blob: b442f0a108594b255804d57239770b1af3833c55 [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 }
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 Tarreauf161d0f2018-02-22 14:05:55 +0100148#ifdef CONFIG_HAP_LOCKLESS_POOLS
Olivier Houchardcf975d42018-01-24 18:38:31 +0100149/* Allocates new entries for pool <pool> until there are at least <avail> + 1
150 * available, then returns the last one for immediate use, so that at least
151 * <avail> are left available in the pool upon return. NULL is returned if the
152 * last entry could not be allocated. It's important to note that at least one
153 * allocation is always performed even if there are enough entries in the pool.
154 * A call to the garbage collector is performed at most once in case malloc()
155 * returns an error, before returning NULL.
156 */
157void *__pool_refill_alloc(struct pool_head *pool, unsigned int avail)
158{
Olivier Houchard8b2c8a72018-10-21 01:52:59 +0200159 void *ptr = NULL, **free_list;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100160 int failed = 0;
161 int size = pool->size;
162 int limit = pool->limit;
163 int allocated = pool->allocated, allocated_orig = allocated;
164
165 /* stop point */
166 avail += pool->used;
167
168 while (1) {
169 if (limit && allocated >= limit) {
Olivier Houchard20872762019-03-08 18:53:35 +0100170 _HA_ATOMIC_ADD(&pool->allocated, allocated - allocated_orig);
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200171 activity[tid].pool_fail++;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100172 return NULL;
173 }
174
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200175 pool_avg_bump(&pool->needed_avg, pool->allocated);
176
Olivier Houchardcf975d42018-01-24 18:38:31 +0100177 ptr = malloc(size + POOL_EXTRA);
178 if (!ptr) {
Olivier Houchard20872762019-03-08 18:53:35 +0100179 _HA_ATOMIC_ADD(&pool->failed, 1);
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200180 if (failed) {
181 activity[tid].pool_fail++;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100182 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200183 }
Olivier Houchardcf975d42018-01-24 18:38:31 +0100184 failed++;
185 pool_gc(pool);
186 continue;
187 }
188 if (++allocated > avail)
189 break;
190
191 free_list = pool->free_list;
192 do {
193 *POOL_LINK(pool, ptr) = free_list;
194 __ha_barrier_store();
Olivier Houchard20872762019-03-08 18:53:35 +0100195 } while (_HA_ATOMIC_CAS(&pool->free_list, &free_list, ptr) == 0);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100196 }
Olivier Houchard20872762019-03-08 18:53:35 +0100197 __ha_barrier_atomic_store();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100198
Olivier Houchard20872762019-03-08 18:53:35 +0100199 _HA_ATOMIC_ADD(&pool->allocated, allocated - allocated_orig);
200 _HA_ATOMIC_ADD(&pool->used, 1);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100201
202#ifdef DEBUG_MEMORY_POOLS
203 /* keep track of where the element was allocated from */
204 *POOL_LINK(pool, ptr) = (void *)pool;
205#endif
206 return ptr;
207}
208void *pool_refill_alloc(struct pool_head *pool, unsigned int avail)
209{
210 void *ptr;
211
212 ptr = __pool_refill_alloc(pool, avail);
213 return ptr;
214}
215/*
216 * This function frees whatever can be freed in pool <pool>.
217 */
218void pool_flush(struct pool_head *pool)
219{
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100220 struct pool_free_list cmp, new;
Olivier Houchard8b2c8a72018-10-21 01:52:59 +0200221 void **next, *temp;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100222 int removed = 0;
223
224 if (!pool)
225 return;
Willy Tarreau21072b92020-05-29 17:23:05 +0200226 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100227 do {
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100228 cmp.free_list = pool->free_list;
229 cmp.seq = pool->seq;
230 new.free_list = NULL;
231 new.seq = cmp.seq + 1;
232 } while (!_HA_ATOMIC_DWCAS(&pool->free_list, &cmp, &new));
Olivier Houchard20872762019-03-08 18:53:35 +0100233 __ha_barrier_atomic_store();
Willy Tarreau21072b92020-05-29 17:23:05 +0200234 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Olivier Houchardb6fa08b2020-02-01 17:37:22 +0100235 next = cmp.free_list;
Olivier Houchardcf975d42018-01-24 18:38:31 +0100236 while (next) {
237 temp = next;
238 next = *POOL_LINK(pool, temp);
239 removed++;
240 free(temp);
241 }
242 pool->free_list = next;
Olivier Houchard20872762019-03-08 18:53:35 +0100243 _HA_ATOMIC_SUB(&pool->allocated, removed);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100244 /* here, we should have pool->allocate == pool->used */
245}
246
247/*
248 * This function frees whatever can be freed in all pools, but respecting
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200249 * the minimum thresholds imposed by owners. It makes sure to be alone to
250 * run by using thread_isolate(). <pool_ctx> is unused.
Olivier Houchardcf975d42018-01-24 18:38:31 +0100251 */
252void pool_gc(struct pool_head *pool_ctx)
253{
Olivier Houchardcf975d42018-01-24 18:38:31 +0100254 struct pool_head *entry;
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200255 int isolated = thread_isolated();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100256
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200257 if (!isolated)
258 thread_isolate();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100259
260 list_for_each_entry(entry, &pools, list) {
261 while ((int)((volatile int)entry->allocated - (volatile int)entry->used) > (int)entry->minavail) {
262 struct pool_free_list cmp, new;
263
264 cmp.seq = entry->seq;
265 __ha_barrier_load();
266 cmp.free_list = entry->free_list;
267 __ha_barrier_load();
268 if (cmp.free_list == NULL)
269 break;
270 new.free_list = *POOL_LINK(entry, cmp.free_list);
271 new.seq = cmp.seq + 1;
Willy Tarreau6a38b322019-05-11 18:04:24 +0200272 if (HA_ATOMIC_DWCAS(&entry->free_list, &cmp, &new) == 0)
Olivier Houchardcf975d42018-01-24 18:38:31 +0100273 continue;
274 free(cmp.free_list);
Olivier Houchard20872762019-03-08 18:53:35 +0100275 _HA_ATOMIC_SUB(&entry->allocated, 1);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100276 }
277 }
278
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200279 if (!isolated)
280 thread_release();
Olivier Houchardcf975d42018-01-24 18:38:31 +0100281}
Willy Tarreaue18db9e2018-10-16 10:28:54 +0200282
283/* frees an object to the local cache, possibly pushing oldest objects to the
284 * global pool. Must not be called directly.
285 */
286void __pool_put_to_cache(struct pool_head *pool, void *ptr, ssize_t idx)
287{
288 struct pool_cache_item *item = (struct pool_cache_item *)ptr;
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100289 struct pool_cache_head *ph = &pool_cache[tid][idx];
Willy Tarreaue18db9e2018-10-16 10:28:54 +0200290
291 LIST_ADD(&ph->list, &item->by_pool);
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100292 LIST_ADD(&pool_lru_head[tid], &item->by_lru);
Willy Tarreaue18db9e2018-10-16 10:28:54 +0200293 ph->count++;
294 pool_cache_count++;
295 pool_cache_bytes += ph->size;
296
297 if (pool_cache_bytes <= CONFIG_HAP_POOL_CACHE_SIZE)
298 return;
299
300 do {
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100301 item = LIST_PREV(&pool_lru_head[tid], struct pool_cache_item *, by_lru);
Willy Tarreaue18db9e2018-10-16 10:28:54 +0200302 /* note: by definition we remove oldest objects so they also are the
303 * oldest in their own pools, thus their next is the pool's head.
304 */
305 ph = LIST_NEXT(&item->by_pool, struct pool_cache_head *, list);
306 LIST_DEL(&item->by_pool);
307 LIST_DEL(&item->by_lru);
308 ph->count--;
309 pool_cache_count--;
310 pool_cache_bytes -= ph->size;
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100311 __pool_free(pool_base_start + (ph - pool_cache[tid]), item);
Willy Tarreaue18db9e2018-10-16 10:28:54 +0200312 } while (pool_cache_bytes > CONFIG_HAP_POOL_CACHE_SIZE * 7 / 8);
313}
314
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100315#else /* CONFIG_HAP_LOCKLESS_POOLS */
Willy Tarreau50e608d2007-05-13 18:26:08 +0200316
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100317/* Allocates new entries for pool <pool> until there are at least <avail> + 1
318 * available, then returns the last one for immediate use, so that at least
319 * <avail> are left available in the pool upon return. NULL is returned if the
320 * last entry could not be allocated. It's important to note that at least one
321 * allocation is always performed even if there are enough entries in the pool.
322 * A call to the garbage collector is performed at most once in case malloc()
323 * returns an error, before returning NULL.
Willy Tarreau50e608d2007-05-13 18:26:08 +0200324 */
Christopher Fauletb349e482017-08-29 09:52:38 +0200325void *__pool_refill_alloc(struct pool_head *pool, unsigned int avail)
Willy Tarreau50e608d2007-05-13 18:26:08 +0200326{
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100327 void *ptr = NULL;
328 int failed = 0;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200329
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100330#ifdef DEBUG_FAIL_ALLOC
331 if (mem_should_fail(pool))
332 return NULL;
333#endif
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100334 /* stop point */
335 avail += pool->used;
336
337 while (1) {
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200338 if (pool->limit && pool->allocated >= pool->limit) {
339 activity[tid].pool_fail++;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200340 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200341 }
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100342
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200343 pool_avg_bump(&pool->needed_avg, pool->allocated);
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200344 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Willy Tarreauf13322e2017-11-22 10:50:54 +0100345 ptr = pool_alloc_area(pool->size + POOL_EXTRA);
Willy Tarreau82867542019-07-04 11:48:16 +0200346#ifdef DEBUG_MEMORY_POOLS
347 /* keep track of where the element was allocated from. This
348 * is done out of the lock so that the system really allocates
349 * the data without harming other threads waiting on the lock.
350 */
351 if (ptr)
352 *POOL_LINK(pool, ptr) = (void *)pool;
353#endif
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200354 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100355 if (!ptr) {
Willy Tarreau58102cf2015-10-28 16:24:21 +0100356 pool->failed++;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200357 if (failed) {
358 activity[tid].pool_fail++;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100359 return NULL;
Willy Tarreaua8b2ce02019-05-28 17:04:16 +0200360 }
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100361 failed++;
Willy Tarreaubafbe012017-11-24 17:34:44 +0100362 pool_gc(pool);
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100363 continue;
364 }
365 if (++pool->allocated > avail)
366 break;
367
Willy Tarreauac421112015-10-28 15:09:29 +0100368 *POOL_LINK(pool, ptr) = (void *)pool->free_list;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100369 pool->free_list = ptr;
Willy Tarreau7dcd46d2007-05-14 00:16:13 +0200370 }
Willy Tarreau50e608d2007-05-13 18:26:08 +0200371 pool->used++;
Willy Tarreaua885f6d2014-12-03 15:25:28 +0100372 return ptr;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200373}
Christopher Fauletb349e482017-08-29 09:52:38 +0200374void *pool_refill_alloc(struct pool_head *pool, unsigned int avail)
375{
376 void *ptr;
Willy Tarreau50e608d2007-05-13 18:26:08 +0200377
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100378 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
Christopher Fauletb349e482017-08-29 09:52:38 +0200379 ptr = __pool_refill_alloc(pool, avail);
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100380 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Christopher Fauletb349e482017-08-29 09:52:38 +0200381 return ptr;
382}
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200383/*
384 * This function frees whatever can be freed in pool <pool>.
385 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100386void pool_flush(struct pool_head *pool)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200387{
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200388 void *temp;
389
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200390 if (!pool)
391 return;
392
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200393 while (1) {
394 HA_SPIN_LOCK(POOL_LOCK, &pool->lock);
395 temp = pool->free_list;
396 if (!temp) {
397 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
398 break;
399 }
400 pool->free_list = *POOL_LINK(pool, temp);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200401 pool->allocated--;
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200402 HA_SPIN_UNLOCK(POOL_LOCK, &pool->lock);
Willy Tarreauf13322e2017-11-22 10:50:54 +0100403 pool_free_area(temp, pool->size + POOL_EXTRA);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200404 }
Willy Tarreau3e853ea2019-07-04 11:30:00 +0200405 /* here, we should have pool->allocated == pool->used */
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200406}
407
408/*
409 * This function frees whatever can be freed in all pools, but respecting
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200410 * the minimum thresholds imposed by owners. It makes sure to be alone to
411 * run by using thread_isolate(). <pool_ctx> is unused.
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200412 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100413void pool_gc(struct pool_head *pool_ctx)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200414{
415 struct pool_head *entry;
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200416 int isolated = thread_isolated();
Willy Tarreaub7f9d122009-04-21 02:17:45 +0200417
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200418 if (!isolated)
419 thread_isolate();
Willy Tarreaub7f9d122009-04-21 02:17:45 +0200420
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200421 list_for_each_entry(entry, &pools, list) {
Olivier Houchard51d93392020-03-12 19:05:39 +0100422 void *temp;
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200423 //qfprintf(stderr, "Flushing pool %s\n", entry->name);
Olivier Houchard51d93392020-03-12 19:05:39 +0100424 while (entry->free_list &&
Willy Tarreau57767b82014-12-22 21:40:55 +0100425 (int)(entry->allocated - entry->used) > (int)entry->minavail) {
Olivier Houchard51d93392020-03-12 19:05:39 +0100426 temp = entry->free_list;
427 entry->free_list = *POOL_LINK(entry, temp);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200428 entry->allocated--;
Willy Tarreauf13322e2017-11-22 10:50:54 +0100429 pool_free_area(temp, entry->size + POOL_EXTRA);
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200430 }
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200431 }
Christopher Fauletb349e482017-08-29 09:52:38 +0200432
Willy Tarreauc0e2ff22020-04-24 06:15:24 +0200433 if (!isolated)
434 thread_release();
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200435}
Olivier Houchardcf975d42018-01-24 18:38:31 +0100436#endif
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200437
438/*
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200439 * This function destroys a pool by freeing it completely, unless it's still
440 * in use. This should be called only under extreme circumstances. It always
441 * returns NULL if the resulting pool is empty, easing the clearing of the old
442 * pointer, otherwise it returns the pool.
443 * .
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200444 */
Willy Tarreaubafbe012017-11-24 17:34:44 +0100445void *pool_destroy(struct pool_head *pool)
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200446{
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200447 if (pool) {
Willy Tarreaubafbe012017-11-24 17:34:44 +0100448 pool_flush(pool);
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200449 if (pool->used)
450 return pool;
451 pool->users--;
452 if (!pool->users) {
453 LIST_DEL(&pool->list);
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100454#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100455 HA_SPIN_DESTROY(&pool->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100456#endif
Willy Tarreau0a93b642018-10-16 07:58:39 +0200457 if ((pool - pool_base_start) < MAX_BASE_POOLS)
458 memset(pool, 0, sizeof(*pool));
459 else
460 free(pool);
Willy Tarreaudae4aa82007-06-16 23:19:53 +0200461 }
Willy Tarreau4d2d0982007-05-14 00:39:29 +0200462 }
463 return NULL;
Willy Tarreaue6ce59d2007-05-13 19:38:49 +0200464}
465
Willy Tarreau2455ceb2018-11-26 15:57:34 +0100466/* This destroys all pools on exit. It is *not* thread safe. */
467void pool_destroy_all()
468{
469 struct pool_head *entry, *back;
470
471 list_for_each_entry_safe(entry, back, &pools, list)
472 pool_destroy(entry);
473}
474
Willy Tarreau12833bb2014-01-28 16:49:56 +0100475/* This function dumps memory usage information into the trash buffer. */
476void dump_pools_to_trash()
Willy Tarreau50e608d2007-05-13 18:26:08 +0200477{
478 struct pool_head *entry;
479 unsigned long allocated, used;
480 int nbpools;
481
482 allocated = used = nbpools = 0;
Willy Tarreau12833bb2014-01-28 16:49:56 +0100483 chunk_printf(&trash, "Dumping pools usage. Use SIGQUIT to flush them.\n");
Willy Tarreau50e608d2007-05-13 18:26:08 +0200484 list_for_each_entry(entry, &pools, list) {
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100485#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100486 HA_SPIN_LOCK(POOL_LOCK, &entry->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100487#endif
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200488 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 +0200489 entry->name, entry->size, entry->allocated,
Willy Tarreaua1e4f8c2020-05-08 08:31:56 +0200490 entry->size * entry->allocated, entry->used,
491 pool_avg(entry->needed_avg), entry->failed,
Willy Tarreau0a93b642018-10-16 07:58:39 +0200492 entry->users, entry, (int)pool_get_index(entry),
493 (entry->flags & MEM_F_SHARED) ? " [SHARED]" : "");
Willy Tarreau50e608d2007-05-13 18:26:08 +0200494
495 allocated += entry->allocated * entry->size;
496 used += entry->used * entry->size;
497 nbpools++;
Willy Tarreauf161d0f2018-02-22 14:05:55 +0100498#ifndef CONFIG_HAP_LOCKLESS_POOLS
Christopher Faulet2a944ee2017-11-07 10:42:54 +0100499 HA_SPIN_UNLOCK(POOL_LOCK, &entry->lock);
Olivier Houchardcf975d42018-01-24 18:38:31 +0100500#endif
Willy Tarreau50e608d2007-05-13 18:26:08 +0200501 }
Willy Tarreau12833bb2014-01-28 16:49:56 +0100502 chunk_appendf(&trash, "Total: %d pools, %lu bytes allocated, %lu used.\n",
Willy Tarreau50e608d2007-05-13 18:26:08 +0200503 nbpools, allocated, used);
504}
505
Willy Tarreau12833bb2014-01-28 16:49:56 +0100506/* Dump statistics on pools usage. */
507void dump_pools(void)
508{
509 dump_pools_to_trash();
Willy Tarreau843b7cb2018-07-13 10:54:26 +0200510 qfprintf(stderr, "%s", trash.area);
Willy Tarreau12833bb2014-01-28 16:49:56 +0100511}
512
Willy Tarreau58102cf2015-10-28 16:24:21 +0100513/* This function returns the total number of failed pool allocations */
514int pool_total_failures()
515{
516 struct pool_head *entry;
517 int failed = 0;
518
519 list_for_each_entry(entry, &pools, list)
520 failed += entry->failed;
521 return failed;
522}
523
524/* This function returns the total amount of memory allocated in pools (in bytes) */
525unsigned long pool_total_allocated()
526{
527 struct pool_head *entry;
528 unsigned long allocated = 0;
529
530 list_for_each_entry(entry, &pools, list)
531 allocated += entry->allocated * entry->size;
532 return allocated;
533}
534
535/* This function returns the total amount of memory used in pools (in bytes) */
536unsigned long pool_total_used()
537{
538 struct pool_head *entry;
539 unsigned long used = 0;
540
541 list_for_each_entry(entry, &pools, list)
542 used += entry->used * entry->size;
543 return used;
544}
545
William Lallemande7ed8852016-11-19 02:25:36 +0100546/* This function dumps memory usage information onto the stream interface's
547 * read buffer. It returns 0 as long as it does not complete, non-zero upon
548 * completion. No state is used.
549 */
550static int cli_io_handler_dump_pools(struct appctx *appctx)
551{
552 struct stream_interface *si = appctx->owner;
553
554 dump_pools_to_trash();
Willy Tarreau06d80a92017-10-19 14:32:15 +0200555 if (ci_putchk(si_ic(si), &trash) == -1) {
Willy Tarreaudb398432018-11-15 11:08:52 +0100556 si_rx_room_blk(si);
William Lallemande7ed8852016-11-19 02:25:36 +0100557 return 0;
558 }
559 return 1;
560}
561
Willy Tarreau7107c8b2018-11-26 11:44:35 +0100562/* callback used to create early pool <name> of size <size> and store the
563 * resulting pointer into <ptr>. If the allocation fails, it quits with after
564 * emitting an error message.
565 */
566void create_pool_callback(struct pool_head **ptr, char *name, unsigned int size)
567{
568 *ptr = create_pool(name, size, MEM_F_SHARED);
569 if (!*ptr) {
570 ha_alert("Failed to allocate pool '%s' of size %u : %s. Aborting.\n",
571 name, size, strerror(errno));
572 exit(1);
573 }
574}
575
Willy Tarreau7f0165e2018-11-26 17:09:46 +0100576/* Initializes all per-thread arrays on startup */
577static void init_pools()
578{
579 int thr, idx;
580
581 for (thr = 0; thr < MAX_THREADS; thr++) {
582 for (idx = 0; idx < MAX_BASE_POOLS; idx++) {
583 LIST_INIT(&pool_cache[thr][idx].list);
584 pool_cache[thr][idx].size = 0;
585 }
586 LIST_INIT(&pool_lru_head[thr]);
587 }
588}
589
590INITCALL0(STG_PREPARE, init_pools);
Willy Tarreau7107c8b2018-11-26 11:44:35 +0100591
William Lallemande7ed8852016-11-19 02:25:36 +0100592/* register cli keywords */
593static struct cli_kw_list cli_kws = {{ },{
Willy Tarreaue9ecec82016-12-16 18:55:23 +0100594 { { "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 +0100595 {{},}
596}};
597
Willy Tarreau0108d902018-11-25 19:14:37 +0100598INITCALL1(STG_REGISTER, cli_register_kw, &cli_kws);
William Lallemande7ed8852016-11-19 02:25:36 +0100599
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100600#ifdef DEBUG_FAIL_ALLOC
601#define MEM_FAIL_MAX_CHAR 32
602#define MEM_FAIL_MAX_STR 128
603static int mem_fail_cur_idx;
604static char mem_fail_str[MEM_FAIL_MAX_CHAR * MEM_FAIL_MAX_STR];
605__decl_hathreads(static HA_SPINLOCK_T mem_fail_lock);
606
607int mem_should_fail(const struct pool_head *pool)
608{
Olivier Houchard9c4f08a2019-02-01 16:28:04 +0100609 int ret = 0;
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100610 int n;
611
612 if (mem_fail_rate > 0 && !(global.mode & MODE_STARTING)) {
Willy Tarreau52bf8392020-03-08 00:42:37 +0100613 int randnb = ha_random() % 100;
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100614
615 if (mem_fail_rate > randnb)
616 ret = 1;
617 else
618 ret = 0;
619 }
Olivier Houchard04f5fe82020-02-01 17:49:31 +0100620 HA_SPIN_LOCK(POOL_LOCK, &mem_fail_lock);
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100621 n = snprintf(&mem_fail_str[mem_fail_cur_idx * MEM_FAIL_MAX_CHAR],
622 MEM_FAIL_MAX_CHAR - 2,
623 "%d %.18s %d %d", mem_fail_cur_idx, pool->name, ret, tid);
624 while (n < MEM_FAIL_MAX_CHAR - 1)
625 mem_fail_str[mem_fail_cur_idx * MEM_FAIL_MAX_CHAR + n++] = ' ';
626 if (mem_fail_cur_idx < MEM_FAIL_MAX_STR - 1)
627 mem_fail_str[mem_fail_cur_idx * MEM_FAIL_MAX_CHAR + n] = '\n';
628 else
629 mem_fail_str[mem_fail_cur_idx * MEM_FAIL_MAX_CHAR + n] = 0;
630 mem_fail_cur_idx++;
631 if (mem_fail_cur_idx == MEM_FAIL_MAX_STR)
632 mem_fail_cur_idx = 0;
Olivier Houchard04f5fe82020-02-01 17:49:31 +0100633 HA_SPIN_UNLOCK(POOL_LOCK, &mem_fail_lock);
Olivier Houcharddc21ff72019-01-29 15:20:16 +0100634 return ret;
635
636}
637
638/* config parser for global "tune.fail-alloc" */
639static int mem_parse_global_fail_alloc(char **args, int section_type, struct proxy *curpx,
640 struct proxy *defpx, const char *file, int line,
641 char **err)
642{
643 if (too_many_args(1, args, err, NULL))
644 return -1;
645 mem_fail_rate = atoi(args[1]);
646 if (mem_fail_rate < 0 || mem_fail_rate > 100) {
647 memprintf(err, "'%s' expects a numeric value between 0 and 100.", args[0]);
648 return -1;
649 }
650 return 0;
651}
652#endif
653
654/* register global config keywords */
655static struct cfg_kw_list mem_cfg_kws = {ILH, {
656#ifdef DEBUG_FAIL_ALLOC
657 { CFG_GLOBAL, "tune.fail-alloc", mem_parse_global_fail_alloc },
658#endif
659 { 0, NULL, NULL }
660}};
661
662INITCALL1(STG_REGISTER, cfg_register_keywords, &mem_cfg_kws);
663
Willy Tarreau50e608d2007-05-13 18:26:08 +0200664/*
665 * Local variables:
666 * c-indent-level: 8
667 * c-basic-offset: 8
668 * End:
669 */