prof.h revision b1941c615023cab9baf0a78a28df1e3b4972434f
1/******************************************************************************/ 2#ifdef JEMALLOC_H_TYPES 3 4typedef struct prof_bt_s prof_bt_t; 5typedef struct prof_cnt_s prof_cnt_t; 6typedef struct prof_thr_cnt_s prof_thr_cnt_t; 7typedef struct prof_ctx_s prof_ctx_t; 8typedef struct prof_tdata_s prof_tdata_t; 9 10/* Option defaults. */ 11#define PROF_PREFIX_DEFAULT "jeprof" 12#define LG_PROF_SAMPLE_DEFAULT 19 13#define LG_PROF_INTERVAL_DEFAULT -1 14 15/* 16 * Hard limit on stack backtrace depth. The version of prof_backtrace() that 17 * is based on __builtin_return_address() necessarily has a hard-coded number 18 * of backtrace frame handlers, and should be kept in sync with this setting. 19 */ 20#define PROF_BT_MAX 128 21 22/* Maximum number of backtraces to store in each per thread LRU cache. */ 23#define PROF_TCMAX 1024 24 25/* Initial hash table size. */ 26#define PROF_CKH_MINITEMS 64 27 28/* Size of memory buffer to use when writing dump files. */ 29#define PROF_DUMP_BUFSIZE 65536 30 31/* Size of stack-allocated buffer used by prof_printf(). */ 32#define PROF_PRINTF_BUFSIZE 128 33 34/* 35 * Number of mutexes shared among all ctx's. No space is allocated for these 36 * unless profiling is enabled, so it's okay to over-provision. 37 */ 38#define PROF_NCTX_LOCKS 1024 39 40/* 41 * prof_tdata pointers close to NULL are used to encode state information that 42 * is used for cleaning up during thread shutdown. 43 */ 44#define PROF_TDATA_STATE_REINCARNATED ((prof_tdata_t *)(uintptr_t)1) 45#define PROF_TDATA_STATE_PURGATORY ((prof_tdata_t *)(uintptr_t)2) 46#define PROF_TDATA_STATE_MAX PROF_TDATA_STATE_PURGATORY 47 48#endif /* JEMALLOC_H_TYPES */ 49/******************************************************************************/ 50#ifdef JEMALLOC_H_STRUCTS 51 52struct prof_bt_s { 53 /* Backtrace, stored as len program counters. */ 54 void **vec; 55 unsigned len; 56}; 57 58#ifdef JEMALLOC_PROF_LIBGCC 59/* Data structure passed to libgcc _Unwind_Backtrace() callback functions. */ 60typedef struct { 61 prof_bt_t *bt; 62 unsigned nignore; 63 unsigned max; 64} prof_unwind_data_t; 65#endif 66 67struct prof_cnt_s { 68 /* 69 * Profiling counters. An allocation/deallocation pair can operate on 70 * different prof_thr_cnt_t objects that are linked into the same 71 * prof_ctx_t cnts_ql, so it is possible for the cur* counters to go 72 * negative. In principle it is possible for the *bytes counters to 73 * overflow/underflow, but a general solution would require something 74 * like 128-bit counters; this implementation doesn't bother to solve 75 * that problem. 76 */ 77 int64_t curobjs; 78 int64_t curbytes; 79 uint64_t accumobjs; 80 uint64_t accumbytes; 81}; 82 83struct prof_thr_cnt_s { 84 /* Linkage into prof_ctx_t's cnts_ql. */ 85 ql_elm(prof_thr_cnt_t) cnts_link; 86 87 /* Linkage into thread's LRU. */ 88 ql_elm(prof_thr_cnt_t) lru_link; 89 90 /* 91 * Associated context. If a thread frees an object that it did not 92 * allocate, it is possible that the context is not cached in the 93 * thread's hash table, in which case it must be able to look up the 94 * context, insert a new prof_thr_cnt_t into the thread's hash table, 95 * and link it into the prof_ctx_t's cnts_ql. 96 */ 97 prof_ctx_t *ctx; 98 99 /* 100 * Threads use memory barriers to update the counters. Since there is 101 * only ever one writer, the only challenge is for the reader to get a 102 * consistent read of the counters. 103 * 104 * The writer uses this series of operations: 105 * 106 * 1) Increment epoch to an odd number. 107 * 2) Update counters. 108 * 3) Increment epoch to an even number. 109 * 110 * The reader must assure 1) that the epoch is even while it reads the 111 * counters, and 2) that the epoch doesn't change between the time it 112 * starts and finishes reading the counters. 113 */ 114 unsigned epoch; 115 116 /* Profiling counters. */ 117 prof_cnt_t cnts; 118}; 119 120struct prof_ctx_s { 121 /* Associated backtrace. */ 122 prof_bt_t *bt; 123 124 /* Protects nlimbo, cnt_merged, and cnts_ql. */ 125 malloc_mutex_t *lock; 126 127 /* 128 * Number of threads that currently cause this ctx to be in a state of 129 * limbo due to one of: 130 * - Initializing per thread counters associated with this ctx. 131 * - Preparing to destroy this ctx. 132 * nlimbo must be 1 (single destroyer) in order to safely destroy the 133 * ctx. 134 */ 135 unsigned nlimbo; 136 137 /* Temporary storage for summation during dump. */ 138 prof_cnt_t cnt_summed; 139 140 /* When threads exit, they merge their stats into cnt_merged. */ 141 prof_cnt_t cnt_merged; 142 143 /* 144 * List of profile counters, one for each thread that has allocated in 145 * this context. 146 */ 147 ql_head(prof_thr_cnt_t) cnts_ql; 148}; 149 150struct prof_tdata_s { 151 /* 152 * Hash of (prof_bt_t *)-->(prof_thr_cnt_t *). Each thread keeps a 153 * cache of backtraces, with associated thread-specific prof_thr_cnt_t 154 * objects. Other threads may read the prof_thr_cnt_t contents, but no 155 * others will ever write them. 156 * 157 * Upon thread exit, the thread must merge all the prof_thr_cnt_t 158 * counter data into the associated prof_ctx_t objects, and unlink/free 159 * the prof_thr_cnt_t objects. 160 */ 161 ckh_t bt2cnt; 162 163 /* LRU for contents of bt2cnt. */ 164 ql_head(prof_thr_cnt_t) lru_ql; 165 166 /* Backtrace vector, used for calls to prof_backtrace(). */ 167 void **vec; 168 169 /* Sampling state. */ 170 uint64_t prng_state; 171 uint64_t threshold; 172 uint64_t accum; 173 174 /* State used to avoid dumping while operating on prof internals. */ 175 bool enq; 176 bool enq_idump; 177 bool enq_gdump; 178}; 179 180#endif /* JEMALLOC_H_STRUCTS */ 181/******************************************************************************/ 182#ifdef JEMALLOC_H_EXTERNS 183 184extern bool opt_prof; 185/* 186 * Even if opt_prof is true, sampling can be temporarily disabled by setting 187 * opt_prof_active to false. No locking is used when updating opt_prof_active, 188 * so there are no guarantees regarding how long it will take for all threads 189 * to notice state changes. 190 */ 191extern bool opt_prof_active; 192extern size_t opt_lg_prof_sample; /* Mean bytes between samples. */ 193extern ssize_t opt_lg_prof_interval; /* lg(prof_interval). */ 194extern bool opt_prof_gdump; /* High-water memory dumping. */ 195extern bool opt_prof_final; /* Final profile dumping. */ 196extern bool opt_prof_leak; /* Dump leak summary at exit. */ 197extern bool opt_prof_accum; /* Report cumulative bytes. */ 198extern char opt_prof_prefix[PATH_MAX + 1]; 199 200/* 201 * Profile dump interval, measured in bytes allocated. Each arena triggers a 202 * profile dump when it reaches this threshold. The effect is that the 203 * interval between profile dumps averages prof_interval, though the actual 204 * interval between dumps will tend to be sporadic, and the interval will be a 205 * maximum of approximately (prof_interval * narenas). 206 */ 207extern uint64_t prof_interval; 208 209/* 210 * If true, promote small sampled objects to large objects, since small run 211 * headers do not have embedded profile context pointers. 212 */ 213extern bool prof_promote; 214 215void bt_init(prof_bt_t *bt, void **vec); 216void prof_backtrace(prof_bt_t *bt, unsigned nignore); 217prof_thr_cnt_t *prof_lookup(prof_bt_t *bt); 218void prof_idump(void); 219bool prof_mdump(const char *filename); 220void prof_gdump(void); 221prof_tdata_t *prof_tdata_init(void); 222void prof_tdata_cleanup(void *arg); 223void prof_boot0(void); 224void prof_boot1(void); 225bool prof_boot2(void); 226void prof_prefork(void); 227void prof_postfork_parent(void); 228void prof_postfork_child(void); 229 230#endif /* JEMALLOC_H_EXTERNS */ 231/******************************************************************************/ 232#ifdef JEMALLOC_H_INLINES 233 234#define PROF_ALLOC_PREP(nignore, size, ret) do { \ 235 prof_tdata_t *prof_tdata; \ 236 prof_bt_t bt; \ 237 \ 238 assert(size == s2u(size)); \ 239 \ 240 prof_tdata = prof_tdata_get(true); \ 241 if ((uintptr_t)prof_tdata <= (uintptr_t)PROF_TDATA_STATE_MAX) { \ 242 if (prof_tdata != NULL) \ 243 ret = (prof_thr_cnt_t *)(uintptr_t)1U; \ 244 else \ 245 ret = NULL; \ 246 break; \ 247 } \ 248 \ 249 if (opt_prof_active == false) { \ 250 /* Sampling is currently inactive, so avoid sampling. */\ 251 ret = (prof_thr_cnt_t *)(uintptr_t)1U; \ 252 } else if (opt_lg_prof_sample == 0) { \ 253 /* Don't bother with sampling logic, since sampling */\ 254 /* interval is 1. */\ 255 bt_init(&bt, prof_tdata->vec); \ 256 prof_backtrace(&bt, nignore); \ 257 ret = prof_lookup(&bt); \ 258 } else { \ 259 if (prof_tdata->threshold == 0) { \ 260 /* Initialize. Seed the prng differently for */\ 261 /* each thread. */\ 262 prof_tdata->prng_state = \ 263 (uint64_t)(uintptr_t)&size; \ 264 prof_sample_threshold_update(prof_tdata); \ 265 } \ 266 \ 267 /* Determine whether to capture a backtrace based on */\ 268 /* whether size is enough for prof_accum to reach */\ 269 /* prof_tdata->threshold. However, delay updating */\ 270 /* these variables until prof_{m,re}alloc(), because */\ 271 /* we don't know for sure that the allocation will */\ 272 /* succeed. */\ 273 /* */\ 274 /* Use subtraction rather than addition to avoid */\ 275 /* potential integer overflow. */\ 276 if (size >= prof_tdata->threshold - \ 277 prof_tdata->accum) { \ 278 bt_init(&bt, prof_tdata->vec); \ 279 prof_backtrace(&bt, nignore); \ 280 ret = prof_lookup(&bt); \ 281 } else \ 282 ret = (prof_thr_cnt_t *)(uintptr_t)1U; \ 283 } \ 284} while (0) 285 286#ifndef JEMALLOC_ENABLE_INLINE 287malloc_tsd_protos(JEMALLOC_ATTR(unused), prof_tdata, prof_tdata_t *) 288 289prof_tdata_t *prof_tdata_get(bool create); 290void prof_sample_threshold_update(prof_tdata_t *prof_tdata); 291prof_ctx_t *prof_ctx_get(const void *ptr); 292void prof_ctx_set(const void *ptr, prof_ctx_t *ctx); 293bool prof_sample_accum_update(size_t size); 294void prof_malloc(const void *ptr, size_t size, prof_thr_cnt_t *cnt); 295void prof_realloc(const void *ptr, size_t size, prof_thr_cnt_t *cnt, 296 size_t old_size, prof_ctx_t *old_ctx); 297void prof_free(const void *ptr, size_t size); 298#endif 299 300#if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_PROF_C_)) 301/* Thread-specific backtrace cache, used to reduce bt2ctx contention. */ 302malloc_tsd_externs(prof_tdata, prof_tdata_t *) 303malloc_tsd_funcs(JEMALLOC_INLINE, prof_tdata, prof_tdata_t *, NULL, 304 prof_tdata_cleanup) 305 306JEMALLOC_INLINE prof_tdata_t * 307prof_tdata_get(bool create) 308{ 309 prof_tdata_t *prof_tdata; 310 311 cassert(config_prof); 312 313 prof_tdata = *prof_tdata_tsd_get(); 314 if (create && prof_tdata == NULL) 315 prof_tdata = prof_tdata_init(); 316 317 return (prof_tdata); 318} 319 320JEMALLOC_INLINE void 321prof_sample_threshold_update(prof_tdata_t *prof_tdata) 322{ 323 /* 324 * The body of this function is compiled out unless heap profiling is 325 * enabled, so that it is possible to compile jemalloc with floating 326 * point support completely disabled. Avoiding floating point code is 327 * important on memory-constrained systems, but it also enables a 328 * workaround for versions of glibc that don't properly save/restore 329 * floating point registers during dynamic lazy symbol loading (which 330 * internally calls into whatever malloc implementation happens to be 331 * integrated into the application). Note that some compilers (e.g. 332 * gcc 4.8) may use floating point registers for fast memory moves, so 333 * jemalloc must be compiled with such optimizations disabled (e.g. 334 * -mno-sse) in order for the workaround to be complete. 335 */ 336#ifdef JEMALLOC_PROF 337 uint64_t r; 338 double u; 339 340 cassert(config_prof); 341 342 /* 343 * Compute sample threshold as a geometrically distributed random 344 * variable with mean (2^opt_lg_prof_sample). 345 * 346 * __ __ 347 * | log(u) | 1 348 * prof_tdata->threshold = | -------- |, where p = ------------------- 349 * | log(1-p) | opt_lg_prof_sample 350 * 2 351 * 352 * For more information on the math, see: 353 * 354 * Non-Uniform Random Variate Generation 355 * Luc Devroye 356 * Springer-Verlag, New York, 1986 357 * pp 500 358 * (http://luc.devroye.org/rnbookindex.html) 359 */ 360 prng64(r, 53, prof_tdata->prng_state, 361 UINT64_C(6364136223846793005), UINT64_C(1442695040888963407)); 362 u = (double)r * (1.0/9007199254740992.0L); 363 prof_tdata->threshold = (uint64_t)(log(u) / 364 log(1.0 - (1.0 / (double)((uint64_t)1U << opt_lg_prof_sample)))) 365 + (uint64_t)1U; 366#endif 367} 368 369JEMALLOC_INLINE prof_ctx_t * 370prof_ctx_get(const void *ptr) 371{ 372 prof_ctx_t *ret; 373 arena_chunk_t *chunk; 374 375 cassert(config_prof); 376 assert(ptr != NULL); 377 378 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 379 if (chunk != ptr) { 380 /* Region. */ 381 ret = arena_prof_ctx_get(ptr); 382 } else 383 ret = huge_prof_ctx_get(ptr); 384 385 return (ret); 386} 387 388JEMALLOC_INLINE void 389prof_ctx_set(const void *ptr, prof_ctx_t *ctx) 390{ 391 arena_chunk_t *chunk; 392 393 cassert(config_prof); 394 assert(ptr != NULL); 395 396 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 397 if (chunk != ptr) { 398 /* Region. */ 399 arena_prof_ctx_set(ptr, ctx); 400 } else 401 huge_prof_ctx_set(ptr, ctx); 402} 403 404JEMALLOC_INLINE bool 405prof_sample_accum_update(size_t size) 406{ 407 prof_tdata_t *prof_tdata; 408 409 cassert(config_prof); 410 /* Sampling logic is unnecessary if the interval is 1. */ 411 assert(opt_lg_prof_sample != 0); 412 413 prof_tdata = prof_tdata_get(false); 414 if ((uintptr_t)prof_tdata <= (uintptr_t)PROF_TDATA_STATE_MAX) 415 return (true); 416 417 /* Take care to avoid integer overflow. */ 418 if (size >= prof_tdata->threshold - prof_tdata->accum) { 419 prof_tdata->accum -= (prof_tdata->threshold - size); 420 /* Compute new sample threshold. */ 421 prof_sample_threshold_update(prof_tdata); 422 while (prof_tdata->accum >= prof_tdata->threshold) { 423 prof_tdata->accum -= prof_tdata->threshold; 424 prof_sample_threshold_update(prof_tdata); 425 } 426 return (false); 427 } else { 428 prof_tdata->accum += size; 429 return (true); 430 } 431} 432 433JEMALLOC_INLINE void 434prof_malloc(const void *ptr, size_t size, prof_thr_cnt_t *cnt) 435{ 436 437 cassert(config_prof); 438 assert(ptr != NULL); 439 assert(size == isalloc(ptr, true)); 440 441 if (opt_lg_prof_sample != 0) { 442 if (prof_sample_accum_update(size)) { 443 /* 444 * Don't sample. For malloc()-like allocation, it is 445 * always possible to tell in advance how large an 446 * object's usable size will be, so there should never 447 * be a difference between the size passed to 448 * PROF_ALLOC_PREP() and prof_malloc(). 449 */ 450 assert((uintptr_t)cnt == (uintptr_t)1U); 451 } 452 } 453 454 if ((uintptr_t)cnt > (uintptr_t)1U) { 455 prof_ctx_set(ptr, cnt->ctx); 456 457 cnt->epoch++; 458 /*********/ 459 mb_write(); 460 /*********/ 461 cnt->cnts.curobjs++; 462 cnt->cnts.curbytes += size; 463 if (opt_prof_accum) { 464 cnt->cnts.accumobjs++; 465 cnt->cnts.accumbytes += size; 466 } 467 /*********/ 468 mb_write(); 469 /*********/ 470 cnt->epoch++; 471 /*********/ 472 mb_write(); 473 /*********/ 474 } else 475 prof_ctx_set(ptr, (prof_ctx_t *)(uintptr_t)1U); 476} 477 478JEMALLOC_INLINE void 479prof_realloc(const void *ptr, size_t size, prof_thr_cnt_t *cnt, 480 size_t old_size, prof_ctx_t *old_ctx) 481{ 482 prof_thr_cnt_t *told_cnt; 483 484 cassert(config_prof); 485 assert(ptr != NULL || (uintptr_t)cnt <= (uintptr_t)1U); 486 487 if (ptr != NULL) { 488 assert(size == isalloc(ptr, true)); 489 if (opt_lg_prof_sample != 0) { 490 if (prof_sample_accum_update(size)) { 491 /* 492 * Don't sample. The size passed to 493 * PROF_ALLOC_PREP() was larger than what 494 * actually got allocated, so a backtrace was 495 * captured for this allocation, even though 496 * its actual size was insufficient to cross 497 * the sample threshold. 498 */ 499 cnt = (prof_thr_cnt_t *)(uintptr_t)1U; 500 } 501 } 502 } 503 504 if ((uintptr_t)old_ctx > (uintptr_t)1U) { 505 told_cnt = prof_lookup(old_ctx->bt); 506 if (told_cnt == NULL) { 507 /* 508 * It's too late to propagate OOM for this realloc(), 509 * so operate directly on old_cnt->ctx->cnt_merged. 510 */ 511 malloc_mutex_lock(old_ctx->lock); 512 old_ctx->cnt_merged.curobjs--; 513 old_ctx->cnt_merged.curbytes -= old_size; 514 malloc_mutex_unlock(old_ctx->lock); 515 told_cnt = (prof_thr_cnt_t *)(uintptr_t)1U; 516 } 517 } else 518 told_cnt = (prof_thr_cnt_t *)(uintptr_t)1U; 519 520 if ((uintptr_t)told_cnt > (uintptr_t)1U) 521 told_cnt->epoch++; 522 if ((uintptr_t)cnt > (uintptr_t)1U) { 523 prof_ctx_set(ptr, cnt->ctx); 524 cnt->epoch++; 525 } else if (ptr != NULL) 526 prof_ctx_set(ptr, (prof_ctx_t *)(uintptr_t)1U); 527 /*********/ 528 mb_write(); 529 /*********/ 530 if ((uintptr_t)told_cnt > (uintptr_t)1U) { 531 told_cnt->cnts.curobjs--; 532 told_cnt->cnts.curbytes -= old_size; 533 } 534 if ((uintptr_t)cnt > (uintptr_t)1U) { 535 cnt->cnts.curobjs++; 536 cnt->cnts.curbytes += size; 537 if (opt_prof_accum) { 538 cnt->cnts.accumobjs++; 539 cnt->cnts.accumbytes += size; 540 } 541 } 542 /*********/ 543 mb_write(); 544 /*********/ 545 if ((uintptr_t)told_cnt > (uintptr_t)1U) 546 told_cnt->epoch++; 547 if ((uintptr_t)cnt > (uintptr_t)1U) 548 cnt->epoch++; 549 /*********/ 550 mb_write(); /* Not strictly necessary. */ 551} 552 553JEMALLOC_INLINE void 554prof_free(const void *ptr, size_t size) 555{ 556 prof_ctx_t *ctx = prof_ctx_get(ptr); 557 558 cassert(config_prof); 559 560 if ((uintptr_t)ctx > (uintptr_t)1) { 561 prof_thr_cnt_t *tcnt; 562 assert(size == isalloc(ptr, true)); 563 tcnt = prof_lookup(ctx->bt); 564 565 if (tcnt != NULL) { 566 tcnt->epoch++; 567 /*********/ 568 mb_write(); 569 /*********/ 570 tcnt->cnts.curobjs--; 571 tcnt->cnts.curbytes -= size; 572 /*********/ 573 mb_write(); 574 /*********/ 575 tcnt->epoch++; 576 /*********/ 577 mb_write(); 578 /*********/ 579 } else { 580 /* 581 * OOM during free() cannot be propagated, so operate 582 * directly on cnt->ctx->cnt_merged. 583 */ 584 malloc_mutex_lock(ctx->lock); 585 ctx->cnt_merged.curobjs--; 586 ctx->cnt_merged.curbytes -= size; 587 malloc_mutex_unlock(ctx->lock); 588 } 589 } 590} 591#endif 592 593#endif /* JEMALLOC_H_INLINES */ 594/******************************************************************************/ 595