blob: 6f19ec783e3d8f9d1521ae9156d68afcf77439bd [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 Tarreaued891fd2020-06-01 19:00:28 +020030#ifdef CONFIG_HAP_LOCAL_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 Tarreaued891fd2020-06-01 19:00:28 +0200110#ifdef CONFIG_HAP_LOCAL_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 Tarreaued891fd2020-06-01 19:00:28 +0200122#ifdef CONFIG_HAP_LOCAL_POOLS
123/* 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
159 if (limit && allocated >= limit) {
Willy Tarreau0bae0752021-03-02 20:05:09 +0100160 activity[tid].pool_fail++;
161 return NULL;
162 }
163
164 ptr = pool_alloc_area(pool->size + POOL_EXTRA);
165 if (!ptr) {
Willy Tarreau4781b152021-04-06 13:53:36 +0200166 _HA_ATOMIC_INC(&pool->failed);
Willy Tarreau0bae0752021-03-02 20:05:09 +0100167 activity[tid].pool_fail++;
168 return NULL;
169 }
170
Willy Tarreau4781b152021-04-06 13:53:36 +0200171 _HA_ATOMIC_INC(&pool->allocated);
172 _HA_ATOMIC_INC(&pool->used);
Willy Tarreau0bae0752021-03-02 20:05:09 +0100173
174#ifdef DEBUG_MEMORY_POOLS
175 /* keep track of where the element was allocated from */
176 *POOL_LINK(pool, ptr) = (void *)pool;
177#endif
178 return ptr;
179}
180
181/* legacy stuff */
182void *pool_refill_alloc(struct pool_head *pool, unsigned int avail)
183{
184 void *ptr;
185
186 ptr = __pool_refill_alloc(pool, avail);
187 return ptr;
188}
189
190/* legacy stuff */
191void pool_flush(struct pool_head *pool)
192{
193}
194
195/* This function might ask the malloc library to trim its buffers. */
196void pool_gc(struct pool_head *pool_ctx)
197{
198#if defined(HA_HAVE_MALLOC_TRIM)
199 malloc_trim(0);
200#endif
201}
202
203#elif defined(CONFIG_HAP_LOCKLESS_POOLS)
204
Olivier Houchardcf975d42018-01-24 18:38:31 +0100205/* Allocates new entries for pool <pool> until there are at least <avail> + 1
206 * available, then returns the last one for immediate use, so that at least
207 * <avail> are left available in the pool upon return. NULL is returned if the
208 * last entry could not be allocated. It's important to note that at least one
209 * allocation is always performed even if there are enough entries in the pool.
210 * A call to the garbage collector is performed at most once in case malloc()
211 * returns an error, before returning NULL.
212 */
213void *__pool_refill_alloc(struct pool_head *pool, unsigned int avail)
214{
Olivier Houchard8b2c8a72018-10-21 01:52:59 +0200215 void *ptr = NULL, **free_list;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100216 int failed = 0;
217 int size = pool->size;
218 int limit = pool->limit;
219 int allocated = pool->allocated, allocated_orig = allocated;
220
221 /* stop point */
222 avail += pool->used;
223
224 while (1) {
225 if (limit && allocated >= limit) {
Olivier Houchard20872762019-03-08 18:53:35 +0100226 _HA_ATOMIC_ADD(&pool->allocated, allocated - allocated_orig);
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200227 activity[tid].pool_fail++;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100228 return NULL;
229 }
230
Willy Tarreau606135a2020-06-01 12:35:03 +0200231 swrate_add_scaled(&pool->needed_avg, POOL_AVG_SAMPLES, pool->allocated, POOL_AVG_SAMPLES/4);
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200232
Willy Tarreau24aa1ee2020-05-30 18:56:17 +0200233 ptr = pool_alloc_area(size + POOL_EXTRA);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100234 if (!ptr) {
Willy Tarreau4781b152021-04-06 13:53:36 +0200235 _HA_ATOMIC_INC(&pool->failed);
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200236 if (failed) {
237 activity[tid].pool_fail++;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100238 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200239 }
Olivier Houchardcf975d42018-01-24 18:38:31 +0100240 failed++;
241 pool_gc(pool);
242 continue;
243 }
244 if (++allocated > avail)
245 break;
246
247 free_list = pool->free_list;
248 do {
249 *POOL_LINK(pool, ptr) = free_list;
250 __ha_barrier_store();
Olivier Houchard20872762019-03-08 18:53:35 +0100251 } while (_HA_ATOMIC_CAS(&pool->free_list, &free_list, ptr) == 0);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100252 }
Olivier Houchard20872762019-03-08 18:53:35 +0100253 __ha_barrier_atomic_store();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100254
Olivier Houchard20872762019-03-08 18:53:35 +0100255 _HA_ATOMIC_ADD(&pool->allocated, allocated - allocated_orig);
Willy Tarreau4781b152021-04-06 13:53:36 +0200256 _HA_ATOMIC_INC(&pool->used);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100257
258#ifdef DEBUG_MEMORY_POOLS
259 /* keep track of where the element was allocated from */
260 *POOL_LINK(pool, ptr) = (void *)pool;
261#endif
262 return ptr;
263}
264void *pool_refill_alloc(struct pool_head *pool, unsigned int avail)
265{
266 void *ptr;
267
268 ptr = __pool_refill_alloc(pool, avail);
269 return ptr;
270}
271/*
272 * This function frees whatever can be freed in pool <pool>.
273 */
274void pool_flush(struct pool_head *pool)
275{
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100276 struct pool_free_list cmp, new;
Olivier Houchard8b2c8a72018-10-21 01:52:59 +0200277 void **next, *temp;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100278 int removed = 0;
279
280 if (!pool)
281 return;
Willy Tarreau21072b92020-05-29 17:23:05 +0200282 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100283 do {
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100284 cmp.free_list = pool->free_list;
285 cmp.seq = pool->seq;
286 new.free_list = NULL;
287 new.seq = cmp.seq + 1;
288 } while (!_HA_ATOMIC_DWCAS(&pool->free_list, &cmp, &new));
Olivier Houchard20872762019-03-08 18:53:35 +0100289 __ha_barrier_atomic_store();
Willy Tarreau21072b92020-05-29 17:23:05 +0200290 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100291 next = cmp.free_list;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100292 while (next) {
293 temp = next;
294 next = *POOL_LINK(pool, temp);
295 removed++;
Willy Tarreau24aa1ee2020-05-30 18:56:17 +0200296 pool_free_area(temp, pool->size + POOL_EXTRA);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100297 }
298 pool->free_list = next;
Olivier Houchard20872762019-03-08 18:53:35 +0100299 _HA_ATOMIC_SUB(&pool->allocated, removed);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100300 /* here, we should have pool->allocate == pool->used */
301}
302
303/*
304 * This function frees whatever can be freed in all pools, but respecting
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200305 * the minimum thresholds imposed by owners. It makes sure to be alone to
306 * run by using thread_isolate(). <pool_ctx> is unused.
Olivier Houchardcf975d42018-01-24 18:38:31 +0100307 */
308void pool_gc(struct pool_head *pool_ctx)
309{
Olivier Houchardcf975d42018-01-24 18:38:31 +0100310 struct pool_head *entry;
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200311 int isolated = thread_isolated();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100312
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200313 if (!isolated)
314 thread_isolate();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100315
316 list_for_each_entry(entry, &pools, list) {
317 while ((int)((volatile int)entry->allocated - (volatile int)entry->used) > (int)entry->minavail) {
318 struct pool_free_list cmp, new;
319
320 cmp.seq = entry->seq;
321 __ha_barrier_load();
322 cmp.free_list = entry->free_list;
323 __ha_barrier_load();
324 if (cmp.free_list == NULL)
325 break;
326 new.free_list = *POOL_LINK(entry, cmp.free_list);
327 new.seq = cmp.seq + 1;
Willy Tarreau6a38b322019-05-11 18:04:24 +0200328 if (HA_ATOMIC_DWCAS(&entry->free_list, &cmp, &new) == 0)
Olivier Houchardcf975d42018-01-24 18:38:31 +0100329 continue;
Willy Tarreau24aa1ee2020-05-30 18:56:17 +0200330 pool_free_area(cmp.free_list, entry->size + POOL_EXTRA);
Willy Tarreau4781b152021-04-06 13:53:36 +0200331 _HA_ATOMIC_DEC(&entry->allocated);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100332 }
333 }
334
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200335 if (!isolated)
336 thread_release();
Willy Tarreau88366c22020-11-03 15:53:34 +0100337
338#if defined(HA_HAVE_MALLOC_TRIM)
339 malloc_trim(0);
340#endif
Olivier Houchardcf975d42018-01-24 18:38:31 +0100341}
Willy Tarreaue18db9e2018-10-16 10:28:54 +0200342
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100343#else /* CONFIG_HAP_LOCKLESS_POOLS */
Willy Tarreau50e608d2007-05-13 18:26:08 +0200344
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100345/* Allocates new entries for pool <pool> until there are at least <avail> + 1
346 * available, then returns the last one for immediate use, so that at least
347 * <avail> are left available in the pool upon return. NULL is returned if the
348 * last entry could not be allocated. It's important to note that at least one
349 * allocation is always performed even if there are enough entries in the pool.
350 * A call to the garbage collector is performed at most once in case malloc()
351 * returns an error, before returning NULL.
Willy Tarreau50e608d2007-05-13 18:26:08 +0200352 */
Christopher Fauletb349e482017-08-29 09:52:38 +0200353void *__pool_refill_alloc(struct pool_head *pool, unsigned int avail)
Willy Tarreau50e608d2007-05-13 18:26:08 +0200354{
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100355 void *ptr = NULL;
356 int failed = 0;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200357
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100358#ifdef DEBUG_FAIL_ALLOC
359 if (mem_should_fail(pool))
360 return NULL;
361#endif
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100362 /* stop point */
363 avail += pool->used;
364
365 while (1) {
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200366 if (pool->limit && pool->allocated >= pool->limit) {
367 activity[tid].pool_fail++;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200368 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200369 }
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100370
Willy Tarreau606135a2020-06-01 12:35:03 +0200371 swrate_add_scaled(&pool->needed_avg, POOL_AVG_SAMPLES, pool->allocated, POOL_AVG_SAMPLES/4);
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200372 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Willy Tarreauf13322e2017-11-22 10:50:54 +0100373 ptr = pool_alloc_area(pool->size + POOL_EXTRA);
Willy Tarreau82867542019-07-04 11:48:16 +0200374#ifdef DEBUG_MEMORY_POOLS
375 /* keep track of where the element was allocated from. This
376 * is done out of the lock so that the system really allocates
377 * the data without harming other threads waiting on the lock.
378 */
379 if (ptr)
380 *POOL_LINK(pool, ptr) = (void *)pool;
381#endif
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200382 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100383 if (!ptr) {
Willy Tarreau58102cf2015-10-28 16:24:21 +0100384 pool->failed++;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200385 if (failed) {
386 activity[tid].pool_fail++;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100387 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200388 }
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100389 failed++;
Willy Tarreaubafbe012017-11-24 17:34:44 +0100390 pool_gc(pool);
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100391 continue;
392 }
393 if (++pool->allocated > avail)
394 break;
395
Willy Tarreauac421112015-10-28 15:09:29 +0100396 *POOL_LINK(pool, ptr) = (void *)pool->free_list;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100397 pool->free_list = ptr;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200398 }
Willy Tarreau50e608d2007-05-13 18:26:08 +0200399 pool->used++;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100400 return ptr;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200401}
Christopher Fauletb349e482017-08-29 09:52:38 +0200402void *pool_refill_alloc(struct pool_head *pool, unsigned int avail)
403{
404 void *ptr;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200405
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100406 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Christopher Fauletb349e482017-08-29 09:52:38 +0200407 ptr = __pool_refill_alloc(pool, avail);
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100408 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Christopher Fauletb349e482017-08-29 09:52:38 +0200409 return ptr;
410}
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200411/*
412 * This function frees whatever can be freed in pool <pool>.
413 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100414void pool_flush(struct pool_head *pool)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200415{
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200416 void *temp;
417
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200418 if (!pool)
419 return;
420
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200421 while (1) {
422 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
423 temp = pool->free_list;
424 if (!temp) {
425 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
426 break;
427 }
428 pool->free_list = *POOL_LINK(pool, temp);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200429 pool->allocated--;
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200430 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Willy Tarreauf13322e2017-11-22 10:50:54 +0100431 pool_free_area(temp, pool->size + POOL_EXTRA);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200432 }
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200433 /* here, we should have pool->allocated == pool->used */
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200434}
435
436/*
437 * This function frees whatever can be freed in all pools, but respecting
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200438 * the minimum thresholds imposed by owners. It makes sure to be alone to
439 * run by using thread_isolate(). <pool_ctx> is unused.
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200440 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100441void pool_gc(struct pool_head *pool_ctx)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200442{
443 struct pool_head *entry;
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200444 int isolated = thread_isolated();
Willy Tarreaub7f9d122009-04-21 02:17:45 +0200445
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200446 if (!isolated)
447 thread_isolate();
Willy Tarreaub7f9d122009-04-21 02:17:45 +0200448
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200449 list_for_each_entry(entry, &pools, list) {
Olivier Houchard51d93392020-03-12 19:05:39 +0100450 void *temp;
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200451 //qfprintf(stderr, "Flushing pool %s\n", entry->name);
Olivier Houchard51d93392020-03-12 19:05:39 +0100452 while (entry->free_list &&
Willy Tarreau57767b82014-12-22 21:40:55 +0100453 (int)(entry->allocated - entry->used) > (int)entry->minavail) {
Olivier Houchard51d93392020-03-12 19:05:39 +0100454 temp = entry->free_list;
455 entry->free_list = *POOL_LINK(entry, temp);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200456 entry->allocated--;
Willy Tarreauf13322e2017-11-22 10:50:54 +0100457 pool_free_area(temp, entry->size + POOL_EXTRA);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200458 }
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200459 }
Christopher Fauletb349e482017-08-29 09:52:38 +0200460
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200461 if (!isolated)
462 thread_release();
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200463}
Olivier Houchardcf975d42018-01-24 18:38:31 +0100464#endif
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200465
466/*
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200467 * This function destroys a pool by freeing it completely, unless it's still
468 * in use. This should be called only under extreme circumstances. It always
469 * returns NULL if the resulting pool is empty, easing the clearing of the old
470 * pointer, otherwise it returns the pool.
471 * .
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200472 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100473void *pool_destroy(struct pool_head *pool)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200474{
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200475 if (pool) {
Willy Tarreaubafbe012017-11-24 17:34:44 +0100476 pool_flush(pool);
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200477 if (pool->used)
478 return pool;
479 pool->users--;
480 if (!pool->users) {
481 LIST_DEL(&pool->list);
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100482#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100483 HA_SPIN_DESTROY(&pool->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100484#endif
Willy Tarreau9f3129e2021-04-17 00:31:38 +0200485 /* note that if used == 0, the cache is empty */
486 free(pool);
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200487 }
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200488 }
489 return NULL;
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200490}
491
Willy Tarreau2455ceb2018-11-26 15:57:34 +0100492/* This destroys all pools on exit. It is *not* thread safe. */
493void pool_destroy_all()
494{
495 struct pool_head *entry, *back;
496
497 list_for_each_entry_safe(entry, back, &pools, list)
498 pool_destroy(entry);
499}
500
Willy Tarreau12833bb2014-01-28 16:49:56 +0100501/* This function dumps memory usage information into the trash buffer. */
502void dump_pools_to_trash()
Willy Tarreau50e608d2007-05-13 18:26:08 +0200503{
504 struct pool_head *entry;
505 unsigned long allocated, used;
506 int nbpools;
507
508 allocated = used = nbpools = 0;
Willy Tarreau12833bb2014-01-28 16:49:56 +0100509 chunk_printf(&trash, "Dumping pools usage. Use SIGQUIT to flush them.\n");
Willy Tarreau50e608d2007-05-13 18:26:08 +0200510 list_for_each_entry(entry, &pools, list) {
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100511#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100512 HA_SPIN_LOCK(POOL_LOCK, &entry->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100513#endif
Willy Tarreau9f3129e2021-04-17 00:31:38 +0200514 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 +0200515 entry->name, entry->size, entry->allocated,
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200516 entry->size * entry->allocated, entry->used,
Willy Tarreau606135a2020-06-01 12:35:03 +0200517 swrate_avg(entry->needed_avg, POOL_AVG_SAMPLES), entry->failed,
Willy Tarreau9f3129e2021-04-17 00:31:38 +0200518 entry->users, entry,
Willy Tarreau0a93b642018-10-16 07:58:39 +0200519 (entry->flags & MEM_F_SHARED) ? " [SHARED]" : "");
Willy Tarreau50e608d2007-05-13 18:26:08 +0200520
521 allocated += entry->allocated * entry->size;
522 used += entry->used * entry->size;
523 nbpools++;
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100524#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100525 HA_SPIN_UNLOCK(POOL_LOCK, &entry->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100526#endif
Willy Tarreau50e608d2007-05-13 18:26:08 +0200527 }
Willy Tarreau12833bb2014-01-28 16:49:56 +0100528 chunk_appendf(&trash, "Total: %d pools, %lu bytes allocated, %lu used.\n",
Willy Tarreau50e608d2007-05-13 18:26:08 +0200529 nbpools, allocated, used);
530}
531
Willy Tarreau12833bb2014-01-28 16:49:56 +0100532/* Dump statistics on pools usage. */
533void dump_pools(void)
534{
535 dump_pools_to_trash();
Willy Tarreau843b7cb2018-07-13 10:54:26 +0200536 qfprintf(stderr, "%s", trash.area);
Willy Tarreau12833bb2014-01-28 16:49:56 +0100537}
538
Willy Tarreau58102cf2015-10-28 16:24:21 +0100539/* This function returns the total number of failed pool allocations */
540int pool_total_failures()
541{
542 struct pool_head *entry;
543 int failed = 0;
544
545 list_for_each_entry(entry, &pools, list)
546 failed += entry->failed;
547 return failed;
548}
549
550/* This function returns the total amount of memory allocated in pools (in bytes) */
551unsigned long pool_total_allocated()
552{
553 struct pool_head *entry;
554 unsigned long allocated = 0;
555
556 list_for_each_entry(entry, &pools, list)
557 allocated += entry->allocated * entry->size;
558 return allocated;
559}
560
561/* This function returns the total amount of memory used in pools (in bytes) */
562unsigned long pool_total_used()
563{
564 struct pool_head *entry;
565 unsigned long used = 0;
566
567 list_for_each_entry(entry, &pools, list)
568 used += entry->used * entry->size;
569 return used;
570}
571
William Lallemande7ed8852016-11-19 02:25:36 +0100572/* This function dumps memory usage information onto the stream interface's
573 * read buffer. It returns 0 as long as it does not complete, non-zero upon
574 * completion. No state is used.
575 */
576static int cli_io_handler_dump_pools(struct appctx *appctx)
577{
578 struct stream_interface *si = appctx->owner;
579
580 dump_pools_to_trash();
Willy Tarreau06d80a92017-10-19 14:32:15 +0200581 if (ci_putchk(si_ic(si), &trash) == -1) {
Willy Tarreaudb398432018-11-15 11:08:52 +0100582 si_rx_room_blk(si);
William Lallemande7ed8852016-11-19 02:25:36 +0100583 return 0;
584 }
585 return 1;
586}
587
Willy Tarreau7107c8b2018-11-26 11:44:35 +0100588/* callback used to create early pool <name> of size <size> and store the
589 * resulting pointer into <ptr>. If the allocation fails, it quits with after
590 * emitting an error message.
591 */
592void create_pool_callback(struct pool_head **ptr, char *name, unsigned int size)
593{
594 *ptr = create_pool(name, size, MEM_F_SHARED);
595 if (!*ptr) {
596 ha_alert("Failed to allocate pool '%s' of size %u : %s. Aborting.\n",
597 name, size, strerror(errno));
598 exit(1);
599 }
600}
601
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100602/* Initializes all per-thread arrays on startup */
603static void init_pools()
604{
Willy Tarreaued891fd2020-06-01 19:00:28 +0200605#ifdef CONFIG_HAP_LOCAL_POOLS
Willy Tarreau9f3129e2021-04-17 00:31:38 +0200606 int thr;
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100607
608 for (thr = 0; thr < MAX_THREADS; thr++) {
Willy Tarreau20dc3cd2020-06-28 00:54:27 +0200609 LIST_INIT(&ha_thread_info[thr].pool_lru_head);
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100610 }
Willy Tarreaued891fd2020-06-01 19:00:28 +0200611#endif
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100612}
613
614INITCALL0(STG_PREPARE, init_pools);
Willy Tarreau7107c8b2018-11-26 11:44:35 +0100615
William Lallemande7ed8852016-11-19 02:25:36 +0100616/* register cli keywords */
617static struct cli_kw_list cli_kws = {{ },{
Willy Tarreaue9ecec82016-12-16 18:55:23 +0100618 { { "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 +0100619 {{},}
620}};
621
Willy Tarreau0108d902018-11-25 19:14:37 +0100622INITCALL1(STG_REGISTER, cli_register_kw, &cli_kws);
William Lallemande7ed8852016-11-19 02:25:36 +0100623
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100624#ifdef DEBUG_FAIL_ALLOC
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100625
626int mem_should_fail(const struct pool_head *pool)
627{
Olivier Houchard9c4f08a2019-02-01 16:28:04 +0100628 int ret = 0;
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100629
630 if (mem_fail_rate > 0 && !(global.mode & MODE_STARTING)) {
Willy Tarreau52bf8392020-03-08 00:42:37 +0100631 int randnb = ha_random() % 100;
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100632
633 if (mem_fail_rate > randnb)
634 ret = 1;
635 else
636 ret = 0;
637 }
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100638 return ret;
639
640}
641
642/* config parser for global "tune.fail-alloc" */
643static int mem_parse_global_fail_alloc(char **args, int section_type, struct proxy *curpx,
Amaury Denoyelle3b1c9a32021-03-22 11:21:36 +0100644 const struct proxy *defpx, const char *file, int line,
645 char **err)
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100646{
647 if (too_many_args(1, args, err, NULL))
648 return -1;
649 mem_fail_rate = atoi(args[1]);
650 if (mem_fail_rate < 0 || mem_fail_rate > 100) {
651 memprintf(err, "'%s' expects a numeric value between 0 and 100.", args[0]);
652 return -1;
653 }
654 return 0;
655}
656#endif
657
658/* register global config keywords */
659static struct cfg_kw_list mem_cfg_kws = {ILH, {
660#ifdef DEBUG_FAIL_ALLOC
661 { CFG_GLOBAL, "tune.fail-alloc", mem_parse_global_fail_alloc },
662#endif
663 { 0, NULL, NULL }
664}};
665
666INITCALL1(STG_REGISTER, cfg_register_keywords, &mem_cfg_kws);
667
Willy Tarreau50e608d2007-05-13 18:26:08 +0200668/*
669 * Local variables:
670 * c-indent-level: 8
671 * c-basic-offset: 8
672 * End:
673 */