tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
lib: objpool added: ring-array based lockless MPMC
objpool is a scalable implementation of high performance queue for
object allocation and reclamation, such as kretprobe instances.
With leveraging percpu ring-array to mitigate hot spots of memory
contention, it delivers near-linear scalability for high parallel
scenarios. The objpool is best suited for the following cases:
1) Memory allocation or reclamation are prohibited or too expensive
2) Consumers are of different priorities, such as irqs and threads
Limitations:
1) Maximum objects (capacity) is fixed after objpool creation
2) All pre-allocated objects are managed in percpu ring array,
which consumes more memory than linked lists
Link: https://lore.kernel.org/all/20231017135654.82270-2-wuqiang.matt@bytedance.com/
Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com>
Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
authored by
wuqiang.matt
and committed by
Masami Hiramatsu (Google)
b4edb8d2
f843249c
+462
-1
+181
include/linux/objpool.h
+181
include/linux/objpool.h
···
1
+
/* SPDX-License-Identifier: GPL-2.0 */
2
+
3
+
#ifndef _LINUX_OBJPOOL_H
4
+
#define _LINUX_OBJPOOL_H
5
+
6
+
#include <linux/types.h>
7
+
#include <linux/refcount.h>
8
+
9
+
/*
10
+
* objpool: ring-array based lockless MPMC queue
11
+
*
12
+
* Copyright: wuqiang.matt@bytedance.com,mhiramat@kernel.org
13
+
*
14
+
* objpool is a scalable implementation of high performance queue for
15
+
* object allocation and reclamation, such as kretprobe instances.
16
+
*
17
+
* With leveraging percpu ring-array to mitigate hot spots of memory
18
+
* contention, it delivers near-linear scalability for high parallel
19
+
* scenarios. The objpool is best suited for the following cases:
20
+
* 1) Memory allocation or reclamation are prohibited or too expensive
21
+
* 2) Consumers are of different priorities, such as irqs and threads
22
+
*
23
+
* Limitations:
24
+
* 1) Maximum objects (capacity) is fixed after objpool creation
25
+
* 2) All pre-allocated objects are managed in percpu ring array,
26
+
* which consumes more memory than linked lists
27
+
*/
28
+
29
+
/**
30
+
* struct objpool_slot - percpu ring array of objpool
31
+
* @head: head sequence of the local ring array (to retrieve at)
32
+
* @tail: tail sequence of the local ring array (to append at)
33
+
* @last: the last sequence number marked as ready for retrieve
34
+
* @mask: bits mask for modulo capacity to compute array indexes
35
+
* @entries: object entries on this slot
36
+
*
37
+
* Represents a cpu-local array-based ring buffer, its size is specialized
38
+
* during initialization of object pool. The percpu objpool node is to be
39
+
* allocated from local memory for NUMA system, and to be kept compact in
40
+
* continuous memory: CPU assigned number of objects are stored just after
41
+
* the body of objpool_node.
42
+
*
43
+
* Real size of the ring array is far too smaller than the value range of
44
+
* head and tail, typed as uint32_t: [0, 2^32), so only lower bits (mask)
45
+
* of head and tail are used as the actual position in the ring array. In
46
+
* general the ring array is acting like a small sliding window, which is
47
+
* always moving forward in the loop of [0, 2^32).
48
+
*/
49
+
struct objpool_slot {
50
+
uint32_t head;
51
+
uint32_t tail;
52
+
uint32_t last;
53
+
uint32_t mask;
54
+
void *entries[];
55
+
} __packed;
56
+
57
+
struct objpool_head;
58
+
59
+
/*
60
+
* caller-specified callback for object initial setup, it's only called
61
+
* once for each object (just after the memory allocation of the object)
62
+
*/
63
+
typedef int (*objpool_init_obj_cb)(void *obj, void *context);
64
+
65
+
/* caller-specified cleanup callback for objpool destruction */
66
+
typedef int (*objpool_fini_cb)(struct objpool_head *head, void *context);
67
+
68
+
/**
69
+
* struct objpool_head - object pooling metadata
70
+
* @obj_size: object size, aligned to sizeof(void *)
71
+
* @nr_objs: total objs (to be pre-allocated with objpool)
72
+
* @nr_cpus: local copy of nr_cpu_ids
73
+
* @capacity: max objs can be managed by one objpool_slot
74
+
* @gfp: gfp flags for kmalloc & vmalloc
75
+
* @ref: refcount of objpool
76
+
* @flags: flags for objpool management
77
+
* @cpu_slots: pointer to the array of objpool_slot
78
+
* @release: resource cleanup callback
79
+
* @context: caller-provided context
80
+
*/
81
+
struct objpool_head {
82
+
int obj_size;
83
+
int nr_objs;
84
+
int nr_cpus;
85
+
int capacity;
86
+
gfp_t gfp;
87
+
refcount_t ref;
88
+
unsigned long flags;
89
+
struct objpool_slot **cpu_slots;
90
+
objpool_fini_cb release;
91
+
void *context;
92
+
};
93
+
94
+
#define OBJPOOL_NR_OBJECT_MAX (1UL << 24) /* maximum numbers of total objects */
95
+
#define OBJPOOL_OBJECT_SIZE_MAX (1UL << 16) /* maximum size of an object */
96
+
97
+
/**
98
+
* objpool_init() - initialize objpool and pre-allocated objects
99
+
* @pool: the object pool to be initialized, declared by caller
100
+
* @nr_objs: total objects to be pre-allocated by this object pool
101
+
* @object_size: size of an object (should be > 0)
102
+
* @gfp: flags for memory allocation (via kmalloc or vmalloc)
103
+
* @context: user context for object initialization callback
104
+
* @objinit: object initialization callback for extra setup
105
+
* @release: cleanup callback for extra cleanup task
106
+
*
107
+
* return value: 0 for success, otherwise error code
108
+
*
109
+
* All pre-allocated objects are to be zeroed after memory allocation.
110
+
* Caller could do extra initialization in objinit callback. objinit()
111
+
* will be called just after slot allocation and called only once for
112
+
* each object. After that the objpool won't touch any content of the
113
+
* objects. It's caller's duty to perform reinitialization after each
114
+
* pop (object allocation) or do clearance before each push (object
115
+
* reclamation).
116
+
*/
117
+
int objpool_init(struct objpool_head *pool, int nr_objs, int object_size,
118
+
gfp_t gfp, void *context, objpool_init_obj_cb objinit,
119
+
objpool_fini_cb release);
120
+
121
+
/**
122
+
* objpool_pop() - allocate an object from objpool
123
+
* @pool: object pool
124
+
*
125
+
* return value: object ptr or NULL if failed
126
+
*/
127
+
void *objpool_pop(struct objpool_head *pool);
128
+
129
+
/**
130
+
* objpool_push() - reclaim the object and return back to objpool
131
+
* @obj: object ptr to be pushed to objpool
132
+
* @pool: object pool
133
+
*
134
+
* return: 0 or error code (it fails only when user tries to push
135
+
* the same object multiple times or wrong "objects" into objpool)
136
+
*/
137
+
int objpool_push(void *obj, struct objpool_head *pool);
138
+
139
+
/**
140
+
* objpool_drop() - discard the object and deref objpool
141
+
* @obj: object ptr to be discarded
142
+
* @pool: object pool
143
+
*
144
+
* return: 0 if objpool was released; -EAGAIN if there are still
145
+
* outstanding objects
146
+
*
147
+
* objpool_drop is normally for the release of outstanding objects
148
+
* after objpool cleanup (objpool_fini). Thinking of this example:
149
+
* kretprobe is unregistered and objpool_fini() is called to release
150
+
* all remained objects, but there are still objects being used by
151
+
* unfinished kretprobes (like blockable function: sys_accept). So
152
+
* only when the last outstanding object is dropped could the whole
153
+
* objpool be released along with the call of objpool_drop()
154
+
*/
155
+
int objpool_drop(void *obj, struct objpool_head *pool);
156
+
157
+
/**
158
+
* objpool_free() - release objpool forcely (all objects to be freed)
159
+
* @pool: object pool to be released
160
+
*/
161
+
void objpool_free(struct objpool_head *pool);
162
+
163
+
/**
164
+
* objpool_fini() - deref object pool (also releasing unused objects)
165
+
* @pool: object pool to be dereferenced
166
+
*
167
+
* objpool_fini() will try to release all remained free objects and
168
+
* then drop an extra reference of the objpool. If all objects are
169
+
* already returned to objpool (so called synchronous use cases),
170
+
* the objpool itself will be freed together. But if there are still
171
+
* outstanding objects (so called asynchronous use cases, such like
172
+
* blockable kretprobe), the objpool won't be released until all
173
+
* the outstanding objects are dropped, but the caller must assure
174
+
* there are no concurrent objpool_push() on the fly. Normally RCU
175
+
* is being required to make sure all ongoing objpool_push() must
176
+
* be finished before calling objpool_fini(), so does test_objpool,
177
+
* kretprobe or rethook
178
+
*/
179
+
void objpool_fini(struct objpool_head *pool);
180
+
181
+
#endif /* _LINUX_OBJPOOL_H */
+1
-1
lib/Makefile
+1
-1
lib/Makefile
+280
lib/objpool.c
+280
lib/objpool.c
···
1
+
// SPDX-License-Identifier: GPL-2.0
2
+
3
+
#include <linux/objpool.h>
4
+
#include <linux/slab.h>
5
+
#include <linux/vmalloc.h>
6
+
#include <linux/atomic.h>
7
+
#include <linux/irqflags.h>
8
+
#include <linux/cpumask.h>
9
+
#include <linux/log2.h>
10
+
11
+
/*
12
+
* objpool: ring-array based lockless MPMC/FIFO queues
13
+
*
14
+
* Copyright: wuqiang.matt@bytedance.com,mhiramat@kernel.org
15
+
*/
16
+
17
+
/* initialize percpu objpool_slot */
18
+
static int
19
+
objpool_init_percpu_slot(struct objpool_head *pool,
20
+
struct objpool_slot *slot,
21
+
int nodes, void *context,
22
+
objpool_init_obj_cb objinit)
23
+
{
24
+
void *obj = (void *)&slot->entries[pool->capacity];
25
+
int i;
26
+
27
+
/* initialize elements of percpu objpool_slot */
28
+
slot->mask = pool->capacity - 1;
29
+
30
+
for (i = 0; i < nodes; i++) {
31
+
if (objinit) {
32
+
int rc = objinit(obj, context);
33
+
if (rc)
34
+
return rc;
35
+
}
36
+
slot->entries[slot->tail & slot->mask] = obj;
37
+
obj = obj + pool->obj_size;
38
+
slot->tail++;
39
+
slot->last = slot->tail;
40
+
pool->nr_objs++;
41
+
}
42
+
43
+
return 0;
44
+
}
45
+
46
+
/* allocate and initialize percpu slots */
47
+
static int
48
+
objpool_init_percpu_slots(struct objpool_head *pool, int nr_objs,
49
+
void *context, objpool_init_obj_cb objinit)
50
+
{
51
+
int i, cpu_count = 0;
52
+
53
+
for (i = 0; i < pool->nr_cpus; i++) {
54
+
55
+
struct objpool_slot *slot;
56
+
int nodes, size, rc;
57
+
58
+
/* skip the cpu node which could never be present */
59
+
if (!cpu_possible(i))
60
+
continue;
61
+
62
+
/* compute how many objects to be allocated with this slot */
63
+
nodes = nr_objs / num_possible_cpus();
64
+
if (cpu_count < (nr_objs % num_possible_cpus()))
65
+
nodes++;
66
+
cpu_count++;
67
+
68
+
size = struct_size(slot, entries, pool->capacity) +
69
+
pool->obj_size * nodes;
70
+
71
+
/*
72
+
* here we allocate percpu-slot & objs together in a single
73
+
* allocation to make it more compact, taking advantage of
74
+
* warm caches and TLB hits. in default vmalloc is used to
75
+
* reduce the pressure of kernel slab system. as we know,
76
+
* mimimal size of vmalloc is one page since vmalloc would
77
+
* always align the requested size to page size
78
+
*/
79
+
if (pool->gfp & GFP_ATOMIC)
80
+
slot = kmalloc_node(size, pool->gfp, cpu_to_node(i));
81
+
else
82
+
slot = __vmalloc_node(size, sizeof(void *), pool->gfp,
83
+
cpu_to_node(i), __builtin_return_address(0));
84
+
if (!slot)
85
+
return -ENOMEM;
86
+
memset(slot, 0, size);
87
+
pool->cpu_slots[i] = slot;
88
+
89
+
/* initialize the objpool_slot of cpu node i */
90
+
rc = objpool_init_percpu_slot(pool, slot, nodes, context, objinit);
91
+
if (rc)
92
+
return rc;
93
+
}
94
+
95
+
return 0;
96
+
}
97
+
98
+
/* cleanup all percpu slots of the object pool */
99
+
static void objpool_fini_percpu_slots(struct objpool_head *pool)
100
+
{
101
+
int i;
102
+
103
+
if (!pool->cpu_slots)
104
+
return;
105
+
106
+
for (i = 0; i < pool->nr_cpus; i++)
107
+
kvfree(pool->cpu_slots[i]);
108
+
kfree(pool->cpu_slots);
109
+
}
110
+
111
+
/* initialize object pool and pre-allocate objects */
112
+
int objpool_init(struct objpool_head *pool, int nr_objs, int object_size,
113
+
gfp_t gfp, void *context, objpool_init_obj_cb objinit,
114
+
objpool_fini_cb release)
115
+
{
116
+
int rc, capacity, slot_size;
117
+
118
+
/* check input parameters */
119
+
if (nr_objs <= 0 || nr_objs > OBJPOOL_NR_OBJECT_MAX ||
120
+
object_size <= 0 || object_size > OBJPOOL_OBJECT_SIZE_MAX)
121
+
return -EINVAL;
122
+
123
+
/* align up to unsigned long size */
124
+
object_size = ALIGN(object_size, sizeof(long));
125
+
126
+
/* calculate capacity of percpu objpool_slot */
127
+
capacity = roundup_pow_of_two(nr_objs);
128
+
if (!capacity)
129
+
return -EINVAL;
130
+
131
+
/* initialize objpool pool */
132
+
memset(pool, 0, sizeof(struct objpool_head));
133
+
pool->nr_cpus = nr_cpu_ids;
134
+
pool->obj_size = object_size;
135
+
pool->capacity = capacity;
136
+
pool->gfp = gfp & ~__GFP_ZERO;
137
+
pool->context = context;
138
+
pool->release = release;
139
+
slot_size = pool->nr_cpus * sizeof(struct objpool_slot);
140
+
pool->cpu_slots = kzalloc(slot_size, pool->gfp);
141
+
if (!pool->cpu_slots)
142
+
return -ENOMEM;
143
+
144
+
/* initialize per-cpu slots */
145
+
rc = objpool_init_percpu_slots(pool, nr_objs, context, objinit);
146
+
if (rc)
147
+
objpool_fini_percpu_slots(pool);
148
+
else
149
+
refcount_set(&pool->ref, pool->nr_objs + 1);
150
+
151
+
return rc;
152
+
}
153
+
EXPORT_SYMBOL_GPL(objpool_init);
154
+
155
+
/* adding object to slot, abort if the slot was already full */
156
+
static inline int
157
+
objpool_try_add_slot(void *obj, struct objpool_head *pool, int cpu)
158
+
{
159
+
struct objpool_slot *slot = pool->cpu_slots[cpu];
160
+
uint32_t head, tail;
161
+
162
+
/* loading tail and head as a local snapshot, tail first */
163
+
tail = READ_ONCE(slot->tail);
164
+
165
+
do {
166
+
head = READ_ONCE(slot->head);
167
+
/* fault caught: something must be wrong */
168
+
WARN_ON_ONCE(tail - head > pool->nr_objs);
169
+
} while (!try_cmpxchg_acquire(&slot->tail, &tail, tail + 1));
170
+
171
+
/* now the tail position is reserved for the given obj */
172
+
WRITE_ONCE(slot->entries[tail & slot->mask], obj);
173
+
/* update sequence to make this obj available for pop() */
174
+
smp_store_release(&slot->last, tail + 1);
175
+
176
+
return 0;
177
+
}
178
+
179
+
/* reclaim an object to object pool */
180
+
int objpool_push(void *obj, struct objpool_head *pool)
181
+
{
182
+
unsigned long flags;
183
+
int rc;
184
+
185
+
/* disable local irq to avoid preemption & interruption */
186
+
raw_local_irq_save(flags);
187
+
rc = objpool_try_add_slot(obj, pool, raw_smp_processor_id());
188
+
raw_local_irq_restore(flags);
189
+
190
+
return rc;
191
+
}
192
+
EXPORT_SYMBOL_GPL(objpool_push);
193
+
194
+
/* try to retrieve object from slot */
195
+
static inline void *objpool_try_get_slot(struct objpool_head *pool, int cpu)
196
+
{
197
+
struct objpool_slot *slot = pool->cpu_slots[cpu];
198
+
/* load head snapshot, other cpus may change it */
199
+
uint32_t head = smp_load_acquire(&slot->head);
200
+
201
+
while (head != READ_ONCE(slot->last)) {
202
+
void *obj;
203
+
204
+
/* obj must be retrieved before moving forward head */
205
+
obj = READ_ONCE(slot->entries[head & slot->mask]);
206
+
207
+
/* move head forward to mark it's consumption */
208
+
if (try_cmpxchg_release(&slot->head, &head, head + 1))
209
+
return obj;
210
+
}
211
+
212
+
return NULL;
213
+
}
214
+
215
+
/* allocate an object from object pool */
216
+
void *objpool_pop(struct objpool_head *pool)
217
+
{
218
+
void *obj = NULL;
219
+
unsigned long flags;
220
+
int i, cpu;
221
+
222
+
/* disable local irq to avoid preemption & interruption */
223
+
raw_local_irq_save(flags);
224
+
225
+
cpu = raw_smp_processor_id();
226
+
for (i = 0; i < num_possible_cpus(); i++) {
227
+
obj = objpool_try_get_slot(pool, cpu);
228
+
if (obj)
229
+
break;
230
+
cpu = cpumask_next_wrap(cpu, cpu_possible_mask, -1, 1);
231
+
}
232
+
raw_local_irq_restore(flags);
233
+
234
+
return obj;
235
+
}
236
+
EXPORT_SYMBOL_GPL(objpool_pop);
237
+
238
+
/* release whole objpool forcely */
239
+
void objpool_free(struct objpool_head *pool)
240
+
{
241
+
if (!pool->cpu_slots)
242
+
return;
243
+
244
+
/* release percpu slots */
245
+
objpool_fini_percpu_slots(pool);
246
+
247
+
/* call user's cleanup callback if provided */
248
+
if (pool->release)
249
+
pool->release(pool, pool->context);
250
+
}
251
+
EXPORT_SYMBOL_GPL(objpool_free);
252
+
253
+
/* drop the allocated object, rather reclaim it to objpool */
254
+
int objpool_drop(void *obj, struct objpool_head *pool)
255
+
{
256
+
if (!obj || !pool)
257
+
return -EINVAL;
258
+
259
+
if (refcount_dec_and_test(&pool->ref)) {
260
+
objpool_free(pool);
261
+
return 0;
262
+
}
263
+
264
+
return -EAGAIN;
265
+
}
266
+
EXPORT_SYMBOL_GPL(objpool_drop);
267
+
268
+
/* drop unused objects and defref objpool for releasing */
269
+
void objpool_fini(struct objpool_head *pool)
270
+
{
271
+
int count = 1; /* extra ref for objpool itself */
272
+
273
+
/* drop all remained objects from objpool */
274
+
while (objpool_pop(pool))
275
+
count++;
276
+
277
+
if (refcount_sub_and_test(count, &pool->ref))
278
+
objpool_free(pool);
279
+
}
280
+
EXPORT_SYMBOL_GPL(objpool_fini);