tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
1RCU on Uniprocessor Systems
2
3
4A common misconception is that, on UP systems, the call_rcu() primitive
5may immediately invoke its function. The basis of this misconception
6is that since there is only one CPU, it should not be necessary to
7wait for anything else to get done, since there are no other CPUs for
8anything else to be happening on. Although this approach will -sort- -of-
9work a surprising amount of the time, it is a very bad idea in general.
10This document presents three examples that demonstrate exactly how bad
11an idea this is.
12
13
14Example 1: softirq Suicide
15
16Suppose that an RCU-based algorithm scans a linked list containing
17elements A, B, and C in process context, and can delete elements from
18this same list in softirq context. Suppose that the process-context scan
19is referencing element B when it is interrupted by softirq processing,
20which deletes element B, and then invokes call_rcu() to free element B
21after a grace period.
22
23Now, if call_rcu() were to directly invoke its arguments, then upon return
24from softirq, the list scan would find itself referencing a newly freed
25element B. This situation can greatly decrease the life expectancy of
26your kernel.
27
28This same problem can occur if call_rcu() is invoked from a hardware
29interrupt handler.
30
31
32Example 2: Function-Call Fatality
33
34Of course, one could avert the suicide described in the preceding example
35by having call_rcu() directly invoke its arguments only if it was called
36from process context. However, this can fail in a similar manner.
37
38Suppose that an RCU-based algorithm again scans a linked list containing
39elements A, B, and C in process contexts, but that it invokes a function
40on each element as it is scanned. Suppose further that this function
41deletes element B from the list, then passes it to call_rcu() for deferred
42freeing. This may be a bit unconventional, but it is perfectly legal
43RCU usage, since call_rcu() must wait for a grace period to elapse.
44Therefore, in this case, allowing call_rcu() to immediately invoke
45its arguments would cause it to fail to make the fundamental guarantee
46underlying RCU, namely that call_rcu() defers invoking its arguments until
47all RCU read-side critical sections currently executing have completed.
48
49Quick Quiz #1: why is it -not- legal to invoke synchronize_rcu() in
50 this case?
51
52
53Example 3: Death by Deadlock
54
55Suppose that call_rcu() is invoked while holding a lock, and that the
56callback function must acquire this same lock. In this case, if
57call_rcu() were to directly invoke the callback, the result would
58be self-deadlock.
59
60In some cases, it would possible to restructure to code so that
61the call_rcu() is delayed until after the lock is released. However,
62there are cases where this can be quite ugly:
63
641. If a number of items need to be passed to call_rcu() within
65 the same critical section, then the code would need to create
66 a list of them, then traverse the list once the lock was
67 released.
68
692. In some cases, the lock will be held across some kernel API,
70 so that delaying the call_rcu() until the lock is released
71 requires that the data item be passed up via a common API.
72 It is far better to guarantee that callbacks are invoked
73 with no locks held than to have to modify such APIs to allow
74 arbitrary data items to be passed back up through them.
75
76If call_rcu() directly invokes the callback, painful locking restrictions
77or API changes would be required.
78
79Quick Quiz #2: What locking restriction must RCU callbacks respect?
80
81
82Summary
83
84Permitting call_rcu() to immediately invoke its arguments breaks RCU,
85even on a UP system. So do not do it! Even on a UP system, the RCU
86infrastructure -must- respect grace periods, and -must- invoke callbacks
87from a known environment in which no locks are held.
88
89Note that it -is- safe for synchronize_rcu() to return immediately on
90UP systems, including !PREEMPT SMP builds running on UP systems.
91
92Quick Quiz #3: Why can't synchronize_rcu() return immediately on
93 UP systems running preemptable RCU?
94
95
96Answer to Quick Quiz #1:
97 Why is it -not- legal to invoke synchronize_rcu() in this case?
98
99 Because the calling function is scanning an RCU-protected linked
100 list, and is therefore within an RCU read-side critical section.
101 Therefore, the called function has been invoked within an RCU
102 read-side critical section, and is not permitted to block.
103
104Answer to Quick Quiz #2:
105 What locking restriction must RCU callbacks respect?
106
107 Any lock that is acquired within an RCU callback must be
108 acquired elsewhere using an _irq variant of the spinlock
109 primitive. For example, if "mylock" is acquired by an
110 RCU callback, then a process-context acquisition of this
111 lock must use something like spin_lock_irqsave() to
112 acquire the lock.
113
114 If the process-context code were to simply use spin_lock(),
115 then, since RCU callbacks can be invoked from softirq context,
116 the callback might be called from a softirq that interrupted
117 the process-context critical section. This would result in
118 self-deadlock.
119
120 This restriction might seem gratuitous, since very few RCU
121 callbacks acquire locks directly. However, a great many RCU
122 callbacks do acquire locks -indirectly-, for example, via
123 the kfree() primitive.
124
125Answer to Quick Quiz #3:
126 Why can't synchronize_rcu() return immediately on UP systems
127 running preemptable RCU?
128
129 Because some other task might have been preempted in the middle
130 of an RCU read-side critical section. If synchronize_rcu()
131 simply immediately returned, it would prematurely signal the
132 end of the grace period, which would come as a nasty shock to
133 that other thread when it started running again.