduler-clock interrupt while executing in usermode, a
429
+
quiescent state will be noted by ``rcu_sched_clock_irq()`` on the right.
430
+
Either way, the passage through a quiescent state will be noted in a
431
+
per-CPU variable.
432
+
433
+
The next time an ``RCU_SOFTIRQ`` handler executes on this CPU (for
434
+
example, after the next scheduler-clock interrupt), ``rcu_core()`` will
435
+
invoke ``rcu_check_quiescent_state()``, which will notice the recorded
436
+
quiescent state, and invoke ``rcu_report_qs_rdp()``. If
437
+
``rcu_report_qs_rdp()`` verifies that the quiescent state really does
438
+
apply to the current grace period, it invokes ``rcu_report_rnp()`` which
439
+
traverses up the ``rcu_node`` tree as shown at the bottom of the
440
+
diagram, clearing bits from each ``rcu_node`` structure's ``->qsmask``
441
+
field, and propagating up the tree when the result is zero.
442
+
443
+
Note that traversal passes upwards out of a given ``rcu_node`` structure
444
+
only if the current CPU is reporting the last quiescent state for the
445
+
subtree headed by that ``rcu_node`` structure. A key point is that if a
446
+
CPU's traversal stops at a given ``rcu_node`` structure, then there will
447
+
be a later traversal by another CPU (or perhaps the same one) that
448
+
proceeds upwards from that point, and the ``rcu_node`` ``->lock``
449
+
guarantees that the first CPU's quiescent state happens before the
450
+
remainder of the second CPU's traversal. Applying this line of thought
451
+
repeatedly shows that all CPUs' quiescent states happen before the last
452
+
CPU traverses through the root ``rcu_node`` structure, the “last CPU”
453
+
being the one that clears the last bit in the root ``rcu_node``
454
+
structure's ``->qsmask`` field.
455
+
456
+
Dynamic Tick Interface
457
+
^^^^^^^^^^^^^^^^^^^^^^
458
+
459
+
Due to energy-efficiency considerations, RCU is forbidden from
460
+
disturbing idle CPUs. CPUs are therefore required to notify RCU when
461
+
entering or leaving idle state, which they do via fully ordered
462
+
value-returning atomic operations on a per-CPU variable. The ordering
463
+
effects are as shown below:
464
+
465
+
.. kernel-figure:: TreeRCU-dyntick.svg
466
+
467
+
The RCU grace-period kernel thread samples the per-CPU idleness variable
468
+
while holding the corresponding CPU's leaf ``rcu_node`` structure's
469
+
``->lock``. This means that any RCU read-side critical sections that
470
+
precede the idle period (the oval near the top of the diagram above)
471
+
will happen before the end of the current grace period. Similarly, the
472
+
beginning of the current grace period will happen before any RCU
473
+
read-side critical sections that follow the idle period (the oval near
474
+
the bottom of the diagram above).
475
+
476
+
Plumbing this into the full grace-period execution is described
477
+
`below <#Forcing%20Quiescent%20States>`__.
478
+
479
+
CPU-Hotplug Interface
480
+
^^^^^^^^^^^^^^^^^^^^^
481
+
482
+
RCU is also forbidden from disturbing offline CPUs, which might well be
483
+
powered off and removed from the system completely. CPUs are therefore
484
+
required to notify RCU of their comings and goings as part of the
485
+
corresponding CPU hotplug operations. The ordering effects are shown
486
+
below:
487
+
488
+
.. kernel-figure:: TreeRCU-hotplug.svg
489
+
490
+
Because CPU hotplug operations are much less frequent than idle
491
+
transitions, they are heavier weight, and thus acquire the CPU's leaf
492
+
``rcu_node`` structure's ``->lock`` and update this structure's
493
+
``->qsmaskinitnext``. The RCU grace-period kernel thread samples this
494
+
mask to detect CPUs having gone offline since the beginning of this
495
+
grace period.
496
+
497
+
Plumbing this into the full grace-period execution is described
498
+
`below <#Forcing%20Quiescent%20States>`__.
499
+
500
+
Forcing Quiescent States
501
+
^^^^^^^^^^^^^^^^^^^^^^^^
502
+
503
+
As noted above, idle and offline CPUs cannot report their own quiescent
504
+
states, and therefore the grace-period kernel thread must do the
505
+
reporting on their behalf. This process is called “forcing quiescent
506
+
states”, it is repeated every few jiffies, and its ordering effects are
507
+
shown below:
508
+
509
+
.. kernel-figure:: TreeRCU-gp-fqs.svg
510
+
511
+
Each pass of quiescent state forcing is guaranteed to traverse the leaf
512
+
``rcu_node`` structures, and if there are no new quiescent states due to
513
+
recently idled and/or offlined CPUs, then only the leaves are traversed.
514
+
However, if there is a newly offlined CPU as illustrated on the left or
515
+
a newly idled CPU as illustrated on the right, the corresponding
516
+
quiescent state will be driven up towards the root. As with
517
+
self-reported quiescent states, the upwards driving stops once it
518
+
reaches an ``rcu_node`` structure that has quiescent states outstanding
519
+
from other CPUs.
520
+
521
+
+-----------------------------------------------------------------------+
522
+
| **Quick Quiz**: |
523
+
+-----------------------------------------------------------------------+
524
+
| The leftmost drive to root stopped before it reached the root |
525
+
| ``rcu_node`` structure, which means that there are still CPUs |
526
+
| subordinate to that structure on which the current grace period is |
527
+
| waiting. Given that, how is it possible that the rightmost drive to |
528
+
| root ended the grace period? |
529
+
+-----------------------------------------------------------------------+
530
+
| **Answer**: |
531
+
+-----------------------------------------------------------------------+
532
+
| Good analysis! It is in fact impossible in the absence of bugs in |
533
+
| RCU. But this diagram is complex enough as it is, so simplicity |
534
+
| overrode accuracy. You can think of it as poetic license, or you can |
535
+
| think of it as misdirection that is resolved in the |
536
+
| `stitched-together diagram <#Putting%20It%20All%20Together>`__. |
537
+
+-----------------------------------------------------------------------+
538
+
539
+
Grace-Period Cleanup
540
+
^^^^^^^^^^^^^^^^^^^^
541
+
542
+
Grace-period cleanup first scans the ``rcu_node`` tree breadth-first
543
+
advancing all the ``->gp_seq`` fields, then it advances the
544
+
``rcu_state`` structure's ``->gp_seq`` field. The ordering effects are
545
+
shown below:
546
+
547
+
.. kernel-figure:: TreeRCU-gp-cleanup.svg
548
+
549
+
As indicated by the oval at the bottom of the diagram, once grace-period
550
+
cleanup is complete, the next grace period can begin.
551
+
552
+
+-----------------------------------------------------------------------+
553
+
| **Quick Quiz**: |
554
+
+-----------------------------------------------------------------------+
555
+
| But when precisely does the grace period end? |
556
+
+-----------------------------------------------------------------------+
557
+
| **Answer**: |
558
+
+-----------------------------------------------------------------------+
559
+
| There is no useful single point at which the grace period can be said |
560
+
| to end. The earliest reasonable candidate is as soon as the last CPU |
561
+
| has reported its quiescent state, but it may be some milliseconds |
562
+
| before RCU becomes aware of this. The latest reasonable candidate is |
563
+
| once the ``rcu_state`` structure's ``->gp_seq`` field has been |
564
+
| updated, but it is quite possible that some CPUs have already |
565
+
| completed phase two of their updates by that time. In short, if you |
566
+
| are going to work with RCU, you need to learn to embrace uncertainty. |
567
+
+-----------------------------------------------------------------------+
568
+
569
+
Callback Invocation
570
+
^^^^^^^^^^^^^^^^^^^
571
+
572
+
Once a given CPU's leaf ``rcu_node`` structure's ``->gp_seq`` field has
573
+
been updated, that CPU can begin invoking its RCU callbacks that were
574
+
waiting for this grace period to end. These callbacks are identified by
575
+
``rcu_advance_cbs()``, which is usually invoked by
576
+
``__note_gp_changes()``. As shown in the diagram below, this invocation
577
+
can be triggered by the scheduling-clock interrupt
578
+
(``rcu_sched_clock_irq()`` on the left) or by idle entry
579
+
(``rcu_cleanup_after_idle()`` on the right, but only for kernels build
580
+
with ``CONFIG_RCU_FAST_NO_HZ=y``). Either way, ``RCU_SOFTIRQ`` is
581
+
raised, which results in ``rcu_do_batch()`` invoking the callbacks,
582
+
which in turn allows those callbacks to carry out (either directly or
583
+
indirectly via wakeup) the needed phase-two processing for each update.
584
+
585
+
.. kernel-figure:: TreeRCU-callback-invocation.svg
586
+
587
+
Please note that callback invocation can also be prompted by any number
588
+
of corner-case code paths, for example, when a CPU notes that it has
589
+
excessive numbers of callbacks queued. In all cases, the CPU acquires
590
+
its leaf ``rcu_node`` structure's ``->lock`` before invoking callbacks,
591
+
which preserves the required ordering against the newly completed grace
592
+
period.
593
+
594
+
However, if the callback function communicates to other CPUs, for
595
+
example, doing a wakeup, then it is that function's responsibility to
596
+
maintain ordering. For example, if the callback function wakes up a task
597
+
that runs on some other CPU, proper ordering must in place in both the
598
+
callback function and the task being awakened. To see why this is
599
+
important, consider the top half of the `grace-period
600
+
cleanup <#Grace-Period%20Cleanup>`__ diagram. The callback might be
601
+
running on a CPU corresponding to the leftmost leaf ``rcu_node``
602
+
structure, and awaken a task that is to run on a CPU corresponding to
603
+
the rightmost leaf ``rcu_node`` structure, and the grace-period kernel
604
+
thread might not yet have reached the rightmost leaf. In this case, the
605
+
grace period's memory ordering might not yet have reached that CPU, so
606
+
again the callback function and the awakened task must supply proper
607
+
ordering.
608
+
609
+
Putting It All Together
610
+
~~~~~~~~~~~~~~~~~~~~~~~
611
+
612
+
A stitched-together diagram is here:
613
+
614
+
.. kernel-figure:: TreeRCU-gp.svg
615
+
616
+
Legal Statement
617
+
~~~~~~~~~~~~~~~
618
+
619
+
This work represents the view of the author and does not necessarily
620
+
represent the view of IBM.
621
+
622
+
Linux is a registered trademark of Linus Torvalds.
623
+
624
+
Other company, product, and service names may be trademarks or service
625
+
marks of others.
-3330
Documentation/RCU/Design/Requirements/Requirements.html
-3330
Documentation/RCU/Design/Requirements/Requirements.html
···
1
-
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN"
2
-
"http://www.w3.org/TR/html4/loose.dtd">
3
-
<html>
4
-
<head><title>A Tour Through RCU's Requirements [LWN.net]</title>
5
-
<meta HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=utf-8">
6
-
7
-
<h1>A Tour Through RCU's Requirements</h1>
8
-
9
-
<p>Copyright IBM Corporation, 2015</p>
10
-
<p>Author: Paul E. McKenney</p>
11
-
<p><i>The initial version of this document appeared in the
12
-
<a href="https://lwn.net/">LWN</a> articles
13
-
<a href="https://lwn.net/Articles/652156/">here</a>,
14
-
<a href="https://lwn.net/Articles/652677/">here</a>, and
15
-
<a href="https://lwn.net/Articles/653326/">here</a>.</i></p>
16
-
17
-
<h2>Introduction</h2>
18
-
19
-
<p>
20
-
Read-copy update (RCU) is a synchronization mechanism that is often
21
-
used as a replacement for reader-writer locking.
22
-
RCU is unusual in that updaters do not block readers,
23
-
which means that RCU's read-side primitives can be exceedingly fast
24
-
and scalable.
25
-
In addition, updaters can make useful forward progress concurrently
26
-
with readers.
27
-
However, all this concurrency between RCU readers and updaters does raise
28
-
the question of exactly what RCU readers are doing, which in turn
29
-
raises the question of exactly what RCU's requirements are.
30
-
31
-
<p>
32
-
This document therefore summarizes RCU's requirements, and can be thought
33
-
of as an informal, high-level specification for RCU.
34
-
It is important to understand that RCU's specification is primarily
35
-
empirical in nature;
36
-
in fact, I learned about many of these requirements the hard way.
37
-
This situation might cause some consternation, however, not only
38
-
has this learning process been a lot of fun, but it has also been
39
-
a great privilege to work with so many people willing to apply
40
-
technologies in interesting new ways.
41
-
42
-
<p>
43
-
All that aside, here are the categories of currently known RCU requirements:
44
-
</p>
45
-
46
-
<ol>
47
-
<li> <a href="#Fundamental Requirements">
48
-
Fundamental Requirements</a>
49
-
<li> <a href="#Fundamental Non-Requirements">Fundamental Non-Requirements</a>
50
-
<li> <a href="#Parallelism Facts of Life">
51
-
Parallelism Facts of Life</a>
52
-
<li> <a href="#Quality-of-Implementation Requirements">
53
-
Quality-of-Implementation Requirements</a>
54
-
<li> <a href="#Linux Kernel Complications">
55
-
Linux Kernel Complications</a>
56
-
<li> <a href="#Software-Engineering Requirements">
57
-
Software-Engineering Requirements</a>
58
-
<li> <a href="#Other RCU Flavors">
59
-
Other RCU Flavors</a>
60
-
<li> <a href="#Possible Future Changes">
61
-
Possible Future Changes</a>
62
-
</ol>
63
-
64
-
<p>
65
-
This is followed by a <a href="#Summary">summary</a>,
66
-
however, the answers to each quick quiz immediately follows the quiz.
67
-
Select the big white space with your mouse to see the answer.
68
-
69
-
<h2><a name="Fundamental Requirements">Fundamental Requirements</a></h2>
70
-
71
-
<p>
72
-
RCU's fundamental requirements are the closest thing RCU has to hard
73
-
mathematical requirements.
74
-
These are:
75
-
76
-
<ol>
77
-
<li> <a href="#Grace-Period Guarantee">
78
-
Grace-Period Guarantee</a>
79
-
<li> <a href="#Publish-Subscribe Guarantee">
80
-
Publish-Subscribe Guarantee</a>
81
-
<li> <a href="#Memory-Barrier Guarantees">
82
-
Memory-Barrier Guarantees</a>
83
-
<li> <a href="#RCU Primitives Guaranteed to Execute Unconditionally">
84
-
RCU Primitives Guaranteed to Execute Unconditionally</a>
85
-
<li> <a href="#Guaranteed Read-to-Write Upgrade">
86
-
Guaranteed Read-to-Write Upgrade</a>
87
-
</ol>
88
-
89
-
<h3><a name="Grace-Period Guarantee">Grace-Period Guarantee</a></h3>
90
-
91
-
<p>
92
-
RCU's grace-period guarantee is unusual in being premeditated:
93
-
Jack Slingwine and I had this guarantee firmly in mind when we started
94
-
work on RCU (then called “rclock”) in the early 1990s.
95
-
That said, the past two decades of experience with RCU have produced
96
-
a much more detailed understanding of this guarantee.
97
-
98
-
<p>
99
-
RCU's grace-period guarantee allows updaters to wait for the completion
100
-
of all pre-existing RCU read-side critical sections.
101
-
An RCU read-side critical section
102
-
begins with the marker <tt>rcu_read_lock()</tt> and ends with
103
-
the marker <tt>rcu_read_unlock()</tt>.
104
-
These markers may be nested, and RCU treats a nested set as one
105
-
big RCU read-side critical section.
106
-
Production-quality implementations of <tt>rcu_read_lock()</tt> and
107
-
<tt>rcu_read_unlock()</tt> are extremely lightweight, and in
108
-
fact have exactly zero overhead in Linux kernels built for production
109
-
use with <tt>CONFIG_PREEMPT=n</tt>.
110
-
111
-
<p>
112
-
This guarantee allows ordering to be enforced with extremely low
113
-
overhead to readers, for example:
114
-
115
-
<blockquote>
116
-
<pre>
117
-
1 int x, y;
118
-
2
119
-
3 void thread0(void)
120
-
4 {
121
-
5 rcu_read_lock();
122
-
6 r1 = READ_ONCE(x);
123
-
7 r2 = READ_ONCE(y);
124
-
8 rcu_read_unlock();
125
-
9 }
126
-
10
127
-
11 void thread1(void)
128
-
12 {
129
-
13 WRITE_ONCE(x, 1);
130
-
14 synchronize_rcu();
131
-
15 WRITE_ONCE(y, 1);
132
-
16 }
133
-
</pre>
134
-
</blockquote>
135
-
136
-
<p>
137
-
Because the <tt>synchronize_rcu()</tt> on line 14 waits for
138
-
all pre-existing readers, any instance of <tt>thread0()</tt> that
139
-
loads a value of zero from <tt>x</tt> must complete before
140
-
<tt>thread1()</tt> stores to <tt>y</tt>, so that instance must
141
-
also load a value of zero from <tt>y</tt>.
142
-
Similarly, any instance of <tt>thread0()</tt> that loads a value of
143
-
one from <tt>y</tt> must have started after the
144
-
<tt>synchronize_rcu()</tt> started, and must therefore also load
145
-
a value of one from <tt>x</tt>.
146
-
Therefore, the outcome:
147
-
<blockquote>
148
-
<pre>
149
-
(r1 == 0 && r2 == 1)
150
-
</pre>
151
-
</blockquote>
152
-
cannot happen.
153
-
154
-
<table>
155
-
<tr><th> </th></tr>
156
-
<tr><th align="left">Quick Quiz:</th></tr>
157
-
<tr><td>
158
-
Wait a minute!
159
-
You said that updaters can make useful forward progress concurrently
160
-
with readers, but pre-existing readers will block
161
-
<tt>synchronize_rcu()</tt>!!!
162
-
Just who are you trying to fool???
163
-
</td></tr>
164
-
<tr><th align="left">Answer:</th></tr>
165
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
166
-
First, if updaters do not wish to be blocked by readers, they can use
167
-
<tt>call_rcu()</tt> or <tt>kfree_rcu()</tt>, which will
168
-
be discussed later.
169
-
Second, even when using <tt>synchronize_rcu()</tt>, the other
170
-
update-side code does run concurrently with readers, whether
171
-
pre-existing or not.
172
-
</font></td></tr>
173
-
<tr><td> </td></tr>
174
-
</table>
175
-
176
-
<p>
177
-
This scenario resembles one of the first uses of RCU in
178
-
<a href="https://en.wikipedia.org/wiki/DYNIX">DYNIX/ptx</a>,
179
-
which managed a distributed lock manager's transition into
180
-
a state suitable for handling recovery from node failure,
181
-
more or less as follows:
182
-
183
-
<blockquote>
184
-
<pre>
185
-
1 #define STATE_NORMAL 0
186
-
2 #define STATE_WANT_RECOVERY 1
187
-
3 #define STATE_RECOVERING 2
188
-
4 #define STATE_WANT_NORMAL 3
189
-
5
190
-
6 int state = STATE_NORMAL;
191
-
7
192
-
8 void do_something_dlm(void)
193
-
9 {
194
-
10 int state_snap;
195
-
11
196
-
12 rcu_read_lock();
197
-
13 state_snap = READ_ONCE(state);
198
-
14 if (state_snap == STATE_NORMAL)
199
-
15 do_something();
200
-
16 else
201
-
17 do_something_carefully();
202
-
18 rcu_read_unlock();
203
-
19 }
204
-
20
205
-
21 void start_recovery(void)
206
-
22 {
207
-
23 WRITE_ONCE(state, STATE_WANT_RECOVERY);
208
-
24 synchronize_rcu();
209
-
25 WRITE_ONCE(state, STATE_RECOVERING);
210
-
26 recovery();
211
-
27 WRITE_ONCE(state, STATE_WANT_NORMAL);
212
-
28 synchronize_rcu();
213
-
29 WRITE_ONCE(state, STATE_NORMAL);
214
-
30 }
215
-
</pre>
216
-
</blockquote>
217
-
218
-
<p>
219
-
The RCU read-side critical section in <tt>do_something_dlm()</tt>
220
-
works with the <tt>synchronize_rcu()</tt> in <tt>start_recovery()</tt>
221
-
to guarantee that <tt>do_something()</tt> never runs concurrently
222
-
with <tt>recovery()</tt>, but with little or no synchronization
223
-
overhead in <tt>do_something_dlm()</tt>.
224
-
225
-
<table>
226
-
<tr><th> </th></tr>
227
-
<tr><th align="left">Quick Quiz:</th></tr>
228
-
<tr><td>
229
-
Why is the <tt>synchronize_rcu()</tt> on line 28 needed?
230
-
</td></tr>
231
-
<tr><th align="left">Answer:</th></tr>
232
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
233
-
Without that extra grace period, memory reordering could result in
234
-
<tt>do_something_dlm()</tt> executing <tt>do_something()</tt>
235
-
concurrently with the last bits of <tt>recovery()</tt>.
236
-
</font></td></tr>
237
-
<tr><td> </td></tr>
238
-
</table>
239
-
240
-
<p>
241
-
In order to avoid fatal problems such as deadlocks,
242
-
an RCU read-side critical section must not contain calls to
243
-
<tt>synchronize_rcu()</tt>.
244
-
Similarly, an RCU read-side critical section must not
245
-
contain anything that waits, directly or indirectly, on completion of
246
-
an invocation of <tt>synchronize_rcu()</tt>.
247
-
248
-
<p>
249
-
Although RCU's grace-period guarantee is useful in and of itself, with
250
-
<a href="https://lwn.net/Articles/573497/">quite a few use cases</a>,
251
-
it would be good to be able to use RCU to coordinate read-side
252
-
access to linked data structures.
253
-
For this, the grace-period guarantee is not sufficient, as can
254
-
be seen in function <tt>add_gp_buggy()</tt> below.
255
-
We will look at the reader's code later, but in the meantime, just think of
256
-
the reader as locklessly picking up the <tt>gp</tt> pointer,
257
-
and, if the value loaded is non-<tt>NULL</tt>, locklessly accessing the
258
-
<tt>->a</tt> and <tt>->b</tt> fields.
259
-
260
-
<blockquote>
261
-
<pre>
262
-
1 bool add_gp_buggy(int a, int b)
263
-
2 {
264
-
3 p = kmalloc(sizeof(*p), GFP_KERNEL);
265
-
4 if (!p)
266
-
5 return -ENOMEM;
267
-
6 spin_lock(&gp_lock);
268
-
7 if (rcu_access_pointer(gp)) {
269
-
8 spin_unlock(&gp_lock);
270
-
9 return false;
271
-
10 }
272
-
11 p->a = a;
273
-
12 p->b = a;
274
-
13 gp = p; /* ORDERING BUG */
275
-
14 spin_unlock(&gp_lock);
276
-
15 return true;
277
-
16 }
278
-
</pre>
279
-
</blockquote>
280
-
281
-
<p>
282
-
The problem is that both the compiler and weakly ordered CPUs are within
283
-
their rights to reorder this code as follows:
284
-
285
-
<blockquote>
286
-
<pre>
287
-
1 bool add_gp_buggy_optimized(int a, int b)
288
-
2 {
289
-
3 p = kmalloc(sizeof(*p), GFP_KERNEL);
290
-
4 if (!p)
291
-
5 return -ENOMEM;
292
-
6 spin_lock(&gp_lock);
293
-
7 if (rcu_access_pointer(gp)) {
294
-
8 spin_unlock(&gp_lock);
295
-
9 return false;
296
-
10 }
297
-
<b>11 gp = p; /* ORDERING BUG */
298
-
12 p->a = a;
299
-
13 p->b = a;</b>
300
-
14 spin_unlock(&gp_lock);
301
-
15 return true;
302
-
16 }
303
-
</pre>
304
-
</blockquote>
305
-
306
-
<p>
307
-
If an RCU reader fetches <tt>gp</tt> just after
308
-
<tt>add_gp_buggy_optimized</tt> executes line 11,
309
-
it will see garbage in the <tt>->a</tt> and <tt>->b</tt>
310
-
fields.
311
-
And this is but one of many ways in which compiler and hardware optimizations
312
-
could cause trouble.
313
-
Therefore, we clearly need some way to prevent the compiler and the CPU from
314
-
reordering in this manner, which brings us to the publish-subscribe
315
-
guarantee discussed in the next section.
316
-
317
-
<h3><a name="Publish-Subscribe Guarantee">Publish/Subscribe Guarantee</a></h3>
318
-
319
-
<p>
320
-
RCU's publish-subscribe guarantee allows data to be inserted
321
-
into a linked data structure without disrupting RCU readers.
322
-
The updater uses <tt>rcu_assign_pointer()</tt> to insert the
323
-
new data, and readers use <tt>rcu_dereference()</tt> to
324
-
access data, whether new or old.
325
-
The following shows an example of insertion:
326
-
327
-
<blockquote>
328
-
<pre>
329
-
1 bool add_gp(int a, int b)
330
-
2 {
331
-
3 p = kmalloc(sizeof(*p), GFP_KERNEL);
332
-
4 if (!p)
333
-
5 return -ENOMEM;
334
-
6 spin_lock(&gp_lock);
335
-
7 if (rcu_access_pointer(gp)) {
336
-
8 spin_unlock(&gp_lock);
337
-
9 return false;
338
-
10 }
339
-
11 p->a = a;
340
-
12 p->b = a;
341
-
13 rcu_assign_pointer(gp, p);
342
-
14 spin_unlock(&gp_lock);
343
-
15 return true;
344
-
16 }
345
-
</pre>
346
-
</blockquote>
347
-
348
-
<p>
349
-
The <tt>rcu_assign_pointer()</tt> on line 13 is conceptually
350
-
equivalent to a simple assignment statement, but also guarantees
351
-
that its assignment will
352
-
happen after the two assignments in lines 11 and 12,
353
-
similar to the C11 <tt>memory_order_release</tt> store operation.
354
-
It also prevents any number of “interesting” compiler
355
-
optimizations, for example, the use of <tt>gp</tt> as a scratch
356
-
location immediately preceding the assignment.
357
-
358
-
<table>
359
-
<tr><th> </th></tr>
360
-
<tr><th align="left">Quick Quiz:</th></tr>
361
-
<tr><td>
362
-
But <tt>rcu_assign_pointer()</tt> does nothing to prevent the
363
-
two assignments to <tt>p->a</tt> and <tt>p->b</tt>
364
-
from being reordered.
365
-
Can't that also cause problems?
366
-
</td></tr>
367
-
<tr><th align="left">Answer:</th></tr>
368
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
369
-
No, it cannot.
370
-
The readers cannot see either of these two fields until
371
-
the assignment to <tt>gp</tt>, by which time both fields are
372
-
fully initialized.
373
-
So reordering the assignments
374
-
to <tt>p->a</tt> and <tt>p->b</tt> cannot possibly
375
-
cause any problems.
376
-
</font></td></tr>
377
-
<tr><td> </td></tr>
378
-
</table>
379
-
380
-
<p>
381
-
It is tempting to assume that the reader need not do anything special
382
-
to control its accesses to the RCU-protected data,
383
-
as shown in <tt>do_something_gp_buggy()</tt> below:
384
-
385
-
<blockquote>
386
-
<pre>
387
-
1 bool do_something_gp_buggy(void)
388
-
2 {
389
-
3 rcu_read_lock();
390
-
4 p = gp; /* OPTIMIZATIONS GALORE!!! */
391
-
5 if (p) {
392
-
6 do_something(p->a, p->b);
393
-
7 rcu_read_unlock();
394
-
8 return true;
395
-
9 }
396
-
10 rcu_read_unlock();
397
-
11 return false;
398
-
12 }
399
-
</pre>
400
-
</blockquote>
401
-
402
-
<p>
403
-
However, this temptation must be resisted because there are a
404
-
surprisingly large number of ways that the compiler
405
-
(to say nothing of
406
-
<a href="https://h71000.www7.hp.com/wizard/wiz_2637.html">DEC Alpha CPUs</a>)
407
-
can trip this code up.
408
-
For but one example, if the compiler were short of registers, it
409
-
might choose to refetch from <tt>gp</tt> rather than keeping
410
-
a separate copy in <tt>p</tt> as follows:
411
-
412
-
<blockquote>
413
-
<pre>
414
-
1 bool do_something_gp_buggy_optimized(void)
415
-
2 {
416
-
3 rcu_read_lock();
417
-
4 if (gp) { /* OPTIMIZATIONS GALORE!!! */
418
-
<b> 5 do_something(gp->a, gp->b);</b>
419
-
6 rcu_read_unlock();
420
-
7 return true;
421
-
8 }
422
-
9 rcu_read_unlock();
423
-
10 return false;
424
-
11 }
425
-
</pre>
426
-
</blockquote>
427
-
428
-
<p>
429
-
If this function ran concurrently with a series of updates that
430
-
replaced the current structure with a new one,
431
-
the fetches of <tt>gp->a</tt>
432
-
and <tt>gp->b</tt> might well come from two different structures,
433
-
which could cause serious confusion.
434
-
To prevent this (and much else besides), <tt>do_something_gp()</tt> uses
435
-
<tt>rcu_dereference()</tt> to fetch from <tt>gp</tt>:
436
-
437
-
<blockquote>
438
-
<pre>
439
-
1 bool do_something_gp(void)
440
-
2 {
441
-
3 rcu_read_lock();
442
-
4 p = rcu_dereference(gp);
443
-
5 if (p) {
444
-
6 do_something(p->a, p->b);
445
-
7 rcu_read_unlock();
446
-
8 return true;
447
-
9 }
448
-
10 rcu_read_unlock();
449
-
11 return false;
450
-
12 }
451
-
</pre>
452
-
</blockquote>
453
-
454
-
<p>
455
-
The <tt>rcu_dereference()</tt> uses volatile casts and (for DEC Alpha)
456
-
memory barriers in the Linux kernel.
457
-
Should a
458
-
<a href="http://www.rdrop.com/users/paulmck/RCU/consume.2015.07.13a.pdf">high-quality implementation of C11 <tt>memory_order_consume</tt> [PDF]</a>
459
-
ever appear, then <tt>rcu_dereference()</tt> could be implemented
460
-
as a <tt>memory_order_consume</tt> load.
461
-
Regardless of the exact implementation, a pointer fetched by
462
-
<tt>rcu_dereference()</tt> may not be used outside of the
463
-
outermost RCU read-side critical section containing that
464
-
<tt>rcu_dereference()</tt>, unless protection of
465
-
the corresponding data element has been passed from RCU to some
466
-
other synchronization mechanism, most commonly locking or
467
-
<a href="https://www.kernel.org/doc/Documentation/RCU/rcuref.txt">reference counting</a>.
468
-
469
-
<p>
470
-
In short, updaters use <tt>rcu_assign_pointer()</tt> and readers
471
-
use <tt>rcu_dereference()</tt>, and these two RCU API elements
472
-
work together to ensure that readers have a consistent view of
473
-
newly added data elements.
474
-
475
-
<p>
476
-
Of course, it is also necessary to remove elements from RCU-protected
477
-
data structures, for example, using the following process:
478
-
479
-
<ol>
480
-
<li> Remove the data element from the enclosing structure.
481
-
<li> Wait for all pre-existing RCU read-side critical sections
482
-
to complete (because only pre-existing readers can possibly have
483
-
a reference to the newly removed data element).
484
-
<li> At this point, only the updater has a reference to the
485
-
newly removed data element, so it can safely reclaim
486
-
the data element, for example, by passing it to <tt>kfree()</tt>.
487
-
</ol>
488
-
489
-
This process is implemented by <tt>remove_gp_synchronous()</tt>:
490
-
491
-
<blockquote>
492
-
<pre>
493
-
1 bool remove_gp_synchronous(void)
494
-
2 {
495
-
3 struct foo *p;
496
-
4
497
-
5 spin_lock(&gp_lock);
498
-
6 p = rcu_access_pointer(gp);
499
-
7 if (!p) {
500
-
8 spin_unlock(&gp_lock);
501
-
9 return false;
502
-
10 }
503
-
11 rcu_assign_pointer(gp, NULL);
504
-
12 spin_unlock(&gp_lock);
505
-
13 synchronize_rcu();
506
-
14 kfree(p);
507
-
15 return true;
508
-
16 }
509
-
</pre>
510
-
</blockquote>
511
-
512
-
<p>
513
-
This function is straightforward, with line 13 waiting for a grace
514
-
period before line 14 frees the old data element.
515
-
This waiting ensures that readers will reach line 7 of
516
-
<tt>do_something_gp()</tt> before the data element referenced by
517
-
<tt>p</tt> is freed.
518
-
The <tt>rcu_access_pointer()</tt> on line 6 is similar to
519
-
<tt>rcu_dereference()</tt>, except that:
520
-
521
-
<ol>
522
-
<li> The value returned by <tt>rcu_access_pointer()</tt>
523
-
cannot be dereferenced.
524
-
If you want to access the value pointed to as well as
525
-
the pointer itself, use <tt>rcu_dereference()</tt>
526
-
instead of <tt>rcu_access_pointer()</tt>.
527
-
<li> The call to <tt>rcu_access_pointer()</tt> need not be
528
-
protected.
529
-
In contrast, <tt>rcu_dereference()</tt> must either be
530
-
within an RCU read-side critical section or in a code
531
-
segment where the pointer cannot change, for example, in
532
-
code protected by the corresponding update-side lock.
533
-
</ol>
534
-
535
-
<table>
536
-
<tr><th> </th></tr>
537
-
<tr><th align="left">Quick Quiz:</th></tr>
538
-
<tr><td>
539
-
Without the <tt>rcu_dereference()</tt> or the
540
-
<tt>rcu_access_pointer()</tt>, what destructive optimizations
541
-
might the compiler make use of?
542
-
</td></tr>
543
-
<tr><th align="left">Answer:</th></tr>
544
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
545
-
Let's start with what happens to <tt>do_something_gp()</tt>
546
-
if it fails to use <tt>rcu_dereference()</tt>.
547
-
It could reuse a value formerly fetched from this same pointer.
548
-
It could also fetch the pointer from <tt>gp</tt> in a byte-at-a-time
549
-
manner, resulting in <i>load tearing</i>, in turn resulting a bytewise
550
-
mash-up of two distinct pointer values.
551
-
It might even use value-speculation optimizations, where it makes
552
-
a wrong guess, but by the time it gets around to checking the
553
-
value, an update has changed the pointer to match the wrong guess.
554
-
Too bad about any dereferences that returned pre-initialization garbage
555
-
in the meantime!
556
-
</font>
557
-
558
-
<p><font color="ffffff">
559
-
For <tt>remove_gp_synchronous()</tt>, as long as all modifications
560
-
to <tt>gp</tt> are carried out while holding <tt>gp_lock</tt>,
561
-
the above optimizations are harmless.
562
-
However, <tt>sparse</tt> will complain if you
563
-
define <tt>gp</tt> with <tt>__rcu</tt> and then
564
-
access it without using
565
-
either <tt>rcu_access_pointer()</tt> or <tt>rcu_dereference()</tt>.
566
-
</font></td></tr>
567
-
<tr><td> </td></tr>
568
-
</table>
569
-
570
-
<p>
571
-
In short, RCU's publish-subscribe guarantee is provided by the combination
572
-
of <tt>rcu_assign_pointer()</tt> and <tt>rcu_dereference()</tt>.
573
-
This guarantee allows data elements to be safely added to RCU-protected
574
-
linked data structures without disrupting RCU readers.
575
-
This guarantee can be used in combination with the grace-period
576
-
guarantee to also allow data elements to be removed from RCU-protected
577
-
linked data structures, again without disrupting RCU readers.
578
-
579
-
<p>
580
-
This guarantee was only partially premeditated.
581
-
DYNIX/ptx used an explicit memory barrier for publication, but had nothing
582
-
resembling <tt>rcu_dereference()</tt> for subscription, nor did it
583
-
have anything resembling the <tt>smp_read_barrier_depends()</tt>
584
-
that was later subsumed into <tt>rcu_dereference()</tt> and later
585
-
still into <tt>READ_ONCE()</tt>.
586
-
The need for these operations made itself known quite suddenly at a
587
-
late-1990s meeting with the DEC Alpha architects, back in the days when
588
-
DEC was still a free-standing company.
589
-
It took the Alpha architects a good hour to convince me that any sort
590
-
of barrier would ever be needed, and it then took me a good <i>two</i> hours
591
-
to convince them that their documentation did not make this point clear.
592
-
More recent work with the C and C++ standards committees have provided
593
-
much education on tricks and traps from the compiler.
594
-
In short, compilers were much less tricky in the early 1990s, but in
595
-
2015, don't even think about omitting <tt>rcu_dereference()</tt>!
596
-
597
-
<h3><a name="Memory-Barrier Guarantees">Memory-Barrier Guarantees</a></h3>
598
-
599
-
<p>
600
-
The previous section's simple linked-data-structure scenario clearly
601
-
demonstrates the need for RCU's stringent memory-ordering guarantees on
602
-
systems with more than one CPU:
603
-
604
-
<ol>
605
-
<li> Each CPU that has an RCU read-side critical section that
606
-
begins before <tt>synchronize_rcu()</tt> starts is
607
-
guaranteed to execute a full memory barrier between the time
608
-
that the RCU read-side critical section ends and the time that
609
-
<tt>synchronize_rcu()</tt> returns.
610
-
Without this guarantee, a pre-existing RCU read-side critical section
611
-
might hold a reference to the newly removed <tt>struct foo</tt>
612
-
after the <tt>kfree()</tt> on line 14 of
613
-
<tt>remove_gp_synchronous()</tt>.
614
-
<li> Each CPU that has an RCU read-side critical section that ends
615
-
after <tt>synchronize_rcu()</tt> returns is guaranteed
616
-
to execute a full memory barrier between the time that
617
-
<tt>synchronize_rcu()</tt> begins and the time that the RCU
618
-
read-side critical section begins.
619
-
Without this guarantee, a later RCU read-side critical section
620
-
running after the <tt>kfree()</tt> on line 14 of
621
-
<tt>remove_gp_synchronous()</tt> might
622
-
later run <tt>do_something_gp()</tt> and find the
623
-
newly deleted <tt>struct foo</tt>.
624
-
<li> If the task invoking <tt>synchronize_rcu()</tt> remains
625
-
on a given CPU, then that CPU is guaranteed to execute a full
626
-
memory barrier sometime during the execution of
627
-
<tt>synchronize_rcu()</tt>.
628
-
This guarantee ensures that the <tt>kfree()</tt> on
629
-
line 14 of <tt>remove_gp_synchronous()</tt> really does
630
-
execute after the removal on line 11.
631
-
<li> If the task invoking <tt>synchronize_rcu()</tt> migrates
632
-
among a group of CPUs during that invocation, then each of the
633
-
CPUs in that group is guaranteed to execute a full memory barrier
634
-
sometime during the execution of <tt>synchronize_rcu()</tt>.
635
-
This guarantee also ensures that the <tt>kfree()</tt> on
636
-
line 14 of <tt>remove_gp_synchronous()</tt> really does
637
-
execute after the removal on
638
-
line 11, but also in the case where the thread executing the
639
-
<tt>synchronize_rcu()</tt> migrates in the meantime.
640
-
</ol>
641
-
642
-
<table>
643
-
<tr><th> </th></tr>
644
-
<tr><th align="left">Quick Quiz:</th></tr>
645
-
<tr><td>
646
-
Given that multiple CPUs can start RCU read-side critical sections
647
-
at any time without any ordering whatsoever, how can RCU possibly
648
-
tell whether or not a given RCU read-side critical section starts
649
-
before a given instance of <tt>synchronize_rcu()</tt>?
650
-
</td></tr>
651
-
<tr><th align="left">Answer:</th></tr>
652
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
653
-
If RCU cannot tell whether or not a given
654
-
RCU read-side critical section starts before a
655
-
given instance of <tt>synchronize_rcu()</tt>,
656
-
then it must assume that the RCU read-side critical section
657
-
started first.
658
-
In other words, a given instance of <tt>synchronize_rcu()</tt>
659
-
can avoid waiting on a given RCU read-side critical section only
660
-
if it can prove that <tt>synchronize_rcu()</tt> started first.
661
-
</font>
662
-
663
-
<p><font color="ffffff">
664
-
A related question is “When <tt>rcu_read_lock()</tt>
665
-
doesn't generate any code, why does it matter how it relates
666
-
to a grace period?”
667
-
The answer is that it is not the relationship of
668
-
<tt>rcu_read_lock()</tt> itself that is important, but rather
669
-
the relationship of the code within the enclosed RCU read-side
670
-
critical section to the code preceding and following the
671
-
grace period.
672
-
If we take this viewpoint, then a given RCU read-side critical
673
-
section begins before a given grace period when some access
674
-
preceding the grace period observes the effect of some access
675
-
within the critical section, in which case none of the accesses
676
-
within the critical section may observe the effects of any
677
-
access following the grace period.
678
-
</font>
679
-
680
-
<p><font color="ffffff">
681
-
As of late 2016, mathematical models of RCU take this
682
-
viewpoint, for example, see slides 62 and 63
683
-
of the
684
-
<a href="http://www2.rdrop.com/users/paulmck/scalability/paper/LinuxMM.2016.10.04c.LCE.pdf">2016 LinuxCon EU</a>
685
-
presentation.
686
-
</font></td></tr>
687
-
<tr><td> </td></tr>
688
-
</table>
689
-
690
-
<table>
691
-
<tr><th> </th></tr>
692
-
<tr><th align="left">Quick Quiz:</th></tr>
693
-
<tr><td>
694
-
The first and second guarantees require unbelievably strict ordering!
695
-
Are all these memory barriers <i> really</i> required?
696
-
</td></tr>
697
-
<tr><th align="left">Answer:</th></tr>
698
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
699
-
Yes, they really are required.
700
-
To see why the first guarantee is required, consider the following
701
-
sequence of events:
702
-
</font>
703
-
704
-
<ol>
705
-
<li> <font color="ffffff">
706
-
CPU 1: <tt>rcu_read_lock()</tt>
707
-
</font>
708
-
<li> <font color="ffffff">
709
-
CPU 1: <tt>q = rcu_dereference(gp);
710
-
/* Very likely to return p. */</tt>
711
-
</font>
712
-
<li> <font color="ffffff">
713
-
CPU 0: <tt>list_del_rcu(p);</tt>
714
-
</font>
715
-
<li> <font color="ffffff">
716
-
CPU 0: <tt>synchronize_rcu()</tt> starts.
717
-
</font>
718
-
<li> <font color="ffffff">
719
-
CPU 1: <tt>do_something_with(q->a);
720
-
/* No smp_mb(), so might happen after kfree(). */</tt>
721
-
</font>
722
-
<li> <font color="ffffff">
723
-
CPU 1: <tt>rcu_read_unlock()</tt>
724
-
</font>
725
-
<li> <font color="ffffff">
726
-
CPU 0: <tt>synchronize_rcu()</tt> returns.
727
-
</font>
728
-
<li> <font color="ffffff">
729
-
CPU 0: <tt>kfree(p);</tt>
730
-
</font>
731
-
</ol>
732
-
733
-
<p><font color="ffffff">
734
-
Therefore, there absolutely must be a full memory barrier between the
735
-
end of the RCU read-side critical section and the end of the
736
-
grace period.
737
-
</font>
738
-
739
-
<p><font color="ffffff">
740
-
The sequence of events demonstrating the necessity of the second rule
741
-
is roughly similar:
742
-
</font>
743
-
744
-
<ol>
745
-
<li> <font color="ffffff">CPU 0: <tt>list_del_rcu(p);</tt>
746
-
</font>
747
-
<li> <font color="ffffff">CPU 0: <tt>synchronize_rcu()</tt> starts.
748
-
</font>
749
-
<li> <font color="ffffff">CPU 1: <tt>rcu_read_lock()</tt>
750
-
</font>
751
-
<li> <font color="ffffff">CPU 1: <tt>q = rcu_dereference(gp);
752
-
/* Might return p if no memory barrier. */</tt>
753
-
</font>
754
-
<li> <font color="ffffff">CPU 0: <tt>synchronize_rcu()</tt> returns.
755
-
</font>
756
-
<li> <font color="ffffff">CPU 0: <tt>kfree(p);</tt>
757
-
</font>
758
-
<li> <font color="ffffff">
759
-
CPU 1: <tt>do_something_with(q->a); /* Boom!!! */</tt>
760
-
</font>
761
-
<li> <font color="ffffff">CPU 1: <tt>rcu_read_unlock()</tt>
762
-
</font>
763
-
</ol>
764
-
765
-
<p><font color="ffffff">
766
-
And similarly, without a memory barrier between the beginning of the
767
-
grace period and the beginning of the RCU read-side critical section,
768
-
CPU 1 might end up accessing the freelist.
769
-
</font>
770
-
771
-
<p><font color="ffffff">
772
-
The “as if” rule of course applies, so that any
773
-
implementation that acts as if the appropriate memory barriers
774
-
were in place is a correct implementation.
775
-
That said, it is much easier to fool yourself into believing
776
-
that you have adhered to the as-if rule than it is to actually
777
-
adhere to it!
778
-
</font></td></tr>
779
-
<tr><td> </td></tr>
780
-
</table>
781
-
782
-
<table>
783
-
<tr><th> </th></tr>
784
-
<tr><th align="left">Quick Quiz:</th></tr>
785
-
<tr><td>
786
-
You claim that <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>
787
-
generate absolutely no code in some kernel builds.
788
-
This means that the compiler might arbitrarily rearrange consecutive
789
-
RCU read-side critical sections.
790
-
Given such rearrangement, if a given RCU read-side critical section
791
-
is done, how can you be sure that all prior RCU read-side critical
792
-
sections are done?
793
-
Won't the compiler rearrangements make that impossible to determine?
794
-
</td></tr>
795
-
<tr><th align="left">Answer:</th></tr>
796
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
797
-
In cases where <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>
798
-
generate absolutely no code, RCU infers quiescent states only at
799
-
special locations, for example, within the scheduler.
800
-
Because calls to <tt>schedule()</tt> had better prevent calling-code
801
-
accesses to shared variables from being rearranged across the call to
802
-
<tt>schedule()</tt>, if RCU detects the end of a given RCU read-side
803
-
critical section, it will necessarily detect the end of all prior
804
-
RCU read-side critical sections, no matter how aggressively the
805
-
compiler scrambles the code.
806
-
</font>
807
-
808
-
<p><font color="ffffff">
809
-
Again, this all assumes that the compiler cannot scramble code across
810
-
calls to the scheduler, out of interrupt handlers, into the idle loop,
811
-
into user-mode code, and so on.
812
-
But if your kernel build allows that sort of scrambling, you have broken
813
-
far more than just RCU!
814
-
</font></td></tr>
815
-
<tr><td> </td></tr>
816
-
</table>
817
-
818
-
<p>
819
-
Note that these memory-barrier requirements do not replace the fundamental
820
-
RCU requirement that a grace period wait for all pre-existing readers.
821
-
On the contrary, the memory barriers called out in this section must operate in
822
-
such a way as to <i>enforce</i> this fundamental requirement.
823
-
Of course, different implementations enforce this requirement in different
824
-
ways, but enforce it they must.
825
-
826
-
<h3><a name="RCU Primitives Guaranteed to Execute Unconditionally">RCU Primitives Guaranteed to Execute Unconditionally</a></h3>
827
-
828
-
<p>
829
-
The common-case RCU primitives are unconditional.
830
-
They are invoked, they do their job, and they return, with no possibility
831
-
of error, and no need to retry.
832
-
This is a key RCU design philosophy.
833
-
834
-
<p>
835
-
However, this philosophy is pragmatic rather than pigheaded.
836
-
If someone comes up with a good justification for a particular conditional
837
-
RCU primitive, it might well be implemented and added.
838
-
After all, this guarantee was reverse-engineered, not premeditated.
839
-
The unconditional nature of the RCU primitives was initially an
840
-
accident of implementation, and later experience with synchronization
841
-
primitives with conditional primitives caused me to elevate this
842
-
accident to a guarantee.
843
-
Therefore, the justification for adding a conditional primitive to
844
-
RCU would need to be based on detailed and compelling use cases.
845
-
846
-
<h3><a name="Guaranteed Read-to-Write Upgrade">Guaranteed Read-to-Write Upgrade</a></h3>
847
-
848
-
<p>
849
-
As far as RCU is concerned, it is always possible to carry out an
850
-
update within an RCU read-side critical section.
851
-
For example, that RCU read-side critical section might search for
852
-
a given data element, and then might acquire the update-side
853
-
spinlock in order to update that element, all while remaining
854
-
in that RCU read-side critical section.
855
-
Of course, it is necessary to exit the RCU read-side critical section
856
-
before invoking <tt>synchronize_rcu()</tt>, however, this
857
-
inconvenience can be avoided through use of the
858
-
<tt>call_rcu()</tt> and <tt>kfree_rcu()</tt> API members
859
-
described later in this document.
860
-
861
-
<table>
862
-
<tr><th> </th></tr>
863
-
<tr><th align="left">Quick Quiz:</th></tr>
864
-
<tr><td>
865
-
But how does the upgrade-to-write operation exclude other readers?
866
-
</td></tr>
867
-
<tr><th align="left">Answer:</th></tr>
868
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
869
-
It doesn't, just like normal RCU updates, which also do not exclude
870
-
RCU readers.
871
-
</font></td></tr>
872
-
<tr><td> </td></tr>
873
-
</table>
874
-
875
-
<p>
876
-
This guarantee allows lookup code to be shared between read-side
877
-
and update-side code, and was premeditated, appearing in the earliest
878
-
DYNIX/ptx RCU documentation.
879
-
880
-
<h2><a name="Fundamental Non-Requirements">Fundamental Non-Requirements</a></h2>
881
-
882
-
<p>
883
-
RCU provides extremely lightweight readers, and its read-side guarantees,
884
-
though quite useful, are correspondingly lightweight.
885
-
It is therefore all too easy to assume that RCU is guaranteeing more
886
-
than it really is.
887
-
Of course, the list of things that RCU does not guarantee is infinitely
888
-
long, however, the following sections list a few non-guarantees that
889
-
have caused confusion.
890
-
Except where otherwise noted, these non-guarantees were premeditated.
891
-
892
-
<ol>
893
-
<li> <a href="#Readers Impose Minimal Ordering">
894
-
Readers Impose Minimal Ordering</a>
895
-
<li> <a href="#Readers Do Not Exclude Updaters">
896
-
Readers Do Not Exclude Updaters</a>
897
-
<li> <a href="#Updaters Only Wait For Old Readers">
898
-
Updaters Only Wait For Old Readers</a>
899
-
<li> <a href="#Grace Periods Don't Partition Read-Side Critical Sections">
900
-
Grace Periods Don't Partition Read-Side Critical Sections</a>
901
-
<li> <a href="#Read-Side Critical Sections Don't Partition Grace Periods">
902
-
Read-Side Critical Sections Don't Partition Grace Periods</a>
903
-
</ol>
904
-
905
-
<h3><a name="Readers Impose Minimal Ordering">Readers Impose Minimal Ordering</a></h3>
906
-
907
-
<p>
908
-
Reader-side markers such as <tt>rcu_read_lock()</tt> and
909
-
<tt>rcu_read_unlock()</tt> provide absolutely no ordering guarantees
910
-
except through their interaction with the grace-period APIs such as
911
-
<tt>synchronize_rcu()</tt>.
912
-
To see this, consider the following pair of threads:
913
-
914
-
<blockquote>
915
-
<pre>
916
-
1 void thread0(void)
917
-
2 {
918
-
3 rcu_read_lock();
919
-
4 WRITE_ONCE(x, 1);
920
-
5 rcu_read_unlock();
921
-
6 rcu_read_lock();
922
-
7 WRITE_ONCE(y, 1);
923
-
8 rcu_read_unlock();
924
-
9 }
925
-
10
926
-
11 void thread1(void)
927
-
12 {
928
-
13 rcu_read_lock();
929
-
14 r1 = READ_ONCE(y);
930
-
15 rcu_read_unlock();
931
-
16 rcu_read_lock();
932
-
17 r2 = READ_ONCE(x);
933
-
18 rcu_read_unlock();
934
-
19 }
935
-
</pre>
936
-
</blockquote>
937
-
938
-
<p>
939
-
After <tt>thread0()</tt> and <tt>thread1()</tt> execute
940
-
concurrently, it is quite possible to have
941
-
942
-
<blockquote>
943
-
<pre>
944
-
(r1 == 1 && r2 == 0)
945
-
</pre>
946
-
</blockquote>
947
-
948
-
(that is, <tt>y</tt> appears to have been assigned before <tt>x</tt>),
949
-
which would not be possible if <tt>rcu_read_lock()</tt> and
950
-
<tt>rcu_read_unlock()</tt> had much in the way of ordering
951
-
properties.
952
-
But they do not, so the CPU is within its rights
953
-
to do significant reordering.
954
-
This is by design: Any significant ordering constraints would slow down
955
-
these fast-path APIs.
956
-
957
-
<table>
958
-
<tr><th> </th></tr>
959
-
<tr><th align="left">Quick Quiz:</th></tr>
960
-
<tr><td>
961
-
Can't the compiler also reorder this code?
962
-
</td></tr>
963
-
<tr><th align="left">Answer:</th></tr>
964
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
965
-
No, the volatile casts in <tt>READ_ONCE()</tt> and
966
-
<tt>WRITE_ONCE()</tt> prevent the compiler from reordering in
967
-
this particular case.
968
-
</font></td></tr>
969
-
<tr><td> </td></tr>
970
-
</table>
971
-
972
-
<h3><a name="Readers Do Not Exclude Updaters">Readers Do Not Exclude Updaters</a></h3>
973
-
974
-
<p>
975
-
Neither <tt>rcu_read_lock()</tt> nor <tt>rcu_read_unlock()</tt>
976
-
exclude updates.
977
-
All they do is to prevent grace periods from ending.
978
-
The following example illustrates this:
979
-
980
-
<blockquote>
981
-
<pre>
982
-
1 void thread0(void)
983
-
2 {
984
-
3 rcu_read_lock();
985
-
4 r1 = READ_ONCE(y);
986
-
5 if (r1) {
987
-
6 do_something_with_nonzero_x();
988
-
7 r2 = READ_ONCE(x);
989
-
8 WARN_ON(!r2); /* BUG!!! */
990
-
9 }
991
-
10 rcu_read_unlock();
992
-
11 }
993
-
12
994
-
13 void thread1(void)
995
-
14 {
996
-
15 spin_lock(&my_lock);
997
-
16 WRITE_ONCE(x, 1);
998
-
17 WRITE_ONCE(y, 1);
999
-
18 spin_unlock(&my_lock);
1000
-
19 }
1001
-
</pre>
1002
-
</blockquote>
1003
-
1004
-
<p>
1005
-
If the <tt>thread0()</tt> function's <tt>rcu_read_lock()</tt>
1006
-
excluded the <tt>thread1()</tt> function's update,
1007
-
the <tt>WARN_ON()</tt> could never fire.
1008
-
But the fact is that <tt>rcu_read_lock()</tt> does not exclude
1009
-
much of anything aside from subsequent grace periods, of which
1010
-
<tt>thread1()</tt> has none, so the
1011
-
<tt>WARN_ON()</tt> can and does fire.
1012
-
1013
-
<h3><a name="Updaters Only Wait For Old Readers">Updaters Only Wait For Old Readers</a></h3>
1014
-
1015
-
<p>
1016
-
It might be tempting to assume that after <tt>synchronize_rcu()</tt>
1017
-
completes, there are no readers executing.
1018
-
This temptation must be avoided because
1019
-
new readers can start immediately after <tt>synchronize_rcu()</tt>
1020
-
starts, and <tt>synchronize_rcu()</tt> is under no
1021
-
obligation to wait for these new readers.
1022
-
1023
-
<table>
1024
-
<tr><th> </th></tr>
1025
-
<tr><th align="left">Quick Quiz:</th></tr>
1026
-
<tr><td>
1027
-
Suppose that synchronize_rcu() did wait until <i>all</i>
1028
-
readers had completed instead of waiting only on
1029
-
pre-existing readers.
1030
-
For how long would the updater be able to rely on there
1031
-
being no readers?
1032
-
</td></tr>
1033
-
<tr><th align="left">Answer:</th></tr>
1034
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
1035
-
For no time at all.
1036
-
Even if <tt>synchronize_rcu()</tt> were to wait until
1037
-
all readers had completed, a new reader might start immediately after
1038
-
<tt>synchronize_rcu()</tt> completed.
1039
-
Therefore, the code following
1040
-
<tt>synchronize_rcu()</tt> can <i>never</i> rely on there being
1041
-
no readers.
1042
-
</font></td></tr>
1043
-
<tr><td> </td></tr>
1044
-
</table>
1045
-
1046
-
<h3><a name="Grace Periods Don't Partition Read-Side Critical Sections">
1047
-
Grace Periods Don't Partition Read-Side Critical Sections</a></h3>
1048
-
1049
-
<p>
1050
-
It is tempting to assume that if any part of one RCU read-side critical
1051
-
section precedes a given grace period, and if any part of another RCU
1052
-
read-side critical section follows that same grace period, then all of
1053
-
the first RCU read-side critical section must precede all of the second.
1054
-
However, this just isn't the case: A single grace period does not
1055
-
partition the set of RCU read-side critical sections.
1056
-
An example of this situation can be illustrated as follows, where
1057
-
<tt>x</tt>, <tt>y</tt>, and <tt>z</tt> are initially all zero:
1058
-
1059
-
<blockquote>
1060
-
<pre>
1061
-
1 void thread0(void)
1062
-
2 {
1063
-
3 rcu_read_lock();
1064
-
4 WRITE_ONCE(a, 1);
1065
-
5 WRITE_ONCE(b, 1);
1066
-
6 rcu_read_unlock();
1067
-
7 }
1068
-
8
1069
-
9 void thread1(void)
1070
-
10 {
1071
-
11 r1 = READ_ONCE(a);
1072
-
12 synchronize_rcu();
1073
-
13 WRITE_ONCE(c, 1);
1074
-
14 }
1075
-
15
1076
-
16 void thread2(void)
1077
-
17 {
1078
-
18 rcu_read_lock();
1079
-
19 r2 = READ_ONCE(b);
1080
-
20 r3 = READ_ONCE(c);
1081
-
21 rcu_read_unlock();
1082
-
22 }
1083
-
</pre>
1084
-
</blockquote>
1085
-
1086
-
<p>
1087
-
It turns out that the outcome:
1088
-
1089
-
<blockquote>
1090
-
<pre>
1091
-
(r1 == 1 && r2 == 0 && r3 == 1)
1092
-
</pre>
1093
-
</blockquote>
1094
-
1095
-
is entirely possible.
1096
-
The following figure show how this can happen, with each circled
1097
-
<tt>QS</tt> indicating the point at which RCU recorded a
1098
-
<i>quiescent state</i> for each thread, that is, a state in which
1099
-
RCU knows that the thread cannot be in the midst of an RCU read-side
1100
-
critical section that started before the current grace period:
1101
-
1102
-
<p><img src="GPpartitionReaders1.svg" alt="GPpartitionReaders1.svg" width="60%"></p>
1103
-
1104
-
<p>
1105
-
If it is necessary to partition RCU read-side critical sections in this
1106
-
manner, it is necessary to use two grace periods, where the first
1107
-
grace period is known to end before the second grace period starts:
1108
-
1109
-
<blockquote>
1110
-
<pre>
1111
-
1 void thread0(void)
1112
-
2 {
1113
-
3 rcu_read_lock();
1114
-
4 WRITE_ONCE(a, 1);
1115
-
5 WRITE_ONCE(b, 1);
1116
-
6 rcu_read_unlock();
1117
-
7 }
1118
-
8
1119
-
9 void thread1(void)
1120
-
10 {
1121
-
11 r1 = READ_ONCE(a);
1122
-
12 synchronize_rcu();
1123
-
13 WRITE_ONCE(c, 1);
1124
-
14 }
1125
-
15
1126
-
16 void thread2(void)
1127
-
17 {
1128
-
18 r2 = READ_ONCE(c);
1129
-
19 synchronize_rcu();
1130
-
20 WRITE_ONCE(d, 1);
1131
-
21 }
1132
-
22
1133
-
23 void thread3(void)
1134
-
24 {
1135
-
25 rcu_read_lock();
1136
-
26 r3 = READ_ONCE(b);
1137
-
27 r4 = READ_ONCE(d);
1138
-
28 rcu_read_unlock();
1139
-
29 }
1140
-
</pre>
1141
-
</blockquote>
1142
-
1143
-
<p>
1144
-
Here, if <tt>(r1 == 1)</tt>, then
1145
-
<tt>thread0()</tt>'s write to <tt>b</tt> must happen
1146
-
before the end of <tt>thread1()</tt>'s grace period.
1147
-
If in addition <tt>(r4 == 1)</tt>, then
1148
-
<tt>thread3()</tt>'s read from <tt>b</tt> must happen
1149
-
after the beginning of <tt>thread2()</tt>'s grace period.
1150
-
If it is also the case that <tt>(r2 == 1)</tt>, then the
1151
-
end of <tt>thread1()</tt>'s grace period must precede the
1152
-
beginning of <tt>thread2()</tt>'s grace period.
1153
-
This mean that the two RCU read-side critical sections cannot overlap,
1154
-
guaranteeing that <tt>(r3 == 1)</tt>.
1155
-
As a result, the outcome:
1156
-
1157
-
<blockquote>
1158
-
<pre>
1159
-
(r1 == 1 && r2 == 1 && r3 == 0 && r4 == 1)
1160
-
</pre>
1161
-
</blockquote>
1162
-
1163
-
cannot happen.
1164
-
1165
-
<p>
1166
-
This non-requirement was also non-premeditated, but became apparent
1167
-
when studying RCU's interaction with memory ordering.
1168
-
1169
-
<h3><a name="Read-Side Critical Sections Don't Partition Grace Periods">
1170
-
Read-Side Critical Sections Don't Partition Grace Periods</a></h3>
1171
-
1172
-
<p>
1173
-
It is also tempting to assume that if an RCU read-side critical section
1174
-
happens between a pair of grace periods, then those grace periods cannot
1175
-
overlap.
1176
-
However, this temptation leads nowhere good, as can be illustrated by
1177
-
the following, with all variables initially zero:
1178
-
1179
-
<blockquote>
1180
-
<pre>
1181
-
1 void thread0(void)
1182
-
2 {
1183
-
3 rcu_read_lock();
1184
-
4 WRITE_ONCE(a, 1);
1185
-
5 WRITE_ONCE(b, 1);
1186
-
6 rcu_read_unlock();
1187
-
7 }
1188
-
8
1189
-
9 void thread1(void)
1190
-
10 {
1191
-
11 r1 = READ_ONCE(a);
1192
-
12 synchronize_rcu();
1193
-
13 WRITE_ONCE(c, 1);
1194
-
14 }
1195
-
15
1196
-
16 void thread2(void)
1197
-
17 {
1198
-
18 rcu_read_lock();
1199
-
19 WRITE_ONCE(d, 1);
1200
-
20 r2 = READ_ONCE(c);
1201
-
21 rcu_read_unlock();
1202
-
22 }
1203
-
23
1204
-
24 void thread3(void)
1205
-
25 {
1206
-
26 r3 = READ_ONCE(d);
1207
-
27 synchronize_rcu();
1208
-
28 WRITE_ONCE(e, 1);
1209
-
29 }
1210
-
30
1211
-
31 void thread4(void)
1212
-
32 {
1213
-
33 rcu_read_lock();
1214
-
34 r4 = READ_ONCE(b);
1215
-
35 r5 = READ_ONCE(e);
1216
-
36 rcu_read_unlock();
1217
-
37 }
1218
-
</pre>
1219
-
</blockquote>
1220
-
1221
-
<p>
1222
-
In this case, the outcome:
1223
-
1224
-
<blockquote>
1225
-
<pre>
1226
-
(r1 == 1 && r2 == 1 && r3 == 1 && r4 == 0 && r5 == 1)
1227
-
</pre>
1228
-
</blockquote>
1229
-
1230
-
is entirely possible, as illustrated below:
1231
-
1232
-
<p><img src="ReadersPartitionGP1.svg" alt="ReadersPartitionGP1.svg" width="100%"></p>
1233
-
1234
-
<p>
1235
-
Again, an RCU read-side critical section can overlap almost all of a
1236
-
given grace period, just so long as it does not overlap the entire
1237
-
grace period.
1238
-
As a result, an RCU read-side critical section cannot partition a pair
1239
-
of RCU grace periods.
1240
-
1241
-
<table>
1242
-
<tr><th> </th></tr>
1243
-
<tr><th align="left">Quick Quiz:</th></tr>
1244
-
<tr><td>
1245
-
How long a sequence of grace periods, each separated by an RCU
1246
-
read-side critical section, would be required to partition the RCU
1247
-
read-side critical sections at the beginning and end of the chain?
1248
-
</td></tr>
1249
-
<tr><th align="left">Answer:</th></tr>
1250
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
1251
-
In theory, an infinite number.
1252
-
In practice, an unknown number that is sensitive to both implementation
1253
-
details and timing considerations.
1254
-
Therefore, even in practice, RCU users must abide by the
1255
-
theoretical rather than the practical answer.
1256
-
</font></td></tr>
1257
-
<tr><td> </td></tr>
1258
-
</table>
1259
-
1260
-
<h2><a name="Parallelism Facts of Life">Parallelism Facts of Life</a></h2>
1261
-
1262
-
<p>
1263
-
These parallelism facts of life are by no means specific to RCU, but
1264
-
the RCU implementation must abide by them.
1265
-
They therefore bear repeating:
1266
-
1267
-
<ol>
1268
-
<li> Any CPU or task may be delayed at any time,
1269
-
and any attempts to avoid these delays by disabling
1270
-
preemption, interrupts, or whatever are completely futile.
1271
-
This is most obvious in preemptible user-level
1272
-
environments and in virtualized environments (where
1273
-
a given guest OS's VCPUs can be preempted at any time by
1274
-
the underlying hypervisor), but can also happen in bare-metal
1275
-
environments due to ECC errors, NMIs, and other hardware
1276
-
events.
1277
-
Although a delay of more than about 20 seconds can result
1278
-
in splats, the RCU implementation is obligated to use
1279
-
algorithms that can tolerate extremely long delays, but where
1280
-
“extremely long” is not long enough to allow
1281
-
wrap-around when incrementing a 64-bit counter.
1282
-
<li> Both the compiler and the CPU can reorder memory accesses.
1283
-
Where it matters, RCU must use compiler directives and
1284
-
memory-barrier instructions to preserve ordering.
1285
-
<li> Conflicting writes to memory locations in any given cache line
1286
-
will result in expensive cache misses.
1287
-
Greater numbers of concurrent writes and more-frequent
1288
-
concurrent writes will result in more dramatic slowdowns.
1289
-
RCU is therefore obligated to use algorithms that have
1290
-
sufficient locality to avoid significant performance and
1291
-
scalability problems.
1292
-
<li> As a rough rule of thumb, only one CPU's worth of processing
1293
-
may be carried out under the protection of any given exclusive
1294
-
lock.
1295
-
RCU must therefore use scalable locking designs.
1296
-
<li> Counters are finite, especially on 32-bit systems.
1297
-
RCU's use of counters must therefore tolerate counter wrap,
1298
-
or be designed such that counter wrap would take way more
1299
-
time than a single system is likely to run.
1300
-
An uptime of ten years is quite possible, a runtime
1301
-
of a century much less so.
1302
-
As an example of the latter, RCU's dyntick-idle nesting counter
1303
-
allows 54 bits for interrupt nesting level (this counter
1304
-
is 64 bits even on a 32-bit system).
1305
-
Overflowing this counter requires 2<sup>54</sup>
1306
-
half-interrupts on a given CPU without that CPU ever going idle.
1307
-
If a half-interrupt happened every microsecond, it would take
1308
-
570 years of runtime to overflow this counter, which is currently
1309
-
believed to be an acceptably long time.
1310
-
<li> Linux systems can have thousands of CPUs running a single
1311
-
Linux kernel in a single shared-memory environment.
1312
-
RCU must therefore pay close attention to high-end scalability.
1313
-
</ol>
1314
-
1315
-
<p>
1316
-
This last parallelism fact of life means that RCU must pay special
1317
-
attention to the preceding facts of life.
1318
-
The idea that Linux might scale to systems with thousands of CPUs would
1319
-
have been met with some skepticism in the 1990s, but these requirements
1320
-
would have otherwise have been unsurprising, even in the early 1990s.
1321
-
1322
-
<h2><a name="Quality-of-Implementation Requirements">Quality-of-Implementation Requirements</a></h2>
1323
-
1324
-
<p>
1325
-
These sections list quality-of-implementation requirements.
1326
-
Although an RCU implementation that ignores these requirements could
1327
-
still be used, it would likely be subject to limitations that would
1328
-
make it inappropriate for industrial-strength production use.
1329
-
Classes of quality-of-implementation requirements are as follows:
1330
-
1331
-
<ol>
1332
-
<li> <a href="#Specialization">Specialization</a>
1333
-
<li> <a href="#Performance and Scalability">Performance and Scalability</a>
1334
-
<li> <a href="#Forward Progress">Forward Progress</a>
1335
-
<li> <a href="#Composability">Composability</a>
1336
-
<li> <a href="#Corner Cases">Corner Cases</a>
1337
-
</ol>
1338
-
1339
-
<p>
1340
-
These classes is covered in the following sections.
1341
-
1342
-
<h3><a name="Specialization">Specialization</a></h3>
1343
-
1344
-
<p>
1345
-
RCU is and always has been intended primarily for read-mostly situations,
1346
-
which means that RCU's read-side primitives are optimized, often at the
1347
-
expense of its update-side primitives.
1348
-
Experience thus far is captured by the following list of situations:
1349
-
1350
-
<ol>
1351
-
<li> Read-mostly data, where stale and inconsistent data is not
1352
-
a problem: RCU works great!
1353
-
<li> Read-mostly data, where data must be consistent:
1354
-
RCU works well.
1355
-
<li> Read-write data, where data must be consistent:
1356
-
RCU <i>might</i> work OK.
1357
-
Or not.
1358
-
<li> Write-mostly data, where data must be consistent:
1359
-
RCU is very unlikely to be the right tool for the job,
1360
-
with the following exceptions, where RCU can provide:
1361
-
<ol type=a>
1362
-
<li> Existence guarantees for update-friendly mechanisms.
1363
-
<li> Wait-free read-side primitives for real-time use.
1364
-
</ol>
1365
-
</ol>
1366
-
1367
-
<p>
1368
-
This focus on read-mostly situations means that RCU must interoperate
1369
-
with other synchronization primitives.
1370
-
For example, the <tt>add_gp()</tt> and <tt>remove_gp_synchronous()</tt>
1371
-
examples discussed earlier use RCU to protect readers and locking to
1372
-
coordinate updaters.
1373
-
However, the need extends much farther, requiring that a variety of
1374
-
synchronization primitives be legal within RCU read-side critical sections,
1375
-
including spinlocks, sequence locks, atomic operations, reference
1376
-
counters, and memory barriers.
1377
-
1378
-
<table>
1379
-
<tr><th> </th></tr>
1380
-
<tr><th align="left">Quick Quiz:</th></tr>
1381
-
<tr><td>
1382
-
What about sleeping locks?
1383
-
</td></tr>
1384
-
<tr><th align="left">Answer:</th></tr>
1385
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
1386
-
These are forbidden within Linux-kernel RCU read-side critical
1387
-
sections because it is not legal to place a quiescent state
1388
-
(in this case, voluntary context switch) within an RCU read-side
1389
-
critical section.
1390
-
However, sleeping locks may be used within userspace RCU read-side
1391
-
critical sections, and also within Linux-kernel sleepable RCU
1392
-
<a href="#Sleepable RCU"><font color="ffffff">(SRCU)</font></a>
1393
-
read-side critical sections.
1394
-
In addition, the -rt patchset turns spinlocks into a
1395
-
sleeping locks so that the corresponding critical sections
1396
-
can be preempted, which also means that these sleeplockified
1397
-
spinlocks (but not other sleeping locks!) may be acquire within
1398
-
-rt-Linux-kernel RCU read-side critical sections.
1399
-
</font>
1400
-
1401
-
<p><font color="ffffff">
1402
-
Note that it <i>is</i> legal for a normal RCU read-side
1403
-
critical section to conditionally acquire a sleeping locks
1404
-
(as in <tt>mutex_trylock()</tt>), but only as long as it does
1405
-
not loop indefinitely attempting to conditionally acquire that
1406
-
sleeping locks.
1407
-
The key point is that things like <tt>mutex_trylock()</tt>
1408
-
either return with the mutex held, or return an error indication if
1409
-
the mutex was not immediately available.
1410
-
Either way, <tt>mutex_trylock()</tt> returns immediately without
1411
-
sleeping.
1412
-
</font></td></tr>
1413
-
<tr><td> </td></tr>
1414
-
</table>
1415
-
1416
-
<p>
1417
-
It often comes as a surprise that many algorithms do not require a
1418
-
consistent view of data, but many can function in that mode,
1419
-
with network routing being the poster child.
1420
-
Internet routing algorithms take significant time to propagate
1421
-
updates, so that by the time an update arrives at a given system,
1422
-
that system has been sending network traffic the wrong way for
1423
-
a considerable length of time.
1424
-
Having a few threads continue to send traffic the wrong way for a
1425
-
few more milliseconds is clearly not a problem: In the worst case,
1426
-
TCP retransmissions will eventually get the data where it needs to go.
1427
-
In general, when tracking the state of the universe outside of the
1428
-
computer, some level of inconsistency must be tolerated due to
1429
-
speed-of-light delays if nothing else.
1430
-
1431
-
<p>
1432
-
Furthermore, uncertainty about external state is inherent in many cases.
1433
-
For example, a pair of veterinarians might use heartbeat to determine
1434
-
whether or not a given cat was alive.
1435
-
But how long should they wait after the last heartbeat to decide that
1436
-
the cat is in fact dead?
1437
-
Waiting less than 400 milliseconds makes no sense because this would
1438
-
mean that a relaxed cat would be considered to cycle between death
1439
-
and life more than 100 times per minute.
1440
-
Moreover, just as with human beings, a cat's heart might stop for
1441
-
some period of time, so the exact wait period is a judgment call.
1442
-
One of our pair of veterinarians might wait 30 seconds before pronouncing
1443
-
the cat dead, while the other might insist on waiting a full minute.
1444
-
The two veterinarians would then disagree on the state of the cat during
1445
-
the final 30 seconds of the minute following the last heartbeat.
1446
-
1447
-
<p>
1448
-
Interestingly enough, this same situation applies to hardware.
1449
-
When push comes to shove, how do we tell whether or not some
1450
-
external server has failed?
1451
-
We send messages to it periodically, and declare it failed if we
1452
-
don't receive a response within a given period of time.
1453
-
Policy decisions can usually tolerate short
1454
-
periods of inconsistency.
1455
-
The policy was decided some time ago, and is only now being put into
1456
-
effect, so a few milliseconds of delay is normally inconsequential.
1457
-
1458
-
<p>
1459
-
However, there are algorithms that absolutely must see consistent data.
1460
-
For example, the translation between a user-level SystemV semaphore
1461
-
ID to the corresponding in-kernel data structure is protected by RCU,
1462
-
but it is absolutely forbidden to update a semaphore that has just been
1463
-
removed.
1464
-
In the Linux kernel, this need for consistency is accommodated by acquiring
1465
-
spinlocks located in the in-kernel data structure from within
1466
-
the RCU read-side critical section, and this is indicated by the
1467
-
green box in the figure above.
1468
-
Many other techniques may be used, and are in fact used within the
1469
-
Linux kernel.
1470
-
1471
-
<p>
1472
-
In short, RCU is not required to maintain consistency, and other
1473
-
mechanisms may be used in concert with RCU when consistency is required.
1474
-
RCU's specialization allows it to do its job extremely well, and its
1475
-
ability to interoperate with other synchronization mechanisms allows
1476
-
the right mix of synchronization tools to be used for a given job.
1477
-
1478
-
<h3><a name="Performance and Scalability">Performance and Scalability</a></h3>
1479
-
1480
-
<p>
1481
-
Energy efficiency is a critical component of performance today,
1482
-
and Linux-kernel RCU implementations must therefore avoid unnecessarily
1483
-
awakening idle CPUs.
1484
-
I cannot claim that this requirement was premeditated.
1485
-
In fact, I learned of it during a telephone conversation in which I
1486
-
was given “frank and open” feedback on the importance
1487
-
of energy efficiency in battery-powered systems and on specific
1488
-
energy-efficiency shortcomings of the Linux-kernel RCU implementation.
1489
-
In my experience, the battery-powered embedded community will consider
1490
-
any unnecessary wakeups to be extremely unfriendly acts.
1491
-
So much so that mere Linux-kernel-mailing-list posts are
1492
-
insufficient to vent their ire.
1493
-
1494
-
<p>
1495
-
Memory consumption is not particularly important for in most
1496
-
situations, and has become decreasingly
1497
-
so as memory sizes have expanded and memory
1498
-
costs have plummeted.
1499
-
However, as I learned from Matt Mackall's
1500
-
<a href="http://elinux.org/Linux_Tiny-FAQ">bloatwatch</a>
1501
-
efforts, memory footprint is critically important on single-CPU systems with
1502
-
non-preemptible (<tt>CONFIG_PREEMPT=n</tt>) kernels, and thus
1503
-
<a href="https://lkml.kernel.org/g/20090113221724.GA15307@linux.vnet.ibm.com">tiny RCU</a>
1504
-
was born.
1505
-
Josh Triplett has since taken over the small-memory banner with his
1506
-
<a href="https://tiny.wiki.kernel.org/">Linux kernel tinification</a>
1507
-
project, which resulted in
1508
-
<a href="#Sleepable RCU">SRCU</a>
1509
-
becoming optional for those kernels not needing it.
1510
-
1511
-
<p>
1512
-
The remaining performance requirements are, for the most part,
1513
-
unsurprising.
1514
-
For example, in keeping with RCU's read-side specialization,
1515
-
<tt>rcu_dereference()</tt> should have negligible overhead (for
1516
-
example, suppression of a few minor compiler optimizations).
1517
-
Similarly, in non-preemptible environments, <tt>rcu_read_lock()</tt> and
1518
-
<tt>rcu_read_unlock()</tt> should have exactly zero overhead.
1519
-
1520
-
<p>
1521
-
In preemptible environments, in the case where the RCU read-side
1522
-
critical section was not preempted (as will be the case for the
1523
-
highest-priority real-time process), <tt>rcu_read_lock()</tt> and
1524
-
<tt>rcu_read_unlock()</tt> should have minimal overhead.
1525
-
In particular, they should not contain atomic read-modify-write
1526
-
operations, memory-barrier instructions, preemption disabling,
1527
-
interrupt disabling, or backwards branches.
1528
-
However, in the case where the RCU read-side critical section was preempted,
1529
-
<tt>rcu_read_unlock()</tt> may acquire spinlocks and disable interrupts.
1530
-
This is why it is better to nest an RCU read-side critical section
1531
-
within a preempt-disable region than vice versa, at least in cases
1532
-
where that critical section is short enough to avoid unduly degrading
1533
-
real-time latencies.
1534
-
1535
-
<p>
1536
-
The <tt>synchronize_rcu()</tt> grace-period-wait primitive is
1537
-
optimized for throughput.
1538
-
It may therefore incur several milliseconds of latency in addition to
1539
-
the duration of the longest RCU read-side critical section.
1540
-
On the other hand, multiple concurrent invocations of
1541
-
<tt>synchronize_rcu()</tt> are required to use batching optimizations
1542
-
so that they can be satisfied by a single underlying grace-period-wait
1543
-
operation.
1544
-
For example, in the Linux kernel, it is not unusual for a single
1545
-
grace-period-wait operation to serve more than
1546
-
<a href="https://www.usenix.org/conference/2004-usenix-annual-technical-conference/making-rcu-safe-deep-sub-millisecond-response">1,000 separate invocations</a>
1547
-
of <tt>synchronize_rcu()</tt>, thus amortizing the per-invocation
1548
-
overhead down to nearly zero.
1549
-
However, the grace-period optimization is also required to avoid
1550
-
measurable degradation of real-time scheduling and interrupt latencies.
1551
-
1552
-
<p>
1553
-
In some cases, the multi-millisecond <tt>synchronize_rcu()</tt>
1554
-
latencies are unacceptable.
1555
-
In these cases, <tt>synchronize_rcu_expedited()</tt> may be used
1556
-
instead, reducing the grace-period latency down to a few tens of
1557
-
microseconds on small systems, at least in cases where the RCU read-side
1558
-
critical sections are short.
1559
-
There are currently no special latency requirements for
1560
-
<tt>synchronize_rcu_expedited()</tt> on large systems, but,
1561
-
consistent with the empirical nature of the RCU specification,
1562
-
that is subject to change.
1563
-
However, there most definitely are scalability requirements:
1564
-
A storm of <tt>synchronize_rcu_expedited()</tt> invocations on 4096
1565
-
CPUs should at least make reasonable forward progress.
1566
-
In return for its shorter latencies, <tt>synchronize_rcu_expedited()</tt>
1567
-
is permitted to impose modest degradation of real-time latency
1568
-
on non-idle online CPUs.
1569
-
Here, “modest” means roughly the same latency
1570
-
degradation as a scheduling-clock interrupt.
1571
-
1572
-
<p>
1573
-
There are a number of situations where even
1574
-
<tt>synchronize_rcu_expedited()</tt>'s reduced grace-period
1575
-
latency is unacceptable.
1576
-
In these situations, the asynchronous <tt>call_rcu()</tt> can be
1577
-
used in place of <tt>synchronize_rcu()</tt> as follows:
1578
-
1579
-
<blockquote>
1580
-
<pre>
1581
-
1 struct foo {
1582
-
2 int a;
1583
-
3 int b;
1584
-
4 struct rcu_head rh;
1585
-
5 };
1586
-
6
1587
-
7 static void remove_gp_cb(struct rcu_head *rhp)
1588
-
8 {
1589
-
9 struct foo *p = container_of(rhp, struct foo, rh);
1590
-
10
1591
-
11 kfree(p);
1592
-
12 }
1593
-
13
1594
-
14 bool remove_gp_asynchronous(void)
1595
-
15 {
1596
-
16 struct foo *p;
1597
-
17
1598
-
18 spin_lock(&gp_lock);
1599
-
19 p = rcu_access_pointer(gp);
1600
-
20 if (!p) {
1601
-
21 spin_unlock(&gp_lock);
1602
-
22 return false;
1603
-
23 }
1604
-
24 rcu_assign_pointer(gp, NULL);
1605
-
25 call_rcu(&p->rh, remove_gp_cb);
1606
-
26 spin_unlock(&gp_lock);
1607
-
27 return true;
1608
-
28 }
1609
-
</pre>
1610
-
</blockquote>
1611
-
1612
-
<p>
1613
-
A definition of <tt>struct foo</tt> is finally needed, and appears
1614
-
on lines 1-5.
1615
-
The function <tt>remove_gp_cb()</tt> is passed to <tt>call_rcu()</tt>
1616
-
on line 25, and will be invoked after the end of a subsequent
1617
-
grace period.
1618
-
This gets the same effect as <tt>remove_gp_synchronous()</tt>,
1619
-
but without forcing the updater to wait for a grace period to elapse.
1620
-
The <tt>call_rcu()</tt> function may be used in a number of
1621
-
situations where neither <tt>synchronize_rcu()</tt> nor
1622
-
<tt>synchronize_rcu_expedited()</tt> would be legal,
1623
-
including within preempt-disable code, <tt>local_bh_disable()</tt> code,
1624
-
interrupt-disable code, and interrupt handlers.
1625
-
However, even <tt>call_rcu()</tt> is illegal within NMI handlers
1626
-
and from idle and offline CPUs.
1627
-
The callback function (<tt>remove_gp_cb()</tt> in this case) will be
1628
-
executed within softirq (software interrupt) environment within the
1629
-
Linux kernel,
1630
-
either within a real softirq handler or under the protection
1631
-
of <tt>local_bh_disable()</tt>.
1632
-
In both the Linux kernel and in userspace, it is bad practice to
1633
-
write an RCU callback function that takes too long.
1634
-
Long-running operations should be relegated to separate threads or
1635
-
(in the Linux kernel) workqueues.
1636
-
1637
-
<table>
1638
-
<tr><th> </th></tr>
1639
-
<tr><th align="left">Quick Quiz:</th></tr>
1640
-
<tr><td>
1641
-
Why does line 19 use <tt>rcu_access_pointer()</tt>?
1642
-
After all, <tt>call_rcu()</tt> on line 25 stores into the
1643
-
structure, which would interact badly with concurrent insertions.
1644
-
Doesn't this mean that <tt>rcu_dereference()</tt> is required?
1645
-
</td></tr>
1646
-
<tr><th align="left">Answer:</th></tr>
1647
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
1648
-
Presumably the <tt>->gp_lock</tt> acquired on line 18 excludes
1649
-
any changes, including any insertions that <tt>rcu_dereference()</tt>
1650
-
would protect against.
1651
-
Therefore, any insertions will be delayed until after
1652
-
<tt>->gp_lock</tt>
1653
-
is released on line 25, which in turn means that
1654
-
<tt>rcu_access_pointer()</tt> suffices.
1655
-
</font></td></tr>
1656
-
<tr><td> </td></tr>
1657
-
</table>
1658
-
1659
-
<p>
1660
-
However, all that <tt>remove_gp_cb()</tt> is doing is
1661
-
invoking <tt>kfree()</tt> on the data element.
1662
-
This is a common idiom, and is supported by <tt>kfree_rcu()</tt>,
1663
-
which allows “fire and forget” operation as shown below:
1664
-
1665
-
<blockquote>
1666
-
<pre>
1667
-
1 struct foo {
1668
-
2 int a;
1669
-
3 int b;
1670
-
4 struct rcu_head rh;
1671
-
5 };
1672
-
6
1673
-
7 bool remove_gp_faf(void)
1674
-
8 {
1675
-
9 struct foo *p;
1676
-
10
1677
-
11 spin_lock(&gp_lock);
1678
-
12 p = rcu_dereference(gp);
1679
-
13 if (!p) {
1680
-
14 spin_unlock(&gp_lock);
1681
-
15 return false;
1682
-
16 }
1683
-
17 rcu_assign_pointer(gp, NULL);
1684
-
18 kfree_rcu(p, rh);
1685
-
19 spin_unlock(&gp_lock);
1686
-
20 return true;
1687
-
21 }
1688
-
</pre>
1689
-
</blockquote>
1690
-
1691
-
<p>
1692
-
Note that <tt>remove_gp_faf()</tt> simply invokes
1693
-
<tt>kfree_rcu()</tt> and proceeds, without any need to pay any
1694
-
further attention to the subsequent grace period and <tt>kfree()</tt>.
1695
-
It is permissible to invoke <tt>kfree_rcu()</tt> from the same
1696
-
environments as for <tt>call_rcu()</tt>.
1697
-
Interestingly enough, DYNIX/ptx had the equivalents of
1698
-
<tt>call_rcu()</tt> and <tt>kfree_rcu()</tt>, but not
1699
-
<tt>synchronize_rcu()</tt>.
1700
-
This was due to the fact that RCU was not heavily used within DYNIX/ptx,
1701
-
so the very few places that needed something like
1702
-
<tt>synchronize_rcu()</tt> simply open-coded it.
1703
-
1704
-
<table>
1705
-
<tr><th> </th></tr>
1706
-
<tr><th align="left">Quick Quiz:</th></tr>
1707
-
<tr><td>
1708
-
Earlier it was claimed that <tt>call_rcu()</tt> and
1709
-
<tt>kfree_rcu()</tt> allowed updaters to avoid being blocked
1710
-
by readers.
1711
-
But how can that be correct, given that the invocation of the callback
1712
-
and the freeing of the memory (respectively) must still wait for
1713
-
a grace period to elapse?
1714
-
</td></tr>
1715
-
<tr><th align="left">Answer:</th></tr>
1716
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
1717
-
We could define things this way, but keep in mind that this sort of
1718
-
definition would say that updates in garbage-collected languages
1719
-
cannot complete until the next time the garbage collector runs,
1720
-
which does not seem at all reasonable.
1721
-
The key point is that in most cases, an updater using either
1722
-
<tt>call_rcu()</tt> or <tt>kfree_rcu()</tt> can proceed to the
1723
-
next update as soon as it has invoked <tt>call_rcu()</tt> or
1724
-
<tt>kfree_rcu()</tt>, without having to wait for a subsequent
1725
-
grace period.
1726
-
</font></td></tr>
1727
-
<tr><td> </td></tr>
1728
-
</table>
1729
-
1730
-
<p>
1731
-
But what if the updater must wait for the completion of code to be
1732
-
executed after the end of the grace period, but has other tasks
1733
-
that can be carried out in the meantime?
1734
-
The polling-style <tt>get_state_synchronize_rcu()</tt> and
1735
-
<tt>cond_synchronize_rcu()</tt> functions may be used for this
1736
-
purpose, as shown below:
1737
-
1738
-
<blockquote>
1739
-
<pre>
1740
-
1 bool remove_gp_poll(void)
1741
-
2 {
1742
-
3 struct foo *p;
1743
-
4 unsigned long s;
1744
-
5
1745
-
6 spin_lock(&gp_lock);
1746
-
7 p = rcu_access_pointer(gp);
1747
-
8 if (!p) {
1748
-
9 spin_unlock(&gp_lock);
1749
-
10 return false;
1750
-
11 }
1751
-
12 rcu_assign_pointer(gp, NULL);
1752
-
13 spin_unlock(&gp_lock);
1753
-
14 s = get_state_synchronize_rcu();
1754
-
15 do_something_while_waiting();
1755
-
16 cond_synchronize_rcu(s);
1756
-
17 kfree(p);
1757
-
18 return true;
1758
-
19 }
1759
-
</pre>
1760
-
</blockquote>
1761
-
1762
-
<p>
1763
-
On line 14, <tt>get_state_synchronize_rcu()</tt> obtains a
1764
-
“cookie” from RCU,
1765
-
then line 15 carries out other tasks,
1766
-
and finally, line 16 returns immediately if a grace period has
1767
-
elapsed in the meantime, but otherwise waits as required.
1768
-
The need for <tt>get_state_synchronize_rcu</tt> and
1769
-
<tt>cond_synchronize_rcu()</tt> has appeared quite recently,
1770
-
so it is too early to tell whether they will stand the test of time.
1771
-
1772
-
<p>
1773
-
RCU thus provides a range of tools to allow updaters to strike the
1774
-
required tradeoff between latency, flexibility and CPU overhead.
1775
-
1776
-
<h3><a name="Forward Progress">Forward Progress</a></h3>
1777
-
1778
-
<p>
1779
-
In theory, delaying grace-period completion and callback invocation
1780
-
is harmless.
1781
-
In practice, not only are memory sizes finite but also callbacks sometimes
1782
-
do wakeups, and sufficiently deferred wakeups can be difficult
1783
-
to distinguish from system hangs.
1784
-
Therefore, RCU must provide a number of mechanisms to promote forward
1785
-
progress.
1786
-
1787
-
<p>
1788
-
These mechanisms are not foolproof, nor can they be.
1789
-
For one simple example, an infinite loop in an RCU read-side critical
1790
-
section must by definition prevent later grace periods from ever completing.
1791
-
For a more involved example, consider a 64-CPU system built with
1792
-
<tt>CONFIG_RCU_NOCB_CPU=y</tt> and booted with <tt>rcu_nocbs=1-63</tt>,
1793
-
where CPUs 1 through 63 spin in tight loops that invoke
1794
-
<tt>call_rcu()</tt>.
1795
-
Even if these tight loops also contain calls to <tt>cond_resched()</tt>
1796
-
(thus allowing grace periods to complete), CPU 0 simply will
1797
-
not be able to invoke callbacks as fast as the other 63 CPUs can
1798
-
register them, at least not until the system runs out of memory.
1799
-
In both of these examples, the Spiderman principle applies: With great
1800
-
power comes great responsibility.
1801
-
However, short of this level of abuse, RCU is required to
1802
-
ensure timely completion of grace periods and timely invocation of
1803
-
callbacks.
1804
-
1805
-
<p>
1806
-
RCU takes the following steps to encourage timely completion of
1807
-
grace periods:
1808
-
1809
-
<ol>
1810
-
<li> If a grace period fails to complete within 100 milliseconds,
1811
-
RCU causes future invocations of <tt>cond_resched()</tt> on
1812
-
the holdout CPUs to provide an RCU quiescent state.
1813
-
RCU also causes those CPUs' <tt>need_resched()</tt> invocations
1814
-
to return <tt>true</tt>, but only after the corresponding CPU's
1815
-
next scheduling-clock.
1816
-
<li> CPUs mentioned in the <tt>nohz_full</tt> kernel boot parameter
1817
-
can run indefinitely in the kernel without scheduling-clock
1818
-
interrupts, which defeats the above <tt>need_resched()</tt>
1819
-
strategem.
1820
-
RCU will therefore invoke <tt>resched_cpu()</tt> on any
1821
-
<tt>nohz_full</tt> CPUs still holding out after
1822
-
109 milliseconds.
1823
-
<li> In kernels built with <tt>CONFIG_RCU_BOOST=y</tt>, if a given
1824
-
task that has been preempted within an RCU read-side critical
1825
-
section is holding out for more than 500 milliseconds,
1826
-
RCU will resort to priority boosting.
1827
-
<li> If a CPU is still holding out 10 seconds into the grace
1828
-
period, RCU will invoke <tt>resched_cpu()</tt> on it regardless
1829
-
of its <tt>nohz_full</tt> state.
1830
-
</ol>
1831
-
1832
-
<p>
1833
-
The above values are defaults for systems running with <tt>HZ=1000</tt>.
1834
-
They will vary as the value of <tt>HZ</tt> varies, and can also be
1835
-
changed using the relevant Kconfig options and kernel boot parameters.
1836
-
RCU currently does not do much sanity checking of these
1837
-
parameters, so please use caution when changing them.
1838
-
Note that these forward-progress measures are provided only for RCU,
1839
-
not for
1840
-
<a href="#Sleepable RCU">SRCU</a> or
1841
-
<a href="#Tasks RCU">Tasks RCU</a>.
1842
-
1843
-
<p>
1844
-
RCU takes the following steps in <tt>call_rcu()</tt> to encourage timely
1845
-
invocation of callbacks when any given non-<tt>rcu_nocbs</tt> CPU has
1846
-
10,000 callbacks, or has 10,000 more callbacks than it had the last time
1847
-
encouragement was provided:
1848
-
1849
-
<ol>
1850
-
<li> Starts a grace period, if one is not already in progress.
1851
-
<li> Forces immediate checking for quiescent states, rather than
1852
-
waiting for three milliseconds to have elapsed since the
1853
-
beginning of the grace period.
1854
-
<li> Immediately tags the CPU's callbacks with their grace period
1855
-
completion numbers, rather than waiting for the <tt>RCU_SOFTIRQ</tt>
1856
-
handler to get around to it.
1857
-
<li> Lifts callback-execution batch limits, which speeds up callback
1858
-
invocation at the expense of degrading realtime response.
1859
-
</ol>
1860
-
1861
-
<p>
1862
-
Again, these are default values when running at <tt>HZ=1000</tt>,
1863
-
and can be overridden.
1864
-
Again, these forward-progress measures are provided only for RCU,
1865
-
not for
1866
-
<a href="#Sleepable RCU">SRCU</a> or
1867
-
<a href="#Tasks RCU">Tasks RCU</a>.
1868
-
Even for RCU, callback-invocation forward progress for <tt>rcu_nocbs</tt>
1869
-
CPUs is much less well-developed, in part because workloads benefiting
1870
-
from <tt>rcu_nocbs</tt> CPUs tend to invoke <tt>call_rcu()</tt>
1871
-
relatively infrequently.
1872
-
If workloads emerge that need both <tt>rcu_nocbs</tt> CPUs and high
1873
-
<tt>call_rcu()</tt> invocation rates, then additional forward-progress
1874
-
work will be required.
1875
-
1876
-
<h3><a name="Composability">Composability</a></h3>
1877
-
1878
-
<p>
1879
-
Composability has received much attention in recent years, perhaps in part
1880
-
due to the collision of multicore hardware with object-oriented techniques
1881
-
designed in single-threaded environments for single-threaded use.
1882
-
And in theory, RCU read-side critical sections may be composed, and in
1883
-
fact may be nested arbitrarily deeply.
1884
-
In practice, as with all real-world implementations of composable
1885
-
constructs, there are limitations.
1886
-
1887
-
<p>
1888
-
Implementations of RCU for which <tt>rcu_read_lock()</tt>
1889
-
and <tt>rcu_read_unlock()</tt> generate no code, such as
1890
-
Linux-kernel RCU when <tt>CONFIG_PREEMPT=n</tt>, can be
1891
-
nested arbitrarily deeply.
1892
-
After all, there is no overhead.
1893
-
Except that if all these instances of <tt>rcu_read_lock()</tt>
1894
-
and <tt>rcu_read_unlock()</tt> are visible to the compiler,
1895
-
compilation will eventually fail due to exhausting memory,
1896
-
mass storage, or user patience, whichever comes first.
1897
-
If the nesting is not visible to the compiler, as is the case with
1898
-
mutually recursive functions each in its own translation unit,
1899
-
stack overflow will result.
1900
-
If the nesting takes the form of loops, perhaps in the guise of tail
1901
-
recursion, either the control variable
1902
-
will overflow or (in the Linux kernel) you will get an RCU CPU stall warning.
1903
-
Nevertheless, this class of RCU implementations is one
1904
-
of the most composable constructs in existence.
1905
-
1906
-
<p>
1907
-
RCU implementations that explicitly track nesting depth
1908
-
are limited by the nesting-depth counter.
1909
-
For example, the Linux kernel's preemptible RCU limits nesting to
1910
-
<tt>INT_MAX</tt>.
1911
-
This should suffice for almost all practical purposes.
1912
-
That said, a consecutive pair of RCU read-side critical sections
1913
-
between which there is an operation that waits for a grace period
1914
-
cannot be enclosed in another RCU read-side critical section.
1915
-
This is because it is not legal to wait for a grace period within
1916
-
an RCU read-side critical section: To do so would result either
1917
-
in deadlock or
1918
-
in RCU implicitly splitting the enclosing RCU read-side critical
1919
-
section, neither of which is conducive to a long-lived and prosperous
1920
-
kernel.
1921
-
1922
-
<p>
1923
-
It is worth noting that RCU is not alone in limiting composability.
1924
-
For example, many transactional-memory implementations prohibit
1925
-
composing a pair of transactions separated by an irrevocable
1926
-
operation (for example, a network receive operation).
1927
-
For another example, lock-based critical sections can be composed
1928
-
surprisingly freely, but only if deadlock is avoided.
1929
-
1930
-
<p>
1931
-
In short, although RCU read-side critical sections are highly composable,
1932
-
care is required in some situations, just as is the case for any other
1933
-
composable synchronization mechanism.
1934
-
1935
-
<h3><a name="Corner Cases">Corner Cases</a></h3>
1936
-
1937
-
<p>
1938
-
A given RCU workload might have an endless and intense stream of
1939
-
RCU read-side critical sections, perhaps even so intense that there
1940
-
was never a point in time during which there was not at least one
1941
-
RCU read-side critical section in flight.
1942
-
RCU cannot allow this situation to block grace periods: As long as
1943
-
all the RCU read-side critical sections are finite, grace periods
1944
-
must also be finite.
1945
-
1946
-
<p>
1947
-
That said, preemptible RCU implementations could potentially result
1948
-
in RCU read-side critical sections being preempted for long durations,
1949
-
which has the effect of creating a long-duration RCU read-side
1950
-
critical section.
1951
-
This situation can arise only in heavily loaded systems, but systems using
1952
-
real-time priorities are of course more vulnerable.
1953
-
Therefore, RCU priority boosting is provided to help deal with this
1954
-
case.
1955
-
That said, the exact requirements on RCU priority boosting will likely
1956
-
evolve as more experience accumulates.
1957
-
1958
-
<p>
1959
-
Other workloads might have very high update rates.
1960
-
Although one can argue that such workloads should instead use
1961
-
something other than RCU, the fact remains that RCU must
1962
-
handle such workloads gracefully.
1963
-
This requirement is another factor driving batching of grace periods,
1964
-
but it is also the driving force behind the checks for large numbers
1965
-
of queued RCU callbacks in the <tt>call_rcu()</tt> code path.
1966
-
Finally, high update rates should not delay RCU read-side critical
1967
-
sections, although some small read-side delays can occur when using
1968
-
<tt>synchronize_rcu_expedited()</tt>, courtesy of this function's use
1969
-
of <tt>smp_call_function_single()</tt>.
1970
-
1971
-
<p>
1972
-
Although all three of these corner cases were understood in the early
1973
-
1990s, a simple user-level test consisting of <tt>close(open(path))</tt>
1974
-
in a tight loop
1975
-
in the early 2000s suddenly provided a much deeper appreciation of the
1976
-
high-update-rate corner case.
1977
-
This test also motivated addition of some RCU code to react to high update
1978
-
rates, for example, if a given CPU finds itself with more than 10,000
1979
-
RCU callbacks queued, it will cause RCU to take evasive action by
1980
-
more aggressively starting grace periods and more aggressively forcing
1981
-
completion of grace-period processing.
1982
-
This evasive action causes the grace period to complete more quickly,
1983
-
but at the cost of restricting RCU's batching optimizations, thus
1984
-
increasing the CPU overhead incurred by that grace period.
1985
-
1986
-
<h2><a name="Software-Engineering Requirements">
1987
-
Software-Engineering Requirements</a></h2>
1988
-
1989
-
<p>
1990
-
Between Murphy's Law and “To err is human”, it is necessary to
1991
-
guard against mishaps and misuse:
1992
-
1993
-
<ol>
1994
-
<li> It is all too easy to forget to use <tt>rcu_read_lock()</tt>
1995
-
everywhere that it is needed, so kernels built with
1996
-
<tt>CONFIG_PROVE_RCU=y</tt> will splat if
1997
-
<tt>rcu_dereference()</tt> is used outside of an
1998
-
RCU read-side critical section.
1999
-
Update-side code can use <tt>rcu_dereference_protected()</tt>,
2000
-
which takes a
2001
-
<a href="https://lwn.net/Articles/371986/">lockdep expression</a>
2002
-
to indicate what is providing the protection.
2003
-
If the indicated protection is not provided, a lockdep splat
2004
-
is emitted.
2005
-
2006
-
<p>
2007
-
Code shared between readers and updaters can use
2008
-
<tt>rcu_dereference_check()</tt>, which also takes a
2009
-
lockdep expression, and emits a lockdep splat if neither
2010
-
<tt>rcu_read_lock()</tt> nor the indicated protection
2011
-
is in place.
2012
-
In addition, <tt>rcu_dereference_raw()</tt> is used in those
2013
-
(hopefully rare) cases where the required protection cannot
2014
-
be easily described.
2015
-
Finally, <tt>rcu_read_lock_held()</tt> is provided to
2016
-
allow a function to verify that it has been invoked within
2017
-
an RCU read-side critical section.
2018
-
I was made aware of this set of requirements shortly after Thomas
2019
-
Gleixner audited a number of RCU uses.
2020
-
<li> A given function might wish to check for RCU-related preconditions
2021
-
upon entry, before using any other RCU API.
2022
-
The <tt>rcu_lockdep_assert()</tt> does this job,
2023
-
asserting the expression in kernels having lockdep enabled
2024
-
and doing nothing otherwise.
2025
-
<li> It is also easy to forget to use <tt>rcu_assign_pointer()</tt>
2026
-
and <tt>rcu_dereference()</tt>, perhaps (incorrectly)
2027
-
substituting a simple assignment.
2028
-
To catch this sort of error, a given RCU-protected pointer may be
2029
-
tagged with <tt>__rcu</tt>, after which sparse
2030
-
will complain about simple-assignment accesses to that pointer.
2031
-
Arnd Bergmann made me aware of this requirement, and also
2032
-
supplied the needed
2033
-
<a href="https://lwn.net/Articles/376011/">patch series</a>.
2034
-
<li> Kernels built with <tt>CONFIG_DEBUG_OBJECTS_RCU_HEAD=y</tt>
2035
-
will splat if a data element is passed to <tt>call_rcu()</tt>
2036
-
twice in a row, without a grace period in between.
2037
-
(This error is similar to a double free.)
2038
-
The corresponding <tt>rcu_head</tt> structures that are
2039
-
dynamically allocated are automatically tracked, but
2040
-
<tt>rcu_head</tt> structures allocated on the stack
2041
-
must be initialized with <tt>init_rcu_head_on_stack()</tt>
2042
-
and cleaned up with <tt>destroy_rcu_head_on_stack()</tt>.
2043
-
Similarly, statically allocated non-stack <tt>rcu_head</tt>
2044
-
structures must be initialized with <tt>init_rcu_head()</tt>
2045
-
and cleaned up with <tt>destroy_rcu_head()</tt>.
2046
-
Mathieu Desnoyers made me aware of this requirement, and also
2047
-
supplied the needed
2048
-
<a href="https://lkml.kernel.org/g/20100319013024.GA28456@Krystal">patch</a>.
2049
-
<li> An infinite loop in an RCU read-side critical section will
2050
-
eventually trigger an RCU CPU stall warning splat, with
2051
-
the duration of “eventually” being controlled by the
2052
-
<tt>RCU_CPU_STALL_TIMEOUT</tt> <tt>Kconfig</tt> option, or,
2053
-
alternatively, by the
2054
-
<tt>rcupdate.rcu_cpu_stall_timeout</tt> boot/sysfs
2055
-
parameter.
2056
-
However, RCU is not obligated to produce this splat
2057
-
unless there is a grace period waiting on that particular
2058
-
RCU read-side critical section.
2059
-
<p>
2060
-
Some extreme workloads might intentionally delay
2061
-
RCU grace periods, and systems running those workloads can
2062
-
be booted with <tt>rcupdate.rcu_cpu_stall_suppress</tt>
2063
-
to suppress the splats.
2064
-
This kernel parameter may also be set via <tt>sysfs</tt>.
2065
-
Furthermore, RCU CPU stall warnings are counter-productive
2066
-
during sysrq dumps and during panics.
2067
-
RCU therefore supplies the <tt>rcu_sysrq_start()</tt> and
2068
-
<tt>rcu_sysrq_end()</tt> API members to be called before
2069
-
and after long sysrq dumps.
2070
-
RCU also supplies the <tt>rcu_panic()</tt> notifier that is
2071
-
automatically invoked at the beginning of a panic to suppress
2072
-
further RCU CPU stall warnings.
2073
-
2074
-
<p>
2075
-
This requirement made itself known in the early 1990s, pretty
2076
-
much the first time that it was necessary to debug a CPU stall.
2077
-
That said, the initial implementation in DYNIX/ptx was quite
2078
-
generic in comparison with that of Linux.
2079
-
<li> Although it would be very good to detect pointers leaking out
2080
-
of RCU read-side critical sections, there is currently no
2081
-
good way of doing this.
2082
-
One complication is the need to distinguish between pointers
2083
-
leaking and pointers that have been handed off from RCU to
2084
-
some other synchronization mechanism, for example, reference
2085
-
counting.
2086
-
<li> In kernels built with <tt>CONFIG_RCU_TRACE=y</tt>, RCU-related
2087
-
information is provided via event tracing.
2088
-
<li> Open-coded use of <tt>rcu_assign_pointer()</tt> and
2089
-
<tt>rcu_dereference()</tt> to create typical linked
2090
-
data structures can be surprisingly error-prone.
2091
-
Therefore, RCU-protected
2092
-
<a href="https://lwn.net/Articles/609973/#RCU List APIs">linked lists</a>
2093
-
and, more recently, RCU-protected
2094
-
<a href="https://lwn.net/Articles/612100/">hash tables</a>
2095
-
are available.
2096
-
Many other special-purpose RCU-protected data structures are
2097
-
available in the Linux kernel and the userspace RCU library.
2098
-
<li> Some linked structures are created at compile time, but still
2099
-
require <tt>__rcu</tt> checking.
2100
-
The <tt>RCU_POINTER_INITIALIZER()</tt> macro serves this
2101
-
purpose.
2102
-
<li> It is not necessary to use <tt>rcu_assign_pointer()</tt>
2103
-
when creating linked structures that are to be published via
2104
-
a single external pointer.
2105
-
The <tt>RCU_INIT_POINTER()</tt> macro is provided for
2106
-
this task and also for assigning <tt>NULL</tt> pointers
2107
-
at runtime.
2108
-
</ol>
2109
-
2110
-
<p>
2111
-
This not a hard-and-fast list: RCU's diagnostic capabilities will
2112
-
continue to be guided by the number and type of usage bugs found
2113
-
in real-world RCU usage.
2114
-
2115
-
<h2><a name="Linux Kernel Complications">Linux Kernel Complications</a></h2>
2116
-
2117
-
<p>
2118
-
The Linux kernel provides an interesting environment for all kinds of
2119
-
software, including RCU.
2120
-
Some of the relevant points of interest are as follows:
2121
-
2122
-
<ol>
2123
-
<li> <a href="#Configuration">Configuration</a>.
2124
-
<li> <a href="#Firmware Interface">Firmware Interface</a>.
2125
-
<li> <a href="#Early Boot">Early Boot</a>.
2126
-
<li> <a href="#Interrupts and NMIs">
2127
-
Interrupts and non-maskable interrupts (NMIs)</a>.
2128
-
<li> <a href="#Loadable Modules">Loadable Modules</a>.
2129
-
<li> <a href="#Hotplug CPU">Hotplug CPU</a>.
2130
-
<li> <a href="#Scheduler and RCU">Scheduler and RCU</a>.
2131
-
<li> <a href="#Tracing and RCU">Tracing and RCU</a>.
2132
-
<li> <a href="#Energy Efficiency">Energy Efficiency</a>.
2133
-
<li> <a href="#Scheduling-Clock Interrupts and RCU">
2134
-
Scheduling-Clock Interrupts and RCU</a>.
2135
-
<li> <a href="#Memory Efficiency">Memory Efficiency</a>.
2136
-
<li> <a href="#Performance, Scalability, Response Time, and Reliability">
2137
-
Performance, Scalability, Response Time, and Reliability</a>.
2138
-
</ol>
2139
-
2140
-
<p>
2141
-
This list is probably incomplete, but it does give a feel for the
2142
-
most notable Linux-kernel complications.
2143
-
Each of the following sections covers one of the above topics.
2144
-
2145
-
<h3><a name="Configuration">Configuration</a></h3>
2146
-
2147
-
<p>
2148
-
RCU's goal is automatic configuration, so that almost nobody
2149
-
needs to worry about RCU's <tt>Kconfig</tt> options.
2150
-
And for almost all users, RCU does in fact work well
2151
-
“out of the box.”
2152
-
2153
-
<p>
2154
-
However, there are specialized use cases that are handled by
2155
-
kernel boot parameters and <tt>Kconfig</tt> options.
2156
-
Unfortunately, the <tt>Kconfig</tt> system will explicitly ask users
2157
-
about new <tt>Kconfig</tt> options, which requires almost all of them
2158
-
be hidden behind a <tt>CONFIG_RCU_EXPERT</tt> <tt>Kconfig</tt> option.
2159
-
2160
-
<p>
2161
-
This all should be quite obvious, but the fact remains that
2162
-
Linus Torvalds recently had to
2163
-
<a href="https://lkml.kernel.org/g/CA+55aFy4wcCwaL4okTs8wXhGZ5h-ibecy_Meg9C4MNQrUnwMcg@mail.gmail.com">remind</a>
2164
-
me of this requirement.
2165
-
2166
-
<h3><a name="Firmware Interface">Firmware Interface</a></h3>
2167
-
2168
-
<p>
2169
-
In many cases, kernel obtains information about the system from the
2170
-
firmware, and sometimes things are lost in translation.
2171
-
Or the translation is accurate, but the original message is bogus.
2172
-
2173
-
<p>
2174
-
For example, some systems' firmware overreports the number of CPUs,
2175
-
sometimes by a large factor.
2176
-
If RCU naively believed the firmware, as it used to do,
2177
-
it would create too many per-CPU kthreads.
2178
-
Although the resulting system will still run correctly, the extra
2179
-
kthreads needlessly consume memory and can cause confusion
2180
-
when they show up in <tt>ps</tt> listings.
2181
-
2182
-
<p>
2183
-
RCU must therefore wait for a given CPU to actually come online before
2184
-
it can allow itself to believe that the CPU actually exists.
2185
-
The resulting “ghost CPUs” (which are never going to
2186
-
come online) cause a number of
2187
-
<a href="https://paulmck.livejournal.com/37494.html">interesting complications</a>.
2188
-
2189
-
<h3><a name="Early Boot">Early Boot</a></h3>
2190
-
2191
-
<p>
2192
-
The Linux kernel's boot sequence is an interesting process,
2193
-
and RCU is used early, even before <tt>rcu_init()</tt>
2194
-
is invoked.
2195
-
In fact, a number of RCU's primitives can be used as soon as the
2196
-
initial task's <tt>task_struct</tt> is available and the
2197
-
boot CPU's per-CPU variables are set up.
2198
-
The read-side primitives (<tt>rcu_read_lock()</tt>,
2199
-
<tt>rcu_read_unlock()</tt>, <tt>rcu_dereference()</tt>,
2200
-
and <tt>rcu_access_pointer()</tt>) will operate normally very early on,
2201
-
as will <tt>rcu_assign_pointer()</tt>.
2202
-
2203
-
<p>
2204
-
Although <tt>call_rcu()</tt> may be invoked at any
2205
-
time during boot, callbacks are not guaranteed to be invoked until after
2206
-
all of RCU's kthreads have been spawned, which occurs at
2207
-
<tt>early_initcall()</tt> time.
2208
-
This delay in callback invocation is due to the fact that RCU does not
2209
-
invoke callbacks until it is fully initialized, and this full initialization
2210
-
cannot occur until after the scheduler has initialized itself to the
2211
-
point where RCU can spawn and run its kthreads.
2212
-
In theory, it would be possible to invoke callbacks earlier,
2213
-
however, this is not a panacea because there would be severe restrictions
2214
-
on what operations those callbacks could invoke.
2215
-
2216
-
<p>
2217
-
Perhaps surprisingly, <tt>synchronize_rcu()</tt> and
2218
-
<tt>synchronize_rcu_expedited()</tt>,
2219
-
will operate normally
2220
-
during very early boot, the reason being that there is only one CPU
2221
-
and preemption is disabled.
2222
-
This means that the call <tt>synchronize_rcu()</tt> (or friends)
2223
-
itself is a quiescent
2224
-
state and thus a grace period, so the early-boot implementation can
2225
-
be a no-op.
2226
-
2227
-
<p>
2228
-
However, once the scheduler has spawned its first kthread, this early
2229
-
boot trick fails for <tt>synchronize_rcu()</tt> (as well as for
2230
-
<tt>synchronize_rcu_expedited()</tt>) in <tt>CONFIG_PREEMPT=y</tt>
2231
-
kernels.
2232
-
The reason is that an RCU read-side critical section might be preempted,
2233
-
which means that a subsequent <tt>synchronize_rcu()</tt> really does have
2234
-
to wait for something, as opposed to simply returning immediately.
2235
-
Unfortunately, <tt>synchronize_rcu()</tt> can't do this until all of
2236
-
its kthreads are spawned, which doesn't happen until some time during
2237
-
<tt>early_initcalls()</tt> time.
2238
-
But this is no excuse: RCU is nevertheless required to correctly handle
2239
-
synchronous grace periods during this time period.
2240
-
Once all of its kthreads are up and running, RCU starts running
2241
-
normally.
2242
-
2243
-
<table>
2244
-
<tr><th> </th></tr>
2245
-
<tr><th align="left">Quick Quiz:</th></tr>
2246
-
<tr><td>
2247
-
How can RCU possibly handle grace periods before all of its
2248
-
kthreads have been spawned???
2249
-
</td></tr>
2250
-
<tr><th align="left">Answer:</th></tr>
2251
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
2252
-
Very carefully!
2253
-
</font>
2254
-
2255
-
<p><font color="ffffff">
2256
-
During the “dead zone” between the time that the
2257
-
scheduler spawns the first task and the time that all of RCU's
2258
-
kthreads have been spawned, all synchronous grace periods are
2259
-
handled by the expedited grace-period mechanism.
2260
-
At runtime, this expedited mechanism relies on workqueues, but
2261
-
during the dead zone the requesting task itself drives the
2262
-
desired expedited grace period.
2263
-
Because dead-zone execution takes place within task context,
2264
-
everything works.
2265
-
Once the dead zone ends, expedited grace periods go back to
2266
-
using workqueues, as is required to avoid problems that would
2267
-
otherwise occur when a user task received a POSIX signal while
2268
-
driving an expedited grace period.
2269
-
</font>
2270
-
2271
-
<p><font color="ffffff">
2272
-
And yes, this does mean that it is unhelpful to send POSIX
2273
-
signals to random tasks between the time that the scheduler
2274
-
spawns its first kthread and the time that RCU's kthreads
2275
-
have all been spawned.
2276
-
If there ever turns out to be a good reason for sending POSIX
2277
-
signals during that time, appropriate adjustments will be made.
2278
-
(If it turns out that POSIX signals are sent during this time for
2279
-
no good reason, other adjustments will be made, appropriate
2280
-
or otherwise.)
2281
-
</font></td></tr>
2282
-
<tr><td> </td></tr>
2283
-
</table>
2284
-
2285
-
<p>
2286
-
I learned of these boot-time requirements as a result of a series of
2287
-
system hangs.
2288
-
2289
-
<h3><a name="Interrupts and NMIs">Interrupts and NMIs</a></h3>
2290
-
2291
-
<p>
2292
-
The Linux kernel has interrupts, and RCU read-side critical sections are
2293
-
legal within interrupt handlers and within interrupt-disabled regions
2294
-
of code, as are invocations of <tt>call_rcu()</tt>.
2295
-
2296
-
<p>
2297
-
Some Linux-kernel architectures can enter an interrupt handler from
2298
-
non-idle process context, and then just never leave it, instead stealthily
2299
-
transitioning back to process context.
2300
-
This trick is sometimes used to invoke system calls from inside the kernel.
2301
-
These “half-interrupts” mean that RCU has to be very careful
2302
-
about how it counts interrupt nesting levels.
2303
-
I learned of this requirement the hard way during a rewrite
2304
-
of RCU's dyntick-idle code.
2305
-
2306
-
<p>
2307
-
The Linux kernel has non-maskable interrupts (NMIs), and
2308
-
RCU read-side critical sections are legal within NMI handlers.
2309
-
Thankfully, RCU update-side primitives, including
2310
-
<tt>call_rcu()</tt>, are prohibited within NMI handlers.
2311
-
2312
-
<p>
2313
-
The name notwithstanding, some Linux-kernel architectures
2314
-
can have nested NMIs, which RCU must handle correctly.
2315
-
Andy Lutomirski
2316
-
<a href="https://lkml.kernel.org/r/CALCETrXLq1y7e_dKFPgou-FKHB6Pu-r8+t-6Ds+8=va7anBWDA@mail.gmail.com">surprised me</a>
2317
-
with this requirement;
2318
-
he also kindly surprised me with
2319
-
<a href="https://lkml.kernel.org/r/CALCETrXSY9JpW3uE6H8WYk81sg56qasA2aqmjMPsq5dOtzso=g@mail.gmail.com">an algorithm</a>
2320
-
that meets this requirement.
2321
-
2322
-
<p>
2323
-
Furthermore, NMI handlers can be interrupted by what appear to RCU
2324
-
to be normal interrupts.
2325
-
One way that this can happen is for code that directly invokes
2326
-
<tt>rcu_irq_enter()</tt> and <tt>rcu_irq_exit()</tt> to be called
2327
-
from an NMI handler.
2328
-
This astonishing fact of life prompted the current code structure,
2329
-
which has <tt>rcu_irq_enter()</tt> invoking <tt>rcu_nmi_enter()</tt>
2330
-
and <tt>rcu_irq_exit()</tt> invoking <tt>rcu_nmi_exit()</tt>.
2331
-
And yes, I also learned of this requirement the hard way.
2332
-
2333
-
<h3><a name="Loadable Modules">Loadable Modules</a></h3>
2334
-
2335
-
<p>
2336
-
The Linux kernel has loadable modules, and these modules can
2337
-
also be unloaded.
2338
-
After a given module has been unloaded, any attempt to call
2339
-
one of its functions results in a segmentation fault.
2340
-
The module-unload functions must therefore cancel any
2341
-
delayed calls to loadable-module functions, for example,
2342
-
any outstanding <tt>mod_timer()</tt> must be dealt with
2343
-
via <tt>del_timer_sync()</tt> or similar.
2344
-
2345
-
<p>
2346
-
Unfortunately, there is no way to cancel an RCU callback;
2347
-
once you invoke <tt>call_rcu()</tt>, the callback function is
2348
-
eventually going to be invoked, unless the system goes down first.
2349
-
Because it is normally considered socially irresponsible to crash the system
2350
-
in response to a module unload request, we need some other way
2351
-
to deal with in-flight RCU callbacks.
2352
-
2353
-
<p>
2354
-
RCU therefore provides
2355
-
<tt><a href="https://lwn.net/Articles/217484/">rcu_barrier()</a></tt>,
2356
-
which waits until all in-flight RCU callbacks have been invoked.
2357
-
If a module uses <tt>call_rcu()</tt>, its exit function should therefore
2358
-
prevent any future invocation of <tt>call_rcu()</tt>, then invoke
2359
-
<tt>rcu_barrier()</tt>.
2360
-
In theory, the underlying module-unload code could invoke
2361
-
<tt>rcu_barrier()</tt> unconditionally, but in practice this would
2362
-
incur unacceptable latencies.
2363
-
2364
-
<p>
2365
-
Nikita Danilov noted this requirement for an analogous filesystem-unmount
2366
-
situation, and Dipankar Sarma incorporated <tt>rcu_barrier()</tt> into RCU.
2367
-
The need for <tt>rcu_barrier()</tt> for module unloading became
2368
-
apparent later.
2369
-
2370
-
<p>
2371
-
<b>Important note</b>: The <tt>rcu_barrier()</tt> function is not,
2372
-
repeat, <i>not</i>, obligated to wait for a grace period.
2373
-
It is instead only required to wait for RCU callbacks that have
2374
-
already been posted.
2375
-
Therefore, if there are no RCU callbacks posted anywhere in the system,
2376
-
<tt>rcu_barrier()</tt> is within its rights to return immediately.
2377
-
Even if there are callbacks posted, <tt>rcu_barrier()</tt> does not
2378
-
necessarily need to wait for a grace period.
2379
-
2380
-
<table>
2381
-
<tr><th> </th></tr>
2382
-
<tr><th align="left">Quick Quiz:</th></tr>
2383
-
<tr><td>
2384
-
Wait a minute!
2385
-
Each RCU callbacks must wait for a grace period to complete,
2386
-
and <tt>rcu_barrier()</tt> must wait for each pre-existing
2387
-
callback to be invoked.
2388
-
Doesn't <tt>rcu_barrier()</tt> therefore need to wait for
2389
-
a full grace period if there is even one callback posted anywhere
2390
-
in the system?
2391
-
</td></tr>
2392
-
<tr><th align="left">Answer:</th></tr>
2393
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
2394
-
Absolutely not!!!
2395
-
</font>
2396
-
2397
-
<p><font color="ffffff">
2398
-
Yes, each RCU callbacks must wait for a grace period to complete,
2399
-
but it might well be partly (or even completely) finished waiting
2400
-
by the time <tt>rcu_barrier()</tt> is invoked.
2401
-
In that case, <tt>rcu_barrier()</tt> need only wait for the
2402
-
remaining portion of the grace period to elapse.
2403
-
So even if there are quite a few callbacks posted,
2404
-
<tt>rcu_barrier()</tt> might well return quite quickly.
2405
-
</font>
2406
-
2407
-
<p><font color="ffffff">
2408
-
So if you need to wait for a grace period as well as for all
2409
-
pre-existing callbacks, you will need to invoke both
2410
-
<tt>synchronize_rcu()</tt> and <tt>rcu_barrier()</tt>.
2411
-
If latency is a concern, you can always use workqueues
2412
-
to invoke them concurrently.
2413
-
</font></td></tr>
2414
-
<tr><td> </td></tr>
2415
-
</table>
2416
-
2417
-
<h3><a name="Hotplug CPU">Hotplug CPU</a></h3>
2418
-
2419
-
<p>
2420
-
The Linux kernel supports CPU hotplug, which means that CPUs
2421
-
can come and go.
2422
-
It is of course illegal to use any RCU API member from an offline CPU,
2423
-
with the exception of <a href="#Sleepable RCU">SRCU</a> read-side
2424
-
critical sections.
2425
-
This requirement was present from day one in DYNIX/ptx, but
2426
-
on the other hand, the Linux kernel's CPU-hotplug implementation
2427
-
is “interesting.”
2428
-
2429
-
<p>
2430
-
The Linux-kernel CPU-hotplug implementation has notifiers that
2431
-
are used to allow the various kernel subsystems (including RCU)
2432
-
to respond appropriately to a given CPU-hotplug operation.
2433
-
Most RCU operations may be invoked from CPU-hotplug notifiers,
2434
-
including even synchronous grace-period operations such as
2435
-
<tt>synchronize_rcu()</tt> and <tt>synchronize_rcu_expedited()</tt>.
2436
-
2437
-
<p>
2438
-
However, all-callback-wait operations such as
2439
-
<tt>rcu_barrier()</tt> are also not supported, due to the
2440
-
fact that there are phases of CPU-hotplug operations where
2441
-
the outgoing CPU's callbacks will not be invoked until after
2442
-
the CPU-hotplug operation ends, which could also result in deadlock.
2443
-
Furthermore, <tt>rcu_barrier()</tt> blocks CPU-hotplug operations
2444
-
during its execution, which results in another type of deadlock
2445
-
when invoked from a CPU-hotplug notifier.
2446
-
2447
-
<h3><a name="Scheduler and RCU">Scheduler and RCU</a></h3>
2448
-
2449
-
<p>
2450
-
RCU depends on the scheduler, and the scheduler uses RCU to
2451
-
protect some of its data structures.
2452
-
The preemptible-RCU <tt>rcu_read_unlock()</tt>
2453
-
implementation must therefore be written carefully to avoid deadlocks
2454
-
involving the scheduler's runqueue and priority-inheritance locks.
2455
-
In particular, <tt>rcu_read_unlock()</tt> must tolerate an
2456
-
interrupt where the interrupt handler invokes both
2457
-
<tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>.
2458
-
This possibility requires <tt>rcu_read_unlock()</tt> to use
2459
-
negative nesting levels to avoid destructive recursion via
2460
-
interrupt handler's use of RCU.
2461
-
2462
-
<p>
2463
-
This scheduler-RCU requirement came as a
2464
-
<a href="https://lwn.net/Articles/453002/">complete surprise</a>.
2465
-
2466
-
<p>
2467
-
As noted above, RCU makes use of kthreads, and it is necessary to
2468
-
avoid excessive CPU-time accumulation by these kthreads.
2469
-
This requirement was no surprise, but RCU's violation of it
2470
-
when running context-switch-heavy workloads when built with
2471
-
<tt>CONFIG_NO_HZ_FULL=y</tt>
2472
-
<a href="http://www.rdrop.com/users/paulmck/scalability/paper/BareMetal.2015.01.15b.pdf">did come as a surprise [PDF]</a>.
2473
-
RCU has made good progress towards meeting this requirement, even
2474
-
for context-switch-heavy <tt>CONFIG_NO_HZ_FULL=y</tt> workloads,
2475
-
but there is room for further improvement.
2476
-
2477
-
<p>
2478
-
It is forbidden to hold any of scheduler's runqueue or priority-inheritance
2479
-
spinlocks across an <tt>rcu_read_unlock()</tt> unless interrupts have been
2480
-
disabled across the entire RCU read-side critical section, that is,
2481
-
up to and including the matching <tt>rcu_read_lock()</tt>.
2482
-
Violating this restriction can result in deadlocks involving these
2483
-
scheduler spinlocks.
2484
-
There was hope that this restriction might be lifted when interrupt-disabled
2485
-
calls to <tt>rcu_read_unlock()</tt> started deferring the reporting of
2486
-
the resulting RCU-preempt quiescent state until the end of the corresponding
2487
-
interrupts-disabled region.
2488
-
Unfortunately, timely reporting of the corresponding quiescent state
2489
-
to expedited grace periods requires a call to <tt>raise_softirq()</tt>,
2490
-
which can acquire these scheduler spinlocks.
2491
-
In addition, real-time systems using RCU priority boosting
2492
-
need this restriction to remain in effect because deferred
2493
-
quiescent-state reporting would also defer deboosting, which in turn
2494
-
would degrade real-time latencies.
2495
-
2496
-
<p>
2497
-
In theory, if a given RCU read-side critical section could be
2498
-
guaranteed to be less than one second in duration, holding a scheduler
2499
-
spinlock across that critical section's <tt>rcu_read_unlock()</tt>
2500
-
would require only that preemption be disabled across the entire
2501
-
RCU read-side critical section, not interrupts.
2502
-
Unfortunately, given the possibility of vCPU preemption, long-running
2503
-
interrupts, and so on, it is not possible in practice to guarantee
2504
-
that a given RCU read-side critical section will complete in less than
2505
-
one second.
2506
-
Therefore, as noted above, if scheduler spinlocks are held across
2507
-
a given call to <tt>rcu_read_unlock()</tt>, interrupts must be
2508
-
disabled across the entire RCU read-side critical section.
2509
-
2510
-
<h3><a name="Tracing and RCU">Tracing and RCU</a></h3>
2511
-
2512
-
<p>
2513
-
It is possible to use tracing on RCU code, but tracing itself
2514
-
uses RCU.
2515
-
For this reason, <tt>rcu_dereference_raw_notrace()</tt>
2516
-
is provided for use by tracing, which avoids the destructive
2517
-
recursion that could otherwise ensue.
2518
-
This API is also used by virtualization in some architectures,
2519
-
where RCU readers execute in environments in which tracing
2520
-
cannot be used.
2521
-
The tracing folks both located the requirement and provided the
2522
-
needed fix, so this surprise requirement was relatively painless.
2523
-
2524
-
<h3><a name="Energy Efficiency">Energy Efficiency</a></h3>
2525
-
2526
-
<p>
2527
-
Interrupting idle CPUs is considered socially unacceptable,
2528
-
especially by people with battery-powered embedded systems.
2529
-
RCU therefore conserves energy by detecting which CPUs are
2530
-
idle, including tracking CPUs that have been interrupted from idle.
2531
-
This is a large part of the energy-efficiency requirement,
2532
-
so I learned of this via an irate phone call.
2533
-
2534
-
<p>
2535
-
Because RCU avoids interrupting idle CPUs, it is illegal to
2536
-
execute an RCU read-side critical section on an idle CPU.
2537
-
(Kernels built with <tt>CONFIG_PROVE_RCU=y</tt> will splat
2538
-
if you try it.)
2539
-
The <tt>RCU_NONIDLE()</tt> macro and <tt>_rcuidle</tt>
2540
-
event tracing is provided to work around this restriction.
2541
-
In addition, <tt>rcu_is_watching()</tt> may be used to
2542
-
test whether or not it is currently legal to run RCU read-side
2543
-
critical sections on this CPU.
2544
-
I learned of the need for diagnostics on the one hand
2545
-
and <tt>RCU_NONIDLE()</tt> on the other while inspecting
2546
-
idle-loop code.
2547
-
Steven Rostedt supplied <tt>_rcuidle</tt> event tracing,
2548
-
which is used quite heavily in the idle loop.
2549
-
However, there are some restrictions on the code placed within
2550
-
<tt>RCU_NONIDLE()</tt>:
2551
-
2552
-
<ol>
2553
-
<li> Blocking is prohibited.
2554
-
In practice, this is not a serious restriction given that idle
2555
-
tasks are prohibited from blocking to begin with.
2556
-
<li> Although nesting <tt>RCU_NONIDLE()</tt> is permitted, they cannot
2557
-
nest indefinitely deeply.
2558
-
However, given that they can be nested on the order of a million
2559
-
deep, even on 32-bit systems, this should not be a serious
2560
-
restriction.
2561
-
This nesting limit would probably be reached long after the
2562
-
compiler OOMed or the stack overflowed.
2563
-
<li> Any code path that enters <tt>RCU_NONIDLE()</tt> must sequence
2564
-
out of that same <tt>RCU_NONIDLE()</tt>.
2565
-
For example, the following is grossly illegal:
2566
-
2567
-
<blockquote>
2568
-
<pre>
2569
-
1 RCU_NONIDLE({
2570
-
2 do_something();
2571
-
3 goto bad_idea; /* BUG!!! */
2572
-
4 do_something_else();});
2573
-
5 bad_idea:
2574
-
</pre>
2575
-
</blockquote>
2576
-
2577
-
<p>
2578
-
It is just as illegal to transfer control into the middle of
2579
-
<tt>RCU_NONIDLE()</tt>'s argument.
2580
-
Yes, in theory, you could transfer in as long as you also
2581
-
transferred out, but in practice you could also expect to get sharply
2582
-
worded review comments.
2583
-
</ol>
2584
-
2585
-
<p>
2586
-
It is similarly socially unacceptable to interrupt an
2587
-
<tt>nohz_full</tt> CPU running in userspace.
2588
-
RCU must therefore track <tt>nohz_full</tt> userspace
2589
-
execution.
2590
-
RCU must therefore be able to sample state at two points in
2591
-
time, and be able to determine whether or not some other CPU spent
2592
-
any time idle and/or executing in userspace.
2593
-
2594
-
<p>
2595
-
These energy-efficiency requirements have proven quite difficult to
2596
-
understand and to meet, for example, there have been more than five
2597
-
clean-sheet rewrites of RCU's energy-efficiency code, the last of
2598
-
which was finally able to demonstrate
2599
-
<a href="http://www.rdrop.com/users/paulmck/realtime/paper/AMPenergy.2013.04.19a.pdf">real energy savings running on real hardware [PDF]</a>.
2600
-
As noted earlier,
2601
-
I learned of many of these requirements via angry phone calls:
2602
-
Flaming me on the Linux-kernel mailing list was apparently not
2603
-
sufficient to fully vent their ire at RCU's energy-efficiency bugs!
2604
-
2605
-
<h3><a name="Scheduling-Clock Interrupts and RCU">
2606
-
Scheduling-Clock Interrupts and RCU</a></h3>
2607
-
2608
-
<p>
2609
-
The kernel transitions between in-kernel non-idle execution, userspace
2610
-
execution, and the idle loop.
2611
-
Depending on kernel configuration, RCU handles these states differently:
2612
-
2613
-
<table border=3>
2614
-
<tr><th><tt>HZ</tt> Kconfig</th>
2615
-
<th>In-Kernel</th>
2616
-
<th>Usermode</th>
2617
-
<th>Idle</th></tr>
2618
-
<tr><th align="left"><tt>HZ_PERIODIC</tt></th>
2619
-
<td>Can rely on scheduling-clock interrupt.</td>
2620
-
<td>Can rely on scheduling-clock interrupt and its
2621
-
detection of interrupt from usermode.</td>
2622
-
<td>Can rely on RCU's dyntick-idle detection.</td></tr>
2623
-
<tr><th align="left"><tt>NO_HZ_IDLE</tt></th>
2624
-
<td>Can rely on scheduling-clock interrupt.</td>
2625
-
<td>Can rely on scheduling-clock interrupt and its
2626
-
detection of interrupt from usermode.</td>
2627
-
<td>Can rely on RCU's dyntick-idle detection.</td></tr>
2628
-
<tr><th align="left"><tt>NO_HZ_FULL</tt></th>
2629
-
<td>Can only sometimes rely on scheduling-clock interrupt.
2630
-
In other cases, it is necessary to bound kernel execution
2631
-
times and/or use IPIs.</td>
2632
-
<td>Can rely on RCU's dyntick-idle detection.</td>
2633
-
<td>Can rely on RCU's dyntick-idle detection.</td></tr>
2634
-
</table>
2635
-
2636
-
<table>
2637
-
<tr><th> </th></tr>
2638
-
<tr><th align="left">Quick Quiz:</th></tr>
2639
-
<tr><td>
2640
-
Why can't <tt>NO_HZ_FULL</tt> in-kernel execution rely on the
2641
-
scheduling-clock interrupt, just like <tt>HZ_PERIODIC</tt>
2642
-
and <tt>NO_HZ_IDLE</tt> do?
2643
-
</td></tr>
2644
-
<tr><th align="left">Answer:</th></tr>
2645
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
2646
-
Because, as a performance optimization, <tt>NO_HZ_FULL</tt>
2647
-
does not necessarily re-enable the scheduling-clock interrupt
2648
-
on entry to each and every system call.
2649
-
</font></td></tr>
2650
-
<tr><td> </td></tr>
2651
-
</table>
2652
-
2653
-
<p>
2654
-
However, RCU must be reliably informed as to whether any given
2655
-
CPU is currently in the idle loop, and, for <tt>NO_HZ_FULL</tt>,
2656
-
also whether that CPU is executing in usermode, as discussed
2657
-
<a href="#Energy Efficiency">earlier</a>.
2658
-
It also requires that the scheduling-clock interrupt be enabled when
2659
-
RCU needs it to be:
2660
-
2661
-
<ol>
2662
-
<li> If a CPU is either idle or executing in usermode, and RCU believes
2663
-
it is non-idle, the scheduling-clock tick had better be running.
2664
-
Otherwise, you will get RCU CPU stall warnings. Or at best,
2665
-
very long (11-second) grace periods, with a pointless IPI waking
2666
-
the CPU from time to time.
2667
-
<li> If a CPU is in a portion of the kernel that executes RCU read-side
2668
-
critical sections, and RCU believes this CPU to be idle, you will get
2669
-
random memory corruption. <b>DON'T DO THIS!!!</b>
2670
-
2671
-
<br>This is one reason to test with lockdep, which will complain
2672
-
about this sort of thing.
2673
-
<li> If a CPU is in a portion of the kernel that is absolutely
2674
-
positively no-joking guaranteed to never execute any RCU read-side
2675
-
critical sections, and RCU believes this CPU to to be idle,
2676
-
no problem. This sort of thing is used by some architectures
2677
-
for light-weight exception handlers, which can then avoid the
2678
-
overhead of <tt>rcu_irq_enter()</tt> and <tt>rcu_irq_exit()</tt>
2679
-
at exception entry and exit, respectively.
2680
-
Some go further and avoid the entireties of <tt>irq_enter()</tt>
2681
-
and <tt>irq_exit()</tt>.
2682
-
2683
-
<br>Just make very sure you are running some of your tests with
2684
-
<tt>CONFIG_PROVE_RCU=y</tt>, just in case one of your code paths
2685
-
was in fact joking about not doing RCU read-side critical sections.
2686
-
<li> If a CPU is executing in the kernel with the scheduling-clock
2687
-
interrupt disabled and RCU believes this CPU to be non-idle,
2688
-
and if the CPU goes idle (from an RCU perspective) every few
2689
-
jiffies, no problem. It is usually OK for there to be the
2690
-
occasional gap between idle periods of up to a second or so.
2691
-
2692
-
<br>If the gap grows too long, you get RCU CPU stall warnings.
2693
-
<li> If a CPU is either idle or executing in usermode, and RCU believes
2694
-
it to be idle, of course no problem.
2695
-
<li> If a CPU is executing in the kernel, the kernel code
2696
-
path is passing through quiescent states at a reasonable
2697
-
frequency (preferably about once per few jiffies, but the
2698
-
occasional excursion to a second or so is usually OK) and the
2699
-
scheduling-clock interrupt is enabled, of course no problem.
2700
-
2701
-
<br>If the gap between a successive pair of quiescent states grows
2702
-
too long, you get RCU CPU stall warnings.
2703
-
</ol>
2704
-
2705
-
<table>
2706
-
<tr><th> </th></tr>
2707
-
<tr><th align="left">Quick Quiz:</th></tr>
2708
-
<tr><td>
2709
-
But what if my driver has a hardware interrupt handler
2710
-
that can run for many seconds?
2711
-
I cannot invoke <tt>schedule()</tt> from an hardware
2712
-
interrupt handler, after all!
2713
-
</td></tr>
2714
-
<tr><th align="left">Answer:</th></tr>
2715
-
<tr><td bgcolor="#ffffff"><font color="ffffff">
2716
-
One approach is to do <tt>rcu_irq_exit();rcu_irq_enter();</tt>
2717
-
every so often.
2718
-
But given that long-running interrupt handlers can cause
2719
-
other problems, not least for response time, shouldn't you
2720
-
work to keep your interrupt handler's runtime within reasonable
2721
-
bounds?
2722
-
</font></td></tr>
2723
-
<tr><td> </td></tr>
2724
-
</table>
2725
-
2726
-
<p>
2727
-
But as long as RCU is properly informed of kernel state transitions between
2728
-
in-kernel execution, usermode execution, and idle, and as long as the
2729
-
scheduling-clock interrupt is enabled when RCU needs it to be, you
2730
-
can rest assured that the bugs you encounter will be in some other
2731
-
part of RCU or some other part of the kernel!
2732
-
2733
-
<h3><a name="Memory Efficiency">Memory Efficiency</a></h3>
2734
-
2735
-
<p>
2736
-
Although small-memory non-realtime systems can simply use Tiny RCU,
2737
-
code size is only one aspect of memory efficiency.
2738
-
Another aspect is the size of the <tt>rcu_head</tt> structure
2739
-
used by <tt>call_rcu()</tt> and <tt>kfree_rcu()</tt>.
2740
-
Although this structure contains nothing more than a pair of pointers,
2741
-
it does appear in many RCU-protected data structures, including
2742
-
some that are size critical.
2743
-
The <tt>page</tt> structure is a case in point, as evidenced by
2744
-
the many occurrences of the <tt>union</tt> keyword within that structure.
2745
-
2746
-
<p>
2747
-
This need for memory efficiency is one reason that RCU uses hand-crafted
2748
-
singly linked lists to track the <tt>rcu_head</tt> structures that
2749
-
are waiting for a grace period to elapse.
2750
-
It is also the reason why <tt>rcu_head</tt> structures do not contain
2751
-
debug information, such as fields tracking the file and line of the
2752
-
<tt>call_rcu()</tt> or <tt>kfree_rcu()</tt> that posted them.
2753
-
Although this information might appear in debug-only kernel builds at some
2754
-
point, in the meantime, the <tt>->func</tt> field will often provide
2755
-
the needed debug information.
2756
-
2757
-
<p>
2758
-
However, in some cases, the need for memory efficiency leads to even
2759
-
more extreme measures.
2760
-
Returning to the <tt>page</tt> structure, the <tt>rcu_head</tt> field
2761
-
shares storage with a great many other structures that are used at
2762
-
various points in the corresponding page's lifetime.
2763
-
In order to correctly resolve certain
2764
-
<a href="https://lkml.kernel.org/g/1439976106-137226-1-git-send-email-kirill.shutemov@linux.intel.com">race conditions</a>,
2765
-
the Linux kernel's memory-management subsystem needs a particular bit
2766
-
to remain zero during all phases of grace-period processing,
2767
-
and that bit happens to map to the bottom bit of the
2768
-
<tt>rcu_head</tt> structure's <tt>->next</tt> field.
2769
-
RCU makes this guarantee as long as <tt>call_rcu()</tt>
2770
-
is used to post the callback, as opposed to <tt>kfree_rcu()</tt>
2771
-
or some future “lazy”
2772
-
variant of <tt>call_rcu()</tt> that might one day be created for
2773
-
energy-efficiency purposes.
2774
-
2775
-
<p>
2776
-
That said, there are limits.
2777
-
RCU requires that the <tt>rcu_head</tt> structure be aligned to a
2778
-
two-byte boundary, and passing a misaligned <tt>rcu_head</tt>
2779
-
structure to one of the <tt>call_rcu()</tt> family of functions
2780
-
will result in a splat.
2781
-
It is therefore necessary to exercise caution when packing
2782
-
structures containing fields of type <tt>rcu_head</tt>.
2783
-
Why not a four-byte or even eight-byte alignment requirement?
2784
-
Because the m68k architecture provides only two-byte alignment,
2785
-
and thus acts as alignment's least common denominator.
2786
-
2787
-
<p>
2788
-
The reason for reserving the bottom bit of pointers to
2789
-
<tt>rcu_head</tt> structures is to leave the door open to
2790
-
“lazy” callbacks whose invocations can safely be deferred.
2791
-
Deferring invocation could potentially have energy-efficiency
2792
-
benefits, but only if the rate of non-lazy callbacks decreases
2793
-
significantly for some important workload.
2794
-
In the meantime, reserving the bottom bit keeps this option open
2795
-
in case it one day becomes useful.
2796
-
2797
-
<h3><a name="Performance, Scalability, Response Time, and Reliability">
2798
-
Performance, Scalability, Response Time, and Reliability</a></h3>
2799
-
2800
-
<p>
2801
-
Expanding on the
2802
-
<a href="#Performance and Scalability">earlier discussion</a>,
2803
-
RCU is used heavily by hot code paths in performance-critical
2804
-
portions of the Linux kernel's networking, security, virtualization,
2805
-
and scheduling code paths.
2806
-
RCU must therefore use efficient implementations, especially in its
2807
-
read-side primitives.
2808
-
To that end, it would be good if preemptible RCU's implementation
2809
-
of <tt>rcu_read_lock()</tt> could be inlined, however, doing
2810
-
this requires resolving <tt>#include</tt> issues with the
2811
-
<tt>task_struct</tt> structure.
2812
-
2813
-
<p>
2814
-
The Linux kernel supports hardware configurations with up to
2815
-
4096 CPUs, which means that RCU must be extremely scalable.
2816
-
Algorithms that involve frequent acquisitions of global locks or
2817
-
frequent atomic operations on global variables simply cannot be
2818
-
tolerated within the RCU implementation.
2819
-
RCU therefore makes heavy use of a combining tree based on the
2820
-
<tt>rcu_node</tt> structure.
2821
-
RCU is required to tolerate all CPUs continuously invoking any
2822
-
combination of RCU's runtime primitives with minimal per-operation
2823
-
overhead.
2824
-
In fact, in many cases, increasing load must <i>decrease</i> the
2825
-
per-operation overhead, witness the batching optimizations for
2826
-
<tt>synchronize_rcu()</tt>, <tt>call_rcu()</tt>,
2827
-
<tt>synchronize_rcu_expedited()</tt>, and <tt>rcu_barrier()</tt>.
2828
-
As a general rule, RCU must cheerfully accept whatever the
2829
-
rest of the Linux kernel decides to throw at it.
2830
-
2831
-
<p>
2832
-
The Linux kernel is used for real-time workloads, especially
2833
-
in conjunction with the
2834
-
<a href="https://rt.wiki.kernel.org/index.php/Main_Page">-rt patchset</a>.
2835
-
The real-time-latency response requirements are such that the
2836
-
traditional approach of disabling preemption across RCU
2837
-
read-side critical sections is inappropriate.
2838
-
Kernels built with <tt>CONFIG_PREEMPT=y</tt> therefore
2839
-
use an RCU implementation that allows RCU read-side critical
2840
-
sections to be preempted.
2841
-
This requirement made its presence known after users made it
2842
-
clear that an earlier
2843
-
<a href="https://lwn.net/Articles/107930/">real-time patch</a>
2844
-
did not meet their needs, in conjunction with some
2845
-
<a href="https://lkml.kernel.org/g/20050318002026.GA2693@us.ibm.com">RCU issues</a>
2846
-
encountered by a very early version of the -rt patchset.
2847
-
2848
-
<p>
2849
-
In addition, RCU must make do with a sub-100-microsecond real-time latency
2850
-
budget.
2851
-
In fact, on smaller systems with the -rt patchset, the Linux kernel
2852
-
provides sub-20-microsecond real-time latencies for the whole kernel,
2853
-
including RCU.
2854
-
RCU's scalability and latency must therefore be sufficient for
2855
-
these sorts of configurations.
2856
-
To my surprise, the sub-100-microsecond real-time latency budget
2857
-
<a href="http://www.rdrop.com/users/paulmck/realtime/paper/bigrt.2013.01.31a.LCA.pdf">
2858
-
applies to even the largest systems [PDF]</a>,
2859
-
up to and including systems with 4096 CPUs.
2860
-
This real-time requirement motivated the grace-period kthread, which
2861
-
also simplified handling of a number of race conditions.
2862
-
2863
-
<p>
2864
-
RCU must avoid degrading real-time response for CPU-bound threads, whether
2865
-
executing in usermode (which is one use case for
2866
-
<tt>CONFIG_NO_HZ_FULL=y</tt>) or in the kernel.
2867
-
That said, CPU-bound loops in the kernel must execute
2868
-
<tt>cond_resched()</tt> at least once per few tens of milliseconds
2869
-
in order to avoid receiving an IPI from RCU.
2870
-
2871
-
<p>
2872
-
Finally, RCU's status as a synchronization primitive means that
2873
-
any RCU failure can result in arbitrary memory corruption that can be
2874
-
extremely difficult to debug.
2875
-
This means that RCU must be extremely reliable, which in
2876
-
practice also means that RCU must have an aggressive stress-test
2877
-
suite.
2878
-
This stress-test suite is called <tt>rcutorture</tt>.
2879
-
2880
-
<p>
2881
-
Although the need for <tt>rcutorture</tt> was no surprise,
2882
-
the current immense popularity of the Linux kernel is posing
2883
-
interesting—and perhaps unprecedented—validation
2884
-
challenges.
2885
-
To see this, keep in mind that there are well over one billion
2886
-
instances of the Linux kernel running today, given Android
2887
-
smartphones, Linux-powered televisions, and servers.
2888
-
This number can be expected to increase sharply with the advent of
2889
-
the celebrated Internet of Things.
2890
-
2891
-
<p>
2892
-
Suppose that RCU contains a race condition that manifests on average
2893
-
once per million years of runtime.
2894
-
This bug will be occurring about three times per <i>day</i> across
2895
-
the installed base.
2896
-
RCU could simply hide behind hardware error rates, given that no one
2897
-
should really expect their smartphone to last for a million years.
2898
-
However, anyone taking too much comfort from this thought should
2899
-
consider the fact that in most jurisdictions, a successful multi-year
2900
-
test of a given mechanism, which might include a Linux kernel,
2901
-
suffices for a number of types of safety-critical certifications.
2902
-
In fact, rumor has it that the Linux kernel is already being used
2903
-
in production for safety-critical applications.
2904
-
I don't know about you, but I would feel quite bad if a bug in RCU
2905
-
killed someone.
2906
-
Which might explain my recent focus on validation and verification.
2907
-
2908
-
<h2><a name="Other RCU Flavors">Other RCU Flavors</a></h2>
2909
-
2910
-
<p>
2911
-
One of the more surprising things about RCU is that there are now
2912
-
no fewer than five <i>flavors</i>, or API families.
2913
-
In addition, the primary flavor that has been the sole focus up to
2914
-
this point has two different implementations, non-preemptible and
2915
-
preemptible.
2916
-
The other four flavors are listed below, with requirements for each
2917
-
described in a separate section.
2918
-
2919
-
<ol>
2920
-
<li> <a href="#Bottom-Half Flavor">Bottom-Half Flavor (Historical)</a>
2921
-
<li> <a href="#Sched Flavor">Sched Flavor (Historical)</a>
2922
-
<li> <a href="#Sleepable RCU">Sleepable RCU</a>
2923
-
<li> <a href="#Tasks RCU">Tasks RCU</a>
2924
-
</ol>
2925
-
2926
-
<h3><a name="Bottom-Half Flavor">Bottom-Half Flavor (Historical)</a></h3>
2927
-
2928
-
<p>
2929
-
The RCU-bh flavor of RCU has since been expressed in terms of
2930
-
the other RCU flavors as part of a consolidation of the three
2931
-
flavors into a single flavor.
2932
-
The read-side API remains, and continues to disable softirq and to
2933
-
be accounted for by lockdep.
2934
-
Much of the material in this section is therefore strictly historical
2935
-
in nature.
2936
-
2937
-
<p>
2938
-
The softirq-disable (AKA “bottom-half”,
2939
-
hence the “_bh” abbreviations)
2940
-
flavor of RCU, or <i>RCU-bh</i>, was developed by
2941
-
Dipankar Sarma to provide a flavor of RCU that could withstand the
2942
-
network-based denial-of-service attacks researched by Robert
2943
-
Olsson.
2944
-
These attacks placed so much networking load on the system
2945
-
that some of the CPUs never exited softirq execution,
2946
-
which in turn prevented those CPUs from ever executing a context switch,
2947
-
which, in the RCU implementation of that time, prevented grace periods
2948
-
from ever ending.
2949
-
The result was an out-of-memory condition and a system hang.
2950
-
2951
-
<p>
2952
-
The solution was the creation of RCU-bh, which does
2953
-
<tt>local_bh_disable()</tt>
2954
-
across its read-side critical sections, and which uses the transition
2955
-
from one type of softirq processing to another as a quiescent state
2956
-
in addition to context switch, idle, user mode, and offline.
2957
-
This means that RCU-bh grace periods can complete even when some of
2958
-
the CPUs execute in softirq indefinitely, thus allowing algorithms
2959
-
based on RCU-bh to withstand network-based denial-of-service attacks.
2960
-
2961
-
<p>
2962
-
Because
2963
-
<tt>rcu_read_lock_bh()</tt> and <tt>rcu_read_unlock_bh()</tt>
2964
-
disable and re-enable softirq handlers, any attempt to start a softirq
2965
-
handlers during the
2966
-
RCU-bh read-side critical section will be deferred.
2967
-
In this case, <tt>rcu_read_unlock_bh()</tt>
2968
-
will invoke softirq processing, which can take considerable time.
2969
-
One can of course argue that this softirq overhead should be associated
2970
-
with the code following the RCU-bh read-side critical section rather
2971
-
than <tt>rcu_read_unlock_bh()</tt>, but the fact
2972
-
is that most profiling tools cannot be expected to make this sort
2973
-
of fine distinction.
2974
-
For example, suppose that a three-millisecond-long RCU-bh read-side
2975
-
critical section executes during a time of heavy networking load.
2976
-
There will very likely be an attempt to invoke at least one softirq
2977
-
handler during that three milliseconds, but any such invocation will
2978
-
be delayed until the time of the <tt>rcu_read_unlock_bh()</tt>.
2979
-
This can of course make it appear at first glance as if
2980
-
<tt>rcu_read_unlock_bh()</tt> was executing very slowly.
2981
-
2982
-
<p>
2983
-
The
2984
-
<a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">RCU-bh API</a>
2985
-
includes
2986
-
<tt>rcu_read_lock_bh()</tt>,
2987
-
<tt>rcu_read_unlock_bh()</tt>,
2988
-
<tt>rcu_dereference_bh()</tt>,
2989
-
<tt>rcu_dereference_bh_check()</tt>,
2990
-
<tt>synchronize_rcu_bh()</tt>,
2991
-
<tt>synchronize_rcu_bh_expedited()</tt>,
2992
-
<tt>call_rcu_bh()</tt>,
2993
-
<tt>rcu_barrier_bh()</tt>, and
2994
-
<tt>rcu_read_lock_bh_held()</tt>.
2995
-
However, the update-side APIs are now simple wrappers for other RCU
2996
-
flavors, namely RCU-sched in CONFIG_PREEMPT=n kernels and RCU-preempt
2997
-
otherwise.
2998
-
2999
-
<h3><a name="Sched Flavor">Sched Flavor (Historical)</a></h3>
3000
-
3001
-
<p>
3002
-
The RCU-sched flavor of RCU has since been expressed in terms of
3003
-
the other RCU flavors as part of a consolidation of the three
3004
-
flavors into a single flavor.
3005
-
The read-side API remains, and continues to disable preemption and to
3006
-
be accounted for by lockdep.
3007
-
Much of the material in this section is therefore strictly historical
3008
-
in nature.
3009
-
3010
-
<p>
3011
-
Before preemptible RCU, waiting for an RCU grace period had the
3012
-
side effect of also waiting for all pre-existing interrupt
3013
-
and NMI handlers.
3014
-
However, there are legitimate preemptible-RCU implementations that
3015
-
do not have this property, given that any point in the code outside
3016
-
of an RCU read-side critical section can be a quiescent state.
3017
-
Therefore, <i>RCU-sched</i> was created, which follows “classic”
3018
-
RCU in that an RCU-sched grace period waits for for pre-existing
3019
-
interrupt and NMI handlers.
3020
-
In kernels built with <tt>CONFIG_PREEMPT=n</tt>, the RCU and RCU-sched
3021
-
APIs have identical implementations, while kernels built with
3022
-
<tt>CONFIG_PREEMPT=y</tt> provide a separate implementation for each.
3023
-
3024
-
<p>
3025
-
Note well that in <tt>CONFIG_PREEMPT=y</tt> kernels,
3026
-
<tt>rcu_read_lock_sched()</tt> and <tt>rcu_read_unlock_sched()</tt>
3027
-
disable and re-enable preemption, respectively.
3028
-
This means that if there was a preemption attempt during the
3029
-
RCU-sched read-side critical section, <tt>rcu_read_unlock_sched()</tt>
3030
-
will enter the scheduler, with all the latency and overhead entailed.
3031
-
Just as with <tt>rcu_read_unlock_bh()</tt>, this can make it look
3032
-
as if <tt>rcu_read_unlock_sched()</tt> was executing very slowly.
3033
-
However, the highest-priority task won't be preempted, so that task
3034
-
will enjoy low-overhead <tt>rcu_read_unlock_sched()</tt> invocations.
3035
-
3036
-
<p>
3037
-
The
3038
-
<a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">RCU-sched API</a>
3039
-
includes
3040
-
<tt>rcu_read_lock_sched()</tt>,
3041
-
<tt>rcu_read_unlock_sched()</tt>,
3042
-
<tt>rcu_read_lock_sched_notrace()</tt>,
3043
-
<tt>rcu_read_unlock_sched_notrace()</tt>,
3044
-
<tt>rcu_dereference_sched()</tt>,
3045
-
<tt>rcu_dereference_sched_check()</tt>,
3046
-
<tt>synchronize_sched()</tt>,
3047
-
<tt>synchronize_rcu_sched_expedited()</tt>,
3048
-
<tt>call_rcu_sched()</tt>,
3049
-
<tt>rcu_barrier_sched()</tt>, and
3050
-
<tt>rcu_read_lock_sched_held()</tt>.
3051
-
However, anything that disables preemption also marks an RCU-sched
3052
-
read-side critical section, including
3053
-
<tt>preempt_disable()</tt> and <tt>preempt_enable()</tt>,
3054
-
<tt>local_irq_save()</tt> and <tt>local_irq_restore()</tt>,
3055
-
and so on.
3056
-
3057
-
<h3><a name="Sleepable RCU">Sleepable RCU</a></h3>
3058
-
3059
-
<p>
3060
-
For well over a decade, someone saying “I need to block within
3061
-
an RCU read-side critical section” was a reliable indication
3062
-
that this someone did not understand RCU.
3063
-
After all, if you are always blocking in an RCU read-side critical
3064
-
section, you can probably afford to use a higher-overhead synchronization
3065
-
mechanism.
3066
-
However, that changed with the advent of the Linux kernel's notifiers,
3067
-
whose RCU read-side critical
3068
-
sections almost never sleep, but sometimes need to.
3069
-
This resulted in the introduction of
3070
-
<a href="https://lwn.net/Articles/202847/">sleepable RCU</a>,
3071
-
or <i>SRCU</i>.
3072
-
3073
-
<p>
3074
-
SRCU allows different domains to be defined, with each such domain
3075
-
defined by an instance of an <tt>srcu_struct</tt> structure.
3076
-
A pointer to this structure must be passed in to each SRCU function,
3077
-
for example, <tt>synchronize_srcu(&ss)</tt>, where
3078
-
<tt>ss</tt> is the <tt>srcu_struct</tt> structure.
3079
-
The key benefit of these domains is that a slow SRCU reader in one
3080
-
domain does not delay an SRCU grace period in some other domain.
3081
-
That said, one consequence of these domains is that read-side code
3082
-
must pass a “cookie” from <tt>srcu_read_lock()</tt>
3083
-
to <tt>srcu_read_unlock()</tt>, for example, as follows:
3084
-
3085
-
<blockquote>
3086
-
<pre>
3087
-
1 int idx;
3088
-
2
3089
-
3 idx = srcu_read_lock(&ss);
3090
-
4 do_something();
3091
-
5 srcu_read_unlock(&ss, idx);
3092
-
</pre>
3093
-
</blockquote>
3094
-
3095
-
<p>
3096
-
As noted above, it is legal to block within SRCU read-side critical sections,
3097
-
however, with great power comes great responsibility.
3098
-
If you block forever in one of a given domain's SRCU read-side critical
3099
-
sections, then that domain's grace periods will also be blocked forever.
3100
-
Of course, one good way to block forever is to deadlock, which can
3101
-
happen if any operation in a given domain's SRCU read-side critical
3102
-
section can wait, either directly or indirectly, for that domain's
3103
-
grace period to elapse.
3104
-
For example, this results in a self-deadlock:
3105
-
3106
-
<blockquote>
3107
-
<pre>
3108
-
1 int idx;
3109
-
2
3110
-
3 idx = srcu_read_lock(&ss);
3111
-
4 do_something();
3112
-
5 synchronize_srcu(&ss);
3113
-
6 srcu_read_unlock(&ss, idx);
3114
-
</pre>
3115
-
</blockquote>
3116
-
3117
-
<p>
3118
-
However, if line 5 acquired a mutex that was held across
3119
-
a <tt>synchronize_srcu()</tt> for domain <tt>ss</tt>,
3120
-
deadlock would still be possible.
3121
-
Furthermore, if line 5 acquired a mutex that was held across
3122
-
a <tt>synchronize_srcu()</tt> for some other domain <tt>ss1</tt>,
3123
-
and if an <tt>ss1</tt>-domain SRCU read-side critical section
3124
-
acquired another mutex that was held across as <tt>ss</tt>-domain
3125
-
<tt>synchronize_srcu()</tt>,
3126
-
deadlock would again be possible.
3127
-
Such a deadlock cycle could extend across an arbitrarily large number
3128
-
of different SRCU domains.
3129
-
Again, with great power comes great responsibility.
3130
-
3131
-
<p>
3132
-
Unlike the other RCU flavors, SRCU read-side critical sections can
3133
-
run on idle and even offline CPUs.
3134
-
This ability requires that <tt>srcu_read_lock()</tt> and
3135
-
<tt>srcu_read_unlock()</tt> contain memory barriers, which means
3136
-
that SRCU readers will run a bit slower than would RCU readers.
3137
-
It also motivates the <tt>smp_mb__after_srcu_read_unlock()</tt>
3138
-
API, which, in combination with <tt>srcu_read_unlock()</tt>,
3139
-
guarantees a full memory barrier.
3140
-
3141
-
<p>
3142
-
Also unlike other RCU flavors, <tt>synchronize_srcu()</tt> may <b>not</b>
3143
-
be invoked from CPU-hotplug notifiers, due to the fact that SRCU grace
3144
-
periods make use of timers and the possibility of timers being temporarily
3145
-
“stranded” on the outgoing CPU.
3146
-
This stranding of timers means that timers posted to the outgoing CPU
3147
-
will not fire until late in the CPU-hotplug process.
3148
-
The problem is that if a notifier is waiting on an SRCU grace period,
3149
-
that grace period is waiting on a timer, and that timer is stranded on the
3150
-
outgoing CPU, then the notifier will never be awakened, in other words,
3151
-
deadlock has occurred.
3152
-
This same situation of course also prohibits <tt>srcu_barrier()</tt>
3153
-
from being invoked from CPU-hotplug notifiers.
3154
-
3155
-
<p>
3156
-
SRCU also differs from other RCU flavors in that SRCU's expedited and
3157
-
non-expedited grace periods are implemented by the same mechanism.
3158
-
This means that in the current SRCU implementation, expediting a
3159
-
future grace period has the side effect of expediting all prior
3160
-
grace periods that have not yet completed.
3161
-
(But please note that this is a property of the current implementation,
3162
-
not necessarily of future implementations.)
3163
-
In addition, if SRCU has been idle for longer than the interval
3164
-
specified by the <tt>srcutree.exp_holdoff</tt> kernel boot parameter
3165
-
(25 microseconds by default),
3166
-
and if a <tt>synchronize_srcu()</tt> invocation ends this idle period,
3167
-
that invocation will be automatically expedited.
3168
-
3169
-
<p>
3170
-
As of v4.12, SRCU's callbacks are maintained per-CPU, eliminating
3171
-
a locking bottleneck present in prior kernel versions.
3172
-
Although this will allow users to put much heavier stress on
3173
-
<tt>call_srcu()</tt>, it is important to note that SRCU does not
3174
-
yet take any special steps to deal with callback flooding.
3175
-
So if you are posting (say) 10,000 SRCU callbacks per second per CPU,
3176
-
you are probably totally OK, but if you intend to post (say) 1,000,000
3177
-
SRCU callbacks per second per CPU, please run some tests first.
3178
-
SRCU just might need a few adjustment to deal with that sort of load.
3179
-
Of course, your mileage may vary based on the speed of your CPUs and
3180
-
the size of your memory.
3181
-
3182
-
<p>
3183
-
The
3184
-
<a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">SRCU API</a>
3185
-
includes
3186
-
<tt>srcu_read_lock()</tt>,
3187
-
<tt>srcu_read_unlock()</tt>,
3188
-
<tt>srcu_dereference()</tt>,
3189
-
<tt>srcu_dereference_check()</tt>,
3190
-
<tt>synchronize_srcu()</tt>,
3191
-
<tt>synchronize_srcu_expedited()</tt>,
3192
-
<tt>call_srcu()</tt>,
3193
-
<tt>srcu_barrier()</tt>, and
3194
-
<tt>srcu_read_lock_held()</tt>.
3195
-
It also includes
3196
-
<tt>DEFINE_SRCU()</tt>,
3197
-
<tt>DEFINE_STATIC_SRCU()</tt>, and
3198
-
<tt>init_srcu_struct()</tt>
3199
-
APIs for defining and initializing <tt>srcu_struct</tt> structures.
3200
-
3201
-
<h3><a name="Tasks RCU">Tasks RCU</a></h3>
3202
-
3203
-
<p>
3204
-
Some forms of tracing use “trampolines” to handle the
3205
-
binary rewriting required to install different types of probes.
3206
-
It would be good to be able to free old trampolines, which sounds
3207
-
like a job for some form of RCU.
3208
-
However, because it is necessary to be able to install a trace
3209
-
anywhere in the code, it is not possible to use read-side markers
3210
-
such as <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>.
3211
-
In addition, it does not work to have these markers in the trampoline
3212
-
itself, because there would need to be instructions following
3213
-
<tt>rcu_read_unlock()</tt>.
3214
-
Although <tt>synchronize_rcu()</tt> would guarantee that execution
3215
-
reached the <tt>rcu_read_unlock()</tt>, it would not be able to
3216
-
guarantee that execution had completely left the trampoline.
3217
-
3218
-
<p>
3219
-
The solution, in the form of
3220
-
<a href="https://lwn.net/Articles/607117/"><i>Tasks RCU</i></a>,
3221
-
is to have implicit
3222
-
read-side critical sections that are delimited by voluntary context
3223
-
switches, that is, calls to <tt>schedule()</tt>,
3224
-
<tt>cond_resched()</tt>, and
3225
-
<tt>synchronize_rcu_tasks()</tt>.
3226
-
In addition, transitions to and from userspace execution also delimit
3227
-
tasks-RCU read-side critical sections.
3228
-
3229
-
<p>
3230
-
The tasks-RCU API is quite compact, consisting only of
3231
-
<tt>call_rcu_tasks()</tt>,
3232
-
<tt>synchronize_rcu_tasks()</tt>, and
3233
-
<tt>rcu_barrier_tasks()</tt>.
3234
-
In <tt>CONFIG_PREEMPT=n</tt> kernels, trampolines cannot be preempted,
3235
-
so these APIs map to
3236
-
<tt>call_rcu()</tt>,
3237
-
<tt>synchronize_rcu()</tt>, and
3238
-
<tt>rcu_barrier()</tt>, respectively.
3239
-
In <tt>CONFIG_PREEMPT=y</tt> kernels, trampolines can be preempted,
3240
-
and these three APIs are therefore implemented by separate functions
3241
-
that check for voluntary context switches.
3242
-
3243
-
<h2><a name="Possible Future Changes">Possible Future Changes</a></h2>
3244
-
3245
-
<p>
3246
-
One of the tricks that RCU uses to attain update-side scalability is
3247
-
to increase grace-period latency with increasing numbers of CPUs.
3248
-
If this becomes a serious problem, it will be necessary to rework the
3249
-
grace-period state machine so as to avoid the need for the additional
3250
-
latency.
3251
-
3252
-
<p>
3253
-
RCU disables CPU hotplug in a few places, perhaps most notably in the
3254
-
<tt>rcu_barrier()</tt> operations.
3255
-
If there is a strong reason to use <tt>rcu_barrier()</tt> in CPU-hotplug
3256
-
notifiers, it will be necessary to avoid disabling CPU hotplug.
3257
-
This would introduce some complexity, so there had better be a <i>very</i>
3258
-
good reason.
3259
-
3260
-
<p>
3261
-
The tradeoff between grace-period latency on the one hand and interruptions
3262
-
of other CPUs on the other hand may need to be re-examined.
3263
-
The desire is of course for zero grace-period latency as well as zero
3264
-
interprocessor interrupts undertaken during an expedited grace period
3265
-
operation.
3266
-
While this ideal is unlikely to be achievable, it is quite possible that
3267
-
further improvements can be made.
3268
-
3269
-
<p>
3270
-
The multiprocessor implementations of RCU use a combining tree that
3271
-
groups CPUs so as to reduce lock contention and increase cache locality.
3272
-
However, this combining tree does not spread its memory across NUMA
3273
-
nodes nor does it align the CPU groups with hardware features such
3274
-
as sockets or cores.
3275
-
Such spreading and alignment is currently believed to be unnecessary
3276
-
because the hotpath read-side primitives do not access the combining
3277
-
tree, nor does <tt>call_rcu()</tt> in the common case.
3278
-
If you believe that your architecture needs such spreading and alignment,
3279
-
then your architecture should also benefit from the
3280
-
<tt>rcutree.rcu_fanout_leaf</tt> boot parameter, which can be set
3281
-
to the number of CPUs in a socket, NUMA node, or whatever.
3282
-
If the number of CPUs is too large, use a fraction of the number of
3283
-
CPUs.
3284
-
If the number of CPUs is a large prime number, well, that certainly
3285
-
is an “interesting” architectural choice!
3286
-
More flexible arrangements might be considered, but only if
3287
-
<tt>rcutree.rcu_fanout_leaf</tt> has proven inadequate, and only
3288
-
if the inadequacy has been demonstrated by a carefully run and
3289
-
realistic system-level workload.
3290
-
3291
-
<p>
3292
-
Please note that arrangements that require RCU to remap CPU numbers will
3293
-
require extremely good demonstration of need and full exploration of
3294
-
alternatives.
3295
-
3296
-
<p>
3297
-
RCU's various kthreads are reasonably recent additions.
3298
-
It is quite likely that adjustments will be required to more gracefully
3299
-
handle extreme loads.
3300
-
It might also be necessary to be able to relate CPU utilization by
3301
-
RCU's kthreads and softirq handlers to the code that instigated this
3302
-
CPU utilization.
3303
-
For example, RCU callback overhead might be charged back to the
3304
-
originating <tt>call_rcu()</tt> instance, though probably not
3305
-
in production kernels.
3306
-
3307
-
<p>
3308
-
Additional work may be required to provide reasonable forward-progress
3309
-
guarantees under heavy load for grace periods and for callback
3310
-
invocation.
3311
-
3312
-
<h2><a name="Summary">Summary</a></h2>
3313
-
3314
-
<p>
3315
-
This document has presented more than two decade's worth of RCU
3316
-
requirements.
3317
-
Given that the requirements keep changing, this will not be the last
3318
-
word on this subject, but at least it serves to get an important
3319
-
subset of the requirements set forth.
3320
-
3321
-
<h2><a name="Acknowledgments">Acknowledgments</a></h2>
3322
-
3323
-
I am grateful to Steven Rostedt, Lai Jiangshan, Ingo Molnar,
3324
-
Oleg Nesterov, Borislav Petkov, Peter Zijlstra, Boqun Feng, and
3325
-
Andy Lutomirski for their help in rendering
3326
-
this article human readable, and to Michelle Rankin for her support
3327
-
of this effort.
3328
-
Other contributions are acknowledged in the Linux kernel's git archive.
3329
-
3330
-
</body></html>
+2662
Documentation/RCU/Design/Requirements/Requirements.rst
+2662
Documentation/RCU/Design/Requirements/Requirements.rst
···
1
+
=================================
2
+
A Tour Through RCU's Requirements
3
+
=================================
4
+
5
+
Copyright IBM Corporation, 2015
6
+
7
+
Author: Paul E. McKenney
8
+
9
+
The initial version of this document appeared in the
10
+
`LWN <https://lwn.net/>`_ on those articles:
11
+
`part 1 <https://lwn.net/Articles/652156/>`_,
12
+
`part 2 <https://lwn.net/Articles/652677/>`_, and
13
+
`part 3 <https://lwn.net/Articles/653326/>`_.
14
+
15
+
Introduction
16
+
------------
17
+
18
+
Read-copy update (RCU) is a synchronization mechanism that is often used
19
+
as a replacement for reader-writer locking. RCU is unusual in that
20
+
updaters do not block readers, which means that RCU's read-side
21
+
primitives can be exceedingly fast and scalable. In addition, updaters
22
+
can make useful forward progress concurrently with readers. However, all
23
+
this concurrency between RCU readers and updaters does raise the
24
+
question of exactly what RCU readers are doing, which in turn raises the
25
+
question of exactly what RCU's requirements are.
26
+
27
+
This document therefore summarizes RCU's requirements, and can be
28
+
thought of as an informal, high-level specification for RCU. It is
29
+
important to understand that RCU's specification is primarily empirical
30
+
in nature; in fact, I learned about many of these requirements the hard
31
+
way. This situation might cause some consternation, however, not only
32
+
has this learning process been a lot of fun, but it has also been a
33
+
great privilege to work with so many people willing to apply
34
+
technologies in interesting new ways.
35
+
36
+
All that aside, here are the categories of currently known RCU
37
+
requirements:
38
+
39
+
#. `Fundamental Requirements <#Fundamental%20Requirements>`__
40
+
#. `Fundamental Non-Requirements <#Fundamental%20Non-Requirements>`__
41
+
#. `Parallelism Facts of Life <#Parallelism%20Facts%20of%20Life>`__
42
+
#. `Quality-of-Implementation
43
+
Requirements <#Quality-of-Implementation%20Requirements>`__
44
+
#. `Linux Kernel Complications <#Linux%20Kernel%20Complications>`__
45
+
#. `Software-Engineering
46
+
Requirements <#Software-Engineering%20Requirements>`__
47
+
#. `Other RCU Flavors <#Other%20RCU%20Flavors>`__
48
+
#. `Possible Future Changes <#Possible%20Future%20Changes>`__
49
+
50
+
This is followed by a `summary <#Summary>`__, however, the answers to
51
+
each quick quiz immediately follows the quiz. Select the big white space
52
+
with your mouse to see the answer.
53
+
54
+
Fundamental Requirements
55
+
------------------------
56
+
57
+
RCU's fundamental requirements are the closest thing RCU has to hard
58
+
mathematical requirements. These are:
59
+
60
+
#. `Grace-Period Guarantee <#Grace-Period%20Guarantee>`__
61
+
#. `Publish-Subscribe Guarantee <#Publish-Subscribe%20Guarantee>`__
62
+
#. `Memory-Barrier Guarantees <#Memory-Barrier%20Guarantees>`__
63
+
#. `RCU Primitives Guaranteed to Execute
64
+
Unconditionally <#RCU%20Primitives%20Guaranteed%20to%20Execute%20Unconditionally>`__
65
+
#. `Guaranteed Read-to-Write
66
+
Upgrade <#Guaranteed%20Read-to-Write%20Upgrade>`__
67
+
68
+
Grace-Period Guarantee
69
+
~~~~~~~~~~~~~~~~~~~~~~
70
+
71
+
RCU's grace-period guarantee is unusual in being premeditated: Jack
72
+
Slingwine and I had this guarantee firmly in mind when we started work
73
+
on RCU (then called “rclock”) in the early 1990s. That said, the past
74
+
two decades of experience with RCU have produced a much more detailed
75
+
understanding of this guarantee.
76
+
77
+
RCU's grace-period guarantee allows updaters to wait for the completion
78
+
of all pre-existing RCU read-side critical sections. An RCU read-side
79
+
critical section begins with the marker ``rcu_read_lock()`` and ends
80
+
with the marker ``rcu_read_unlock()``. These markers may be nested, and
81
+
RCU treats a nested set as one big RCU read-side critical section.
82
+
Production-quality implementations of ``rcu_read_lock()`` and
83
+
``rcu_read_unlock()`` are extremely lightweight, and in fact have
84
+
exactly zero overhead in Linux kernels built for production use with
85
+
``CONFIG_PREEMPT=n``.
86
+
87
+
This guarantee allows ordering to be enforced with extremely low
88
+
overhead to readers, for example:
89
+
90
+
::
91
+
92
+
1 int x, y;
93
+
2
94
+
3 void thread0(void)
95
+
4 {
96
+
5 rcu_read_lock();
97
+
6 r1 = READ_ONCE(x);
98
+
7 r2 = READ_ONCE(y);
99
+
8 rcu_read_unlock();
100
+
9 }
101
+
10
102
+
11 void thread1(void)
103
+
12 {
104
+
13 WRITE_ONCE(x, 1);
105
+
14 synchronize_rcu();
106
+
15 WRITE_ONCE(y, 1);
107
+
16 }
108
+
109
+
Because the ``synchronize_rcu()`` on line 14 waits for all pre-existing
110
+
readers, any instance of ``thread0()`` that loads a value of zero from
111
+
``x`` must complete before ``thread1()`` stores to ``y``, so that
112
+
instance must also load a value of zero from ``y``. Similarly, any
113
+
instance of ``thread0()`` that loads a value of one from ``y`` must have
114
+
started after the ``synchronize_rcu()`` started, and must therefore also
115
+
load a value of one from ``x``. Therefore, the outcome:
116
+
117
+
::
118
+
119
+
(r1 == 0 && r2 == 1)
120
+
121
+
cannot happen.
122
+
123
+
+-----------------------------------------------------------------------+
124
+
| **Quick Quiz**: |
125
+
+-----------------------------------------------------------------------+
126
+
| Wait a minute! You said that updaters can make useful forward |
127
+
| progress concurrently with readers, but pre-existing readers will |
128
+
| block ``synchronize_rcu()``!!! |
129
+
| Just who are you trying to fool??? |
130
+
+-----------------------------------------------------------------------+
131
+
| **Answer**: |
132
+
+-----------------------------------------------------------------------+
133
+
| First, if updaters do not wish to be blocked by readers, they can use |
134
+
| ``call_rcu()`` or ``kfree_rcu()``, which will be discussed later. |
135
+
| Second, even when using ``synchronize_rcu()``, the other update-side |
136
+
| code does run concurrently with readers, whether pre-existing or not. |
137
+
+-----------------------------------------------------------------------+
138
+
139
+
This scenario resembles one of the first uses of RCU in
140
+
`DYNIX/ptx <https://en.wikipedia.org/wiki/DYNIX>`__, which managed a
141
+
distributed lock manager's transition into a state suitable for handling
142
+
recovery from node failure, more or less as follows:
143
+
144
+
::
145
+
146
+
1 #define STATE_NORMAL 0
147
+
2 #define STATE_WANT_RECOVERY 1
148
+
3 #define STATE_RECOVERING 2
149
+
4 #define STATE_WANT_NORMAL 3
150
+
5
151
+
6 int state = STATE_NORMAL;
152
+
7
153
+
8 void do_something_dlm(void)
154
+
9 {
155
+
10 int state_snap;
156
+
11
157
+
12 rcu_read_lock();
158
+
13 state_snap = READ_ONCE(state);
159
+
14 if (state_snap == STATE_NORMAL)
160
+
15 do_something();
161
+
16 else
162
+
17 do_something_carefully();
163
+
18 rcu_read_unlock();
164
+
19 }
165
+
20
166
+
21 void start_recovery(void)
167
+
22 {
168
+
23 WRITE_ONCE(state, STATE_WANT_RECOVERY);
169
+
24 synchronize_rcu();
170
+
25 WRITE_ONCE(state, STATE_RECOVERING);
171
+
26 recovery();
172
+
27 WRITE_ONCE(state, STATE_WANT_NORMAL);
173
+
28 synchronize_rcu();
174
+
29 WRITE_ONCE(state, STATE_NORMAL);
175
+
30 }
176
+
177
+
The RCU read-side critical section in ``do_something_dlm()`` works with
178
+
the ``synchronize_rcu()`` in ``start_recovery()`` to guarantee that
179
+
``do_something()`` never runs concurrently with ``recovery()``, but with
180
+
little or no synchronization overhead in ``do_something_dlm()``.
181
+
182
+
+-----------------------------------------------------------------------+
183
+
| **Quick Quiz**: |
184
+
+-----------------------------------------------------------------------+
185
+
| Why is the ``synchronize_rcu()`` on line 28 needed? |
186
+
+-----------------------------------------------------------------------+
187
+
| **Answer**: |
188
+
+-----------------------------------------------------------------------+
189
+
| Without that extra grace period, memory reordering could result in |
190
+
| ``do_something_dlm()`` executing ``do_something()`` concurrently with |
191
+
| the last bits of ``recovery()``. |
192
+
+-----------------------------------------------------------------------+
193
+
194
+
In order to avoid fatal problems such as deadlocks, an RCU read-side
195
+
critical section must not contain calls to ``synchronize_rcu()``.
196
+
Similarly, an RCU read-side critical section must not contain anything
197
+
that waits, directly or indirectly, on completion of an invocation of
198
+
``synchronize_rcu()``.
199
+
200
+
Although RCU's grace-period guarantee is useful in and of itself, with
201
+
`quite a few use cases <https://lwn.net/Articles/573497/>`__, it would
202
+
be good to be able to use RCU to coordinate read-side access to linked
203
+
data structures. For this, the grace-period guarantee is not sufficient,
204
+
as can be seen in function ``add_gp_buggy()`` below. We will look at the
205
+
reader's code later, but in the meantime, just think of the reader as
206
+
locklessly picking up the ``gp`` pointer, and, if the value loaded is
207
+
non-\ ``NULL``, locklessly accessing the ``->a`` and ``->b`` fields.
208
+
209
+
::
210
+
211
+
1 bool add_gp_buggy(int a, int b)
212
+
2 {
213
+
3 p = kmalloc(sizeof(*p), GFP_KERNEL);
214
+
4 if (!p)
215
+
5 return -ENOMEM;
216
+
6 spin_lock(&gp_lock);
217
+
7 if (rcu_access_pointer(gp)) {
218
+
8 spin_unlock(&gp_lock);
219
+
9 return false;
220
+
10 }
221
+
11 p->a = a;
222
+
12 p->b = a;
223
+
13 gp = p; /* ORDERING BUG */
224
+
14 spin_unlock(&gp_lock);
225
+
15 return true;
226
+
16 }
227
+
228
+
The problem is that both the compiler and weakly ordered CPUs are within
229
+
their rights to reorder this code as follows:
230
+
231
+
::
232
+
233
+
1 bool add_gp_buggy_optimized(int a, int b)
234
+
2 {
235
+
3 p = kmalloc(sizeof(*p), GFP_KERNEL);
236
+
4 if (!p)
237
+
5 return -ENOMEM;
238
+
6 spin_lock(&gp_lock);
239
+
7 if (rcu_access_pointer(gp)) {
240
+
8 spin_unlock(&gp_lock);
241
+
9 return false;
242
+
10 }
243
+
11 gp = p; /* ORDERING BUG */
244
+
12 p->a = a;
245
+
13 p->b = a;
246
+
14 spin_unlock(&gp_lock);
247
+
15 return true;
248
+
16 }
249
+
250
+
If an RCU reader fetches ``gp`` just after ``add_gp_buggy_optimized``
251
+
executes line 11, it will see garbage in the ``->a`` and ``->b`` fields.
252
+
And this is but one of many ways in which compiler and hardware
253
+
optimizations could cause trouble. Therefore, we clearly need some way
254
+
to prevent the compiler and the CPU from reordering in this manner,
255
+
which brings us to the publish-subscribe guarantee discussed in the next
256
+
section.
257
+
258
+
Publish/Subscribe Guarantee
259
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~
260
+
261
+
RCU's publish-subscribe guarantee allows data to be inserted into a
262
+
linked data structure without disrupting RCU readers. The updater uses
263
+
``rcu_assign_pointer()`` to insert the new data, and readers use
264
+
``rcu_dereference()`` to access data, whether new or old. The following
265
+
shows an example of insertion:
266
+
267
+
::
268
+
269
+
1 bool add_gp(int a, int b)
270
+
2 {
271
+
3 p = kmalloc(sizeof(*p), GFP_KERNEL);
272
+
4 if (!p)
273
+
5 return -ENOMEM;
274
+
6 spin_lock(&gp_lock);
275
+
7 if (rcu_access_pointer(gp)) {
276
+
8 spin_unlock(&gp_lock);
277
+
9 return false;
278
+
10 }
279
+
11 p->a = a;
280
+
12 p->b = a;
281
+
13 rcu_assign_pointer(gp, p);
282
+
14 spin_unlock(&gp_lock);
283
+
15 return true;
284
+
16 }
285
+
286
+
The ``rcu_assign_pointer()`` on line 13 is conceptually equivalent to a
287
+
simple assignment statement, but also guarantees that its assignment
288
+
will happen after the two assignments in lines 11 and 12, similar to the
289
+
C11 ``memory_order_release`` store operation. It also prevents any
290
+
number of “interesting” compiler optimizations, for example, the use of
291
+
``gp`` as a scratch location immediately preceding the assignment.
292
+
293
+
+-----------------------------------------------------------------------+
294
+
| **Quick Quiz**: |
295
+
+-----------------------------------------------------------------------+
296
+
| But ``rcu_assign_pointer()`` does nothing to prevent the two |
297
+
| assignments to ``p->a`` and ``p->b`` from being reordered. Can't that |
298
+
| also cause problems? |
299
+
+-----------------------------------------------------------------------+
300
+
| **Answer**: |
301
+
+-----------------------------------------------------------------------+
302
+
| No, it cannot. The readers cannot see either of these two fields |
303
+
| until the assignment to ``gp``, by which time both fields are fully |
304
+
| initialized. So reordering the assignments to ``p->a`` and ``p->b`` |
305
+
| cannot possibly cause any problems. |
306
+
+-----------------------------------------------------------------------+
307
+
308
+
It is tempting to assume that the reader need not do anything special to
309
+
control its accesses to the RCU-protected data, as shown in
310
+
``do_something_gp_buggy()`` below:
311
+
312
+
::
313
+
314
+
1 bool do_something_gp_buggy(void)
315
+
2 {
316
+
3 rcu_read_lock();
317
+
4 p = gp; /* OPTIMIZATIONS GALORE!!! */
318
+
5 if (p) {
319
+
6 do_something(p->a, p->b);
320
+
7 rcu_read_unlock();
321
+
8 return true;
322
+
9 }
323
+
10 rcu_read_unlock();
324
+
11 return false;
325
+
12 }
326
+
327
+
However, this temptation must be resisted because there are a
328
+
surprisingly large number of ways that the compiler (to say nothing of
329
+
`DEC Alpha CPUs <https://h71000.www7.hp.com/wizard/wiz_2637.html>`__)
330
+
can trip this code up. For but one example, if the compiler were short
331
+
of registers, it might choose to refetch from ``gp`` rather than keeping
332
+
a separate copy in ``p`` as follows:
333
+
334
+
::
335
+
336
+
1 bool do_something_gp_buggy_optimized(void)
337
+
2 {
338
+
3 rcu_read_lock();
339
+
4 if (gp) { /* OPTIMIZATIONS GALORE!!! */
340
+
5 do_something(gp->a, gp->b);
341
+
6 rcu_read_unlock();
342
+
7 return true;
343
+
8 }
344
+
9 rcu_read_unlock();
345
+
10 return false;
346
+
11 }
347
+
348
+
If this function ran concurrently with a series of updates that replaced
349
+
the current structure with a new one, the fetches of ``gp->a`` and
350
+
``gp->b`` might well come from two different structures, which could
351
+
cause serious confusion. To prevent this (and much else besides),
352
+
``do_something_gp()`` uses ``rcu_dereference()`` to fetch from ``gp``:
353
+
354
+
::
355
+
356
+
1 bool do_something_gp(void)
357
+
2 {
358
+
3 rcu_read_lock();
359
+
4 p = rcu_dereference(gp);
360
+
5 if (p) {
361
+
6 do_something(p->a, p->b);
362
+
7 rcu_read_unlock();
363
+
8 return true;
364
+
9 }
365
+
10 rcu_read_unlock();
366
+
11 return false;
367
+
12 }
368
+
369
+
The ``rcu_dereference()`` uses volatile casts and (for DEC Alpha) memory
370
+
barriers in the Linux kernel. Should a `high-quality implementation of
371
+
C11 ``memory_order_consume``
372
+
[PDF] <http://www.rdrop.com/users/paulmck/RCU/consume.2015.07.13a.pdf>`__
373
+
ever appear, then ``rcu_dereference()`` could be implemented as a
374
+
``memory_order_consume`` load. Regardless of the exact implementation, a
375
+
pointer fetched by ``rcu_dereference()`` may not be used outside of the
376
+
outermost RCU read-side critical section containing that
377
+
``rcu_dereference()``, unless protection of the corresponding data
378
+
element has been passed from RCU to some other synchronization
379
+
mechanism, most commonly locking or `reference
380
+
counting <https://www.kernel.org/doc/Documentation/RCU/rcuref.txt>`__.
381
+
382
+
In short, updaters use ``rcu_assign_pointer()`` and readers use
383
+
``rcu_dereference()``, and these two RCU API elements work together to
384
+
ensure that readers have a consistent view of newly added data elements.
385
+
386
+
Of course, it is also necessary to remove elements from RCU-protected
387
+
data structures, for example, using the following process:
388
+
389
+
#. Remove the data element from the enclosing structure.
390
+
#. Wait for all pre-existing RCU read-side critical sections to complete
391
+
(because only pre-existing readers can possibly have a reference to
392
+
the newly removed data element).
393
+
#. At this point, only the updater has a reference to the newly removed
394
+
data element, so it can safely reclaim the data element, for example,
395
+
by passing it to ``kfree()``.
396
+
397
+
This process is implemented by ``remove_gp_synchronous()``:
398
+
399
+
::
400
+
401
+
1 bool remove_gp_synchronous(void)
402
+
2 {
403
+
3 struct foo *p;
404
+
4
405
+
5 spin_lock(&gp_lock);
406
+
6 p = rcu_access_pointer(gp);
407
+
7 if (!p) {
408
+
8 spin_unlock(&gp_lock);
409
+
9 return false;
410
+
10 }
411
+
11 rcu_assign_pointer(gp, NULL);
412
+
12 spin_unlock(&gp_lock);
413
+
13 synchronize_rcu();
414
+
14 kfree(p);
415
+
15 return true;
416
+
16 }
417
+
418
+
This function is straightforward, with line 13 waiting for a grace
419
+
period before line 14 frees the old data element. This waiting ensures
420
+
that readers will reach line 7 of ``do_something_gp()`` before the data
421
+
element referenced by ``p`` is freed. The ``rcu_access_pointer()`` on
422
+
line 6 is similar to ``rcu_dereference()``, except that:
423
+
424
+
#. The value returned by ``rcu_access_pointer()`` cannot be
425
+
dereferenced. If you want to access the value pointed to as well as
426
+
the pointer itself, use ``rcu_dereference()`` instead of
427
+
``rcu_access_pointer()``.
428
+
#. The call to ``rcu_access_pointer()`` need not be protected. In
429
+
contrast, ``rcu_dereference()`` must either be within an RCU
430
+
read-side critical section or in a code segment where the pointer
431
+
cannot change, for example, in code protected by the corresponding
432
+
update-side lock.
433
+
434
+
+-----------------------------------------------------------------------+
435
+
| **Quick Quiz**: |
436
+
+-----------------------------------------------------------------------+
437
+
| Without the ``rcu_dereference()`` or the ``rcu_access_pointer()``, |
438
+
| what destructive optimizations might the compiler make use of? |
439
+
+-----------------------------------------------------------------------+
440
+
| **Answer**: |
441
+
+-----------------------------------------------------------------------+
442
+
| Let's start with what happens to ``do_something_gp()`` if it fails to |
443
+
| use ``rcu_dereference()``. It could reuse a value formerly fetched |
444
+
| from this same pointer. It could also fetch the pointer from ``gp`` |
445
+
| in a byte-at-a-time manner, resulting in *load tearing*, in turn |
446
+
| resulting a bytewise mash-up of two distinct pointer values. It might |
447
+
| even use value-speculation optimizations, where it makes a wrong |
448
+
| guess, but by the time it gets around to checking the value, an |
449
+
| update has changed the pointer to match the wrong guess. Too bad |
450
+
| about any dereferences that returned pre-initialization garbage in |
451
+
| the meantime! |
452
+
| For ``remove_gp_synchronous()``, as long as all modifications to |
453
+
| ``gp`` are carried out while holding ``gp_lock``, the above |
454
+
| optimizations are harmless. However, ``sparse`` will complain if you |
455
+
| define ``gp`` with ``__rcu`` and then access it without using either |
456
+
| ``rcu_access_pointer()`` or ``rcu_dereference()``. |
457
+
+-----------------------------------------------------------------------+
458
+
459
+
In short, RCU's publish-subscribe guarantee is provided by the
460
+
combination of ``rcu_assign_pointer()`` and ``rcu_dereference()``. This
461
+
guarantee allows data elements to be safely added to RCU-protected
462
+
linked data structures without disrupting RCU readers. This guarantee
463
+
can be used in combination with the grace-period guarantee to also allow
464
+
data elements to be removed from RCU-protected linked data structures,
465
+
again without disrupting RCU readers.
466
+
467
+
This guarantee was only partially premeditated. DYNIX/ptx used an
468
+
explicit memory barrier for publication, but had nothing resembling
469
+
``rcu_dereference()`` for subscription, nor did it have anything
470
+
resembling the ``smp_read_barrier_depends()`` that was later subsumed
471
+
into ``rcu_dereference()`` and later still into ``READ_ONCE()``. The
472
+
need for these operations made itself known quite suddenly at a
473
+
late-1990s meeting with the DEC Alpha architects, back in the days when
474
+
DEC was still a free-standing company. It took the Alpha architects a
475
+
good hour to convince me that any sort of barrier would ever be needed,
476
+
and it then took me a good *two* hours to convince them that their
477
+
documentation did not make this point clear. More recent work with the C
478
+
and C++ standards committees have provided much education on tricks and
479
+
traps from the compiler. In short, compilers were much less tricky in
480
+
the early 1990s, but in 2015, don't even think about omitting
481
+
``rcu_dereference()``!
482
+
483
+
Memory-Barrier Guarantees
484
+
~~~~~~~~~~~~~~~~~~~~~~~~~
485
+
486
+
The previous section's simple linked-data-structure scenario clearly
487
+
demonstrates the need for RCU's stringent memory-ordering guarantees on
488
+
systems with more than one CPU:
489
+
490
+
#. Each CPU that has an RCU read-side critical section that begins
491
+
before ``synchronize_rcu()`` starts is guaranteed to execute a full
492
+
memory barrier between the time that the RCU read-side critical
493
+
section ends and the time that ``synchronize_rcu()`` returns. Without
494
+
this guarantee, a pre-existing RCU read-side critical section might
495
+
hold a reference to the newly removed ``struct foo`` after the
496
+
``kfree()`` on line 14 of ``remove_gp_synchronous()``.
497
+
#. Each CPU that has an RCU read-side critical section that ends after
498
+
``synchronize_rcu()`` returns is guaranteed to execute a full memory
499
+
barrier between the time that ``synchronize_rcu()`` begins and the
500
+
time that the RCU read-side critical section begins. Without this
501
+
guarantee, a later RCU read-side critical section running after the
502
+
``kfree()`` on line 14 of ``remove_gp_synchronous()`` might later run
503
+
``do_something_gp()`` and find the newly deleted ``struct foo``.
504
+
#. If the task invoking ``synchronize_rcu()`` remains on a given CPU,
505
+
then that CPU is guaranteed to execute a full memory barrier sometime
506
+
during the execution of ``synchronize_rcu()``. This guarantee ensures
507
+
that the ``kfree()`` on line 14 of ``remove_gp_synchronous()`` really
508
+
does execute after the removal on line 11.
509
+
#. If the task invoking ``synchronize_rcu()`` migrates among a group of
510
+
CPUs during that invocation, then each of the CPUs in that group is
511
+
guaranteed to execute a full memory barrier sometime during the
512
+
execution of ``synchronize_rcu()``. This guarantee also ensures that
513
+
the ``kfree()`` on line 14 of ``remove_gp_synchronous()`` really does
514
+
execute after the removal on line 11, but also in the case where the
515
+
thread executing the ``synchronize_rcu()`` migrates in the meantime.
516
+
517
+
+-----------------------------------------------------------------------+
518
+
| **Quick Quiz**: |
519
+
+-----------------------------------------------------------------------+
520
+
| Given that multiple CPUs can start RCU read-side critical sections at |
521
+
| any time without any ordering whatsoever, how can RCU possibly tell |
522
+
| whether or not a given RCU read-side critical section starts before a |
523
+
| given instance of ``synchronize_rcu()``? |
524
+
+-----------------------------------------------------------------------+
525
+
| **Answer**: |
526
+
+-----------------------------------------------------------------------+
527
+
| If RCU cannot tell whether or not a given RCU read-side critical |
528
+
| section starts before a given instance of ``synchronize_rcu()``, then |
529
+
| it must assume that the RCU read-side critical section started first. |
530
+
| In other words, a given instance of ``synchronize_rcu()`` can avoid |
531
+
| waiting on a given RCU read-side critical section only if it can |
532
+
| prove that ``synchronize_rcu()`` started first. |
533
+
| A related question is “When ``rcu_read_lock()`` doesn't generate any |
534
+
| code, why does it matter how it relates to a grace period?” The |
535
+
| answer is that it is not the relationship of ``rcu_read_lock()`` |
536
+
| itself that is important, but rather the relationship of the code |
537
+
| within the enclosed RCU read-side critical section to the code |
538
+
| preceding and following the grace period. If we take this viewpoint, |
539
+
| then a given RCU read-side critical section begins before a given |
540
+
| grace period when some access preceding the grace period observes the |
541
+
| effect of some access within the critical section, in which case none |
542
+
| of the accesses within the critical section may observe the effects |
543
+
| of any access following the grace period. |
544
+
| |
545
+
| As of late 2016, mathematical models of RCU take this viewpoint, for |
546
+
| example, see slides 62 and 63 of the `2016 LinuxCon |
547
+
| EU <http://www2.rdrop.com/users/paulmck/scalability/paper/LinuxMM.201 |
548
+
| 6.10.04c.LCE.pdf>`__ |
549
+
| presentation. |
550
+
+-----------------------------------------------------------------------+
551
+
552
+
+-----------------------------------------------------------------------+
553
+
| **Quick Quiz**: |
554
+
+-----------------------------------------------------------------------+
555
+
| The first and second guarantees require unbelievably strict ordering! |
556
+
| Are all these memory barriers *really* required? |
557
+
+-----------------------------------------------------------------------+
558
+
| **Answer**: |
559
+
+-----------------------------------------------------------------------+
560
+
| Yes, they really are required. To see why the first guarantee is |
561
+
| required, consider the following sequence of events: |
562
+
| |
563
+
| #. CPU 1: ``rcu_read_lock()`` |
564
+
| #. CPU 1: ``q = rcu_dereference(gp); /* Very likely to return p. */`` |
565
+
| #. CPU 0: ``list_del_rcu(p);`` |
566
+
| #. CPU 0: ``synchronize_rcu()`` starts. |
567
+
| #. CPU 1: ``do_something_with(q->a);`` |
568
+
| ``/* No smp_mb(), so might happen after kfree(). */`` |
569
+
| #. CPU 1: ``rcu_read_unlock()`` |
570
+
| #. CPU 0: ``synchronize_rcu()`` returns. |
571
+
| #. CPU 0: ``kfree(p);`` |
572
+
| |
573
+
| Therefore, there absolutely must be a full memory barrier between the |
574
+
| end of the RCU read-side critical section and the end of the grace |
575
+
| period. |
576
+
| |
577
+
| The sequence of events demonstrating the necessity of the second rule |
578
+
| is roughly similar: |
579
+
| |
580
+
| #. CPU 0: ``list_del_rcu(p);`` |
581
+
| #. CPU 0: ``synchronize_rcu()`` starts. |
582
+
| #. CPU 1: ``rcu_read_lock()`` |
583
+
| #. CPU 1: ``q = rcu_dereference(gp);`` |
584
+
| ``/* Might return p if no memory barrier. */`` |
585
+
| #. CPU 0: ``synchronize_rcu()`` returns. |
586
+
| #. CPU 0: ``kfree(p);`` |
587
+
| #. CPU 1: ``do_something_with(q->a); /* Boom!!! */`` |
588
+
| #. CPU 1: ``rcu_read_unlock()`` |
589
+
| |
590
+
| And similarly, without a memory barrier between the beginning of the |
591
+
| grace period and the beginning of the RCU read-side critical section, |
592
+
| CPU 1 might end up accessing the freelist. |
593
+
| |
594
+
| The “as if” rule of course applies, so that any implementation that |
595
+
| acts as if the appropriate memory barriers were in place is a correct |
596
+
| implementation. That said, it is much easier to fool yourself into |
597
+
| believing that you have adhered to the as-if rule than it is to |
598
+
| actually adhere to it! |
599
+
+-----------------------------------------------------------------------+
600
+
601
+
+-----------------------------------------------------------------------+
602
+
| **Quick Quiz**: |
603
+
+-----------------------------------------------------------------------+
604
+
| You claim that ``rcu_read_lock()`` and ``rcu_read_unlock()`` generate |
605
+
| absolutely no code in some kernel builds. This means that the |
606
+
| compiler might arbitrarily rearrange consecutive RCU read-side |
607
+
| critical sections. Given such rearrangement, if a given RCU read-side |
608
+
| critical section is done, how can you be sure that all prior RCU |
609
+
| read-side critical sections are done? Won't the compiler |
610
+
| rearrangements make that impossible to determine? |
611
+
+-----------------------------------------------------------------------+
612
+
| **Answer**: |
613
+
+-----------------------------------------------------------------------+
614
+
| In cases where ``rcu_read_lock()`` and ``rcu_read_unlock()`` generate |
615
+
| absolutely no code, RCU infers quiescent states only at special |
616
+
| locations, for example, within the scheduler. Because calls to |
617
+
| ``schedule()`` had better prevent calling-code accesses to shared |
618
+
| variables from being rearranged across the call to ``schedule()``, if |
619
+
| RCU detects the end of a given RCU read-side critical section, it |
620
+
| will necessarily detect the end of all prior RCU read-side critical |
621
+
| sections, no matter how aggressively the compiler scrambles the code. |
622
+
| Again, this all assumes that the compiler cannot scramble code across |
623
+
| calls to the scheduler, out of interrupt handlers, into the idle |
624
+
| loop, into user-mode code, and so on. But if your kernel build allows |
625
+
| that sort of scrambling, you have broken far more than just RCU! |
626
+
+-----------------------------------------------------------------------+
627
+
628
+
Note that these memory-barrier requirements do not replace the
629
+
fundamental RCU requirement that a grace period wait for all
630
+
pre-existing readers. On the contrary, the memory barriers called out in
631
+
this section must operate in such a way as to *enforce* this fundamental
632
+
requirement. Of course, different implementations enforce this
633
+
requirement in different ways, but enforce it they must.
634
+
635
+
RCU Primitives Guaranteed to Execute Unconditionally
636
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
637
+
638
+
The common-case RCU primitives are unconditional. They are invoked, they
639
+
do their job, and they return, with no possibility of error, and no need
640
+
to retry. This is a key RCU design philosophy.
641
+
642
+
However, this philosophy is pragmatic rather than pigheaded. If someone
643
+
comes up with a good justification for a particular conditional RCU
644
+
primitive, it might well be implemented and added. After all, this
645
+
guarantee was reverse-engineered, not premeditated. The unconditional
646
+
nature of the RCU primitives was initially an accident of
647
+
implementation, and later experience with synchronization primitives
648
+
with conditional primitives caused me to elevate this accident to a
649
+
guarantee. Therefore, the justification for adding a conditional
650
+
primitive to RCU would need to be based on detailed and compelling use
651
+
cases.
652
+
653
+
Guaranteed Read-to-Write Upgrade
654
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
655
+
656
+
As far as RCU is concerned, it is always possible to carry out an update
657
+
within an RCU read-side critical section. For example, that RCU
658
+
read-side critical section might search for a given data element, and
659
+
then might acquire the update-side spinlock in order to update that
660
+
element, all while remaining in that RCU read-side critical section. Of
661
+
course, it is necessary to exit the RCU read-side critical section
662
+
before invoking ``synchronize_rcu()``, however, this inconvenience can
663
+
be avoided through use of the ``call_rcu()`` and ``kfree_rcu()`` API
664
+
members described later in this document.
665
+
666
+
+-----------------------------------------------------------------------+
667
+
| **Quick Quiz**: |
668
+
+-----------------------------------------------------------------------+
669
+
| But how does the upgrade-to-write operation exclude other readers? |
670
+
+-----------------------------------------------------------------------+
671
+
| **Answer**: |
672
+
+-----------------------------------------------------------------------+
673
+
| It doesn't, just like normal RCU updates, which also do not exclude |
674
+
| RCU readers. |
675
+
+-----------------------------------------------------------------------+
676
+
677
+
This guarantee allows lookup code to be shared between read-side and
678
+
update-side code, and was premeditated, appearing in the earliest
679
+
DYNIX/ptx RCU documentation.
680
+
681
+
Fundamental Non-Requirements
682
+
----------------------------
683
+
684
+
RCU provides extremely lightweight readers, and its read-side
685
+
guarantees, though quite useful, are correspondingly lightweight. It is
686
+
therefore all too easy to assume that RCU is guaranteeing more than it
687
+
really is. Of course, the list of things that RCU does not guarantee is
688
+
infinitely long, however, the following sections list a few
689
+
non-guarantees that have caused confusion. Except where otherwise noted,
690
+
these non-guarantees were premeditated.
691
+
692
+
#. `Readers Impose Minimal
693
+
Ordering <#Readers%20Impose%20Minimal%20Ordering>`__
694
+
#. `Readers Do Not Exclude
695
+
Updaters <#Readers%20Do%20Not%20Exclude%20Updaters>`__
696
+
#. `Updaters Only Wait For Old
697
+
Readers <#Updaters%20Only%20Wait%20For%20Old%20Readers>`__
698
+
#. `Grace Periods Don't Partition Read-Side Critical
699
+
Sections <#Grace%20Periods%20Don't%20Partition%20Read-Side%20Critical%20Sections>`__
700
+
#. `Read-Side Critical Sections Don't Partition Grace
701
+
Periods <#Read-Side%20Critical%20Sections%20Don't%20Partition%20Grace%20Periods>`__
702
+
703
+
Readers Impose Minimal Ordering
704
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
705
+
706
+
Reader-side markers such as ``rcu_read_lock()`` and
707
+
``rcu_read_unlock()`` provide absolutely no ordering guarantees except
708
+
through their interaction with the grace-period APIs such as
709
+
``synchronize_rcu()``. To see this, consider the following pair of
710
+
threads:
711
+
712
+
::
713
+
714
+
1 void thread0(void)
715
+
2 {
716
+
3 rcu_read_lock();
717
+
4 WRITE_ONCE(x, 1);
718
+
5 rcu_read_unlock();
719
+
6 rcu_read_lock();
720
+
7 WRITE_ONCE(y, 1);
721
+
8 rcu_read_unlock();
722
+
9 }
723
+
10
724
+
11 void thread1(void)
725
+
12 {
726
+
13 rcu_read_lock();
727
+
14 r1 = READ_ONCE(y);
728
+
15 rcu_read_unlock();
729
+
16 rcu_read_lock();
730
+
17 r2 = READ_ONCE(x);
731
+
18 rcu_read_unlock();
732
+
19 }
733
+
734
+
After ``thread0()`` and ``thread1()`` execute concurrently, it is quite
735
+
possible to have
736
+
737
+
::
738
+
739
+
(r1 == 1 && r2 == 0)
740
+
741
+
(that is, ``y`` appears to have been assigned before ``x``), which would
742
+
not be possible if ``rcu_read_lock()`` and ``rcu_read_unlock()`` had
743
+
much in the way of ordering properties. But they do not, so the CPU is
744
+
within its rights to do significant reordering. This is by design: Any
745
+
significant ordering constraints would slow down these fast-path APIs.
746
+
747
+
+-----------------------------------------------------------------------+
748
+
| **Quick Quiz**: |
749
+
+-----------------------------------------------------------------------+
750
+
| Can't the compiler also reorder this code? |
751
+
+-----------------------------------------------------------------------+
752
+
| **Answer**: |
753
+
+-----------------------------------------------------------------------+
754
+
| No, the volatile casts in ``READ_ONCE()`` and ``WRITE_ONCE()`` |
755
+
| prevent the compiler from reordering in this particular case. |
756
+
+-----------------------------------------------------------------------+
757
+
758
+
Readers Do Not Exclude Updaters
759
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
760
+
761
+
Neither ``rcu_read_lock()`` nor ``rcu_read_unlock()`` exclude updates.
762
+
All they do is to prevent grace periods from ending. The following
763
+
example illustrates this:
764
+
765
+
::
766
+
767
+
1 void thread0(void)
768
+
2 {
769
+
3 rcu_read_lock();
770
+
4 r1 = READ_ONCE(y);
771
+
5 if (r1) {
772
+
6 do_something_with_nonzero_x();
773
+
7 r2 = READ_ONCE(x);
774
+
8 WARN_ON(!r2); /* BUG!!! */
775
+
9 }
776
+
10 rcu_read_unlock();
777
+
11 }
778
+
12
779
+
13 void thread1(void)
780
+
14 {
781
+
15 spin_lock(&my_lock);
782
+
16 WRITE_ONCE(x, 1);
783
+
17 WRITE_ONCE(y, 1);
784
+
18 spin_unlock(&my_lock);
785
+
19 }
786
+
787
+
If the ``thread0()`` function's ``rcu_read_lock()`` excluded the
788
+
``thread1()`` function's update, the ``WARN_ON()`` could never fire. But
789
+
the fact is that ``rcu_read_lock()`` does not exclude much of anything
790
+
aside from subsequent grace periods, of which ``thread1()`` has none, so
791
+
the ``WARN_ON()`` can and does fire.
792
+
793
+
Updaters Only Wait For Old Readers
794
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
795
+
796
+
It might be tempting to assume that after ``synchronize_rcu()``
797
+
completes, there are no readers executing. This temptation must be
798
+
avoided because new readers can start immediately after
799
+
``synchronize_rcu()`` starts, and ``synchronize_rcu()`` is under no
800
+
obligation to wait for these new readers.
801
+
802
+
+-----------------------------------------------------------------------+
803
+
| **Quick Quiz**: |
804
+
+-----------------------------------------------------------------------+
805
+
| Suppose that synchronize_rcu() did wait until *all* readers had |
806
+
| completed instead of waiting only on pre-existing readers. For how |
807
+
| long would the updater be able to rely on there being no readers? |
808
+
+-----------------------------------------------------------------------+
809
+
| **Answer**: |
810
+
+-----------------------------------------------------------------------+
811
+
| For no time at all. Even if ``synchronize_rcu()`` were to wait until |
812
+
| all readers had completed, a new reader might start immediately after |
813
+
| ``synchronize_rcu()`` completed. Therefore, the code following |
814
+
| ``synchronize_rcu()`` can *never* rely on there being no readers. |
815
+
+-----------------------------------------------------------------------+
816
+
817
+
Grace Periods Don't Partition Read-Side Critical Sections
818
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
819
+
820
+
It is tempting to assume that if any part of one RCU read-side critical
821
+
section precedes a given grace period, and if any part of another RCU
822
+
read-side critical section follows that same grace period, then all of
823
+
the first RCU read-side critical section must precede all of the second.
824
+
However, this just isn't the case: A single grace period does not
825
+
partition the set of RCU read-side critical sections. An example of this
826
+
situation can be illustrated as follows, where ``x``, ``y``, and ``z``
827
+
are initially all zero:
828
+
829
+
::
830
+
831
+
1 void thread0(void)
832
+
2 {
833
+
3 rcu_read_lock();
834
+
4 WRITE_ONCE(a, 1);
835
+
5 WRITE_ONCE(b, 1);
836
+
6 rcu_read_unlock();
837
+
7 }
838
+
8
839
+
9 void thread1(void)
840
+
10 {
841
+
11 r1 = READ_ONCE(a);
842
+
12 synchronize_rcu();
843
+
13 WRITE_ONCE(c, 1);
844
+
14 }
845
+
15
846
+
16 void thread2(void)
847
+
17 {
848
+
18 rcu_read_lock();
849
+
19 r2 = READ_ONCE(b);
850
+
20 r3 = READ_ONCE(c);
851
+
21 rcu_read_unlock();
852
+
22 }
853
+
854
+
It turns out that the outcome:
855
+
856
+
::
857
+
858
+
(r1 == 1 && r2 == 0 && r3 == 1)
859
+
860
+
is entirely possible. The following figure show how this can happen,
861
+
with each circled ``QS`` indicating the point at which RCU recorded a
862
+
*quiescent state* for each thread, that is, a state in which RCU knows
863
+
that the thread cannot be in the midst of an RCU read-side critical
864
+
section that started before the current grace period:
865
+
866
+
.. kernel-figure:: GPpartitionReaders1.svg
867
+
868
+
If it is necessary to partition RCU read-side critical sections in this
869
+
manner, it is necessary to use two grace periods, where the first grace
870
+
period is known to end before the second grace period starts:
871
+
872
+
::
873
+
874
+
1 void thread0(void)
875
+
2 {
876
+
3 rcu_read_lock();
877
+
4 WRITE_ONCE(a, 1);
878
+
5 WRITE_ONCE(b, 1);
879
+
6 rcu_read_unlock();
880
+
7 }
881
+
8
882
+
9 void thread1(void)
883
+
10 {
884
+
11 r1 = READ_ONCE(a);
885
+
12 synchronize_rcu();
886
+
13 WRITE_ONCE(c, 1);
887
+
14 }
888
+
15
889
+
16 void thread2(void)
890
+
17 {
891
+
18 r2 = READ_ONCE(c);
892
+
19 synchronize_rcu();
893
+
20 WRITE_ONCE(d, 1);
894
+
21 }
895
+
22
896
+
23 void thread3(void)
897
+
24 {
898
+
25 rcu_read_lock();
899
+
26 r3 = READ_ONCE(b);
900
+
27 r4 = READ_ONCE(d);
901
+
28 rcu_read_unlock();
902
+
29 }
903
+
904
+
Here, if ``(r1 == 1)``, then ``thread0()``'s write to ``b`` must happen
905
+
before the end of ``thread1()``'s grace period. If in addition
906
+
``(r4 == 1)``, then ``thread3()``'s read from ``b`` must happen after
907
+
the beginning of ``thread2()``'s grace period. If it is also the case
908
+
that ``(r2 == 1)``, then the end of ``thread1()``'s grace period must
909
+
precede the beginning of ``thread2()``'s grace period. This mean that
910
+
the two RCU read-side critical sections cannot overlap, guaranteeing
911
+
that ``(r3 == 1)``. As a result, the outcome:
912
+
913
+
::
914
+
915
+
(r1 == 1 && r2 == 1 && r3 == 0 && r4 == 1)
916
+
917
+
cannot happen.
918
+
919
+
This non-requirement was also non-premeditated, but became apparent when
920
+
studying RCU's interaction with memory ordering.
921
+
922
+
Read-Side Critical Sections Don't Partition Grace Periods
923
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
924
+
925
+
It is also tempting to assume that if an RCU read-side critical section
926
+
happens between a pair of grace periods, then those grace periods cannot
927
+
overlap. However, this temptation leads nowhere good, as can be
928
+
illustrated by the following, with all variables initially zero:
929
+
930
+
::
931
+
932
+
1 void thread0(void)
933
+
2 {
934
+
3 rcu_read_lock();
935
+
4 WRITE_ONCE(a, 1);
936
+
5 WRITE_ONCE(b, 1);
937
+
6 rcu_read_unlock();
938
+
7 }
939
+
8
940
+
9 void thread1(void)
941
+
10 {
942
+
11 r1 = READ_ONCE(a);
943
+
12 synchronize_rcu();
944
+
13 WRITE_ONCE(c, 1);
945
+
14 }
946
+
15
947
+
16 void thread2(void)
948
+
17 {
949
+
18 rcu_read_lock();
950
+
19 WRITE_ONCE(d, 1);
951
+
20 r2 = READ_ONCE(c);
952
+
21 rcu_read_unlock();
953
+
22 }
954
+
23
955
+
24 void thread3(void)
956
+
25 {
957
+
26 r3 = READ_ONCE(d);
958
+
27 synchronize_rcu();
959
+
28 WRITE_ONCE(e, 1);
960
+
29 }
961
+
30
962
+
31 void thread4(void)
963
+
32 {
964
+
33 rcu_read_lock();
965
+
34 r4 = READ_ONCE(b);
966
+
35 r5 = READ_ONCE(e);
967
+
36 rcu_read_unlock();
968
+
37 }
969
+
970
+
In this case, the outcome:
971
+
972
+
::
973
+
974
+
(r1 == 1 && r2 == 1 && r3 == 1 && r4 == 0 && r5 == 1)
975
+
976
+
is entirely possible, as illustrated below:
977
+
978
+
.. kernel-figure:: ReadersPartitionGP1.svg
979
+
980
+
Again, an RCU read-side critical section can overlap almost all of a
981
+
given grace period, just so long as it does not overlap the entire grace
982
+
period. As a result, an RCU read-side critical section cannot partition
983
+
a pair of RCU grace periods.
984
+
985
+
+-----------------------------------------------------------------------+
986
+
| **Quick Quiz**: |
987
+
+-----------------------------------------------------------------------+
988
+
| How long a sequence of grace periods, each separated by an RCU |
989
+
| read-side critical section, would be required to partition the RCU |
990
+
| read-side critical sections at the beginning and end of the chain? |
991
+
+-----------------------------------------------------------------------+
992
+
| **Answer**: |
993
+
+-----------------------------------------------------------------------+
994
+
| In theory, an infinite number. In practice, an unknown number that is |
995
+
| sensitive to both implementation details and timing considerations. |
996
+
| Therefore, even in practice, RCU users must abide by the theoretical |
997
+
| rather than the practical answer. |
998
+
+-----------------------------------------------------------------------+
999
+
1000
+
Parallelism Facts of Life
1001
+
-------------------------
1002
+
1003
+
These parallelism facts of life are by no means specific to RCU, but the
1004
+
RCU implementation must abide by them. They therefore bear repeating:
1005
+
1006
+
#. Any CPU or task may be delayed at any time, and any attempts to avoid
1007
+
these delays by disabling preemption, interrupts, or whatever are
1008
+
completely futile. This is most obvious in preemptible user-level
1009
+
environments and in virtualized environments (where a given guest
1010
+
OS's VCPUs can be preempted at any time by the underlying
1011
+
hypervisor), but can also happen in bare-metal environments due to
1012
+
ECC errors, NMIs, and other hardware events. Although a delay of more
1013
+
than about 20 seconds can result in splats, the RCU implementation is
1014
+
obligated to use algorithms that can tolerate extremely long delays,
1015
+
but where “extremely long” is not long enough to allow wrap-around
1016
+
when incrementing a 64-bit counter.
1017
+
#. Both the compiler and the CPU can reorder memory accesses. Where it
1018
+
matters, RCU must use compiler directives and memory-barrier
1019
+
instructions to preserve ordering.
1020
+
#. Conflicting writes to memory locations in any given cache line will
1021
+
result in expensive cache misses. Greater numbers of concurrent
1022
+
writes and more-frequent concurrent writes will result in more
1023
+
dramatic slowdowns. RCU is therefore obligated to use algorithms that
1024
+
have sufficient locality to avoid significant performance and
1025
+
scalability problems.
1026
+
#. As a rough rule of thumb, only one CPU's worth of processing may be
1027
+
carried out under the protection of any given exclusive lock. RCU
1028
+
must therefore use scalable locking designs.
1029
+
#. Counters are finite, especially on 32-bit systems. RCU's use of
1030
+
counters must therefore tolerate counter wrap, or be designed such
1031
+
that counter wrap would take way more time than a single system is
1032
+
likely to run. An uptime of ten years is quite possible, a runtime of
1033
+
a century much less so. As an example of the latter, RCU's
1034
+
dyntick-idle nesting counter allows 54 bits for interrupt nesting
1035
+
level (this counter is 64 bits even on a 32-bit system). Overflowing
1036
+
this counter requires 2\ :sup:`54` half-interrupts on a given CPU
1037
+
without that CPU ever going idle. If a half-interrupt happened every
1038
+
microsecond, it would take 570 years of runtime to overflow this
1039
+
counter, which is currently believed to be an acceptably long time.
1040
+
#. Linux systems can have thousands of CPUs running a single Linux
1041
+
kernel in a single shared-memory environment. RCU must therefore pay
1042
+
close attention to high-end scalability.
1043
+
1044
+
This last parallelism fact of life means that RCU must pay special
1045
+
attention to the preceding facts of life. The idea that Linux might
1046
+
scale to systems with thousands of CPUs would have been met with some
1047
+
skepticism in the 1990s, but these requirements would have otherwise
1048
+
have been unsurprising, even in the early 1990s.
1049
+
1050
+
Quality-of-Implementation Requirements
1051
+
--------------------------------------
1052
+
1053
+
These sections list quality-of-implementation requirements. Although an
1054
+
RCU implementation that ignores these requirements could still be used,
1055
+
it would likely be subject to limitations that would make it
1056
+
inappropriate for industrial-strength production use. Classes of
1057
+
quality-of-implementation requirements are as follows:
1058
+
1059
+
#. `Specialization <#Specialization>`__
1060
+
#. `Performance and Scalability <#Performance%20and%20Scalability>`__
1061
+
#. `Forward Progress <#Forward%20Progress>`__
1062
+
#. `Composability <#Composability>`__
1063
+
#. `Corner Cases <#Corner%20Cases>`__
1064
+
1065
+
These classes is covered in the following sections.
1066
+
1067
+
Specialization
1068
+
~~~~~~~~~~~~~~
1069
+
1070
+
RCU is and always has been intended primarily for read-mostly
1071
+
situations, which means that RCU's read-side primitives are optimized,
1072
+
often at the expense of its update-side primitives. Experience thus far
1073
+
is captured by the following list of situations:
1074
+
1075
+
#. Read-mostly data, where stale and inconsistent data is not a problem:
1076
+
RCU works great!
1077
+
#. Read-mostly data, where data must be consistent: RCU works well.
1078
+
#. Read-write data, where data must be consistent: RCU *might* work OK.
1079
+
Or not.
1080
+
#. Write-mostly data, where data must be consistent: RCU is very
1081
+
unlikely to be the right tool for the job, with the following
1082
+
exceptions, where RCU can provide:
1083
+
1084
+
a. Existence guarantees for update-friendly mechanisms.
1085
+
b. Wait-free read-side primitives for real-time use.
1086
+
1087
+
This focus on read-mostly situations means that RCU must interoperate
1088
+
with other synchronization primitives. For example, the ``add_gp()`` and
1089
+
``remove_gp_synchronous()`` examples discussed earlier use RCU to
1090
+
protect readers and locking to coordinate updaters. However, the need
1091
+
extends much farther, requiring that a variety of synchronization
1092
+
primitives be legal within RCU read-side critical sections, including
1093
+
spinlocks, sequence locks, atomic operations, reference counters, and
1094
+
memory barriers.
1095
+
1096
+
+-----------------------------------------------------------------------+
1097
+
| **Quick Quiz**: |
1098
+
+-----------------------------------------------------------------------+
1099
+
| What about sleeping locks? |
1100
+
+-----------------------------------------------------------------------+
1101
+
| **Answer**: |
1102
+
+-----------------------------------------------------------------------+
1103
+
| These are forbidden within Linux-kernel RCU read-side critical |
1104
+
| sections because it is not legal to place a quiescent state (in this |
1105
+
| case, voluntary context switch) within an RCU read-side critical |
1106
+
| section. However, sleeping locks may be used within userspace RCU |
1107
+
| read-side critical sections, and also within Linux-kernel sleepable |
1108
+
| RCU `(SRCU) <#Sleepable%20RCU>`__ read-side critical sections. In |
1109
+
| addition, the -rt patchset turns spinlocks into a sleeping locks so |
1110
+
| that the corresponding critical sections can be preempted, which also |
1111
+
| means that these sleeplockified spinlocks (but not other sleeping |
1112
+
| locks!) may be acquire within -rt-Linux-kernel RCU read-side critical |
1113
+
| sections. |
1114
+
| Note that it *is* legal for a normal RCU read-side critical section |
1115
+
| to conditionally acquire a sleeping locks (as in |
1116
+
| ``mutex_trylock()``), but only as long as it does not loop |
1117
+
| indefinitely attempting to conditionally acquire that sleeping locks. |
1118
+
| The key point is that things like ``mutex_trylock()`` either return |
1119
+
| with the mutex held, or return an error indication if the mutex was |
1120
+
| not immediately available. Either way, ``mutex_trylock()`` returns |
1121
+
| immediately without sleeping. |
1122
+
+-----------------------------------------------------------------------+
1123
+
1124
+
It often comes as a surprise that many algorithms do not require a
1125
+
consistent view of data, but many can function in that mode, with
1126
+
network routing being the poster child. Internet routing algorithms take
1127
+
significant time to propagate updates, so that by the time an update
1128
+
arrives at a given system, that system has been sending network traffic
1129
+
the wrong way for a considerable length of time. Having a few threads
1130
+
continue to send traffic the wrong way for a few more milliseconds is
1131
+
clearly not a problem: In the worst case, TCP retransmissions will
1132
+
eventually get the data where it needs to go. In general, when tracking
1133
+
the state of the universe outside of the computer, some level of
1134
+
inconsistency must be tolerated due to speed-of-light delays if nothing
1135
+
else.
1136
+
1137
+
Furthermore, uncertainty about external state is inherent in many cases.
1138
+
For example, a pair of veterinarians might use heartbeat to determine
1139
+
whether or not a given cat was alive. But how long should they wait
1140
+
after the last heartbeat to decide that the cat is in fact dead? Waiting
1141
+
less than 400 milliseconds makes no sense because this would mean that a
1142
+
relaxed cat would be considered to cycle between death and life more
1143
+
than 100 times per minute. Moreover, just as with human beings, a cat's
1144
+
heart might stop for some period of time, so the exact wait period is a
1145
+
judgment call. One of our pair of veterinarians might wait 30 seconds
1146
+
before pronouncing the cat dead, while the other might insist on waiting
1147
+
a full minute. The two veterinarians would then disagree on the state of
1148
+
the cat during the final 30 seconds of the minute following the last
1149
+
heartbeat.
1150
+
1151
+
Interestingly enough, this same situation applies to hardware. When push
1152
+
comes to shove, how do we tell whether or not some external server has
1153
+
failed? We send messages to it periodically, and declare it failed if we
1154
+
don't receive a response within a given period of time. Policy decisions
1155
+
can usually tolerate short periods of inconsistency. The policy was
1156
+
decided some time ago, and is only now being put into effect, so a few
1157
+
milliseconds of delay is normally inconsequential.
1158
+
1159
+
However, there are algorithms that absolutely must see consistent data.
1160
+
For example, the translation between a user-level SystemV semaphore ID
1161
+
to the corresponding in-kernel data structure is protected by RCU, but
1162
+
it is absolutely forbidden to update a semaphore that has just been
1163
+
removed. In the Linux kernel, this need for consistency is accommodated
1164
+
by acquiring spinlocks located in the in-kernel data structure from
1165
+
within the RCU read-side critical section, and this is indicated by the
1166
+
green box in the figure above. Many other techniques may be used, and
1167
+
are in fact used within the Linux kernel.
1168
+
1169
+
In short, RCU is not required to maintain consistency, and other
1170
+
mechanisms may be used in concert with RCU when consistency is required.
1171
+
RCU's specialization allows it to do its job extremely well, and its
1172
+
ability to interoperate with other synchronization mechanisms allows the
1173
+
right mix of synchronization tools to be used for a given job.
1174
+
1175
+
Performance and Scalability
1176
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~
1177
+
1178
+
Energy efficiency is a critical component of performance today, and
1179
+
Linux-kernel RCU implementations must therefore avoid unnecessarily
1180
+
awakening idle CPUs. I cannot claim that this requirement was
1181
+
premeditated. In fact, I learned of it during a telephone conversation
1182
+
in which I was given “frank and open” feedback on the importance of
1183
+
energy efficiency in battery-powered systems and on specific
1184
+
energy-efficiency shortcomings of the Linux-kernel RCU implementation.
1185
+
In my experience, the battery-powered embedded community will consider
1186
+
any unnecessary wakeups to be extremely unfriendly acts. So much so that
1187
+
mere Linux-kernel-mailing-list posts are insufficient to vent their ire.
1188
+
1189
+
Memory consumption is not particularly important for in most situations,
1190
+
and has become decreasingly so as memory sizes have expanded and memory
1191
+
costs have plummeted. However, as I learned from Matt Mackall's
1192
+
`bloatwatch <http://elinux.org/Linux_Tiny-FAQ>`__ efforts, memory
1193
+
footprint is critically important on single-CPU systems with
1194
+
non-preemptible (``CONFIG_PREEMPT=n``) kernels, and thus `tiny
1195
+
RCU <https://lkml.kernel.org/g/20090113221724.GA15307@linux.vnet.ibm.com>`__
1196
+
was born. Josh Triplett has since taken over the small-memory banner
1197
+
with his `Linux kernel tinification <https://tiny.wiki.kernel.org/>`__
1198
+
project, which resulted in `SRCU <#Sleepable%20RCU>`__ becoming optional
1199
+
for those kernels not needing it.
1200
+
1201
+
The remaining performance requirements are, for the most part,
1202
+
unsurprising. For example, in keeping with RCU's read-side
1203
+
specialization, ``rcu_dereference()`` should have negligible overhead
1204
+
(for example, suppression of a few minor compiler optimizations).
1205
+
Similarly, in non-preemptible environments, ``rcu_read_lock()`` and
1206
+
``rcu_read_unlock()`` should have exactly zero overhead.
1207
+
1208
+
In preemptible environments, in the case where the RCU read-side
1209
+
critical section was not preempted (as will be the case for the
1210
+
highest-priority real-time process), ``rcu_read_lock()`` and
1211
+
``rcu_read_unlock()`` should have minimal overhead. In particular, they
1212
+
should not contain atomic read-modify-write operations, memory-barrier
1213
+
instructions, preemption disabling, interrupt disabling, or backwards
1214
+
branches. However, in the case where the RCU read-side critical section
1215
+
was preempted, ``rcu_read_unlock()`` may acquire spinlocks and disable
1216
+
interrupts. This is why it is better to nest an RCU read-side critical
1217
+
section within a preempt-disable region than vice versa, at least in
1218
+
cases where that critical section is short enough to avoid unduly
1219
+
degrading real-time latencies.
1220
+
1221
+
The ``synchronize_rcu()`` grace-period-wait primitive is optimized for
1222
+
throughput. It may therefore incur several milliseconds of latency in
1223
+
addition to the duration of the longest RCU read-side critical section.
1224
+
On the other hand, multiple concurrent invocations of
1225
+
``synchronize_rcu()`` are required to use batching optimizations so that
1226
+
they can be satisfied by a single underlying grace-period-wait
1227
+
operation. For example, in the Linux kernel, it is not unusual for a
1228
+
single grace-period-wait operation to serve more than `1,000 separate
1229
+
invocations <https://www.usenix.org/conference/2004-usenix-annual-technical-conference/making-rcu-safe-deep-sub-millisecond-response>`__
1230
+
of ``synchronize_rcu()``, thus amortizing the per-invocation overhead
1231
+
down to nearly zero. However, the grace-period optimization is also
1232
+
required to avoid measurable degradation of real-time scheduling and
1233
+
interrupt latencies.
1234
+
1235
+
In some cases, the multi-millisecond ``synchronize_rcu()`` latencies are
1236
+
unacceptable. In these cases, ``synchronize_rcu_expedited()`` may be
1237
+
used instead, reducing the grace-period latency down to a few tens of
1238
+
microseconds on small systems, at least in cases where the RCU read-side
1239
+
critical sections are short. There are currently no special latency
1240
+
requirements for ``synchronize_rcu_expedited()`` on large systems, but,
1241
+
consistent with the empirical nature of the RCU specification, that is
1242
+
subject to change. However, there most definitely are scalability
1243
+
requirements: A storm of ``synchronize_rcu_expedited()`` invocations on
1244
+
4096 CPUs should at least make reasonable forward progress. In return
1245
+
for its shorter latencies, ``synchronize_rcu_expedited()`` is permitted
1246
+
to impose modest degradation of real-time latency on non-idle online
1247
+
CPUs. Here, “modest” means roughly the same latency degradation as a
1248
+
scheduling-clock interrupt.
1249
+
1250
+
There are a number of situations where even
1251
+
``synchronize_rcu_expedited()``'s reduced grace-period latency is
1252
+
unacceptable. In these situations, the asynchronous ``call_rcu()`` can
1253
+
be used in place of ``synchronize_rcu()`` as follows:
1254
+
1255
+
::
1256
+
1257
+
1 struct foo {
1258
+
2 int a;
1259
+
3 int b;
1260
+
4 struct rcu_head rh;
1261
+
5 };
1262
+
6
1263
+
7 static void remove_gp_cb(struct rcu_head *rhp)
1264
+
8 {
1265
+
9 struct foo *p = container_of(rhp, struct foo, rh);
1266
+
10
1267
+
11 kfree(p);
1268
+
12 }
1269
+
13
1270
+
14 bool remove_gp_asynchronous(void)
1271
+
15 {
1272
+
16 struct foo *p;
1273
+
17
1274
+
18 spin_lock(&gp_lock);
1275
+
19 p = rcu_access_pointer(gp);
1276
+
20 if (!p) {
1277
+
21 spin_unlock(&gp_lock);
1278
+
22 return false;
1279
+
23 }
1280
+
24 rcu_assign_pointer(gp, NULL);
1281
+
25 call_rcu(&p->rh, remove_gp_cb);
1282
+
26 spin_unlock(&gp_lock);
1283
+
27 return true;
1284
+
28 }
1285
+
1286
+
A definition of ``struct foo`` is finally needed, and appears on
1287
+
lines 1-5. The function ``remove_gp_cb()`` is passed to ``call_rcu()``
1288
+
on line 25, and will be invoked after the end of a subsequent grace
1289
+
period. This gets the same effect as ``remove_gp_synchronous()``, but
1290
+
without forcing the updater to wait for a grace period to elapse. The
1291
+
``call_rcu()`` function may be used in a number of situations where
1292
+
neither ``synchronize_rcu()`` nor ``synchronize_rcu_expedited()`` would
1293
+
be legal, including within preempt-disable code, ``local_bh_disable()``
1294
+
code, interrupt-disable code, and interrupt handlers. However, even
1295
+
``call_rcu()`` is illegal within NMI handlers and from idle and offline
1296
+
CPUs. The callback function (``remove_gp_cb()`` in this case) will be
1297
+
executed within softirq (software interrupt) environment within the
1298
+
Linux kernel, either within a real softirq handler or under the
1299
+
protection of ``local_bh_disable()``. In both the Linux kernel and in
1300
+
userspace, it is bad practice to write an RCU callback function that
1301
+
takes too long. Long-running operations should be relegated to separate
1302
+
threads or (in the Linux kernel) workqueues.
1303
+
1304
+
+-----------------------------------------------------------------------+
1305
+
| **Quick Quiz**: |
1306
+
+-----------------------------------------------------------------------+
1307
+
| Why does line 19 use ``rcu_access_pointer()``? After all, |
1308
+
| ``call_rcu()`` on line 25 stores into the structure, which would |
1309
+
| interact badly with concurrent insertions. Doesn't this mean that |
1310
+
| ``rcu_dereference()`` is required? |
1311
+
+-----------------------------------------------------------------------+
1312
+
| **Answer**: |
1313
+
+-----------------------------------------------------------------------+
1314
+
| Presumably the ``->gp_lock`` acquired on line 18 excludes any |
1315
+
| changes, including any insertions that ``rcu_dereference()`` would |
1316
+
| protect against. Therefore, any insertions will be delayed until |
1317
+
| after ``->gp_lock`` is released on line 25, which in turn means that |
1318
+
| ``rcu_access_pointer()`` suffices. |
1319
+
+-----------------------------------------------------------------------+
1320
+
1321
+
However, all that ``remove_gp_cb()`` is doing is invoking ``kfree()`` on
1322
+
the data element. This is a common idiom, and is supported by
1323
+
``kfree_rcu()``, which allows “fire and forget” operation as shown
1324
+
below:
1325
+
1326
+
::
1327
+
1328
+
1 struct foo {
1329
+
2 int a;
1330
+
3 int b;
1331
+
4 struct rcu_head rh;
1332
+
5 };
1333
+
6
1334
+
7 bool remove_gp_faf(void)
1335
+
8 {
1336
+
9 struct foo *p;
1337
+
10
1338
+
11 spin_lock(&gp_lock);
1339
+
12 p = rcu_dereference(gp);
1340
+
13 if (!p) {
1341
+
14 spin_unlock(&gp_lock);
1342
+
15 return false;
1343
+
16 }
1344
+
17 rcu_assign_pointer(gp, NULL);
1345
+
18 kfree_rcu(p, rh);
1346
+
19 spin_unlock(&gp_lock);
1347
+
20 return true;
1348
+
21 }
1349
+
1350
+
Note that ``remove_gp_faf()`` simply invokes ``kfree_rcu()`` and
1351
+
proceeds, without any need to pay any further attention to the
1352
+
subsequent grace period and ``kfree()``. It is permissible to invoke
1353
+
``kfree_rcu()`` from the same environments as for ``call_rcu()``.
1354
+
Interestingly enough, DYNIX/ptx had the equivalents of ``call_rcu()``
1355
+
and ``kfree_rcu()``, but not ``synchronize_rcu()``. This was due to the
1356
+
fact that RCU was not heavily used within DYNIX/ptx, so the very few
1357
+
places that needed something like ``synchronize_rcu()`` simply
1358
+
open-coded it.
1359
+
1360
+
+-----------------------------------------------------------------------+
1361
+
| **Quick Quiz**: |
1362
+
+-----------------------------------------------------------------------+
1363
+
| Earlier it was claimed that ``call_rcu()`` and ``kfree_rcu()`` |
1364
+
| allowed updaters to avoid being blocked by readers. But how can that |
1365
+
| be correct, given that the invocation of the callback and the freeing |
1366
+
| of the memory (respectively) must still wait for a grace period to |
1367
+
| elapse? |
1368
+
+-----------------------------------------------------------------------+
1369
+
| **Answer**: |
1370
+
+-----------------------------------------------------------------------+
1371
+
| We could define things this way, but keep in mind that this sort of |
1372
+
| definition would say that updates in garbage-collected languages |
1373
+
| cannot complete until the next time the garbage collector runs, which |
1374
+
| does not seem at all reasonable. The key point is that in most cases, |
1375
+
| an updater using either ``call_rcu()`` or ``kfree_rcu()`` can proceed |
1376
+
| to the next update as soon as it has invoked ``call_rcu()`` or |
1377
+
| ``kfree_rcu()``, without having to wait for a subsequent grace |
1378
+
| period. |
1379
+
+-----------------------------------------------------------------------+
1380
+
1381
+
But what if the updater must wait for the completion of code to be
1382
+
executed after the end of the grace period, but has other tasks that can
1383
+
be carried out in the meantime? The polling-style
1384
+
``get_state_synchronize_rcu()`` and ``cond_synchronize_rcu()`` functions
1385
+
may be used for this purpose, as shown below:
1386
+
1387
+
::
1388
+
1389
+
1 bool remove_gp_poll(void)
1390
+
2 {
1391
+
3 struct foo *p;
1392
+
4 unsigned long s;
1393
+
5
1394
+
6 spin_lock(&gp_lock);
1395
+
7 p = rcu_access_pointer(gp);
1396
+
8 if (!p) {
1397
+
9 spin_unlock(&gp_lock);
1398
+
10 return false;
1399
+
11 }
1400
+
12 rcu_assign_pointer(gp, NULL);
1401
+
13 spin_unlock(&gp_lock);
1402
+
14 s = get_state_synchronize_rcu();
1403
+
15 do_something_while_waiting();
1404
+
16 cond_synchronize_rcu(s);
1405
+
17 kfree(p);
1406
+
18 return true;
1407
+
19 }
1408
+
1409
+
On line 14, ``get_state_synchronize_rcu()`` obtains a “cookie” from RCU,
1410
+
then line 15 carries out other tasks, and finally, line 16 returns
1411
+
immediately if a grace period has elapsed in the meantime, but otherwise
1412
+
waits as required. The need for ``get_state_synchronize_rcu`` and
1413
+
``cond_synchronize_rcu()`` has appeared quite recently, so it is too
1414
+
early to tell whether they will stand the test of time.
1415
+
1416
+
RCU thus provides a range of tools to allow updaters to strike the
1417
+
required tradeoff between latency, flexibility and CPU overhead.
1418
+
1419
+
Forward Progress
1420
+
~~~~~~~~~~~~~~~~
1421
+
1422
+
In theory, delaying grace-period completion and callback invocation is
1423
+
harmless. In practice, not only are memory sizes finite but also
1424
+
callbacks sometimes do wakeups, and sufficiently deferred wakeups can be
1425
+
difficult to distinguish from system hangs. Therefore, RCU must provide
1426
+
a number of mechanisms to promote forward progress.
1427
+
1428
+
These mechanisms are not foolproof, nor can they be. For one simple
1429
+
example, an infinite loop in an RCU read-side critical section must by
1430
+
definition prevent later grace periods from ever completing. For a more
1431
+
involved example, consider a 64-CPU system built with
1432
+
``CONFIG_RCU_NOCB_CPU=y`` and booted with ``rcu_nocbs=1-63``, where
1433
+
CPUs 1 through 63 spin in tight loops that invoke ``call_rcu()``. Even
1434
+
if these tight loops also contain calls to ``cond_resched()`` (thus
1435
+
allowing grace periods to complete), CPU 0 simply will not be able to
1436
+
invoke callbacks as fast as the other 63 CPUs can register them, at
1437
+
least not until the system runs out of memory. In both of these
1438
+
examples, the Spiderman principle applies: With great power comes great
1439
+
responsibility. However, short of this level of abuse, RCU is required
1440
+
to ensure timely completion of grace periods and timely invocation of
1441
+
callbacks.
1442
+
1443
+
RCU takes the following steps to encourage timely completion of grace
1444
+
periods:
1445
+
1446
+
#. If a grace period fails to complete within 100 milliseconds, RCU
1447
+
causes future invocations of ``cond_resched()`` on the holdout CPUs
1448
+
to provide an RCU quiescent state. RCU also causes those CPUs'
1449
+
``need_resched()`` invocations to return ``true``, but only after the
1450
+
corresponding CPU's next scheduling-clock.
1451
+
#. CPUs mentioned in the ``nohz_full`` kernel boot parameter can run
1452
+
indefinitely in the kernel without scheduling-clock interrupts, which
1453
+
defeats the above ``need_resched()`` strategem. RCU will therefore
1454
+
invoke ``resched_cpu()`` on any ``nohz_full`` CPUs still holding out
1455
+
after 109 milliseconds.
1456
+
#. In kernels built with ``CONFIG_RCU_BOOST=y``, if a given task that
1457
+
has been preempted within an RCU read-side critical section is
1458
+
holding out for more than 500 milliseconds, RCU will resort to
1459
+
priority boosting.
1460
+
#. If a CPU is still holding out 10 seconds into the grace period, RCU
1461
+
will invoke ``resched_cpu()`` on it regardless of its ``nohz_full``
1462
+
state.
1463
+
1464
+
The above values are defaults for systems running with ``HZ=1000``. They
1465
+
will vary as the value of ``HZ`` varies, and can also be changed using
1466
+
the relevant Kconfig options and kernel boot parameters. RCU currently
1467
+
does not do much sanity checking of these parameters, so please use
1468
+
caution when changing them. Note that these forward-progress measures
1469
+
are provided only for RCU, not for `SRCU <#Sleepable%20RCU>`__ or `Tasks
1470
+
RCU <#Tasks%20RCU>`__.
1471
+
1472
+
RCU takes the following steps in ``call_rcu()`` to encourage timely
1473
+
invocation of callbacks when any given non-\ ``rcu_nocbs`` CPU has
1474
+
10,000 callbacks, or has 10,000 more callbacks than it had the last time
1475
+
encouragement was provided:
1476
+
1477
+
#. Starts a grace period, if one is not already in progress.
1478
+
#. Forces immediate checking for quiescent states, rather than waiting
1479
+
for three milliseconds to have elapsed since the beginning of the
1480
+
grace period.
1481
+
#. Immediately tags the CPU's callbacks with their grace period
1482
+
completion numbers, rather than waiting for the ``RCU_SOFTIRQ``
1483
+
handler to get around to it.
1484
+
#. Lifts callback-execution batch limits, which speeds up callback
1485
+
invocation at the expense of degrading realtime response.
1486
+
1487
+
Again, these are default values when running at ``HZ=1000``, and can be
1488
+
overridden. Again, these forward-progress measures are provided only for
1489
+
RCU, not for `SRCU <#Sleepable%20RCU>`__ or `Tasks
1490
+
RCU <#Tasks%20RCU>`__. Even for RCU, callback-invocation forward
1491
+
progress for ``rcu_nocbs`` CPUs is much less well-developed, in part
1492
+
because workloads benefiting from ``rcu_nocbs`` CPUs tend to invoke
1493
+
``call_rcu()`` relatively infrequently. If workloads emerge that need
1494
+
both ``rcu_nocbs`` CPUs and high ``call_rcu()`` invocation rates, then
1495
+
additional forward-progress work will be required.
1496
+
1497
+
Composability
1498
+
~~~~~~~~~~~~~
1499
+
1500
+
Composability has received much attention in recent years, perhaps in
1501
+
part due to the collision of multicore hardware with object-oriented
1502
+
techniques designed in single-threaded environments for single-threaded
1503
+
use. And in theory, RCU read-side critical sections may be composed, and
1504
+
in fact may be nested arbitrarily deeply. In practice, as with all
1505
+
real-world implementations of composable constructs, there are
1506
+
limitations.
1507
+
1508
+
Implementations of RCU for which ``rcu_read_lock()`` and
1509
+
``rcu_read_unlock()`` generate no code, such as Linux-kernel RCU when
1510
+
``CONFIG_PREEMPT=n``, can be nested arbitrarily deeply. After all, there
1511
+
is no overhead. Except that if all these instances of
1512
+
``rcu_read_lock()`` and ``rcu_read_unlock()`` are visible to the
1513
+
compiler, compilation will eventually fail due to exhausting memory,
1514
+
mass storage, or user patience, whichever comes first. If the nesting is
1515
+
not visible to the compiler, as is the case with mutually recursive
1516
+
functions each in its own translation unit, stack overflow will result.
1517
+
If the nesting takes the form of loops, perhaps in the guise of tail
1518
+
recursion, either the control variable will overflow or (in the Linux
1519
+
kernel) you will get an RCU CPU stall warning. Nevertheless, this class
1520
+
of RCU implementations is one of the most composable constructs in
1521
+
existence.
1522
+
1523
+
RCU implementations that explicitly track nesting depth are limited by
1524
+
the nesting-depth counter. For example, the Linux kernel's preemptible
1525
+
RCU limits nesting to ``INT_MAX``. This should suffice for almost all
1526
+
practical purposes. That said, a consecutive pair of RCU read-side
1527
+
critical sections between which there is an operation that waits for a
1528
+
grace period cannot be enclosed in another RCU read-side critical
1529
+
section. This is because it is not legal to wait for a grace period
1530
+
within an RCU read-side critical section: To do so would result either
1531
+
in deadlock or in RCU implicitly splitting the enclosing RCU read-side
1532
+
critical section, neither of which is conducive to a long-lived and
1533
+
prosperous kernel.
1534
+
1535
+
It is worth noting that RCU is not alone in limiting composability. For
1536
+
example, many transactional-memory implementations prohibit composing a
1537
+
pair of transactions separated by an irrevocable operation (for example,
1538
+
a network receive operation). For another example, lock-based critical
1539
+
sections can be composed surprisingly freely, but only if deadlock is
1540
+
avoided.
1541
+
1542
+
In short, although RCU read-side critical sections are highly
1543
+
composable, care is required in some situations, just as is the case for
1544
+
any other composable synchronization mechanism.
1545
+
1546
+
Corner Cases
1547
+
~~~~~~~~~~~~
1548
+
1549
+
A given RCU workload might have an endless and intense stream of RCU
1550
+
read-side critical sections, perhaps even so intense that there was
1551
+
never a point in time during which there was not at least one RCU
1552
+
read-side critical section in flight. RCU cannot allow this situation to
1553
+
block grace periods: As long as all the RCU read-side critical sections
1554
+
are finite, grace periods must also be finite.
1555
+
1556
+
That said, preemptible RCU implementations could potentially result in
1557
+
RCU read-side critical sections being preempted for long durations,
1558
+
which has the effect of creating a long-duration RCU read-side critical
1559
+
section. This situation can arise only in heavily loaded systems, but
1560
+
systems using real-time priorities are of course more vulnerable.
1561
+
Therefore, RCU priority boosting is provided to help deal with this
1562
+
case. That said, the exact requirements on RCU priority boosting will
1563
+
likely evolve as more experience accumulates.
1564
+
1565
+
Other workloads might have very high update rates. Although one can
1566
+
argue that such workloads should instead use something other than RCU,
1567
+
the fact remains that RCU must handle such workloads gracefully. This
1568
+
requirement is another factor driving batching of grace periods, but it
1569
+
is also the driving force behind the checks for large numbers of queued
1570
+
RCU callbacks in the ``call_rcu()`` code path. Finally, high update
1571
+
rates should not delay RCU read-side critical sections, although some
1572
+
small read-side delays can occur when using
1573
+
``synchronize_rcu_expedited()``, courtesy of this function's use of
1574
+
``smp_call_function_single()``.
1575
+
1576
+
Although all three of these corner cases were understood in the early
1577
+
1990s, a simple user-level test consisting of ``close(open(path))`` in a
1578
+
tight loop in the early 2000s suddenly provided a much deeper
1579
+
appreciation of the high-update-rate corner case. This test also
1580
+
motivated addition of some RCU code to react to high update rates, for
1581
+
example, if a given CPU finds itself with more than 10,000 RCU callbacks
1582
+
queued, it will cause RCU to take evasive action by more aggressively
1583
+
starting grace periods and more aggressively forcing completion of
1584
+
grace-period processing. This evasive action causes the grace period to
1585
+
complete more quickly, but at the cost of restricting RCU's batching
1586
+
optimizations, thus increasing the CPU overhead incurred by that grace
1587
+
period.
1588
+
1589
+
Software-Engineering Requirements
1590
+
---------------------------------
1591
+
1592
+
Between Murphy's Law and “To err is human”, it is necessary to guard
1593
+
against mishaps and misuse:
1594
+
1595
+
#. It is all too easy to forget to use ``rcu_read_lock()`` everywhere
1596
+
that it is needed, so kernels built with ``CONFIG_PROVE_RCU=y`` will
1597
+
splat if ``rcu_dereference()`` is used outside of an RCU read-side
1598
+
critical section. Update-side code can use
1599
+
``rcu_dereference_protected()``, which takes a `lockdep
1600
+
expression <https://lwn.net/Articles/371986/>`__ to indicate what is
1601
+
providing the protection. If the indicated protection is not
1602
+
provided, a lockdep splat is emitted.
1603
+
Code shared between readers and updaters can use
1604
+
``rcu_dereference_check()``, which also takes a lockdep expression,
1605
+
and emits a lockdep splat if neither ``rcu_read_lock()`` nor the
1606
+
indicated protection is in place. In addition,
1607
+
``rcu_dereference_raw()`` is used in those (hopefully rare) cases
1608
+
where the required protection cannot be easily described. Finally,
1609
+
``rcu_read_lock_held()`` is provided to allow a function to verify
1610
+
that it has been invoked within an RCU read-side critical section. I
1611
+
was made aware of this set of requirements shortly after Thomas
1612
+
Gleixner audited a number of RCU uses.
1613
+
#. A given function might wish to check for RCU-related preconditions
1614
+
upon entry, before using any other RCU API. The
1615
+
``rcu_lockdep_assert()`` does this job, asserting the expression in
1616
+
kernels having lockdep enabled and doing nothing otherwise.
1617
+
#. It is also easy to forget to use ``rcu_assign_pointer()`` and
1618
+
``rcu_dereference()``, perhaps (incorrectly) substituting a simple
1619
+
assignment. To catch this sort of error, a given RCU-protected
1620
+
pointer may be tagged with ``__rcu``, after which sparse will
1621
+
complain about simple-assignment accesses to that pointer. Arnd
1622
+
Bergmann made me aware of this requirement, and also supplied the
1623
+
needed `patch series <https://lwn.net/Articles/376011/>`__.
1624
+
#. Kernels built with ``CONFIG_DEBUG_OBJECTS_RCU_HEAD=y`` will splat if
1625
+
a data element is passed to ``call_rcu()`` twice in a row, without a
1626
+
grace period in between. (This error is similar to a double free.)
1627
+
The corresponding ``rcu_head`` structures that are dynamically
1628
+
allocated are automatically tracked, but ``rcu_head`` structures
1629
+
allocated on the stack must be initialized with
1630
+
``init_rcu_head_on_stack()`` and cleaned up with
1631
+
``destroy_rcu_head_on_stack()``. Similarly, statically allocated
1632
+
non-stack ``rcu_head`` structures must be initialized with
1633
+
``init_rcu_head()`` and cleaned up with ``destroy_rcu_head()``.
1634
+
Mathieu Desnoyers made me aware of this requirement, and also
1635
+
supplied the needed
1636
+
`patch <https://lkml.kernel.org/g/20100319013024.GA28456@Krystal>`__.
1637
+
#. An infinite loop in an RCU read-side critical section will eventually
1638
+
trigger an RCU CPU stall warning splat, with the duration of
1639
+
“eventually” being controlled by the ``RCU_CPU_STALL_TIMEOUT``
1640
+
``Kconfig`` option, or, alternatively, by the
1641
+
``rcupdate.rcu_cpu_stall_timeout`` boot/sysfs parameter. However, RCU
1642
+
is not obligated to produce this splat unless there is a grace period
1643
+
waiting on that particular RCU read-side critical section.
1644
+
1645
+
Some extreme workloads might intentionally delay RCU grace periods,
1646
+
and systems running those workloads can be booted with
1647
+
``rcupdate.rcu_cpu_stall_suppress`` to suppress the splats. This
1648
+
kernel parameter may also be set via ``sysfs``. Furthermore, RCU CPU
1649
+
stall warnings are counter-productive during sysrq dumps and during
1650
+
panics. RCU therefore supplies the ``rcu_sysrq_start()`` and
1651
+
``rcu_sysrq_end()`` API members to be called before and after long
1652
+
sysrq dumps. RCU also supplies the ``rcu_panic()`` notifier that is
1653
+
automatically invoked at the beginning of a panic to suppress further
1654
+
RCU CPU stall warnings.
1655
+
1656
+
This requirement made itself known in the early 1990s, pretty much
1657
+
the first time that it was necessary to debug a CPU stall. That said,
1658
+
the initial implementation in DYNIX/ptx was quite generic in
1659
+
comparison with that of Linux.
1660
+
1661
+
#. Although it would be very good to detect pointers leaking out of RCU
1662
+
read-side critical sections, there is currently no good way of doing
1663
+
this. One complication is the need to distinguish between pointers
1664
+
leaking and pointers that have been handed off from RCU to some other
1665
+
synchronization mechanism, for example, reference counting.
1666
+
#. In kernels built with ``CONFIG_RCU_TRACE=y``, RCU-related information
1667
+
is provided via event tracing.
1668
+
#. Open-coded use of ``rcu_assign_pointer()`` and ``rcu_dereference()``
1669
+
to create typical linked data structures can be surprisingly
1670
+
error-prone. Therefore, RCU-protected `linked
1671
+
lists <https://lwn.net/Articles/609973/#RCU%20List%20APIs>`__ and,
1672
+
more recently, RCU-protected `hash
1673
+
tables <https://lwn.net/Articles/612100/>`__ are available. Many
1674
+
other special-purpose RCU-protected data structures are available in
1675
+
the Linux kernel and the userspace RCU library.
1676
+
#. Some linked structures are created at compile time, but still require
1677
+
``__rcu`` checking. The ``RCU_POINTER_INITIALIZER()`` macro serves
1678
+
this purpose.
1679
+
#. It is not necessary to use ``rcu_assign_pointer()`` when creating
1680
+
linked structures that are to be published via a single external
1681
+
pointer. The ``RCU_INIT_POINTER()`` macro is provided for this task
1682
+
and also for assigning ``NULL`` pointers at runtime.
1683
+
1684
+
This not a hard-and-fast list: RCU's diagnostic capabilities will
1685
+
continue to be guided by the number and type of usage bugs found in
1686
+
real-world RCU usage.
1687
+
1688
+
Linux Kernel Complications
1689
+
--------------------------
1690
+
1691
+
The Linux kernel provides an interesting environment for all kinds of
1692
+
software, including RCU. Some of the relevant points of interest are as
1693
+
follows:
1694
+
1695
+
#. `Configuration <#Configuration>`__.
1696
+
#. `Firmware Interface <#Firmware%20Interface>`__.
1697
+
#. `Early Boot <#Early%20Boot>`__.
1698
+
#. `Interrupts and non-maskable interrupts
1699
+
(NMIs) <#Interrupts%20and%20NMIs>`__.
1700
+
#. `Loadable Modules <#Loadable%20Modules>`__.
1701
+
#. `Hotplug CPU <#Hotplug%20CPU>`__.
1702
+
#. `Scheduler and RCU <#Scheduler%20and%20RCU>`__.
1703
+
#. `Tracing and RCU <#Tracing%20and%20RCU>`__.
1704
+
#. `Energy Efficiency <#Energy%20Efficiency>`__.
1705
+
#. `Scheduling-Clock Interrupts and
1706
+
RCU <#Scheduling-Clock%20Interrupts%20and%20RCU>`__.
1707
+
#. `Memory Efficiency <#Memory%20Efficiency>`__.
1708
+
#. `Performance, Scalability, Response Time, and
1709
+
Reliability <#Performance,%20Scalability,%20Response%20Time,%20and%20Reliability>`__.
1710
+
1711
+
This list is probably incomplete, but it does give a feel for the most
1712
+
notable Linux-kernel complications. Each of the following sections
1713
+
covers one of the above topics.
1714
+
1715
+
Configuration
1716
+
~~~~~~~~~~~~~
1717
+
1718
+
RCU's goal is automatic configuration, so that almost nobody needs to
1719
+
worry about RCU's ``Kconfig`` options. And for almost all users, RCU
1720
+
does in fact work well “out of the box.”
1721
+
1722
+
However, there are specialized use cases that are handled by kernel boot
1723
+
parameters and ``Kconfig`` options. Unfortunately, the ``Kconfig``
1724
+
system will explicitly ask users about new ``Kconfig`` options, which
1725
+
requires almost all of them be hidden behind a ``CONFIG_RCU_EXPERT``
1726
+
``Kconfig`` option.
1727
+
1728
+
This all should be quite obvious, but the fact remains that Linus
1729
+
Torvalds recently had to
1730
+
`remind <https://lkml.kernel.org/g/CA+55aFy4wcCwaL4okTs8wXhGZ5h-ibecy_Meg9C4MNQrUnwMcg@mail.gmail.com>`__
1731
+
me of this requirement.
1732
+
1733
+
Firmware Interface
1734
+
~~~~~~~~~~~~~~~~~~
1735
+
1736
+
In many cases, kernel obtains information about the system from the
1737
+
firmware, and sometimes things are lost in translation. Or the
1738
+
translation is accurate, but the original message is bogus.
1739
+
1740
+
For example, some systems' firmware overreports the number of CPUs,
1741
+
sometimes by a large factor. If RCU naively believed the firmware, as it
1742
+
used to do, it would create too many per-CPU kthreads. Although the
1743
+
resulting system will still run correctly, the extra kthreads needlessly
1744
+
consume memory and can cause confusion when they show up in ``ps``
1745
+
listings.
1746
+
1747
+
RCU must therefore wait for a given CPU to actually come online before
1748
+
it can allow itself to believe that the CPU actually exists. The
1749
+
resulting “ghost CPUs” (which are never going to come online) cause a
1750
+
number of `interesting
1751
+
complications <https://paulmck.livejournal.com/37494.html>`__.
1752
+
1753
+
Early Boot
1754
+
~~~~~~~~~~
1755
+
1756
+
The Linux kernel's boot sequence is an interesting process, and RCU is
1757
+
used early, even before ``rcu_init()`` is invoked. In fact, a number of
1758
+
RCU's primitives can be used as soon as the initial task's
1759
+
``task_struct`` is available and the boot CPU's per-CPU variables are
1760
+
set up. The read-side primitives (``rcu_read_lock()``,
1761
+
``rcu_read_unlock()``, ``rcu_dereference()``, and
1762
+
``rcu_access_pointer()``) will operate normally very early on, as will
1763
+
``rcu_assign_pointer()``.
1764
+
1765
+
Although ``call_rcu()`` may be invoked at any time during boot,
1766
+
callbacks are not guaranteed to be invoked until after all of RCU's
1767
+
kthreads have been spawned, which occurs at ``early_initcall()`` time.
1768
+
This delay in callback invocation is due to the fact that RCU does not
1769
+
invoke callbacks until it is fully initialized, and this full
1770
+
initialization cannot occur until after the scheduler has initialized
1771
+
itself to the point where RCU can spawn and run its kthreads. In theory,
1772
+
it would be possible to invoke callbacks earlier, however, this is not a
1773
+
panacea because there would be severe restrictions on what operations
1774
+
those callbacks could invoke.
1775
+
1776
+
Perhaps surprisingly, ``synchronize_rcu()`` and
1777
+
``synchronize_rcu_expedited()``, will operate normally during very early
1778
+
boot, the reason being that there is only one CPU and preemption is
1779
+
disabled. This means that the call ``synchronize_rcu()`` (or friends)
1780
+
itself is a quiescent state and thus a grace period, so the early-boot
1781
+
implementation can be a no-op.
1782
+
1783
+
However, once the scheduler has spawned its first kthread, this early
1784
+
boot trick fails for ``synchronize_rcu()`` (as well as for
1785
+
``synchronize_rcu_expedited()``) in ``CONFIG_PREEMPT=y`` kernels. The
1786
+
reason is that an RCU read-side critical section might be preempted,
1787
+
which means that a subsequent ``synchronize_rcu()`` really does have to
1788
+
wait for something, as opposed to simply returning immediately.
1789
+
Unfortunately, ``synchronize_rcu()`` can't do this until all of its
1790
+
kthreads are spawned, which doesn't happen until some time during
1791
+
``early_initcalls()`` time. But this is no excuse: RCU is nevertheless
1792
+
required to correctly handle synchronous grace periods during this time
1793
+
period. Once all of its kthreads are up and running, RCU starts running
1794
+
normally.
1795
+
1796
+
+-----------------------------------------------------------------------+
1797
+
| **Quick Quiz**: |
1798
+
+-----------------------------------------------------------------------+
1799
+
| How can RCU possibly handle grace periods before all of its kthreads |
1800
+
| have been spawned??? |
1801
+
+-----------------------------------------------------------------------+
1802
+
| **Answer**: |
1803
+
+-----------------------------------------------------------------------+
1804
+
| Very carefully! |
1805
+
| During the “dead zone” between the time that the scheduler spawns the |
1806
+
| first task and the time that all of RCU's kthreads have been spawned, |
1807
+
| all synchronous grace periods are handled by the expedited |
1808
+
| grace-period mechanism. At runtime, this expedited mechanism relies |
1809
+
| on workqueues, but during the dead zone the requesting task itself |
1810
+
| drives the desired expedited grace period. Because dead-zone |
1811
+
| execution takes place within task context, everything works. Once the |
1812
+
| dead zone ends, expedited grace periods go back to using workqueues, |
1813
+
| as is required to avoid problems that would otherwise occur when a |
1814
+
| user task received a POSIX signal while driving an expedited grace |
1815
+
| period. |
1816
+
| |
1817
+
| And yes, this does mean that it is unhelpful to send POSIX signals to |
1818
+
| random tasks between the time that the scheduler spawns its first |
1819
+
| kthread and the time that RCU's kthreads have all been spawned. If |
1820
+
| there ever turns out to be a good reason for sending POSIX signals |
1821
+
| during that time, appropriate adjustments will be made. (If it turns |
1822
+
| out that POSIX signals are sent during this time for no good reason, |
1823
+
| other adjustments will be made, appropriate or otherwise.) |
1824
+
+-----------------------------------------------------------------------+
1825
+
1826
+
I learned of these boot-time requirements as a result of a series of
1827
+
system hangs.
1828
+
1829
+
Interrupts and NMIs
1830
+
~~~~~~~~~~~~~~~~~~~
1831
+
1832
+
The Linux kernel has interrupts, and RCU read-side critical sections are
1833
+
legal within interrupt handlers and within interrupt-disabled regions of
1834
+
code, as are invocations of ``call_rcu()``.
1835
+
1836
+
Some Linux-kernel architectures can enter an interrupt handler from
1837
+
non-idle process context, and then just never leave it, instead
1838
+
stealthily transitioning back to process context. This trick is
1839
+
sometimes used to invoke system calls from inside the kernel. These
1840
+
“half-interrupts” mean that RCU has to be very careful about how it
1841
+
counts interrupt nesting levels. I learned of this requirement the hard
1842
+
way during a rewrite of RCU's dyntick-idle code.
1843
+
1844
+
The Linux kernel has non-maskable interrupts (NMIs), and RCU read-side
1845
+
critical sections are legal within NMI handlers. Thankfully, RCU
1846
+
update-side primitives, including ``call_rcu()``, are prohibited within
1847
+
NMI handlers.
1848
+
1849
+
The name notwithstanding, some Linux-kernel architectures can have
1850
+
nested NMIs, which RCU must handle correctly. Andy Lutomirski `surprised
1851
+
me <https://lkml.kernel.org/r/CALCETrXLq1y7e_dKFPgou-FKHB6Pu-r8+t-6Ds+8=va7anBWDA@mail.gmail.com>`__
1852
+
with this requirement; he also kindly surprised me with `an
1853
+
algorithm <https://lkml.kernel.org/r/CALCETrXSY9JpW3uE6H8WYk81sg56qasA2aqmjMPsq5dOtzso=g@mail.gmail.com>`__
1854
+
that meets this requirement.
1855
+
1856
+
Furthermore, NMI handlers can be interrupted by what appear to RCU to be
1857
+
normal interrupts. One way that this can happen is for code that
1858
+
directly invokes ``rcu_irq_enter()`` and ``rcu_irq_exit()`` to be called
1859
+
from an NMI handler. This astonishing fact of life prompted the current
1860
+
code structure, which has ``rcu_irq_enter()`` invoking
1861
+
``rcu_nmi_enter()`` and ``rcu_irq_exit()`` invoking ``rcu_nmi_exit()``.
1862
+
And yes, I also learned of this requirement the hard way.
1863
+
1864
+
Loadable Modules
1865
+
~~~~~~~~~~~~~~~~
1866
+
1867
+
The Linux kernel has loadable modules, and these modules can also be
1868
+
unloaded. After a given module has been unloaded, any attempt to call
1869
+
one of its functions results in a segmentation fault. The module-unload
1870
+
functions must therefore cancel any delayed calls to loadable-module
1871
+
functions, for example, any outstanding ``mod_timer()`` must be dealt
1872
+
with via ``del_timer_sync()`` or similar.
1873
+
1874
+
Unfortunately, there is no way to cancel an RCU callback; once you
1875
+
invoke ``call_rcu()``, the callback function is eventually going to be
1876
+
invoked, unless the system goes down first. Because it is normally
1877
+
considered socially irresponsible to crash the system in response to a
1878
+
module unload request, we need some other way to deal with in-flight RCU
1879
+
callbacks.
1880
+
1881
+
RCU therefore provides ``rcu_barrier()``, which waits until all
1882
+
in-flight RCU callbacks have been invoked. If a module uses
1883
+
``call_rcu()``, its exit function should therefore prevent any future
1884
+
invocation of ``call_rcu()``, then invoke ``rcu_barrier()``. In theory,
1885
+
the underlying module-unload code could invoke ``rcu_barrier()``
1886
+
unconditionally, but in practice this would incur unacceptable
1887
+
latencies.
1888
+
1889
+
Nikita Danilov noted this requirement for an analogous
1890
+
filesystem-unmount situation, and Dipankar Sarma incorporated
1891
+
``rcu_barrier()`` into RCU. The need for ``rcu_barrier()`` for module
1892
+
unloading became apparent later.
1893
+
1894
+
.. important::
1895
+
1896
+
The ``rcu_barrier()`` function is not, repeat,
1897
+
*not*, obligated to wait for a grace period. It is instead only required
1898
+
to wait for RCU callbacks that have already been posted. Therefore, if
1899
+
there are no RCU callbacks posted anywhere in the system,
1900
+
``rcu_barrier()`` is within its rights to return immediately. Even if
1901
+
there are callbacks posted, ``rcu_barrier()`` does not necessarily need
1902
+
to wait for a grace period.
1903
+
1904
+
+-----------------------------------------------------------------------+
1905
+
| **Quick Quiz**: |
1906
+
+-----------------------------------------------------------------------+
1907
+
| Wait a minute! Each RCU callbacks must wait for a grace period to |
1908
+
| complete, and ``rcu_barrier()`` must wait for each pre-existing |
1909
+
| callback to be invoked. Doesn't ``rcu_barrier()`` therefore need to |
1910
+
| wait for a full grace period if there is even one callback posted |
1911
+
| anywhere in the system? |
1912
+
+-----------------------------------------------------------------------+
1913
+
| **Answer**: |
1914
+
+-----------------------------------------------------------------------+
1915
+
| Absolutely not!!! |
1916
+
| Yes, each RCU callbacks must wait for a grace period to complete, but |
1917
+
| it might well be partly (or even completely) finished waiting by the |
1918
+
| time ``rcu_barrier()`` is invoked. In that case, ``rcu_barrier()`` |
1919
+
| need only wait for the remaining portion of the grace period to |
1920
+
| elapse. So even if there are quite a few callbacks posted, |
1921
+
| ``rcu_barrier()`` might well return quite quickly. |
1922
+
| |
1923
+
| So if you need to wait for a grace period as well as for all |
1924
+
| pre-existing callbacks, you will need to invoke both |
1925
+
| ``synchronize_rcu()`` and ``rcu_barrier()``. If latency is a concern, |
1926
+
| you can always use workqueues to invoke them concurrently. |
1927
+
+-----------------------------------------------------------------------+
1928
+
1929
+
Hotplug CPU
1930
+
~~~~~~~~~~~
1931
+
1932
+
The Linux kernel supports CPU hotplug, which means that CPUs can come
1933
+
and go. It is of course illegal to use any RCU API member from an
1934
+
offline CPU, with the exception of `SRCU <#Sleepable%20RCU>`__ read-side
1935
+
critical sections. This requirement was present from day one in
1936
+
DYNIX/ptx, but on the other hand, the Linux kernel's CPU-hotplug
1937
+
implementation is “interesting.”
1938
+
1939
+
The Linux-kernel CPU-hotplug implementation has notifiers that are used
1940
+
to allow the various kernel subsystems (including RCU) to respond
1941
+
appropriately to a given CPU-hotplug operation. Most RCU operations may
1942
+
be invoked from CPU-hotplug notifiers, including even synchronous
1943
+
grace-period operations such as ``synchronize_rcu()`` and
1944
+
``synchronize_rcu_expedited()``.
1945
+
1946
+
However, all-callback-wait operations such as ``rcu_barrier()`` are also
1947
+
not supported, due to the fact that there are phases of CPU-hotplug
1948
+
operations where the outgoing CPU's callbacks will not be invoked until
1949
+
after the CPU-hotplug operation ends, which could also result in
1950
+
deadlock. Furthermore, ``rcu_barrier()`` blocks CPU-hotplug operations
1951
+
during its execution, which results in another type of deadlock when
1952
+
invoked from a CPU-hotplug notifier.
1953
+
1954
+
Scheduler and RCU
1955
+
~~~~~~~~~~~~~~~~~
1956
+
1957
+
RCU depends on the scheduler, and the scheduler uses RCU to protect some
1958
+
of its data structures. The preemptible-RCU ``rcu_read_unlock()``
1959
+
implementation must therefore be written carefully to avoid deadlocks
1960
+
involving the scheduler's runqueue and priority-inheritance locks. In
1961
+
particular, ``rcu_read_unlock()`` must tolerate an interrupt where the
1962
+
interrupt handler invokes both ``rcu_read_lock()`` and
1963
+
``rcu_read_unlock()``. This possibility requires ``rcu_read_unlock()``
1964
+
to use negative nesting levels to avoid destructive recursion via
1965
+
interrupt handler's use of RCU.
1966
+
1967
+
This scheduler-RCU requirement came as a `complete
1968
+
surprise <https://lwn.net/Articles/453002/>`__.
1969
+
1970
+
As noted above, RCU makes use of kthreads, and it is necessary to avoid
1971
+
excessive CPU-time accumulation by these kthreads. This requirement was
1972
+
no surprise, but RCU's violation of it when running context-switch-heavy
1973
+
workloads when built with ``CONFIG_NO_HZ_FULL=y`` `did come as a
1974
+
surprise
1975
+
[PDF] <http://www.rdrop.com/users/paulmck/scalability/paper/BareMetal.2015.01.15b.pdf>`__.
1976
+
RCU has made good progress towards meeting this requirement, even for
1977
+
context-switch-heavy ``CONFIG_NO_HZ_FULL=y`` workloads, but there is
1978
+
room for further improvement.
1979
+
1980
+
It is forbidden to hold any of scheduler's runqueue or
1981
+
priority-inheritance spinlocks across an ``rcu_read_unlock()`` unless
1982
+
interrupts have been disabled across the entire RCU read-side critical
1983
+
section, that is, up to and including the matching ``rcu_read_lock()``.
1984
+
Violating this restriction can result in deadlocks involving these
1985
+
scheduler spinlocks. There was hope that this restriction might be
1986
+
lifted when interrupt-disabled calls to ``rcu_read_unlock()`` started
1987
+
deferring the reporting of the resulting RCU-preempt quiescent state
1988
+
until the end of the corresponding interrupts-disabled region.
1989
+
Unfortunately, timely reporting of the corresponding quiescent state to
1990
+
expedited grace periods requires a call to ``raise_softirq()``, which
1991
+
can acquire these scheduler spinlocks. In addition, real-time systems
1992
+
using RCU priority boosting need this restriction to remain in effect
1993
+
because deferred quiescent-state reporting would also defer deboosting,
1994
+
which in turn would degrade real-time latencies.
1995
+
1996
+
In theory, if a given RCU read-side critical section could be guaranteed
1997
+
to be less than one second in duration, holding a scheduler spinlock
1998
+
across that critical section's ``rcu_read_unlock()`` would require only
1999
+
that preemption be disabled across the entire RCU read-side critical
2000
+
section, not interrupts. Unfortunately, given the possibility of vCPU
2001
+
preemption, long-running interrupts, and so on, it is not possible in
2002
+
practice to guarantee that a given RCU read-side critical section will
2003
+
complete in less than one second. Therefore, as noted above, if
2004
+
scheduler spinlocks are held across a given call to
2005
+
``rcu_read_unlock()``, interrupts must be disabled across the entire RCU
2006
+
read-side critical section.
2007
+
2008
+
Tracing and RCU
2009
+
~~~~~~~~~~~~~~~
2010
+
2011
+
It is possible to use tracing on RCU code, but tracing itself uses RCU.
2012
+
For this reason, ``rcu_dereference_raw_notrace()`` is provided for use
2013
+
by tracing, which avoids the destructive recursion that could otherwise
2014
+
ensue. This API is also used by virtualization in some architectures,
2015
+
where RCU readers execute in environments in which tracing cannot be
2016
+
used. The tracing folks both located the requirement and provided the
2017
+
needed fix, so this surprise requirement was relatively painless.
2018
+
2019
+
Energy Efficiency
2020
+
~~~~~~~~~~~~~~~~~
2021
+
2022
+
Interrupting idle CPUs is considered socially unacceptable, especially
2023
+
by people with battery-powered embedded systems. RCU therefore conserves
2024
+
energy by detecting which CPUs are idle, including tracking CPUs that
2025
+
have been interrupted from idle. This is a large part of the
2026
+
energy-efficiency requirement, so I learned of this via an irate phone
2027
+
call.
2028
+
2029
+
Because RCU avoids interrupting idle CPUs, it is illegal to execute an
2030
+
RCU read-side critical section on an idle CPU. (Kernels built with
2031
+
``CONFIG_PROVE_RCU=y`` will splat if you try it.) The ``RCU_NONIDLE()``
2032
+
macro and ``_rcuidle`` event tracing is provided to work around this
2033
+
restriction. In addition, ``rcu_is_watching()`` may be used to test
2034
+
whether or not it is currently legal to run RCU read-side critical
2035
+
sections on this CPU. I learned of the need for diagnostics on the one
2036
+
hand and ``RCU_NONIDLE()`` on the other while inspecting idle-loop code.
2037
+
Steven Rostedt supplied ``_rcuidle`` event tracing, which is used quite
2038
+
heavily in the idle loop. However, there are some restrictions on the
2039
+
code placed within ``RCU_NONIDLE()``:
2040
+
2041
+
#. Blocking is prohibited. In practice, this is not a serious
2042
+
restriction given that idle tasks are prohibited from blocking to
2043
+
begin with.
2044
+
#. Although nesting ``RCU_NONIDLE()`` is permitted, they cannot nest
2045
+
indefinitely deeply. However, given that they can be nested on the
2046
+
order of a million deep, even on 32-bit systems, this should not be a
2047
+
serious restriction. This nesting limit would probably be reached
2048
+
long after the compiler OOMed or the stack overflowed.
2049
+
#. Any code path that enters ``RCU_NONIDLE()`` must sequence out of that
2050
+
same ``RCU_NONIDLE()``. For example, the following is grossly
2051
+
illegal:
2052
+
2053
+
::
2054
+
2055
+
1 RCU_NONIDLE({
2056
+
2 do_something();
2057
+
3 goto bad_idea; /* BUG!!! */
2058
+
4 do_something_else();});
2059
+
5 bad_idea:
2060
+
2061
+
2062
+
It is just as illegal to transfer control into the middle of
2063
+
``RCU_NONIDLE()``'s argument. Yes, in theory, you could transfer in
2064
+
as long as you also transferred out, but in practice you could also
2065
+
expect to get sharply worded review comments.
2066
+
2067
+
It is similarly socially unacceptable to interrupt an ``nohz_full`` CPU
2068
+
running in userspace. RCU must therefore track ``nohz_full`` userspace
2069
+
execution. RCU must therefore be able to sample state at two points in
2070
+
time, and be able to determine whether or not some other CPU spent any
2071
+
time idle and/or executing in userspace.
2072
+
2073
+
These energy-efficiency requirements have proven quite difficult to
2074
+
understand and to meet, for example, there have been more than five
2075
+
clean-sheet rewrites of RCU's energy-efficiency code, the last of which
2076
+
was finally able to demonstrate `real energy savings running on real
2077
+
hardware
2078
+
[PDF] <http://www.rdrop.com/users/paulmck/realtime/paper/AMPenergy.2013.04.19a.pdf>`__.
2079
+
As noted earlier, I learned of many of these requirements via angry
2080
+
phone calls: Flaming me on the Linux-kernel mailing list was apparently
2081
+
not sufficient to fully vent their ire at RCU's energy-efficiency bugs!
2082
+
2083
+
Scheduling-Clock Interrupts and RCU
2084
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2085
+
2086
+
The kernel transitions between in-kernel non-idle execution, userspace
2087
+
execution, and the idle loop. Depending on kernel configuration, RCU
2088
+
handles these states differently:
2089
+
2090
+
+-----------------+------------------+------------------+-----------------+
2091
+
| ``HZ`` Kconfig | In-Kernel | Usermode | Idle |
2092
+
+=================+==================+==================+=================+
2093
+
| ``HZ_PERIODIC`` | Can rely on | Can rely on | Can rely on |
2094
+
| | scheduling-clock | scheduling-clock | RCU's |
2095
+
| | interrupt. | interrupt and | dyntick-idle |
2096
+
| | | its detection | detection. |
2097
+
| | | of interrupt | |
2098
+
| | | from usermode. | |
2099
+
+-----------------+------------------+------------------+-----------------+
2100
+
| ``NO_HZ_IDLE`` | Can rely on | Can rely on | Can rely on |
2101
+
| | scheduling-clock | scheduling-clock | RCU's |
2102
+
| | interrupt. | interrupt and | dyntick-idle |
2103
+
| | | its detection | detection. |
2104
+
| | | of interrupt | |
2105
+
| | | from usermode. | |
2106
+
+-----------------+------------------+------------------+-----------------+
2107
+
| ``NO_HZ_FULL`` | Can only | Can rely on | Can rely on |
2108
+
| | sometimes rely | RCU's | RCU's |
2109
+
| | on | dyntick-idle | dyntick-idle |
2110
+
| | scheduling-clock | detection. | detection. |
2111
+
| | interrupt. In | | |
2112
+
| | other cases, it | | |
2113
+
| | is necessary to | | |
2114
+
| | bound kernel | | |
2115
+
| | execution times | | |
2116
+
| | and/or use | | |
2117
+
| | IPIs. | | |
2118
+
+-----------------+------------------+------------------+-----------------+
2119
+
2120
+
+-----------------------------------------------------------------------+
2121
+
| **Quick Quiz**: |
2122
+
+-----------------------------------------------------------------------+
2123
+
| Why can't ``NO_HZ_FULL`` in-kernel execution rely on the |
2124
+
| scheduling-clock interrupt, just like ``HZ_PERIODIC`` and |
2125
+
| ``NO_HZ_IDLE`` do? |
2126
+
+-----------------------------------------------------------------------+
2127
+
| **Answer**: |
2128
+
+-----------------------------------------------------------------------+
2129
+
| Because, as a performance optimization, ``NO_HZ_FULL`` does not |
2130
+
| necessarily re-enable the scheduling-clock interrupt on entry to each |
2131
+
| and every system call. |
2132
+
+-----------------------------------------------------------------------+
2133
+
2134
+
However, RCU must be reliably informed as to whether any given CPU is
2135
+
currently in the idle loop, and, for ``NO_HZ_FULL``, also whether that
2136
+
CPU is executing in usermode, as discussed
2137
+
`earlier <#Energy%20Efficiency>`__. It also requires that the
2138
+
scheduling-clock interrupt be enabled when RCU needs it to be:
2139
+
2140
+
#. If a CPU is either idle or executing in usermode, and RCU believes it
2141
+
is non-idle, the scheduling-clock tick had better be running.
2142
+
Otherwise, you will get RCU CPU stall warnings. Or at best, very long
2143
+
(11-second) grace periods, with a pointless IPI waking the CPU from
2144
+
time to time.
2145
+
#. If a CPU is in a portion of the kernel that executes RCU read-side
2146
+
critical sections, and RCU believes this CPU to be idle, you will get
2147
+
random memory corruption. **DON'T DO THIS!!!**
2148
+
This is one reason to test with lockdep, which will complain about
2149
+
this sort of thing.
2150
+
#. If a CPU is in a portion of the kernel that is absolutely positively
2151
+
no-joking guaranteed to never execute any RCU read-side critical
2152
+
sections, and RCU believes this CPU to to be idle, no problem. This
2153
+
sort of thing is used by some architectures for light-weight
2154
+
exception handlers, which can then avoid the overhead of
2155
+
``rcu_irq_enter()`` and ``rcu_irq_exit()`` at exception entry and
2156
+
exit, respectively. Some go further and avoid the entireties of
2157
+
``irq_enter()`` and ``irq_exit()``.
2158
+
Just make very sure you are running some of your tests with
2159
+
``CONFIG_PROVE_RCU=y``, just in case one of your code paths was in
2160
+
fact joking about not doing RCU read-side critical sections.
2161
+
#. If a CPU is executing in the kernel with the scheduling-clock
2162
+
interrupt disabled and RCU believes this CPU to be non-idle, and if
2163
+
the CPU goes idle (from an RCU perspective) every few jiffies, no
2164
+
problem. It is usually OK for there to be the occasional gap between
2165
+
idle periods of up to a second or so.
2166
+
If the gap grows too long, you get RCU CPU stall warnings.
2167
+
#. If a CPU is either idle or executing in usermode, and RCU believes it
2168
+
to be idle, of course no problem.
2169
+
#. If a CPU is executing in the kernel, the kernel code path is passing
2170
+
through quiescent states at a reasonable frequency (preferably about
2171
+
once per few jiffies, but the occasional excursion to a second or so
2172
+
is usually OK) and the scheduling-clock interrupt is enabled, of
2173
+
course no problem.
2174
+
If the gap between a successive pair of quiescent states grows too
2175
+
long, you get RCU CPU stall warnings.
2176
+
2177
+
+-----------------------------------------------------------------------+
2178
+
| **Quick Quiz**: |
2179
+
+-----------------------------------------------------------------------+
2180
+
| But what if my driver has a hardware interrupt handler that can run |
2181
+
| for many seconds? I cannot invoke ``schedule()`` from an hardware |
2182
+
| interrupt handler, after all! |
2183
+
+-----------------------------------------------------------------------+
2184
+
| **Answer**: |
2185
+
+-----------------------------------------------------------------------+
2186
+
| One approach is to do ``rcu_irq_exit();rcu_irq_enter();`` every so |
2187
+
| often. But given that long-running interrupt handlers can cause other |
2188
+
| problems, not least for response time, shouldn't you work to keep |
2189
+
| your interrupt handler's runtime within reasonable bounds? |
2190
+
+-----------------------------------------------------------------------+
2191
+
2192
+
But as long as RCU is properly informed of kernel state transitions
2193
+
between in-kernel execution, usermode execution, and idle, and as long
2194
+
as the scheduling-clock interrupt is enabled when RCU needs it to be,
2195
+
you can rest assured that the bugs you encounter will be in some other
2196
+
part of RCU or some other part of the kernel!
2197
+
2198
+
Memory Efficiency
2199
+
~~~~~~~~~~~~~~~~~
2200
+
2201
+
Although small-memory non-realtime systems can simply use Tiny RCU, code
2202
+
size is only one aspect of memory efficiency. Another aspect is the size
2203
+
of the ``rcu_head`` structure used by ``call_rcu()`` and
2204
+
``kfree_rcu()``. Although this structure contains nothing more than a
2205
+
pair of pointers, it does appear in many RCU-protected data structures,
2206
+
including some that are size critical. The ``page`` structure is a case
2207
+
in point, as evidenced by the many occurrences of the ``union`` keyword
2208
+
within that structure.
2209
+
2210
+
This need for memory efficiency is one reason that RCU uses hand-crafted
2211
+
singly linked lists to track the ``rcu_head`` structures that are
2212
+
waiting for a grace period to elapse. It is also the reason why
2213
+
``rcu_head`` structures do not contain debug information, such as fields
2214
+
tracking the file and line of the ``call_rcu()`` or ``kfree_rcu()`` that
2215
+
posted them. Although this information might appear in debug-only kernel
2216
+
builds at some point, in the meantime, the ``->func`` field will often
2217
+
provide the needed debug information.
2218
+
2219
+
However, in some cases, the need for memory efficiency leads to even
2220
+
more extreme measures. Returning to the ``page`` structure, the
2221
+
``rcu_head`` field shares storage with a great many other structures
2222
+
that are used at various points in the corresponding page's lifetime. In
2223
+
order to correctly resolve certain `race
2224
+
conditions <https://lkml.kernel.org/g/1439976106-137226-1-git-send-email-kirill.shutemov@linux.intel.com>`__,
2225
+
the Linux kernel's memory-management subsystem needs a particular bit to
2226
+
remain zero during all phases of grace-period processing, and that bit
2227
+
happens to map to the bottom bit of the ``rcu_head`` structure's
2228
+
``->next`` field. RCU makes this guarantee as long as ``call_rcu()`` is
2229
+
used to post the callback, as opposed to ``kfree_rcu()`` or some future
2230
+
“lazy” variant of ``call_rcu()`` that might one day be created for
2231
+
energy-efficiency purposes.
2232
+
2233
+
That said, there are limits. RCU requires that the ``rcu_head``
2234
+
structure be aligned to a two-byte boundary, and passing a misaligned
2235
+
``rcu_head`` structure to one of the ``call_rcu()`` family of functions
2236
+
will result in a splat. It is therefore necessary to exercise caution
2237
+
when packing structures containing fields of type ``rcu_head``. Why not
2238
+
a four-byte or even eight-byte alignment requirement? Because the m68k
2239
+
architecture provides only two-byte alignment, and thus acts as
2240
+
alignment's least common denominator.
2241
+
2242
+
The reason for reserving the bottom bit of pointers to ``rcu_head``
2243
+
structures is to leave the door open to “lazy” callbacks whose
2244
+
invocations can safely be deferred. Deferring invocation could
2245
+
potentially have energy-efficiency benefits, but only if the rate of
2246
+
non-lazy callbacks decreases significantly for some important workload.
2247
+
In the meantime, reserving the bottom bit keeps this option open in case
2248
+
it one day becomes useful.
2249
+
2250
+
Performance, Scalability, Response Time, and Reliability
2251
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2252
+
2253
+
Expanding on the `earlier
2254
+
discussion <#Performance%20and%20Scalability>`__, RCU is used heavily by
2255
+
hot code paths in performance-critical portions of the Linux kernel's
2256
+
networking, security, virtualization, and scheduling code paths. RCU
2257
+
must therefore use efficient implementations, especially in its
2258
+
read-side primitives. To that end, it would be good if preemptible RCU's
2259
+
implementation of ``rcu_read_lock()`` could be inlined, however, doing
2260
+
this requires resolving ``#include`` issues with the ``task_struct``
2261
+
structure.
2262
+
2263
+
The Linux kernel supports hardware configurations with up to 4096 CPUs,
2264
+
which means that RCU must be extremely scalable. Algorithms that involve
2265
+
frequent acquisitions of global locks or frequent atomic operations on
2266
+
global variables simply cannot be tolerated within the RCU
2267
+
implementation. RCU therefore makes heavy use of a combining tree based
2268
+
on the ``rcu_node`` structure. RCU is required to tolerate all CPUs
2269
+
continuously invoking any combination of RCU's runtime primitives with
2270
+
minimal per-operation overhead. In fact, in many cases, increasing load
2271
+
must *decrease* the per-operation overhead, witness the batching
2272
+
optimizations for ``synchronize_rcu()``, ``call_rcu()``,
2273
+
``synchronize_rcu_expedited()``, and ``rcu_barrier()``. As a general
2274
+
rule, RCU must cheerfully accept whatever the rest of the Linux kernel
2275
+
decides to throw at it.
2276
+
2277
+
The Linux kernel is used for real-time workloads, especially in
2278
+
conjunction with the `-rt
2279
+
patchset <https://rt.wiki.kernel.org/index.php/Main_Page>`__. The
2280
+
real-time-latency response requirements are such that the traditional
2281
+
approach of disabling preemption across RCU read-side critical sections
2282
+
is inappropriate. Kernels built with ``CONFIG_PREEMPT=y`` therefore use
2283
+
an RCU implementation that allows RCU read-side critical sections to be
2284
+
preempted. This requirement made its presence known after users made it
2285
+
clear that an earlier `real-time
2286
+
patch <https://lwn.net/Articles/107930/>`__ did not meet their needs, in
2287
+
conjunction with some `RCU
2288
+
issues <https://lkml.kernel.org/g/20050318002026.GA2693@us.ibm.com>`__
2289
+
encountered by a very early version of the -rt patchset.
2290
+
2291
+
In addition, RCU must make do with a sub-100-microsecond real-time
2292
+
latency budget. In fact, on smaller systems with the -rt patchset, the
2293
+
Linux kernel provides sub-20-microsecond real-time latencies for the
2294
+
whole kernel, including RCU. RCU's scalability and latency must
2295
+
therefore be sufficient for these sorts of configurations. To my
2296
+
surprise, the sub-100-microsecond real-time latency budget `applies to
2297
+
even the largest systems
2298
+
[PDF] <http://www.rdrop.com/users/paulmck/realtime/paper/bigrt.2013.01.31a.LCA.pdf>`__,
2299
+
up to and including systems with 4096 CPUs. This real-time requirement
2300
+
motivated the grace-period kthread, which also simplified handling of a
2301
+
number of race conditions.
2302
+
2303
+
RCU must avoid degrading real-time response for CPU-bound threads,
2304
+
whether executing in usermode (which is one use case for
2305
+
``CONFIG_NO_HZ_FULL=y``) or in the kernel. That said, CPU-bound loops in
2306
+
the kernel must execute ``cond_resched()`` at least once per few tens of
2307
+
milliseconds in order to avoid receiving an IPI from RCU.
2308
+
2309
+
Finally, RCU's status as a synchronization primitive means that any RCU
2310
+
failure can result in arbitrary memory corruption that can be extremely
2311
+
difficult to debug. This means that RCU must be extremely reliable,
2312
+
which in practice also means that RCU must have an aggressive
2313
+
stress-test suite. This stress-test suite is called ``rcutorture``.
2314
+
2315
+
Although the need for ``rcutorture`` was no surprise, the current
2316
+
immense popularity of the Linux kernel is posing interesting—and perhaps
2317
+
unprecedented—validation challenges. To see this, keep in mind that
2318
+
there are well over one billion instances of the Linux kernel running
2319
+
today, given Android smartphones, Linux-powered televisions, and
2320
+
servers. This number can be expected to increase sharply with the advent
2321
+
of the celebrated Internet of Things.
2322
+
2323
+
Suppose that RCU contains a race condition that manifests on average
2324
+
once per million years of runtime. This bug will be occurring about
2325
+
three times per *day* across the installed base. RCU could simply hide
2326
+
behind hardware error rates, given that no one should really expect
2327
+
their smartphone to last for a million years. However, anyone taking too
2328
+
much comfort from this thought should consider the fact that in most
2329
+
jurisdictions, a successful multi-year test of a given mechanism, which
2330
+
might include a Linux kernel, suffices for a number of types of
2331
+
safety-critical certifications. In fact, rumor has it that the Linux
2332
+
kernel is already being used in production for safety-critical
2333
+
applications. I don't know about you, but I would feel quite bad if a
2334
+
bug in RCU killed someone. Which might explain my recent focus on
2335
+
validation and verification.
2336
+
2337
+
Other RCU Flavors
2338
+
-----------------
2339
+
2340
+
One of the more surprising things about RCU is that there are now no
2341
+
fewer than five *flavors*, or API families. In addition, the primary
2342
+
flavor that has been the sole focus up to this point has two different
2343
+
implementations, non-preemptible and preemptible. The other four flavors
2344
+
are listed below, with requirements for each described in a separate
2345
+
section.
2346
+
2347
+
#. `Bottom-Half Flavor (Historical) <#Bottom-Half%20Flavor>`__
2348
+
#. `Sched Flavor (Historical) <#Sched%20Flavor>`__
2349
+
#. `Sleepable RCU <#Sleepable%20RCU>`__
2350
+
#. `Tasks RCU <#Tasks%20RCU>`__
2351
+
2352
+
Bottom-Half Flavor (Historical)
2353
+
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2354
+
2355
+
The RCU-bh flavor of RCU has since been expressed in terms of the other
2356
+
RCU flavors as part of a consolidation of the three flavors into a
2357
+
single flavor. The read-side API remains, and continues to disable
2358
+
softirq and to be accounted for by lockdep. Much of the material in this
2359
+
section is therefore strictly historical in nature.
2360
+
2361
+
The softirq-disable (AKA “bottom-half”, hence the “_bh” abbreviations)
2362
+
flavor of RCU, or *RCU-bh*, was developed by Dipankar Sarma to provide a
2363
+
flavor of RCU that could withstand the network-based denial-of-service
2364
+
attacks researched by Robert Olsson. These attacks placed so much
2365
+
networking load on the system that some of the CPUs never exited softirq
2366
+
execution, which in turn prevented those CPUs from ever executing a
2367
+
context switch, which, in the RCU implementation of that time, prevented
2368
+
grace periods from ever ending. The result was an out-of-memory
2369
+
condition and a system hang.
2370
+
2371
+
The solution was the creation of RCU-bh, which does
2372
+
``local_bh_disable()`` across its read-side critical sections, and which
2373
+
uses the transition from one type of softirq processing to another as a
2374
+
quiescent state in addition to context switch, idle, user mode, and
2375
+
offline. This means that RCU-bh grace periods can complete even when
2376
+
some of the CPUs execute in softirq indefinitely, thus allowing
2377
+
algorithms based on RCU-bh to withstand network-based denial-of-service
2378
+
attacks.
2379
+
2380
+
Because ``rcu_read_lock_bh()`` and ``rcu_read_unlock_bh()`` disable and
2381
+
re-enable softirq handlers, any attempt to start a softirq handlers
2382
+
during the RCU-bh read-side critical section will be deferred. In this
2383
+
case, ``rcu_read_unlock_bh()`` will invoke softirq processing, which can
2384
+
take considerable time. One can of course argue that this softirq
2385
+
overhead should be associated with the code following the RCU-bh
2386
+
read-side critical section rather than ``rcu_read_unlock_bh()``, but the
2387
+
fact is that most profiling tools cannot be expected to make this sort
2388
+
of fine distinction. For example, suppose that a three-millisecond-long
2389
+
RCU-bh read-side critical section executes during a time of heavy
2390
+
networking load. There will very likely be an attempt to invoke at least
2391
+
one softirq handler during that three milliseconds, but any such
2392
+
invocation will be delayed until the time of the
2393
+
``rcu_read_unlock_bh()``. This can of course make it appear at first
2394
+
glance as if ``rcu_read_unlock_bh()`` was executing very slowly.
2395
+
2396
+
The `RCU-bh
2397
+
API <https://lwn.net/Articles/609973/#RCU%20Per-Flavor%20API%20Table>`__
2398
+
includes ``rcu_read_lock_bh()``, ``rcu_read_unlock_bh()``,
2399
+
``rcu_dereference_bh()``, ``rcu_dereference_bh_check()``,
2400
+
``synchronize_rcu_bh()``, ``synchronize_rcu_bh_expedited()``,
2401
+
``call_rcu_bh()``, ``rcu_barrier_bh()``, and
2402
+
``rcu_read_lock_bh_held()``. However, the update-side APIs are now
2403
+
simple wrappers for other RCU flavors, namely RCU-sched in
2404
+
CONFIG_PREEMPT=n kernels and RCU-preempt otherwise.
2405
+
2406
+
Sched Flavor (Historical)
2407
+
~~~~~~~~~~~~~~~~~~~~~~~~~
2408
+
2409
+
The RCU-sched flavor of RCU has since been expressed in terms of the
2410
+
other RCU flavors as part of a consolidation of the three flavors into a
2411
+
single flavor. The read-side API remains, and continues to disable
2412
+
preemption and to be accounted for by lockdep. Much of the material in
2413
+
this section is therefore strictly historical in nature.
2414
+
2415
+
Before preemptible RCU, waiting for an RCU grace period had the side
2416
+
effect of also waiting for all pre-existing interrupt and NMI handlers.
2417
+
However, there are legitimate preemptible-RCU implementations that do
2418
+
not have this property, given that any point in the code outside of an
2419
+
RCU read-side critical section can be a quiescent state. Therefore,
2420
+
*RCU-sched* was created, which follows “classic” RCU in that an
2421
+
RCU-sched grace period waits for for pre-existing interrupt and NMI
2422
+
handlers. In kernels built with ``CONFIG_PREEMPT=n``, the RCU and
2423
+
RCU-sched APIs have identical implementations, while kernels built with
2424
+
``CONFIG_PREEMPT=y`` provide a separate implementation for each.
2425
+
2426
+
Note well that in ``CONFIG_PREEMPT=y`` kernels,
2427
+
``rcu_read_lock_sched()`` and ``rcu_read_unlock_sched()`` disable and
2428
+
re-enable preemption, respectively. This means that if there was a
2429
+
preemption attempt during the RCU-sched read-side critical section,
2430
+
``rcu_read_unlock_sched()`` will enter the scheduler, with all the
2431
+
latency and overhead entailed. Just as with ``rcu_read_unlock_bh()``,
2432
+
this can make it look as if ``rcu_read_unlock_sched()`` was executing
2433
+
very slowly. However, the highest-priority task won't be preempted, so
2434
+
that task will enjoy low-overhead ``rcu_read_unlock_sched()``
2435
+
invocations.
2436
+
2437
+
The `RCU-sched
2438
+
API <https://lwn.net/Articles/609973/#RCU%20Per-Flavor%20API%20Table>`__
2439
+
includes ``rcu_read_lock_sched()``, ``rcu_read_unlock_sched()``,
2440
+
``rcu_read_lock_sched_notrace()``, ``rcu_read_unlock_sched_notrace()``,
2441
+
``rcu_dereference_sched()``, ``rcu_dereference_sched_check()``,
2442
+
``synchronize_sched()``, ``synchronize_rcu_sched_expedited()``,
2443
+
``call_rcu_sched()``, ``rcu_barrier_sched()``, and
2444
+
``rcu_read_lock_sched_held()``. However, anything that disables
2445
+
preemption also marks an RCU-sched read-side critical section, including
2446
+
``preempt_disable()`` and ``preempt_enable()``, ``local_irq_save()`` and
2447
+
``local_irq_restore()``, and so on.
2448
+
2449
+
Sleepable RCU
2450
+
~~~~~~~~~~~~~
2451
+
2452
+
For well over a decade, someone saying “I need to block within an RCU
2453
+
read-side critical section” was a reliable indication that this someone
2454
+
did not understand RCU. After all, if you are always blocking in an RCU
2455
+
read-side critical section, you can probably afford to use a
2456
+
higher-overhead synchronization mechanism. However, that changed with
2457
+
the advent of the Linux kernel's notifiers, whose RCU read-side critical
2458
+
sections almost never sleep, but sometimes need to. This resulted in the
2459
+
introduction of `sleepable RCU <https://lwn.net/Articles/202847/>`__, or
2460
+
*SRCU*.
2461
+
2462
+
SRCU allows different domains to be defined, with each such domain
2463
+
defined by an instance of an ``srcu_struct`` structure. A pointer to
2464
+
this structure must be passed in to each SRCU function, for example,
2465
+
``synchronize_srcu(&ss)``, where ``ss`` is the ``srcu_struct``
2466
+
structure. The key benefit of these domains is that a slow SRCU reader
2467
+
in one domain does not delay an SRCU grace period in some other domain.
2468
+
That said, one consequence of these domains is that read-side code must
2469
+
pass a “cookie” from ``srcu_read_lock()`` to ``srcu_read_unlock()``, for
2470
+
example, as follows:
2471
+
2472
+
::
2473
+
2474
+
1 int idx;
2475
+
2
2476
+
3 idx = srcu_read_lock(&ss);
2477
+
4 do_something();
2478
+
5 srcu_read_unlock(&ss, idx);
2479
+
2480
+
As noted above, it is legal to block within SRCU read-side critical
2481
+
sections, however, with great power comes great responsibility. If you
2482
+
block forever in one of a given domain's SRCU read-side critical
2483
+
sections, then that domain's grace periods will also be blocked forever.
2484
+
Of course, one good way to block forever is to deadlock, which can
2485
+
happen if any operation in a given domain's SRCU read-side critical
2486
+
section can wait, either directly or indirectly, for that domain's grace
2487
+
period to elapse. For example, this results in a self-deadlock:
2488
+
2489
+
::
2490
+
2491
+
1 int idx;
2492
+
2
2493
+
3 idx = srcu_read_lock(&ss);
2494
+
4 do_something();
2495
+
5 synchronize_srcu(&ss);
2496
+
6 srcu_read_unlock(&ss, idx);
2497
+
2498
+
However, if line 5 acquired a mutex that was held across a
2499
+
``synchronize_srcu()`` for domain ``ss``, deadlock would still be
2500
+
possible. Furthermore, if line 5 acquired a mutex that was held across a
2501
+
``synchronize_srcu()`` for some other domain ``ss1``, and if an
2502
+
``ss1``-domain SRCU read-side critical section acquired another mutex
2503
+
that was held across as ``ss``-domain ``synchronize_srcu()``, deadlock
2504
+
would again be possible. Such a deadlock cycle could extend across an
2505
+
arbitrarily large number of different SRCU domains. Again, with great
2506
+
power comes great responsibility.
2507
+
2508
+
Unlike the other RCU flavors, SRCU read-side critical sections can run
2509
+
on idle and even offline CPUs. This ability requires that
2510
+
``srcu_read_lock()`` and ``srcu_read_unlock()`` contain memory barriers,
2511
+
which means that SRCU readers will run a bit slower than would RCU
2512
+
readers. It also motivates the ``smp_mb__after_srcu_read_unlock()`` API,
2513
+
which, in combination with ``srcu_read_unlock()``, guarantees a full
2514
+
memory barrier.
2515
+
2516
+
Also unlike other RCU flavors, ``synchronize_srcu()`` may **not** be
2517
+
invoked from CPU-hotplug notifiers, due to the fact that SRCU grace
2518
+
periods make use of timers and the possibility of timers being
2519
+
temporarily “stranded” on the outgoing CPU. This stranding of timers
2520
+
means that timers posted to the outgoing CPU will not fire until late in
2521
+
the CPU-hotplug process. The problem is that if a notifier is waiting on
2522
+
an SRCU grace period, that grace period is waiting on a timer, and that
2523
+
timer is stranded on the outgoing CPU, then the notifier will never be
2524
+
awakened, in other words, deadlock has occurred. This same situation of
2525
+
course also prohibits ``srcu_barrier()`` from being invoked from
2526
+
CPU-hotplug notifiers.
2527
+
2528
+
SRCU also differs from other RCU flavors in that SRCU's expedited and
2529
+
non-expedited grace periods are implemented by the same mechanism. This
2530
+
means that in the current SRCU implementation, expediting a future grace
2531
+
period has the side effect of expediting all prior grace periods that
2532
+
have not yet completed. (But please note that this is a property of the
2533
+
current implementation, not necessarily of future implementations.) In
2534
+
addition, if SRCU has been idle for longer than the interval specified
2535
+
by the ``srcutree.exp_holdoff`` kernel boot parameter (25 microseconds
2536
+
by default), and if a ``synchronize_srcu()`` invocation ends this idle
2537
+
period, that invocation will be automatically expedited.
2538
+
2539
+
As of v4.12, SRCU's callbacks are maintained per-CPU, eliminating a
2540
+
locking bottleneck present in prior kernel versions. Although this will
2541
+
allow users to put much heavier stress on ``call_srcu()``, it is
2542
+
important to note that SRCU does not yet take any special steps to deal
2543
+
with callback flooding. So if you are posting (say) 10,000 SRCU
2544
+
callbacks per second per CPU, you are probably totally OK, but if you
2545
+
intend to post (say) 1,000,000 SRCU callbacks per second per CPU, please
2546
+
run some tests first. SRCU just might need a few adjustment to deal with
2547
+
that sort of load. Of course, your mileage may vary based on the speed
2548
+
of your CPUs and the size of your memory.
2549
+
2550
+
The `SRCU
2551
+
API <https://lwn.net/Articles/609973/#RCU%20Per-Flavor%20API%20Table>`__
2552
+
includes ``srcu_read_lock()``, ``srcu_read_unlock()``,
2553
+
``srcu_dereference()``, ``srcu_dereference_check()``,
2554
+
``synchronize_srcu()``, ``synchronize_srcu_expedited()``,
2555
+
``call_srcu()``, ``srcu_barrier()``, and ``srcu_read_lock_held()``. It
2556
+
also includes ``DEFINE_SRCU()``, ``DEFINE_STATIC_SRCU()``, and
2557
+
``init_srcu_struct()`` APIs for defining and initializing
2558
+
``srcu_struct`` structures.
2559
+
2560
+
Tasks RCU
2561
+
~~~~~~~~~
2562
+
2563
+
Some forms of tracing use “trampolines” to handle the binary rewriting
2564
+
required to install different types of probes. It would be good to be
2565
+
able to free old trampolines, which sounds like a job for some form of
2566
+
RCU. However, because it is necessary to be able to install a trace
2567
+
anywhere in the code, it is not possible to use read-side markers such
2568
+
as ``rcu_read_lock()`` and ``rcu_read_unlock()``. In addition, it does
2569
+
not work to have these markers in the trampoline itself, because there
2570
+
would need to be instructions following ``rcu_read_unlock()``. Although
2571
+
``synchronize_rcu()`` would guarantee that execution reached the
2572
+
``rcu_read_unlock()``, it would not be able to guarantee that execution
2573
+
had completely left the trampoline.
2574
+
2575
+
The solution, in the form of `Tasks
2576
+
RCU <https://lwn.net/Articles/607117/>`__, is to have implicit read-side
2577
+
critical sections that are delimited by voluntary context switches, that
2578
+
is, calls to ``schedule()``, ``cond_resched()``, and
2579
+
``synchronize_rcu_tasks()``. In addition, transitions to and from
2580
+
userspace execution also delimit tasks-RCU read-side critical sections.
2581
+
2582
+
The tasks-RCU API is quite compact, consisting only of
2583
+
``call_rcu_tasks()``, ``synchronize_rcu_tasks()``, and
2584
+
``rcu_barrier_tasks()``. In ``CONFIG_PREEMPT=n`` kernels, trampolines
2585
+
cannot be preempted, so these APIs map to ``call_rcu()``,
2586
+
``synchronize_rcu()``, and ``rcu_barrier()``, respectively. In
2587
+
``CONFIG_PREEMPT=y`` kernels, trampolines can be preempted, and these
2588
+
three APIs are therefore implemented by separate functions that check
2589
+
for voluntary context switches.
2590
+
2591
+
Possible Future Changes
2592
+
-----------------------
2593
+
2594
+
One of the tricks that RCU uses to attain update-side scalability is to
2595
+
increase grace-period latency with increasing numbers of CPUs. If this
2596
+
becomes a serious problem, it will be necessary to rework the
2597
+
grace-period state machine so as to avoid the need for the additional
2598
+
latency.
2599
+
2600
+
RCU disables CPU hotplug in a few places, perhaps most notably in the
2601
+
``rcu_barrier()`` operations. If there is a strong reason to use
2602
+
``rcu_barrier()`` in CPU-hotplug notifiers, it will be necessary to
2603
+
avoid disabling CPU hotplug. This would introduce some complexity, so
2604
+
there had better be a *very* good reason.
2605
+
2606
+
The tradeoff between grace-period latency on the one hand and
2607
+
interruptions of other CPUs on the other hand may need to be
2608
+
re-examined. The desire is of course for zero grace-period latency as
2609
+
well as zero interprocessor interrupts undertaken during an expedited
2610
+
grace period operation. While this ideal is unlikely to be achievable,
2611
+
it is quite possible that further improvements can be made.
2612
+
2613
+
The multiprocessor implementations of RCU use a combining tree that
2614
+
groups CPUs so as to reduce lock contention and increase cache locality.
2615
+
However, this combining tree does not spread its memory across NUMA
2616
+
nodes nor does it align the CPU groups with hardware features such as
2617
+
sockets or cores. Such spreading and alignment is currently believed to
2618
+
be unnecessary because the hotpath read-side primitives do not access
2619
+
the combining tree, nor does ``call_rcu()`` in the common case. If you
2620
+
believe that your architecture needs such spreading and alignment, then
2621
+
your architecture should also benefit from the
2622
+
``rcutree.rcu_fanout_leaf`` boot parameter, which can be set to the
2623
+
number of CPUs in a socket, NUMA node, or whatever. If the number of
2624
+
CPUs is too large, use a fraction of the number of CPUs. If the number
2625
+
of CPUs is a large prime number, well, that certainly is an
2626
+
“interesting” architectural choice! More flexible arrangements might be
2627
+
considered, but only if ``rcutree.rcu_fanout_leaf`` has proven
2628
+
inadequate, and only if the inadequacy has been demonstrated by a
2629
+
carefully run and realistic system-level workload.
2630
+
2631
+
Please note that arrangements that require RCU to remap CPU numbers will
2632
+
require extremely good demonstration of need and full exploration of
2633
+
alternatives.
2634
+
2635
+
RCU's various kthreads are reasonably recent additions. It is quite
2636
+
likely that adjustments will be required to more gracefully handle
2637
+
extreme loads. It might also be necessary to be able to relate CPU
2638
+
utilization by RCU's kthreads and softirq handlers to the code that
2639
+
instigated this CPU utilization. For example, RCU callback overhead
2640
+
might be charged back to the originating ``call_rcu()`` instance, though
2641
+
probably not in production kernels.
2642
+
2643
+
Additional work may be required to provide reasonable forward-progress
2644
+
guarantees under heavy load for grace periods and for callback
2645
+
invocation.
2646
+
2647
+
Summary
2648
+
-------
2649
+
2650
+
This document has presented more than two decade's worth of RCU
2651
+
requirements. Given that the requirements keep changing, this will not
2652
+
be the last word on this subject, but at least it serves to get an
2653
+
important subset of the requirements set forth.
2654
+
2655
+
Acknowledgments
2656
+
---------------
2657
+
2658
+
I am grateful to Steven Rostedt, Lai Jiangshan, Ingo Molnar, Oleg
2659
+
Nesterov, Borislav Petkov, Peter Zijlstra, Boqun Feng, and Andy
2660
+
Lutomirski for their help in rendering this article human readable, and
2661
+
to Michelle Rankin for her support of this effort. Other contributions
2662
+
are acknowledged in the Linux kernel's git archive.
+5
Documentation/RCU/index.rst
+5
Documentation/RCU/index.rst
+2
-2
Documentation/RCU/whatisRCU.txt
+2
-2
Documentation/RCU/whatisRCU.txt
···
302
302
must prohibit. The rcu_dereference_protected() variant takes
303
303
a lockdep expression to indicate which locks must be acquired
304
304
by the caller. If the indicated protection is not provided,
305
-
a lockdep splat is emitted. See RCU/Design/Requirements/Requirements.html
305
+
a lockdep splat is emitted. See Documentation/RCU/Design/Requirements/Requirements.rst
306
306
and the API's code comments for more details and example usage.
307
307
308
308
The following diagram shows how each API communicates among the
···
630
630
promotes synchronize_rcu() to a full memory barrier in compliance with
631
631
the "Memory-Barrier Guarantees" listed in:
632
632
633
-
Documentation/RCU/Design/Requirements/Requirements.html.
633
+
Documentation/RCU/Design/Requirements/Requirements.rst
634
634
635
635
It is possible to nest rcu_read_lock(), since reader-writer locks may
636
636
be recursively acquired. Note also that rcu_read_lock() is immune