Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 1 | /* |
| 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 Tarreau | f14d190 | 2021-10-05 18:14:11 +0200 | [diff] [blame] | 12 | |
| 13 | #include <sys/mman.h> |
Willy Tarreau | 7107c8b | 2018-11-26 11:44:35 +0100 | [diff] [blame] | 14 | #include <errno.h> |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 15 | |
Willy Tarreau | 5d9ddc5 | 2021-10-06 19:54:09 +0200 | [diff] [blame] | 16 | #include <haproxy/activity.h> |
Willy Tarreau | 4c7e4b7 | 2020-05-27 12:58:42 +0200 | [diff] [blame] | 17 | #include <haproxy/api.h> |
Willy Tarreau | b255105 | 2020-06-09 09:07:15 +0200 | [diff] [blame] | 18 | #include <haproxy/applet-t.h> |
Willy Tarreau | 6be7849 | 2020-06-05 00:00:29 +0200 | [diff] [blame] | 19 | #include <haproxy/cfgparse.h> |
Willy Tarreau | f1d32c4 | 2020-06-04 21:07:02 +0200 | [diff] [blame] | 20 | #include <haproxy/channel.h> |
Willy Tarreau | 83487a8 | 2020-06-04 20:19:54 +0200 | [diff] [blame] | 21 | #include <haproxy/cli.h> |
Willy Tarreau | 36979d9 | 2020-06-05 17:27:29 +0200 | [diff] [blame] | 22 | #include <haproxy/errors.h> |
Willy Tarreau | f268ee8 | 2020-06-04 17:05:57 +0200 | [diff] [blame] | 23 | #include <haproxy/global.h> |
Willy Tarreau | 853b297 | 2020-05-27 18:01:47 +0200 | [diff] [blame] | 24 | #include <haproxy/list.h> |
Willy Tarreau | b255105 | 2020-06-09 09:07:15 +0200 | [diff] [blame] | 25 | #include <haproxy/pool.h> |
Willy Tarreau | 2eec9b5 | 2020-06-04 19:58:55 +0200 | [diff] [blame] | 26 | #include <haproxy/stats-t.h> |
Willy Tarreau | 5e539c9 | 2020-06-04 20:45:39 +0200 | [diff] [blame] | 27 | #include <haproxy/stream_interface.h> |
Willy Tarreau | b255105 | 2020-06-09 09:07:15 +0200 | [diff] [blame] | 28 | #include <haproxy/thread.h> |
Willy Tarreau | 48fbcae | 2020-06-03 18:09:46 +0200 | [diff] [blame] | 29 | #include <haproxy/tools.h> |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 30 | |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 31 | |
Willy Tarreau | 2d6f628 | 2021-04-15 16:24:00 +0200 | [diff] [blame] | 32 | #ifdef CONFIG_HAP_POOLS |
Willy Tarreau | 7f0165e | 2018-11-26 17:09:46 +0100 | [diff] [blame] | 33 | /* These ones are initialized per-thread on startup by init_pools() */ |
Willy Tarreau | e18db9e | 2018-10-16 10:28:54 +0200 | [diff] [blame] | 34 | THREAD_LOCAL size_t pool_cache_bytes = 0; /* total cache size */ |
| 35 | THREAD_LOCAL size_t pool_cache_count = 0; /* #cache objects */ |
Willy Tarreau | ed891fd | 2020-06-01 19:00:28 +0200 | [diff] [blame] | 36 | #endif |
Willy Tarreau | e18db9e | 2018-10-16 10:28:54 +0200 | [diff] [blame] | 37 | |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 38 | static struct list pools = LIST_HEAD_INIT(pools); |
Willy Tarreau | 067ac9f | 2015-10-08 14:12:13 +0200 | [diff] [blame] | 39 | int mem_poison_byte = -1; |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 40 | |
Olivier Houchard | dc21ff7 | 2019-01-29 15:20:16 +0100 | [diff] [blame] | 41 | #ifdef DEBUG_FAIL_ALLOC |
| 42 | static int mem_fail_rate = 0; |
Olivier Houchard | dc21ff7 | 2019-01-29 15:20:16 +0100 | [diff] [blame] | 43 | #endif |
| 44 | |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 45 | static int using_default_allocator = 1; |
| 46 | static int(*my_mallctl)(const char *, void *, size_t *, void *, size_t) = NULL; |
Willy Tarreau | 157e393 | 2021-09-15 10:05:48 +0200 | [diff] [blame] | 47 | |
Willy Tarreau | 0d93a81 | 2021-12-23 09:26:30 +0100 | [diff] [blame] | 48 | /* ask the allocator to trim memory pools. |
| 49 | * This must run under thread isolation so that competing threads trying to |
| 50 | * allocate or release memory do not prevent the allocator from completing |
| 51 | * its job. We just have to be careful as callers might already be isolated |
| 52 | * themselves. |
| 53 | */ |
Willy Tarreau | ea3323f | 2021-09-15 10:38:21 +0200 | [diff] [blame] | 54 | static void trim_all_pools(void) |
| 55 | { |
Willy Tarreau | 0d93a81 | 2021-12-23 09:26:30 +0100 | [diff] [blame] | 56 | int isolated = thread_isolated(); |
| 57 | |
| 58 | if (!isolated) |
| 59 | thread_isolate(); |
| 60 | |
David Carlier | d450ff6 | 2021-11-25 16:14:38 +0000 | [diff] [blame] | 61 | if (my_mallctl) { |
| 62 | unsigned int i, narenas = 0; |
| 63 | size_t len = sizeof(narenas); |
| 64 | |
| 65 | if (my_mallctl("arenas.narenas", &narenas, &len, NULL, 0) == 0) { |
| 66 | for (i = 0; i < narenas; i ++) { |
| 67 | char mib[32] = {0}; |
| 68 | snprintf(mib, sizeof(mib), "arena.%u.purge", i); |
| 69 | (void)my_mallctl(mib, NULL, NULL, NULL, 0); |
| 70 | } |
| 71 | } |
| 72 | } else { |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 73 | #if defined(HA_HAVE_MALLOC_TRIM) |
David Carlier | d450ff6 | 2021-11-25 16:14:38 +0000 | [diff] [blame] | 74 | if (using_default_allocator) |
| 75 | malloc_trim(0); |
David CARLIER | b1e190a | 2021-11-26 20:44:44 +0000 | [diff] [blame] | 76 | #elif defined(HA_HAVE_MALLOC_ZONE) |
| 77 | if (using_default_allocator) { |
| 78 | vm_address_t *zones; |
| 79 | unsigned int i, nzones; |
| 80 | |
| 81 | if (malloc_get_all_zones(0, NULL, &zones, &nzones) == KERN_SUCCESS) { |
| 82 | for (i = 0; i < nzones; i ++) { |
| 83 | malloc_zone_t *zone = (malloc_zone_t *)zones[i]; |
| 84 | |
| 85 | /* we cannot purge anonymous zones */ |
| 86 | if (zone->zone_name) |
| 87 | malloc_zone_pressure_relief(zone, 0); |
| 88 | } |
| 89 | } |
| 90 | } |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 91 | #endif |
David Carlier | d450ff6 | 2021-11-25 16:14:38 +0000 | [diff] [blame] | 92 | } |
Willy Tarreau | 0d93a81 | 2021-12-23 09:26:30 +0100 | [diff] [blame] | 93 | |
| 94 | if (!isolated) |
| 95 | thread_release(); |
Willy Tarreau | ea3323f | 2021-09-15 10:38:21 +0200 | [diff] [blame] | 96 | } |
| 97 | |
Willy Tarreau | 157e393 | 2021-09-15 10:05:48 +0200 | [diff] [blame] | 98 | /* check if we're using the same allocator as the one that provides |
| 99 | * malloc_trim() and mallinfo(). The principle is that on glibc, both |
| 100 | * malloc_trim() and mallinfo() are provided, and using mallinfo() we |
| 101 | * can check if malloc() is performed through glibc or any other one |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 102 | * the executable was linked against (e.g. jemalloc). Prior to this we |
| 103 | * have to check whether we're running on jemalloc by verifying if the |
| 104 | * mallctl() function is provided. Its pointer will be used later. |
Willy Tarreau | 157e393 | 2021-09-15 10:05:48 +0200 | [diff] [blame] | 105 | */ |
| 106 | static void detect_allocator(void) |
| 107 | { |
Willy Tarreau | 781f07a | 2021-11-26 15:55:55 +0100 | [diff] [blame] | 108 | #if defined(__ELF__) |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 109 | extern int mallctl(const char *, void *, size_t *, void *, size_t) __attribute__((weak)); |
| 110 | |
| 111 | my_mallctl = mallctl; |
Willy Tarreau | 781f07a | 2021-11-26 15:55:55 +0100 | [diff] [blame] | 112 | #endif |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 113 | |
| 114 | if (!my_mallctl) { |
| 115 | my_mallctl = get_sym_curr_addr("mallctl"); |
| 116 | using_default_allocator = (my_mallctl == NULL); |
| 117 | } |
| 118 | |
| 119 | if (!my_mallctl) { |
| 120 | #if defined(HA_HAVE_MALLOC_TRIM) |
Willy Tarreau | c2afb86 | 2021-09-16 09:18:21 +0200 | [diff] [blame] | 121 | #ifdef HA_HAVE_MALLINFO2 |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 122 | struct mallinfo2 mi1, mi2; |
Willy Tarreau | c2afb86 | 2021-09-16 09:18:21 +0200 | [diff] [blame] | 123 | #else |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 124 | struct mallinfo mi1, mi2; |
Willy Tarreau | c2afb86 | 2021-09-16 09:18:21 +0200 | [diff] [blame] | 125 | #endif |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 126 | void *ptr; |
Willy Tarreau | 157e393 | 2021-09-15 10:05:48 +0200 | [diff] [blame] | 127 | |
Willy Tarreau | c2afb86 | 2021-09-16 09:18:21 +0200 | [diff] [blame] | 128 | #ifdef HA_HAVE_MALLINFO2 |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 129 | mi1 = mallinfo2(); |
Willy Tarreau | c2afb86 | 2021-09-16 09:18:21 +0200 | [diff] [blame] | 130 | #else |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 131 | mi1 = mallinfo(); |
Willy Tarreau | c2afb86 | 2021-09-16 09:18:21 +0200 | [diff] [blame] | 132 | #endif |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 133 | ptr = DISGUISE(malloc(1)); |
Willy Tarreau | c2afb86 | 2021-09-16 09:18:21 +0200 | [diff] [blame] | 134 | #ifdef HA_HAVE_MALLINFO2 |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 135 | mi2 = mallinfo2(); |
Willy Tarreau | c2afb86 | 2021-09-16 09:18:21 +0200 | [diff] [blame] | 136 | #else |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 137 | mi2 = mallinfo(); |
Willy Tarreau | c2afb86 | 2021-09-16 09:18:21 +0200 | [diff] [blame] | 138 | #endif |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 139 | free(DISGUISE(ptr)); |
Willy Tarreau | ea3323f | 2021-09-15 10:38:21 +0200 | [diff] [blame] | 140 | |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 141 | using_default_allocator = !!memcmp(&mi1, &mi2, sizeof(mi1)); |
David CARLIER | b1e190a | 2021-11-26 20:44:44 +0000 | [diff] [blame] | 142 | #elif defined(HA_HAVE_MALLOC_ZONE) |
| 143 | using_default_allocator = (malloc_default_zone() != NULL); |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 144 | #endif |
| 145 | } |
Willy Tarreau | 845b560 | 2021-09-15 10:41:24 +0200 | [diff] [blame] | 146 | } |
| 147 | |
| 148 | static int is_trim_enabled(void) |
| 149 | { |
David Carlier | ed23214 | 2021-11-25 16:09:45 +0000 | [diff] [blame] | 150 | return using_default_allocator; |
Willy Tarreau | 157e393 | 2021-09-15 10:05:48 +0200 | [diff] [blame] | 151 | } |
Willy Tarreau | ea3323f | 2021-09-15 10:38:21 +0200 | [diff] [blame] | 152 | |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 153 | /* Try to find an existing shared pool with the same characteristics and |
| 154 | * returns it, otherwise creates this one. NULL is returned if no memory |
Willy Tarreau | 581bf81 | 2016-01-25 02:19:13 +0100 | [diff] [blame] | 155 | * is available for a new creation. Two flags are supported : |
| 156 | * - MEM_F_SHARED to indicate that the pool may be shared with other users |
| 157 | * - MEM_F_EXACT to indicate that the size must not be rounded up |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 158 | */ |
| 159 | struct pool_head *create_pool(char *name, unsigned int size, unsigned int flags) |
| 160 | { |
| 161 | struct pool_head *pool; |
Willy Tarreau | 7dcd46d | 2007-05-14 00:16:13 +0200 | [diff] [blame] | 162 | struct pool_head *entry; |
| 163 | struct list *start; |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 164 | unsigned int align; |
Willy Tarreau | 9f3129e | 2021-04-17 00:31:38 +0200 | [diff] [blame] | 165 | int thr __maybe_unused; |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 166 | |
Willy Tarreau | ac42111 | 2015-10-28 15:09:29 +0100 | [diff] [blame] | 167 | /* We need to store a (void *) at the end of the chunks. Since we know |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 168 | * that the malloc() function will never return such a small size, |
| 169 | * let's round the size up to something slightly bigger, in order to |
| 170 | * ease merging of entries. Note that the rounding is a power of two. |
Willy Tarreau | ac42111 | 2015-10-28 15:09:29 +0100 | [diff] [blame] | 171 | * This extra (void *) is not accounted for in the size computation |
| 172 | * so that the visible parts outside are not affected. |
Willy Tarreau | 30f931e | 2018-10-23 14:40:23 +0200 | [diff] [blame] | 173 | * |
| 174 | * Note: for the LRU cache, we need to store 2 doubly-linked lists. |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 175 | */ |
| 176 | |
Willy Tarreau | 581bf81 | 2016-01-25 02:19:13 +0100 | [diff] [blame] | 177 | if (!(flags & MEM_F_EXACT)) { |
Willy Tarreau | 30f931e | 2018-10-23 14:40:23 +0200 | [diff] [blame] | 178 | align = 4 * sizeof(void *); // 2 lists = 4 pointers min |
Willy Tarreau | 581bf81 | 2016-01-25 02:19:13 +0100 | [diff] [blame] | 179 | size = ((size + POOL_EXTRA + align - 1) & -align) - POOL_EXTRA; |
| 180 | } |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 181 | |
Christopher Faulet | b349e48 | 2017-08-29 09:52:38 +0200 | [diff] [blame] | 182 | /* TODO: thread: we do not lock pool list for now because all pools are |
| 183 | * created during HAProxy startup (so before threads creation) */ |
Willy Tarreau | 7dcd46d | 2007-05-14 00:16:13 +0200 | [diff] [blame] | 184 | start = &pools; |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 185 | pool = NULL; |
Willy Tarreau | 7dcd46d | 2007-05-14 00:16:13 +0200 | [diff] [blame] | 186 | |
| 187 | list_for_each_entry(entry, &pools, list) { |
| 188 | if (entry->size == size) { |
| 189 | /* either we can share this place and we take it, or |
Ilya Shipitsin | 47d1718 | 2020-06-21 21:42:57 +0500 | [diff] [blame] | 190 | * we look for a shareable one or for the next position |
Willy Tarreau | 7dcd46d | 2007-05-14 00:16:13 +0200 | [diff] [blame] | 191 | * before which we will insert a new one. |
| 192 | */ |
Willy Tarreau | 1ab6c0b | 2021-05-05 07:29:01 +0200 | [diff] [blame] | 193 | if ((flags & entry->flags & MEM_F_SHARED) |
| 194 | #ifdef DEBUG_DONT_SHARE_POOLS |
| 195 | && strcmp(name, entry->name) == 0 |
| 196 | #endif |
| 197 | ) { |
Willy Tarreau | 7dcd46d | 2007-05-14 00:16:13 +0200 | [diff] [blame] | 198 | /* we can share this one */ |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 199 | pool = entry; |
Krzysztof Piotr Oledzki | a643baf | 2008-05-29 23:53:44 +0200 | [diff] [blame] | 200 | DPRINTF(stderr, "Sharing %s with %s\n", name, pool->name); |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 201 | break; |
| 202 | } |
| 203 | } |
Willy Tarreau | 7dcd46d | 2007-05-14 00:16:13 +0200 | [diff] [blame] | 204 | else if (entry->size > size) { |
| 205 | /* insert before this one */ |
| 206 | start = &entry->list; |
| 207 | break; |
| 208 | } |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 209 | } |
| 210 | |
| 211 | if (!pool) { |
Willy Tarreau | 0a93b64 | 2018-10-16 07:58:39 +0200 | [diff] [blame] | 212 | if (!pool) |
| 213 | pool = calloc(1, sizeof(*pool)); |
| 214 | |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 215 | if (!pool) |
| 216 | return NULL; |
| 217 | if (name) |
| 218 | strlcpy2(pool->name, name, sizeof(pool->name)); |
| 219 | pool->size = size; |
| 220 | pool->flags = flags; |
Willy Tarreau | 2b71810 | 2021-04-21 07:32:39 +0200 | [diff] [blame] | 221 | LIST_APPEND(start, &pool->list); |
Christopher Faulet | 2f6d3c0 | 2019-06-25 21:45:59 +0200 | [diff] [blame] | 222 | |
Willy Tarreau | 2d6f628 | 2021-04-15 16:24:00 +0200 | [diff] [blame] | 223 | #ifdef CONFIG_HAP_POOLS |
Christopher Faulet | 2f6d3c0 | 2019-06-25 21:45:59 +0200 | [diff] [blame] | 224 | /* update per-thread pool cache if necessary */ |
Willy Tarreau | 9f3129e | 2021-04-17 00:31:38 +0200 | [diff] [blame] | 225 | for (thr = 0; thr < MAX_THREADS; thr++) { |
| 226 | LIST_INIT(&pool->cache[thr].list); |
Christopher Faulet | 2f6d3c0 | 2019-06-25 21:45:59 +0200 | [diff] [blame] | 227 | } |
Willy Tarreau | ed891fd | 2020-06-01 19:00:28 +0200 | [diff] [blame] | 228 | #endif |
Olivier Houchard | 8af97eb | 2020-02-01 17:45:32 +0100 | [diff] [blame] | 229 | } |
| 230 | pool->users++; |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 231 | return pool; |
| 232 | } |
Olivier Houchard | cf975d4 | 2018-01-24 18:38:31 +0100 | [diff] [blame] | 233 | |
Willy Tarreau | 1384364 | 2021-04-17 16:57:25 +0200 | [diff] [blame] | 234 | /* Tries to allocate an object for the pool <pool> using the system's allocator |
| 235 | * and directly returns it. The pool's allocated counter is checked and updated, |
Willy Tarreau | 8715dec | 2021-06-10 17:31:48 +0200 | [diff] [blame] | 236 | * but no other checks are performed. |
Willy Tarreau | 1384364 | 2021-04-17 16:57:25 +0200 | [diff] [blame] | 237 | */ |
| 238 | void *pool_get_from_os(struct pool_head *pool) |
| 239 | { |
| 240 | if (!pool->limit || pool->allocated < pool->limit) { |
| 241 | void *ptr = pool_alloc_area(pool->size + POOL_EXTRA); |
| 242 | if (ptr) { |
| 243 | _HA_ATOMIC_INC(&pool->allocated); |
| 244 | return ptr; |
| 245 | } |
| 246 | _HA_ATOMIC_INC(&pool->failed); |
| 247 | } |
| 248 | activity[tid].pool_fail++; |
| 249 | return NULL; |
| 250 | |
| 251 | } |
| 252 | |
Willy Tarreau | 45e4e28 | 2021-04-17 17:48:40 +0200 | [diff] [blame] | 253 | /* Releases a pool item back to the operating system and atomically updates |
| 254 | * the allocation counter. |
| 255 | */ |
| 256 | void pool_put_to_os(struct pool_head *pool, void *ptr) |
| 257 | { |
Willy Tarreau | 9a7aa3b | 2021-06-10 17:20:19 +0200 | [diff] [blame] | 258 | #ifdef DEBUG_UAF |
| 259 | /* This object will be released for real in order to detect a use after |
| 260 | * free. We also force a write to the area to ensure we crash on double |
| 261 | * free or free of a const area. |
| 262 | */ |
| 263 | *(uint32_t *)ptr = 0xDEADADD4; |
| 264 | #endif /* DEBUG_UAF */ |
| 265 | |
Willy Tarreau | 45e4e28 | 2021-04-17 17:48:40 +0200 | [diff] [blame] | 266 | pool_free_area(ptr, pool->size + POOL_EXTRA); |
| 267 | _HA_ATOMIC_DEC(&pool->allocated); |
| 268 | } |
| 269 | |
Willy Tarreau | 8fe726f | 2021-04-15 18:20:12 +0200 | [diff] [blame] | 270 | /* Tries to allocate an object for the pool <pool> using the system's allocator |
| 271 | * and directly returns it. The pool's counters are updated but the object is |
| 272 | * never cached, so this is usable with and without local or shared caches. |
Willy Tarreau | 8fe726f | 2021-04-15 18:20:12 +0200 | [diff] [blame] | 273 | */ |
| 274 | void *pool_alloc_nocache(struct pool_head *pool) |
Willy Tarreau | 0bae075 | 2021-03-02 20:05:09 +0100 | [diff] [blame] | 275 | { |
Willy Tarreau | 0bae075 | 2021-03-02 20:05:09 +0100 | [diff] [blame] | 276 | void *ptr = NULL; |
| 277 | |
Willy Tarreau | 1384364 | 2021-04-17 16:57:25 +0200 | [diff] [blame] | 278 | ptr = pool_get_from_os(pool); |
| 279 | if (!ptr) |
Willy Tarreau | 0bae075 | 2021-03-02 20:05:09 +0100 | [diff] [blame] | 280 | return NULL; |
Willy Tarreau | 0bae075 | 2021-03-02 20:05:09 +0100 | [diff] [blame] | 281 | |
Willy Tarreau | 1384364 | 2021-04-17 16:57:25 +0200 | [diff] [blame] | 282 | swrate_add_scaled(&pool->needed_avg, POOL_AVG_SAMPLES, pool->used, POOL_AVG_SAMPLES/4); |
Willy Tarreau | 4781b15 | 2021-04-06 13:53:36 +0200 | [diff] [blame] | 283 | _HA_ATOMIC_INC(&pool->used); |
Willy Tarreau | 0bae075 | 2021-03-02 20:05:09 +0100 | [diff] [blame] | 284 | |
Willy Tarreau | 0bae075 | 2021-03-02 20:05:09 +0100 | [diff] [blame] | 285 | /* keep track of where the element was allocated from */ |
Willy Tarreau | 8c49270 | 2022-01-01 17:10:50 +0100 | [diff] [blame] | 286 | POOL_DEBUG_SET_MARK(pool, ptr); |
Willy Tarreau | 0bae075 | 2021-03-02 20:05:09 +0100 | [diff] [blame] | 287 | return ptr; |
| 288 | } |
| 289 | |
Willy Tarreau | 45e4e28 | 2021-04-17 17:48:40 +0200 | [diff] [blame] | 290 | /* Release a pool item back to the OS and keeps the pool's counters up to date. |
| 291 | * This is always defined even when pools are not enabled (their usage stats |
| 292 | * are maintained). |
| 293 | */ |
| 294 | void pool_free_nocache(struct pool_head *pool, void *ptr) |
| 295 | { |
| 296 | _HA_ATOMIC_DEC(&pool->used); |
| 297 | swrate_add(&pool->needed_avg, POOL_AVG_SAMPLES, pool->used); |
| 298 | pool_put_to_os(pool, ptr); |
| 299 | } |
| 300 | |
Willy Tarreau | b8498e9 | 2021-04-18 10:23:02 +0200 | [diff] [blame] | 301 | |
| 302 | #ifdef CONFIG_HAP_POOLS |
| 303 | |
Willy Tarreau | a0b5831 | 2022-01-02 17:19:14 +0100 | [diff] [blame] | 304 | /* removes up to <count> items from the end of the local pool cache <ph> for |
| 305 | * pool <pool>. The shared pool is refilled with these objects in the limit |
| 306 | * of the number of acceptable objects, and the rest will be released to the |
| 307 | * OS. It is not a problem is <count> is larger than the number of objects in |
| 308 | * the local cache. The counters are automatically updated. |
Willy Tarreau | 8721203 | 2021-04-19 08:14:03 +0200 | [diff] [blame] | 309 | */ |
Willy Tarreau | a0b5831 | 2022-01-02 17:19:14 +0100 | [diff] [blame] | 310 | static void pool_evict_last_items(struct pool_head *pool, struct pool_cache_head *ph, uint count) |
Willy Tarreau | 8721203 | 2021-04-19 08:14:03 +0200 | [diff] [blame] | 311 | { |
Willy Tarreau | 8721203 | 2021-04-19 08:14:03 +0200 | [diff] [blame] | 312 | struct pool_cache_item *item; |
Willy Tarreau | 1513c54 | 2022-01-02 17:53:02 +0100 | [diff] [blame] | 313 | struct pool_item *pi, *head = NULL; |
Willy Tarreau | a0b5831 | 2022-01-02 17:19:14 +0100 | [diff] [blame] | 314 | uint released = 0; |
Willy Tarreau | 1513c54 | 2022-01-02 17:53:02 +0100 | [diff] [blame] | 315 | uint cluster = 0; |
Willy Tarreau | 361e31e | 2022-01-02 00:27:06 +0100 | [diff] [blame] | 316 | uint to_free_max; |
| 317 | |
| 318 | to_free_max = pool_releasable(pool); |
Willy Tarreau | 8721203 | 2021-04-19 08:14:03 +0200 | [diff] [blame] | 319 | |
Willy Tarreau | a0b5831 | 2022-01-02 17:19:14 +0100 | [diff] [blame] | 320 | while (released < count && !LIST_ISEMPTY(&ph->list)) { |
Willy Tarreau | d5ec100 | 2022-01-02 12:40:14 +0100 | [diff] [blame] | 321 | item = LIST_PREV(&ph->list, typeof(item), by_pool); |
Willy Tarreau | 0575d8f | 2022-01-21 19:00:25 +0100 | [diff] [blame] | 322 | pool_check_pattern(ph, item, pool->size); |
Willy Tarreau | 2b71810 | 2021-04-21 07:32:39 +0200 | [diff] [blame] | 323 | LIST_DELETE(&item->by_pool); |
| 324 | LIST_DELETE(&item->by_lru); |
Willy Tarreau | b46674a | 2021-12-30 17:37:33 +0100 | [diff] [blame] | 325 | |
Willy Tarreau | 1513c54 | 2022-01-02 17:53:02 +0100 | [diff] [blame] | 326 | if (to_free_max > released || cluster) { |
Willy Tarreau | 361e31e | 2022-01-02 00:27:06 +0100 | [diff] [blame] | 327 | pi = (struct pool_item *)item; |
Willy Tarreau | 1513c54 | 2022-01-02 17:53:02 +0100 | [diff] [blame] | 328 | pi->next = NULL; |
| 329 | pi->down = head; |
| 330 | head = pi; |
| 331 | cluster++; |
| 332 | if (cluster >= CONFIG_HAP_POOL_CLUSTER_SIZE) { |
| 333 | /* enough to make a cluster */ |
| 334 | pool_put_to_shared_cache(pool, head, cluster); |
| 335 | cluster = 0; |
| 336 | head = NULL; |
| 337 | } |
Willy Tarreau | 361e31e | 2022-01-02 00:27:06 +0100 | [diff] [blame] | 338 | } else |
Willy Tarreau | b46674a | 2021-12-30 17:37:33 +0100 | [diff] [blame] | 339 | pool_free_nocache(pool, item); |
Willy Tarreau | 1513c54 | 2022-01-02 17:53:02 +0100 | [diff] [blame] | 340 | |
| 341 | released++; |
Willy Tarreau | 361e31e | 2022-01-02 00:27:06 +0100 | [diff] [blame] | 342 | } |
| 343 | |
Willy Tarreau | 1513c54 | 2022-01-02 17:53:02 +0100 | [diff] [blame] | 344 | /* incomplete cluster left */ |
| 345 | if (cluster) |
| 346 | pool_put_to_shared_cache(pool, head, cluster); |
| 347 | |
Willy Tarreau | a0b5831 | 2022-01-02 17:19:14 +0100 | [diff] [blame] | 348 | ph->count -= released; |
| 349 | pool_cache_count -= released; |
| 350 | pool_cache_bytes -= released * pool->size; |
| 351 | } |
| 352 | |
| 353 | /* Evicts some of the oldest objects from one local cache, until its number of |
| 354 | * objects is no more than 16+1/8 of the total number of locally cached objects |
| 355 | * or the total size of the local cache is no more than 75% of its maximum (i.e. |
| 356 | * we don't want a single cache to use all the cache for itself). For this, the |
| 357 | * list is scanned in reverse. |
| 358 | */ |
| 359 | void pool_evict_from_local_cache(struct pool_head *pool) |
| 360 | { |
| 361 | struct pool_cache_head *ph = &pool->cache[tid]; |
| 362 | |
Willy Tarreau | 43937e9 | 2022-01-02 17:24:55 +0100 | [diff] [blame] | 363 | while (ph->count >= CONFIG_HAP_POOL_CLUSTER_SIZE && |
| 364 | ph->count >= 16 + pool_cache_count / 8 && |
Willy Tarreau | a0b5831 | 2022-01-02 17:19:14 +0100 | [diff] [blame] | 365 | pool_cache_bytes > CONFIG_HAP_POOL_CACHE_SIZE * 3 / 4) { |
Willy Tarreau | 43937e9 | 2022-01-02 17:24:55 +0100 | [diff] [blame] | 366 | pool_evict_last_items(pool, ph, CONFIG_HAP_POOL_CLUSTER_SIZE); |
Willy Tarreau | 8721203 | 2021-04-19 08:14:03 +0200 | [diff] [blame] | 367 | } |
| 368 | } |
| 369 | |
Willy Tarreau | b8498e9 | 2021-04-18 10:23:02 +0200 | [diff] [blame] | 370 | /* Evicts some of the oldest objects from the local cache, pushing them to the |
| 371 | * global pool. |
| 372 | */ |
| 373 | void pool_evict_from_local_caches() |
| 374 | { |
| 375 | struct pool_cache_item *item; |
| 376 | struct pool_cache_head *ph; |
| 377 | struct pool_head *pool; |
| 378 | |
| 379 | do { |
Willy Tarreau | b4e3476 | 2021-09-30 19:02:18 +0200 | [diff] [blame] | 380 | item = LIST_PREV(&th_ctx->pool_lru_head, struct pool_cache_item *, by_lru); |
Willy Tarreau | b8498e9 | 2021-04-18 10:23:02 +0200 | [diff] [blame] | 381 | /* note: by definition we remove oldest objects so they also are the |
| 382 | * oldest in their own pools, thus their next is the pool's head. |
| 383 | */ |
| 384 | ph = LIST_NEXT(&item->by_pool, struct pool_cache_head *, list); |
| 385 | pool = container_of(ph - tid, struct pool_head, cache); |
Willy Tarreau | 43937e9 | 2022-01-02 17:24:55 +0100 | [diff] [blame] | 386 | pool_evict_last_items(pool, ph, CONFIG_HAP_POOL_CLUSTER_SIZE); |
Willy Tarreau | b8498e9 | 2021-04-18 10:23:02 +0200 | [diff] [blame] | 387 | } while (pool_cache_bytes > CONFIG_HAP_POOL_CACHE_SIZE * 7 / 8); |
| 388 | } |
Willy Tarreau | 0bae075 | 2021-03-02 20:05:09 +0100 | [diff] [blame] | 389 | |
Willy Tarreau | b2a853d | 2021-04-19 11:49:26 +0200 | [diff] [blame] | 390 | /* Frees an object to the local cache, possibly pushing oldest objects to the |
| 391 | * shared cache, which itself may decide to release some of them to the OS. |
| 392 | * While it is unspecified what the object becomes past this point, it is |
| 393 | * guaranteed to be released from the users' perpective. |
| 394 | */ |
| 395 | void pool_put_to_cache(struct pool_head *pool, void *ptr) |
| 396 | { |
| 397 | struct pool_cache_item *item = (struct pool_cache_item *)ptr; |
| 398 | struct pool_cache_head *ph = &pool->cache[tid]; |
| 399 | |
Willy Tarreau | 2b71810 | 2021-04-21 07:32:39 +0200 | [diff] [blame] | 400 | LIST_INSERT(&ph->list, &item->by_pool); |
Willy Tarreau | b4e3476 | 2021-09-30 19:02:18 +0200 | [diff] [blame] | 401 | LIST_INSERT(&th_ctx->pool_lru_head, &item->by_lru); |
Willy Tarreau | b2a853d | 2021-04-19 11:49:26 +0200 | [diff] [blame] | 402 | ph->count++; |
Willy Tarreau | 0575d8f | 2022-01-21 19:00:25 +0100 | [diff] [blame] | 403 | pool_fill_pattern(ph, item, pool->size); |
Willy Tarreau | b2a853d | 2021-04-19 11:49:26 +0200 | [diff] [blame] | 404 | pool_cache_count++; |
| 405 | pool_cache_bytes += pool->size; |
| 406 | |
| 407 | if (unlikely(pool_cache_bytes > CONFIG_HAP_POOL_CACHE_SIZE * 3 / 4)) { |
Willy Tarreau | 43937e9 | 2022-01-02 17:24:55 +0100 | [diff] [blame] | 408 | if (ph->count >= 16 + pool_cache_count / 8 + CONFIG_HAP_POOL_CLUSTER_SIZE) |
Willy Tarreau | b2a853d | 2021-04-19 11:49:26 +0200 | [diff] [blame] | 409 | pool_evict_from_local_cache(pool); |
| 410 | if (pool_cache_bytes > CONFIG_HAP_POOL_CACHE_SIZE) |
| 411 | pool_evict_from_local_caches(); |
| 412 | } |
| 413 | } |
| 414 | |
Willy Tarreau | eb3cc29 | 2021-04-15 18:13:13 +0200 | [diff] [blame] | 415 | #if defined(CONFIG_HAP_NO_GLOBAL_POOLS) |
| 416 | |
Willy Tarreau | 0bae075 | 2021-03-02 20:05:09 +0100 | [diff] [blame] | 417 | /* legacy stuff */ |
| 418 | void pool_flush(struct pool_head *pool) |
| 419 | { |
| 420 | } |
| 421 | |
| 422 | /* This function might ask the malloc library to trim its buffers. */ |
| 423 | void pool_gc(struct pool_head *pool_ctx) |
| 424 | { |
Willy Tarreau | ea3323f | 2021-09-15 10:38:21 +0200 | [diff] [blame] | 425 | trim_all_pools(); |
Willy Tarreau | 0bae075 | 2021-03-02 20:05:09 +0100 | [diff] [blame] | 426 | } |
| 427 | |
Willy Tarreau | 9b3ed51 | 2021-06-10 10:21:35 +0200 | [diff] [blame] | 428 | #else /* CONFIG_HAP_NO_GLOBAL_POOLS */ |
| 429 | |
Willy Tarreau | afe2c4a | 2021-12-30 17:09:31 +0100 | [diff] [blame] | 430 | /* Tries to refill the local cache <pch> from the shared one for pool <pool>. |
| 431 | * This is only used when pools are in use and shared pools are enabled. No |
| 432 | * malloc() is attempted, and poisonning is never performed. The purpose is to |
| 433 | * get the fastest possible refilling so that the caller can easily check if |
| 434 | * the cache has enough objects for its use. |
| 435 | */ |
| 436 | void pool_refill_local_from_shared(struct pool_head *pool, struct pool_cache_head *pch) |
| 437 | { |
| 438 | struct pool_cache_item *item; |
Willy Tarreau | 148160b | 2022-01-02 14:35:57 +0100 | [diff] [blame] | 439 | struct pool_item *ret, *down; |
| 440 | uint count; |
Willy Tarreau | afe2c4a | 2021-12-30 17:09:31 +0100 | [diff] [blame] | 441 | |
| 442 | /* we'll need to reference the first element to figure the next one. We |
| 443 | * must temporarily lock it so that nobody allocates then releases it, |
| 444 | * or the dereference could fail. |
| 445 | */ |
| 446 | ret = _HA_ATOMIC_LOAD(&pool->free_list); |
| 447 | do { |
| 448 | while (unlikely(ret == POOL_BUSY)) { |
| 449 | __ha_cpu_relax(); |
| 450 | ret = _HA_ATOMIC_LOAD(&pool->free_list); |
| 451 | } |
| 452 | if (ret == NULL) |
| 453 | return; |
| 454 | } while (unlikely((ret = _HA_ATOMIC_XCHG(&pool->free_list, POOL_BUSY)) == POOL_BUSY)); |
| 455 | |
| 456 | if (unlikely(ret == NULL)) { |
| 457 | HA_ATOMIC_STORE(&pool->free_list, NULL); |
| 458 | return; |
| 459 | } |
| 460 | |
| 461 | /* this releases the lock */ |
Willy Tarreau | c16ed3b | 2022-01-01 18:22:20 +0100 | [diff] [blame] | 462 | HA_ATOMIC_STORE(&pool->free_list, ret->next); |
Willy Tarreau | afe2c4a | 2021-12-30 17:09:31 +0100 | [diff] [blame] | 463 | |
Willy Tarreau | 148160b | 2022-01-02 14:35:57 +0100 | [diff] [blame] | 464 | /* now store the retrieved object(s) into the local cache */ |
| 465 | count = 0; |
| 466 | for (; ret; ret = down) { |
| 467 | down = ret->down; |
| 468 | /* keep track of where the element was allocated from */ |
| 469 | POOL_DEBUG_SET_MARK(pool, ret); |
Willy Tarreau | afe2c4a | 2021-12-30 17:09:31 +0100 | [diff] [blame] | 470 | |
Willy Tarreau | 148160b | 2022-01-02 14:35:57 +0100 | [diff] [blame] | 471 | item = (struct pool_cache_item *)ret; |
| 472 | LIST_INSERT(&pch->list, &item->by_pool); |
| 473 | LIST_INSERT(&th_ctx->pool_lru_head, &item->by_lru); |
| 474 | count++; |
Willy Tarreau | 0575d8f | 2022-01-21 19:00:25 +0100 | [diff] [blame] | 475 | pool_fill_pattern(pch, item, pool->size); |
Willy Tarreau | 148160b | 2022-01-02 14:35:57 +0100 | [diff] [blame] | 476 | } |
| 477 | HA_ATOMIC_ADD(&pool->used, count); |
| 478 | pch->count += count; |
| 479 | pool_cache_count += count; |
| 480 | pool_cache_bytes += count * pool->size; |
Willy Tarreau | afe2c4a | 2021-12-30 17:09:31 +0100 | [diff] [blame] | 481 | } |
| 482 | |
Willy Tarreau | 337410c | 2022-01-02 15:15:54 +0100 | [diff] [blame] | 483 | /* Adds pool item cluster <item> to the shared cache, which contains <count> |
| 484 | * elements. The caller is advised to first check using pool_releasable() if |
| 485 | * it's wise to add this series of objects there. Both the pool and the item's |
| 486 | * head must be valid. |
Willy Tarreau | b46674a | 2021-12-30 17:37:33 +0100 | [diff] [blame] | 487 | */ |
Willy Tarreau | 337410c | 2022-01-02 15:15:54 +0100 | [diff] [blame] | 488 | void pool_put_to_shared_cache(struct pool_head *pool, struct pool_item *item, uint count) |
Willy Tarreau | b46674a | 2021-12-30 17:37:33 +0100 | [diff] [blame] | 489 | { |
Willy Tarreau | c16ed3b | 2022-01-01 18:22:20 +0100 | [diff] [blame] | 490 | struct pool_item *free_list; |
Willy Tarreau | b46674a | 2021-12-30 17:37:33 +0100 | [diff] [blame] | 491 | |
Willy Tarreau | 337410c | 2022-01-02 15:15:54 +0100 | [diff] [blame] | 492 | _HA_ATOMIC_SUB(&pool->used, count); |
Willy Tarreau | b46674a | 2021-12-30 17:37:33 +0100 | [diff] [blame] | 493 | free_list = _HA_ATOMIC_LOAD(&pool->free_list); |
| 494 | do { |
| 495 | while (unlikely(free_list == POOL_BUSY)) { |
| 496 | __ha_cpu_relax(); |
| 497 | free_list = _HA_ATOMIC_LOAD(&pool->free_list); |
| 498 | } |
Willy Tarreau | c16ed3b | 2022-01-01 18:22:20 +0100 | [diff] [blame] | 499 | _HA_ATOMIC_STORE(&item->next, free_list); |
Willy Tarreau | b46674a | 2021-12-30 17:37:33 +0100 | [diff] [blame] | 500 | __ha_barrier_atomic_store(); |
| 501 | } while (!_HA_ATOMIC_CAS(&pool->free_list, &free_list, item)); |
| 502 | __ha_barrier_atomic_store(); |
| 503 | swrate_add(&pool->needed_avg, POOL_AVG_SAMPLES, pool->used); |
| 504 | } |
| 505 | |
Olivier Houchard | cf975d4 | 2018-01-24 18:38:31 +0100 | [diff] [blame] | 506 | /* |
| 507 | * This function frees whatever can be freed in pool <pool>. |
| 508 | */ |
| 509 | void pool_flush(struct pool_head *pool) |
| 510 | { |
Willy Tarreau | 148160b | 2022-01-02 14:35:57 +0100 | [diff] [blame] | 511 | struct pool_item *next, *temp, *down; |
Olivier Houchard | cf975d4 | 2018-01-24 18:38:31 +0100 | [diff] [blame] | 512 | |
| 513 | if (!pool) |
| 514 | return; |
Willy Tarreau | 2a4523f | 2021-06-09 18:59:58 +0200 | [diff] [blame] | 515 | |
| 516 | /* The loop below atomically detaches the head of the free list and |
| 517 | * replaces it with a NULL. Then the list can be released. |
| 518 | */ |
| 519 | next = pool->free_list; |
Olivier Houchard | cf975d4 | 2018-01-24 18:38:31 +0100 | [diff] [blame] | 520 | do { |
Willy Tarreau | 2a4523f | 2021-06-09 18:59:58 +0200 | [diff] [blame] | 521 | while (unlikely(next == POOL_BUSY)) { |
| 522 | __ha_cpu_relax(); |
| 523 | next = _HA_ATOMIC_LOAD(&pool->free_list); |
| 524 | } |
| 525 | if (next == NULL) |
| 526 | return; |
| 527 | } while (unlikely((next = _HA_ATOMIC_XCHG(&pool->free_list, POOL_BUSY)) == POOL_BUSY)); |
| 528 | _HA_ATOMIC_STORE(&pool->free_list, NULL); |
Olivier Houchard | 2087276 | 2019-03-08 18:53:35 +0100 | [diff] [blame] | 529 | __ha_barrier_atomic_store(); |
Willy Tarreau | 2a4523f | 2021-06-09 18:59:58 +0200 | [diff] [blame] | 530 | |
Olivier Houchard | cf975d4 | 2018-01-24 18:38:31 +0100 | [diff] [blame] | 531 | while (next) { |
| 532 | temp = next; |
Willy Tarreau | c16ed3b | 2022-01-01 18:22:20 +0100 | [diff] [blame] | 533 | next = temp->next; |
Willy Tarreau | 148160b | 2022-01-02 14:35:57 +0100 | [diff] [blame] | 534 | for (; temp; temp = down) { |
| 535 | down = temp->down; |
| 536 | pool_put_to_os(pool, temp); |
| 537 | } |
Olivier Houchard | cf975d4 | 2018-01-24 18:38:31 +0100 | [diff] [blame] | 538 | } |
Willy Tarreau | c239cde | 2021-06-10 06:54:22 +0200 | [diff] [blame] | 539 | /* here, we should have pool->allocated == pool->used */ |
Olivier Houchard | cf975d4 | 2018-01-24 18:38:31 +0100 | [diff] [blame] | 540 | } |
| 541 | |
Willy Tarreau | e6ce59d | 2007-05-13 19:38:49 +0200 | [diff] [blame] | 542 | /* |
Willy Tarreau | e6ce59d | 2007-05-13 19:38:49 +0200 | [diff] [blame] | 543 | * This function frees whatever can be freed in all pools, but respecting |
Willy Tarreau | c0e2ff2 | 2020-04-24 06:15:24 +0200 | [diff] [blame] | 544 | * the minimum thresholds imposed by owners. It makes sure to be alone to |
| 545 | * run by using thread_isolate(). <pool_ctx> is unused. |
Willy Tarreau | e6ce59d | 2007-05-13 19:38:49 +0200 | [diff] [blame] | 546 | */ |
Willy Tarreau | bafbe01 | 2017-11-24 17:34:44 +0100 | [diff] [blame] | 547 | void pool_gc(struct pool_head *pool_ctx) |
Willy Tarreau | e6ce59d | 2007-05-13 19:38:49 +0200 | [diff] [blame] | 548 | { |
| 549 | struct pool_head *entry; |
Willy Tarreau | c0e2ff2 | 2020-04-24 06:15:24 +0200 | [diff] [blame] | 550 | int isolated = thread_isolated(); |
Willy Tarreau | b7f9d12 | 2009-04-21 02:17:45 +0200 | [diff] [blame] | 551 | |
Willy Tarreau | c0e2ff2 | 2020-04-24 06:15:24 +0200 | [diff] [blame] | 552 | if (!isolated) |
| 553 | thread_isolate(); |
Willy Tarreau | b7f9d12 | 2009-04-21 02:17:45 +0200 | [diff] [blame] | 554 | |
Willy Tarreau | e6ce59d | 2007-05-13 19:38:49 +0200 | [diff] [blame] | 555 | list_for_each_entry(entry, &pools, list) { |
Willy Tarreau | 148160b | 2022-01-02 14:35:57 +0100 | [diff] [blame] | 556 | struct pool_item *temp, *down; |
Willy Tarreau | c16ed3b | 2022-01-01 18:22:20 +0100 | [diff] [blame] | 557 | |
Olivier Houchard | 51d9339 | 2020-03-12 19:05:39 +0100 | [diff] [blame] | 558 | while (entry->free_list && |
Willy Tarreau | 57767b8 | 2014-12-22 21:40:55 +0100 | [diff] [blame] | 559 | (int)(entry->allocated - entry->used) > (int)entry->minavail) { |
Olivier Houchard | 51d9339 | 2020-03-12 19:05:39 +0100 | [diff] [blame] | 560 | temp = entry->free_list; |
Willy Tarreau | c16ed3b | 2022-01-01 18:22:20 +0100 | [diff] [blame] | 561 | entry->free_list = temp->next; |
Willy Tarreau | 148160b | 2022-01-02 14:35:57 +0100 | [diff] [blame] | 562 | for (; temp; temp = down) { |
| 563 | down = temp->down; |
| 564 | pool_put_to_os(entry, temp); |
| 565 | } |
Willy Tarreau | e6ce59d | 2007-05-13 19:38:49 +0200 | [diff] [blame] | 566 | } |
Willy Tarreau | e6ce59d | 2007-05-13 19:38:49 +0200 | [diff] [blame] | 567 | } |
Christopher Faulet | b349e48 | 2017-08-29 09:52:38 +0200 | [diff] [blame] | 568 | |
Willy Tarreau | ea3323f | 2021-09-15 10:38:21 +0200 | [diff] [blame] | 569 | trim_all_pools(); |
Willy Tarreau | 26ed183 | 2021-06-10 08:40:16 +0200 | [diff] [blame] | 570 | |
Willy Tarreau | c0e2ff2 | 2020-04-24 06:15:24 +0200 | [diff] [blame] | 571 | if (!isolated) |
| 572 | thread_release(); |
Willy Tarreau | e6ce59d | 2007-05-13 19:38:49 +0200 | [diff] [blame] | 573 | } |
Willy Tarreau | 9b3ed51 | 2021-06-10 10:21:35 +0200 | [diff] [blame] | 574 | #endif /* CONFIG_HAP_NO_GLOBAL_POOLS */ |
Willy Tarreau | b8498e9 | 2021-04-18 10:23:02 +0200 | [diff] [blame] | 575 | |
| 576 | #else /* CONFIG_HAP_POOLS */ |
| 577 | |
| 578 | /* legacy stuff */ |
| 579 | void pool_flush(struct pool_head *pool) |
| 580 | { |
| 581 | } |
| 582 | |
| 583 | /* This function might ask the malloc library to trim its buffers. */ |
| 584 | void pool_gc(struct pool_head *pool_ctx) |
| 585 | { |
Willy Tarreau | ea3323f | 2021-09-15 10:38:21 +0200 | [diff] [blame] | 586 | trim_all_pools(); |
Willy Tarreau | b8498e9 | 2021-04-18 10:23:02 +0200 | [diff] [blame] | 587 | } |
| 588 | |
| 589 | #endif /* CONFIG_HAP_POOLS */ |
Willy Tarreau | e6ce59d | 2007-05-13 19:38:49 +0200 | [diff] [blame] | 590 | |
Willy Tarreau | f14d190 | 2021-10-05 18:14:11 +0200 | [diff] [blame] | 591 | |
| 592 | #ifdef DEBUG_UAF |
| 593 | |
| 594 | /************* use-after-free allocator *************/ |
| 595 | |
| 596 | /* allocates an area of size <size> and returns it. The semantics are similar |
| 597 | * to those of malloc(). However the allocation is rounded up to 4kB so that a |
| 598 | * full page is allocated. This ensures the object can be freed alone so that |
| 599 | * future dereferences are easily detected. The returned object is always |
| 600 | * 16-bytes aligned to avoid issues with unaligned structure objects. In case |
| 601 | * some padding is added, the area's start address is copied at the end of the |
| 602 | * padding to help detect underflows. |
| 603 | */ |
| 604 | void *pool_alloc_area_uaf(size_t size) |
| 605 | { |
| 606 | size_t pad = (4096 - size) & 0xFF0; |
Willy Tarreau | f14d190 | 2021-10-05 18:14:11 +0200 | [diff] [blame] | 607 | void *ret; |
| 608 | |
Willy Tarreau | f14d190 | 2021-10-05 18:14:11 +0200 | [diff] [blame] | 609 | ret = mmap(NULL, (size + 4095) & -4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0); |
| 610 | if (ret != MAP_FAILED) { |
| 611 | /* let's dereference the page before returning so that the real |
| 612 | * allocation in the system is performed without holding the lock. |
| 613 | */ |
| 614 | *(int *)ret = 0; |
| 615 | if (pad >= sizeof(void *)) |
| 616 | *(void **)(ret + pad - sizeof(void *)) = ret + pad; |
| 617 | ret += pad; |
| 618 | } else { |
| 619 | ret = NULL; |
| 620 | } |
Willy Tarreau | f14d190 | 2021-10-05 18:14:11 +0200 | [diff] [blame] | 621 | return ret; |
| 622 | } |
| 623 | |
| 624 | /* frees an area <area> of size <size> allocated by pool_alloc_area(). The |
| 625 | * semantics are identical to free() except that the size must absolutely match |
| 626 | * the one passed to pool_alloc_area(). In case some padding is added, the |
| 627 | * area's start address is compared to the one at the end of the padding, and |
| 628 | * a segfault is triggered if they don't match, indicating an underflow. |
| 629 | */ |
| 630 | void pool_free_area_uaf(void *area, size_t size) |
| 631 | { |
| 632 | size_t pad = (4096 - size) & 0xFF0; |
| 633 | |
| 634 | if (pad >= sizeof(void *) && *(void **)(area - sizeof(void *)) != area) |
| 635 | ABORT_NOW(); |
| 636 | |
Willy Tarreau | f14d190 | 2021-10-05 18:14:11 +0200 | [diff] [blame] | 637 | munmap(area - pad, (size + 4095) & -4096); |
Willy Tarreau | f14d190 | 2021-10-05 18:14:11 +0200 | [diff] [blame] | 638 | } |
| 639 | |
| 640 | #endif /* DEBUG_UAF */ |
| 641 | |
Willy Tarreau | e6ce59d | 2007-05-13 19:38:49 +0200 | [diff] [blame] | 642 | /* |
Willy Tarreau | dae4aa8 | 2007-06-16 23:19:53 +0200 | [diff] [blame] | 643 | * This function destroys a pool by freeing it completely, unless it's still |
| 644 | * in use. This should be called only under extreme circumstances. It always |
| 645 | * returns NULL if the resulting pool is empty, easing the clearing of the old |
| 646 | * pointer, otherwise it returns the pool. |
| 647 | * . |
Willy Tarreau | e6ce59d | 2007-05-13 19:38:49 +0200 | [diff] [blame] | 648 | */ |
Willy Tarreau | bafbe01 | 2017-11-24 17:34:44 +0100 | [diff] [blame] | 649 | void *pool_destroy(struct pool_head *pool) |
Willy Tarreau | e6ce59d | 2007-05-13 19:38:49 +0200 | [diff] [blame] | 650 | { |
Willy Tarreau | 4d2d098 | 2007-05-14 00:39:29 +0200 | [diff] [blame] | 651 | if (pool) { |
Willy Tarreau | bafbe01 | 2017-11-24 17:34:44 +0100 | [diff] [blame] | 652 | pool_flush(pool); |
Willy Tarreau | dae4aa8 | 2007-06-16 23:19:53 +0200 | [diff] [blame] | 653 | if (pool->used) |
| 654 | return pool; |
| 655 | pool->users--; |
| 656 | if (!pool->users) { |
Willy Tarreau | 2b71810 | 2021-04-21 07:32:39 +0200 | [diff] [blame] | 657 | LIST_DELETE(&pool->list); |
Willy Tarreau | 9f3129e | 2021-04-17 00:31:38 +0200 | [diff] [blame] | 658 | /* note that if used == 0, the cache is empty */ |
| 659 | free(pool); |
Willy Tarreau | dae4aa8 | 2007-06-16 23:19:53 +0200 | [diff] [blame] | 660 | } |
Willy Tarreau | 4d2d098 | 2007-05-14 00:39:29 +0200 | [diff] [blame] | 661 | } |
| 662 | return NULL; |
Willy Tarreau | e6ce59d | 2007-05-13 19:38:49 +0200 | [diff] [blame] | 663 | } |
| 664 | |
Willy Tarreau | 2455ceb | 2018-11-26 15:57:34 +0100 | [diff] [blame] | 665 | /* This destroys all pools on exit. It is *not* thread safe. */ |
| 666 | void pool_destroy_all() |
| 667 | { |
| 668 | struct pool_head *entry, *back; |
| 669 | |
| 670 | list_for_each_entry_safe(entry, back, &pools, list) |
| 671 | pool_destroy(entry); |
| 672 | } |
| 673 | |
Willy Tarreau | 12833bb | 2014-01-28 16:49:56 +0100 | [diff] [blame] | 674 | /* This function dumps memory usage information into the trash buffer. */ |
| 675 | void dump_pools_to_trash() |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 676 | { |
| 677 | struct pool_head *entry; |
| 678 | unsigned long allocated, used; |
| 679 | int nbpools; |
Willy Tarreau | 1b4a714 | 2021-10-07 16:29:31 +0200 | [diff] [blame] | 680 | #ifdef CONFIG_HAP_POOLS |
| 681 | unsigned long cached_bytes = 0; |
| 682 | uint cached = 0; |
| 683 | #endif |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 684 | |
| 685 | allocated = used = nbpools = 0; |
Willy Tarreau | 12833bb | 2014-01-28 16:49:56 +0100 | [diff] [blame] | 686 | chunk_printf(&trash, "Dumping pools usage. Use SIGQUIT to flush them.\n"); |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 687 | list_for_each_entry(entry, &pools, list) { |
Willy Tarreau | 1b4a714 | 2021-10-07 16:29:31 +0200 | [diff] [blame] | 688 | #ifdef CONFIG_HAP_POOLS |
| 689 | int i; |
| 690 | for (cached = i = 0; i < global.nbthread; i++) |
| 691 | cached += entry->cache[i].count; |
| 692 | cached_bytes += cached * entry->size; |
| 693 | #endif |
| 694 | chunk_appendf(&trash, " - Pool %s (%u bytes) : %u allocated (%u bytes), %u used" |
| 695 | #ifdef CONFIG_HAP_POOLS |
| 696 | " (~%u by thread caches)" |
| 697 | #endif |
| 698 | ", needed_avg %u, %u failures, %u users, @%p%s\n", |
| 699 | entry->name, entry->size, entry->allocated, |
| 700 | entry->size * entry->allocated, entry->used, |
| 701 | #ifdef CONFIG_HAP_POOLS |
| 702 | cached, |
| 703 | #endif |
| 704 | swrate_avg(entry->needed_avg, POOL_AVG_SAMPLES), entry->failed, |
| 705 | entry->users, entry, |
| 706 | (entry->flags & MEM_F_SHARED) ? " [SHARED]" : ""); |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 707 | |
| 708 | allocated += entry->allocated * entry->size; |
| 709 | used += entry->used * entry->size; |
| 710 | nbpools++; |
| 711 | } |
Willy Tarreau | 1b4a714 | 2021-10-07 16:29:31 +0200 | [diff] [blame] | 712 | chunk_appendf(&trash, "Total: %d pools, %lu bytes allocated, %lu used" |
| 713 | #ifdef CONFIG_HAP_POOLS |
| 714 | " (~%lu by thread caches)" |
| 715 | #endif |
| 716 | ".\n", |
| 717 | nbpools, allocated, used |
| 718 | #ifdef CONFIG_HAP_POOLS |
| 719 | , cached_bytes |
| 720 | #endif |
| 721 | ); |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 722 | } |
| 723 | |
Willy Tarreau | 12833bb | 2014-01-28 16:49:56 +0100 | [diff] [blame] | 724 | /* Dump statistics on pools usage. */ |
| 725 | void dump_pools(void) |
| 726 | { |
| 727 | dump_pools_to_trash(); |
Willy Tarreau | 843b7cb | 2018-07-13 10:54:26 +0200 | [diff] [blame] | 728 | qfprintf(stderr, "%s", trash.area); |
Willy Tarreau | 12833bb | 2014-01-28 16:49:56 +0100 | [diff] [blame] | 729 | } |
| 730 | |
Willy Tarreau | 58102cf | 2015-10-28 16:24:21 +0100 | [diff] [blame] | 731 | /* This function returns the total number of failed pool allocations */ |
| 732 | int pool_total_failures() |
| 733 | { |
| 734 | struct pool_head *entry; |
| 735 | int failed = 0; |
| 736 | |
| 737 | list_for_each_entry(entry, &pools, list) |
| 738 | failed += entry->failed; |
| 739 | return failed; |
| 740 | } |
| 741 | |
| 742 | /* This function returns the total amount of memory allocated in pools (in bytes) */ |
| 743 | unsigned long pool_total_allocated() |
| 744 | { |
| 745 | struct pool_head *entry; |
| 746 | unsigned long allocated = 0; |
| 747 | |
| 748 | list_for_each_entry(entry, &pools, list) |
| 749 | allocated += entry->allocated * entry->size; |
| 750 | return allocated; |
| 751 | } |
| 752 | |
| 753 | /* This function returns the total amount of memory used in pools (in bytes) */ |
| 754 | unsigned long pool_total_used() |
| 755 | { |
| 756 | struct pool_head *entry; |
| 757 | unsigned long used = 0; |
| 758 | |
| 759 | list_for_each_entry(entry, &pools, list) |
| 760 | used += entry->used * entry->size; |
| 761 | return used; |
| 762 | } |
| 763 | |
William Lallemand | e7ed885 | 2016-11-19 02:25:36 +0100 | [diff] [blame] | 764 | /* This function dumps memory usage information onto the stream interface's |
| 765 | * read buffer. It returns 0 as long as it does not complete, non-zero upon |
| 766 | * completion. No state is used. |
| 767 | */ |
| 768 | static int cli_io_handler_dump_pools(struct appctx *appctx) |
| 769 | { |
| 770 | struct stream_interface *si = appctx->owner; |
| 771 | |
| 772 | dump_pools_to_trash(); |
Willy Tarreau | 06d80a9 | 2017-10-19 14:32:15 +0200 | [diff] [blame] | 773 | if (ci_putchk(si_ic(si), &trash) == -1) { |
Willy Tarreau | db39843 | 2018-11-15 11:08:52 +0100 | [diff] [blame] | 774 | si_rx_room_blk(si); |
William Lallemand | e7ed885 | 2016-11-19 02:25:36 +0100 | [diff] [blame] | 775 | return 0; |
| 776 | } |
| 777 | return 1; |
| 778 | } |
| 779 | |
Willy Tarreau | 7107c8b | 2018-11-26 11:44:35 +0100 | [diff] [blame] | 780 | /* callback used to create early pool <name> of size <size> and store the |
| 781 | * resulting pointer into <ptr>. If the allocation fails, it quits with after |
| 782 | * emitting an error message. |
| 783 | */ |
| 784 | void create_pool_callback(struct pool_head **ptr, char *name, unsigned int size) |
| 785 | { |
| 786 | *ptr = create_pool(name, size, MEM_F_SHARED); |
| 787 | if (!*ptr) { |
| 788 | ha_alert("Failed to allocate pool '%s' of size %u : %s. Aborting.\n", |
| 789 | name, size, strerror(errno)); |
| 790 | exit(1); |
| 791 | } |
| 792 | } |
| 793 | |
Willy Tarreau | 7f0165e | 2018-11-26 17:09:46 +0100 | [diff] [blame] | 794 | /* Initializes all per-thread arrays on startup */ |
| 795 | static void init_pools() |
| 796 | { |
Willy Tarreau | 2d6f628 | 2021-04-15 16:24:00 +0200 | [diff] [blame] | 797 | #ifdef CONFIG_HAP_POOLS |
Willy Tarreau | 9f3129e | 2021-04-17 00:31:38 +0200 | [diff] [blame] | 798 | int thr; |
Willy Tarreau | 7f0165e | 2018-11-26 17:09:46 +0100 | [diff] [blame] | 799 | |
| 800 | for (thr = 0; thr < MAX_THREADS; thr++) { |
Willy Tarreau | b4e3476 | 2021-09-30 19:02:18 +0200 | [diff] [blame] | 801 | LIST_INIT(&ha_thread_ctx[thr].pool_lru_head); |
Willy Tarreau | 7f0165e | 2018-11-26 17:09:46 +0100 | [diff] [blame] | 802 | } |
Willy Tarreau | ed891fd | 2020-06-01 19:00:28 +0200 | [diff] [blame] | 803 | #endif |
Willy Tarreau | 157e393 | 2021-09-15 10:05:48 +0200 | [diff] [blame] | 804 | detect_allocator(); |
Willy Tarreau | 7f0165e | 2018-11-26 17:09:46 +0100 | [diff] [blame] | 805 | } |
| 806 | |
| 807 | INITCALL0(STG_PREPARE, init_pools); |
Willy Tarreau | 7107c8b | 2018-11-26 11:44:35 +0100 | [diff] [blame] | 808 | |
Willy Tarreau | 845b560 | 2021-09-15 10:41:24 +0200 | [diff] [blame] | 809 | /* Report in build options if trim is supported */ |
| 810 | static void pools_register_build_options(void) |
| 811 | { |
| 812 | if (is_trim_enabled()) { |
| 813 | char *ptr = NULL; |
| 814 | memprintf(&ptr, "Support for malloc_trim() is enabled."); |
| 815 | hap_register_build_opts(ptr, 1); |
| 816 | } |
| 817 | } |
| 818 | INITCALL0(STG_REGISTER, pools_register_build_options); |
| 819 | |
William Lallemand | e7ed885 | 2016-11-19 02:25:36 +0100 | [diff] [blame] | 820 | /* register cli keywords */ |
| 821 | static struct cli_kw_list cli_kws = {{ },{ |
Willy Tarreau | b205bfd | 2021-05-07 11:38:37 +0200 | [diff] [blame] | 822 | { { "show", "pools", NULL }, "show pools : report information about the memory pools usage", NULL, cli_io_handler_dump_pools }, |
William Lallemand | e7ed885 | 2016-11-19 02:25:36 +0100 | [diff] [blame] | 823 | {{},} |
| 824 | }}; |
| 825 | |
Willy Tarreau | 0108d90 | 2018-11-25 19:14:37 +0100 | [diff] [blame] | 826 | INITCALL1(STG_REGISTER, cli_register_kw, &cli_kws); |
William Lallemand | e7ed885 | 2016-11-19 02:25:36 +0100 | [diff] [blame] | 827 | |
Olivier Houchard | dc21ff7 | 2019-01-29 15:20:16 +0100 | [diff] [blame] | 828 | #ifdef DEBUG_FAIL_ALLOC |
Olivier Houchard | dc21ff7 | 2019-01-29 15:20:16 +0100 | [diff] [blame] | 829 | |
| 830 | int mem_should_fail(const struct pool_head *pool) |
| 831 | { |
Olivier Houchard | 9c4f08a | 2019-02-01 16:28:04 +0100 | [diff] [blame] | 832 | int ret = 0; |
Olivier Houchard | dc21ff7 | 2019-01-29 15:20:16 +0100 | [diff] [blame] | 833 | |
| 834 | if (mem_fail_rate > 0 && !(global.mode & MODE_STARTING)) { |
Willy Tarreau | 20f88ab | 2021-04-17 15:50:28 +0200 | [diff] [blame] | 835 | if (mem_fail_rate > statistical_prng_range(100)) |
Olivier Houchard | dc21ff7 | 2019-01-29 15:20:16 +0100 | [diff] [blame] | 836 | ret = 1; |
| 837 | else |
| 838 | ret = 0; |
| 839 | } |
Olivier Houchard | dc21ff7 | 2019-01-29 15:20:16 +0100 | [diff] [blame] | 840 | return ret; |
| 841 | |
| 842 | } |
| 843 | |
| 844 | /* config parser for global "tune.fail-alloc" */ |
| 845 | static int mem_parse_global_fail_alloc(char **args, int section_type, struct proxy *curpx, |
Amaury Denoyelle | 3b1c9a3 | 2021-03-22 11:21:36 +0100 | [diff] [blame] | 846 | const struct proxy *defpx, const char *file, int line, |
| 847 | char **err) |
Olivier Houchard | dc21ff7 | 2019-01-29 15:20:16 +0100 | [diff] [blame] | 848 | { |
| 849 | if (too_many_args(1, args, err, NULL)) |
| 850 | return -1; |
| 851 | mem_fail_rate = atoi(args[1]); |
| 852 | if (mem_fail_rate < 0 || mem_fail_rate > 100) { |
| 853 | memprintf(err, "'%s' expects a numeric value between 0 and 100.", args[0]); |
| 854 | return -1; |
| 855 | } |
| 856 | return 0; |
| 857 | } |
| 858 | #endif |
| 859 | |
| 860 | /* register global config keywords */ |
| 861 | static struct cfg_kw_list mem_cfg_kws = {ILH, { |
| 862 | #ifdef DEBUG_FAIL_ALLOC |
| 863 | { CFG_GLOBAL, "tune.fail-alloc", mem_parse_global_fail_alloc }, |
| 864 | #endif |
| 865 | { 0, NULL, NULL } |
| 866 | }}; |
| 867 | |
| 868 | INITCALL1(STG_REGISTER, cfg_register_keywords, &mem_cfg_kws); |
| 869 | |
Willy Tarreau | 50e608d | 2007-05-13 18:26:08 +0200 | [diff] [blame] | 870 | /* |
| 871 | * Local variables: |
| 872 | * c-indent-level: 8 |
| 873 | * c-basic-offset: 8 |
| 874 | * End: |
| 875 | */ |