Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/mm/mempool.c
4 *
5 * memory buffer pool support. Such pools are mostly used
6 * for guaranteed, deadlock-free memory allocations during
7 * extreme VM load.
8 *
9 * started by Ingo Molnar, Copyright (C) 2001
10 * debugging by David Rientjes, Copyright (C) 2015
11 */
12
13#include <linux/mm.h>
14#include <linux/slab.h>
15#include <linux/highmem.h>
16#include <linux/kasan.h>
17#include <linux/kmemleak.h>
18#include <linux/export.h>
19#include <linux/mempool.h>
20#include <linux/writeback.h>
21#include "slab.h"
22
23#ifdef CONFIG_SLUB_DEBUG_ON
24static void poison_error(mempool_t *pool, void *element, size_t size,
25 size_t byte)
26{
27 const int nr = pool->curr_nr;
28 const int start = max_t(int, byte - (BITS_PER_LONG / 8), 0);
29 const int end = min_t(int, byte + (BITS_PER_LONG / 8), size);
30 int i;
31
32 pr_err("BUG: mempool element poison mismatch\n");
33 pr_err("Mempool %p size %zu\n", pool, size);
34 pr_err(" nr=%d @ %p: %s0x", nr, element, start > 0 ? "... " : "");
35 for (i = start; i < end; i++)
36 pr_cont("%x ", *(u8 *)(element + i));
37 pr_cont("%s\n", end < size ? "..." : "");
38 dump_stack();
39}
40
41static void __check_element(mempool_t *pool, void *element, size_t size)
42{
43 u8 *obj = element;
44 size_t i;
45
46 for (i = 0; i < size; i++) {
47 u8 exp = (i < size - 1) ? POISON_FREE : POISON_END;
48
49 if (obj[i] != exp) {
50 poison_error(pool, element, size, i);
51 return;
52 }
53 }
54 memset(obj, POISON_INUSE, size);
55}
56
57static void check_element(mempool_t *pool, void *element)
58{
59 /* Skip checking: KASAN might save its metadata in the element. */
60 if (kasan_enabled())
61 return;
62
63 /* Mempools backed by slab allocator */
64 if (pool->free == mempool_kfree) {
65 __check_element(pool, element, (size_t)pool->pool_data);
66 } else if (pool->free == mempool_free_slab) {
67 __check_element(pool, element, kmem_cache_size(pool->pool_data));
68 } else if (pool->free == mempool_free_pages) {
69 /* Mempools backed by page allocator */
70 int order = (int)(long)pool->pool_data;
71 void *addr = kmap_local_page((struct page *)element);
72
73 __check_element(pool, addr, 1UL << (PAGE_SHIFT + order));
74 kunmap_local(addr);
75 }
76}
77
78static void __poison_element(void *element, size_t size)
79{
80 u8 *obj = element;
81
82 memset(obj, POISON_FREE, size - 1);
83 obj[size - 1] = POISON_END;
84}
85
86static void poison_element(mempool_t *pool, void *element)
87{
88 /* Skip poisoning: KASAN might save its metadata in the element. */
89 if (kasan_enabled())
90 return;
91
92 /* Mempools backed by slab allocator */
93 if (pool->alloc == mempool_kmalloc) {
94 __poison_element(element, (size_t)pool->pool_data);
95 } else if (pool->alloc == mempool_alloc_slab) {
96 __poison_element(element, kmem_cache_size(pool->pool_data));
97 } else if (pool->alloc == mempool_alloc_pages) {
98 /* Mempools backed by page allocator */
99 int order = (int)(long)pool->pool_data;
100 void *addr = kmap_local_page((struct page *)element);
101
102 __poison_element(addr, 1UL << (PAGE_SHIFT + order));
103 kunmap_local(addr);
104 }
105}
106#else /* CONFIG_SLUB_DEBUG_ON */
107static inline void check_element(mempool_t *pool, void *element)
108{
109}
110static inline void poison_element(mempool_t *pool, void *element)
111{
112}
113#endif /* CONFIG_SLUB_DEBUG_ON */
114
115static __always_inline bool kasan_poison_element(mempool_t *pool, void *element)
116{
117 if (pool->alloc == mempool_alloc_slab || pool->alloc == mempool_kmalloc)
118 return kasan_mempool_poison_object(element);
119 else if (pool->alloc == mempool_alloc_pages)
120 return kasan_mempool_poison_pages(element,
121 (unsigned long)pool->pool_data);
122 return true;
123}
124
125static void kasan_unpoison_element(mempool_t *pool, void *element)
126{
127 if (pool->alloc == mempool_kmalloc)
128 kasan_mempool_unpoison_object(element, (size_t)pool->pool_data);
129 else if (pool->alloc == mempool_alloc_slab)
130 kasan_mempool_unpoison_object(element,
131 kmem_cache_size(pool->pool_data));
132 else if (pool->alloc == mempool_alloc_pages)
133 kasan_mempool_unpoison_pages(element,
134 (unsigned long)pool->pool_data);
135}
136
137static __always_inline void add_element(mempool_t *pool, void *element)
138{
139 BUG_ON(pool->curr_nr >= pool->min_nr);
140 poison_element(pool, element);
141 if (kasan_poison_element(pool, element))
142 pool->elements[pool->curr_nr++] = element;
143}
144
145static void *remove_element(mempool_t *pool)
146{
147 void *element = pool->elements[--pool->curr_nr];
148
149 BUG_ON(pool->curr_nr < 0);
150 kasan_unpoison_element(pool, element);
151 check_element(pool, element);
152 return element;
153}
154
155/**
156 * mempool_exit - exit a mempool initialized with mempool_init()
157 * @pool: pointer to the memory pool which was initialized with
158 * mempool_init().
159 *
160 * Free all reserved elements in @pool and @pool itself. This function
161 * only sleeps if the free_fn() function sleeps.
162 *
163 * May be called on a zeroed but uninitialized mempool (i.e. allocated with
164 * kzalloc()).
165 */
166void mempool_exit(mempool_t *pool)
167{
168 while (pool->curr_nr) {
169 void *element = remove_element(pool);
170 pool->free(element, pool->pool_data);
171 }
172 kfree(pool->elements);
173 pool->elements = NULL;
174}
175EXPORT_SYMBOL(mempool_exit);
176
177/**
178 * mempool_destroy - deallocate a memory pool
179 * @pool: pointer to the memory pool which was allocated via
180 * mempool_create().
181 *
182 * Free all reserved elements in @pool and @pool itself. This function
183 * only sleeps if the free_fn() function sleeps.
184 */
185void mempool_destroy(mempool_t *pool)
186{
187 if (unlikely(!pool))
188 return;
189
190 mempool_exit(pool);
191 kfree(pool);
192}
193EXPORT_SYMBOL(mempool_destroy);
194
195int mempool_init_node(mempool_t *pool, int min_nr, mempool_alloc_t *alloc_fn,
196 mempool_free_t *free_fn, void *pool_data,
197 gfp_t gfp_mask, int node_id)
198{
199 spin_lock_init(&pool->lock);
200 pool->min_nr = min_nr;
201 pool->pool_data = pool_data;
202 pool->alloc = alloc_fn;
203 pool->free = free_fn;
204 init_waitqueue_head(&pool->wait);
205
206 pool->elements = kmalloc_array_node(min_nr, sizeof(void *),
207 gfp_mask, node_id);
208 if (!pool->elements)
209 return -ENOMEM;
210
211 /*
212 * First pre-allocate the guaranteed number of buffers.
213 */
214 while (pool->curr_nr < pool->min_nr) {
215 void *element;
216
217 element = pool->alloc(gfp_mask, pool->pool_data);
218 if (unlikely(!element)) {
219 mempool_exit(pool);
220 return -ENOMEM;
221 }
222 add_element(pool, element);
223 }
224
225 return 0;
226}
227EXPORT_SYMBOL(mempool_init_node);
228
229/**
230 * mempool_init - initialize a memory pool
231 * @pool: pointer to the memory pool that should be initialized
232 * @min_nr: the minimum number of elements guaranteed to be
233 * allocated for this pool.
234 * @alloc_fn: user-defined element-allocation function.
235 * @free_fn: user-defined element-freeing function.
236 * @pool_data: optional private data available to the user-defined functions.
237 *
238 * Like mempool_create(), but initializes the pool in (i.e. embedded in another
239 * structure).
240 *
241 * Return: %0 on success, negative error code otherwise.
242 */
243int mempool_init(mempool_t *pool, int min_nr, mempool_alloc_t *alloc_fn,
244 mempool_free_t *free_fn, void *pool_data)
245{
246 return mempool_init_node(pool, min_nr, alloc_fn, free_fn,
247 pool_data, GFP_KERNEL, NUMA_NO_NODE);
248
249}
250EXPORT_SYMBOL(mempool_init);
251
252/**
253 * mempool_create - create a memory pool
254 * @min_nr: the minimum number of elements guaranteed to be
255 * allocated for this pool.
256 * @alloc_fn: user-defined element-allocation function.
257 * @free_fn: user-defined element-freeing function.
258 * @pool_data: optional private data available to the user-defined functions.
259 *
260 * this function creates and allocates a guaranteed size, preallocated
261 * memory pool. The pool can be used from the mempool_alloc() and mempool_free()
262 * functions. This function might sleep. Both the alloc_fn() and the free_fn()
263 * functions might sleep - as long as the mempool_alloc() function is not called
264 * from IRQ contexts.
265 *
266 * Return: pointer to the created memory pool object or %NULL on error.
267 */
268mempool_t *mempool_create(int min_nr, mempool_alloc_t *alloc_fn,
269 mempool_free_t *free_fn, void *pool_data)
270{
271 return mempool_create_node(min_nr, alloc_fn, free_fn, pool_data,
272 GFP_KERNEL, NUMA_NO_NODE);
273}
274EXPORT_SYMBOL(mempool_create);
275
276mempool_t *mempool_create_node(int min_nr, mempool_alloc_t *alloc_fn,
277 mempool_free_t *free_fn, void *pool_data,
278 gfp_t gfp_mask, int node_id)
279{
280 mempool_t *pool;
281
282 pool = kzalloc_node(sizeof(*pool), gfp_mask, node_id);
283 if (!pool)
284 return NULL;
285
286 if (mempool_init_node(pool, min_nr, alloc_fn, free_fn, pool_data,
287 gfp_mask, node_id)) {
288 kfree(pool);
289 return NULL;
290 }
291
292 return pool;
293}
294EXPORT_SYMBOL(mempool_create_node);
295
296/**
297 * mempool_resize - resize an existing memory pool
298 * @pool: pointer to the memory pool which was allocated via
299 * mempool_create().
300 * @new_min_nr: the new minimum number of elements guaranteed to be
301 * allocated for this pool.
302 *
303 * This function shrinks/grows the pool. In the case of growing,
304 * it cannot be guaranteed that the pool will be grown to the new
305 * size immediately, but new mempool_free() calls will refill it.
306 * This function may sleep.
307 *
308 * Note, the caller must guarantee that no mempool_destroy is called
309 * while this function is running. mempool_alloc() & mempool_free()
310 * might be called (eg. from IRQ contexts) while this function executes.
311 *
312 * Return: %0 on success, negative error code otherwise.
313 */
314int mempool_resize(mempool_t *pool, int new_min_nr)
315{
316 void *element;
317 void **new_elements;
318 unsigned long flags;
319
320 BUG_ON(new_min_nr <= 0);
321 might_sleep();
322
323 spin_lock_irqsave(&pool->lock, flags);
324 if (new_min_nr <= pool->min_nr) {
325 while (new_min_nr < pool->curr_nr) {
326 element = remove_element(pool);
327 spin_unlock_irqrestore(&pool->lock, flags);
328 pool->free(element, pool->pool_data);
329 spin_lock_irqsave(&pool->lock, flags);
330 }
331 pool->min_nr = new_min_nr;
332 goto out_unlock;
333 }
334 spin_unlock_irqrestore(&pool->lock, flags);
335
336 /* Grow the pool */
337 new_elements = kmalloc_array(new_min_nr, sizeof(*new_elements),
338 GFP_KERNEL);
339 if (!new_elements)
340 return -ENOMEM;
341
342 spin_lock_irqsave(&pool->lock, flags);
343 if (unlikely(new_min_nr <= pool->min_nr)) {
344 /* Raced, other resize will do our work */
345 spin_unlock_irqrestore(&pool->lock, flags);
346 kfree(new_elements);
347 goto out;
348 }
349 memcpy(new_elements, pool->elements,
350 pool->curr_nr * sizeof(*new_elements));
351 kfree(pool->elements);
352 pool->elements = new_elements;
353 pool->min_nr = new_min_nr;
354
355 while (pool->curr_nr < pool->min_nr) {
356 spin_unlock_irqrestore(&pool->lock, flags);
357 element = pool->alloc(GFP_KERNEL, pool->pool_data);
358 if (!element)
359 goto out;
360 spin_lock_irqsave(&pool->lock, flags);
361 if (pool->curr_nr < pool->min_nr) {
362 add_element(pool, element);
363 } else {
364 spin_unlock_irqrestore(&pool->lock, flags);
365 pool->free(element, pool->pool_data); /* Raced */
366 goto out;
367 }
368 }
369out_unlock:
370 spin_unlock_irqrestore(&pool->lock, flags);
371out:
372 return 0;
373}
374EXPORT_SYMBOL(mempool_resize);
375
376/**
377 * mempool_alloc - allocate an element from a specific memory pool
378 * @pool: pointer to the memory pool which was allocated via
379 * mempool_create().
380 * @gfp_mask: the usual allocation bitmask.
381 *
382 * this function only sleeps if the alloc_fn() function sleeps or
383 * returns NULL. Note that due to preallocation, this function
384 * *never* fails when called from process contexts. (it might
385 * fail if called from an IRQ context.)
386 * Note: using __GFP_ZERO is not supported.
387 *
388 * Return: pointer to the allocated element or %NULL on error.
389 */
390void *mempool_alloc(mempool_t *pool, gfp_t gfp_mask)
391{
392 void *element;
393 unsigned long flags;
394 wait_queue_entry_t wait;
395 gfp_t gfp_temp;
396
397 VM_WARN_ON_ONCE(gfp_mask & __GFP_ZERO);
398 might_alloc(gfp_mask);
399
400 gfp_mask |= __GFP_NOMEMALLOC; /* don't allocate emergency reserves */
401 gfp_mask |= __GFP_NORETRY; /* don't loop in __alloc_pages */
402 gfp_mask |= __GFP_NOWARN; /* failures are OK */
403
404 gfp_temp = gfp_mask & ~(__GFP_DIRECT_RECLAIM|__GFP_IO);
405
406repeat_alloc:
407
408 element = pool->alloc(gfp_temp, pool->pool_data);
409 if (likely(element != NULL))
410 return element;
411
412 spin_lock_irqsave(&pool->lock, flags);
413 if (likely(pool->curr_nr)) {
414 element = remove_element(pool);
415 spin_unlock_irqrestore(&pool->lock, flags);
416 /* paired with rmb in mempool_free(), read comment there */
417 smp_wmb();
418 /*
419 * Update the allocation stack trace as this is more useful
420 * for debugging.
421 */
422 kmemleak_update_trace(element);
423 return element;
424 }
425
426 /*
427 * We use gfp mask w/o direct reclaim or IO for the first round. If
428 * alloc failed with that and @pool was empty, retry immediately.
429 */
430 if (gfp_temp != gfp_mask) {
431 spin_unlock_irqrestore(&pool->lock, flags);
432 gfp_temp = gfp_mask;
433 goto repeat_alloc;
434 }
435
436 /* We must not sleep if !__GFP_DIRECT_RECLAIM */
437 if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
438 spin_unlock_irqrestore(&pool->lock, flags);
439 return NULL;
440 }
441
442 /* Let's wait for someone else to return an element to @pool */
443 init_wait(&wait);
444 prepare_to_wait(&pool->wait, &wait, TASK_UNINTERRUPTIBLE);
445
446 spin_unlock_irqrestore(&pool->lock, flags);
447
448 /*
449 * FIXME: this should be io_schedule(). The timeout is there as a
450 * workaround for some DM problems in 2.6.18.
451 */
452 io_schedule_timeout(5*HZ);
453
454 finish_wait(&pool->wait, &wait);
455 goto repeat_alloc;
456}
457EXPORT_SYMBOL(mempool_alloc);
458
459/**
460 * mempool_alloc_preallocated - allocate an element from preallocated elements
461 * belonging to a specific memory pool
462 * @pool: pointer to the memory pool which was allocated via
463 * mempool_create().
464 *
465 * This function is similar to mempool_alloc, but it only attempts allocating
466 * an element from the preallocated elements. It does not sleep and immediately
467 * returns if no preallocated elements are available.
468 *
469 * Return: pointer to the allocated element or %NULL if no elements are
470 * available.
471 */
472void *mempool_alloc_preallocated(mempool_t *pool)
473{
474 void *element;
475 unsigned long flags;
476
477 spin_lock_irqsave(&pool->lock, flags);
478 if (likely(pool->curr_nr)) {
479 element = remove_element(pool);
480 spin_unlock_irqrestore(&pool->lock, flags);
481 /* paired with rmb in mempool_free(), read comment there */
482 smp_wmb();
483 /*
484 * Update the allocation stack trace as this is more useful
485 * for debugging.
486 */
487 kmemleak_update_trace(element);
488 return element;
489 }
490 spin_unlock_irqrestore(&pool->lock, flags);
491
492 return NULL;
493}
494EXPORT_SYMBOL(mempool_alloc_preallocated);
495
496/**
497 * mempool_free - return an element to the pool.
498 * @element: pool element pointer.
499 * @pool: pointer to the memory pool which was allocated via
500 * mempool_create().
501 *
502 * this function only sleeps if the free_fn() function sleeps.
503 */
504void mempool_free(void *element, mempool_t *pool)
505{
506 unsigned long flags;
507
508 if (unlikely(element == NULL))
509 return;
510
511 /*
512 * Paired with the wmb in mempool_alloc(). The preceding read is
513 * for @element and the following @pool->curr_nr. This ensures
514 * that the visible value of @pool->curr_nr is from after the
515 * allocation of @element. This is necessary for fringe cases
516 * where @element was passed to this task without going through
517 * barriers.
518 *
519 * For example, assume @p is %NULL at the beginning and one task
520 * performs "p = mempool_alloc(...);" while another task is doing
521 * "while (!p) cpu_relax(); mempool_free(p, ...);". This function
522 * may end up using curr_nr value which is from before allocation
523 * of @p without the following rmb.
524 */
525 smp_rmb();
526
527 /*
528 * For correctness, we need a test which is guaranteed to trigger
529 * if curr_nr + #allocated == min_nr. Testing curr_nr < min_nr
530 * without locking achieves that and refilling as soon as possible
531 * is desirable.
532 *
533 * Because curr_nr visible here is always a value after the
534 * allocation of @element, any task which decremented curr_nr below
535 * min_nr is guaranteed to see curr_nr < min_nr unless curr_nr gets
536 * incremented to min_nr afterwards. If curr_nr gets incremented
537 * to min_nr after the allocation of @element, the elements
538 * allocated after that are subject to the same guarantee.
539 *
540 * Waiters happen iff curr_nr is 0 and the above guarantee also
541 * ensures that there will be frees which return elements to the
542 * pool waking up the waiters.
543 */
544 if (unlikely(READ_ONCE(pool->curr_nr) < pool->min_nr)) {
545 spin_lock_irqsave(&pool->lock, flags);
546 if (likely(pool->curr_nr < pool->min_nr)) {
547 add_element(pool, element);
548 spin_unlock_irqrestore(&pool->lock, flags);
549 wake_up(&pool->wait);
550 return;
551 }
552 spin_unlock_irqrestore(&pool->lock, flags);
553 }
554 pool->free(element, pool->pool_data);
555}
556EXPORT_SYMBOL(mempool_free);
557
558/*
559 * A commonly used alloc and free fn.
560 */
561void *mempool_alloc_slab(gfp_t gfp_mask, void *pool_data)
562{
563 struct kmem_cache *mem = pool_data;
564 VM_BUG_ON(mem->ctor);
565 return kmem_cache_alloc(mem, gfp_mask);
566}
567EXPORT_SYMBOL(mempool_alloc_slab);
568
569void mempool_free_slab(void *element, void *pool_data)
570{
571 struct kmem_cache *mem = pool_data;
572 kmem_cache_free(mem, element);
573}
574EXPORT_SYMBOL(mempool_free_slab);
575
576/*
577 * A commonly used alloc and free fn that kmalloc/kfrees the amount of memory
578 * specified by pool_data
579 */
580void *mempool_kmalloc(gfp_t gfp_mask, void *pool_data)
581{
582 size_t size = (size_t)pool_data;
583 return kmalloc(size, gfp_mask);
584}
585EXPORT_SYMBOL(mempool_kmalloc);
586
587void mempool_kfree(void *element, void *pool_data)
588{
589 kfree(element);
590}
591EXPORT_SYMBOL(mempool_kfree);
592
593/*
594 * A simple mempool-backed page allocator that allocates pages
595 * of the order specified by pool_data.
596 */
597void *mempool_alloc_pages(gfp_t gfp_mask, void *pool_data)
598{
599 int order = (int)(long)pool_data;
600 return alloc_pages(gfp_mask, order);
601}
602EXPORT_SYMBOL(mempool_alloc_pages);
603
604void mempool_free_pages(void *element, void *pool_data)
605{
606 int order = (int)(long)pool_data;
607 __free_pages(element, order);
608}
609EXPORT_SYMBOL(mempool_free_pages);
1/*
2 * linux/mm/mempool.c
3 *
4 * memory buffer pool support. Such pools are mostly used
5 * for guaranteed, deadlock-free memory allocations during
6 * extreme VM load.
7 *
8 * started by Ingo Molnar, Copyright (C) 2001
9 */
10
11#include <linux/mm.h>
12#include <linux/slab.h>
13#include <linux/export.h>
14#include <linux/mempool.h>
15#include <linux/blkdev.h>
16#include <linux/writeback.h>
17
18static void add_element(mempool_t *pool, void *element)
19{
20 BUG_ON(pool->curr_nr >= pool->min_nr);
21 pool->elements[pool->curr_nr++] = element;
22}
23
24static void *remove_element(mempool_t *pool)
25{
26 BUG_ON(pool->curr_nr <= 0);
27 return pool->elements[--pool->curr_nr];
28}
29
30/**
31 * mempool_destroy - deallocate a memory pool
32 * @pool: pointer to the memory pool which was allocated via
33 * mempool_create().
34 *
35 * Free all reserved elements in @pool and @pool itself. This function
36 * only sleeps if the free_fn() function sleeps.
37 */
38void mempool_destroy(mempool_t *pool)
39{
40 while (pool->curr_nr) {
41 void *element = remove_element(pool);
42 pool->free(element, pool->pool_data);
43 }
44 kfree(pool->elements);
45 kfree(pool);
46}
47EXPORT_SYMBOL(mempool_destroy);
48
49/**
50 * mempool_create - create a memory pool
51 * @min_nr: the minimum number of elements guaranteed to be
52 * allocated for this pool.
53 * @alloc_fn: user-defined element-allocation function.
54 * @free_fn: user-defined element-freeing function.
55 * @pool_data: optional private data available to the user-defined functions.
56 *
57 * this function creates and allocates a guaranteed size, preallocated
58 * memory pool. The pool can be used from the mempool_alloc() and mempool_free()
59 * functions. This function might sleep. Both the alloc_fn() and the free_fn()
60 * functions might sleep - as long as the mempool_alloc() function is not called
61 * from IRQ contexts.
62 */
63mempool_t *mempool_create(int min_nr, mempool_alloc_t *alloc_fn,
64 mempool_free_t *free_fn, void *pool_data)
65{
66 return mempool_create_node(min_nr,alloc_fn,free_fn, pool_data,
67 GFP_KERNEL, NUMA_NO_NODE);
68}
69EXPORT_SYMBOL(mempool_create);
70
71mempool_t *mempool_create_node(int min_nr, mempool_alloc_t *alloc_fn,
72 mempool_free_t *free_fn, void *pool_data,
73 gfp_t gfp_mask, int node_id)
74{
75 mempool_t *pool;
76 pool = kzalloc_node(sizeof(*pool), gfp_mask, node_id);
77 if (!pool)
78 return NULL;
79 pool->elements = kmalloc_node(min_nr * sizeof(void *),
80 gfp_mask, node_id);
81 if (!pool->elements) {
82 kfree(pool);
83 return NULL;
84 }
85 spin_lock_init(&pool->lock);
86 pool->min_nr = min_nr;
87 pool->pool_data = pool_data;
88 init_waitqueue_head(&pool->wait);
89 pool->alloc = alloc_fn;
90 pool->free = free_fn;
91
92 /*
93 * First pre-allocate the guaranteed number of buffers.
94 */
95 while (pool->curr_nr < pool->min_nr) {
96 void *element;
97
98 element = pool->alloc(gfp_mask, pool->pool_data);
99 if (unlikely(!element)) {
100 mempool_destroy(pool);
101 return NULL;
102 }
103 add_element(pool, element);
104 }
105 return pool;
106}
107EXPORT_SYMBOL(mempool_create_node);
108
109/**
110 * mempool_resize - resize an existing memory pool
111 * @pool: pointer to the memory pool which was allocated via
112 * mempool_create().
113 * @new_min_nr: the new minimum number of elements guaranteed to be
114 * allocated for this pool.
115 * @gfp_mask: the usual allocation bitmask.
116 *
117 * This function shrinks/grows the pool. In the case of growing,
118 * it cannot be guaranteed that the pool will be grown to the new
119 * size immediately, but new mempool_free() calls will refill it.
120 *
121 * Note, the caller must guarantee that no mempool_destroy is called
122 * while this function is running. mempool_alloc() & mempool_free()
123 * might be called (eg. from IRQ contexts) while this function executes.
124 */
125int mempool_resize(mempool_t *pool, int new_min_nr, gfp_t gfp_mask)
126{
127 void *element;
128 void **new_elements;
129 unsigned long flags;
130
131 BUG_ON(new_min_nr <= 0);
132
133 spin_lock_irqsave(&pool->lock, flags);
134 if (new_min_nr <= pool->min_nr) {
135 while (new_min_nr < pool->curr_nr) {
136 element = remove_element(pool);
137 spin_unlock_irqrestore(&pool->lock, flags);
138 pool->free(element, pool->pool_data);
139 spin_lock_irqsave(&pool->lock, flags);
140 }
141 pool->min_nr = new_min_nr;
142 goto out_unlock;
143 }
144 spin_unlock_irqrestore(&pool->lock, flags);
145
146 /* Grow the pool */
147 new_elements = kmalloc(new_min_nr * sizeof(*new_elements), gfp_mask);
148 if (!new_elements)
149 return -ENOMEM;
150
151 spin_lock_irqsave(&pool->lock, flags);
152 if (unlikely(new_min_nr <= pool->min_nr)) {
153 /* Raced, other resize will do our work */
154 spin_unlock_irqrestore(&pool->lock, flags);
155 kfree(new_elements);
156 goto out;
157 }
158 memcpy(new_elements, pool->elements,
159 pool->curr_nr * sizeof(*new_elements));
160 kfree(pool->elements);
161 pool->elements = new_elements;
162 pool->min_nr = new_min_nr;
163
164 while (pool->curr_nr < pool->min_nr) {
165 spin_unlock_irqrestore(&pool->lock, flags);
166 element = pool->alloc(gfp_mask, pool->pool_data);
167 if (!element)
168 goto out;
169 spin_lock_irqsave(&pool->lock, flags);
170 if (pool->curr_nr < pool->min_nr) {
171 add_element(pool, element);
172 } else {
173 spin_unlock_irqrestore(&pool->lock, flags);
174 pool->free(element, pool->pool_data); /* Raced */
175 goto out;
176 }
177 }
178out_unlock:
179 spin_unlock_irqrestore(&pool->lock, flags);
180out:
181 return 0;
182}
183EXPORT_SYMBOL(mempool_resize);
184
185/**
186 * mempool_alloc - allocate an element from a specific memory pool
187 * @pool: pointer to the memory pool which was allocated via
188 * mempool_create().
189 * @gfp_mask: the usual allocation bitmask.
190 *
191 * this function only sleeps if the alloc_fn() function sleeps or
192 * returns NULL. Note that due to preallocation, this function
193 * *never* fails when called from process contexts. (it might
194 * fail if called from an IRQ context.)
195 */
196void * mempool_alloc(mempool_t *pool, gfp_t gfp_mask)
197{
198 void *element;
199 unsigned long flags;
200 wait_queue_t wait;
201 gfp_t gfp_temp;
202
203 might_sleep_if(gfp_mask & __GFP_WAIT);
204
205 gfp_mask |= __GFP_NOMEMALLOC; /* don't allocate emergency reserves */
206 gfp_mask |= __GFP_NORETRY; /* don't loop in __alloc_pages */
207 gfp_mask |= __GFP_NOWARN; /* failures are OK */
208
209 gfp_temp = gfp_mask & ~(__GFP_WAIT|__GFP_IO);
210
211repeat_alloc:
212
213 element = pool->alloc(gfp_temp, pool->pool_data);
214 if (likely(element != NULL))
215 return element;
216
217 spin_lock_irqsave(&pool->lock, flags);
218 if (likely(pool->curr_nr)) {
219 element = remove_element(pool);
220 spin_unlock_irqrestore(&pool->lock, flags);
221 /* paired with rmb in mempool_free(), read comment there */
222 smp_wmb();
223 return element;
224 }
225
226 /*
227 * We use gfp mask w/o __GFP_WAIT or IO for the first round. If
228 * alloc failed with that and @pool was empty, retry immediately.
229 */
230 if (gfp_temp != gfp_mask) {
231 spin_unlock_irqrestore(&pool->lock, flags);
232 gfp_temp = gfp_mask;
233 goto repeat_alloc;
234 }
235
236 /* We must not sleep if !__GFP_WAIT */
237 if (!(gfp_mask & __GFP_WAIT)) {
238 spin_unlock_irqrestore(&pool->lock, flags);
239 return NULL;
240 }
241
242 /* Let's wait for someone else to return an element to @pool */
243 init_wait(&wait);
244 prepare_to_wait(&pool->wait, &wait, TASK_UNINTERRUPTIBLE);
245
246 spin_unlock_irqrestore(&pool->lock, flags);
247
248 /*
249 * FIXME: this should be io_schedule(). The timeout is there as a
250 * workaround for some DM problems in 2.6.18.
251 */
252 io_schedule_timeout(5*HZ);
253
254 finish_wait(&pool->wait, &wait);
255 goto repeat_alloc;
256}
257EXPORT_SYMBOL(mempool_alloc);
258
259/**
260 * mempool_free - return an element to the pool.
261 * @element: pool element pointer.
262 * @pool: pointer to the memory pool which was allocated via
263 * mempool_create().
264 *
265 * this function only sleeps if the free_fn() function sleeps.
266 */
267void mempool_free(void *element, mempool_t *pool)
268{
269 unsigned long flags;
270
271 if (unlikely(element == NULL))
272 return;
273
274 /*
275 * Paired with the wmb in mempool_alloc(). The preceding read is
276 * for @element and the following @pool->curr_nr. This ensures
277 * that the visible value of @pool->curr_nr is from after the
278 * allocation of @element. This is necessary for fringe cases
279 * where @element was passed to this task without going through
280 * barriers.
281 *
282 * For example, assume @p is %NULL at the beginning and one task
283 * performs "p = mempool_alloc(...);" while another task is doing
284 * "while (!p) cpu_relax(); mempool_free(p, ...);". This function
285 * may end up using curr_nr value which is from before allocation
286 * of @p without the following rmb.
287 */
288 smp_rmb();
289
290 /*
291 * For correctness, we need a test which is guaranteed to trigger
292 * if curr_nr + #allocated == min_nr. Testing curr_nr < min_nr
293 * without locking achieves that and refilling as soon as possible
294 * is desirable.
295 *
296 * Because curr_nr visible here is always a value after the
297 * allocation of @element, any task which decremented curr_nr below
298 * min_nr is guaranteed to see curr_nr < min_nr unless curr_nr gets
299 * incremented to min_nr afterwards. If curr_nr gets incremented
300 * to min_nr after the allocation of @element, the elements
301 * allocated after that are subject to the same guarantee.
302 *
303 * Waiters happen iff curr_nr is 0 and the above guarantee also
304 * ensures that there will be frees which return elements to the
305 * pool waking up the waiters.
306 */
307 if (unlikely(pool->curr_nr < pool->min_nr)) {
308 spin_lock_irqsave(&pool->lock, flags);
309 if (likely(pool->curr_nr < pool->min_nr)) {
310 add_element(pool, element);
311 spin_unlock_irqrestore(&pool->lock, flags);
312 wake_up(&pool->wait);
313 return;
314 }
315 spin_unlock_irqrestore(&pool->lock, flags);
316 }
317 pool->free(element, pool->pool_data);
318}
319EXPORT_SYMBOL(mempool_free);
320
321/*
322 * A commonly used alloc and free fn.
323 */
324void *mempool_alloc_slab(gfp_t gfp_mask, void *pool_data)
325{
326 struct kmem_cache *mem = pool_data;
327 return kmem_cache_alloc(mem, gfp_mask);
328}
329EXPORT_SYMBOL(mempool_alloc_slab);
330
331void mempool_free_slab(void *element, void *pool_data)
332{
333 struct kmem_cache *mem = pool_data;
334 kmem_cache_free(mem, element);
335}
336EXPORT_SYMBOL(mempool_free_slab);
337
338/*
339 * A commonly used alloc and free fn that kmalloc/kfrees the amount of memory
340 * specified by pool_data
341 */
342void *mempool_kmalloc(gfp_t gfp_mask, void *pool_data)
343{
344 size_t size = (size_t)pool_data;
345 return kmalloc(size, gfp_mask);
346}
347EXPORT_SYMBOL(mempool_kmalloc);
348
349void mempool_kfree(void *element, void *pool_data)
350{
351 kfree(element);
352}
353EXPORT_SYMBOL(mempool_kfree);
354
355/*
356 * A simple mempool-backed page allocator that allocates pages
357 * of the order specified by pool_data.
358 */
359void *mempool_alloc_pages(gfp_t gfp_mask, void *pool_data)
360{
361 int order = (int)(long)pool_data;
362 return alloc_pages(gfp_mask, order);
363}
364EXPORT_SYMBOL(mempool_alloc_pages);
365
366void mempool_free_pages(void *element, void *pool_data)
367{
368 int order = (int)(long)pool_data;
369 __free_pages(element, order);
370}
371EXPORT_SYMBOL(mempool_free_pages);