prof.h revision eefdd02e70ec1b9cf11920fcff585835dcbd766b
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 * - Dumping a heap profile that includes this ctx. 133 * nlimbo must be 1 (single destroyer) in order to safely destroy the 134 * ctx. 135 */ 136 unsigned nlimbo; 137 138 /* Temporary storage for summation during dump. */ 139 prof_cnt_t cnt_summed; 140 141 /* When threads exit, they merge their stats into cnt_merged. */ 142 prof_cnt_t cnt_merged; 143 144 /* 145 * List of profile counters, one for each thread that has allocated in 146 * this context. 147 */ 148 ql_head(prof_thr_cnt_t) cnts_ql; 149 150 /* Linkage for list of contexts to be dumped. */ 151 ql_elm(prof_ctx_t) dump_link; 152}; 153typedef ql_head(prof_ctx_t) prof_ctx_list_t; 154 155struct prof_tdata_s { 156 /* 157 * Hash of (prof_bt_t *)-->(prof_thr_cnt_t *). Each thread keeps a 158 * cache of backtraces, with associated thread-specific prof_thr_cnt_t 159 * objects. Other threads may read the prof_thr_cnt_t contents, but no 160 * others will ever write them. 161 * 162 * Upon thread exit, the thread must merge all the prof_thr_cnt_t 163 * counter data into the associated prof_ctx_t objects, and unlink/free 164 * the prof_thr_cnt_t objects. 165 */ 166 ckh_t bt2cnt; 167 168 /* LRU for contents of bt2cnt. */ 169 ql_head(prof_thr_cnt_t) lru_ql; 170 171 /* Backtrace vector, used for calls to prof_backtrace(). */ 172 void **vec; 173 174 /* Sampling state. */ 175 uint64_t prng_state; 176 uint64_t threshold; 177 uint64_t accum; 178 179 /* State used to avoid dumping while operating on prof internals. */ 180 bool enq; 181 bool enq_idump; 182 bool enq_gdump; 183}; 184 185#endif /* JEMALLOC_H_STRUCTS */ 186/******************************************************************************/ 187#ifdef JEMALLOC_H_EXTERNS 188 189extern bool opt_prof; 190/* 191 * Even if opt_prof is true, sampling can be temporarily disabled by setting 192 * opt_prof_active to false. No locking is used when updating opt_prof_active, 193 * so there are no guarantees regarding how long it will take for all threads 194 * to notice state changes. 195 */ 196extern bool opt_prof_active; 197extern size_t opt_lg_prof_sample; /* Mean bytes between samples. */ 198extern ssize_t opt_lg_prof_interval; /* lg(prof_interval). */ 199extern bool opt_prof_gdump; /* High-water memory dumping. */ 200extern bool opt_prof_final; /* Final profile dumping. */ 201extern bool opt_prof_leak; /* Dump leak summary at exit. */ 202extern bool opt_prof_accum; /* Report cumulative bytes. */ 203extern char opt_prof_prefix[ 204 /* Minimize memory bloat for non-prof builds. */ 205#ifdef JEMALLOC_PROF 206 PATH_MAX + 207#endif 208 1]; 209 210/* 211 * Profile dump interval, measured in bytes allocated. Each arena triggers a 212 * profile dump when it reaches this threshold. The effect is that the 213 * interval between profile dumps averages prof_interval, though the actual 214 * interval between dumps will tend to be sporadic, and the interval will be a 215 * maximum of approximately (prof_interval * narenas). 216 */ 217extern uint64_t prof_interval; 218 219/* 220 * If true, promote small sampled objects to large objects, since small run 221 * headers do not have embedded profile context pointers. 222 */ 223extern bool prof_promote; 224 225void bt_init(prof_bt_t *bt, void **vec); 226void prof_backtrace(prof_bt_t *bt, unsigned nignore); 227prof_thr_cnt_t *prof_lookup(prof_bt_t *bt); 228void prof_idump(void); 229bool prof_mdump(const char *filename); 230void prof_gdump(void); 231prof_tdata_t *prof_tdata_init(void); 232void prof_tdata_cleanup(void *arg); 233void prof_boot0(void); 234void prof_boot1(void); 235bool prof_boot2(void); 236void prof_prefork(void); 237void prof_postfork_parent(void); 238void prof_postfork_child(void); 239 240#endif /* JEMALLOC_H_EXTERNS */ 241/******************************************************************************/ 242#ifdef JEMALLOC_H_INLINES 243 244#define PROF_ALLOC_PREP(nignore, size, ret) do { \ 245 prof_tdata_t *prof_tdata; \ 246 prof_bt_t bt; \ 247 \ 248 assert(size == s2u(size)); \ 249 \ 250 prof_tdata = prof_tdata_get(true); \ 251 if ((uintptr_t)prof_tdata <= (uintptr_t)PROF_TDATA_STATE_MAX) { \ 252 if (prof_tdata != NULL) \ 253 ret = (prof_thr_cnt_t *)(uintptr_t)1U; \ 254 else \ 255 ret = NULL; \ 256 break; \ 257 } \ 258 \ 259 if (opt_prof_active == false) { \ 260 /* Sampling is currently inactive, so avoid sampling. */\ 261 ret = (prof_thr_cnt_t *)(uintptr_t)1U; \ 262 } else if (opt_lg_prof_sample == 0) { \ 263 /* Don't bother with sampling logic, since sampling */\ 264 /* interval is 1. */\ 265 bt_init(&bt, prof_tdata->vec); \ 266 prof_backtrace(&bt, nignore); \ 267 ret = prof_lookup(&bt); \ 268 } else { \ 269 if (prof_tdata->threshold == 0) { \ 270 /* Initialize. Seed the prng differently for */\ 271 /* each thread. */\ 272 prof_tdata->prng_state = \ 273 (uint64_t)(uintptr_t)&size; \ 274 prof_sample_threshold_update(prof_tdata); \ 275 } \ 276 \ 277 /* Determine whether to capture a backtrace based on */\ 278 /* whether size is enough for prof_accum to reach */\ 279 /* prof_tdata->threshold. However, delay updating */\ 280 /* these variables until prof_{m,re}alloc(), because */\ 281 /* we don't know for sure that the allocation will */\ 282 /* succeed. */\ 283 /* */\ 284 /* Use subtraction rather than addition to avoid */\ 285 /* potential integer overflow. */\ 286 if (size >= prof_tdata->threshold - \ 287 prof_tdata->accum) { \ 288 bt_init(&bt, prof_tdata->vec); \ 289 prof_backtrace(&bt, nignore); \ 290 ret = prof_lookup(&bt); \ 291 } else \ 292 ret = (prof_thr_cnt_t *)(uintptr_t)1U; \ 293 } \ 294} while (0) 295 296#ifndef JEMALLOC_ENABLE_INLINE 297malloc_tsd_protos(JEMALLOC_ATTR(unused), prof_tdata, prof_tdata_t *) 298 299prof_tdata_t *prof_tdata_get(bool create); 300void prof_sample_threshold_update(prof_tdata_t *prof_tdata); 301prof_ctx_t *prof_ctx_get(const void *ptr); 302void prof_ctx_set(const void *ptr, size_t usize, prof_ctx_t *ctx); 303bool prof_sample_accum_update(size_t size); 304void prof_malloc(const void *ptr, size_t usize, prof_thr_cnt_t *cnt); 305void prof_realloc(const void *ptr, size_t usize, prof_thr_cnt_t *cnt, 306 size_t old_usize, prof_ctx_t *old_ctx); 307void prof_free(const void *ptr, size_t size); 308#endif 309 310#if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_PROF_C_)) 311/* Thread-specific backtrace cache, used to reduce bt2ctx contention. */ 312malloc_tsd_externs(prof_tdata, prof_tdata_t *) 313malloc_tsd_funcs(JEMALLOC_INLINE, prof_tdata, prof_tdata_t *, NULL, 314 prof_tdata_cleanup) 315 316JEMALLOC_INLINE prof_tdata_t * 317prof_tdata_get(bool create) 318{ 319 prof_tdata_t *prof_tdata; 320 321 cassert(config_prof); 322 323 prof_tdata = *prof_tdata_tsd_get(); 324 if (create && prof_tdata == NULL) 325 prof_tdata = prof_tdata_init(); 326 327 return (prof_tdata); 328} 329 330JEMALLOC_INLINE void 331prof_sample_threshold_update(prof_tdata_t *prof_tdata) 332{ 333 /* 334 * The body of this function is compiled out unless heap profiling is 335 * enabled, so that it is possible to compile jemalloc with floating 336 * point support completely disabled. Avoiding floating point code is 337 * important on memory-constrained systems, but it also enables a 338 * workaround for versions of glibc that don't properly save/restore 339 * floating point registers during dynamic lazy symbol loading (which 340 * internally calls into whatever malloc implementation happens to be 341 * integrated into the application). Note that some compilers (e.g. 342 * gcc 4.8) may use floating point registers for fast memory moves, so 343 * jemalloc must be compiled with such optimizations disabled (e.g. 344 * -mno-sse) in order for the workaround to be complete. 345 */ 346#ifdef JEMALLOC_PROF 347 uint64_t r; 348 double u; 349 350 cassert(config_prof); 351 352 /* 353 * Compute sample threshold as a geometrically distributed random 354 * variable with mean (2^opt_lg_prof_sample). 355 * 356 * __ __ 357 * | log(u) | 1 358 * prof_tdata->threshold = | -------- |, where p = ------------------- 359 * | log(1-p) | opt_lg_prof_sample 360 * 2 361 * 362 * For more information on the math, see: 363 * 364 * Non-Uniform Random Variate Generation 365 * Luc Devroye 366 * Springer-Verlag, New York, 1986 367 * pp 500 368 * (http://luc.devroye.org/rnbookindex.html) 369 */ 370 prng64(r, 53, prof_tdata->prng_state, 371 UINT64_C(6364136223846793005), UINT64_C(1442695040888963407)); 372 u = (double)r * (1.0/9007199254740992.0L); 373 prof_tdata->threshold = (uint64_t)(log(u) / 374 log(1.0 - (1.0 / (double)((uint64_t)1U << opt_lg_prof_sample)))) 375 + (uint64_t)1U; 376#endif 377} 378 379JEMALLOC_INLINE prof_ctx_t * 380prof_ctx_get(const void *ptr) 381{ 382 prof_ctx_t *ret; 383 arena_chunk_t *chunk; 384 385 cassert(config_prof); 386 assert(ptr != NULL); 387 388 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 389 if (chunk != ptr) { 390 /* Region. */ 391 ret = arena_prof_ctx_get(ptr); 392 } else 393 ret = huge_prof_ctx_get(ptr); 394 395 return (ret); 396} 397 398JEMALLOC_INLINE void 399prof_ctx_set(const void *ptr, size_t usize, prof_ctx_t *ctx) 400{ 401 arena_chunk_t *chunk; 402 403 cassert(config_prof); 404 assert(ptr != NULL); 405 406 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 407 if (chunk != ptr) { 408 /* Region. */ 409 arena_prof_ctx_set(ptr, usize, ctx); 410 } else 411 huge_prof_ctx_set(ptr, ctx); 412} 413 414JEMALLOC_INLINE bool 415prof_sample_accum_update(size_t size) 416{ 417 prof_tdata_t *prof_tdata; 418 419 cassert(config_prof); 420 /* Sampling logic is unnecessary if the interval is 1. */ 421 assert(opt_lg_prof_sample != 0); 422 423 prof_tdata = prof_tdata_get(false); 424 if ((uintptr_t)prof_tdata <= (uintptr_t)PROF_TDATA_STATE_MAX) 425 return (true); 426 427 /* Take care to avoid integer overflow. */ 428 if (size >= prof_tdata->threshold - prof_tdata->accum) { 429 prof_tdata->accum -= (prof_tdata->threshold - size); 430 /* Compute new sample threshold. */ 431 prof_sample_threshold_update(prof_tdata); 432 while (prof_tdata->accum >= prof_tdata->threshold) { 433 prof_tdata->accum -= prof_tdata->threshold; 434 prof_sample_threshold_update(prof_tdata); 435 } 436 return (false); 437 } else { 438 prof_tdata->accum += size; 439 return (true); 440 } 441} 442 443JEMALLOC_INLINE void 444prof_malloc(const void *ptr, size_t usize, prof_thr_cnt_t *cnt) 445{ 446 447 cassert(config_prof); 448 assert(ptr != NULL); 449 assert(usize == isalloc(ptr, true)); 450 451 if (opt_lg_prof_sample != 0) { 452 if (prof_sample_accum_update(usize)) { 453 /* 454 * Don't sample. For malloc()-like allocation, it is 455 * always possible to tell in advance how large an 456 * object's usable size will be, so there should never 457 * be a difference between the usize passed to 458 * PROF_ALLOC_PREP() and prof_malloc(). 459 */ 460 assert((uintptr_t)cnt == (uintptr_t)1U); 461 } 462 } 463 464 if ((uintptr_t)cnt > (uintptr_t)1U) { 465 prof_ctx_set(ptr, usize, cnt->ctx); 466 467 cnt->epoch++; 468 /*********/ 469 mb_write(); 470 /*********/ 471 cnt->cnts.curobjs++; 472 cnt->cnts.curbytes += usize; 473 if (opt_prof_accum) { 474 cnt->cnts.accumobjs++; 475 cnt->cnts.accumbytes += usize; 476 } 477 /*********/ 478 mb_write(); 479 /*********/ 480 cnt->epoch++; 481 /*********/ 482 mb_write(); 483 /*********/ 484 } else 485 prof_ctx_set(ptr, usize, (prof_ctx_t *)(uintptr_t)1U); 486} 487 488JEMALLOC_INLINE void 489prof_realloc(const void *ptr, size_t usize, prof_thr_cnt_t *cnt, 490 size_t old_usize, prof_ctx_t *old_ctx) 491{ 492 prof_thr_cnt_t *told_cnt; 493 494 cassert(config_prof); 495 assert(ptr != NULL || (uintptr_t)cnt <= (uintptr_t)1U); 496 497 if (ptr != NULL) { 498 assert(usize == isalloc(ptr, true)); 499 if (opt_lg_prof_sample != 0) { 500 if (prof_sample_accum_update(usize)) { 501 /* 502 * Don't sample. The usize passed to 503 * PROF_ALLOC_PREP() was larger than what 504 * actually got allocated, so a backtrace was 505 * captured for this allocation, even though 506 * its actual usize was insufficient to cross 507 * the sample threshold. 508 */ 509 cnt = (prof_thr_cnt_t *)(uintptr_t)1U; 510 } 511 } 512 } 513 514 if ((uintptr_t)old_ctx > (uintptr_t)1U) { 515 told_cnt = prof_lookup(old_ctx->bt); 516 if (told_cnt == NULL) { 517 /* 518 * It's too late to propagate OOM for this realloc(), 519 * so operate directly on old_cnt->ctx->cnt_merged. 520 */ 521 malloc_mutex_lock(old_ctx->lock); 522 old_ctx->cnt_merged.curobjs--; 523 old_ctx->cnt_merged.curbytes -= old_usize; 524 malloc_mutex_unlock(old_ctx->lock); 525 told_cnt = (prof_thr_cnt_t *)(uintptr_t)1U; 526 } 527 } else 528 told_cnt = (prof_thr_cnt_t *)(uintptr_t)1U; 529 530 if ((uintptr_t)told_cnt > (uintptr_t)1U) 531 told_cnt->epoch++; 532 if ((uintptr_t)cnt > (uintptr_t)1U) { 533 prof_ctx_set(ptr, usize, cnt->ctx); 534 cnt->epoch++; 535 } else if (ptr != NULL) 536 prof_ctx_set(ptr, usize, (prof_ctx_t *)(uintptr_t)1U); 537 /*********/ 538 mb_write(); 539 /*********/ 540 if ((uintptr_t)told_cnt > (uintptr_t)1U) { 541 told_cnt->cnts.curobjs--; 542 told_cnt->cnts.curbytes -= old_usize; 543 } 544 if ((uintptr_t)cnt > (uintptr_t)1U) { 545 cnt->cnts.curobjs++; 546 cnt->cnts.curbytes += usize; 547 if (opt_prof_accum) { 548 cnt->cnts.accumobjs++; 549 cnt->cnts.accumbytes += usize; 550 } 551 } 552 /*********/ 553 mb_write(); 554 /*********/ 555 if ((uintptr_t)told_cnt > (uintptr_t)1U) 556 told_cnt->epoch++; 557 if ((uintptr_t)cnt > (uintptr_t)1U) 558 cnt->epoch++; 559 /*********/ 560 mb_write(); /* Not strictly necessary. */ 561} 562 563JEMALLOC_INLINE void 564prof_free(const void *ptr, size_t size) 565{ 566 prof_ctx_t *ctx = prof_ctx_get(ptr); 567 568 cassert(config_prof); 569 570 if ((uintptr_t)ctx > (uintptr_t)1) { 571 prof_thr_cnt_t *tcnt; 572 assert(size == isalloc(ptr, true)); 573 tcnt = prof_lookup(ctx->bt); 574 575 if (tcnt != NULL) { 576 tcnt->epoch++; 577 /*********/ 578 mb_write(); 579 /*********/ 580 tcnt->cnts.curobjs--; 581 tcnt->cnts.curbytes -= size; 582 /*********/ 583 mb_write(); 584 /*********/ 585 tcnt->epoch++; 586 /*********/ 587 mb_write(); 588 /*********/ 589 } else { 590 /* 591 * OOM during free() cannot be propagated, so operate 592 * directly on cnt->ctx->cnt_merged. 593 */ 594 malloc_mutex_lock(ctx->lock); 595 ctx->cnt_merged.curobjs--; 596 ctx->cnt_merged.curbytes -= size; 597 malloc_mutex_unlock(ctx->lock); 598 } 599 } 600} 601#endif 602 603#endif /* JEMALLOC_H_INLINES */ 604/******************************************************************************/ 605