blob: 59cc5083e960b0a571e356e5262e140be4e0844a [file] [log] [blame]
Willy Tarreau50e608d2007-05-13 18:26:08 +02001/*
2 * Memory management functions.
3 *
4 * Copyright 2000-2007 Willy Tarreau <w@1wt.eu>
5 *
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
10 *
11 */
Willy Tarreau7107c8b2018-11-26 11:44:35 +010012#include <errno.h>
Willy Tarreau50e608d2007-05-13 18:26:08 +020013
Willy Tarreaub2551052020-06-09 09:07:15 +020014#include <haproxy/activity-t.h>
Willy Tarreau4c7e4b72020-05-27 12:58:42 +020015#include <haproxy/api.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020016#include <haproxy/applet-t.h>
Willy Tarreau6be78492020-06-05 00:00:29 +020017#include <haproxy/cfgparse.h>
Willy Tarreauf1d32c42020-06-04 21:07:02 +020018#include <haproxy/channel.h>
Willy Tarreau83487a82020-06-04 20:19:54 +020019#include <haproxy/cli.h>
Willy Tarreau36979d92020-06-05 17:27:29 +020020#include <haproxy/errors.h>
Willy Tarreauf268ee82020-06-04 17:05:57 +020021#include <haproxy/global.h>
Willy Tarreau853b2972020-05-27 18:01:47 +020022#include <haproxy/list.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020023#include <haproxy/pool.h>
Willy Tarreau2eec9b52020-06-04 19:58:55 +020024#include <haproxy/stats-t.h>
Willy Tarreau5e539c92020-06-04 20:45:39 +020025#include <haproxy/stream_interface.h>
Willy Tarreaub2551052020-06-09 09:07:15 +020026#include <haproxy/thread.h>
Willy Tarreau48fbcae2020-06-03 18:09:46 +020027#include <haproxy/tools.h>
Willy Tarreau50e608d2007-05-13 18:26:08 +020028
Willy Tarreau50e608d2007-05-13 18:26:08 +020029
Willy Tarreau2d6f6282021-04-15 16:24:00 +020030#ifdef CONFIG_HAP_POOLS
Willy Tarreau7f0165e2018-11-26 17:09:46 +010031/* These ones are initialized per-thread on startup by init_pools() */
Willy Tarreaue18db9e2018-10-16 10:28:54 +020032THREAD_LOCAL size_t pool_cache_bytes = 0; /* total cache size */
33THREAD_LOCAL size_t pool_cache_count = 0; /* #cache objects */
Willy Tarreaued891fd2020-06-01 19:00:28 +020034#endif
Willy Tarreaue18db9e2018-10-16 10:28:54 +020035
Willy Tarreau50e608d2007-05-13 18:26:08 +020036static struct list pools = LIST_HEAD_INIT(pools);
Willy Tarreau067ac9f2015-10-08 14:12:13 +020037int mem_poison_byte = -1;
Willy Tarreau50e608d2007-05-13 18:26:08 +020038
Olivier Houcharddc21ff72019-01-29 15:20:16 +010039#ifdef DEBUG_FAIL_ALLOC
40static int mem_fail_rate = 0;
41static int mem_should_fail(const struct pool_head *);
42#endif
43
Willy Tarreau50e608d2007-05-13 18:26:08 +020044/* Try to find an existing shared pool with the same characteristics and
45 * returns it, otherwise creates this one. NULL is returned if no memory
Willy Tarreau581bf812016-01-25 02:19:13 +010046 * is available for a new creation. Two flags are supported :
47 * - MEM_F_SHARED to indicate that the pool may be shared with other users
48 * - MEM_F_EXACT to indicate that the size must not be rounded up
Willy Tarreau50e608d2007-05-13 18:26:08 +020049 */
50struct pool_head *create_pool(char *name, unsigned int size, unsigned int flags)
51{
52 struct pool_head *pool;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +020053 struct pool_head *entry;
54 struct list *start;
Willy Tarreau50e608d2007-05-13 18:26:08 +020055 unsigned int align;
Willy Tarreau9f3129e2021-04-17 00:31:38 +020056 int thr __maybe_unused;
Willy Tarreau50e608d2007-05-13 18:26:08 +020057
Willy Tarreauac421112015-10-28 15:09:29 +010058 /* We need to store a (void *) at the end of the chunks. Since we know
Willy Tarreau50e608d2007-05-13 18:26:08 +020059 * that the malloc() function will never return such a small size,
60 * let's round the size up to something slightly bigger, in order to
61 * ease merging of entries. Note that the rounding is a power of two.
Willy Tarreauac421112015-10-28 15:09:29 +010062 * This extra (void *) is not accounted for in the size computation
63 * so that the visible parts outside are not affected.
Willy Tarreau30f931e2018-10-23 14:40:23 +020064 *
65 * Note: for the LRU cache, we need to store 2 doubly-linked lists.
Willy Tarreau50e608d2007-05-13 18:26:08 +020066 */
67
Willy Tarreau581bf812016-01-25 02:19:13 +010068 if (!(flags & MEM_F_EXACT)) {
Willy Tarreau30f931e2018-10-23 14:40:23 +020069 align = 4 * sizeof(void *); // 2 lists = 4 pointers min
Willy Tarreau581bf812016-01-25 02:19:13 +010070 size = ((size + POOL_EXTRA + align - 1) & -align) - POOL_EXTRA;
71 }
Willy Tarreau50e608d2007-05-13 18:26:08 +020072
Christopher Fauletb349e482017-08-29 09:52:38 +020073 /* TODO: thread: we do not lock pool list for now because all pools are
74 * created during HAProxy startup (so before threads creation) */
Willy Tarreau7dcd46d2007-05-14 00:16:13 +020075 start = &pools;
Willy Tarreau50e608d2007-05-13 18:26:08 +020076 pool = NULL;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +020077
78 list_for_each_entry(entry, &pools, list) {
79 if (entry->size == size) {
80 /* either we can share this place and we take it, or
Ilya Shipitsin47d17182020-06-21 21:42:57 +050081 * we look for a shareable one or for the next position
Willy Tarreau7dcd46d2007-05-14 00:16:13 +020082 * before which we will insert a new one.
83 */
84 if (flags & entry->flags & MEM_F_SHARED) {
85 /* we can share this one */
Willy Tarreau50e608d2007-05-13 18:26:08 +020086 pool = entry;
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +020087 DPRINTF(stderr, "Sharing %s with %s\n", name, pool->name);
Willy Tarreau50e608d2007-05-13 18:26:08 +020088 break;
89 }
90 }
Willy Tarreau7dcd46d2007-05-14 00:16:13 +020091 else if (entry->size > size) {
92 /* insert before this one */
93 start = &entry->list;
94 break;
95 }
Willy Tarreau50e608d2007-05-13 18:26:08 +020096 }
97
98 if (!pool) {
Willy Tarreau0a93b642018-10-16 07:58:39 +020099 if (!pool)
100 pool = calloc(1, sizeof(*pool));
101
Willy Tarreau50e608d2007-05-13 18:26:08 +0200102 if (!pool)
103 return NULL;
104 if (name)
105 strlcpy2(pool->name, name, sizeof(pool->name));
106 pool->size = size;
107 pool->flags = flags;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200108 LIST_ADDQ(start, &pool->list);
Christopher Faulet2f6d3c02019-06-25 21:45:59 +0200109
Willy Tarreau2d6f6282021-04-15 16:24:00 +0200110#ifdef CONFIG_HAP_POOLS
Christopher Faulet2f6d3c02019-06-25 21:45:59 +0200111 /* update per-thread pool cache if necessary */
Willy Tarreau9f3129e2021-04-17 00:31:38 +0200112 for (thr = 0; thr < MAX_THREADS; thr++) {
113 LIST_INIT(&pool->cache[thr].list);
Christopher Faulet2f6d3c02019-06-25 21:45:59 +0200114 }
Willy Tarreaued891fd2020-06-01 19:00:28 +0200115#endif
Olivier Houchard8af97eb2020-02-01 17:45:32 +0100116 HA_SPIN_INIT(&pool->lock);
Olivier Houchard8af97eb2020-02-01 17:45:32 +0100117 }
118 pool->users++;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200119 return pool;
120}
Olivier Houchardcf975d42018-01-24 18:38:31 +0100121
Willy Tarreau2d6f6282021-04-15 16:24:00 +0200122#ifdef CONFIG_HAP_POOLS
Willy Tarreaued891fd2020-06-01 19:00:28 +0200123/* Evicts some of the oldest objects from the local cache, pushing them to the
124 * global pool.
125 */
126void pool_evict_from_cache()
127{
128 struct pool_cache_item *item;
129 struct pool_cache_head *ph;
Willy Tarreau9f3129e2021-04-17 00:31:38 +0200130 struct pool_head *pool;
Willy Tarreaued891fd2020-06-01 19:00:28 +0200131
132 do {
Willy Tarreau20dc3cd2020-06-28 00:54:27 +0200133 item = LIST_PREV(&ti->pool_lru_head, struct pool_cache_item *, by_lru);
Willy Tarreaued891fd2020-06-01 19:00:28 +0200134 /* note: by definition we remove oldest objects so they also are the
135 * oldest in their own pools, thus their next is the pool's head.
136 */
137 ph = LIST_NEXT(&item->by_pool, struct pool_cache_head *, list);
Willy Tarreau9f3129e2021-04-17 00:31:38 +0200138 pool = container_of(ph - tid, struct pool_head, cache);
Willy Tarreaued891fd2020-06-01 19:00:28 +0200139 LIST_DEL(&item->by_pool);
140 LIST_DEL(&item->by_lru);
141 ph->count--;
142 pool_cache_count--;
Willy Tarreaud5140e72021-04-17 14:05:10 +0200143 pool_cache_bytes -= pool->size;
Willy Tarreau9f3129e2021-04-17 00:31:38 +0200144 __pool_free(pool, item);
Willy Tarreaued891fd2020-06-01 19:00:28 +0200145 } while (pool_cache_bytes > CONFIG_HAP_POOL_CACHE_SIZE * 7 / 8);
146}
147#endif
148
Willy Tarreau0bae0752021-03-02 20:05:09 +0100149#if defined(CONFIG_HAP_NO_GLOBAL_POOLS)
150
151/* simply fall back on the default OS' allocator */
152
153void *__pool_refill_alloc(struct pool_head *pool, unsigned int avail)
154{
155 int allocated = pool->allocated;
156 int limit = pool->limit;
157 void *ptr = NULL;
158
Willy Tarreau53a7fe42021-04-15 16:43:18 +0200159#ifdef DEBUG_FAIL_ALLOC
160 if (mem_should_fail(pool))
161 return NULL;
162#endif
Willy Tarreau0bae0752021-03-02 20:05:09 +0100163 if (limit && allocated >= limit) {
Willy Tarreau0bae0752021-03-02 20:05:09 +0100164 activity[tid].pool_fail++;
165 return NULL;
166 }
167
168 ptr = pool_alloc_area(pool->size + POOL_EXTRA);
169 if (!ptr) {
Willy Tarreau4781b152021-04-06 13:53:36 +0200170 _HA_ATOMIC_INC(&pool->failed);
Willy Tarreau0bae0752021-03-02 20:05:09 +0100171 activity[tid].pool_fail++;
172 return NULL;
173 }
174
Willy Tarreau4781b152021-04-06 13:53:36 +0200175 _HA_ATOMIC_INC(&pool->allocated);
176 _HA_ATOMIC_INC(&pool->used);
Willy Tarreau0bae0752021-03-02 20:05:09 +0100177
178#ifdef DEBUG_MEMORY_POOLS
179 /* keep track of where the element was allocated from */
180 *POOL_LINK(pool, ptr) = (void *)pool;
181#endif
182 return ptr;
183}
184
185/* legacy stuff */
186void *pool_refill_alloc(struct pool_head *pool, unsigned int avail)
187{
188 void *ptr;
189
190 ptr = __pool_refill_alloc(pool, avail);
191 return ptr;
192}
193
194/* legacy stuff */
195void pool_flush(struct pool_head *pool)
196{
197}
198
199/* This function might ask the malloc library to trim its buffers. */
200void pool_gc(struct pool_head *pool_ctx)
201{
202#if defined(HA_HAVE_MALLOC_TRIM)
203 malloc_trim(0);
204#endif
205}
206
207#elif defined(CONFIG_HAP_LOCKLESS_POOLS)
208
Olivier Houchardcf975d42018-01-24 18:38:31 +0100209/* Allocates new entries for pool <pool> until there are at least <avail> + 1
210 * available, then returns the last one for immediate use, so that at least
211 * <avail> are left available in the pool upon return. NULL is returned if the
212 * last entry could not be allocated. It's important to note that at least one
213 * allocation is always performed even if there are enough entries in the pool.
214 * A call to the garbage collector is performed at most once in case malloc()
215 * returns an error, before returning NULL.
216 */
217void *__pool_refill_alloc(struct pool_head *pool, unsigned int avail)
218{
Olivier Houchard8b2c8a72018-10-21 01:52:59 +0200219 void *ptr = NULL, **free_list;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100220 int failed = 0;
221 int size = pool->size;
222 int limit = pool->limit;
223 int allocated = pool->allocated, allocated_orig = allocated;
224
Willy Tarreau53a7fe42021-04-15 16:43:18 +0200225#ifdef DEBUG_FAIL_ALLOC
226 if (mem_should_fail(pool))
227 return NULL;
228#endif
Olivier Houchardcf975d42018-01-24 18:38:31 +0100229 /* stop point */
230 avail += pool->used;
231
232 while (1) {
233 if (limit && allocated >= limit) {
Olivier Houchard20872762019-03-08 18:53:35 +0100234 _HA_ATOMIC_ADD(&pool->allocated, allocated - allocated_orig);
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200235 activity[tid].pool_fail++;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100236 return NULL;
237 }
238
Willy Tarreau606135a2020-06-01 12:35:03 +0200239 swrate_add_scaled(&pool->needed_avg, POOL_AVG_SAMPLES, pool->allocated, POOL_AVG_SAMPLES/4);
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200240
Willy Tarreau24aa1ee2020-05-30 18:56:17 +0200241 ptr = pool_alloc_area(size + POOL_EXTRA);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100242 if (!ptr) {
Willy Tarreau4781b152021-04-06 13:53:36 +0200243 _HA_ATOMIC_INC(&pool->failed);
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200244 if (failed) {
245 activity[tid].pool_fail++;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100246 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200247 }
Olivier Houchardcf975d42018-01-24 18:38:31 +0100248 failed++;
249 pool_gc(pool);
250 continue;
251 }
252 if (++allocated > avail)
253 break;
254
255 free_list = pool->free_list;
256 do {
257 *POOL_LINK(pool, ptr) = free_list;
258 __ha_barrier_store();
Olivier Houchard20872762019-03-08 18:53:35 +0100259 } while (_HA_ATOMIC_CAS(&pool->free_list, &free_list, ptr) == 0);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100260 }
Olivier Houchard20872762019-03-08 18:53:35 +0100261 __ha_barrier_atomic_store();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100262
Olivier Houchard20872762019-03-08 18:53:35 +0100263 _HA_ATOMIC_ADD(&pool->allocated, allocated - allocated_orig);
Willy Tarreau4781b152021-04-06 13:53:36 +0200264 _HA_ATOMIC_INC(&pool->used);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100265
266#ifdef DEBUG_MEMORY_POOLS
267 /* keep track of where the element was allocated from */
268 *POOL_LINK(pool, ptr) = (void *)pool;
269#endif
270 return ptr;
271}
272void *pool_refill_alloc(struct pool_head *pool, unsigned int avail)
273{
274 void *ptr;
275
276 ptr = __pool_refill_alloc(pool, avail);
277 return ptr;
278}
279/*
280 * This function frees whatever can be freed in pool <pool>.
281 */
282void pool_flush(struct pool_head *pool)
283{
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100284 struct pool_free_list cmp, new;
Olivier Houchard8b2c8a72018-10-21 01:52:59 +0200285 void **next, *temp;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100286 int removed = 0;
287
288 if (!pool)
289 return;
Willy Tarreau21072b92020-05-29 17:23:05 +0200290 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100291 do {
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100292 cmp.free_list = pool->free_list;
293 cmp.seq = pool->seq;
294 new.free_list = NULL;
295 new.seq = cmp.seq + 1;
296 } while (!_HA_ATOMIC_DWCAS(&pool->free_list, &cmp, &new));
Olivier Houchard20872762019-03-08 18:53:35 +0100297 __ha_barrier_atomic_store();
Willy Tarreau21072b92020-05-29 17:23:05 +0200298 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100299 next = cmp.free_list;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100300 while (next) {
301 temp = next;
302 next = *POOL_LINK(pool, temp);
303 removed++;
Willy Tarreau24aa1ee2020-05-30 18:56:17 +0200304 pool_free_area(temp, pool->size + POOL_EXTRA);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100305 }
306 pool->free_list = next;
Olivier Houchard20872762019-03-08 18:53:35 +0100307 _HA_ATOMIC_SUB(&pool->allocated, removed);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100308 /* here, we should have pool->allocate == pool->used */
309}
310
311/*
312 * This function frees whatever can be freed in all pools, but respecting
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200313 * the minimum thresholds imposed by owners. It makes sure to be alone to
314 * run by using thread_isolate(). <pool_ctx> is unused.
Olivier Houchardcf975d42018-01-24 18:38:31 +0100315 */
316void pool_gc(struct pool_head *pool_ctx)
317{
Olivier Houchardcf975d42018-01-24 18:38:31 +0100318 struct pool_head *entry;
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200319 int isolated = thread_isolated();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100320
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200321 if (!isolated)
322 thread_isolate();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100323
324 list_for_each_entry(entry, &pools, list) {
325 while ((int)((volatile int)entry->allocated - (volatile int)entry->used) > (int)entry->minavail) {
326 struct pool_free_list cmp, new;
327
328 cmp.seq = entry->seq;
329 __ha_barrier_load();
330 cmp.free_list = entry->free_list;
331 __ha_barrier_load();
332 if (cmp.free_list == NULL)
333 break;
334 new.free_list = *POOL_LINK(entry, cmp.free_list);
335 new.seq = cmp.seq + 1;
Willy Tarreau6a38b322019-05-11 18:04:24 +0200336 if (HA_ATOMIC_DWCAS(&entry->free_list, &cmp, &new) == 0)
Olivier Houchardcf975d42018-01-24 18:38:31 +0100337 continue;
Willy Tarreau24aa1ee2020-05-30 18:56:17 +0200338 pool_free_area(cmp.free_list, entry->size + POOL_EXTRA);
Willy Tarreau4781b152021-04-06 13:53:36 +0200339 _HA_ATOMIC_DEC(&entry->allocated);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100340 }
341 }
342
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200343 if (!isolated)
344 thread_release();
Willy Tarreau88366c22020-11-03 15:53:34 +0100345
346#if defined(HA_HAVE_MALLOC_TRIM)
347 malloc_trim(0);
348#endif
Olivier Houchardcf975d42018-01-24 18:38:31 +0100349}
Willy Tarreaue18db9e2018-10-16 10:28:54 +0200350
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100351#else /* CONFIG_HAP_LOCKLESS_POOLS */
Willy Tarreau50e608d2007-05-13 18:26:08 +0200352
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100353/* Allocates new entries for pool <pool> until there are at least <avail> + 1
354 * available, then returns the last one for immediate use, so that at least
355 * <avail> are left available in the pool upon return. NULL is returned if the
356 * last entry could not be allocated. It's important to note that at least one
357 * allocation is always performed even if there are enough entries in the pool.
358 * A call to the garbage collector is performed at most once in case malloc()
359 * returns an error, before returning NULL.
Willy Tarreau50e608d2007-05-13 18:26:08 +0200360 */
Christopher Fauletb349e482017-08-29 09:52:38 +0200361void *__pool_refill_alloc(struct pool_head *pool, unsigned int avail)
Willy Tarreau50e608d2007-05-13 18:26:08 +0200362{
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100363 void *ptr = NULL;
364 int failed = 0;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200365
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100366#ifdef DEBUG_FAIL_ALLOC
367 if (mem_should_fail(pool))
368 return NULL;
369#endif
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100370 /* stop point */
371 avail += pool->used;
372
373 while (1) {
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200374 if (pool->limit && pool->allocated >= pool->limit) {
375 activity[tid].pool_fail++;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200376 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200377 }
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100378
Willy Tarreau606135a2020-06-01 12:35:03 +0200379 swrate_add_scaled(&pool->needed_avg, POOL_AVG_SAMPLES, pool->allocated, POOL_AVG_SAMPLES/4);
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200380 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Willy Tarreauf13322e2017-11-22 10:50:54 +0100381 ptr = pool_alloc_area(pool->size + POOL_EXTRA);
Willy Tarreau82867542019-07-04 11:48:16 +0200382#ifdef DEBUG_MEMORY_POOLS
383 /* keep track of where the element was allocated from. This
384 * is done out of the lock so that the system really allocates
385 * the data without harming other threads waiting on the lock.
386 */
387 if (ptr)
388 *POOL_LINK(pool, ptr) = (void *)pool;
389#endif
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200390 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100391 if (!ptr) {
Willy Tarreau58102cf2015-10-28 16:24:21 +0100392 pool->failed++;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200393 if (failed) {
394 activity[tid].pool_fail++;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100395 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200396 }
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100397 failed++;
Willy Tarreaubafbe012017-11-24 17:34:44 +0100398 pool_gc(pool);
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100399 continue;
400 }
401 if (++pool->allocated > avail)
402 break;
403
Willy Tarreauac421112015-10-28 15:09:29 +0100404 *POOL_LINK(pool, ptr) = (void *)pool->free_list;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100405 pool->free_list = ptr;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200406 }
Willy Tarreau50e608d2007-05-13 18:26:08 +0200407 pool->used++;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100408 return ptr;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200409}
Christopher Fauletb349e482017-08-29 09:52:38 +0200410void *pool_refill_alloc(struct pool_head *pool, unsigned int avail)
411{
412 void *ptr;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200413
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100414 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Christopher Fauletb349e482017-08-29 09:52:38 +0200415 ptr = __pool_refill_alloc(pool, avail);
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100416 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Christopher Fauletb349e482017-08-29 09:52:38 +0200417 return ptr;
418}
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200419/*
420 * This function frees whatever can be freed in pool <pool>.
421 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100422void pool_flush(struct pool_head *pool)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200423{
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200424 void *temp;
425
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200426 if (!pool)
427 return;
428
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200429 while (1) {
430 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
431 temp = pool->free_list;
432 if (!temp) {
433 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
434 break;
435 }
436 pool->free_list = *POOL_LINK(pool, temp);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200437 pool->allocated--;
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200438 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Willy Tarreauf13322e2017-11-22 10:50:54 +0100439 pool_free_area(temp, pool->size + POOL_EXTRA);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200440 }
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200441 /* here, we should have pool->allocated == pool->used */
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200442}
443
444/*
445 * This function frees whatever can be freed in all pools, but respecting
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200446 * the minimum thresholds imposed by owners. It makes sure to be alone to
447 * run by using thread_isolate(). <pool_ctx> is unused.
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200448 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100449void pool_gc(struct pool_head *pool_ctx)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200450{
451 struct pool_head *entry;
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200452 int isolated = thread_isolated();
Willy Tarreaub7f9d122009-04-21 02:17:45 +0200453
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200454 if (!isolated)
455 thread_isolate();
Willy Tarreaub7f9d122009-04-21 02:17:45 +0200456
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200457 list_for_each_entry(entry, &pools, list) {
Olivier Houchard51d93392020-03-12 19:05:39 +0100458 void *temp;
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200459 //qfprintf(stderr, "Flushing pool %s\n", entry->name);
Olivier Houchard51d93392020-03-12 19:05:39 +0100460 while (entry->free_list &&
Willy Tarreau57767b82014-12-22 21:40:55 +0100461 (int)(entry->allocated - entry->used) > (int)entry->minavail) {
Olivier Houchard51d93392020-03-12 19:05:39 +0100462 temp = entry->free_list;
463 entry->free_list = *POOL_LINK(entry, temp);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200464 entry->allocated--;
Willy Tarreauf13322e2017-11-22 10:50:54 +0100465 pool_free_area(temp, entry->size + POOL_EXTRA);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200466 }
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200467 }
Christopher Fauletb349e482017-08-29 09:52:38 +0200468
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200469 if (!isolated)
470 thread_release();
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200471}
Olivier Houchardcf975d42018-01-24 18:38:31 +0100472#endif
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200473
474/*
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200475 * This function destroys a pool by freeing it completely, unless it's still
476 * in use. This should be called only under extreme circumstances. It always
477 * returns NULL if the resulting pool is empty, easing the clearing of the old
478 * pointer, otherwise it returns the pool.
479 * .
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200480 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100481void *pool_destroy(struct pool_head *pool)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200482{
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200483 if (pool) {
Willy Tarreaubafbe012017-11-24 17:34:44 +0100484 pool_flush(pool);
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200485 if (pool->used)
486 return pool;
487 pool->users--;
488 if (!pool->users) {
489 LIST_DEL(&pool->list);
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100490#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100491 HA_SPIN_DESTROY(&pool->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100492#endif
Willy Tarreau9f3129e2021-04-17 00:31:38 +0200493 /* note that if used == 0, the cache is empty */
494 free(pool);
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200495 }
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200496 }
497 return NULL;
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200498}
499
Willy Tarreau2455ceb2018-11-26 15:57:34 +0100500/* This destroys all pools on exit. It is *not* thread safe. */
501void pool_destroy_all()
502{
503 struct pool_head *entry, *back;
504
505 list_for_each_entry_safe(entry, back, &pools, list)
506 pool_destroy(entry);
507}
508
Willy Tarreau12833bb2014-01-28 16:49:56 +0100509/* This function dumps memory usage information into the trash buffer. */
510void dump_pools_to_trash()
Willy Tarreau50e608d2007-05-13 18:26:08 +0200511{
512 struct pool_head *entry;
513 unsigned long allocated, used;
514 int nbpools;
515
516 allocated = used = nbpools = 0;
Willy Tarreau12833bb2014-01-28 16:49:56 +0100517 chunk_printf(&trash, "Dumping pools usage. Use SIGQUIT to flush them.\n");
Willy Tarreau50e608d2007-05-13 18:26:08 +0200518 list_for_each_entry(entry, &pools, list) {
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100519#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100520 HA_SPIN_LOCK(POOL_LOCK, &entry->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100521#endif
Willy Tarreau9f3129e2021-04-17 00:31:38 +0200522 chunk_appendf(&trash, " - Pool %s (%u bytes) : %u allocated (%u bytes), %u used, needed_avg %u, %u failures, %u users, @%p%s\n",
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200523 entry->name, entry->size, entry->allocated,
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200524 entry->size * entry->allocated, entry->used,
Willy Tarreau606135a2020-06-01 12:35:03 +0200525 swrate_avg(entry->needed_avg, POOL_AVG_SAMPLES), entry->failed,
Willy Tarreau9f3129e2021-04-17 00:31:38 +0200526 entry->users, entry,
Willy Tarreau0a93b642018-10-16 07:58:39 +0200527 (entry->flags & MEM_F_SHARED) ? " [SHARED]" : "");
Willy Tarreau50e608d2007-05-13 18:26:08 +0200528
529 allocated += entry->allocated * entry->size;
530 used += entry->used * entry->size;
531 nbpools++;
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100532#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100533 HA_SPIN_UNLOCK(POOL_LOCK, &entry->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100534#endif
Willy Tarreau50e608d2007-05-13 18:26:08 +0200535 }
Willy Tarreau12833bb2014-01-28 16:49:56 +0100536 chunk_appendf(&trash, "Total: %d pools, %lu bytes allocated, %lu used.\n",
Willy Tarreau50e608d2007-05-13 18:26:08 +0200537 nbpools, allocated, used);
538}
539
Willy Tarreau12833bb2014-01-28 16:49:56 +0100540/* Dump statistics on pools usage. */
541void dump_pools(void)
542{
543 dump_pools_to_trash();
Willy Tarreau843b7cb2018-07-13 10:54:26 +0200544 qfprintf(stderr, "%s", trash.area);
Willy Tarreau12833bb2014-01-28 16:49:56 +0100545}
546
Willy Tarreau58102cf2015-10-28 16:24:21 +0100547/* This function returns the total number of failed pool allocations */
548int pool_total_failures()
549{
550 struct pool_head *entry;
551 int failed = 0;
552
553 list_for_each_entry(entry, &pools, list)
554 failed += entry->failed;
555 return failed;
556}
557
558/* This function returns the total amount of memory allocated in pools (in bytes) */
559unsigned long pool_total_allocated()
560{
561 struct pool_head *entry;
562 unsigned long allocated = 0;
563
564 list_for_each_entry(entry, &pools, list)
565 allocated += entry->allocated * entry->size;
566 return allocated;
567}
568
569/* This function returns the total amount of memory used in pools (in bytes) */
570unsigned long pool_total_used()
571{
572 struct pool_head *entry;
573 unsigned long used = 0;
574
575 list_for_each_entry(entry, &pools, list)
576 used += entry->used * entry->size;
577 return used;
578}
579
William Lallemande7ed8852016-11-19 02:25:36 +0100580/* This function dumps memory usage information onto the stream interface's
581 * read buffer. It returns 0 as long as it does not complete, non-zero upon
582 * completion. No state is used.
583 */
584static int cli_io_handler_dump_pools(struct appctx *appctx)
585{
586 struct stream_interface *si = appctx->owner;
587
588 dump_pools_to_trash();
Willy Tarreau06d80a92017-10-19 14:32:15 +0200589 if (ci_putchk(si_ic(si), &trash) == -1) {
Willy Tarreaudb398432018-11-15 11:08:52 +0100590 si_rx_room_blk(si);
William Lallemande7ed8852016-11-19 02:25:36 +0100591 return 0;
592 }
593 return 1;
594}
595
Willy Tarreau7107c8b2018-11-26 11:44:35 +0100596/* callback used to create early pool <name> of size <size> and store the
597 * resulting pointer into <ptr>. If the allocation fails, it quits with after
598 * emitting an error message.
599 */
600void create_pool_callback(struct pool_head **ptr, char *name, unsigned int size)
601{
602 *ptr = create_pool(name, size, MEM_F_SHARED);
603 if (!*ptr) {
604 ha_alert("Failed to allocate pool '%s' of size %u : %s. Aborting.\n",
605 name, size, strerror(errno));
606 exit(1);
607 }
608}
609
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100610/* Initializes all per-thread arrays on startup */
611static void init_pools()
612{
Willy Tarreau2d6f6282021-04-15 16:24:00 +0200613#ifdef CONFIG_HAP_POOLS
Willy Tarreau9f3129e2021-04-17 00:31:38 +0200614 int thr;
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100615
616 for (thr = 0; thr < MAX_THREADS; thr++) {
Willy Tarreau20dc3cd2020-06-28 00:54:27 +0200617 LIST_INIT(&ha_thread_info[thr].pool_lru_head);
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100618 }
Willy Tarreaued891fd2020-06-01 19:00:28 +0200619#endif
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100620}
621
622INITCALL0(STG_PREPARE, init_pools);
Willy Tarreau7107c8b2018-11-26 11:44:35 +0100623
William Lallemande7ed8852016-11-19 02:25:36 +0100624/* register cli keywords */
625static struct cli_kw_list cli_kws = {{ },{
Willy Tarreaue9ecec82016-12-16 18:55:23 +0100626 { { "show", "pools", NULL }, "show pools : report information about the memory pools usage", NULL, cli_io_handler_dump_pools },
William Lallemande7ed8852016-11-19 02:25:36 +0100627 {{},}
628}};
629
Willy Tarreau0108d902018-11-25 19:14:37 +0100630INITCALL1(STG_REGISTER, cli_register_kw, &cli_kws);
William Lallemande7ed8852016-11-19 02:25:36 +0100631
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100632#ifdef DEBUG_FAIL_ALLOC
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100633
634int mem_should_fail(const struct pool_head *pool)
635{
Olivier Houchard9c4f08a2019-02-01 16:28:04 +0100636 int ret = 0;
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100637
638 if (mem_fail_rate > 0 && !(global.mode & MODE_STARTING)) {
Willy Tarreau52bf8392020-03-08 00:42:37 +0100639 int randnb = ha_random() % 100;
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100640
641 if (mem_fail_rate > randnb)
642 ret = 1;
643 else
644 ret = 0;
645 }
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100646 return ret;
647
648}
649
650/* config parser for global "tune.fail-alloc" */
651static int mem_parse_global_fail_alloc(char **args, int section_type, struct proxy *curpx,
Amaury Denoyelle3b1c9a32021-03-22 11:21:36 +0100652 const struct proxy *defpx, const char *file, int line,
653 char **err)
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100654{
655 if (too_many_args(1, args, err, NULL))
656 return -1;
657 mem_fail_rate = atoi(args[1]);
658 if (mem_fail_rate < 0 || mem_fail_rate > 100) {
659 memprintf(err, "'%s' expects a numeric value between 0 and 100.", args[0]);
660 return -1;
661 }
662 return 0;
663}
664#endif
665
666/* register global config keywords */
667static struct cfg_kw_list mem_cfg_kws = {ILH, {
668#ifdef DEBUG_FAIL_ALLOC
669 { CFG_GLOBAL, "tune.fail-alloc", mem_parse_global_fail_alloc },
670#endif
671 { 0, NULL, NULL }
672}};
673
674INITCALL1(STG_REGISTER, cfg_register_keywords, &mem_cfg_kws);
675
Willy Tarreau50e608d2007-05-13 18:26:08 +0200676/*
677 * Local variables:
678 * c-indent-level: 8
679 * c-basic-offset: 8
680 * End:
681 */