| 1 |
/* -*- Mode: C; tab-width: 4; c-basic-offset: 4; indent-tabs-mode: nil -*- */ |
| 2 |
/* |
| 3 |
* Slabs memory allocation, based on powers-of-N. Slabs are up to 1MB in size |
| 4 |
* and are divided into chunks. The chunk sizes start off at the size of the |
| 5 |
* "item" structure plus space for a small key and value. They increase by |
| 6 |
* a multiplier factor from there, up to half the maximum slab size. The last |
| 7 |
* slab size is always 1MB, since that's the maximum item size allowed by the |
| 8 |
* memcached protocol. |
| 9 |
*/ |
| 10 |
#include "config.h" |
| 11 |
|
| 12 |
#include <fcntl.h> |
| 13 |
#include <errno.h> |
| 14 |
#include <stdlib.h> |
| 15 |
#include <stdio.h> |
| 16 |
#include <string.h> |
| 17 |
#include <assert.h> |
| 18 |
#include <pthread.h> |
| 19 |
#include <inttypes.h> |
| 20 |
#include <stdarg.h> |
| 21 |
|
| 22 |
#include "default_engine.h" |
| 23 |
|
| 24 |
/* |
| 25 |
* Forward Declarations |
| 26 |
*/ |
| 27 |
static int do_slabs_newslab(struct default_engine *engine, const unsigned int id); |
| 28 |
static void *memory_allocate(struct default_engine *engine, size_t size); |
| 29 |
|
| 30 |
#ifndef DONT_PREALLOC_SLABS |
| 31 |
/* Preallocate as many slab pages as possible (called from slabs_init) |
| 32 |
on start-up, so users don't get confused out-of-memory errors when |
| 33 |
they do have free (in-slab) space, but no space to make new slabs. |
| 34 |
if maxslabs is 18 (POWER_LARGEST - POWER_SMALLEST + 1), then all |
| 35 |
slab types can be made. if max memory is less than 18 MB, only the |
| 36 |
smaller ones will be made. */ |
| 37 |
static void slabs_preallocate (const unsigned int maxslabs); |
| 38 |
#endif |
| 39 |
|
| 40 |
/* |
| 41 |
* Figures out which slab class (chunk size) is required to store an item of |
| 42 |
* a given size. |
| 43 |
* |
| 44 |
* Given object size, return id to use when allocating/freeing memory for object |
| 45 |
* 0 means error: can't store such a large object |
| 46 |
*/ |
| 47 |
|
| 48 |
unsigned int slabs_clsid(struct default_engine *engine, const size_t size) { |
| 49 |
int res = POWER_SMALLEST; |
| 50 |
|
| 51 |
if (size == 0) |
| 52 |
return 0; |
| 53 |
while (size > engine->slabs.slabclass[res].size) |
| 54 |
if (res++ == engine->slabs.power_largest) /* won't fit in the biggest slab */ |
| 55 |
return 0; |
| 56 |
return res; |
| 57 |
} |
| 58 |
|
| 59 |
/** |
| 60 |
* Determines the chunk sizes and initializes the slab class descriptors |
| 61 |
* accordingly. |
| 62 |
*/ |
| 63 |
ENGINE_ERROR_CODE slabs_init(struct default_engine *engine, |
| 64 |
const size_t limit, |
| 65 |
const double factor, |
| 66 |
const bool prealloc) { |
| 67 |
int i = POWER_SMALLEST - 1; |
| 68 |
unsigned int size = sizeof(hash_item) + engine->config.chunk_size; |
| 69 |
|
| 70 |
engine->slabs.mem_limit = limit; |
| 71 |
|
| 72 |
if (prealloc) { |
| 73 |
/* Allocate everything in a big chunk with malloc */ |
| 74 |
engine->slabs.mem_base = malloc(engine->slabs.mem_limit); |
| 75 |
if (engine->slabs.mem_base != NULL) { |
| 76 |
engine->slabs.mem_current = engine->slabs.mem_base; |
| 77 |
engine->slabs.mem_avail = engine->slabs.mem_limit; |
| 78 |
} else { |
| 79 |
return ENGINE_ENOMEM; |
| 80 |
} |
| 81 |
} |
| 82 |
|
| 83 |
memset(engine->slabs.slabclass, 0, sizeof(engine->slabs.slabclass)); |
| 84 |
|
| 85 |
while (++i < POWER_LARGEST && size <= engine->config.item_size_max / factor) { |
| 86 |
/* Make sure items are always n-byte aligned */ |
| 87 |
if (size % CHUNK_ALIGN_BYTES) |
| 88 |
size += CHUNK_ALIGN_BYTES - (size % CHUNK_ALIGN_BYTES); |
| 89 |
|
| 90 |
engine->slabs.slabclass[i].size = size; |
| 91 |
engine->slabs.slabclass[i].perslab = engine->config.item_size_max / engine->slabs.slabclass[i].size; |
| 92 |
size *= factor; |
| 93 |
if (engine->config.verbose > 1) { |
| 94 |
EXTENSION_LOGGER_DESCRIPTOR *logger; |
| 95 |
logger = (void*)engine->server.extension->get_extension(EXTENSION_LOGGER); |
| 96 |
logger->log(EXTENSION_LOG_INFO, NULL, |
| 97 |
"slab class %3d: chunk size %9u perslab %7u\n", |
| 98 |
i, engine->slabs.slabclass[i].size, |
| 99 |
engine->slabs.slabclass[i].perslab); |
| 100 |
} |
| 101 |
} |
| 102 |
|
| 103 |
engine->slabs.power_largest = i; |
| 104 |
engine->slabs.slabclass[engine->slabs.power_largest].size = engine->config.item_size_max; |
| 105 |
engine->slabs.slabclass[engine->slabs.power_largest].perslab = 1; |
| 106 |
if (engine->config.verbose > 1) { |
| 107 |
EXTENSION_LOGGER_DESCRIPTOR *logger; |
| 108 |
logger = (void*)engine->server.extension->get_extension(EXTENSION_LOGGER); |
| 109 |
logger->log(EXTENSION_LOG_INFO, NULL, |
| 110 |
"slab class %3d: chunk size %9u perslab %7u\n", |
| 111 |
i, engine->slabs.slabclass[i].size, |
| 112 |
engine->slabs.slabclass[i].perslab); |
| 113 |
} |
| 114 |
|
| 115 |
/* for the test suite: faking of how much we've already malloc'd */ |
| 116 |
{ |
| 117 |
char *t_initial_malloc = getenv("T_MEMD_INITIAL_MALLOC"); |
| 118 |
if (t_initial_malloc) { |
| 119 |
engine->slabs.mem_malloced = (size_t)atol(t_initial_malloc); |
| 120 |
} |
| 121 |
|
| 122 |
} |
| 123 |
|
| 124 |
#ifndef DONT_PREALLOC_SLABS |
| 125 |
{ |
| 126 |
char *pre_alloc = getenv("T_MEMD_SLABS_ALLOC"); |
| 127 |
|
| 128 |
if (pre_alloc == NULL || atoi(pre_alloc) != 0) { |
| 129 |
slabs_preallocate(power_largest); |
| 130 |
} |
| 131 |
} |
| 132 |
#endif |
| 133 |
|
| 134 |
return ENGINE_SUCCESS; |
| 135 |
} |
| 136 |
|
| 137 |
#ifndef DONT_PREALLOC_SLABS |
| 138 |
static void slabs_preallocate (const unsigned int maxslabs) { |
| 139 |
int i; |
| 140 |
unsigned int prealloc = 0; |
| 141 |
|
| 142 |
/* pre-allocate a 1MB slab in every size class so people don't get |
| 143 |
confused by non-intuitive "SERVER_ERROR out of memory" |
| 144 |
messages. this is the most common question on the mailing |
| 145 |
list. if you really don't want this, you can rebuild without |
| 146 |
these three lines. */ |
| 147 |
|
| 148 |
for (i = POWER_SMALLEST; i <= POWER_LARGEST; i++) { |
| 149 |
if (++prealloc > maxslabs) |
| 150 |
return; |
| 151 |
do_slabs_newslab(i); |
| 152 |
} |
| 153 |
|
| 154 |
} |
| 155 |
#endif |
| 156 |
|
| 157 |
static int grow_slab_list (struct default_engine *engine, const unsigned int id) { |
| 158 |
slabclass_t *p = &engine->slabs.slabclass[id]; |
| 159 |
if (p->slabs == p->list_size) { |
| 160 |
size_t new_size = (p->list_size != 0) ? p->list_size * 2 : 16; |
| 161 |
void *new_list = realloc(p->slab_list, new_size * sizeof(void *)); |
| 162 |
if (new_list == 0) return 0; |
| 163 |
p->list_size = new_size; |
| 164 |
p->slab_list = new_list; |
| 165 |
} |
| 166 |
return 1; |
| 167 |
} |
| 168 |
|
| 169 |
static int do_slabs_newslab(struct default_engine *engine, const unsigned int id) { |
| 170 |
slabclass_t *p = &engine->slabs.slabclass[id]; |
| 171 |
int len = p->size * p->perslab; |
| 172 |
char *ptr; |
| 173 |
|
| 174 |
if ((engine->slabs.mem_limit && engine->slabs.mem_malloced + len > engine->slabs.mem_limit && p->slabs > 0) || |
| 175 |
(grow_slab_list(engine, id) == 0) || |
| 176 |
((ptr = memory_allocate(engine, (size_t)len)) == 0)) { |
| 177 |
|
| 178 |
MEMCACHED_SLABS_SLABCLASS_ALLOCATE_FAILED(id); |
| 179 |
return 0; |
| 180 |
} |
| 181 |
|
| 182 |
memset(ptr, 0, (size_t)len); |
| 183 |
p->end_page_ptr = ptr; |
| 184 |
p->end_page_free = p->perslab; |
| 185 |
|
| 186 |
p->slab_list[p->slabs++] = ptr; |
| 187 |
engine->slabs.mem_malloced += len; |
| 188 |
MEMCACHED_SLABS_SLABCLASS_ALLOCATE(id); |
| 189 |
|
| 190 |
return 1; |
| 191 |
} |
| 192 |
|
| 193 |
/*@null@*/ |
| 194 |
static void *do_slabs_alloc(struct default_engine *engine, const size_t size, unsigned int id) { |
| 195 |
slabclass_t *p; |
| 196 |
void *ret = NULL; |
| 197 |
|
| 198 |
if (id < POWER_SMALLEST || id > engine->slabs.power_largest) { |
| 199 |
MEMCACHED_SLABS_ALLOCATE_FAILED(size, 0); |
| 200 |
return NULL; |
| 201 |
} |
| 202 |
|
| 203 |
p = &engine->slabs.slabclass[id]; |
| 204 |
|
| 205 |
#ifdef USE_SYSTEM_MALLOC |
| 206 |
if (engine->slabs.mem_limit && engine->slabs.mem_malloced + size > engine->slabs.mem_limit) { |
| 207 |
MEMCACHED_SLABS_ALLOCATE_FAILED(size, id); |
| 208 |
return 0; |
| 209 |
} |
| 210 |
engine->slabs.mem_malloced += size; |
| 211 |
ret = malloc(size); |
| 212 |
MEMCACHED_SLABS_ALLOCATE(size, id, 0, ret); |
| 213 |
return ret; |
| 214 |
#endif |
| 215 |
|
| 216 |
/* fail unless we have space at the end of a recently allocated page, |
| 217 |
we have something on our freelist, or we could allocate a new page */ |
| 218 |
if (! (p->end_page_ptr != 0 || p->sl_curr != 0 || |
| 219 |
do_slabs_newslab(engine, id) != 0)) { |
| 220 |
/* We don't have more memory available */ |
| 221 |
ret = NULL; |
| 222 |
} else if (p->sl_curr != 0) { |
| 223 |
/* return off our freelist */ |
| 224 |
ret = p->slots[--p->sl_curr]; |
| 225 |
} else { |
| 226 |
/* if we recently allocated a whole page, return from that */ |
| 227 |
assert(p->end_page_ptr != NULL); |
| 228 |
ret = p->end_page_ptr; |
| 229 |
if (--p->end_page_free != 0) { |
| 230 |
p->end_page_ptr = ((caddr_t)p->end_page_ptr) + p->size; |
| 231 |
} else { |
| 232 |
p->end_page_ptr = 0; |
| 233 |
} |
| 234 |
} |
| 235 |
|
| 236 |
if (ret) { |
| 237 |
p->requested += size; |
| 238 |
MEMCACHED_SLABS_ALLOCATE(size, id, p->size, ret); |
| 239 |
} else { |
| 240 |
MEMCACHED_SLABS_ALLOCATE_FAILED(size, id); |
| 241 |
} |
| 242 |
|
| 243 |
return ret; |
| 244 |
} |
| 245 |
|
| 246 |
static void do_slabs_free(struct default_engine *engine, void *ptr, const size_t size, unsigned int id) { |
| 247 |
slabclass_t *p; |
| 248 |
|
| 249 |
if (id < POWER_SMALLEST || id > engine->slabs.power_largest) |
| 250 |
return; |
| 251 |
|
| 252 |
MEMCACHED_SLABS_FREE(size, id, ptr); |
| 253 |
p = &engine->slabs.slabclass[id]; |
| 254 |
|
| 255 |
#ifdef USE_SYSTEM_MALLOC |
| 256 |
engine->slabs.mem_malloced -= size; |
| 257 |
free(ptr); |
| 258 |
return; |
| 259 |
#endif |
| 260 |
|
| 261 |
if (p->sl_curr == p->sl_total) { /* need more space on the free list */ |
| 262 |
int new_size = (p->sl_total != 0) ? p->sl_total * 2 : 16; /* 16 is arbitrary */ |
| 263 |
void **new_slots = realloc(p->slots, new_size * sizeof(void *)); |
| 264 |
if (new_slots == 0) |
| 265 |
return; |
| 266 |
p->slots = new_slots; |
| 267 |
p->sl_total = new_size; |
| 268 |
} |
| 269 |
p->slots[p->sl_curr++] = ptr; |
| 270 |
p->requested -= size; |
| 271 |
return; |
| 272 |
} |
| 273 |
|
| 274 |
void add_statistics(const void *cookie, ADD_STAT add_stats, |
| 275 |
const char* prefix, int num, const char *key, |
| 276 |
const char *fmt, ...) { |
| 277 |
char name[80], val[80]; |
| 278 |
int klen = 0, vlen; |
| 279 |
va_list ap; |
| 280 |
|
| 281 |
assert(cookie); |
| 282 |
assert(add_stats); |
| 283 |
assert(key); |
| 284 |
|
| 285 |
va_start(ap, fmt); |
| 286 |
vlen = vsnprintf(val, sizeof(val) - 1, fmt, ap); |
| 287 |
va_end(ap); |
| 288 |
|
| 289 |
if (prefix != NULL) { |
| 290 |
klen = snprintf(name, sizeof(name), "%s:", prefix); |
| 291 |
} |
| 292 |
|
| 293 |
if (num != -1) { |
| 294 |
klen += snprintf(name + klen, sizeof(name) - klen, "%d:", num); |
| 295 |
} |
| 296 |
|
| 297 |
klen += snprintf(name + klen, sizeof(name) - klen, "%s", key); |
| 298 |
|
| 299 |
add_stats(name, klen, val, vlen, cookie); |
| 300 |
} |
| 301 |
|
| 302 |
/*@null@*/ |
| 303 |
static void do_slabs_stats(struct default_engine *engine, ADD_STAT add_stats, const void *cookie) { |
| 304 |
int i, total; |
| 305 |
/* Get the per-thread stats which contain some interesting aggregates */ |
| 306 |
#ifdef FUTURE |
| 307 |
struct conn *conn = (struct conn*)cookie; |
| 308 |
struct thread_stats thread_stats; |
| 309 |
threadlocal_stats_aggregate(c, &thread_stats); |
| 310 |
#endif |
| 311 |
|
| 312 |
total = 0; |
| 313 |
for(i = POWER_SMALLEST; i <= engine->slabs.power_largest; i++) { |
| 314 |
slabclass_t *p = &engine->slabs.slabclass[i]; |
| 315 |
if (p->slabs != 0) { |
| 316 |
uint32_t perslab, slabs; |
| 317 |
slabs = p->slabs; |
| 318 |
perslab = p->perslab; |
| 319 |
|
| 320 |
add_statistics(cookie, add_stats, NULL, i, "chunk_size", "%u", |
| 321 |
p->size); |
| 322 |
add_statistics(cookie, add_stats, NULL, i, "chunks_per_page", "%u", |
| 323 |
perslab); |
| 324 |
add_statistics(cookie, add_stats, NULL, i, "total_pages", "%u", |
| 325 |
slabs); |
| 326 |
add_statistics(cookie, add_stats, NULL, i, "total_chunks", "%u", |
| 327 |
slabs * perslab); |
| 328 |
add_statistics(cookie, add_stats, NULL, i, "used_chunks", "%u", |
| 329 |
slabs*perslab - p->sl_curr - p->end_page_free); |
| 330 |
add_statistics(cookie, add_stats, NULL, i, "free_chunks", "%u", |
| 331 |
p->sl_curr); |
| 332 |
add_statistics(cookie, add_stats, NULL, i, "free_chunks_end", "%u", |
| 333 |
p->end_page_free); |
| 334 |
add_statistics(cookie, add_stats, NULL, i, "mem_requested", "%zu", |
| 335 |
p->requested); |
| 336 |
#ifdef FUTURE |
| 337 |
add_statistics(cookie, add_stats, NULL, i, "get_hits", "%"PRIu64, |
| 338 |
thread_stats.slab_stats[i].get_hits); |
| 339 |
add_statistics(cookie, add_stats, NULL, i, "cmd_set", "%"PRIu64, |
| 340 |
thread_stats.slab_stats[i].set_cmds); |
| 341 |
add_statistics(cookie, add_stats, NULL, i, "delete_hits", "%"PRIu64, |
| 342 |
thread_stats.slab_stats[i].delete_hits); |
| 343 |
add_statistics(cookie, add_stats, NULL, i, "cas_hits", "%"PRIu64, |
| 344 |
thread_stats.slab_stats[i].cas_hits); |
| 345 |
add_statistics(cookie, add_stats, NULL, i, "cas_badval", "%"PRIu64, |
| 346 |
thread_stats.slab_stats[i].cas_badval); |
| 347 |
#endif |
| 348 |
total++; |
| 349 |
} |
| 350 |
} |
| 351 |
|
| 352 |
/* add overall slab stats and append terminator */ |
| 353 |
|
| 354 |
add_statistics(cookie, add_stats, NULL, -1, "active_slabs", "%d", total); |
| 355 |
add_statistics(cookie, add_stats, NULL, -1, "total_malloced", "%zu", |
| 356 |
engine->slabs.mem_malloced); |
| 357 |
} |
| 358 |
|
| 359 |
static void *memory_allocate(struct default_engine *engine, size_t size) { |
| 360 |
void *ret; |
| 361 |
|
| 362 |
if (engine->slabs.mem_base == NULL) { |
| 363 |
/* We are not using a preallocated large memory chunk */ |
| 364 |
ret = malloc(size); |
| 365 |
} else { |
| 366 |
ret = engine->slabs.mem_current; |
| 367 |
|
| 368 |
if (size > engine->slabs.mem_avail) { |
| 369 |
return NULL; |
| 370 |
} |
| 371 |
|
| 372 |
/* mem_current pointer _must_ be aligned!!! */ |
| 373 |
if (size % CHUNK_ALIGN_BYTES) { |
| 374 |
size += CHUNK_ALIGN_BYTES - (size % CHUNK_ALIGN_BYTES); |
| 375 |
} |
| 376 |
|
| 377 |
engine->slabs.mem_current = ((char*)engine->slabs.mem_current) + size; |
| 378 |
if (size < engine->slabs.mem_avail) { |
| 379 |
engine->slabs.mem_avail -= size; |
| 380 |
} else { |
| 381 |
engine->slabs.mem_avail = 0; |
| 382 |
} |
| 383 |
} |
| 384 |
|
| 385 |
return ret; |
| 386 |
} |
| 387 |
|
| 388 |
void *slabs_alloc(struct default_engine *engine, size_t size, unsigned int id) { |
| 389 |
void *ret; |
| 390 |
|
| 391 |
pthread_mutex_lock(&engine->slabs.lock); |
| 392 |
ret = do_slabs_alloc(engine, size, id); |
| 393 |
pthread_mutex_unlock(&engine->slabs.lock); |
| 394 |
return ret; |
| 395 |
} |
| 396 |
|
| 397 |
void slabs_free(struct default_engine *engine, void *ptr, size_t size, unsigned int id) { |
| 398 |
pthread_mutex_lock(&engine->slabs.lock); |
| 399 |
do_slabs_free(engine, ptr, size, id); |
| 400 |
pthread_mutex_unlock(&engine->slabs.lock); |
| 401 |
} |
| 402 |
|
| 403 |
void slabs_stats(struct default_engine *engine, ADD_STAT add_stats, const void *c) { |
| 404 |
pthread_mutex_lock(&engine->slabs.lock); |
| 405 |
do_slabs_stats(engine, add_stats, c); |
| 406 |
pthread_mutex_unlock(&engine->slabs.lock); |
| 407 |
} |
| 408 |
|
| 409 |
void slabs_adjust_mem_requested(struct default_engine *engine, unsigned int id, size_t old, size_t ntotal) |
| 410 |
{ |
| 411 |
pthread_mutex_lock(&engine->slabs.lock); |
| 412 |
slabclass_t *p; |
| 413 |
if (id < POWER_SMALLEST || id > engine->slabs.power_largest) { |
| 414 |
EXTENSION_LOGGER_DESCRIPTOR *logger; |
| 415 |
logger = (void*)engine->server.extension->get_extension(EXTENSION_LOGGER); |
| 416 |
logger->log(EXTENSION_LOG_WARNING, NULL, |
| 417 |
"Internal error! Invalid slab class\n"); |
| 418 |
abort(); |
| 419 |
} |
| 420 |
|
| 421 |
p = &engine->slabs.slabclass[id]; |
| 422 |
p->requested = p->requested - old + ntotal; |
| 423 |
pthread_mutex_unlock(&engine->slabs.lock); |
| 424 |
} |