1 #define JEMALLOC_TCACHE_C_
2 #include "jemalloc/internal/jemalloc_internal.h"
3
4 /******************************************************************************/
5 /* Data. */
6
7 bool opt_tcache = true;
8 ssize_t opt_lg_tcache_max = LG_TCACHE_MAXCLASS_DEFAULT;
9
10 tcache_bin_info_t *tcache_bin_info;
11 static unsigned stack_nelms; /* Total stack elms per tcache. */
12
13 unsigned nhbins;
14 size_t tcache_maxclass;
15
16 tcaches_t *tcaches;
17
18 /* Index of first element within tcaches that has never been used. */
19 static unsigned tcaches_past;
20
21 /* Head of singly linked list tracking available tcaches elements. */
22 static tcaches_t *tcaches_avail;
23
24 /******************************************************************************/
25
tcache_salloc(const void * ptr)26 size_t tcache_salloc(const void *ptr)
27 {
28
29 return (arena_salloc(ptr, false));
30 }
31
32 void
tcache_event_hard(tsd_t * tsd,tcache_t * tcache)33 tcache_event_hard(tsd_t *tsd, tcache_t *tcache)
34 {
35 szind_t binind = tcache->next_gc_bin;
36 tcache_bin_t *tbin = &tcache->tbins[binind];
37 tcache_bin_info_t *tbin_info = &tcache_bin_info[binind];
38
39 if (tbin->low_water > 0) {
40 /*
41 * Flush (ceiling) 3/4 of the objects below the low water mark.
42 */
43 if (binind < NBINS) {
44 tcache_bin_flush_small(tsd, tcache, tbin, binind,
45 tbin->ncached - tbin->low_water + (tbin->low_water
46 >> 2));
47 } else {
48 tcache_bin_flush_large(tsd, tbin, binind, tbin->ncached
49 - tbin->low_water + (tbin->low_water >> 2), tcache);
50 }
51 /*
52 * Reduce fill count by 2X. Limit lg_fill_div such that the
53 * fill count is always at least 1.
54 */
55 if ((tbin_info->ncached_max >> (tbin->lg_fill_div+1)) >= 1)
56 tbin->lg_fill_div++;
57 } else if (tbin->low_water < 0) {
58 /*
59 * Increase fill count by 2X. Make sure lg_fill_div stays
60 * greater than 0.
61 */
62 if (tbin->lg_fill_div > 1)
63 tbin->lg_fill_div--;
64 }
65 tbin->low_water = tbin->ncached;
66
67 tcache->next_gc_bin++;
68 if (tcache->next_gc_bin == nhbins)
69 tcache->next_gc_bin = 0;
70 }
71
72 void *
tcache_alloc_small_hard(tsd_t * tsd,arena_t * arena,tcache_t * tcache,tcache_bin_t * tbin,szind_t binind,bool * tcache_success)73 tcache_alloc_small_hard(tsd_t *tsd, arena_t *arena, tcache_t *tcache,
74 tcache_bin_t *tbin, szind_t binind, bool *tcache_success)
75 {
76 void *ret;
77
78 arena_tcache_fill_small(tsd, arena, tbin, binind, config_prof ?
79 tcache->prof_accumbytes : 0);
80 if (config_prof)
81 tcache->prof_accumbytes = 0;
82 ret = tcache_alloc_easy(tbin, tcache_success);
83
84 return (ret);
85 }
86
87 void
tcache_bin_flush_small(tsd_t * tsd,tcache_t * tcache,tcache_bin_t * tbin,szind_t binind,unsigned rem)88 tcache_bin_flush_small(tsd_t *tsd, tcache_t *tcache, tcache_bin_t *tbin,
89 szind_t binind, unsigned rem)
90 {
91 arena_t *arena;
92 void *ptr;
93 unsigned i, nflush, ndeferred;
94 bool merged_stats = false;
95
96 assert(binind < NBINS);
97 assert(rem <= tbin->ncached);
98
99 arena = arena_choose(tsd, NULL);
100 assert(arena != NULL);
101 for (nflush = tbin->ncached - rem; nflush > 0; nflush = ndeferred) {
102 /* Lock the arena bin associated with the first object. */
103 arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(
104 *(tbin->avail - 1));
105 arena_t *bin_arena = extent_node_arena_get(&chunk->node);
106 arena_bin_t *bin = &bin_arena->bins[binind];
107
108 if (config_prof && bin_arena == arena) {
109 if (arena_prof_accum(arena, tcache->prof_accumbytes))
110 prof_idump();
111 tcache->prof_accumbytes = 0;
112 }
113
114 malloc_mutex_lock(&bin->lock);
115 if (config_stats && bin_arena == arena) {
116 assert(!merged_stats);
117 merged_stats = true;
118 bin->stats.nflushes++;
119 bin->stats.nrequests += tbin->tstats.nrequests;
120 tbin->tstats.nrequests = 0;
121 }
122 ndeferred = 0;
123 for (i = 0; i < nflush; i++) {
124 ptr = *(tbin->avail - 1 - i);
125 assert(ptr != NULL);
126 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
127 if (extent_node_arena_get(&chunk->node) == bin_arena) {
128 size_t pageind = ((uintptr_t)ptr -
129 (uintptr_t)chunk) >> LG_PAGE;
130 arena_chunk_map_bits_t *bitselm =
131 arena_bitselm_get(chunk, pageind);
132 arena_dalloc_bin_junked_locked(bin_arena, chunk,
133 ptr, bitselm);
134 } else {
135 /*
136 * This object was allocated via a different
137 * arena bin than the one that is currently
138 * locked. Stash the object, so that it can be
139 * handled in a future pass.
140 */
141 *(tbin->avail - 1 - ndeferred) = ptr;
142 ndeferred++;
143 }
144 }
145 malloc_mutex_unlock(&bin->lock);
146 arena_decay_ticks(tsd, bin_arena, nflush - ndeferred);
147 }
148 if (config_stats && !merged_stats) {
149 /*
150 * The flush loop didn't happen to flush to this thread's
151 * arena, so the stats didn't get merged. Manually do so now.
152 */
153 arena_bin_t *bin = &arena->bins[binind];
154 malloc_mutex_lock(&bin->lock);
155 bin->stats.nflushes++;
156 bin->stats.nrequests += tbin->tstats.nrequests;
157 tbin->tstats.nrequests = 0;
158 malloc_mutex_unlock(&bin->lock);
159 }
160
161 memmove(tbin->avail - rem, tbin->avail - tbin->ncached, rem *
162 sizeof(void *));
163 tbin->ncached = rem;
164 if ((int)tbin->ncached < tbin->low_water)
165 tbin->low_water = tbin->ncached;
166 }
167
168 void
tcache_bin_flush_large(tsd_t * tsd,tcache_bin_t * tbin,szind_t binind,unsigned rem,tcache_t * tcache)169 tcache_bin_flush_large(tsd_t *tsd, tcache_bin_t *tbin, szind_t binind,
170 unsigned rem, tcache_t *tcache)
171 {
172 arena_t *arena;
173 void *ptr;
174 unsigned i, nflush, ndeferred;
175 bool merged_stats = false;
176
177 assert(binind < nhbins);
178 assert(rem <= tbin->ncached);
179
180 arena = arena_choose(tsd, NULL);
181 assert(arena != NULL);
182 for (nflush = tbin->ncached - rem; nflush > 0; nflush = ndeferred) {
183 /* Lock the arena associated with the first object. */
184 arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(
185 *(tbin->avail - 1));
186 arena_t *locked_arena = extent_node_arena_get(&chunk->node);
187 UNUSED bool idump;
188
189 if (config_prof)
190 idump = false;
191 malloc_mutex_lock(&locked_arena->lock);
192 if ((config_prof || config_stats) && locked_arena == arena) {
193 if (config_prof) {
194 idump = arena_prof_accum_locked(arena,
195 tcache->prof_accumbytes);
196 tcache->prof_accumbytes = 0;
197 }
198 if (config_stats) {
199 merged_stats = true;
200 arena->stats.nrequests_large +=
201 tbin->tstats.nrequests;
202 arena->stats.lstats[binind - NBINS].nrequests +=
203 tbin->tstats.nrequests;
204 tbin->tstats.nrequests = 0;
205 }
206 }
207 ndeferred = 0;
208 for (i = 0; i < nflush; i++) {
209 ptr = *(tbin->avail - 1 - i);
210 assert(ptr != NULL);
211 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
212 if (extent_node_arena_get(&chunk->node) ==
213 locked_arena) {
214 arena_dalloc_large_junked_locked(locked_arena,
215 chunk, ptr);
216 } else {
217 /*
218 * This object was allocated via a different
219 * arena than the one that is currently locked.
220 * Stash the object, so that it can be handled
221 * in a future pass.
222 */
223 *(tbin->avail - 1 - ndeferred) = ptr;
224 ndeferred++;
225 }
226 }
227 malloc_mutex_unlock(&locked_arena->lock);
228 if (config_prof && idump)
229 prof_idump();
230 arena_decay_ticks(tsd, locked_arena, nflush - ndeferred);
231 }
232 if (config_stats && !merged_stats) {
233 /*
234 * The flush loop didn't happen to flush to this thread's
235 * arena, so the stats didn't get merged. Manually do so now.
236 */
237 malloc_mutex_lock(&arena->lock);
238 arena->stats.nrequests_large += tbin->tstats.nrequests;
239 arena->stats.lstats[binind - NBINS].nrequests +=
240 tbin->tstats.nrequests;
241 tbin->tstats.nrequests = 0;
242 malloc_mutex_unlock(&arena->lock);
243 }
244
245 memmove(tbin->avail - rem, tbin->avail - tbin->ncached, rem *
246 sizeof(void *));
247 tbin->ncached = rem;
248 if ((int)tbin->ncached < tbin->low_water)
249 tbin->low_water = tbin->ncached;
250 }
251
252 void
tcache_arena_associate(tcache_t * tcache,arena_t * arena)253 tcache_arena_associate(tcache_t *tcache, arena_t *arena)
254 {
255
256 if (config_stats) {
257 /* Link into list of extant tcaches. */
258 malloc_mutex_lock(&arena->lock);
259 ql_elm_new(tcache, link);
260 ql_tail_insert(&arena->tcache_ql, tcache, link);
261 malloc_mutex_unlock(&arena->lock);
262 }
263 }
264
265 void
tcache_arena_reassociate(tcache_t * tcache,arena_t * oldarena,arena_t * newarena)266 tcache_arena_reassociate(tcache_t *tcache, arena_t *oldarena, arena_t *newarena)
267 {
268
269 tcache_arena_dissociate(tcache, oldarena);
270 tcache_arena_associate(tcache, newarena);
271 }
272
273 void
tcache_arena_dissociate(tcache_t * tcache,arena_t * arena)274 tcache_arena_dissociate(tcache_t *tcache, arena_t *arena)
275 {
276
277 if (config_stats) {
278 /* Unlink from list of extant tcaches. */
279 malloc_mutex_lock(&arena->lock);
280 if (config_debug) {
281 bool in_ql = false;
282 tcache_t *iter;
283 ql_foreach(iter, &arena->tcache_ql, link) {
284 if (iter == tcache) {
285 in_ql = true;
286 break;
287 }
288 }
289 assert(in_ql);
290 }
291 ql_remove(&arena->tcache_ql, tcache, link);
292 tcache_stats_merge(tcache, arena);
293 malloc_mutex_unlock(&arena->lock);
294 }
295 }
296
297 tcache_t *
tcache_get_hard(tsd_t * tsd)298 tcache_get_hard(tsd_t *tsd)
299 {
300 arena_t *arena;
301
302 if (!tcache_enabled_get()) {
303 if (tsd_nominal(tsd))
304 tcache_enabled_set(false); /* Memoize. */
305 return (NULL);
306 }
307 arena = arena_choose(tsd, NULL);
308 if (unlikely(arena == NULL))
309 return (NULL);
310 return (tcache_create(tsd, arena));
311 }
312
313 tcache_t *
tcache_create(tsd_t * tsd,arena_t * arena)314 tcache_create(tsd_t *tsd, arena_t *arena)
315 {
316 tcache_t *tcache;
317 size_t size, stack_offset;
318 unsigned i;
319
320 size = offsetof(tcache_t, tbins) + (sizeof(tcache_bin_t) * nhbins);
321 /* Naturally align the pointer stacks. */
322 size = PTR_CEILING(size);
323 stack_offset = size;
324 size += stack_nelms * sizeof(void *);
325 /* Avoid false cacheline sharing. */
326 size = sa2u(size, CACHELINE);
327
328 tcache = ipallocztm(tsd, size, CACHELINE, true, false, true,
329 arena_get(0, false));
330 if (tcache == NULL)
331 return (NULL);
332
333 tcache_arena_associate(tcache, arena);
334
335 ticker_init(&tcache->gc_ticker, TCACHE_GC_INCR);
336
337 assert((TCACHE_NSLOTS_SMALL_MAX & 1U) == 0);
338 for (i = 0; i < nhbins; i++) {
339 tcache->tbins[i].lg_fill_div = 1;
340 stack_offset += tcache_bin_info[i].ncached_max * sizeof(void *);
341 /*
342 * avail points past the available space. Allocations will
343 * access the slots toward higher addresses (for the benefit of
344 * prefetch).
345 */
346 tcache->tbins[i].avail = (void **)((uintptr_t)tcache +
347 (uintptr_t)stack_offset);
348 }
349
350 return (tcache);
351 }
352
353 static void
tcache_destroy(tsd_t * tsd,tcache_t * tcache)354 tcache_destroy(tsd_t *tsd, tcache_t *tcache)
355 {
356 arena_t *arena;
357 unsigned i;
358
359 arena = arena_choose(tsd, NULL);
360 tcache_arena_dissociate(tcache, arena);
361
362 for (i = 0; i < NBINS; i++) {
363 tcache_bin_t *tbin = &tcache->tbins[i];
364 tcache_bin_flush_small(tsd, tcache, tbin, i, 0);
365
366 if (config_stats && tbin->tstats.nrequests != 0) {
367 arena_bin_t *bin = &arena->bins[i];
368 malloc_mutex_lock(&bin->lock);
369 bin->stats.nrequests += tbin->tstats.nrequests;
370 malloc_mutex_unlock(&bin->lock);
371 }
372 }
373
374 for (; i < nhbins; i++) {
375 tcache_bin_t *tbin = &tcache->tbins[i];
376 tcache_bin_flush_large(tsd, tbin, i, 0, tcache);
377
378 if (config_stats && tbin->tstats.nrequests != 0) {
379 malloc_mutex_lock(&arena->lock);
380 arena->stats.nrequests_large += tbin->tstats.nrequests;
381 arena->stats.lstats[i - NBINS].nrequests +=
382 tbin->tstats.nrequests;
383 malloc_mutex_unlock(&arena->lock);
384 }
385 }
386
387 if (config_prof && tcache->prof_accumbytes > 0 &&
388 arena_prof_accum(arena, tcache->prof_accumbytes))
389 prof_idump();
390
391 idalloctm(tsd, tcache, false, true, true);
392 }
393
394 void
tcache_cleanup(tsd_t * tsd)395 tcache_cleanup(tsd_t *tsd)
396 {
397 tcache_t *tcache;
398
399 if (!config_tcache)
400 return;
401
402 if ((tcache = tsd_tcache_get(tsd)) != NULL) {
403 tcache_destroy(tsd, tcache);
404 tsd_tcache_set(tsd, NULL);
405 }
406 }
407
408 void
tcache_enabled_cleanup(tsd_t * tsd)409 tcache_enabled_cleanup(tsd_t *tsd)
410 {
411
412 /* Do nothing. */
413 }
414
415 /* Caller must own arena->lock. */
416 void
tcache_stats_merge(tcache_t * tcache,arena_t * arena)417 tcache_stats_merge(tcache_t *tcache, arena_t *arena)
418 {
419 unsigned i;
420
421 cassert(config_stats);
422
423 /* Merge and reset tcache stats. */
424 for (i = 0; i < NBINS; i++) {
425 arena_bin_t *bin = &arena->bins[i];
426 tcache_bin_t *tbin = &tcache->tbins[i];
427 malloc_mutex_lock(&bin->lock);
428 bin->stats.nrequests += tbin->tstats.nrequests;
429 malloc_mutex_unlock(&bin->lock);
430 tbin->tstats.nrequests = 0;
431 }
432
433 for (; i < nhbins; i++) {
434 malloc_large_stats_t *lstats = &arena->stats.lstats[i - NBINS];
435 tcache_bin_t *tbin = &tcache->tbins[i];
436 arena->stats.nrequests_large += tbin->tstats.nrequests;
437 lstats->nrequests += tbin->tstats.nrequests;
438 tbin->tstats.nrequests = 0;
439 }
440 }
441
442 bool
tcaches_create(tsd_t * tsd,unsigned * r_ind)443 tcaches_create(tsd_t *tsd, unsigned *r_ind)
444 {
445 tcache_t *tcache;
446 tcaches_t *elm;
447
448 if (tcaches == NULL) {
449 tcaches = base_alloc(sizeof(tcache_t *) *
450 (MALLOCX_TCACHE_MAX+1));
451 if (tcaches == NULL)
452 return (true);
453 }
454
455 if (tcaches_avail == NULL && tcaches_past > MALLOCX_TCACHE_MAX)
456 return (true);
457 tcache = tcache_create(tsd, arena_get(0, false));
458 if (tcache == NULL)
459 return (true);
460
461 if (tcaches_avail != NULL) {
462 elm = tcaches_avail;
463 tcaches_avail = tcaches_avail->next;
464 elm->tcache = tcache;
465 *r_ind = (unsigned)(elm - tcaches);
466 } else {
467 elm = &tcaches[tcaches_past];
468 elm->tcache = tcache;
469 *r_ind = tcaches_past;
470 tcaches_past++;
471 }
472
473 return (false);
474 }
475
476 static void
tcaches_elm_flush(tsd_t * tsd,tcaches_t * elm)477 tcaches_elm_flush(tsd_t *tsd, tcaches_t *elm)
478 {
479
480 if (elm->tcache == NULL)
481 return;
482 tcache_destroy(tsd, elm->tcache);
483 elm->tcache = NULL;
484 }
485
486 void
tcaches_flush(tsd_t * tsd,unsigned ind)487 tcaches_flush(tsd_t *tsd, unsigned ind)
488 {
489
490 tcaches_elm_flush(tsd, &tcaches[ind]);
491 }
492
493 void
tcaches_destroy(tsd_t * tsd,unsigned ind)494 tcaches_destroy(tsd_t *tsd, unsigned ind)
495 {
496 tcaches_t *elm = &tcaches[ind];
497 tcaches_elm_flush(tsd, elm);
498 elm->next = tcaches_avail;
499 tcaches_avail = elm;
500 }
501
502 bool
tcache_boot(void)503 tcache_boot(void)
504 {
505 unsigned i;
506
507 /*
508 * If necessary, clamp opt_lg_tcache_max, now that large_maxclass is
509 * known.
510 */
511 if (opt_lg_tcache_max < 0 || (1U << opt_lg_tcache_max) < SMALL_MAXCLASS)
512 tcache_maxclass = SMALL_MAXCLASS;
513 else if ((1U << opt_lg_tcache_max) > large_maxclass)
514 tcache_maxclass = large_maxclass;
515 else
516 tcache_maxclass = (1U << opt_lg_tcache_max);
517
518 nhbins = size2index(tcache_maxclass) + 1;
519
520 /* Initialize tcache_bin_info. */
521 tcache_bin_info = (tcache_bin_info_t *)base_alloc(nhbins *
522 sizeof(tcache_bin_info_t));
523 if (tcache_bin_info == NULL)
524 return (true);
525 stack_nelms = 0;
526 for (i = 0; i < NBINS; i++) {
527 if ((arena_bin_info[i].nregs << 1) <= TCACHE_NSLOTS_SMALL_MIN) {
528 tcache_bin_info[i].ncached_max =
529 TCACHE_NSLOTS_SMALL_MIN;
530 } else if ((arena_bin_info[i].nregs << 1) <=
531 TCACHE_NSLOTS_SMALL_MAX) {
532 tcache_bin_info[i].ncached_max =
533 (arena_bin_info[i].nregs << 1);
534 } else {
535 tcache_bin_info[i].ncached_max =
536 TCACHE_NSLOTS_SMALL_MAX;
537 }
538 stack_nelms += tcache_bin_info[i].ncached_max;
539 }
540 for (; i < nhbins; i++) {
541 tcache_bin_info[i].ncached_max = TCACHE_NSLOTS_LARGE;
542 stack_nelms += tcache_bin_info[i].ncached_max;
543 }
544
545 return (false);
546 }
547