blob: 321f8bc677f7addd45a685e03dadcd399bff4ae6 [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 Tarreau0a93b642018-10-16 07:58:39 +020031/* These are the most common pools, expected to be initialized first. These
32 * ones are allocated from an array, allowing to map them to an index.
33 */
34struct pool_head pool_base_start[MAX_BASE_POOLS] = { };
35unsigned int pool_base_count = 0;
36
Willy Tarreau7f0165e2018-11-26 17:09:46 +010037/* These ones are initialized per-thread on startup by init_pools() */
38struct pool_cache_head pool_cache[MAX_THREADS][MAX_BASE_POOLS];
Willy Tarreaue18db9e2018-10-16 10:28:54 +020039THREAD_LOCAL size_t pool_cache_bytes = 0; /* total cache size */
40THREAD_LOCAL size_t pool_cache_count = 0; /* #cache objects */
Willy Tarreaued891fd2020-06-01 19:00:28 +020041#endif
Willy Tarreaue18db9e2018-10-16 10:28:54 +020042
Willy Tarreau50e608d2007-05-13 18:26:08 +020043static struct list pools = LIST_HEAD_INIT(pools);
Willy Tarreau067ac9f2015-10-08 14:12:13 +020044int mem_poison_byte = -1;
Willy Tarreau50e608d2007-05-13 18:26:08 +020045
Olivier Houcharddc21ff72019-01-29 15:20:16 +010046#ifdef DEBUG_FAIL_ALLOC
47static int mem_fail_rate = 0;
48static int mem_should_fail(const struct pool_head *);
49#endif
50
Willy Tarreau50e608d2007-05-13 18:26:08 +020051/* Try to find an existing shared pool with the same characteristics and
52 * returns it, otherwise creates this one. NULL is returned if no memory
Willy Tarreau581bf812016-01-25 02:19:13 +010053 * is available for a new creation. Two flags are supported :
54 * - MEM_F_SHARED to indicate that the pool may be shared with other users
55 * - MEM_F_EXACT to indicate that the size must not be rounded up
Willy Tarreau50e608d2007-05-13 18:26:08 +020056 */
57struct pool_head *create_pool(char *name, unsigned int size, unsigned int flags)
58{
59 struct pool_head *pool;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +020060 struct pool_head *entry;
61 struct list *start;
Willy Tarreau50e608d2007-05-13 18:26:08 +020062 unsigned int align;
Willy Tarreaued891fd2020-06-01 19:00:28 +020063 int idx __maybe_unused;
Willy Tarreau50e608d2007-05-13 18:26:08 +020064
Willy Tarreauac421112015-10-28 15:09:29 +010065 /* We need to store a (void *) at the end of the chunks. Since we know
Willy Tarreau50e608d2007-05-13 18:26:08 +020066 * that the malloc() function will never return such a small size,
67 * let's round the size up to something slightly bigger, in order to
68 * ease merging of entries. Note that the rounding is a power of two.
Willy Tarreauac421112015-10-28 15:09:29 +010069 * This extra (void *) is not accounted for in the size computation
70 * so that the visible parts outside are not affected.
Willy Tarreau30f931e2018-10-23 14:40:23 +020071 *
72 * Note: for the LRU cache, we need to store 2 doubly-linked lists.
Willy Tarreau50e608d2007-05-13 18:26:08 +020073 */
74
Willy Tarreau581bf812016-01-25 02:19:13 +010075 if (!(flags & MEM_F_EXACT)) {
Willy Tarreau30f931e2018-10-23 14:40:23 +020076 align = 4 * sizeof(void *); // 2 lists = 4 pointers min
Willy Tarreau581bf812016-01-25 02:19:13 +010077 size = ((size + POOL_EXTRA + align - 1) & -align) - POOL_EXTRA;
78 }
Willy Tarreau50e608d2007-05-13 18:26:08 +020079
Christopher Fauletb349e482017-08-29 09:52:38 +020080 /* TODO: thread: we do not lock pool list for now because all pools are
81 * created during HAProxy startup (so before threads creation) */
Willy Tarreau7dcd46d2007-05-14 00:16:13 +020082 start = &pools;
Willy Tarreau50e608d2007-05-13 18:26:08 +020083 pool = NULL;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +020084
85 list_for_each_entry(entry, &pools, list) {
86 if (entry->size == size) {
87 /* either we can share this place and we take it, or
Ilya Shipitsin47d17182020-06-21 21:42:57 +050088 * we look for a shareable one or for the next position
Willy Tarreau7dcd46d2007-05-14 00:16:13 +020089 * before which we will insert a new one.
90 */
91 if (flags & entry->flags & MEM_F_SHARED) {
92 /* we can share this one */
Willy Tarreau50e608d2007-05-13 18:26:08 +020093 pool = entry;
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +020094 DPRINTF(stderr, "Sharing %s with %s\n", name, pool->name);
Willy Tarreau50e608d2007-05-13 18:26:08 +020095 break;
96 }
97 }
Willy Tarreau7dcd46d2007-05-14 00:16:13 +020098 else if (entry->size > size) {
99 /* insert before this one */
100 start = &entry->list;
101 break;
102 }
Willy Tarreau50e608d2007-05-13 18:26:08 +0200103 }
104
105 if (!pool) {
Willy Tarreaued891fd2020-06-01 19:00:28 +0200106#ifdef CONFIG_HAP_LOCAL_POOLS
Willy Tarreau0a93b642018-10-16 07:58:39 +0200107 if (pool_base_count < MAX_BASE_POOLS)
108 pool = &pool_base_start[pool_base_count++];
109
110 if (!pool) {
111 /* look for a freed entry */
112 for (entry = pool_base_start; entry != pool_base_start + MAX_BASE_POOLS; entry++) {
113 if (!entry->size) {
114 pool = entry;
115 break;
116 }
117 }
118 }
Willy Tarreaued891fd2020-06-01 19:00:28 +0200119#endif
Willy Tarreau0a93b642018-10-16 07:58:39 +0200120
121 if (!pool)
122 pool = calloc(1, sizeof(*pool));
123
Willy Tarreau50e608d2007-05-13 18:26:08 +0200124 if (!pool)
125 return NULL;
126 if (name)
127 strlcpy2(pool->name, name, sizeof(pool->name));
128 pool->size = size;
129 pool->flags = flags;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200130 LIST_ADDQ(start, &pool->list);
Christopher Faulet2f6d3c02019-06-25 21:45:59 +0200131
Willy Tarreaued891fd2020-06-01 19:00:28 +0200132#ifdef CONFIG_HAP_LOCAL_POOLS
Christopher Faulet2f6d3c02019-06-25 21:45:59 +0200133 /* update per-thread pool cache if necessary */
134 idx = pool_get_index(pool);
135 if (idx >= 0) {
Willy Tarreaued891fd2020-06-01 19:00:28 +0200136 int thr;
137
Christopher Faulet2f6d3c02019-06-25 21:45:59 +0200138 for (thr = 0; thr < MAX_THREADS; thr++)
139 pool_cache[thr][idx].size = size;
140 }
Willy Tarreaued891fd2020-06-01 19:00:28 +0200141#endif
Olivier Houchard8af97eb2020-02-01 17:45:32 +0100142 HA_SPIN_INIT(&pool->lock);
Olivier Houchard8af97eb2020-02-01 17:45:32 +0100143 }
144 pool->users++;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200145 return pool;
146}
Olivier Houchardcf975d42018-01-24 18:38:31 +0100147
Willy Tarreaued891fd2020-06-01 19:00:28 +0200148#ifdef CONFIG_HAP_LOCAL_POOLS
149/* Evicts some of the oldest objects from the local cache, pushing them to the
150 * global pool.
151 */
152void pool_evict_from_cache()
153{
154 struct pool_cache_item *item;
155 struct pool_cache_head *ph;
156
157 do {
Willy Tarreau20dc3cd2020-06-28 00:54:27 +0200158 item = LIST_PREV(&ti->pool_lru_head, struct pool_cache_item *, by_lru);
Willy Tarreaued891fd2020-06-01 19:00:28 +0200159 /* note: by definition we remove oldest objects so they also are the
160 * oldest in their own pools, thus their next is the pool's head.
161 */
162 ph = LIST_NEXT(&item->by_pool, struct pool_cache_head *, list);
163 LIST_DEL(&item->by_pool);
164 LIST_DEL(&item->by_lru);
165 ph->count--;
166 pool_cache_count--;
167 pool_cache_bytes -= ph->size;
168 __pool_free(pool_base_start + (ph - pool_cache[tid]), item);
169 } while (pool_cache_bytes > CONFIG_HAP_POOL_CACHE_SIZE * 7 / 8);
170}
171#endif
172
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100173#ifdef CONFIG_HAP_LOCKLESS_POOLS
Olivier Houchardcf975d42018-01-24 18:38:31 +0100174/* Allocates new entries for pool <pool> until there are at least <avail> + 1
175 * available, then returns the last one for immediate use, so that at least
176 * <avail> are left available in the pool upon return. NULL is returned if the
177 * last entry could not be allocated. It's important to note that at least one
178 * allocation is always performed even if there are enough entries in the pool.
179 * A call to the garbage collector is performed at most once in case malloc()
180 * returns an error, before returning NULL.
181 */
182void *__pool_refill_alloc(struct pool_head *pool, unsigned int avail)
183{
Olivier Houchard8b2c8a72018-10-21 01:52:59 +0200184 void *ptr = NULL, **free_list;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100185 int failed = 0;
186 int size = pool->size;
187 int limit = pool->limit;
188 int allocated = pool->allocated, allocated_orig = allocated;
189
190 /* stop point */
191 avail += pool->used;
192
193 while (1) {
194 if (limit && allocated >= limit) {
Olivier Houchard20872762019-03-08 18:53:35 +0100195 _HA_ATOMIC_ADD(&pool->allocated, allocated - allocated_orig);
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200196 activity[tid].pool_fail++;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100197 return NULL;
198 }
199
Willy Tarreau606135a2020-06-01 12:35:03 +0200200 swrate_add_scaled(&pool->needed_avg, POOL_AVG_SAMPLES, pool->allocated, POOL_AVG_SAMPLES/4);
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200201
Willy Tarreau24aa1ee2020-05-30 18:56:17 +0200202 ptr = pool_alloc_area(size + POOL_EXTRA);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100203 if (!ptr) {
Olivier Houchard20872762019-03-08 18:53:35 +0100204 _HA_ATOMIC_ADD(&pool->failed, 1);
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200205 if (failed) {
206 activity[tid].pool_fail++;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100207 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200208 }
Olivier Houchardcf975d42018-01-24 18:38:31 +0100209 failed++;
210 pool_gc(pool);
211 continue;
212 }
213 if (++allocated > avail)
214 break;
215
216 free_list = pool->free_list;
217 do {
218 *POOL_LINK(pool, ptr) = free_list;
219 __ha_barrier_store();
Olivier Houchard20872762019-03-08 18:53:35 +0100220 } while (_HA_ATOMIC_CAS(&pool->free_list, &free_list, ptr) == 0);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100221 }
Olivier Houchard20872762019-03-08 18:53:35 +0100222 __ha_barrier_atomic_store();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100223
Olivier Houchard20872762019-03-08 18:53:35 +0100224 _HA_ATOMIC_ADD(&pool->allocated, allocated - allocated_orig);
225 _HA_ATOMIC_ADD(&pool->used, 1);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100226
227#ifdef DEBUG_MEMORY_POOLS
228 /* keep track of where the element was allocated from */
229 *POOL_LINK(pool, ptr) = (void *)pool;
230#endif
231 return ptr;
232}
233void *pool_refill_alloc(struct pool_head *pool, unsigned int avail)
234{
235 void *ptr;
236
237 ptr = __pool_refill_alloc(pool, avail);
238 return ptr;
239}
240/*
241 * This function frees whatever can be freed in pool <pool>.
242 */
243void pool_flush(struct pool_head *pool)
244{
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100245 struct pool_free_list cmp, new;
Olivier Houchard8b2c8a72018-10-21 01:52:59 +0200246 void **next, *temp;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100247 int removed = 0;
248
249 if (!pool)
250 return;
Willy Tarreau21072b92020-05-29 17:23:05 +0200251 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100252 do {
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100253 cmp.free_list = pool->free_list;
254 cmp.seq = pool->seq;
255 new.free_list = NULL;
256 new.seq = cmp.seq + 1;
257 } while (!_HA_ATOMIC_DWCAS(&pool->free_list, &cmp, &new));
Olivier Houchard20872762019-03-08 18:53:35 +0100258 __ha_barrier_atomic_store();
Willy Tarreau21072b92020-05-29 17:23:05 +0200259 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100260 next = cmp.free_list;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100261 while (next) {
262 temp = next;
263 next = *POOL_LINK(pool, temp);
264 removed++;
Willy Tarreau24aa1ee2020-05-30 18:56:17 +0200265 pool_free_area(temp, pool->size + POOL_EXTRA);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100266 }
267 pool->free_list = next;
Olivier Houchard20872762019-03-08 18:53:35 +0100268 _HA_ATOMIC_SUB(&pool->allocated, removed);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100269 /* here, we should have pool->allocate == pool->used */
270}
271
272/*
273 * This function frees whatever can be freed in all pools, but respecting
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200274 * the minimum thresholds imposed by owners. It makes sure to be alone to
275 * run by using thread_isolate(). <pool_ctx> is unused.
Olivier Houchardcf975d42018-01-24 18:38:31 +0100276 */
277void pool_gc(struct pool_head *pool_ctx)
278{
Olivier Houchardcf975d42018-01-24 18:38:31 +0100279 struct pool_head *entry;
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200280 int isolated = thread_isolated();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100281
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200282 if (!isolated)
283 thread_isolate();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100284
285 list_for_each_entry(entry, &pools, list) {
286 while ((int)((volatile int)entry->allocated - (volatile int)entry->used) > (int)entry->minavail) {
287 struct pool_free_list cmp, new;
288
289 cmp.seq = entry->seq;
290 __ha_barrier_load();
291 cmp.free_list = entry->free_list;
292 __ha_barrier_load();
293 if (cmp.free_list == NULL)
294 break;
295 new.free_list = *POOL_LINK(entry, cmp.free_list);
296 new.seq = cmp.seq + 1;
Willy Tarreau6a38b322019-05-11 18:04:24 +0200297 if (HA_ATOMIC_DWCAS(&entry->free_list, &cmp, &new) == 0)
Olivier Houchardcf975d42018-01-24 18:38:31 +0100298 continue;
Willy Tarreau24aa1ee2020-05-30 18:56:17 +0200299 pool_free_area(cmp.free_list, entry->size + POOL_EXTRA);
Olivier Houchard20872762019-03-08 18:53:35 +0100300 _HA_ATOMIC_SUB(&entry->allocated, 1);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100301 }
302 }
303
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200304 if (!isolated)
305 thread_release();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100306}
Willy Tarreaue18db9e2018-10-16 10:28:54 +0200307
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100308#else /* CONFIG_HAP_LOCKLESS_POOLS */
Willy Tarreau50e608d2007-05-13 18:26:08 +0200309
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100310/* Allocates new entries for pool <pool> until there are at least <avail> + 1
311 * available, then returns the last one for immediate use, so that at least
312 * <avail> are left available in the pool upon return. NULL is returned if the
313 * last entry could not be allocated. It's important to note that at least one
314 * allocation is always performed even if there are enough entries in the pool.
315 * A call to the garbage collector is performed at most once in case malloc()
316 * returns an error, before returning NULL.
Willy Tarreau50e608d2007-05-13 18:26:08 +0200317 */
Christopher Fauletb349e482017-08-29 09:52:38 +0200318void *__pool_refill_alloc(struct pool_head *pool, unsigned int avail)
Willy Tarreau50e608d2007-05-13 18:26:08 +0200319{
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100320 void *ptr = NULL;
321 int failed = 0;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200322
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100323#ifdef DEBUG_FAIL_ALLOC
324 if (mem_should_fail(pool))
325 return NULL;
326#endif
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100327 /* stop point */
328 avail += pool->used;
329
330 while (1) {
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200331 if (pool->limit && pool->allocated >= pool->limit) {
332 activity[tid].pool_fail++;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200333 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200334 }
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100335
Willy Tarreau606135a2020-06-01 12:35:03 +0200336 swrate_add_scaled(&pool->needed_avg, POOL_AVG_SAMPLES, pool->allocated, POOL_AVG_SAMPLES/4);
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200337 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Willy Tarreauf13322e2017-11-22 10:50:54 +0100338 ptr = pool_alloc_area(pool->size + POOL_EXTRA);
Willy Tarreau82867542019-07-04 11:48:16 +0200339#ifdef DEBUG_MEMORY_POOLS
340 /* keep track of where the element was allocated from. This
341 * is done out of the lock so that the system really allocates
342 * the data without harming other threads waiting on the lock.
343 */
344 if (ptr)
345 *POOL_LINK(pool, ptr) = (void *)pool;
346#endif
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200347 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100348 if (!ptr) {
Willy Tarreau58102cf2015-10-28 16:24:21 +0100349 pool->failed++;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200350 if (failed) {
351 activity[tid].pool_fail++;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100352 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200353 }
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100354 failed++;
Willy Tarreaubafbe012017-11-24 17:34:44 +0100355 pool_gc(pool);
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100356 continue;
357 }
358 if (++pool->allocated > avail)
359 break;
360
Willy Tarreauac421112015-10-28 15:09:29 +0100361 *POOL_LINK(pool, ptr) = (void *)pool->free_list;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100362 pool->free_list = ptr;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200363 }
Willy Tarreau50e608d2007-05-13 18:26:08 +0200364 pool->used++;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100365 return ptr;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200366}
Christopher Fauletb349e482017-08-29 09:52:38 +0200367void *pool_refill_alloc(struct pool_head *pool, unsigned int avail)
368{
369 void *ptr;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200370
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100371 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Christopher Fauletb349e482017-08-29 09:52:38 +0200372 ptr = __pool_refill_alloc(pool, avail);
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100373 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Christopher Fauletb349e482017-08-29 09:52:38 +0200374 return ptr;
375}
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200376/*
377 * This function frees whatever can be freed in pool <pool>.
378 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100379void pool_flush(struct pool_head *pool)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200380{
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200381 void *temp;
382
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200383 if (!pool)
384 return;
385
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200386 while (1) {
387 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
388 temp = pool->free_list;
389 if (!temp) {
390 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
391 break;
392 }
393 pool->free_list = *POOL_LINK(pool, temp);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200394 pool->allocated--;
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200395 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Willy Tarreauf13322e2017-11-22 10:50:54 +0100396 pool_free_area(temp, pool->size + POOL_EXTRA);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200397 }
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200398 /* here, we should have pool->allocated == pool->used */
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200399}
400
401/*
402 * This function frees whatever can be freed in all pools, but respecting
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200403 * the minimum thresholds imposed by owners. It makes sure to be alone to
404 * run by using thread_isolate(). <pool_ctx> is unused.
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200405 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100406void pool_gc(struct pool_head *pool_ctx)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200407{
408 struct pool_head *entry;
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200409 int isolated = thread_isolated();
Willy Tarreaub7f9d122009-04-21 02:17:45 +0200410
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200411 if (!isolated)
412 thread_isolate();
Willy Tarreaub7f9d122009-04-21 02:17:45 +0200413
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200414 list_for_each_entry(entry, &pools, list) {
Olivier Houchard51d93392020-03-12 19:05:39 +0100415 void *temp;
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200416 //qfprintf(stderr, "Flushing pool %s\n", entry->name);
Olivier Houchard51d93392020-03-12 19:05:39 +0100417 while (entry->free_list &&
Willy Tarreau57767b82014-12-22 21:40:55 +0100418 (int)(entry->allocated - entry->used) > (int)entry->minavail) {
Olivier Houchard51d93392020-03-12 19:05:39 +0100419 temp = entry->free_list;
420 entry->free_list = *POOL_LINK(entry, temp);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200421 entry->allocated--;
Willy Tarreauf13322e2017-11-22 10:50:54 +0100422 pool_free_area(temp, entry->size + POOL_EXTRA);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200423 }
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200424 }
Christopher Fauletb349e482017-08-29 09:52:38 +0200425
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200426 if (!isolated)
427 thread_release();
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200428}
Olivier Houchardcf975d42018-01-24 18:38:31 +0100429#endif
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200430
431/*
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200432 * This function destroys a pool by freeing it completely, unless it's still
433 * in use. This should be called only under extreme circumstances. It always
434 * returns NULL if the resulting pool is empty, easing the clearing of the old
435 * pointer, otherwise it returns the pool.
436 * .
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200437 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100438void *pool_destroy(struct pool_head *pool)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200439{
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200440 if (pool) {
Willy Tarreaubafbe012017-11-24 17:34:44 +0100441 pool_flush(pool);
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200442 if (pool->used)
443 return pool;
444 pool->users--;
445 if (!pool->users) {
446 LIST_DEL(&pool->list);
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100447#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100448 HA_SPIN_DESTROY(&pool->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100449#endif
Willy Tarreaued891fd2020-06-01 19:00:28 +0200450
451#ifdef CONFIG_HAP_LOCAL_POOLS
Willy Tarreau0a93b642018-10-16 07:58:39 +0200452 if ((pool - pool_base_start) < MAX_BASE_POOLS)
453 memset(pool, 0, sizeof(*pool));
454 else
Willy Tarreaued891fd2020-06-01 19:00:28 +0200455#endif
Willy Tarreau0a93b642018-10-16 07:58:39 +0200456 free(pool);
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200457 }
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200458 }
459 return NULL;
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200460}
461
Willy Tarreau2455ceb2018-11-26 15:57:34 +0100462/* This destroys all pools on exit. It is *not* thread safe. */
463void pool_destroy_all()
464{
465 struct pool_head *entry, *back;
466
467 list_for_each_entry_safe(entry, back, &pools, list)
468 pool_destroy(entry);
469}
470
Willy Tarreau12833bb2014-01-28 16:49:56 +0100471/* This function dumps memory usage information into the trash buffer. */
472void dump_pools_to_trash()
Willy Tarreau50e608d2007-05-13 18:26:08 +0200473{
474 struct pool_head *entry;
475 unsigned long allocated, used;
476 int nbpools;
477
478 allocated = used = nbpools = 0;
Willy Tarreau12833bb2014-01-28 16:49:56 +0100479 chunk_printf(&trash, "Dumping pools usage. Use SIGQUIT to flush them.\n");
Willy Tarreau50e608d2007-05-13 18:26:08 +0200480 list_for_each_entry(entry, &pools, list) {
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100481#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100482 HA_SPIN_LOCK(POOL_LOCK, &entry->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100483#endif
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200484 chunk_appendf(&trash, " - Pool %s (%u bytes) : %u allocated (%u bytes), %u used, needed_avg %u, %u failures, %u users, @%p=%02d%s\n",
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200485 entry->name, entry->size, entry->allocated,
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200486 entry->size * entry->allocated, entry->used,
Willy Tarreau606135a2020-06-01 12:35:03 +0200487 swrate_avg(entry->needed_avg, POOL_AVG_SAMPLES), entry->failed,
Willy Tarreau0a93b642018-10-16 07:58:39 +0200488 entry->users, entry, (int)pool_get_index(entry),
489 (entry->flags & MEM_F_SHARED) ? " [SHARED]" : "");
Willy Tarreau50e608d2007-05-13 18:26:08 +0200490
491 allocated += entry->allocated * entry->size;
492 used += entry->used * entry->size;
493 nbpools++;
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100494#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100495 HA_SPIN_UNLOCK(POOL_LOCK, &entry->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100496#endif
Willy Tarreau50e608d2007-05-13 18:26:08 +0200497 }
Willy Tarreau12833bb2014-01-28 16:49:56 +0100498 chunk_appendf(&trash, "Total: %d pools, %lu bytes allocated, %lu used.\n",
Willy Tarreau50e608d2007-05-13 18:26:08 +0200499 nbpools, allocated, used);
500}
501
Willy Tarreau12833bb2014-01-28 16:49:56 +0100502/* Dump statistics on pools usage. */
503void dump_pools(void)
504{
505 dump_pools_to_trash();
Willy Tarreau843b7cb2018-07-13 10:54:26 +0200506 qfprintf(stderr, "%s", trash.area);
Willy Tarreau12833bb2014-01-28 16:49:56 +0100507}
508
Willy Tarreau58102cf2015-10-28 16:24:21 +0100509/* This function returns the total number of failed pool allocations */
510int pool_total_failures()
511{
512 struct pool_head *entry;
513 int failed = 0;
514
515 list_for_each_entry(entry, &pools, list)
516 failed += entry->failed;
517 return failed;
518}
519
520/* This function returns the total amount of memory allocated in pools (in bytes) */
521unsigned long pool_total_allocated()
522{
523 struct pool_head *entry;
524 unsigned long allocated = 0;
525
526 list_for_each_entry(entry, &pools, list)
527 allocated += entry->allocated * entry->size;
528 return allocated;
529}
530
531/* This function returns the total amount of memory used in pools (in bytes) */
532unsigned long pool_total_used()
533{
534 struct pool_head *entry;
535 unsigned long used = 0;
536
537 list_for_each_entry(entry, &pools, list)
538 used += entry->used * entry->size;
539 return used;
540}
541
William Lallemande7ed8852016-11-19 02:25:36 +0100542/* This function dumps memory usage information onto the stream interface's
543 * read buffer. It returns 0 as long as it does not complete, non-zero upon
544 * completion. No state is used.
545 */
546static int cli_io_handler_dump_pools(struct appctx *appctx)
547{
548 struct stream_interface *si = appctx->owner;
549
550 dump_pools_to_trash();
Willy Tarreau06d80a92017-10-19 14:32:15 +0200551 if (ci_putchk(si_ic(si), &trash) == -1) {
Willy Tarreaudb398432018-11-15 11:08:52 +0100552 si_rx_room_blk(si);
William Lallemande7ed8852016-11-19 02:25:36 +0100553 return 0;
554 }
555 return 1;
556}
557
Willy Tarreau7107c8b2018-11-26 11:44:35 +0100558/* callback used to create early pool <name> of size <size> and store the
559 * resulting pointer into <ptr>. If the allocation fails, it quits with after
560 * emitting an error message.
561 */
562void create_pool_callback(struct pool_head **ptr, char *name, unsigned int size)
563{
564 *ptr = create_pool(name, size, MEM_F_SHARED);
565 if (!*ptr) {
566 ha_alert("Failed to allocate pool '%s' of size %u : %s. Aborting.\n",
567 name, size, strerror(errno));
568 exit(1);
569 }
570}
571
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100572/* Initializes all per-thread arrays on startup */
573static void init_pools()
574{
Willy Tarreaued891fd2020-06-01 19:00:28 +0200575#ifdef CONFIG_HAP_LOCAL_POOLS
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100576 int thr, idx;
577
578 for (thr = 0; thr < MAX_THREADS; thr++) {
579 for (idx = 0; idx < MAX_BASE_POOLS; idx++) {
580 LIST_INIT(&pool_cache[thr][idx].list);
581 pool_cache[thr][idx].size = 0;
582 }
Willy Tarreau20dc3cd2020-06-28 00:54:27 +0200583 LIST_INIT(&ha_thread_info[thr].pool_lru_head);
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100584 }
Willy Tarreaued891fd2020-06-01 19:00:28 +0200585#endif
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100586}
587
588INITCALL0(STG_PREPARE, init_pools);
Willy Tarreau7107c8b2018-11-26 11:44:35 +0100589
William Lallemande7ed8852016-11-19 02:25:36 +0100590/* register cli keywords */
591static struct cli_kw_list cli_kws = {{ },{
Willy Tarreaue9ecec82016-12-16 18:55:23 +0100592 { { "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 +0100593 {{},}
594}};
595
Willy Tarreau0108d902018-11-25 19:14:37 +0100596INITCALL1(STG_REGISTER, cli_register_kw, &cli_kws);
William Lallemande7ed8852016-11-19 02:25:36 +0100597
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100598#ifdef DEBUG_FAIL_ALLOC
599#define MEM_FAIL_MAX_CHAR 32
600#define MEM_FAIL_MAX_STR 128
601static int mem_fail_cur_idx;
602static char mem_fail_str[MEM_FAIL_MAX_CHAR * MEM_FAIL_MAX_STR];
Willy Tarreauaf613e82020-06-05 08:40:51 +0200603__decl_thread(static HA_SPINLOCK_T mem_fail_lock);
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100604
605int mem_should_fail(const struct pool_head *pool)
606{
Olivier Houchard9c4f08a2019-02-01 16:28:04 +0100607 int ret = 0;
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100608 int n;
609
610 if (mem_fail_rate > 0 && !(global.mode & MODE_STARTING)) {
Willy Tarreau52bf8392020-03-08 00:42:37 +0100611 int randnb = ha_random() % 100;
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100612
613 if (mem_fail_rate > randnb)
614 ret = 1;
615 else
616 ret = 0;
617 }
Olivier Houchard04f5fe82020-02-01 17:49:31 +0100618 HA_SPIN_LOCK(POOL_LOCK, &mem_fail_lock);
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100619 n = snprintf(&mem_fail_str[mem_fail_cur_idx * MEM_FAIL_MAX_CHAR],
620 MEM_FAIL_MAX_CHAR - 2,
621 "%d %.18s %d %d", mem_fail_cur_idx, pool->name, ret, tid);
622 while (n < MEM_FAIL_MAX_CHAR - 1)
623 mem_fail_str[mem_fail_cur_idx * MEM_FAIL_MAX_CHAR + n++] = ' ';
624 if (mem_fail_cur_idx < MEM_FAIL_MAX_STR - 1)
625 mem_fail_str[mem_fail_cur_idx * MEM_FAIL_MAX_CHAR + n] = '\n';
626 else
627 mem_fail_str[mem_fail_cur_idx * MEM_FAIL_MAX_CHAR + n] = 0;
628 mem_fail_cur_idx++;
629 if (mem_fail_cur_idx == MEM_FAIL_MAX_STR)
630 mem_fail_cur_idx = 0;
Olivier Houchard04f5fe82020-02-01 17:49:31 +0100631 HA_SPIN_UNLOCK(POOL_LOCK, &mem_fail_lock);
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100632 return ret;
633
634}
635
636/* config parser for global "tune.fail-alloc" */
637static int mem_parse_global_fail_alloc(char **args, int section_type, struct proxy *curpx,
638 struct proxy *defpx, const char *file, int line,
639 char **err)
640{
641 if (too_many_args(1, args, err, NULL))
642 return -1;
643 mem_fail_rate = atoi(args[1]);
644 if (mem_fail_rate < 0 || mem_fail_rate > 100) {
645 memprintf(err, "'%s' expects a numeric value between 0 and 100.", args[0]);
646 return -1;
647 }
648 return 0;
649}
650#endif
651
652/* register global config keywords */
653static struct cfg_kw_list mem_cfg_kws = {ILH, {
654#ifdef DEBUG_FAIL_ALLOC
655 { CFG_GLOBAL, "tune.fail-alloc", mem_parse_global_fail_alloc },
656#endif
657 { 0, NULL, NULL }
658}};
659
660INITCALL1(STG_REGISTER, cfg_register_keywords, &mem_cfg_kws);
661
Willy Tarreau50e608d2007-05-13 18:26:08 +0200662/*
663 * Local variables:
664 * c-indent-level: 8
665 * c-basic-offset: 8
666 * End:
667 */