blob: 3a4b34cf5c620104790f14af6cce90dee98d44b4 [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 Tarreau4c7e4b72020-05-27 12:58:42 +020014#include <haproxy/api.h>
William Lallemande7ed8852016-11-19 02:25:36 +010015#include <types/applet.h>
16#include <types/cli.h>
Willy Tarreau12833bb2014-01-28 16:49:56 +010017#include <types/global.h>
William Lallemande7ed8852016-11-19 02:25:36 +010018#include <types/stats.h>
19
Olivier Houcharddc21ff72019-01-29 15:20:16 +010020#include <common/cfgparse.h>
Willy Tarreau3f567e42020-05-28 15:29:19 +020021#include <haproxy/thread.h>
Willy Tarreaud0ef4392020-06-02 09:38:52 +020022#include <haproxy/pool.h>
Willy Tarreau853b2972020-05-27 18:01:47 +020023#include <haproxy/list.h>
Willy Tarreau50e608d2007-05-13 18:26:08 +020024#include <common/standard.h>
25
Willy Tarreaua8b2ce02019-05-28 17:04:16 +020026#include <types/activity.h>
27
William Lallemande7ed8852016-11-19 02:25:36 +010028#include <proto/applet.h>
29#include <proto/cli.h>
30#include <proto/channel.h>
Willy Tarreau50e608d2007-05-13 18:26:08 +020031#include <proto/log.h>
William Lallemande7ed8852016-11-19 02:25:36 +010032#include <proto/stream_interface.h>
33#include <proto/stats.h>
Willy Tarreau50e608d2007-05-13 18:26:08 +020034
Willy Tarreaued891fd2020-06-01 19:00:28 +020035#ifdef CONFIG_HAP_LOCAL_POOLS
Willy Tarreau0a93b642018-10-16 07:58:39 +020036/* These are the most common pools, expected to be initialized first. These
37 * ones are allocated from an array, allowing to map them to an index.
38 */
39struct pool_head pool_base_start[MAX_BASE_POOLS] = { };
40unsigned int pool_base_count = 0;
41
Willy Tarreau7f0165e2018-11-26 17:09:46 +010042/* These ones are initialized per-thread on startup by init_pools() */
43struct pool_cache_head pool_cache[MAX_THREADS][MAX_BASE_POOLS];
Willy Tarreaufb117e62020-06-01 18:16:57 +020044struct list pool_lru_head[MAX_THREADS]; /* oldest objects */
Willy Tarreaue18db9e2018-10-16 10:28:54 +020045THREAD_LOCAL size_t pool_cache_bytes = 0; /* total cache size */
46THREAD_LOCAL size_t pool_cache_count = 0; /* #cache objects */
Willy Tarreaued891fd2020-06-01 19:00:28 +020047#endif
Willy Tarreaue18db9e2018-10-16 10:28:54 +020048
Willy Tarreau50e608d2007-05-13 18:26:08 +020049static struct list pools = LIST_HEAD_INIT(pools);
Willy Tarreau067ac9f2015-10-08 14:12:13 +020050int mem_poison_byte = -1;
Willy Tarreau50e608d2007-05-13 18:26:08 +020051
Olivier Houcharddc21ff72019-01-29 15:20:16 +010052#ifdef DEBUG_FAIL_ALLOC
53static int mem_fail_rate = 0;
54static int mem_should_fail(const struct pool_head *);
55#endif
56
Willy Tarreau50e608d2007-05-13 18:26:08 +020057/* Try to find an existing shared pool with the same characteristics and
58 * returns it, otherwise creates this one. NULL is returned if no memory
Willy Tarreau581bf812016-01-25 02:19:13 +010059 * is available for a new creation. Two flags are supported :
60 * - MEM_F_SHARED to indicate that the pool may be shared with other users
61 * - MEM_F_EXACT to indicate that the size must not be rounded up
Willy Tarreau50e608d2007-05-13 18:26:08 +020062 */
63struct pool_head *create_pool(char *name, unsigned int size, unsigned int flags)
64{
65 struct pool_head *pool;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +020066 struct pool_head *entry;
67 struct list *start;
Willy Tarreau50e608d2007-05-13 18:26:08 +020068 unsigned int align;
Willy Tarreaued891fd2020-06-01 19:00:28 +020069 int idx __maybe_unused;
Willy Tarreau50e608d2007-05-13 18:26:08 +020070
Willy Tarreauac421112015-10-28 15:09:29 +010071 /* We need to store a (void *) at the end of the chunks. Since we know
Willy Tarreau50e608d2007-05-13 18:26:08 +020072 * that the malloc() function will never return such a small size,
73 * let's round the size up to something slightly bigger, in order to
74 * ease merging of entries. Note that the rounding is a power of two.
Willy Tarreauac421112015-10-28 15:09:29 +010075 * This extra (void *) is not accounted for in the size computation
76 * so that the visible parts outside are not affected.
Willy Tarreau30f931e2018-10-23 14:40:23 +020077 *
78 * Note: for the LRU cache, we need to store 2 doubly-linked lists.
Willy Tarreau50e608d2007-05-13 18:26:08 +020079 */
80
Willy Tarreau581bf812016-01-25 02:19:13 +010081 if (!(flags & MEM_F_EXACT)) {
Willy Tarreau30f931e2018-10-23 14:40:23 +020082 align = 4 * sizeof(void *); // 2 lists = 4 pointers min
Willy Tarreau581bf812016-01-25 02:19:13 +010083 size = ((size + POOL_EXTRA + align - 1) & -align) - POOL_EXTRA;
84 }
Willy Tarreau50e608d2007-05-13 18:26:08 +020085
Christopher Fauletb349e482017-08-29 09:52:38 +020086 /* TODO: thread: we do not lock pool list for now because all pools are
87 * created during HAProxy startup (so before threads creation) */
Willy Tarreau7dcd46d2007-05-14 00:16:13 +020088 start = &pools;
Willy Tarreau50e608d2007-05-13 18:26:08 +020089 pool = NULL;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +020090
91 list_for_each_entry(entry, &pools, list) {
92 if (entry->size == size) {
93 /* either we can share this place and we take it, or
94 * we look for a sharable one or for the next position
95 * before which we will insert a new one.
96 */
97 if (flags & entry->flags & MEM_F_SHARED) {
98 /* we can share this one */
Willy Tarreau50e608d2007-05-13 18:26:08 +020099 pool = entry;
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +0200100 DPRINTF(stderr, "Sharing %s with %s\n", name, pool->name);
Willy Tarreau50e608d2007-05-13 18:26:08 +0200101 break;
102 }
103 }
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200104 else if (entry->size > size) {
105 /* insert before this one */
106 start = &entry->list;
107 break;
108 }
Willy Tarreau50e608d2007-05-13 18:26:08 +0200109 }
110
111 if (!pool) {
Willy Tarreaued891fd2020-06-01 19:00:28 +0200112#ifdef CONFIG_HAP_LOCAL_POOLS
Willy Tarreau0a93b642018-10-16 07:58:39 +0200113 if (pool_base_count < MAX_BASE_POOLS)
114 pool = &pool_base_start[pool_base_count++];
115
116 if (!pool) {
117 /* look for a freed entry */
118 for (entry = pool_base_start; entry != pool_base_start + MAX_BASE_POOLS; entry++) {
119 if (!entry->size) {
120 pool = entry;
121 break;
122 }
123 }
124 }
Willy Tarreaued891fd2020-06-01 19:00:28 +0200125#endif
Willy Tarreau0a93b642018-10-16 07:58:39 +0200126
127 if (!pool)
128 pool = calloc(1, sizeof(*pool));
129
Willy Tarreau50e608d2007-05-13 18:26:08 +0200130 if (!pool)
131 return NULL;
132 if (name)
133 strlcpy2(pool->name, name, sizeof(pool->name));
134 pool->size = size;
135 pool->flags = flags;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200136 LIST_ADDQ(start, &pool->list);
Christopher Faulet2f6d3c02019-06-25 21:45:59 +0200137
Willy Tarreaued891fd2020-06-01 19:00:28 +0200138#ifdef CONFIG_HAP_LOCAL_POOLS
Christopher Faulet2f6d3c02019-06-25 21:45:59 +0200139 /* update per-thread pool cache if necessary */
140 idx = pool_get_index(pool);
141 if (idx >= 0) {
Willy Tarreaued891fd2020-06-01 19:00:28 +0200142 int thr;
143
Christopher Faulet2f6d3c02019-06-25 21:45:59 +0200144 for (thr = 0; thr < MAX_THREADS; thr++)
145 pool_cache[thr][idx].size = size;
146 }
Willy Tarreaued891fd2020-06-01 19:00:28 +0200147#endif
Olivier Houchard8af97eb2020-02-01 17:45:32 +0100148 HA_SPIN_INIT(&pool->lock);
Olivier Houchard8af97eb2020-02-01 17:45:32 +0100149 }
150 pool->users++;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200151 return pool;
152}
Olivier Houchardcf975d42018-01-24 18:38:31 +0100153
Willy Tarreaued891fd2020-06-01 19:00:28 +0200154#ifdef CONFIG_HAP_LOCAL_POOLS
155/* Evicts some of the oldest objects from the local cache, pushing them to the
156 * global pool.
157 */
158void pool_evict_from_cache()
159{
160 struct pool_cache_item *item;
161 struct pool_cache_head *ph;
162
163 do {
164 item = LIST_PREV(&pool_lru_head[tid], struct pool_cache_item *, by_lru);
165 /* note: by definition we remove oldest objects so they also are the
166 * oldest in their own pools, thus their next is the pool's head.
167 */
168 ph = LIST_NEXT(&item->by_pool, struct pool_cache_head *, list);
169 LIST_DEL(&item->by_pool);
170 LIST_DEL(&item->by_lru);
171 ph->count--;
172 pool_cache_count--;
173 pool_cache_bytes -= ph->size;
174 __pool_free(pool_base_start + (ph - pool_cache[tid]), item);
175 } while (pool_cache_bytes > CONFIG_HAP_POOL_CACHE_SIZE * 7 / 8);
176}
177#endif
178
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100179#ifdef CONFIG_HAP_LOCKLESS_POOLS
Olivier Houchardcf975d42018-01-24 18:38:31 +0100180/* Allocates new entries for pool <pool> until there are at least <avail> + 1
181 * available, then returns the last one for immediate use, so that at least
182 * <avail> are left available in the pool upon return. NULL is returned if the
183 * last entry could not be allocated. It's important to note that at least one
184 * allocation is always performed even if there are enough entries in the pool.
185 * A call to the garbage collector is performed at most once in case malloc()
186 * returns an error, before returning NULL.
187 */
188void *__pool_refill_alloc(struct pool_head *pool, unsigned int avail)
189{
Olivier Houchard8b2c8a72018-10-21 01:52:59 +0200190 void *ptr = NULL, **free_list;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100191 int failed = 0;
192 int size = pool->size;
193 int limit = pool->limit;
194 int allocated = pool->allocated, allocated_orig = allocated;
195
196 /* stop point */
197 avail += pool->used;
198
199 while (1) {
200 if (limit && allocated >= limit) {
Olivier Houchard20872762019-03-08 18:53:35 +0100201 _HA_ATOMIC_ADD(&pool->allocated, allocated - allocated_orig);
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200202 activity[tid].pool_fail++;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100203 return NULL;
204 }
205
Willy Tarreau606135a2020-06-01 12:35:03 +0200206 swrate_add_scaled(&pool->needed_avg, POOL_AVG_SAMPLES, pool->allocated, POOL_AVG_SAMPLES/4);
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200207
Willy Tarreau24aa1ee2020-05-30 18:56:17 +0200208 ptr = pool_alloc_area(size + POOL_EXTRA);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100209 if (!ptr) {
Olivier Houchard20872762019-03-08 18:53:35 +0100210 _HA_ATOMIC_ADD(&pool->failed, 1);
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200211 if (failed) {
212 activity[tid].pool_fail++;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100213 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200214 }
Olivier Houchardcf975d42018-01-24 18:38:31 +0100215 failed++;
216 pool_gc(pool);
217 continue;
218 }
219 if (++allocated > avail)
220 break;
221
222 free_list = pool->free_list;
223 do {
224 *POOL_LINK(pool, ptr) = free_list;
225 __ha_barrier_store();
Olivier Houchard20872762019-03-08 18:53:35 +0100226 } while (_HA_ATOMIC_CAS(&pool->free_list, &free_list, ptr) == 0);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100227 }
Olivier Houchard20872762019-03-08 18:53:35 +0100228 __ha_barrier_atomic_store();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100229
Olivier Houchard20872762019-03-08 18:53:35 +0100230 _HA_ATOMIC_ADD(&pool->allocated, allocated - allocated_orig);
231 _HA_ATOMIC_ADD(&pool->used, 1);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100232
233#ifdef DEBUG_MEMORY_POOLS
234 /* keep track of where the element was allocated from */
235 *POOL_LINK(pool, ptr) = (void *)pool;
236#endif
237 return ptr;
238}
239void *pool_refill_alloc(struct pool_head *pool, unsigned int avail)
240{
241 void *ptr;
242
243 ptr = __pool_refill_alloc(pool, avail);
244 return ptr;
245}
246/*
247 * This function frees whatever can be freed in pool <pool>.
248 */
249void pool_flush(struct pool_head *pool)
250{
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100251 struct pool_free_list cmp, new;
Olivier Houchard8b2c8a72018-10-21 01:52:59 +0200252 void **next, *temp;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100253 int removed = 0;
254
255 if (!pool)
256 return;
Willy Tarreau21072b92020-05-29 17:23:05 +0200257 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100258 do {
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100259 cmp.free_list = pool->free_list;
260 cmp.seq = pool->seq;
261 new.free_list = NULL;
262 new.seq = cmp.seq + 1;
263 } while (!_HA_ATOMIC_DWCAS(&pool->free_list, &cmp, &new));
Olivier Houchard20872762019-03-08 18:53:35 +0100264 __ha_barrier_atomic_store();
Willy Tarreau21072b92020-05-29 17:23:05 +0200265 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100266 next = cmp.free_list;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100267 while (next) {
268 temp = next;
269 next = *POOL_LINK(pool, temp);
270 removed++;
Willy Tarreau24aa1ee2020-05-30 18:56:17 +0200271 pool_free_area(temp, pool->size + POOL_EXTRA);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100272 }
273 pool->free_list = next;
Olivier Houchard20872762019-03-08 18:53:35 +0100274 _HA_ATOMIC_SUB(&pool->allocated, removed);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100275 /* here, we should have pool->allocate == pool->used */
276}
277
278/*
279 * This function frees whatever can be freed in all pools, but respecting
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200280 * the minimum thresholds imposed by owners. It makes sure to be alone to
281 * run by using thread_isolate(). <pool_ctx> is unused.
Olivier Houchardcf975d42018-01-24 18:38:31 +0100282 */
283void pool_gc(struct pool_head *pool_ctx)
284{
Olivier Houchardcf975d42018-01-24 18:38:31 +0100285 struct pool_head *entry;
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200286 int isolated = thread_isolated();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100287
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200288 if (!isolated)
289 thread_isolate();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100290
291 list_for_each_entry(entry, &pools, list) {
292 while ((int)((volatile int)entry->allocated - (volatile int)entry->used) > (int)entry->minavail) {
293 struct pool_free_list cmp, new;
294
295 cmp.seq = entry->seq;
296 __ha_barrier_load();
297 cmp.free_list = entry->free_list;
298 __ha_barrier_load();
299 if (cmp.free_list == NULL)
300 break;
301 new.free_list = *POOL_LINK(entry, cmp.free_list);
302 new.seq = cmp.seq + 1;
Willy Tarreau6a38b322019-05-11 18:04:24 +0200303 if (HA_ATOMIC_DWCAS(&entry->free_list, &cmp, &new) == 0)
Olivier Houchardcf975d42018-01-24 18:38:31 +0100304 continue;
Willy Tarreau24aa1ee2020-05-30 18:56:17 +0200305 pool_free_area(cmp.free_list, entry->size + POOL_EXTRA);
Olivier Houchard20872762019-03-08 18:53:35 +0100306 _HA_ATOMIC_SUB(&entry->allocated, 1);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100307 }
308 }
309
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200310 if (!isolated)
311 thread_release();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100312}
Willy Tarreaue18db9e2018-10-16 10:28:54 +0200313
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100314#else /* CONFIG_HAP_LOCKLESS_POOLS */
Willy Tarreau50e608d2007-05-13 18:26:08 +0200315
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100316/* Allocates new entries for pool <pool> until there are at least <avail> + 1
317 * available, then returns the last one for immediate use, so that at least
318 * <avail> are left available in the pool upon return. NULL is returned if the
319 * last entry could not be allocated. It's important to note that at least one
320 * allocation is always performed even if there are enough entries in the pool.
321 * A call to the garbage collector is performed at most once in case malloc()
322 * returns an error, before returning NULL.
Willy Tarreau50e608d2007-05-13 18:26:08 +0200323 */
Christopher Fauletb349e482017-08-29 09:52:38 +0200324void *__pool_refill_alloc(struct pool_head *pool, unsigned int avail)
Willy Tarreau50e608d2007-05-13 18:26:08 +0200325{
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100326 void *ptr = NULL;
327 int failed = 0;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200328
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100329#ifdef DEBUG_FAIL_ALLOC
330 if (mem_should_fail(pool))
331 return NULL;
332#endif
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100333 /* stop point */
334 avail += pool->used;
335
336 while (1) {
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200337 if (pool->limit && pool->allocated >= pool->limit) {
338 activity[tid].pool_fail++;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200339 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200340 }
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100341
Willy Tarreau606135a2020-06-01 12:35:03 +0200342 swrate_add_scaled(&pool->needed_avg, POOL_AVG_SAMPLES, pool->allocated, POOL_AVG_SAMPLES/4);
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200343 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Willy Tarreauf13322e2017-11-22 10:50:54 +0100344 ptr = pool_alloc_area(pool->size + POOL_EXTRA);
Willy Tarreau82867542019-07-04 11:48:16 +0200345#ifdef DEBUG_MEMORY_POOLS
346 /* keep track of where the element was allocated from. This
347 * is done out of the lock so that the system really allocates
348 * the data without harming other threads waiting on the lock.
349 */
350 if (ptr)
351 *POOL_LINK(pool, ptr) = (void *)pool;
352#endif
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200353 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100354 if (!ptr) {
Willy Tarreau58102cf2015-10-28 16:24:21 +0100355 pool->failed++;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200356 if (failed) {
357 activity[tid].pool_fail++;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100358 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200359 }
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100360 failed++;
Willy Tarreaubafbe012017-11-24 17:34:44 +0100361 pool_gc(pool);
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100362 continue;
363 }
364 if (++pool->allocated > avail)
365 break;
366
Willy Tarreauac421112015-10-28 15:09:29 +0100367 *POOL_LINK(pool, ptr) = (void *)pool->free_list;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100368 pool->free_list = ptr;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200369 }
Willy Tarreau50e608d2007-05-13 18:26:08 +0200370 pool->used++;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100371 return ptr;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200372}
Christopher Fauletb349e482017-08-29 09:52:38 +0200373void *pool_refill_alloc(struct pool_head *pool, unsigned int avail)
374{
375 void *ptr;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200376
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100377 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Christopher Fauletb349e482017-08-29 09:52:38 +0200378 ptr = __pool_refill_alloc(pool, avail);
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100379 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Christopher Fauletb349e482017-08-29 09:52:38 +0200380 return ptr;
381}
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200382/*
383 * This function frees whatever can be freed in pool <pool>.
384 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100385void pool_flush(struct pool_head *pool)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200386{
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200387 void *temp;
388
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200389 if (!pool)
390 return;
391
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200392 while (1) {
393 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
394 temp = pool->free_list;
395 if (!temp) {
396 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
397 break;
398 }
399 pool->free_list = *POOL_LINK(pool, temp);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200400 pool->allocated--;
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200401 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Willy Tarreauf13322e2017-11-22 10:50:54 +0100402 pool_free_area(temp, pool->size + POOL_EXTRA);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200403 }
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200404 /* here, we should have pool->allocated == pool->used */
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200405}
406
407/*
408 * This function frees whatever can be freed in all pools, but respecting
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200409 * the minimum thresholds imposed by owners. It makes sure to be alone to
410 * run by using thread_isolate(). <pool_ctx> is unused.
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200411 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100412void pool_gc(struct pool_head *pool_ctx)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200413{
414 struct pool_head *entry;
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200415 int isolated = thread_isolated();
Willy Tarreaub7f9d122009-04-21 02:17:45 +0200416
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200417 if (!isolated)
418 thread_isolate();
Willy Tarreaub7f9d122009-04-21 02:17:45 +0200419
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200420 list_for_each_entry(entry, &pools, list) {
Olivier Houchard51d93392020-03-12 19:05:39 +0100421 void *temp;
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200422 //qfprintf(stderr, "Flushing pool %s\n", entry->name);
Olivier Houchard51d93392020-03-12 19:05:39 +0100423 while (entry->free_list &&
Willy Tarreau57767b82014-12-22 21:40:55 +0100424 (int)(entry->allocated - entry->used) > (int)entry->minavail) {
Olivier Houchard51d93392020-03-12 19:05:39 +0100425 temp = entry->free_list;
426 entry->free_list = *POOL_LINK(entry, temp);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200427 entry->allocated--;
Willy Tarreauf13322e2017-11-22 10:50:54 +0100428 pool_free_area(temp, entry->size + POOL_EXTRA);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200429 }
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200430 }
Christopher Fauletb349e482017-08-29 09:52:38 +0200431
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200432 if (!isolated)
433 thread_release();
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200434}
Olivier Houchardcf975d42018-01-24 18:38:31 +0100435#endif
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200436
437/*
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200438 * This function destroys a pool by freeing it completely, unless it's still
439 * in use. This should be called only under extreme circumstances. It always
440 * returns NULL if the resulting pool is empty, easing the clearing of the old
441 * pointer, otherwise it returns the pool.
442 * .
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200443 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100444void *pool_destroy(struct pool_head *pool)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200445{
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200446 if (pool) {
Willy Tarreaubafbe012017-11-24 17:34:44 +0100447 pool_flush(pool);
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200448 if (pool->used)
449 return pool;
450 pool->users--;
451 if (!pool->users) {
452 LIST_DEL(&pool->list);
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100453#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100454 HA_SPIN_DESTROY(&pool->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100455#endif
Willy Tarreaued891fd2020-06-01 19:00:28 +0200456
457#ifdef CONFIG_HAP_LOCAL_POOLS
Willy Tarreau0a93b642018-10-16 07:58:39 +0200458 if ((pool - pool_base_start) < MAX_BASE_POOLS)
459 memset(pool, 0, sizeof(*pool));
460 else
Willy Tarreaued891fd2020-06-01 19:00:28 +0200461#endif
Willy Tarreau0a93b642018-10-16 07:58:39 +0200462 free(pool);
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200463 }
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200464 }
465 return NULL;
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200466}
467
Willy Tarreau2455ceb2018-11-26 15:57:34 +0100468/* This destroys all pools on exit. It is *not* thread safe. */
469void pool_destroy_all()
470{
471 struct pool_head *entry, *back;
472
473 list_for_each_entry_safe(entry, back, &pools, list)
474 pool_destroy(entry);
475}
476
Willy Tarreau12833bb2014-01-28 16:49:56 +0100477/* This function dumps memory usage information into the trash buffer. */
478void dump_pools_to_trash()
Willy Tarreau50e608d2007-05-13 18:26:08 +0200479{
480 struct pool_head *entry;
481 unsigned long allocated, used;
482 int nbpools;
483
484 allocated = used = nbpools = 0;
Willy Tarreau12833bb2014-01-28 16:49:56 +0100485 chunk_printf(&trash, "Dumping pools usage. Use SIGQUIT to flush them.\n");
Willy Tarreau50e608d2007-05-13 18:26:08 +0200486 list_for_each_entry(entry, &pools, list) {
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100487#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100488 HA_SPIN_LOCK(POOL_LOCK, &entry->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100489#endif
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200490 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 +0200491 entry->name, entry->size, entry->allocated,
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200492 entry->size * entry->allocated, entry->used,
Willy Tarreau606135a2020-06-01 12:35:03 +0200493 swrate_avg(entry->needed_avg, POOL_AVG_SAMPLES), entry->failed,
Willy Tarreau0a93b642018-10-16 07:58:39 +0200494 entry->users, entry, (int)pool_get_index(entry),
495 (entry->flags & MEM_F_SHARED) ? " [SHARED]" : "");
Willy Tarreau50e608d2007-05-13 18:26:08 +0200496
497 allocated += entry->allocated * entry->size;
498 used += entry->used * entry->size;
499 nbpools++;
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100500#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100501 HA_SPIN_UNLOCK(POOL_LOCK, &entry->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100502#endif
Willy Tarreau50e608d2007-05-13 18:26:08 +0200503 }
Willy Tarreau12833bb2014-01-28 16:49:56 +0100504 chunk_appendf(&trash, "Total: %d pools, %lu bytes allocated, %lu used.\n",
Willy Tarreau50e608d2007-05-13 18:26:08 +0200505 nbpools, allocated, used);
506}
507
Willy Tarreau12833bb2014-01-28 16:49:56 +0100508/* Dump statistics on pools usage. */
509void dump_pools(void)
510{
511 dump_pools_to_trash();
Willy Tarreau843b7cb2018-07-13 10:54:26 +0200512 qfprintf(stderr, "%s", trash.area);
Willy Tarreau12833bb2014-01-28 16:49:56 +0100513}
514
Willy Tarreau58102cf2015-10-28 16:24:21 +0100515/* This function returns the total number of failed pool allocations */
516int pool_total_failures()
517{
518 struct pool_head *entry;
519 int failed = 0;
520
521 list_for_each_entry(entry, &pools, list)
522 failed += entry->failed;
523 return failed;
524}
525
526/* This function returns the total amount of memory allocated in pools (in bytes) */
527unsigned long pool_total_allocated()
528{
529 struct pool_head *entry;
530 unsigned long allocated = 0;
531
532 list_for_each_entry(entry, &pools, list)
533 allocated += entry->allocated * entry->size;
534 return allocated;
535}
536
537/* This function returns the total amount of memory used in pools (in bytes) */
538unsigned long pool_total_used()
539{
540 struct pool_head *entry;
541 unsigned long used = 0;
542
543 list_for_each_entry(entry, &pools, list)
544 used += entry->used * entry->size;
545 return used;
546}
547
William Lallemande7ed8852016-11-19 02:25:36 +0100548/* This function dumps memory usage information onto the stream interface's
549 * read buffer. It returns 0 as long as it does not complete, non-zero upon
550 * completion. No state is used.
551 */
552static int cli_io_handler_dump_pools(struct appctx *appctx)
553{
554 struct stream_interface *si = appctx->owner;
555
556 dump_pools_to_trash();
Willy Tarreau06d80a92017-10-19 14:32:15 +0200557 if (ci_putchk(si_ic(si), &trash) == -1) {
Willy Tarreaudb398432018-11-15 11:08:52 +0100558 si_rx_room_blk(si);
William Lallemande7ed8852016-11-19 02:25:36 +0100559 return 0;
560 }
561 return 1;
562}
563
Willy Tarreau7107c8b2018-11-26 11:44:35 +0100564/* callback used to create early pool <name> of size <size> and store the
565 * resulting pointer into <ptr>. If the allocation fails, it quits with after
566 * emitting an error message.
567 */
568void create_pool_callback(struct pool_head **ptr, char *name, unsigned int size)
569{
570 *ptr = create_pool(name, size, MEM_F_SHARED);
571 if (!*ptr) {
572 ha_alert("Failed to allocate pool '%s' of size %u : %s. Aborting.\n",
573 name, size, strerror(errno));
574 exit(1);
575 }
576}
577
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100578/* Initializes all per-thread arrays on startup */
579static void init_pools()
580{
Willy Tarreaued891fd2020-06-01 19:00:28 +0200581#ifdef CONFIG_HAP_LOCAL_POOLS
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100582 int thr, idx;
583
584 for (thr = 0; thr < MAX_THREADS; thr++) {
585 for (idx = 0; idx < MAX_BASE_POOLS; idx++) {
586 LIST_INIT(&pool_cache[thr][idx].list);
587 pool_cache[thr][idx].size = 0;
588 }
589 LIST_INIT(&pool_lru_head[thr]);
590 }
Willy Tarreaued891fd2020-06-01 19:00:28 +0200591#endif
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100592}
593
594INITCALL0(STG_PREPARE, init_pools);
Willy Tarreau7107c8b2018-11-26 11:44:35 +0100595
William Lallemande7ed8852016-11-19 02:25:36 +0100596/* register cli keywords */
597static struct cli_kw_list cli_kws = {{ },{
Willy Tarreaue9ecec82016-12-16 18:55:23 +0100598 { { "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 +0100599 {{},}
600}};
601
Willy Tarreau0108d902018-11-25 19:14:37 +0100602INITCALL1(STG_REGISTER, cli_register_kw, &cli_kws);
William Lallemande7ed8852016-11-19 02:25:36 +0100603
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100604#ifdef DEBUG_FAIL_ALLOC
605#define MEM_FAIL_MAX_CHAR 32
606#define MEM_FAIL_MAX_STR 128
607static int mem_fail_cur_idx;
608static char mem_fail_str[MEM_FAIL_MAX_CHAR * MEM_FAIL_MAX_STR];
Willy Tarreauaf613e82020-06-05 08:40:51 +0200609__decl_thread(static HA_SPINLOCK_T mem_fail_lock);
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100610
611int mem_should_fail(const struct pool_head *pool)
612{
Olivier Houchard9c4f08a2019-02-01 16:28:04 +0100613 int ret = 0;
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100614 int n;
615
616 if (mem_fail_rate > 0 && !(global.mode & MODE_STARTING)) {
Willy Tarreau52bf8392020-03-08 00:42:37 +0100617 int randnb = ha_random() % 100;
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100618
619 if (mem_fail_rate > randnb)
620 ret = 1;
621 else
622 ret = 0;
623 }
Olivier Houchard04f5fe82020-02-01 17:49:31 +0100624 HA_SPIN_LOCK(POOL_LOCK, &mem_fail_lock);
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100625 n = snprintf(&mem_fail_str[mem_fail_cur_idx * MEM_FAIL_MAX_CHAR],
626 MEM_FAIL_MAX_CHAR - 2,
627 "%d %.18s %d %d", mem_fail_cur_idx, pool->name, ret, tid);
628 while (n < MEM_FAIL_MAX_CHAR - 1)
629 mem_fail_str[mem_fail_cur_idx * MEM_FAIL_MAX_CHAR + n++] = ' ';
630 if (mem_fail_cur_idx < MEM_FAIL_MAX_STR - 1)
631 mem_fail_str[mem_fail_cur_idx * MEM_FAIL_MAX_CHAR + n] = '\n';
632 else
633 mem_fail_str[mem_fail_cur_idx * MEM_FAIL_MAX_CHAR + n] = 0;
634 mem_fail_cur_idx++;
635 if (mem_fail_cur_idx == MEM_FAIL_MAX_STR)
636 mem_fail_cur_idx = 0;
Olivier Houchard04f5fe82020-02-01 17:49:31 +0100637 HA_SPIN_UNLOCK(POOL_LOCK, &mem_fail_lock);
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,
644 struct proxy *defpx, const char *file, int line,
645 char **err)
646{
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 */