blob: 4552e4559b4133ede016cef5282cfb70e1a6f05e [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
William Lallemande7ed8852016-11-19 02:25:36 +010014#include <types/applet.h>
15#include <types/cli.h>
Willy Tarreau12833bb2014-01-28 16:49:56 +010016#include <types/global.h>
William Lallemande7ed8852016-11-19 02:25:36 +010017#include <types/stats.h>
18
Olivier Houcharddc21ff72019-01-29 15:20:16 +010019#include <common/cfgparse.h>
Willy Tarreau50e608d2007-05-13 18:26:08 +020020#include <common/config.h>
Krzysztof Piotr Oledzkia643baf2008-05-29 23:53:44 +020021#include <common/debug.h>
Willy Tarreaue18db9e2018-10-16 10:28:54 +020022#include <common/hathreads.h>
Willy Tarreau0108d902018-11-25 19:14:37 +010023#include <common/initcall.h>
Willy Tarreau50e608d2007-05-13 18:26:08 +020024#include <common/memory.h>
25#include <common/mini-clist.h>
26#include <common/standard.h>
27
Willy Tarreaua8b2ce02019-05-28 17:04:16 +020028#include <types/activity.h>
29
William Lallemande7ed8852016-11-19 02:25:36 +010030#include <proto/applet.h>
31#include <proto/cli.h>
32#include <proto/channel.h>
Willy Tarreau50e608d2007-05-13 18:26:08 +020033#include <proto/log.h>
William Lallemande7ed8852016-11-19 02:25:36 +010034#include <proto/stream_interface.h>
35#include <proto/stats.h>
Willy Tarreau50e608d2007-05-13 18:26:08 +020036
Willy Tarreau0a93b642018-10-16 07:58:39 +020037/* These are the most common pools, expected to be initialized first. These
38 * ones are allocated from an array, allowing to map them to an index.
39 */
40struct pool_head pool_base_start[MAX_BASE_POOLS] = { };
41unsigned int pool_base_count = 0;
42
Willy Tarreau7f0165e2018-11-26 17:09:46 +010043/* These ones are initialized per-thread on startup by init_pools() */
44struct pool_cache_head pool_cache[MAX_THREADS][MAX_BASE_POOLS];
45static struct list pool_lru_head[MAX_THREADS]; /* oldest objects */
Willy Tarreaue18db9e2018-10-16 10:28:54 +020046THREAD_LOCAL size_t pool_cache_bytes = 0; /* total cache size */
47THREAD_LOCAL size_t pool_cache_count = 0; /* #cache objects */
48
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;
Christopher Faulet2f6d3c02019-06-25 21:45:59 +020069 int thr, idx;
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 Tarreau0a93b642018-10-16 07:58:39 +0200112 if (pool_base_count < MAX_BASE_POOLS)
113 pool = &pool_base_start[pool_base_count++];
114
115 if (!pool) {
116 /* look for a freed entry */
117 for (entry = pool_base_start; entry != pool_base_start + MAX_BASE_POOLS; entry++) {
118 if (!entry->size) {
119 pool = entry;
120 break;
121 }
122 }
123 }
124
125 if (!pool)
126 pool = calloc(1, sizeof(*pool));
127
Willy Tarreau50e608d2007-05-13 18:26:08 +0200128 if (!pool)
129 return NULL;
130 if (name)
131 strlcpy2(pool->name, name, sizeof(pool->name));
132 pool->size = size;
133 pool->flags = flags;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200134 LIST_ADDQ(start, &pool->list);
Christopher Faulet2f6d3c02019-06-25 21:45:59 +0200135
136 /* update per-thread pool cache if necessary */
137 idx = pool_get_index(pool);
138 if (idx >= 0) {
139 for (thr = 0; thr < MAX_THREADS; thr++)
140 pool_cache[thr][idx].size = size;
141 }
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100142#ifndef CONFIG_HAP_LOCKLESS_POOLS
Olivier Houchard8af97eb2020-02-01 17:45:32 +0100143 HA_SPIN_INIT(&pool->lock);
Olivier Houchard04f5fe82020-02-01 17:49:31 +0100144#else
Olivier Houchard899fb8a2020-03-18 15:48:29 +0100145 HA_SPIN_INIT(&pool->flush_lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100146#endif
Olivier Houchard8af97eb2020-02-01 17:45:32 +0100147 }
148 pool->users++;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200149 return pool;
150}
Olivier Houchardcf975d42018-01-24 18:38:31 +0100151
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100152#ifdef CONFIG_HAP_LOCKLESS_POOLS
Olivier Houchardcf975d42018-01-24 18:38:31 +0100153/* Allocates new entries for pool <pool> until there are at least <avail> + 1
154 * available, then returns the last one for immediate use, so that at least
155 * <avail> are left available in the pool upon return. NULL is returned if the
156 * last entry could not be allocated. It's important to note that at least one
157 * allocation is always performed even if there are enough entries in the pool.
158 * A call to the garbage collector is performed at most once in case malloc()
159 * returns an error, before returning NULL.
160 */
161void *__pool_refill_alloc(struct pool_head *pool, unsigned int avail)
162{
Olivier Houchard8b2c8a72018-10-21 01:52:59 +0200163 void *ptr = NULL, **free_list;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100164 int failed = 0;
165 int size = pool->size;
166 int limit = pool->limit;
167 int allocated = pool->allocated, allocated_orig = allocated;
168
169 /* stop point */
170 avail += pool->used;
171
172 while (1) {
173 if (limit && allocated >= limit) {
Olivier Houchard20872762019-03-08 18:53:35 +0100174 _HA_ATOMIC_ADD(&pool->allocated, allocated - allocated_orig);
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200175 activity[tid].pool_fail++;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100176 return NULL;
177 }
178
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200179 pool_avg_bump(&pool->needed_avg, pool->allocated);
180
Olivier Houchardcf975d42018-01-24 18:38:31 +0100181 ptr = malloc(size + POOL_EXTRA);
182 if (!ptr) {
Olivier Houchard20872762019-03-08 18:53:35 +0100183 _HA_ATOMIC_ADD(&pool->failed, 1);
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200184 if (failed) {
185 activity[tid].pool_fail++;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100186 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200187 }
Olivier Houchardcf975d42018-01-24 18:38:31 +0100188 failed++;
189 pool_gc(pool);
190 continue;
191 }
192 if (++allocated > avail)
193 break;
194
195 free_list = pool->free_list;
196 do {
197 *POOL_LINK(pool, ptr) = free_list;
198 __ha_barrier_store();
Olivier Houchard20872762019-03-08 18:53:35 +0100199 } while (_HA_ATOMIC_CAS(&pool->free_list, &free_list, ptr) == 0);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100200 }
Olivier Houchard20872762019-03-08 18:53:35 +0100201 __ha_barrier_atomic_store();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100202
Olivier Houchard20872762019-03-08 18:53:35 +0100203 _HA_ATOMIC_ADD(&pool->allocated, allocated - allocated_orig);
204 _HA_ATOMIC_ADD(&pool->used, 1);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100205
206#ifdef DEBUG_MEMORY_POOLS
207 /* keep track of where the element was allocated from */
208 *POOL_LINK(pool, ptr) = (void *)pool;
209#endif
210 return ptr;
211}
212void *pool_refill_alloc(struct pool_head *pool, unsigned int avail)
213{
214 void *ptr;
215
216 ptr = __pool_refill_alloc(pool, avail);
217 return ptr;
218}
219/*
220 * This function frees whatever can be freed in pool <pool>.
221 */
222void pool_flush(struct pool_head *pool)
223{
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100224 struct pool_free_list cmp, new;
Olivier Houchard8b2c8a72018-10-21 01:52:59 +0200225 void **next, *temp;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100226 int removed = 0;
227
228 if (!pool)
229 return;
Olivier Houchard899fb8a2020-03-18 15:48:29 +0100230 HA_SPIN_LOCK(POOL_LOCK, &pool->flush_lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100231 do {
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100232 cmp.free_list = pool->free_list;
233 cmp.seq = pool->seq;
234 new.free_list = NULL;
235 new.seq = cmp.seq + 1;
236 } while (!_HA_ATOMIC_DWCAS(&pool->free_list, &cmp, &new));
Olivier Houchard20872762019-03-08 18:53:35 +0100237 __ha_barrier_atomic_store();
Olivier Houchard899fb8a2020-03-18 15:48:29 +0100238 HA_SPIN_UNLOCK(POOL_LOCK, &pool->flush_lock);
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100239 next = cmp.free_list;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100240 while (next) {
241 temp = next;
242 next = *POOL_LINK(pool, temp);
243 removed++;
244 free(temp);
245 }
246 pool->free_list = next;
Olivier Houchard20872762019-03-08 18:53:35 +0100247 _HA_ATOMIC_SUB(&pool->allocated, removed);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100248 /* here, we should have pool->allocate == pool->used */
249}
250
251/*
252 * This function frees whatever can be freed in all pools, but respecting
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200253 * the minimum thresholds imposed by owners. It makes sure to be alone to
254 * run by using thread_isolate(). <pool_ctx> is unused.
Olivier Houchardcf975d42018-01-24 18:38:31 +0100255 */
256void pool_gc(struct pool_head *pool_ctx)
257{
Olivier Houchardcf975d42018-01-24 18:38:31 +0100258 struct pool_head *entry;
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200259 int isolated = thread_isolated();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100260
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200261 if (!isolated)
262 thread_isolate();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100263
264 list_for_each_entry(entry, &pools, list) {
265 while ((int)((volatile int)entry->allocated - (volatile int)entry->used) > (int)entry->minavail) {
266 struct pool_free_list cmp, new;
267
268 cmp.seq = entry->seq;
269 __ha_barrier_load();
270 cmp.free_list = entry->free_list;
271 __ha_barrier_load();
272 if (cmp.free_list == NULL)
273 break;
274 new.free_list = *POOL_LINK(entry, cmp.free_list);
275 new.seq = cmp.seq + 1;
Willy Tarreau6a38b322019-05-11 18:04:24 +0200276 if (HA_ATOMIC_DWCAS(&entry->free_list, &cmp, &new) == 0)
Olivier Houchardcf975d42018-01-24 18:38:31 +0100277 continue;
278 free(cmp.free_list);
Olivier Houchard20872762019-03-08 18:53:35 +0100279 _HA_ATOMIC_SUB(&entry->allocated, 1);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100280 }
281 }
282
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200283 if (!isolated)
284 thread_release();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100285}
Willy Tarreaue18db9e2018-10-16 10:28:54 +0200286
287/* frees an object to the local cache, possibly pushing oldest objects to the
288 * global pool. Must not be called directly.
289 */
290void __pool_put_to_cache(struct pool_head *pool, void *ptr, ssize_t idx)
291{
292 struct pool_cache_item *item = (struct pool_cache_item *)ptr;
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100293 struct pool_cache_head *ph = &pool_cache[tid][idx];
Willy Tarreaue18db9e2018-10-16 10:28:54 +0200294
295 LIST_ADD(&ph->list, &item->by_pool);
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100296 LIST_ADD(&pool_lru_head[tid], &item->by_lru);
Willy Tarreaue18db9e2018-10-16 10:28:54 +0200297 ph->count++;
298 pool_cache_count++;
299 pool_cache_bytes += ph->size;
300
301 if (pool_cache_bytes <= CONFIG_HAP_POOL_CACHE_SIZE)
302 return;
303
304 do {
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100305 item = LIST_PREV(&pool_lru_head[tid], struct pool_cache_item *, by_lru);
Willy Tarreaue18db9e2018-10-16 10:28:54 +0200306 /* note: by definition we remove oldest objects so they also are the
307 * oldest in their own pools, thus their next is the pool's head.
308 */
309 ph = LIST_NEXT(&item->by_pool, struct pool_cache_head *, list);
310 LIST_DEL(&item->by_pool);
311 LIST_DEL(&item->by_lru);
312 ph->count--;
313 pool_cache_count--;
314 pool_cache_bytes -= ph->size;
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100315 __pool_free(pool_base_start + (ph - pool_cache[tid]), item);
Willy Tarreaue18db9e2018-10-16 10:28:54 +0200316 } while (pool_cache_bytes > CONFIG_HAP_POOL_CACHE_SIZE * 7 / 8);
317}
318
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100319#else /* CONFIG_HAP_LOCKLESS_POOLS */
Willy Tarreau50e608d2007-05-13 18:26:08 +0200320
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100321/* Allocates new entries for pool <pool> until there are at least <avail> + 1
322 * available, then returns the last one for immediate use, so that at least
323 * <avail> are left available in the pool upon return. NULL is returned if the
324 * last entry could not be allocated. It's important to note that at least one
325 * allocation is always performed even if there are enough entries in the pool.
326 * A call to the garbage collector is performed at most once in case malloc()
327 * returns an error, before returning NULL.
Willy Tarreau50e608d2007-05-13 18:26:08 +0200328 */
Christopher Fauletb349e482017-08-29 09:52:38 +0200329void *__pool_refill_alloc(struct pool_head *pool, unsigned int avail)
Willy Tarreau50e608d2007-05-13 18:26:08 +0200330{
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100331 void *ptr = NULL;
332 int failed = 0;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200333
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100334#ifdef DEBUG_FAIL_ALLOC
335 if (mem_should_fail(pool))
336 return NULL;
337#endif
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100338 /* stop point */
339 avail += pool->used;
340
341 while (1) {
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200342 if (pool->limit && pool->allocated >= pool->limit) {
343 activity[tid].pool_fail++;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200344 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200345 }
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100346
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200347 pool_avg_bump(&pool->needed_avg, pool->allocated);
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200348 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Willy Tarreauf13322e2017-11-22 10:50:54 +0100349 ptr = pool_alloc_area(pool->size + POOL_EXTRA);
Willy Tarreau82867542019-07-04 11:48:16 +0200350#ifdef DEBUG_MEMORY_POOLS
351 /* keep track of where the element was allocated from. This
352 * is done out of the lock so that the system really allocates
353 * the data without harming other threads waiting on the lock.
354 */
355 if (ptr)
356 *POOL_LINK(pool, ptr) = (void *)pool;
357#endif
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200358 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100359 if (!ptr) {
Willy Tarreau58102cf2015-10-28 16:24:21 +0100360 pool->failed++;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200361 if (failed) {
362 activity[tid].pool_fail++;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100363 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200364 }
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100365 failed++;
Willy Tarreaubafbe012017-11-24 17:34:44 +0100366 pool_gc(pool);
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100367 continue;
368 }
369 if (++pool->allocated > avail)
370 break;
371
Willy Tarreauac421112015-10-28 15:09:29 +0100372 *POOL_LINK(pool, ptr) = (void *)pool->free_list;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100373 pool->free_list = ptr;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200374 }
Willy Tarreau50e608d2007-05-13 18:26:08 +0200375 pool->used++;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100376 return ptr;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200377}
Christopher Fauletb349e482017-08-29 09:52:38 +0200378void *pool_refill_alloc(struct pool_head *pool, unsigned int avail)
379{
380 void *ptr;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200381
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100382 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Christopher Fauletb349e482017-08-29 09:52:38 +0200383 ptr = __pool_refill_alloc(pool, avail);
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100384 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Christopher Fauletb349e482017-08-29 09:52:38 +0200385 return ptr;
386}
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200387/*
388 * This function frees whatever can be freed in pool <pool>.
389 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100390void pool_flush(struct pool_head *pool)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200391{
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200392 void *temp;
393
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200394 if (!pool)
395 return;
396
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200397 while (1) {
398 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
399 temp = pool->free_list;
400 if (!temp) {
401 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
402 break;
403 }
404 pool->free_list = *POOL_LINK(pool, temp);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200405 pool->allocated--;
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200406 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Willy Tarreauf13322e2017-11-22 10:50:54 +0100407 pool_free_area(temp, pool->size + POOL_EXTRA);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200408 }
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200409 /* here, we should have pool->allocated == pool->used */
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200410}
411
412/*
413 * This function frees whatever can be freed in all pools, but respecting
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200414 * the minimum thresholds imposed by owners. It makes sure to be alone to
415 * run by using thread_isolate(). <pool_ctx> is unused.
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200416 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100417void pool_gc(struct pool_head *pool_ctx)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200418{
419 struct pool_head *entry;
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200420 int isolated = thread_isolated();
Willy Tarreaub7f9d122009-04-21 02:17:45 +0200421
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200422 if (!isolated)
423 thread_isolate();
Willy Tarreaub7f9d122009-04-21 02:17:45 +0200424
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200425 list_for_each_entry(entry, &pools, list) {
Olivier Houchard51d93392020-03-12 19:05:39 +0100426 void *temp;
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200427 //qfprintf(stderr, "Flushing pool %s\n", entry->name);
Olivier Houchard51d93392020-03-12 19:05:39 +0100428 while (entry->free_list &&
Willy Tarreau57767b82014-12-22 21:40:55 +0100429 (int)(entry->allocated - entry->used) > (int)entry->minavail) {
Olivier Houchard51d93392020-03-12 19:05:39 +0100430 temp = entry->free_list;
431 entry->free_list = *POOL_LINK(entry, temp);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200432 entry->allocated--;
Willy Tarreauf13322e2017-11-22 10:50:54 +0100433 pool_free_area(temp, entry->size + POOL_EXTRA);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200434 }
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200435 }
Christopher Fauletb349e482017-08-29 09:52:38 +0200436
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200437 if (!isolated)
438 thread_release();
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200439}
Olivier Houchardcf975d42018-01-24 18:38:31 +0100440#endif
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200441
442/*
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200443 * This function destroys a pool by freeing it completely, unless it's still
444 * in use. This should be called only under extreme circumstances. It always
445 * returns NULL if the resulting pool is empty, easing the clearing of the old
446 * pointer, otherwise it returns the pool.
447 * .
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200448 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100449void *pool_destroy(struct pool_head *pool)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200450{
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200451 if (pool) {
Willy Tarreaubafbe012017-11-24 17:34:44 +0100452 pool_flush(pool);
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200453 if (pool->used)
454 return pool;
455 pool->users--;
456 if (!pool->users) {
457 LIST_DEL(&pool->list);
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100458#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100459 HA_SPIN_DESTROY(&pool->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100460#endif
Willy Tarreau0a93b642018-10-16 07:58:39 +0200461 if ((pool - pool_base_start) < MAX_BASE_POOLS)
462 memset(pool, 0, sizeof(*pool));
463 else
464 free(pool);
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200465 }
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200466 }
467 return NULL;
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200468}
469
Willy Tarreau2455ceb2018-11-26 15:57:34 +0100470/* This destroys all pools on exit. It is *not* thread safe. */
471void pool_destroy_all()
472{
473 struct pool_head *entry, *back;
474
475 list_for_each_entry_safe(entry, back, &pools, list)
476 pool_destroy(entry);
477}
478
Willy Tarreau12833bb2014-01-28 16:49:56 +0100479/* This function dumps memory usage information into the trash buffer. */
480void dump_pools_to_trash()
Willy Tarreau50e608d2007-05-13 18:26:08 +0200481{
482 struct pool_head *entry;
483 unsigned long allocated, used;
484 int nbpools;
485
486 allocated = used = nbpools = 0;
Willy Tarreau12833bb2014-01-28 16:49:56 +0100487 chunk_printf(&trash, "Dumping pools usage. Use SIGQUIT to flush them.\n");
Willy Tarreau50e608d2007-05-13 18:26:08 +0200488 list_for_each_entry(entry, &pools, list) {
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100489#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100490 HA_SPIN_LOCK(POOL_LOCK, &entry->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100491#endif
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200492 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 +0200493 entry->name, entry->size, entry->allocated,
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200494 entry->size * entry->allocated, entry->used,
495 pool_avg(entry->needed_avg), entry->failed,
Willy Tarreau0a93b642018-10-16 07:58:39 +0200496 entry->users, entry, (int)pool_get_index(entry),
497 (entry->flags & MEM_F_SHARED) ? " [SHARED]" : "");
Willy Tarreau50e608d2007-05-13 18:26:08 +0200498
499 allocated += entry->allocated * entry->size;
500 used += entry->used * entry->size;
501 nbpools++;
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100502#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100503 HA_SPIN_UNLOCK(POOL_LOCK, &entry->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100504#endif
Willy Tarreau50e608d2007-05-13 18:26:08 +0200505 }
Willy Tarreau12833bb2014-01-28 16:49:56 +0100506 chunk_appendf(&trash, "Total: %d pools, %lu bytes allocated, %lu used.\n",
Willy Tarreau50e608d2007-05-13 18:26:08 +0200507 nbpools, allocated, used);
508}
509
Willy Tarreau12833bb2014-01-28 16:49:56 +0100510/* Dump statistics on pools usage. */
511void dump_pools(void)
512{
513 dump_pools_to_trash();
Willy Tarreau843b7cb2018-07-13 10:54:26 +0200514 qfprintf(stderr, "%s", trash.area);
Willy Tarreau12833bb2014-01-28 16:49:56 +0100515}
516
Willy Tarreau58102cf2015-10-28 16:24:21 +0100517/* This function returns the total number of failed pool allocations */
518int pool_total_failures()
519{
520 struct pool_head *entry;
521 int failed = 0;
522
523 list_for_each_entry(entry, &pools, list)
524 failed += entry->failed;
525 return failed;
526}
527
528/* This function returns the total amount of memory allocated in pools (in bytes) */
529unsigned long pool_total_allocated()
530{
531 struct pool_head *entry;
532 unsigned long allocated = 0;
533
534 list_for_each_entry(entry, &pools, list)
535 allocated += entry->allocated * entry->size;
536 return allocated;
537}
538
539/* This function returns the total amount of memory used in pools (in bytes) */
540unsigned long pool_total_used()
541{
542 struct pool_head *entry;
543 unsigned long used = 0;
544
545 list_for_each_entry(entry, &pools, list)
546 used += entry->used * entry->size;
547 return used;
548}
549
William Lallemande7ed8852016-11-19 02:25:36 +0100550/* This function dumps memory usage information onto the stream interface's
551 * read buffer. It returns 0 as long as it does not complete, non-zero upon
552 * completion. No state is used.
553 */
554static int cli_io_handler_dump_pools(struct appctx *appctx)
555{
556 struct stream_interface *si = appctx->owner;
557
558 dump_pools_to_trash();
Willy Tarreau06d80a92017-10-19 14:32:15 +0200559 if (ci_putchk(si_ic(si), &trash) == -1) {
Willy Tarreaudb398432018-11-15 11:08:52 +0100560 si_rx_room_blk(si);
William Lallemande7ed8852016-11-19 02:25:36 +0100561 return 0;
562 }
563 return 1;
564}
565
Willy Tarreau7107c8b2018-11-26 11:44:35 +0100566/* callback used to create early pool <name> of size <size> and store the
567 * resulting pointer into <ptr>. If the allocation fails, it quits with after
568 * emitting an error message.
569 */
570void create_pool_callback(struct pool_head **ptr, char *name, unsigned int size)
571{
572 *ptr = create_pool(name, size, MEM_F_SHARED);
573 if (!*ptr) {
574 ha_alert("Failed to allocate pool '%s' of size %u : %s. Aborting.\n",
575 name, size, strerror(errno));
576 exit(1);
577 }
578}
579
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100580/* Initializes all per-thread arrays on startup */
581static void init_pools()
582{
583 int thr, idx;
584
585 for (thr = 0; thr < MAX_THREADS; thr++) {
586 for (idx = 0; idx < MAX_BASE_POOLS; idx++) {
587 LIST_INIT(&pool_cache[thr][idx].list);
588 pool_cache[thr][idx].size = 0;
589 }
590 LIST_INIT(&pool_lru_head[thr]);
591 }
592}
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];
609__decl_hathreads(static HA_SPINLOCK_T mem_fail_lock);
610
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 */