1#define JEMALLOC_ARENA_C_
2#include "jemalloc/internal/jemalloc_internal.h"
3
4/******************************************************************************/
5/* Data. */
6
7purge_mode_t opt_purge = PURGE_DEFAULT;
8const char *purge_mode_names[] = {
9 "ratio",
10 "decay",
11 "N/A"
12};
13ssize_t opt_lg_dirty_mult = LG_DIRTY_MULT_DEFAULT;
14static ssize_t lg_dirty_mult_default;
15ssize_t opt_decay_time = DECAY_TIME_DEFAULT;
16static ssize_t decay_time_default;
17
18arena_bin_info_t arena_bin_info[NBINS];
19
20size_t map_bias;
21size_t map_misc_offset;
22size_t arena_maxrun; /* Max run size for arenas. */
23size_t large_maxclass; /* Max large size class. */
24size_t run_quantize_max; /* Max run_quantize_*() input. */
25static size_t small_maxrun; /* Max run size for small size classes. */
26static bool *small_run_tab; /* Valid small run page multiples. */
27static size_t *run_quantize_floor_tab; /* run_quantize_floor() memoization. */
28static size_t *run_quantize_ceil_tab; /* run_quantize_ceil() memoization. */
29unsigned nlclasses; /* Number of large size classes. */
30unsigned nhclasses; /* Number of huge size classes. */
31static szind_t runs_avail_bias; /* Size index for first runs_avail tree. */
32static szind_t runs_avail_nclasses; /* Number of runs_avail trees. */
33
34/******************************************************************************/
35/*
36 * Function prototypes for static functions that are referenced prior to
37 * definition.
38 */
39
40static void arena_purge_to_limit(arena_t *arena, size_t ndirty_limit);
41static void arena_run_dalloc(arena_t *arena, arena_run_t *run, bool dirty,
42 bool cleaned, bool decommitted);
43static void arena_dalloc_bin_run(arena_t *arena, arena_chunk_t *chunk,
44 arena_run_t *run, arena_bin_t *bin);
45static void arena_bin_lower_run(arena_t *arena, arena_chunk_t *chunk,
46 arena_run_t *run, arena_bin_t *bin);
47
48/******************************************************************************/
49
50JEMALLOC_INLINE_C size_t
51arena_miscelm_size_get(const arena_chunk_map_misc_t *miscelm)
52{
53 arena_chunk_t *chunk;
54 size_t pageind, mapbits;
55
56 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(miscelm);
57 pageind = arena_miscelm_to_pageind(miscelm);
58 mapbits = arena_mapbits_get(chunk, pageind);
59 return (arena_mapbits_size_decode(mapbits));
60}
61
62JEMALLOC_INLINE_C int
63arena_run_addr_comp(const arena_chunk_map_misc_t *a,
64 const arena_chunk_map_misc_t *b)
65{
66 uintptr_t a_miscelm = (uintptr_t)a;
67 uintptr_t b_miscelm = (uintptr_t)b;
68
69 assert(a != NULL);
70 assert(b != NULL);
71
72 return ((a_miscelm > b_miscelm) - (a_miscelm < b_miscelm));
73}
74
75/* Generate red-black tree functions. */
76rb_gen(static UNUSED, arena_run_tree_, arena_run_tree_t, arena_chunk_map_misc_t,
77 rb_link, arena_run_addr_comp)
78
79static size_t
80run_quantize_floor_compute(size_t size)
81{
82 size_t qsize;
83
84 assert(size != 0);
85 assert(size == PAGE_CEILING(size));
86
87 /* Don't change sizes that are valid small run sizes. */
88 if (size <= small_maxrun && small_run_tab[size >> LG_PAGE])
89 return (size);
90
91 /*
92 * Round down to the nearest run size that can actually be requested
93 * during normal large allocation. Add large_pad so that cache index
94 * randomization can offset the allocation from the page boundary.
95 */
96 qsize = index2size(size2index(size - large_pad + 1) - 1) + large_pad;
97 if (qsize <= SMALL_MAXCLASS + large_pad)
98 return (run_quantize_floor_compute(size - large_pad));
99 assert(qsize <= size);
100 return (qsize);
101}
102
103static size_t
104run_quantize_ceil_compute_hard(size_t size)
105{
106 size_t large_run_size_next;
107
108 assert(size != 0);
109 assert(size == PAGE_CEILING(size));
110
111 /*
112 * Return the next quantized size greater than the input size.
--- 17 unchanged lines hidden (view full) ---
130 if (large_run_size_next < size)
131 return (large_run_size_next);
132 return (size);
133 }
134 }
135}
136
137static size_t
138run_quantize_ceil_compute(size_t size)
139{
140 size_t qsize = run_quantize_floor_compute(size);
141
142 if (qsize < size) {
143 /*
144 * Skip a quantization that may have an adequately large run,
145 * because under-sized runs may be mixed in. This only happens
146 * when an unusual size is requested, i.e. for aligned
147 * allocation, and is just one of several places where linear
148 * search would potentially find sufficiently aligned available
149 * memory somewhere lower.
150 */
151 qsize = run_quantize_ceil_compute_hard(qsize);
152 }
153 return (qsize);
154}
155
156#ifdef JEMALLOC_JET
157#undef run_quantize_floor
158#define run_quantize_floor JEMALLOC_N(run_quantize_floor_impl)
159#endif
160static size_t
161run_quantize_floor(size_t size)
162{
163 size_t ret;
164
165 assert(size > 0);
166 assert(size <= run_quantize_max);
167 assert((size & PAGE_MASK) == 0);
168
169 ret = run_quantize_floor_tab[(size >> LG_PAGE) - 1];
170 assert(ret == run_quantize_floor_compute(size));
171 return (ret);
172}
173#ifdef JEMALLOC_JET
174#undef run_quantize_floor
175#define run_quantize_floor JEMALLOC_N(run_quantize_floor)
176run_quantize_t *run_quantize_floor = JEMALLOC_N(run_quantize_floor_impl);
177#endif
178
179#ifdef JEMALLOC_JET
180#undef run_quantize_ceil
181#define run_quantize_ceil JEMALLOC_N(run_quantize_ceil_impl)
182#endif
183static size_t
184run_quantize_ceil(size_t size)
185{
186 size_t ret;
187
188 assert(size > 0);
189 assert(size <= run_quantize_max);
190 assert((size & PAGE_MASK) == 0);
191
192 ret = run_quantize_ceil_tab[(size >> LG_PAGE) - 1];
193 assert(ret == run_quantize_ceil_compute(size));
194 return (ret);
195}
196#ifdef JEMALLOC_JET
197#undef run_quantize_ceil
198#define run_quantize_ceil JEMALLOC_N(run_quantize_ceil)
199run_quantize_t *run_quantize_ceil = JEMALLOC_N(run_quantize_ceil_impl);
200#endif
201
202static arena_run_tree_t *
203arena_runs_avail_get(arena_t *arena, szind_t ind)
204{
205
206 assert(ind >= runs_avail_bias);
207 assert(ind - runs_avail_bias < runs_avail_nclasses);
208
209 return (&arena->runs_avail[ind - runs_avail_bias]);
210}
211
212static void
213arena_avail_insert(arena_t *arena, arena_chunk_t *chunk, size_t pageind,
214 size_t npages)
215{
216 szind_t ind = size2index(run_quantize_floor(arena_miscelm_size_get(
217 arena_miscelm_get(chunk, pageind))));
218 assert(npages == (arena_mapbits_unallocated_size_get(chunk, pageind) >>
219 LG_PAGE));
220 arena_run_tree_insert(arena_runs_avail_get(arena, ind),
221 arena_miscelm_get(chunk, pageind));
222}
223
224static void
225arena_avail_remove(arena_t *arena, arena_chunk_t *chunk, size_t pageind,
226 size_t npages)
227{
228 szind_t ind = size2index(run_quantize_floor(arena_miscelm_size_get(
229 arena_miscelm_get(chunk, pageind))));
230 assert(npages == (arena_mapbits_unallocated_size_get(chunk, pageind) >>
231 LG_PAGE));
232 arena_run_tree_remove(arena_runs_avail_get(arena, ind),
233 arena_miscelm_get(chunk, pageind));
234}
235
236static void
237arena_run_dirty_insert(arena_t *arena, arena_chunk_t *chunk, size_t pageind,
238 size_t npages)
239{
240 arena_chunk_map_misc_t *miscelm = arena_miscelm_get(chunk, pageind);
241
--- 54 unchanged lines hidden (view full) ---
296 arena->ndirty -= arena_chunk_dirty_npages(node);
297 }
298}
299
300JEMALLOC_INLINE_C void *
301arena_run_reg_alloc(arena_run_t *run, arena_bin_info_t *bin_info)
302{
303 void *ret;
304 size_t regind;
305 arena_chunk_map_misc_t *miscelm;
306 void *rpages;
307
308 assert(run->nfree > 0);
309 assert(!bitmap_full(run->bitmap, &bin_info->bitmap_info));
310
311 regind = (unsigned)bitmap_sfu(run->bitmap, &bin_info->bitmap_info);
312 miscelm = arena_run_to_miscelm(run);
313 rpages = arena_miscelm_to_rpages(miscelm);
314 ret = (void *)((uintptr_t)rpages + (uintptr_t)bin_info->reg0_offset +
315 (uintptr_t)(bin_info->reg_interval * regind));
316 run->nfree--;
317 return (ret);
318}
319
320JEMALLOC_INLINE_C void
321arena_run_reg_dalloc(arena_run_t *run, void *ptr)
322{
323 arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run);
324 size_t pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
325 size_t mapbits = arena_mapbits_get(chunk, pageind);
326 szind_t binind = arena_ptr_small_binind_get(ptr, mapbits);
327 arena_bin_info_t *bin_info = &arena_bin_info[binind];
328 size_t regind = arena_run_regind(run, bin_info, ptr);
329
330 assert(run->nfree < bin_info->nregs);
331 /* Freeing an interior pointer can cause assertion failure. */
332 assert(((uintptr_t)ptr -
333 ((uintptr_t)arena_miscelm_to_rpages(arena_run_to_miscelm(run)) +
334 (uintptr_t)bin_info->reg0_offset)) %
335 (uintptr_t)bin_info->reg_interval == 0);
336 assert((uintptr_t)ptr >=
--- 31 unchanged lines hidden (view full) ---
368 UNUSED size_t *p = (size_t *)((uintptr_t)chunk + (run_ind << LG_PAGE));
369
370 arena_run_page_mark_zeroed(chunk, run_ind);
371 for (i = 0; i < PAGE / sizeof(size_t); i++)
372 assert(p[i] == 0);
373}
374
375static void
376arena_nactive_add(arena_t *arena, size_t add_pages)
377{
378
379 if (config_stats) {
380 size_t cactive_add = CHUNK_CEILING((arena->nactive +
381 add_pages) << LG_PAGE) - CHUNK_CEILING(arena->nactive <<
382 LG_PAGE);
383 if (cactive_add != 0)
384 stats_cactive_add(cactive_add);
385 }
386 arena->nactive += add_pages;
387}
388
389static void
390arena_nactive_sub(arena_t *arena, size_t sub_pages)
391{
392
393 if (config_stats) {
394 size_t cactive_sub = CHUNK_CEILING(arena->nactive << LG_PAGE) -
395 CHUNK_CEILING((arena->nactive - sub_pages) << LG_PAGE);
396 if (cactive_sub != 0)
397 stats_cactive_sub(cactive_sub);
398 }
399 arena->nactive -= sub_pages;
400}
401
402static void
403arena_run_split_remove(arena_t *arena, arena_chunk_t *chunk, size_t run_ind,
404 size_t flag_dirty, size_t flag_decommitted, size_t need_pages)
405{
406 size_t total_pages, rem_pages;
407
408 assert(flag_dirty == 0 || flag_decommitted == 0);
409
410 total_pages = arena_mapbits_unallocated_size_get(chunk, run_ind) >>
411 LG_PAGE;
412 assert(arena_mapbits_dirty_get(chunk, run_ind+total_pages-1) ==
413 flag_dirty);
414 assert(need_pages <= total_pages);
415 rem_pages = total_pages - need_pages;
416
417 arena_avail_remove(arena, chunk, run_ind, total_pages);
418 if (flag_dirty != 0)
419 arena_run_dirty_remove(arena, chunk, run_ind, total_pages);
420 arena_nactive_add(arena, need_pages);
421
422 /* Keep track of trailing unused pages for later use. */
423 if (rem_pages > 0) {
424 size_t flags = flag_dirty | flag_decommitted;
425 size_t flag_unzeroed_mask = (flags == 0) ? CHUNK_MAP_UNZEROED :
426 0;
427
428 arena_mapbits_unallocated_set(chunk, run_ind+need_pages,
--- 299 unchanged lines hidden (view full) ---
728 if (arena->spare != NULL)
729 chunk = arena_chunk_init_spare(arena);
730 else {
731 chunk = arena_chunk_init_hard(arena);
732 if (chunk == NULL)
733 return (NULL);
734 }
735
736 arena_avail_insert(arena, chunk, map_bias, chunk_npages-map_bias);
737
738 return (chunk);
739}
740
741static void
742arena_chunk_dalloc(arena_t *arena, arena_chunk_t *chunk)
743{
--- 4 unchanged lines hidden (view full) ---
748 arena_maxrun);
749 assert(arena_mapbits_unallocated_size_get(chunk, chunk_npages-1) ==
750 arena_maxrun);
751 assert(arena_mapbits_dirty_get(chunk, map_bias) ==
752 arena_mapbits_dirty_get(chunk, chunk_npages-1));
753 assert(arena_mapbits_decommitted_get(chunk, map_bias) ==
754 arena_mapbits_decommitted_get(chunk, chunk_npages-1));
755
756 /* Remove run from runs_avail, so that the arena does not use it. */
757 arena_avail_remove(arena, chunk, map_bias, chunk_npages-map_bias);
758
759 if (arena->spare != NULL) {
760 arena_chunk_t *spare = arena->spare;
761 chunk_hooks_t chunk_hooks = CHUNK_HOOKS_INITIALIZER;
762 bool committed;
763
764 arena->spare = chunk;
--- 136 unchanged lines hidden (view full) ---
901 zero, &commit);
902 if (ret == NULL) {
903 /* Revert optimistic stats updates. */
904 malloc_mutex_lock(&arena->lock);
905 if (config_stats) {
906 arena_huge_malloc_stats_update_undo(arena, usize);
907 arena->stats.mapped -= usize;
908 }
909 arena_nactive_sub(arena, usize >> LG_PAGE);
910 malloc_mutex_unlock(&arena->lock);
911 }
912
913 return (ret);
914}
915
916void *
917arena_chunk_alloc_huge(arena_t *arena, size_t usize, size_t alignment,
--- 5 unchanged lines hidden (view full) ---
923
924 malloc_mutex_lock(&arena->lock);
925
926 /* Optimistically update stats. */
927 if (config_stats) {
928 arena_huge_malloc_stats_update(arena, usize);
929 arena->stats.mapped += usize;
930 }
931 arena_nactive_add(arena, usize >> LG_PAGE);
932
933 ret = chunk_alloc_cache(arena, &chunk_hooks, NULL, csize, alignment,
934 zero, true);
935 malloc_mutex_unlock(&arena->lock);
936 if (ret == NULL) {
937 ret = arena_chunk_alloc_huge_hard(arena, &chunk_hooks, usize,
938 alignment, zero, csize);
939 }
940
941 return (ret);
942}
943
944void
945arena_chunk_dalloc_huge(arena_t *arena, void *chunk, size_t usize)
946{
947 chunk_hooks_t chunk_hooks = CHUNK_HOOKS_INITIALIZER;
948 size_t csize;
949
950 csize = CHUNK_CEILING(usize);
951 malloc_mutex_lock(&arena->lock);
952 if (config_stats) {
953 arena_huge_dalloc_stats_update(arena, usize);
954 arena->stats.mapped -= usize;
955 }
956 arena_nactive_sub(arena, usize >> LG_PAGE);
957
958 chunk_dalloc_cache(arena, &chunk_hooks, chunk, csize, true);
959 malloc_mutex_unlock(&arena->lock);
960}
961
962void
963arena_chunk_ralloc_huge_similar(arena_t *arena, void *chunk, size_t oldsize,
964 size_t usize)
965{
966
967 assert(CHUNK_CEILING(oldsize) == CHUNK_CEILING(usize));
968 assert(oldsize != usize);
969
970 malloc_mutex_lock(&arena->lock);
971 if (config_stats)
972 arena_huge_ralloc_stats_update(arena, oldsize, usize);
973 if (oldsize < usize)
974 arena_nactive_add(arena, (usize - oldsize) >> LG_PAGE);
975 else
976 arena_nactive_sub(arena, (oldsize - usize) >> LG_PAGE);
977 malloc_mutex_unlock(&arena->lock);
978}
979
980void
981arena_chunk_ralloc_huge_shrink(arena_t *arena, void *chunk, size_t oldsize,
982 size_t usize)
983{
984 size_t udiff = oldsize - usize;
985 size_t cdiff = CHUNK_CEILING(oldsize) - CHUNK_CEILING(usize);
986
987 malloc_mutex_lock(&arena->lock);
988 if (config_stats) {
989 arena_huge_ralloc_stats_update(arena, oldsize, usize);
990 if (cdiff != 0)
991 arena->stats.mapped -= cdiff;
992 }
993 arena_nactive_sub(arena, udiff >> LG_PAGE);
994
995 if (cdiff != 0) {
996 chunk_hooks_t chunk_hooks = CHUNK_HOOKS_INITIALIZER;
997 void *nchunk = (void *)((uintptr_t)chunk +
998 CHUNK_CEILING(usize));
999
1000 chunk_dalloc_cache(arena, &chunk_hooks, nchunk, cdiff, true);
1001 }
--- 13 unchanged lines hidden (view full) ---
1015 if (err) {
1016 /* Revert optimistic stats updates. */
1017 malloc_mutex_lock(&arena->lock);
1018 if (config_stats) {
1019 arena_huge_ralloc_stats_update_undo(arena, oldsize,
1020 usize);
1021 arena->stats.mapped -= cdiff;
1022 }
1023 arena_nactive_sub(arena, udiff >> LG_PAGE);
1024 malloc_mutex_unlock(&arena->lock);
1025 } else if (chunk_hooks->merge(chunk, CHUNK_CEILING(oldsize), nchunk,
1026 cdiff, true, arena->ind)) {
1027 chunk_dalloc_arena(arena, chunk_hooks, nchunk, cdiff, *zero,
1028 true);
1029 err = true;
1030 }
1031 return (err);
--- 11 unchanged lines hidden (view full) ---
1043
1044 malloc_mutex_lock(&arena->lock);
1045
1046 /* Optimistically update stats. */
1047 if (config_stats) {
1048 arena_huge_ralloc_stats_update(arena, oldsize, usize);
1049 arena->stats.mapped += cdiff;
1050 }
1051 arena_nactive_add(arena, udiff >> LG_PAGE);
1052
1053 err = (chunk_alloc_cache(arena, &arena->chunk_hooks, nchunk, cdiff,
1054 chunksize, zero, true) == NULL);
1055 malloc_mutex_unlock(&arena->lock);
1056 if (err) {
1057 err = arena_chunk_ralloc_huge_expand_hard(arena, &chunk_hooks,
1058 chunk, oldsize, usize, zero, nchunk, udiff,
1059 cdiff);
1060 } else if (chunk_hooks.merge(chunk, CHUNK_CEILING(oldsize), nchunk,
1061 cdiff, true, arena->ind)) {
1062 chunk_dalloc_arena(arena, &chunk_hooks, nchunk, cdiff, *zero,
1063 true);
1064 err = true;
1065 }
1066
1067 return (err);
1068}
1069
1070/*
1071 * Do first-best-fit run selection, i.e. select the lowest run that best fits.
1072 * Run sizes are indexed, so not all candidate runs are necessarily exactly the
1073 * same size.
1074 */
1075static arena_run_t *
1076arena_run_first_best_fit(arena_t *arena, size_t size)
1077{
1078 szind_t ind, i;
1079
1080 ind = size2index(run_quantize_ceil(size));
1081 for (i = ind; i < runs_avail_nclasses + runs_avail_bias; i++) {
1082 arena_chunk_map_misc_t *miscelm = arena_run_tree_first(
1083 arena_runs_avail_get(arena, i));
1084 if (miscelm != NULL)
1085 return (&miscelm->run);
1086 }
1087
1088 return (NULL);
1089}
1090
1091static arena_run_t *
1092arena_run_alloc_large_helper(arena_t *arena, size_t size, bool zero)
1093{
1094 arena_run_t *run = arena_run_first_best_fit(arena, s2u(size));
1095 if (run != NULL) {
1096 if (arena_run_split_large(arena, run, size, zero))
--- 110 unchanged lines hidden (view full) ---
1207 malloc_mutex_lock(&arena->lock);
1208 arena->lg_dirty_mult = lg_dirty_mult;
1209 arena_maybe_purge(arena);
1210 malloc_mutex_unlock(&arena->lock);
1211
1212 return (false);
1213}
1214
1215static void
1216arena_decay_deadline_init(arena_t *arena)
1217{
1218
1219 assert(opt_purge == purge_mode_decay);
1220
1221 /*
1222 * Generate a new deadline that is uniformly random within the next
1223 * epoch after the current one.
1224 */
1225 nstime_copy(&arena->decay_deadline, &arena->decay_epoch);
1226 nstime_add(&arena->decay_deadline, &arena->decay_interval);
1227 if (arena->decay_time > 0) {
1228 nstime_t jitter;
1229
1230 nstime_init(&jitter, prng_range(&arena->decay_jitter_state,
1231 nstime_ns(&arena->decay_interval)));
1232 nstime_add(&arena->decay_deadline, &jitter);
1233 }
1234}
1235
1236static bool
1237arena_decay_deadline_reached(const arena_t *arena, const nstime_t *time)
1238{
1239
1240 assert(opt_purge == purge_mode_decay);
1241
1242 return (nstime_compare(&arena->decay_deadline, time) <= 0);
1243}
1244
1245static size_t
1246arena_decay_backlog_npages_limit(const arena_t *arena)
1247{
1248 static const uint64_t h_steps[] = {
1249#define STEP(step, h, x, y) \
1250 h,
1251 SMOOTHSTEP
1252#undef STEP
1253 };
1254 uint64_t sum;
1255 size_t npages_limit_backlog;
1256 unsigned i;
1257
1258 assert(opt_purge == purge_mode_decay);
1259
1260 /*
1261 * For each element of decay_backlog, multiply by the corresponding
1262 * fixed-point smoothstep decay factor. Sum the products, then divide
1263 * to round down to the nearest whole number of pages.
1264 */
1265 sum = 0;
1266 for (i = 0; i < SMOOTHSTEP_NSTEPS; i++)
1267 sum += arena->decay_backlog[i] * h_steps[i];
1268 npages_limit_backlog = (sum >> SMOOTHSTEP_BFP);
1269
1270 return (npages_limit_backlog);
1271}
1272
1273static void
1274arena_decay_epoch_advance(arena_t *arena, const nstime_t *time)
1275{
1276 uint64_t nadvance;
1277 nstime_t delta;
1278 size_t ndirty_delta;
1279
1280 assert(opt_purge == purge_mode_decay);
1281 assert(arena_decay_deadline_reached(arena, time));
1282
1283 nstime_copy(&delta, time);
1284 nstime_subtract(&delta, &arena->decay_epoch);
1285 nadvance = nstime_divide(&delta, &arena->decay_interval);
1286 assert(nadvance > 0);
1287
1288 /* Add nadvance decay intervals to epoch. */
1289 nstime_copy(&delta, &arena->decay_interval);
1290 nstime_imultiply(&delta, nadvance);
1291 nstime_add(&arena->decay_epoch, &delta);
1292
1293 /* Set a new deadline. */
1294 arena_decay_deadline_init(arena);
1295
1296 /* Update the backlog. */
1297 if (nadvance >= SMOOTHSTEP_NSTEPS) {
1298 memset(arena->decay_backlog, 0, (SMOOTHSTEP_NSTEPS-1) *
1299 sizeof(size_t));
1300 } else {
1301 memmove(arena->decay_backlog, &arena->decay_backlog[nadvance],
1302 (SMOOTHSTEP_NSTEPS - nadvance) * sizeof(size_t));
1303 if (nadvance > 1) {
1304 memset(&arena->decay_backlog[SMOOTHSTEP_NSTEPS -
1305 nadvance], 0, (nadvance-1) * sizeof(size_t));
1306 }
1307 }
1308 ndirty_delta = (arena->ndirty > arena->decay_ndirty) ? arena->ndirty -
1309 arena->decay_ndirty : 0;
1310 arena->decay_ndirty = arena->ndirty;
1311 arena->decay_backlog[SMOOTHSTEP_NSTEPS-1] = ndirty_delta;
1312 arena->decay_backlog_npages_limit =
1313 arena_decay_backlog_npages_limit(arena);
1314}
1315
1316static size_t
1317arena_decay_npages_limit(arena_t *arena)
1318{
1319 size_t npages_limit;
1320
1321 assert(opt_purge == purge_mode_decay);
1322
1323 npages_limit = arena->decay_backlog_npages_limit;
1324
1325 /* Add in any dirty pages created during the current epoch. */
1326 if (arena->ndirty > arena->decay_ndirty)
1327 npages_limit += arena->ndirty - arena->decay_ndirty;
1328
1329 return (npages_limit);
1330}
1331
1332static void
1333arena_decay_init(arena_t *arena, ssize_t decay_time)
1334{
1335
1336 arena->decay_time = decay_time;
1337 if (decay_time > 0) {
1338 nstime_init2(&arena->decay_interval, decay_time, 0);
1339 nstime_idivide(&arena->decay_interval, SMOOTHSTEP_NSTEPS);
1340 }
1341
1342 nstime_init(&arena->decay_epoch, 0);
1343 nstime_update(&arena->decay_epoch);
1344 arena->decay_jitter_state = (uint64_t)(uintptr_t)arena;
1345 arena_decay_deadline_init(arena);
1346 arena->decay_ndirty = arena->ndirty;
1347 arena->decay_backlog_npages_limit = 0;
1348 memset(arena->decay_backlog, 0, SMOOTHSTEP_NSTEPS * sizeof(size_t));
1349}
1350
1351static bool
1352arena_decay_time_valid(ssize_t decay_time)
1353{
1354
1355 return (decay_time >= -1 && decay_time <= NSTIME_SEC_MAX);
1356}
1357
1358ssize_t
1359arena_decay_time_get(arena_t *arena)
1360{
1361 ssize_t decay_time;
1362
1363 malloc_mutex_lock(&arena->lock);
1364 decay_time = arena->decay_time;
1365 malloc_mutex_unlock(&arena->lock);
1366
1367 return (decay_time);
1368}
1369
1370bool
1371arena_decay_time_set(arena_t *arena, ssize_t decay_time)
1372{
1373
1374 if (!arena_decay_time_valid(decay_time))
1375 return (true);
1376
1377 malloc_mutex_lock(&arena->lock);
1378 /*
1379 * Restart decay backlog from scratch, which may cause many dirty pages
1380 * to be immediately purged. It would conceptually be possible to map
1381 * the old backlog onto the new backlog, but there is no justification
1382 * for such complexity since decay_time changes are intended to be
1383 * infrequent, either between the {-1, 0, >0} states, or a one-time
1384 * arbitrary change during initial arena configuration.
1385 */
1386 arena_decay_init(arena, decay_time);
1387 arena_maybe_purge(arena);
1388 malloc_mutex_unlock(&arena->lock);
1389
1390 return (false);
1391}
1392
1393static void
1394arena_maybe_purge_ratio(arena_t *arena)
1395{
1396
1397 assert(opt_purge == purge_mode_ratio);
1398
1399 /* Don't purge if the option is disabled. */
1400 if (arena->lg_dirty_mult < 0)
1401 return;
1402
1403 /*
1404 * Iterate, since preventing recursive purging could otherwise leave too
1405 * many dirty pages.
1406 */
1407 while (true) {
1408 size_t threshold = (arena->nactive >> arena->lg_dirty_mult);
1409 if (threshold < chunk_npages)
1410 threshold = chunk_npages;
1411 /*
1412 * Don't purge unless the number of purgeable pages exceeds the
1413 * threshold.
1414 */
1415 if (arena->ndirty <= threshold)
1416 return;
1417 arena_purge_to_limit(arena, threshold);
1418 }
1419}
1420
1421static void
1422arena_maybe_purge_decay(arena_t *arena)
1423{
1424 nstime_t time;
1425 size_t ndirty_limit;
1426
1427 assert(opt_purge == purge_mode_decay);
1428
1429 /* Purge all or nothing if the option is disabled. */
1430 if (arena->decay_time <= 0) {
1431 if (arena->decay_time == 0)
1432 arena_purge_to_limit(arena, 0);
1433 return;
1434 }
1435
1436 nstime_copy(&time, &arena->decay_epoch);
1437 if (unlikely(nstime_update(&time))) {
1438 /* Time went backwards. Force an epoch advance. */
1439 nstime_copy(&time, &arena->decay_deadline);
1440 }
1441
1442 if (arena_decay_deadline_reached(arena, &time))
1443 arena_decay_epoch_advance(arena, &time);
1444
1445 ndirty_limit = arena_decay_npages_limit(arena);
1446
1447 /*
1448 * Don't try to purge unless the number of purgeable pages exceeds the
1449 * current limit.
1450 */
1451 if (arena->ndirty <= ndirty_limit)
1452 return;
1453 arena_purge_to_limit(arena, ndirty_limit);
1454}
1455
1456void
1457arena_maybe_purge(arena_t *arena)
1458{
1459
1460 /* Don't recursively purge. */
1461 if (arena->purging)
1462 return;
1463
1464 if (opt_purge == purge_mode_ratio)
1465 arena_maybe_purge_ratio(arena);
1466 else
1467 arena_maybe_purge_decay(arena);
1468}
1469
1470static size_t
1471arena_dirty_count(arena_t *arena)
1472{
1473 size_t ndirty = 0;
1474 arena_runs_dirty_link_t *rdelm;
1475 extent_node_t *chunkselm;
1476
1477 for (rdelm = qr_next(&arena->runs_dirty, rd_link),
--- 19 unchanged lines hidden (view full) ---
1497 }
1498 ndirty += npages;
1499 }
1500
1501 return (ndirty);
1502}
1503
1504static size_t
1505arena_stash_dirty(arena_t *arena, chunk_hooks_t *chunk_hooks,
1506 size_t ndirty_limit, arena_runs_dirty_link_t *purge_runs_sentinel,
1507 extent_node_t *purge_chunks_sentinel)
1508{
1509 arena_runs_dirty_link_t *rdelm, *rdelm_next;
1510 extent_node_t *chunkselm;
1511 size_t nstashed = 0;
1512
1513 /* Stash runs/chunks according to ndirty_limit. */
1514 for (rdelm = qr_next(&arena->runs_dirty, rd_link),
1515 chunkselm = qr_next(&arena->chunks_cache, cc_link);
1516 rdelm != &arena->runs_dirty; rdelm = rdelm_next) {
1517 size_t npages;
1518 rdelm_next = qr_next(rdelm, rd_link);
1519
1520 if (rdelm == &chunkselm->rd) {
1521 extent_node_t *chunkselm_next;
1522 bool zero;
1523 UNUSED void *chunk;
1524
1525 npages = extent_node_size_get(chunkselm) >> LG_PAGE;
1526 if (opt_purge == purge_mode_decay && arena->ndirty -
1527 (nstashed + npages) < ndirty_limit)
1528 break;
1529
1530 chunkselm_next = qr_next(chunkselm, cc_link);
1531 /*
1532 * Allocate. chunkselm remains valid due to the
1533 * dalloc_node=false argument to chunk_alloc_cache().
1534 */
1535 zero = false;
1536 chunk = chunk_alloc_cache(arena, chunk_hooks,
1537 extent_node_addr_get(chunkselm),
1538 extent_node_size_get(chunkselm), chunksize, &zero,
1539 false);
1540 assert(chunk == extent_node_addr_get(chunkselm));
1541 assert(zero == extent_node_zeroed_get(chunkselm));
1542 extent_node_dirty_insert(chunkselm, purge_runs_sentinel,
1543 purge_chunks_sentinel);
1544 assert(npages == (extent_node_size_get(chunkselm) >>
1545 LG_PAGE));
1546 chunkselm = chunkselm_next;
1547 } else {
1548 arena_chunk_t *chunk =
1549 (arena_chunk_t *)CHUNK_ADDR2BASE(rdelm);
1550 arena_chunk_map_misc_t *miscelm =
1551 arena_rd_to_miscelm(rdelm);
1552 size_t pageind = arena_miscelm_to_pageind(miscelm);
1553 arena_run_t *run = &miscelm->run;
1554 size_t run_size =
1555 arena_mapbits_unallocated_size_get(chunk, pageind);
1556
1557 npages = run_size >> LG_PAGE;
1558 if (opt_purge == purge_mode_decay && arena->ndirty -
1559 (nstashed + npages) < ndirty_limit)
1560 break;
1561
1562 assert(pageind + npages <= chunk_npages);
1563 assert(arena_mapbits_dirty_get(chunk, pageind) ==
1564 arena_mapbits_dirty_get(chunk, pageind+npages-1));
1565
1566 /*
1567 * If purging the spare chunk's run, make it available
1568 * prior to allocation.
--- 9 unchanged lines hidden (view full) ---
1578 else {
1579 assert(qr_next(rdelm, rd_link) == rdelm);
1580 assert(qr_prev(rdelm, rd_link) == rdelm);
1581 }
1582 qr_meld(purge_runs_sentinel, rdelm, rd_link);
1583 }
1584
1585 nstashed += npages;
1586 if (opt_purge == purge_mode_ratio && arena->ndirty - nstashed <=
1587 ndirty_limit)
1588 break;
1589 }
1590
1591 return (nstashed);
1592}
1593
1594static size_t
1595arena_purge_stashed(arena_t *arena, chunk_hooks_t *chunk_hooks,
--- 123 unchanged lines hidden (view full) ---
1719 pageind) != 0);
1720 arena_run_t *run = &miscelm->run;
1721 qr_remove(rdelm, rd_link);
1722 arena_run_dalloc(arena, run, false, true, decommitted);
1723 }
1724 }
1725}
1726
1727/*
1728 * NB: ndirty_limit is interpreted differently depending on opt_purge:
1729 * - purge_mode_ratio: Purge as few dirty run/chunks as possible to reach the
1730 * desired state:
1731 * (arena->ndirty <= ndirty_limit)
1732 * - purge_mode_decay: Purge as many dirty runs/chunks as possible without
1733 * violating the invariant:
1734 * (arena->ndirty >= ndirty_limit)
1735 */
1736static void
1737arena_purge_to_limit(arena_t *arena, size_t ndirty_limit)
1738{
1739 chunk_hooks_t chunk_hooks = chunk_hooks_get(arena);
1740 size_t npurge, npurged;
1741 arena_runs_dirty_link_t purge_runs_sentinel;
1742 extent_node_t purge_chunks_sentinel;
1743
1744 arena->purging = true;
1745
1746 /*
1747 * Calls to arena_dirty_count() are disabled even for debug builds
1748 * because overhead grows nonlinearly as memory usage increases.
1749 */
1750 if (false && config_debug) {
1751 size_t ndirty = arena_dirty_count(arena);
1752 assert(ndirty == arena->ndirty);
1753 }
1754 assert(opt_purge != purge_mode_ratio || (arena->nactive >>
1755 arena->lg_dirty_mult) < arena->ndirty || ndirty_limit == 0);
1756
1757 qr_new(&purge_runs_sentinel, rd_link);
1758 extent_node_dirty_linkage_init(&purge_chunks_sentinel);
1759
1760 npurge = arena_stash_dirty(arena, &chunk_hooks, ndirty_limit,
1761 &purge_runs_sentinel, &purge_chunks_sentinel);
1762 if (npurge == 0)
1763 goto label_return;
1764 npurged = arena_purge_stashed(arena, &chunk_hooks, &purge_runs_sentinel,
1765 &purge_chunks_sentinel);
1766 assert(npurged == npurge);
1767 arena_unstash_purged(arena, &chunk_hooks, &purge_runs_sentinel,
1768 &purge_chunks_sentinel);
1769
1770 if (config_stats)
1771 arena->stats.npurge++;
1772
1773label_return:
1774 arena->purging = false;
1775}
1776
1777void
1778arena_purge(arena_t *arena, bool all)
1779{
1780
1781 malloc_mutex_lock(&arena->lock);
1782 if (all)
1783 arena_purge_to_limit(arena, 0);
1784 else
1785 arena_maybe_purge(arena);
1786 malloc_mutex_unlock(&arena->lock);
1787}
1788
1789static void
1790arena_run_coalesce(arena_t *arena, arena_chunk_t *chunk, size_t *p_size,
1791 size_t *p_run_ind, size_t *p_run_pages, size_t flag_dirty,
1792 size_t flag_decommitted)
1793{
--- 100 unchanged lines hidden (view full) ---
1894 } else {
1895 arena_bin_info_t *bin_info = &arena_bin_info[run->binind];
1896 size = bin_info->run_size;
1897 }
1898
1899 return (size);
1900}
1901
1902static void
1903arena_run_dalloc(arena_t *arena, arena_run_t *run, bool dirty, bool cleaned,
1904 bool decommitted)
1905{
1906 arena_chunk_t *chunk;
1907 arena_chunk_map_misc_t *miscelm;
1908 size_t size, run_ind, run_pages, flag_dirty, flag_decommitted;
1909
1910 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run);
1911 miscelm = arena_run_to_miscelm(run);
1912 run_ind = arena_miscelm_to_pageind(miscelm);
1913 assert(run_ind >= map_bias);
1914 assert(run_ind < chunk_npages);
1915 size = arena_run_size_get(arena, chunk, run, run_ind);
1916 run_pages = (size >> LG_PAGE);
1917 arena_nactive_sub(arena, run_pages);
1918
1919 /*
1920 * The run is dirty if the caller claims to have dirtied it, as well as
1921 * if it was already dirty before being allocated and the caller
1922 * doesn't claim to have cleaned it.
1923 */
1924 assert(arena_mapbits_dirty_get(chunk, run_ind) ==
1925 arena_mapbits_dirty_get(chunk, run_ind+run_pages-1));
--- 45 unchanged lines hidden (view full) ---
1971 * allows for an old spare to be fully deallocated, thus decreasing the
1972 * chances of spuriously crossing the dirty page purging threshold.
1973 */
1974 if (dirty)
1975 arena_maybe_purge(arena);
1976}
1977
1978static void
1979arena_run_trim_head(arena_t *arena, arena_chunk_t *chunk, arena_run_t *run,
1980 size_t oldsize, size_t newsize)
1981{
1982 arena_chunk_map_misc_t *miscelm = arena_run_to_miscelm(run);
1983 size_t pageind = arena_miscelm_to_pageind(miscelm);
1984 size_t head_npages = (oldsize - newsize) >> LG_PAGE;
1985 size_t flag_dirty = arena_mapbits_dirty_get(chunk, pageind);
1986 size_t flag_decommitted = arena_mapbits_decommitted_get(chunk, pageind);
--- 211 unchanged lines hidden (view full) ---
2198 bin->runcur = run;
2199
2200 assert(bin->runcur->nfree > 0);
2201
2202 return (arena_run_reg_alloc(bin->runcur, bin_info));
2203}
2204
2205void
2206arena_tcache_fill_small(tsd_t *tsd, arena_t *arena, tcache_bin_t *tbin,
2207 szind_t binind, uint64_t prof_accumbytes)
2208{
2209 unsigned i, nfill;
2210 arena_bin_t *bin;
2211
2212 assert(tbin->ncached == 0);
2213
2214 if (config_prof && arena_prof_accum(arena, prof_accumbytes))
2215 prof_idump();
--- 6 unchanged lines hidden (view full) ---
2222 if ((run = bin->runcur) != NULL && run->nfree > 0)
2223 ptr = arena_run_reg_alloc(run, &arena_bin_info[binind]);
2224 else
2225 ptr = arena_bin_malloc_hard(arena, bin);
2226 if (ptr == NULL) {
2227 /*
2228 * OOM. tbin->avail isn't yet filled down to its first
2229 * element, so the successful allocations (if any) must
2230 * be moved just before tbin->avail before bailing out.
2231 */
2232 if (i > 0) {
2233 memmove(tbin->avail - i, tbin->avail - nfill,
2234 i * sizeof(void *));
2235 }
2236 break;
2237 }
2238 if (config_fill && unlikely(opt_junk_alloc)) {
2239 arena_alloc_junk_small(ptr, &arena_bin_info[binind],
2240 true);
2241 }
2242 /* Insert such that low regions get used first. */
2243 *(tbin->avail - nfill + i) = ptr;
2244 }
2245 if (config_stats) {
2246 bin->stats.nmalloc += i;
2247 bin->stats.nrequests += tbin->tstats.nrequests;
2248 bin->stats.curregs += i;
2249 bin->stats.nfills++;
2250 tbin->tstats.nrequests = 0;
2251 }
2252 malloc_mutex_unlock(&bin->lock);
2253 tbin->ncached = i;
2254 arena_decay_tick(tsd, arena);
2255}
2256
2257void
2258arena_alloc_junk_small(void *ptr, arena_bin_info_t *bin_info, bool zero)
2259{
2260
2261 if (zero) {
2262 size_t redzone_size = bin_info->redzone_size;
--- 93 unchanged lines hidden (view full) ---
2356 assert(opt_quarantine);
2357 assert(usize <= SMALL_MAXCLASS);
2358
2359 binind = size2index(usize);
2360 bin_info = &arena_bin_info[binind];
2361 arena_redzones_validate(ptr, bin_info, true);
2362}
2363
2364static void *
2365arena_malloc_small(tsd_t *tsd, arena_t *arena, szind_t binind, bool zero)
2366{
2367 void *ret;
2368 arena_bin_t *bin;
2369 size_t usize;
2370 arena_run_t *run;
2371
2372 assert(binind < NBINS);
2373 bin = &arena->bins[binind];
2374 usize = index2size(binind);
2375
2376 malloc_mutex_lock(&bin->lock);
2377 if ((run = bin->runcur) != NULL && run->nfree > 0)
2378 ret = arena_run_reg_alloc(run, &arena_bin_info[binind]);
2379 else
2380 ret = arena_bin_malloc_hard(arena, bin);
2381
2382 if (ret == NULL) {
2383 malloc_mutex_unlock(&bin->lock);
2384 return (NULL);
2385 }
2386
2387 if (config_stats) {
2388 bin->stats.nmalloc++;
2389 bin->stats.nrequests++;
2390 bin->stats.curregs++;
2391 }
2392 malloc_mutex_unlock(&bin->lock);
2393 if (config_prof && !isthreaded && arena_prof_accum(arena, usize))
2394 prof_idump();
2395
2396 if (!zero) {
2397 if (config_fill) {
2398 if (unlikely(opt_junk_alloc)) {
2399 arena_alloc_junk_small(ret,
2400 &arena_bin_info[binind], false);
2401 } else if (unlikely(opt_zero))
2402 memset(ret, 0, usize);
2403 }
2404 JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED(ret, usize);
2405 } else {
2406 if (config_fill && unlikely(opt_junk_alloc)) {
2407 arena_alloc_junk_small(ret, &arena_bin_info[binind],
2408 true);
2409 }
2410 JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED(ret, usize);
2411 memset(ret, 0, usize);
2412 }
2413
2414 arena_decay_tick(tsd, arena);
2415 return (ret);
2416}
2417
2418void *
2419arena_malloc_large(tsd_t *tsd, arena_t *arena, szind_t binind, bool zero)
2420{
2421 void *ret;
2422 size_t usize;
2423 uintptr_t random_offset;
2424 arena_run_t *run;
2425 arena_chunk_map_misc_t *miscelm;
2426 UNUSED bool idump;
2427
2428 /* Large allocation. */
2429 usize = index2size(binind);
2430 malloc_mutex_lock(&arena->lock);
2431 if (config_cache_oblivious) {
2432 uint64_t r;
2433
2434 /*
2435 * Compute a uniformly distributed offset within the first page
2436 * that is a multiple of the cacheline size, e.g. [0 .. 63) * 64
2437 * for 4 KiB pages and 64-byte cachelines.
2438 */
2439 r = prng_lg_range(&arena->offset_state, LG_PAGE - LG_CACHELINE);
2440 random_offset = ((uintptr_t)r) << LG_CACHELINE;
2441 } else
2442 random_offset = 0;
2443 run = arena_run_alloc_large(arena, usize + large_pad, zero);
2444 if (run == NULL) {
2445 malloc_mutex_unlock(&arena->lock);
2446 return (NULL);
2447 }
2448 miscelm = arena_run_to_miscelm(run);
2449 ret = (void *)((uintptr_t)arena_miscelm_to_rpages(miscelm) +
2450 random_offset);
2451 if (config_stats) {
2452 szind_t index = binind - NBINS;
2453
2454 arena->stats.nmalloc_large++;
2455 arena->stats.nrequests_large++;
2456 arena->stats.allocated_large += usize;
2457 arena->stats.lstats[index].nmalloc++;
2458 arena->stats.lstats[index].nrequests++;
2459 arena->stats.lstats[index].curruns++;
2460 }
--- 7 unchanged lines hidden (view full) ---
2468 if (config_fill) {
2469 if (unlikely(opt_junk_alloc))
2470 memset(ret, 0xa5, usize);
2471 else if (unlikely(opt_zero))
2472 memset(ret, 0, usize);
2473 }
2474 }
2475
2476 arena_decay_tick(tsd, arena);
2477 return (ret);
2478}
2479
2480void *
2481arena_malloc_hard(tsd_t *tsd, arena_t *arena, size_t size, szind_t ind,
2482 bool zero, tcache_t *tcache)
2483{
2484
2485 arena = arena_choose(tsd, arena);
2486 if (unlikely(arena == NULL))
2487 return (NULL);
2488
2489 if (likely(size <= SMALL_MAXCLASS))
2490 return (arena_malloc_small(tsd, arena, ind, zero));
2491 if (likely(size <= large_maxclass))
2492 return (arena_malloc_large(tsd, arena, ind, zero));
2493 return (huge_malloc(tsd, arena, index2size(ind), zero, tcache));
2494}
2495
2496/* Only handles large allocations that require more than page alignment. */
2497static void *
2498arena_palloc_large(tsd_t *tsd, arena_t *arena, size_t usize, size_t alignment,
2499 bool zero)
2500{
2501 void *ret;
2502 size_t alloc_size, leadsize, trailsize;
2503 arena_run_t *run;
--- 67 unchanged lines hidden (view full) ---
2571 malloc_mutex_unlock(&arena->lock);
2572
2573 if (config_fill && !zero) {
2574 if (unlikely(opt_junk_alloc))
2575 memset(ret, 0xa5, usize);
2576 else if (unlikely(opt_zero))
2577 memset(ret, 0, usize);
2578 }
2579 arena_decay_tick(tsd, arena);
2580 return (ret);
2581}
2582
2583void *
2584arena_palloc(tsd_t *tsd, arena_t *arena, size_t usize, size_t alignment,
2585 bool zero, tcache_t *tcache)
2586{
2587 void *ret;
2588
2589 if (usize <= SMALL_MAXCLASS && (alignment < PAGE || (alignment == PAGE
2590 && (usize & PAGE_MASK) == 0))) {
2591 /* Small; alignment doesn't require special run placement. */
2592 ret = arena_malloc(tsd, arena, usize, size2index(usize), zero,
2593 tcache, true);
2594 } else if (usize <= large_maxclass && alignment <= PAGE) {
2595 /*
2596 * Large; alignment doesn't require special run placement.
2597 * However, the cached pointer may be at a random offset from
2598 * the base of the run, so do some bit manipulation to retrieve
2599 * the base.
2600 */
2601 ret = arena_malloc(tsd, arena, usize, size2index(usize), zero,
2602 tcache, true);
2603 if (config_cache_oblivious)
2604 ret = (void *)((uintptr_t)ret & ~PAGE_MASK);
2605 } else {
2606 if (likely(usize <= large_maxclass)) {
2607 ret = arena_palloc_large(tsd, arena, usize, alignment,
2608 zero);
2609 } else if (likely(alignment <= chunksize))
2610 ret = huge_malloc(tsd, arena, usize, zero, tcache);
--- 60 unchanged lines hidden (view full) ---
2671
2672 assert(run != bin->runcur);
2673 assert(arena_run_tree_search(&bin->runs, arena_run_to_miscelm(run)) ==
2674 NULL);
2675
2676 malloc_mutex_unlock(&bin->lock);
2677 /******************************/
2678 malloc_mutex_lock(&arena->lock);
2679 arena_run_dalloc(arena, run, true, false, false);
2680 malloc_mutex_unlock(&arena->lock);
2681 /****************************/
2682 malloc_mutex_lock(&bin->lock);
2683 if (config_stats)
2684 bin->stats.curruns--;
2685}
2686
2687static void
--- 70 unchanged lines hidden (view full) ---
2758 run = &arena_miscelm_get(chunk, rpages_ind)->run;
2759 bin = &arena->bins[run->binind];
2760 malloc_mutex_lock(&bin->lock);
2761 arena_dalloc_bin_locked_impl(arena, chunk, ptr, bitselm, false);
2762 malloc_mutex_unlock(&bin->lock);
2763}
2764
2765void
2766arena_dalloc_small(tsd_t *tsd, arena_t *arena, arena_chunk_t *chunk, void *ptr,
2767 size_t pageind)
2768{
2769 arena_chunk_map_bits_t *bitselm;
2770
2771 if (config_debug) {
2772 /* arena_ptr_small_binind_get() does extra sanity checking. */
2773 assert(arena_ptr_small_binind_get(ptr, arena_mapbits_get(chunk,
2774 pageind)) != BININD_INVALID);
2775 }
2776 bitselm = arena_bitselm_get(chunk, pageind);
2777 arena_dalloc_bin(arena, chunk, ptr, pageind, bitselm);
2778 arena_decay_tick(tsd, arena);
2779}
2780
2781#ifdef JEMALLOC_JET
2782#undef arena_dalloc_junk_large
2783#define arena_dalloc_junk_large JEMALLOC_N(arena_dalloc_junk_large_impl)
2784#endif
2785void
2786arena_dalloc_junk_large(void *ptr, size_t usize)
--- 28 unchanged lines hidden (view full) ---
2815
2816 arena->stats.ndalloc_large++;
2817 arena->stats.allocated_large -= usize;
2818 arena->stats.lstats[index].ndalloc++;
2819 arena->stats.lstats[index].curruns--;
2820 }
2821 }
2822
2823 arena_run_dalloc(arena, run, true, false, false);
2824}
2825
2826void
2827arena_dalloc_large_junked_locked(arena_t *arena, arena_chunk_t *chunk,
2828 void *ptr)
2829{
2830
2831 arena_dalloc_large_locked_impl(arena, chunk, ptr, true);
2832}
2833
2834void
2835arena_dalloc_large(tsd_t *tsd, arena_t *arena, arena_chunk_t *chunk, void *ptr)
2836{
2837
2838 malloc_mutex_lock(&arena->lock);
2839 arena_dalloc_large_locked_impl(arena, chunk, ptr, false);
2840 malloc_mutex_unlock(&arena->lock);
2841 arena_decay_tick(tsd, arena);
2842}
2843
2844static void
2845arena_ralloc_large_shrink(arena_t *arena, arena_chunk_t *chunk, void *ptr,
2846 size_t oldsize, size_t size)
2847{
2848 size_t pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
2849 arena_chunk_map_misc_t *miscelm = arena_miscelm_get(chunk, pageind);
--- 184 unchanged lines hidden (view full) ---
3034 assert(oldsize > usize_max);
3035 /* Fill before shrinking in order avoid a race. */
3036 arena_ralloc_junk_large(ptr, oldsize, usize_max);
3037 arena_ralloc_large_shrink(arena, chunk, ptr, oldsize, usize_max);
3038 return (false);
3039}
3040
3041bool
3042arena_ralloc_no_move(tsd_t *tsd, void *ptr, size_t oldsize, size_t size,
3043 size_t extra, bool zero)
3044{
3045 size_t usize_min, usize_max;
3046
3047 /* Calls with non-zero extra had to clamp extra. */
3048 assert(extra == 0 || size + extra <= HUGE_MAXCLASS);
3049
3050 if (unlikely(size > HUGE_MAXCLASS))
3051 return (true);
3052
3053 usize_min = s2u(size);
3054 usize_max = s2u(size + extra);
3055 if (likely(oldsize <= large_maxclass && usize_min <= large_maxclass)) {
3056 arena_chunk_t *chunk;
3057
3058 /*
3059 * Avoid moving the allocation if the size class can be left the
3060 * same.
3061 */
3062 if (oldsize <= SMALL_MAXCLASS) {
3063 assert(arena_bin_info[size2index(oldsize)].reg_size ==
3064 oldsize);
3065 if ((usize_max > SMALL_MAXCLASS ||
3066 size2index(usize_max) != size2index(oldsize)) &&
3067 (size > oldsize || usize_max < oldsize))
3068 return (true);
3069 } else {
3070 if (usize_max <= SMALL_MAXCLASS)
3071 return (true);
3072 if (arena_ralloc_large(ptr, oldsize, usize_min,
3073 usize_max, zero))
3074 return (true);
3075 }
3076
3077 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
3078 arena_decay_tick(tsd, extent_node_arena_get(&chunk->node));
3079 return (false);
3080 } else {
3081 return (huge_ralloc_no_move(tsd, ptr, oldsize, usize_min,
3082 usize_max, zero));
3083 }
3084}
3085
3086static void *
3087arena_ralloc_move_helper(tsd_t *tsd, arena_t *arena, size_t usize,
3088 size_t alignment, bool zero, tcache_t *tcache)
3089{
3090
3091 if (alignment == 0)
3092 return (arena_malloc(tsd, arena, usize, size2index(usize), zero,
3093 tcache, true));
3094 usize = sa2u(usize, alignment);
3095 if (unlikely(usize == 0 || usize > HUGE_MAXCLASS))
3096 return (NULL);
3097 return (ipalloct(tsd, usize, alignment, zero, tcache, arena));
3098}
3099
3100void *
3101arena_ralloc(tsd_t *tsd, arena_t *arena, void *ptr, size_t oldsize, size_t size,
3102 size_t alignment, bool zero, tcache_t *tcache)
3103{
3104 void *ret;
3105 size_t usize;
3106
3107 usize = s2u(size);
3108 if (unlikely(usize == 0 || size > HUGE_MAXCLASS))
3109 return (NULL);
3110
3111 if (likely(usize <= large_maxclass)) {
3112 size_t copysize;
3113
3114 /* Try to avoid moving the allocation. */
3115 if (!arena_ralloc_no_move(tsd, ptr, oldsize, usize, 0, zero))
3116 return (ptr);
3117
3118 /*
3119 * size and oldsize are different enough that we need to move
3120 * the object. In that case, fall back to allocating new space
3121 * and copying.
3122 */
3123 ret = arena_ralloc_move_helper(tsd, arena, usize, alignment,
--- 46 unchanged lines hidden (view full) ---
3170
3171 return ((ssize_t)atomic_read_z((size_t *)&lg_dirty_mult_default));
3172}
3173
3174bool
3175arena_lg_dirty_mult_default_set(ssize_t lg_dirty_mult)
3176{
3177
3178 if (opt_purge != purge_mode_ratio)
3179 return (true);
3180 if (!arena_lg_dirty_mult_valid(lg_dirty_mult))
3181 return (true);
3182 atomic_write_z((size_t *)&lg_dirty_mult_default, (size_t)lg_dirty_mult);
3183 return (false);
3184}
3185
3186ssize_t
3187arena_decay_time_default_get(void)
3188{
3189
3190 return ((ssize_t)atomic_read_z((size_t *)&decay_time_default));
3191}
3192
3193bool
3194arena_decay_time_default_set(ssize_t decay_time)
3195{
3196
3197 if (opt_purge != purge_mode_decay)
3198 return (true);
3199 if (!arena_decay_time_valid(decay_time))
3200 return (true);
3201 atomic_write_z((size_t *)&decay_time_default, (size_t)decay_time);
3202 return (false);
3203}
3204
3205static void
3206arena_basic_stats_merge_locked(arena_t *arena, unsigned *nthreads,
3207 const char **dss, ssize_t *lg_dirty_mult, ssize_t *decay_time,
3208 size_t *nactive, size_t *ndirty)
3209{
3210
3211 *nthreads += arena_nthreads_get(arena);
3212 *dss = dss_prec_names[arena->dss_prec];
3213 *lg_dirty_mult = arena->lg_dirty_mult;
3214 *decay_time = arena->decay_time;
3215 *nactive += arena->nactive;
3216 *ndirty += arena->ndirty;
3217}
3218
3219void
3220arena_basic_stats_merge(arena_t *arena, unsigned *nthreads, const char **dss,
3221 ssize_t *lg_dirty_mult, ssize_t *decay_time, size_t *nactive,
3222 size_t *ndirty)
3223{
3224
3225 malloc_mutex_lock(&arena->lock);
3226 arena_basic_stats_merge_locked(arena, nthreads, dss, lg_dirty_mult,
3227 decay_time, nactive, ndirty);
3228 malloc_mutex_unlock(&arena->lock);
3229}
3230
3231void
3232arena_stats_merge(arena_t *arena, unsigned *nthreads, const char **dss,
3233 ssize_t *lg_dirty_mult, ssize_t *decay_time, size_t *nactive,
3234 size_t *ndirty, arena_stats_t *astats, malloc_bin_stats_t *bstats,
3235 malloc_large_stats_t *lstats, malloc_huge_stats_t *hstats)
3236{
3237 unsigned i;
3238
3239 cassert(config_stats);
3240
3241 malloc_mutex_lock(&arena->lock);
3242 arena_basic_stats_merge_locked(arena, nthreads, dss, lg_dirty_mult,
3243 decay_time, nactive, ndirty);
3244
3245 astats->mapped += arena->stats.mapped;
3246 astats->npurge += arena->stats.npurge;
3247 astats->nmadvise += arena->stats.nmadvise;
3248 astats->purged += arena->stats.purged;
3249 astats->metadata_mapped += arena->stats.metadata_mapped;
3250 astats->metadata_allocated += arena_metadata_allocated_get(arena);
3251 astats->allocated_large += arena->stats.allocated_large;
3252 astats->nmalloc_large += arena->stats.nmalloc_large;
--- 31 unchanged lines hidden (view full) ---
3284 }
3285 bstats[i].nruns += bin->stats.nruns;
3286 bstats[i].reruns += bin->stats.reruns;
3287 bstats[i].curruns += bin->stats.curruns;
3288 malloc_mutex_unlock(&bin->lock);
3289 }
3290}
3291
3292unsigned
3293arena_nthreads_get(arena_t *arena)
3294{
3295
3296 return (atomic_read_u(&arena->nthreads));
3297}
3298
3299void
3300arena_nthreads_inc(arena_t *arena)
3301{
3302
3303 atomic_add_u(&arena->nthreads, 1);
3304}
3305
3306void
3307arena_nthreads_dec(arena_t *arena)
3308{
3309
3310 atomic_sub_u(&arena->nthreads, 1);
3311}
3312
3313arena_t *
3314arena_new(unsigned ind)
3315{
3316 arena_t *arena;
3317 size_t arena_size;
3318 unsigned i;
3319 arena_bin_t *bin;
3320
3321 /* Compute arena size to incorporate sufficient runs_avail elements. */
3322 arena_size = offsetof(arena_t, runs_avail) + (sizeof(arena_run_tree_t) *
3323 runs_avail_nclasses);
3324 /*
3325 * Allocate arena, arena->lstats, and arena->hstats contiguously, mainly
3326 * because there is no way to clean up if base_alloc() OOMs.
3327 */
3328 if (config_stats) {
3329 arena = (arena_t *)base_alloc(CACHELINE_CEILING(arena_size) +
3330 QUANTUM_CEILING(nlclasses * sizeof(malloc_large_stats_t) +
3331 nhclasses) * sizeof(malloc_huge_stats_t));
3332 } else
3333 arena = (arena_t *)base_alloc(arena_size);
3334 if (arena == NULL)
3335 return (NULL);
3336
3337 arena->ind = ind;
3338 arena->nthreads = 0;
3339 if (malloc_mutex_init(&arena->lock))
3340 return (NULL);
3341
3342 if (config_stats) {
3343 memset(&arena->stats, 0, sizeof(arena_stats_t));
3344 arena->stats.lstats = (malloc_large_stats_t *)((uintptr_t)arena
3345 + CACHELINE_CEILING(arena_size));
3346 memset(arena->stats.lstats, 0, nlclasses *
3347 sizeof(malloc_large_stats_t));
3348 arena->stats.hstats = (malloc_huge_stats_t *)((uintptr_t)arena
3349 + CACHELINE_CEILING(arena_size) +
3350 QUANTUM_CEILING(nlclasses * sizeof(malloc_large_stats_t)));
3351 memset(arena->stats.hstats, 0, nhclasses *
3352 sizeof(malloc_huge_stats_t));
3353 if (config_tcache)
3354 ql_new(&arena->tcache_ql);
3355 }
3356
3357 if (config_prof)
--- 15 unchanged lines hidden (view full) ---
3373
3374 arena->spare = NULL;
3375
3376 arena->lg_dirty_mult = arena_lg_dirty_mult_default_get();
3377 arena->purging = false;
3378 arena->nactive = 0;
3379 arena->ndirty = 0;
3380
3381 for(i = 0; i < runs_avail_nclasses; i++)
3382 arena_run_tree_new(&arena->runs_avail[i]);
3383 qr_new(&arena->runs_dirty, rd_link);
3384 qr_new(&arena->chunks_cache, cc_link);
3385
3386 if (opt_purge == purge_mode_decay)
3387 arena_decay_init(arena, arena_decay_time_default_get());
3388
3389 ql_new(&arena->huge);
3390 if (malloc_mutex_init(&arena->huge_mtx))
3391 return (NULL);
3392
3393 extent_tree_szad_new(&arena->chunks_szad_cached);
3394 extent_tree_ad_new(&arena->chunks_ad_cached);
3395 extent_tree_szad_new(&arena->chunks_szad_retained);
3396 extent_tree_ad_new(&arena->chunks_ad_retained);
--- 38 unchanged lines hidden (view full) ---
3435 /*
3436 * Determine redzone size based on minimum alignment and minimum
3437 * redzone size. Add padding to the end of the run if it is needed to
3438 * align the regions. The padding allows each redzone to be half the
3439 * minimum alignment; without the padding, each redzone would have to
3440 * be twice as large in order to maintain alignment.
3441 */
3442 if (config_fill && unlikely(opt_redzone)) {
3443 size_t align_min = ZU(1) << (ffs_zu(bin_info->reg_size) - 1);
3444 if (align_min <= REDZONE_MINSIZE) {
3445 bin_info->redzone_size = REDZONE_MINSIZE;
3446 pad_size = 0;
3447 } else {
3448 bin_info->redzone_size = align_min >> 1;
3449 pad_size = bin_info->redzone_size;
3450 }
3451 } else {
3452 bin_info->redzone_size = 0;
3453 pad_size = 0;
3454 }
3455 bin_info->reg_interval = bin_info->reg_size +
3456 (bin_info->redzone_size << 1);
3457
3458 /*
3459 * Compute run size under ideal conditions (no redzones, no limit on run
3460 * size).
3461 */
3462 try_run_size = PAGE;
3463 try_nregs = (uint32_t)(try_run_size / bin_info->reg_size);
3464 do {
3465 perfect_run_size = try_run_size;
3466 perfect_nregs = try_nregs;
3467
3468 try_run_size += PAGE;
3469 try_nregs = (uint32_t)(try_run_size / bin_info->reg_size);
3470 } while (perfect_run_size != perfect_nregs * bin_info->reg_size);
3471 assert(perfect_nregs <= RUN_MAXREGS);
3472
3473 actual_run_size = perfect_run_size;
3474 actual_nregs = (uint32_t)((actual_run_size - pad_size) /
3475 bin_info->reg_interval);
3476
3477 /*
3478 * Redzones can require enough padding that not even a single region can
3479 * fit within the number of pages that would normally be dedicated to a
3480 * run for this size class. Increase the run size until at least one
3481 * region fits.
3482 */
3483 while (actual_nregs == 0) {
3484 assert(config_fill && unlikely(opt_redzone));
3485
3486 actual_run_size += PAGE;
3487 actual_nregs = (uint32_t)((actual_run_size - pad_size) /
3488 bin_info->reg_interval);
3489 }
3490
3491 /*
3492 * Make sure that the run will fit within an arena chunk.
3493 */
3494 while (actual_run_size > arena_maxrun) {
3495 actual_run_size -= PAGE;
3496 actual_nregs = (uint32_t)((actual_run_size - pad_size) /
3497 bin_info->reg_interval);
3498 }
3499 assert(actual_nregs > 0);
3500 assert(actual_run_size == s2u(actual_run_size));
3501
3502 /* Copy final settings. */
3503 bin_info->run_size = actual_run_size;
3504 bin_info->nregs = actual_nregs;
3505 bin_info->reg0_offset = (uint32_t)(actual_run_size - (actual_nregs *
3506 bin_info->reg_interval) - pad_size + bin_info->redzone_size);
3507
3508 if (actual_run_size > small_maxrun)
3509 small_maxrun = actual_run_size;
3510
3511 assert(bin_info->reg0_offset - bin_info->redzone_size + (bin_info->nregs
3512 * bin_info->reg_interval) + pad_size == bin_info->run_size);
3513}
3514
--- 37 unchanged lines hidden (view full) ---
3552 SIZE_CLASSES
3553#undef TAB_INIT_bin_yes
3554#undef TAB_INIT_bin_no
3555#undef SC
3556
3557 return (false);
3558}
3559
3560static bool
3561run_quantize_init(void)
3562{
3563 unsigned i;
3564
3565 run_quantize_max = chunksize + large_pad;
3566
3567 run_quantize_floor_tab = (size_t *)base_alloc(sizeof(size_t) *
3568 (run_quantize_max >> LG_PAGE));
3569 if (run_quantize_floor_tab == NULL)
3570 return (true);
3571
3572 run_quantize_ceil_tab = (size_t *)base_alloc(sizeof(size_t) *
3573 (run_quantize_max >> LG_PAGE));
3574 if (run_quantize_ceil_tab == NULL)
3575 return (true);
3576
3577 for (i = 1; i <= run_quantize_max >> LG_PAGE; i++) {
3578 size_t run_size = i << LG_PAGE;
3579
3580 run_quantize_floor_tab[i-1] =
3581 run_quantize_floor_compute(run_size);
3582 run_quantize_ceil_tab[i-1] =
3583 run_quantize_ceil_compute(run_size);
3584 }
3585
3586 return (false);
3587}
3588
3589bool
3590arena_boot(void)
3591{
3592 unsigned i;
3593
3594 arena_lg_dirty_mult_default_set(opt_lg_dirty_mult);
3595 arena_decay_time_default_set(opt_decay_time);
3596
3597 /*
3598 * Compute the header size such that it is large enough to contain the
3599 * page map. The page map is biased to omit entries for the header
3600 * itself, so some iteration is necessary to compute the map bias.
3601 *
3602 * 1) Compute safe header_size and map_bias values that include enough
3603 * space for an unbiased page map.
--- 25 unchanged lines hidden (view full) ---
3629 */
3630 large_maxclass = arena_maxrun;
3631 }
3632 assert(large_maxclass > 0);
3633 nlclasses = size2index(large_maxclass) - size2index(SMALL_MAXCLASS);
3634 nhclasses = NSIZES - nlclasses - NBINS;
3635
3636 bin_info_init();
3637 if (small_run_size_init())
3638 return (true);
3639 if (run_quantize_init())
3640 return (true);
3641
3642 runs_avail_bias = size2index(PAGE);
3643 runs_avail_nclasses = size2index(run_quantize_max)+1 - runs_avail_bias;
3644
3645 return (false);
3646}
3647
3648void
3649arena_prefork(arena_t *arena)
3650{
3651 unsigned i;
3652
3653 malloc_mutex_prefork(&arena->lock);
--- 32 unchanged lines hidden ---
2#include "jemalloc/internal/jemalloc_internal.h"
3
4/******************************************************************************/
5/* Data. */
6
7purge_mode_t opt_purge = PURGE_DEFAULT;
8const char *purge_mode_names[] = {
9 "ratio",
10 "decay",
11 "N/A"
12};
13ssize_t opt_lg_dirty_mult = LG_DIRTY_MULT_DEFAULT;
14static ssize_t lg_dirty_mult_default;
15ssize_t opt_decay_time = DECAY_TIME_DEFAULT;
16static ssize_t decay_time_default;
17
18arena_bin_info_t arena_bin_info[NBINS];
19
20size_t map_bias;
21size_t map_misc_offset;
22size_t arena_maxrun; /* Max run size for arenas. */
23size_t large_maxclass; /* Max large size class. */
24size_t run_quantize_max; /* Max run_quantize_*() input. */
25static size_t small_maxrun; /* Max run size for small size classes. */
26static bool *small_run_tab; /* Valid small run page multiples. */
27static size_t *run_quantize_floor_tab; /* run_quantize_floor() memoization. */
28static size_t *run_quantize_ceil_tab; /* run_quantize_ceil() memoization. */
29unsigned nlclasses; /* Number of large size classes. */
30unsigned nhclasses; /* Number of huge size classes. */
31static szind_t runs_avail_bias; /* Size index for first runs_avail tree. */
32static szind_t runs_avail_nclasses; /* Number of runs_avail trees. */
33
34/******************************************************************************/
35/*
36 * Function prototypes for static functions that are referenced prior to
37 * definition.
38 */
39
40static void arena_purge_to_limit(arena_t *arena, size_t ndirty_limit);
41static void arena_run_dalloc(arena_t *arena, arena_run_t *run, bool dirty,
42 bool cleaned, bool decommitted);
43static void arena_dalloc_bin_run(arena_t *arena, arena_chunk_t *chunk,
44 arena_run_t *run, arena_bin_t *bin);
45static void arena_bin_lower_run(arena_t *arena, arena_chunk_t *chunk,
46 arena_run_t *run, arena_bin_t *bin);
47
48/******************************************************************************/
49
50JEMALLOC_INLINE_C size_t
51arena_miscelm_size_get(const arena_chunk_map_misc_t *miscelm)
52{
53 arena_chunk_t *chunk;
54 size_t pageind, mapbits;
55
56 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(miscelm);
57 pageind = arena_miscelm_to_pageind(miscelm);
58 mapbits = arena_mapbits_get(chunk, pageind);
59 return (arena_mapbits_size_decode(mapbits));
60}
61
62JEMALLOC_INLINE_C int
63arena_run_addr_comp(const arena_chunk_map_misc_t *a,
64 const arena_chunk_map_misc_t *b)
65{
66 uintptr_t a_miscelm = (uintptr_t)a;
67 uintptr_t b_miscelm = (uintptr_t)b;
68
69 assert(a != NULL);
70 assert(b != NULL);
71
72 return ((a_miscelm > b_miscelm) - (a_miscelm < b_miscelm));
73}
74
75/* Generate red-black tree functions. */
76rb_gen(static UNUSED, arena_run_tree_, arena_run_tree_t, arena_chunk_map_misc_t,
77 rb_link, arena_run_addr_comp)
78
79static size_t
80run_quantize_floor_compute(size_t size)
81{
82 size_t qsize;
83
84 assert(size != 0);
85 assert(size == PAGE_CEILING(size));
86
87 /* Don't change sizes that are valid small run sizes. */
88 if (size <= small_maxrun && small_run_tab[size >> LG_PAGE])
89 return (size);
90
91 /*
92 * Round down to the nearest run size that can actually be requested
93 * during normal large allocation. Add large_pad so that cache index
94 * randomization can offset the allocation from the page boundary.
95 */
96 qsize = index2size(size2index(size - large_pad + 1) - 1) + large_pad;
97 if (qsize <= SMALL_MAXCLASS + large_pad)
98 return (run_quantize_floor_compute(size - large_pad));
99 assert(qsize <= size);
100 return (qsize);
101}
102
103static size_t
104run_quantize_ceil_compute_hard(size_t size)
105{
106 size_t large_run_size_next;
107
108 assert(size != 0);
109 assert(size == PAGE_CEILING(size));
110
111 /*
112 * Return the next quantized size greater than the input size.
--- 17 unchanged lines hidden (view full) ---
130 if (large_run_size_next < size)
131 return (large_run_size_next);
132 return (size);
133 }
134 }
135}
136
137static size_t
138run_quantize_ceil_compute(size_t size)
139{
140 size_t qsize = run_quantize_floor_compute(size);
141
142 if (qsize < size) {
143 /*
144 * Skip a quantization that may have an adequately large run,
145 * because under-sized runs may be mixed in. This only happens
146 * when an unusual size is requested, i.e. for aligned
147 * allocation, and is just one of several places where linear
148 * search would potentially find sufficiently aligned available
149 * memory somewhere lower.
150 */
151 qsize = run_quantize_ceil_compute_hard(qsize);
152 }
153 return (qsize);
154}
155
156#ifdef JEMALLOC_JET
157#undef run_quantize_floor
158#define run_quantize_floor JEMALLOC_N(run_quantize_floor_impl)
159#endif
160static size_t
161run_quantize_floor(size_t size)
162{
163 size_t ret;
164
165 assert(size > 0);
166 assert(size <= run_quantize_max);
167 assert((size & PAGE_MASK) == 0);
168
169 ret = run_quantize_floor_tab[(size >> LG_PAGE) - 1];
170 assert(ret == run_quantize_floor_compute(size));
171 return (ret);
172}
173#ifdef JEMALLOC_JET
174#undef run_quantize_floor
175#define run_quantize_floor JEMALLOC_N(run_quantize_floor)
176run_quantize_t *run_quantize_floor = JEMALLOC_N(run_quantize_floor_impl);
177#endif
178
179#ifdef JEMALLOC_JET
180#undef run_quantize_ceil
181#define run_quantize_ceil JEMALLOC_N(run_quantize_ceil_impl)
182#endif
183static size_t
184run_quantize_ceil(size_t size)
185{
186 size_t ret;
187
188 assert(size > 0);
189 assert(size <= run_quantize_max);
190 assert((size & PAGE_MASK) == 0);
191
192 ret = run_quantize_ceil_tab[(size >> LG_PAGE) - 1];
193 assert(ret == run_quantize_ceil_compute(size));
194 return (ret);
195}
196#ifdef JEMALLOC_JET
197#undef run_quantize_ceil
198#define run_quantize_ceil JEMALLOC_N(run_quantize_ceil)
199run_quantize_t *run_quantize_ceil = JEMALLOC_N(run_quantize_ceil_impl);
200#endif
201
202static arena_run_tree_t *
203arena_runs_avail_get(arena_t *arena, szind_t ind)
204{
205
206 assert(ind >= runs_avail_bias);
207 assert(ind - runs_avail_bias < runs_avail_nclasses);
208
209 return (&arena->runs_avail[ind - runs_avail_bias]);
210}
211
212static void
213arena_avail_insert(arena_t *arena, arena_chunk_t *chunk, size_t pageind,
214 size_t npages)
215{
216 szind_t ind = size2index(run_quantize_floor(arena_miscelm_size_get(
217 arena_miscelm_get(chunk, pageind))));
218 assert(npages == (arena_mapbits_unallocated_size_get(chunk, pageind) >>
219 LG_PAGE));
220 arena_run_tree_insert(arena_runs_avail_get(arena, ind),
221 arena_miscelm_get(chunk, pageind));
222}
223
224static void
225arena_avail_remove(arena_t *arena, arena_chunk_t *chunk, size_t pageind,
226 size_t npages)
227{
228 szind_t ind = size2index(run_quantize_floor(arena_miscelm_size_get(
229 arena_miscelm_get(chunk, pageind))));
230 assert(npages == (arena_mapbits_unallocated_size_get(chunk, pageind) >>
231 LG_PAGE));
232 arena_run_tree_remove(arena_runs_avail_get(arena, ind),
233 arena_miscelm_get(chunk, pageind));
234}
235
236static void
237arena_run_dirty_insert(arena_t *arena, arena_chunk_t *chunk, size_t pageind,
238 size_t npages)
239{
240 arena_chunk_map_misc_t *miscelm = arena_miscelm_get(chunk, pageind);
241
--- 54 unchanged lines hidden (view full) ---
296 arena->ndirty -= arena_chunk_dirty_npages(node);
297 }
298}
299
300JEMALLOC_INLINE_C void *
301arena_run_reg_alloc(arena_run_t *run, arena_bin_info_t *bin_info)
302{
303 void *ret;
304 size_t regind;
305 arena_chunk_map_misc_t *miscelm;
306 void *rpages;
307
308 assert(run->nfree > 0);
309 assert(!bitmap_full(run->bitmap, &bin_info->bitmap_info));
310
311 regind = (unsigned)bitmap_sfu(run->bitmap, &bin_info->bitmap_info);
312 miscelm = arena_run_to_miscelm(run);
313 rpages = arena_miscelm_to_rpages(miscelm);
314 ret = (void *)((uintptr_t)rpages + (uintptr_t)bin_info->reg0_offset +
315 (uintptr_t)(bin_info->reg_interval * regind));
316 run->nfree--;
317 return (ret);
318}
319
320JEMALLOC_INLINE_C void
321arena_run_reg_dalloc(arena_run_t *run, void *ptr)
322{
323 arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run);
324 size_t pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
325 size_t mapbits = arena_mapbits_get(chunk, pageind);
326 szind_t binind = arena_ptr_small_binind_get(ptr, mapbits);
327 arena_bin_info_t *bin_info = &arena_bin_info[binind];
328 size_t regind = arena_run_regind(run, bin_info, ptr);
329
330 assert(run->nfree < bin_info->nregs);
331 /* Freeing an interior pointer can cause assertion failure. */
332 assert(((uintptr_t)ptr -
333 ((uintptr_t)arena_miscelm_to_rpages(arena_run_to_miscelm(run)) +
334 (uintptr_t)bin_info->reg0_offset)) %
335 (uintptr_t)bin_info->reg_interval == 0);
336 assert((uintptr_t)ptr >=
--- 31 unchanged lines hidden (view full) ---
368 UNUSED size_t *p = (size_t *)((uintptr_t)chunk + (run_ind << LG_PAGE));
369
370 arena_run_page_mark_zeroed(chunk, run_ind);
371 for (i = 0; i < PAGE / sizeof(size_t); i++)
372 assert(p[i] == 0);
373}
374
375static void
376arena_nactive_add(arena_t *arena, size_t add_pages)
377{
378
379 if (config_stats) {
380 size_t cactive_add = CHUNK_CEILING((arena->nactive +
381 add_pages) << LG_PAGE) - CHUNK_CEILING(arena->nactive <<
382 LG_PAGE);
383 if (cactive_add != 0)
384 stats_cactive_add(cactive_add);
385 }
386 arena->nactive += add_pages;
387}
388
389static void
390arena_nactive_sub(arena_t *arena, size_t sub_pages)
391{
392
393 if (config_stats) {
394 size_t cactive_sub = CHUNK_CEILING(arena->nactive << LG_PAGE) -
395 CHUNK_CEILING((arena->nactive - sub_pages) << LG_PAGE);
396 if (cactive_sub != 0)
397 stats_cactive_sub(cactive_sub);
398 }
399 arena->nactive -= sub_pages;
400}
401
402static void
403arena_run_split_remove(arena_t *arena, arena_chunk_t *chunk, size_t run_ind,
404 size_t flag_dirty, size_t flag_decommitted, size_t need_pages)
405{
406 size_t total_pages, rem_pages;
407
408 assert(flag_dirty == 0 || flag_decommitted == 0);
409
410 total_pages = arena_mapbits_unallocated_size_get(chunk, run_ind) >>
411 LG_PAGE;
412 assert(arena_mapbits_dirty_get(chunk, run_ind+total_pages-1) ==
413 flag_dirty);
414 assert(need_pages <= total_pages);
415 rem_pages = total_pages - need_pages;
416
417 arena_avail_remove(arena, chunk, run_ind, total_pages);
418 if (flag_dirty != 0)
419 arena_run_dirty_remove(arena, chunk, run_ind, total_pages);
420 arena_nactive_add(arena, need_pages);
421
422 /* Keep track of trailing unused pages for later use. */
423 if (rem_pages > 0) {
424 size_t flags = flag_dirty | flag_decommitted;
425 size_t flag_unzeroed_mask = (flags == 0) ? CHUNK_MAP_UNZEROED :
426 0;
427
428 arena_mapbits_unallocated_set(chunk, run_ind+need_pages,
--- 299 unchanged lines hidden (view full) ---
728 if (arena->spare != NULL)
729 chunk = arena_chunk_init_spare(arena);
730 else {
731 chunk = arena_chunk_init_hard(arena);
732 if (chunk == NULL)
733 return (NULL);
734 }
735
736 arena_avail_insert(arena, chunk, map_bias, chunk_npages-map_bias);
737
738 return (chunk);
739}
740
741static void
742arena_chunk_dalloc(arena_t *arena, arena_chunk_t *chunk)
743{
--- 4 unchanged lines hidden (view full) ---
748 arena_maxrun);
749 assert(arena_mapbits_unallocated_size_get(chunk, chunk_npages-1) ==
750 arena_maxrun);
751 assert(arena_mapbits_dirty_get(chunk, map_bias) ==
752 arena_mapbits_dirty_get(chunk, chunk_npages-1));
753 assert(arena_mapbits_decommitted_get(chunk, map_bias) ==
754 arena_mapbits_decommitted_get(chunk, chunk_npages-1));
755
756 /* Remove run from runs_avail, so that the arena does not use it. */
757 arena_avail_remove(arena, chunk, map_bias, chunk_npages-map_bias);
758
759 if (arena->spare != NULL) {
760 arena_chunk_t *spare = arena->spare;
761 chunk_hooks_t chunk_hooks = CHUNK_HOOKS_INITIALIZER;
762 bool committed;
763
764 arena->spare = chunk;
--- 136 unchanged lines hidden (view full) ---
901 zero, &commit);
902 if (ret == NULL) {
903 /* Revert optimistic stats updates. */
904 malloc_mutex_lock(&arena->lock);
905 if (config_stats) {
906 arena_huge_malloc_stats_update_undo(arena, usize);
907 arena->stats.mapped -= usize;
908 }
909 arena_nactive_sub(arena, usize >> LG_PAGE);
910 malloc_mutex_unlock(&arena->lock);
911 }
912
913 return (ret);
914}
915
916void *
917arena_chunk_alloc_huge(arena_t *arena, size_t usize, size_t alignment,
--- 5 unchanged lines hidden (view full) ---
923
924 malloc_mutex_lock(&arena->lock);
925
926 /* Optimistically update stats. */
927 if (config_stats) {
928 arena_huge_malloc_stats_update(arena, usize);
929 arena->stats.mapped += usize;
930 }
931 arena_nactive_add(arena, usize >> LG_PAGE);
932
933 ret = chunk_alloc_cache(arena, &chunk_hooks, NULL, csize, alignment,
934 zero, true);
935 malloc_mutex_unlock(&arena->lock);
936 if (ret == NULL) {
937 ret = arena_chunk_alloc_huge_hard(arena, &chunk_hooks, usize,
938 alignment, zero, csize);
939 }
940
941 return (ret);
942}
943
944void
945arena_chunk_dalloc_huge(arena_t *arena, void *chunk, size_t usize)
946{
947 chunk_hooks_t chunk_hooks = CHUNK_HOOKS_INITIALIZER;
948 size_t csize;
949
950 csize = CHUNK_CEILING(usize);
951 malloc_mutex_lock(&arena->lock);
952 if (config_stats) {
953 arena_huge_dalloc_stats_update(arena, usize);
954 arena->stats.mapped -= usize;
955 }
956 arena_nactive_sub(arena, usize >> LG_PAGE);
957
958 chunk_dalloc_cache(arena, &chunk_hooks, chunk, csize, true);
959 malloc_mutex_unlock(&arena->lock);
960}
961
962void
963arena_chunk_ralloc_huge_similar(arena_t *arena, void *chunk, size_t oldsize,
964 size_t usize)
965{
966
967 assert(CHUNK_CEILING(oldsize) == CHUNK_CEILING(usize));
968 assert(oldsize != usize);
969
970 malloc_mutex_lock(&arena->lock);
971 if (config_stats)
972 arena_huge_ralloc_stats_update(arena, oldsize, usize);
973 if (oldsize < usize)
974 arena_nactive_add(arena, (usize - oldsize) >> LG_PAGE);
975 else
976 arena_nactive_sub(arena, (oldsize - usize) >> LG_PAGE);
977 malloc_mutex_unlock(&arena->lock);
978}
979
980void
981arena_chunk_ralloc_huge_shrink(arena_t *arena, void *chunk, size_t oldsize,
982 size_t usize)
983{
984 size_t udiff = oldsize - usize;
985 size_t cdiff = CHUNK_CEILING(oldsize) - CHUNK_CEILING(usize);
986
987 malloc_mutex_lock(&arena->lock);
988 if (config_stats) {
989 arena_huge_ralloc_stats_update(arena, oldsize, usize);
990 if (cdiff != 0)
991 arena->stats.mapped -= cdiff;
992 }
993 arena_nactive_sub(arena, udiff >> LG_PAGE);
994
995 if (cdiff != 0) {
996 chunk_hooks_t chunk_hooks = CHUNK_HOOKS_INITIALIZER;
997 void *nchunk = (void *)((uintptr_t)chunk +
998 CHUNK_CEILING(usize));
999
1000 chunk_dalloc_cache(arena, &chunk_hooks, nchunk, cdiff, true);
1001 }
--- 13 unchanged lines hidden (view full) ---
1015 if (err) {
1016 /* Revert optimistic stats updates. */
1017 malloc_mutex_lock(&arena->lock);
1018 if (config_stats) {
1019 arena_huge_ralloc_stats_update_undo(arena, oldsize,
1020 usize);
1021 arena->stats.mapped -= cdiff;
1022 }
1023 arena_nactive_sub(arena, udiff >> LG_PAGE);
1024 malloc_mutex_unlock(&arena->lock);
1025 } else if (chunk_hooks->merge(chunk, CHUNK_CEILING(oldsize), nchunk,
1026 cdiff, true, arena->ind)) {
1027 chunk_dalloc_arena(arena, chunk_hooks, nchunk, cdiff, *zero,
1028 true);
1029 err = true;
1030 }
1031 return (err);
--- 11 unchanged lines hidden (view full) ---
1043
1044 malloc_mutex_lock(&arena->lock);
1045
1046 /* Optimistically update stats. */
1047 if (config_stats) {
1048 arena_huge_ralloc_stats_update(arena, oldsize, usize);
1049 arena->stats.mapped += cdiff;
1050 }
1051 arena_nactive_add(arena, udiff >> LG_PAGE);
1052
1053 err = (chunk_alloc_cache(arena, &arena->chunk_hooks, nchunk, cdiff,
1054 chunksize, zero, true) == NULL);
1055 malloc_mutex_unlock(&arena->lock);
1056 if (err) {
1057 err = arena_chunk_ralloc_huge_expand_hard(arena, &chunk_hooks,
1058 chunk, oldsize, usize, zero, nchunk, udiff,
1059 cdiff);
1060 } else if (chunk_hooks.merge(chunk, CHUNK_CEILING(oldsize), nchunk,
1061 cdiff, true, arena->ind)) {
1062 chunk_dalloc_arena(arena, &chunk_hooks, nchunk, cdiff, *zero,
1063 true);
1064 err = true;
1065 }
1066
1067 return (err);
1068}
1069
1070/*
1071 * Do first-best-fit run selection, i.e. select the lowest run that best fits.
1072 * Run sizes are indexed, so not all candidate runs are necessarily exactly the
1073 * same size.
1074 */
1075static arena_run_t *
1076arena_run_first_best_fit(arena_t *arena, size_t size)
1077{
1078 szind_t ind, i;
1079
1080 ind = size2index(run_quantize_ceil(size));
1081 for (i = ind; i < runs_avail_nclasses + runs_avail_bias; i++) {
1082 arena_chunk_map_misc_t *miscelm = arena_run_tree_first(
1083 arena_runs_avail_get(arena, i));
1084 if (miscelm != NULL)
1085 return (&miscelm->run);
1086 }
1087
1088 return (NULL);
1089}
1090
1091static arena_run_t *
1092arena_run_alloc_large_helper(arena_t *arena, size_t size, bool zero)
1093{
1094 arena_run_t *run = arena_run_first_best_fit(arena, s2u(size));
1095 if (run != NULL) {
1096 if (arena_run_split_large(arena, run, size, zero))
--- 110 unchanged lines hidden (view full) ---
1207 malloc_mutex_lock(&arena->lock);
1208 arena->lg_dirty_mult = lg_dirty_mult;
1209 arena_maybe_purge(arena);
1210 malloc_mutex_unlock(&arena->lock);
1211
1212 return (false);
1213}
1214
1215static void
1216arena_decay_deadline_init(arena_t *arena)
1217{
1218
1219 assert(opt_purge == purge_mode_decay);
1220
1221 /*
1222 * Generate a new deadline that is uniformly random within the next
1223 * epoch after the current one.
1224 */
1225 nstime_copy(&arena->decay_deadline, &arena->decay_epoch);
1226 nstime_add(&arena->decay_deadline, &arena->decay_interval);
1227 if (arena->decay_time > 0) {
1228 nstime_t jitter;
1229
1230 nstime_init(&jitter, prng_range(&arena->decay_jitter_state,
1231 nstime_ns(&arena->decay_interval)));
1232 nstime_add(&arena->decay_deadline, &jitter);
1233 }
1234}
1235
1236static bool
1237arena_decay_deadline_reached(const arena_t *arena, const nstime_t *time)
1238{
1239
1240 assert(opt_purge == purge_mode_decay);
1241
1242 return (nstime_compare(&arena->decay_deadline, time) <= 0);
1243}
1244
1245static size_t
1246arena_decay_backlog_npages_limit(const arena_t *arena)
1247{
1248 static const uint64_t h_steps[] = {
1249#define STEP(step, h, x, y) \
1250 h,
1251 SMOOTHSTEP
1252#undef STEP
1253 };
1254 uint64_t sum;
1255 size_t npages_limit_backlog;
1256 unsigned i;
1257
1258 assert(opt_purge == purge_mode_decay);
1259
1260 /*
1261 * For each element of decay_backlog, multiply by the corresponding
1262 * fixed-point smoothstep decay factor. Sum the products, then divide
1263 * to round down to the nearest whole number of pages.
1264 */
1265 sum = 0;
1266 for (i = 0; i < SMOOTHSTEP_NSTEPS; i++)
1267 sum += arena->decay_backlog[i] * h_steps[i];
1268 npages_limit_backlog = (sum >> SMOOTHSTEP_BFP);
1269
1270 return (npages_limit_backlog);
1271}
1272
1273static void
1274arena_decay_epoch_advance(arena_t *arena, const nstime_t *time)
1275{
1276 uint64_t nadvance;
1277 nstime_t delta;
1278 size_t ndirty_delta;
1279
1280 assert(opt_purge == purge_mode_decay);
1281 assert(arena_decay_deadline_reached(arena, time));
1282
1283 nstime_copy(&delta, time);
1284 nstime_subtract(&delta, &arena->decay_epoch);
1285 nadvance = nstime_divide(&delta, &arena->decay_interval);
1286 assert(nadvance > 0);
1287
1288 /* Add nadvance decay intervals to epoch. */
1289 nstime_copy(&delta, &arena->decay_interval);
1290 nstime_imultiply(&delta, nadvance);
1291 nstime_add(&arena->decay_epoch, &delta);
1292
1293 /* Set a new deadline. */
1294 arena_decay_deadline_init(arena);
1295
1296 /* Update the backlog. */
1297 if (nadvance >= SMOOTHSTEP_NSTEPS) {
1298 memset(arena->decay_backlog, 0, (SMOOTHSTEP_NSTEPS-1) *
1299 sizeof(size_t));
1300 } else {
1301 memmove(arena->decay_backlog, &arena->decay_backlog[nadvance],
1302 (SMOOTHSTEP_NSTEPS - nadvance) * sizeof(size_t));
1303 if (nadvance > 1) {
1304 memset(&arena->decay_backlog[SMOOTHSTEP_NSTEPS -
1305 nadvance], 0, (nadvance-1) * sizeof(size_t));
1306 }
1307 }
1308 ndirty_delta = (arena->ndirty > arena->decay_ndirty) ? arena->ndirty -
1309 arena->decay_ndirty : 0;
1310 arena->decay_ndirty = arena->ndirty;
1311 arena->decay_backlog[SMOOTHSTEP_NSTEPS-1] = ndirty_delta;
1312 arena->decay_backlog_npages_limit =
1313 arena_decay_backlog_npages_limit(arena);
1314}
1315
1316static size_t
1317arena_decay_npages_limit(arena_t *arena)
1318{
1319 size_t npages_limit;
1320
1321 assert(opt_purge == purge_mode_decay);
1322
1323 npages_limit = arena->decay_backlog_npages_limit;
1324
1325 /* Add in any dirty pages created during the current epoch. */
1326 if (arena->ndirty > arena->decay_ndirty)
1327 npages_limit += arena->ndirty - arena->decay_ndirty;
1328
1329 return (npages_limit);
1330}
1331
1332static void
1333arena_decay_init(arena_t *arena, ssize_t decay_time)
1334{
1335
1336 arena->decay_time = decay_time;
1337 if (decay_time > 0) {
1338 nstime_init2(&arena->decay_interval, decay_time, 0);
1339 nstime_idivide(&arena->decay_interval, SMOOTHSTEP_NSTEPS);
1340 }
1341
1342 nstime_init(&arena->decay_epoch, 0);
1343 nstime_update(&arena->decay_epoch);
1344 arena->decay_jitter_state = (uint64_t)(uintptr_t)arena;
1345 arena_decay_deadline_init(arena);
1346 arena->decay_ndirty = arena->ndirty;
1347 arena->decay_backlog_npages_limit = 0;
1348 memset(arena->decay_backlog, 0, SMOOTHSTEP_NSTEPS * sizeof(size_t));
1349}
1350
1351static bool
1352arena_decay_time_valid(ssize_t decay_time)
1353{
1354
1355 return (decay_time >= -1 && decay_time <= NSTIME_SEC_MAX);
1356}
1357
1358ssize_t
1359arena_decay_time_get(arena_t *arena)
1360{
1361 ssize_t decay_time;
1362
1363 malloc_mutex_lock(&arena->lock);
1364 decay_time = arena->decay_time;
1365 malloc_mutex_unlock(&arena->lock);
1366
1367 return (decay_time);
1368}
1369
1370bool
1371arena_decay_time_set(arena_t *arena, ssize_t decay_time)
1372{
1373
1374 if (!arena_decay_time_valid(decay_time))
1375 return (true);
1376
1377 malloc_mutex_lock(&arena->lock);
1378 /*
1379 * Restart decay backlog from scratch, which may cause many dirty pages
1380 * to be immediately purged. It would conceptually be possible to map
1381 * the old backlog onto the new backlog, but there is no justification
1382 * for such complexity since decay_time changes are intended to be
1383 * infrequent, either between the {-1, 0, >0} states, or a one-time
1384 * arbitrary change during initial arena configuration.
1385 */
1386 arena_decay_init(arena, decay_time);
1387 arena_maybe_purge(arena);
1388 malloc_mutex_unlock(&arena->lock);
1389
1390 return (false);
1391}
1392
1393static void
1394arena_maybe_purge_ratio(arena_t *arena)
1395{
1396
1397 assert(opt_purge == purge_mode_ratio);
1398
1399 /* Don't purge if the option is disabled. */
1400 if (arena->lg_dirty_mult < 0)
1401 return;
1402
1403 /*
1404 * Iterate, since preventing recursive purging could otherwise leave too
1405 * many dirty pages.
1406 */
1407 while (true) {
1408 size_t threshold = (arena->nactive >> arena->lg_dirty_mult);
1409 if (threshold < chunk_npages)
1410 threshold = chunk_npages;
1411 /*
1412 * Don't purge unless the number of purgeable pages exceeds the
1413 * threshold.
1414 */
1415 if (arena->ndirty <= threshold)
1416 return;
1417 arena_purge_to_limit(arena, threshold);
1418 }
1419}
1420
1421static void
1422arena_maybe_purge_decay(arena_t *arena)
1423{
1424 nstime_t time;
1425 size_t ndirty_limit;
1426
1427 assert(opt_purge == purge_mode_decay);
1428
1429 /* Purge all or nothing if the option is disabled. */
1430 if (arena->decay_time <= 0) {
1431 if (arena->decay_time == 0)
1432 arena_purge_to_limit(arena, 0);
1433 return;
1434 }
1435
1436 nstime_copy(&time, &arena->decay_epoch);
1437 if (unlikely(nstime_update(&time))) {
1438 /* Time went backwards. Force an epoch advance. */
1439 nstime_copy(&time, &arena->decay_deadline);
1440 }
1441
1442 if (arena_decay_deadline_reached(arena, &time))
1443 arena_decay_epoch_advance(arena, &time);
1444
1445 ndirty_limit = arena_decay_npages_limit(arena);
1446
1447 /*
1448 * Don't try to purge unless the number of purgeable pages exceeds the
1449 * current limit.
1450 */
1451 if (arena->ndirty <= ndirty_limit)
1452 return;
1453 arena_purge_to_limit(arena, ndirty_limit);
1454}
1455
1456void
1457arena_maybe_purge(arena_t *arena)
1458{
1459
1460 /* Don't recursively purge. */
1461 if (arena->purging)
1462 return;
1463
1464 if (opt_purge == purge_mode_ratio)
1465 arena_maybe_purge_ratio(arena);
1466 else
1467 arena_maybe_purge_decay(arena);
1468}
1469
1470static size_t
1471arena_dirty_count(arena_t *arena)
1472{
1473 size_t ndirty = 0;
1474 arena_runs_dirty_link_t *rdelm;
1475 extent_node_t *chunkselm;
1476
1477 for (rdelm = qr_next(&arena->runs_dirty, rd_link),
--- 19 unchanged lines hidden (view full) ---
1497 }
1498 ndirty += npages;
1499 }
1500
1501 return (ndirty);
1502}
1503
1504static size_t
1505arena_stash_dirty(arena_t *arena, chunk_hooks_t *chunk_hooks,
1506 size_t ndirty_limit, arena_runs_dirty_link_t *purge_runs_sentinel,
1507 extent_node_t *purge_chunks_sentinel)
1508{
1509 arena_runs_dirty_link_t *rdelm, *rdelm_next;
1510 extent_node_t *chunkselm;
1511 size_t nstashed = 0;
1512
1513 /* Stash runs/chunks according to ndirty_limit. */
1514 for (rdelm = qr_next(&arena->runs_dirty, rd_link),
1515 chunkselm = qr_next(&arena->chunks_cache, cc_link);
1516 rdelm != &arena->runs_dirty; rdelm = rdelm_next) {
1517 size_t npages;
1518 rdelm_next = qr_next(rdelm, rd_link);
1519
1520 if (rdelm == &chunkselm->rd) {
1521 extent_node_t *chunkselm_next;
1522 bool zero;
1523 UNUSED void *chunk;
1524
1525 npages = extent_node_size_get(chunkselm) >> LG_PAGE;
1526 if (opt_purge == purge_mode_decay && arena->ndirty -
1527 (nstashed + npages) < ndirty_limit)
1528 break;
1529
1530 chunkselm_next = qr_next(chunkselm, cc_link);
1531 /*
1532 * Allocate. chunkselm remains valid due to the
1533 * dalloc_node=false argument to chunk_alloc_cache().
1534 */
1535 zero = false;
1536 chunk = chunk_alloc_cache(arena, chunk_hooks,
1537 extent_node_addr_get(chunkselm),
1538 extent_node_size_get(chunkselm), chunksize, &zero,
1539 false);
1540 assert(chunk == extent_node_addr_get(chunkselm));
1541 assert(zero == extent_node_zeroed_get(chunkselm));
1542 extent_node_dirty_insert(chunkselm, purge_runs_sentinel,
1543 purge_chunks_sentinel);
1544 assert(npages == (extent_node_size_get(chunkselm) >>
1545 LG_PAGE));
1546 chunkselm = chunkselm_next;
1547 } else {
1548 arena_chunk_t *chunk =
1549 (arena_chunk_t *)CHUNK_ADDR2BASE(rdelm);
1550 arena_chunk_map_misc_t *miscelm =
1551 arena_rd_to_miscelm(rdelm);
1552 size_t pageind = arena_miscelm_to_pageind(miscelm);
1553 arena_run_t *run = &miscelm->run;
1554 size_t run_size =
1555 arena_mapbits_unallocated_size_get(chunk, pageind);
1556
1557 npages = run_size >> LG_PAGE;
1558 if (opt_purge == purge_mode_decay && arena->ndirty -
1559 (nstashed + npages) < ndirty_limit)
1560 break;
1561
1562 assert(pageind + npages <= chunk_npages);
1563 assert(arena_mapbits_dirty_get(chunk, pageind) ==
1564 arena_mapbits_dirty_get(chunk, pageind+npages-1));
1565
1566 /*
1567 * If purging the spare chunk's run, make it available
1568 * prior to allocation.
--- 9 unchanged lines hidden (view full) ---
1578 else {
1579 assert(qr_next(rdelm, rd_link) == rdelm);
1580 assert(qr_prev(rdelm, rd_link) == rdelm);
1581 }
1582 qr_meld(purge_runs_sentinel, rdelm, rd_link);
1583 }
1584
1585 nstashed += npages;
1586 if (opt_purge == purge_mode_ratio && arena->ndirty - nstashed <=
1587 ndirty_limit)
1588 break;
1589 }
1590
1591 return (nstashed);
1592}
1593
1594static size_t
1595arena_purge_stashed(arena_t *arena, chunk_hooks_t *chunk_hooks,
--- 123 unchanged lines hidden (view full) ---
1719 pageind) != 0);
1720 arena_run_t *run = &miscelm->run;
1721 qr_remove(rdelm, rd_link);
1722 arena_run_dalloc(arena, run, false, true, decommitted);
1723 }
1724 }
1725}
1726
1727/*
1728 * NB: ndirty_limit is interpreted differently depending on opt_purge:
1729 * - purge_mode_ratio: Purge as few dirty run/chunks as possible to reach the
1730 * desired state:
1731 * (arena->ndirty <= ndirty_limit)
1732 * - purge_mode_decay: Purge as many dirty runs/chunks as possible without
1733 * violating the invariant:
1734 * (arena->ndirty >= ndirty_limit)
1735 */
1736static void
1737arena_purge_to_limit(arena_t *arena, size_t ndirty_limit)
1738{
1739 chunk_hooks_t chunk_hooks = chunk_hooks_get(arena);
1740 size_t npurge, npurged;
1741 arena_runs_dirty_link_t purge_runs_sentinel;
1742 extent_node_t purge_chunks_sentinel;
1743
1744 arena->purging = true;
1745
1746 /*
1747 * Calls to arena_dirty_count() are disabled even for debug builds
1748 * because overhead grows nonlinearly as memory usage increases.
1749 */
1750 if (false && config_debug) {
1751 size_t ndirty = arena_dirty_count(arena);
1752 assert(ndirty == arena->ndirty);
1753 }
1754 assert(opt_purge != purge_mode_ratio || (arena->nactive >>
1755 arena->lg_dirty_mult) < arena->ndirty || ndirty_limit == 0);
1756
1757 qr_new(&purge_runs_sentinel, rd_link);
1758 extent_node_dirty_linkage_init(&purge_chunks_sentinel);
1759
1760 npurge = arena_stash_dirty(arena, &chunk_hooks, ndirty_limit,
1761 &purge_runs_sentinel, &purge_chunks_sentinel);
1762 if (npurge == 0)
1763 goto label_return;
1764 npurged = arena_purge_stashed(arena, &chunk_hooks, &purge_runs_sentinel,
1765 &purge_chunks_sentinel);
1766 assert(npurged == npurge);
1767 arena_unstash_purged(arena, &chunk_hooks, &purge_runs_sentinel,
1768 &purge_chunks_sentinel);
1769
1770 if (config_stats)
1771 arena->stats.npurge++;
1772
1773label_return:
1774 arena->purging = false;
1775}
1776
1777void
1778arena_purge(arena_t *arena, bool all)
1779{
1780
1781 malloc_mutex_lock(&arena->lock);
1782 if (all)
1783 arena_purge_to_limit(arena, 0);
1784 else
1785 arena_maybe_purge(arena);
1786 malloc_mutex_unlock(&arena->lock);
1787}
1788
1789static void
1790arena_run_coalesce(arena_t *arena, arena_chunk_t *chunk, size_t *p_size,
1791 size_t *p_run_ind, size_t *p_run_pages, size_t flag_dirty,
1792 size_t flag_decommitted)
1793{
--- 100 unchanged lines hidden (view full) ---
1894 } else {
1895 arena_bin_info_t *bin_info = &arena_bin_info[run->binind];
1896 size = bin_info->run_size;
1897 }
1898
1899 return (size);
1900}
1901
1902static void
1903arena_run_dalloc(arena_t *arena, arena_run_t *run, bool dirty, bool cleaned,
1904 bool decommitted)
1905{
1906 arena_chunk_t *chunk;
1907 arena_chunk_map_misc_t *miscelm;
1908 size_t size, run_ind, run_pages, flag_dirty, flag_decommitted;
1909
1910 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run);
1911 miscelm = arena_run_to_miscelm(run);
1912 run_ind = arena_miscelm_to_pageind(miscelm);
1913 assert(run_ind >= map_bias);
1914 assert(run_ind < chunk_npages);
1915 size = arena_run_size_get(arena, chunk, run, run_ind);
1916 run_pages = (size >> LG_PAGE);
1917 arena_nactive_sub(arena, run_pages);
1918
1919 /*
1920 * The run is dirty if the caller claims to have dirtied it, as well as
1921 * if it was already dirty before being allocated and the caller
1922 * doesn't claim to have cleaned it.
1923 */
1924 assert(arena_mapbits_dirty_get(chunk, run_ind) ==
1925 arena_mapbits_dirty_get(chunk, run_ind+run_pages-1));
--- 45 unchanged lines hidden (view full) ---
1971 * allows for an old spare to be fully deallocated, thus decreasing the
1972 * chances of spuriously crossing the dirty page purging threshold.
1973 */
1974 if (dirty)
1975 arena_maybe_purge(arena);
1976}
1977
1978static void
1979arena_run_trim_head(arena_t *arena, arena_chunk_t *chunk, arena_run_t *run,
1980 size_t oldsize, size_t newsize)
1981{
1982 arena_chunk_map_misc_t *miscelm = arena_run_to_miscelm(run);
1983 size_t pageind = arena_miscelm_to_pageind(miscelm);
1984 size_t head_npages = (oldsize - newsize) >> LG_PAGE;
1985 size_t flag_dirty = arena_mapbits_dirty_get(chunk, pageind);
1986 size_t flag_decommitted = arena_mapbits_decommitted_get(chunk, pageind);
--- 211 unchanged lines hidden (view full) ---
2198 bin->runcur = run;
2199
2200 assert(bin->runcur->nfree > 0);
2201
2202 return (arena_run_reg_alloc(bin->runcur, bin_info));
2203}
2204
2205void
2206arena_tcache_fill_small(tsd_t *tsd, arena_t *arena, tcache_bin_t *tbin,
2207 szind_t binind, uint64_t prof_accumbytes)
2208{
2209 unsigned i, nfill;
2210 arena_bin_t *bin;
2211
2212 assert(tbin->ncached == 0);
2213
2214 if (config_prof && arena_prof_accum(arena, prof_accumbytes))
2215 prof_idump();
--- 6 unchanged lines hidden (view full) ---
2222 if ((run = bin->runcur) != NULL && run->nfree > 0)
2223 ptr = arena_run_reg_alloc(run, &arena_bin_info[binind]);
2224 else
2225 ptr = arena_bin_malloc_hard(arena, bin);
2226 if (ptr == NULL) {
2227 /*
2228 * OOM. tbin->avail isn't yet filled down to its first
2229 * element, so the successful allocations (if any) must
2230 * be moved just before tbin->avail before bailing out.
2231 */
2232 if (i > 0) {
2233 memmove(tbin->avail - i, tbin->avail - nfill,
2234 i * sizeof(void *));
2235 }
2236 break;
2237 }
2238 if (config_fill && unlikely(opt_junk_alloc)) {
2239 arena_alloc_junk_small(ptr, &arena_bin_info[binind],
2240 true);
2241 }
2242 /* Insert such that low regions get used first. */
2243 *(tbin->avail - nfill + i) = ptr;
2244 }
2245 if (config_stats) {
2246 bin->stats.nmalloc += i;
2247 bin->stats.nrequests += tbin->tstats.nrequests;
2248 bin->stats.curregs += i;
2249 bin->stats.nfills++;
2250 tbin->tstats.nrequests = 0;
2251 }
2252 malloc_mutex_unlock(&bin->lock);
2253 tbin->ncached = i;
2254 arena_decay_tick(tsd, arena);
2255}
2256
2257void
2258arena_alloc_junk_small(void *ptr, arena_bin_info_t *bin_info, bool zero)
2259{
2260
2261 if (zero) {
2262 size_t redzone_size = bin_info->redzone_size;
--- 93 unchanged lines hidden (view full) ---
2356 assert(opt_quarantine);
2357 assert(usize <= SMALL_MAXCLASS);
2358
2359 binind = size2index(usize);
2360 bin_info = &arena_bin_info[binind];
2361 arena_redzones_validate(ptr, bin_info, true);
2362}
2363
2364static void *
2365arena_malloc_small(tsd_t *tsd, arena_t *arena, szind_t binind, bool zero)
2366{
2367 void *ret;
2368 arena_bin_t *bin;
2369 size_t usize;
2370 arena_run_t *run;
2371
2372 assert(binind < NBINS);
2373 bin = &arena->bins[binind];
2374 usize = index2size(binind);
2375
2376 malloc_mutex_lock(&bin->lock);
2377 if ((run = bin->runcur) != NULL && run->nfree > 0)
2378 ret = arena_run_reg_alloc(run, &arena_bin_info[binind]);
2379 else
2380 ret = arena_bin_malloc_hard(arena, bin);
2381
2382 if (ret == NULL) {
2383 malloc_mutex_unlock(&bin->lock);
2384 return (NULL);
2385 }
2386
2387 if (config_stats) {
2388 bin->stats.nmalloc++;
2389 bin->stats.nrequests++;
2390 bin->stats.curregs++;
2391 }
2392 malloc_mutex_unlock(&bin->lock);
2393 if (config_prof && !isthreaded && arena_prof_accum(arena, usize))
2394 prof_idump();
2395
2396 if (!zero) {
2397 if (config_fill) {
2398 if (unlikely(opt_junk_alloc)) {
2399 arena_alloc_junk_small(ret,
2400 &arena_bin_info[binind], false);
2401 } else if (unlikely(opt_zero))
2402 memset(ret, 0, usize);
2403 }
2404 JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED(ret, usize);
2405 } else {
2406 if (config_fill && unlikely(opt_junk_alloc)) {
2407 arena_alloc_junk_small(ret, &arena_bin_info[binind],
2408 true);
2409 }
2410 JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED(ret, usize);
2411 memset(ret, 0, usize);
2412 }
2413
2414 arena_decay_tick(tsd, arena);
2415 return (ret);
2416}
2417
2418void *
2419arena_malloc_large(tsd_t *tsd, arena_t *arena, szind_t binind, bool zero)
2420{
2421 void *ret;
2422 size_t usize;
2423 uintptr_t random_offset;
2424 arena_run_t *run;
2425 arena_chunk_map_misc_t *miscelm;
2426 UNUSED bool idump;
2427
2428 /* Large allocation. */
2429 usize = index2size(binind);
2430 malloc_mutex_lock(&arena->lock);
2431 if (config_cache_oblivious) {
2432 uint64_t r;
2433
2434 /*
2435 * Compute a uniformly distributed offset within the first page
2436 * that is a multiple of the cacheline size, e.g. [0 .. 63) * 64
2437 * for 4 KiB pages and 64-byte cachelines.
2438 */
2439 r = prng_lg_range(&arena->offset_state, LG_PAGE - LG_CACHELINE);
2440 random_offset = ((uintptr_t)r) << LG_CACHELINE;
2441 } else
2442 random_offset = 0;
2443 run = arena_run_alloc_large(arena, usize + large_pad, zero);
2444 if (run == NULL) {
2445 malloc_mutex_unlock(&arena->lock);
2446 return (NULL);
2447 }
2448 miscelm = arena_run_to_miscelm(run);
2449 ret = (void *)((uintptr_t)arena_miscelm_to_rpages(miscelm) +
2450 random_offset);
2451 if (config_stats) {
2452 szind_t index = binind - NBINS;
2453
2454 arena->stats.nmalloc_large++;
2455 arena->stats.nrequests_large++;
2456 arena->stats.allocated_large += usize;
2457 arena->stats.lstats[index].nmalloc++;
2458 arena->stats.lstats[index].nrequests++;
2459 arena->stats.lstats[index].curruns++;
2460 }
--- 7 unchanged lines hidden (view full) ---
2468 if (config_fill) {
2469 if (unlikely(opt_junk_alloc))
2470 memset(ret, 0xa5, usize);
2471 else if (unlikely(opt_zero))
2472 memset(ret, 0, usize);
2473 }
2474 }
2475
2476 arena_decay_tick(tsd, arena);
2477 return (ret);
2478}
2479
2480void *
2481arena_malloc_hard(tsd_t *tsd, arena_t *arena, size_t size, szind_t ind,
2482 bool zero, tcache_t *tcache)
2483{
2484
2485 arena = arena_choose(tsd, arena);
2486 if (unlikely(arena == NULL))
2487 return (NULL);
2488
2489 if (likely(size <= SMALL_MAXCLASS))
2490 return (arena_malloc_small(tsd, arena, ind, zero));
2491 if (likely(size <= large_maxclass))
2492 return (arena_malloc_large(tsd, arena, ind, zero));
2493 return (huge_malloc(tsd, arena, index2size(ind), zero, tcache));
2494}
2495
2496/* Only handles large allocations that require more than page alignment. */
2497static void *
2498arena_palloc_large(tsd_t *tsd, arena_t *arena, size_t usize, size_t alignment,
2499 bool zero)
2500{
2501 void *ret;
2502 size_t alloc_size, leadsize, trailsize;
2503 arena_run_t *run;
--- 67 unchanged lines hidden (view full) ---
2571 malloc_mutex_unlock(&arena->lock);
2572
2573 if (config_fill && !zero) {
2574 if (unlikely(opt_junk_alloc))
2575 memset(ret, 0xa5, usize);
2576 else if (unlikely(opt_zero))
2577 memset(ret, 0, usize);
2578 }
2579 arena_decay_tick(tsd, arena);
2580 return (ret);
2581}
2582
2583void *
2584arena_palloc(tsd_t *tsd, arena_t *arena, size_t usize, size_t alignment,
2585 bool zero, tcache_t *tcache)
2586{
2587 void *ret;
2588
2589 if (usize <= SMALL_MAXCLASS && (alignment < PAGE || (alignment == PAGE
2590 && (usize & PAGE_MASK) == 0))) {
2591 /* Small; alignment doesn't require special run placement. */
2592 ret = arena_malloc(tsd, arena, usize, size2index(usize), zero,
2593 tcache, true);
2594 } else if (usize <= large_maxclass && alignment <= PAGE) {
2595 /*
2596 * Large; alignment doesn't require special run placement.
2597 * However, the cached pointer may be at a random offset from
2598 * the base of the run, so do some bit manipulation to retrieve
2599 * the base.
2600 */
2601 ret = arena_malloc(tsd, arena, usize, size2index(usize), zero,
2602 tcache, true);
2603 if (config_cache_oblivious)
2604 ret = (void *)((uintptr_t)ret & ~PAGE_MASK);
2605 } else {
2606 if (likely(usize <= large_maxclass)) {
2607 ret = arena_palloc_large(tsd, arena, usize, alignment,
2608 zero);
2609 } else if (likely(alignment <= chunksize))
2610 ret = huge_malloc(tsd, arena, usize, zero, tcache);
--- 60 unchanged lines hidden (view full) ---
2671
2672 assert(run != bin->runcur);
2673 assert(arena_run_tree_search(&bin->runs, arena_run_to_miscelm(run)) ==
2674 NULL);
2675
2676 malloc_mutex_unlock(&bin->lock);
2677 /******************************/
2678 malloc_mutex_lock(&arena->lock);
2679 arena_run_dalloc(arena, run, true, false, false);
2680 malloc_mutex_unlock(&arena->lock);
2681 /****************************/
2682 malloc_mutex_lock(&bin->lock);
2683 if (config_stats)
2684 bin->stats.curruns--;
2685}
2686
2687static void
--- 70 unchanged lines hidden (view full) ---
2758 run = &arena_miscelm_get(chunk, rpages_ind)->run;
2759 bin = &arena->bins[run->binind];
2760 malloc_mutex_lock(&bin->lock);
2761 arena_dalloc_bin_locked_impl(arena, chunk, ptr, bitselm, false);
2762 malloc_mutex_unlock(&bin->lock);
2763}
2764
2765void
2766arena_dalloc_small(tsd_t *tsd, arena_t *arena, arena_chunk_t *chunk, void *ptr,
2767 size_t pageind)
2768{
2769 arena_chunk_map_bits_t *bitselm;
2770
2771 if (config_debug) {
2772 /* arena_ptr_small_binind_get() does extra sanity checking. */
2773 assert(arena_ptr_small_binind_get(ptr, arena_mapbits_get(chunk,
2774 pageind)) != BININD_INVALID);
2775 }
2776 bitselm = arena_bitselm_get(chunk, pageind);
2777 arena_dalloc_bin(arena, chunk, ptr, pageind, bitselm);
2778 arena_decay_tick(tsd, arena);
2779}
2780
2781#ifdef JEMALLOC_JET
2782#undef arena_dalloc_junk_large
2783#define arena_dalloc_junk_large JEMALLOC_N(arena_dalloc_junk_large_impl)
2784#endif
2785void
2786arena_dalloc_junk_large(void *ptr, size_t usize)
--- 28 unchanged lines hidden (view full) ---
2815
2816 arena->stats.ndalloc_large++;
2817 arena->stats.allocated_large -= usize;
2818 arena->stats.lstats[index].ndalloc++;
2819 arena->stats.lstats[index].curruns--;
2820 }
2821 }
2822
2823 arena_run_dalloc(arena, run, true, false, false);
2824}
2825
2826void
2827arena_dalloc_large_junked_locked(arena_t *arena, arena_chunk_t *chunk,
2828 void *ptr)
2829{
2830
2831 arena_dalloc_large_locked_impl(arena, chunk, ptr, true);
2832}
2833
2834void
2835arena_dalloc_large(tsd_t *tsd, arena_t *arena, arena_chunk_t *chunk, void *ptr)
2836{
2837
2838 malloc_mutex_lock(&arena->lock);
2839 arena_dalloc_large_locked_impl(arena, chunk, ptr, false);
2840 malloc_mutex_unlock(&arena->lock);
2841 arena_decay_tick(tsd, arena);
2842}
2843
2844static void
2845arena_ralloc_large_shrink(arena_t *arena, arena_chunk_t *chunk, void *ptr,
2846 size_t oldsize, size_t size)
2847{
2848 size_t pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
2849 arena_chunk_map_misc_t *miscelm = arena_miscelm_get(chunk, pageind);
--- 184 unchanged lines hidden (view full) ---
3034 assert(oldsize > usize_max);
3035 /* Fill before shrinking in order avoid a race. */
3036 arena_ralloc_junk_large(ptr, oldsize, usize_max);
3037 arena_ralloc_large_shrink(arena, chunk, ptr, oldsize, usize_max);
3038 return (false);
3039}
3040
3041bool
3042arena_ralloc_no_move(tsd_t *tsd, void *ptr, size_t oldsize, size_t size,
3043 size_t extra, bool zero)
3044{
3045 size_t usize_min, usize_max;
3046
3047 /* Calls with non-zero extra had to clamp extra. */
3048 assert(extra == 0 || size + extra <= HUGE_MAXCLASS);
3049
3050 if (unlikely(size > HUGE_MAXCLASS))
3051 return (true);
3052
3053 usize_min = s2u(size);
3054 usize_max = s2u(size + extra);
3055 if (likely(oldsize <= large_maxclass && usize_min <= large_maxclass)) {
3056 arena_chunk_t *chunk;
3057
3058 /*
3059 * Avoid moving the allocation if the size class can be left the
3060 * same.
3061 */
3062 if (oldsize <= SMALL_MAXCLASS) {
3063 assert(arena_bin_info[size2index(oldsize)].reg_size ==
3064 oldsize);
3065 if ((usize_max > SMALL_MAXCLASS ||
3066 size2index(usize_max) != size2index(oldsize)) &&
3067 (size > oldsize || usize_max < oldsize))
3068 return (true);
3069 } else {
3070 if (usize_max <= SMALL_MAXCLASS)
3071 return (true);
3072 if (arena_ralloc_large(ptr, oldsize, usize_min,
3073 usize_max, zero))
3074 return (true);
3075 }
3076
3077 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
3078 arena_decay_tick(tsd, extent_node_arena_get(&chunk->node));
3079 return (false);
3080 } else {
3081 return (huge_ralloc_no_move(tsd, ptr, oldsize, usize_min,
3082 usize_max, zero));
3083 }
3084}
3085
3086static void *
3087arena_ralloc_move_helper(tsd_t *tsd, arena_t *arena, size_t usize,
3088 size_t alignment, bool zero, tcache_t *tcache)
3089{
3090
3091 if (alignment == 0)
3092 return (arena_malloc(tsd, arena, usize, size2index(usize), zero,
3093 tcache, true));
3094 usize = sa2u(usize, alignment);
3095 if (unlikely(usize == 0 || usize > HUGE_MAXCLASS))
3096 return (NULL);
3097 return (ipalloct(tsd, usize, alignment, zero, tcache, arena));
3098}
3099
3100void *
3101arena_ralloc(tsd_t *tsd, arena_t *arena, void *ptr, size_t oldsize, size_t size,
3102 size_t alignment, bool zero, tcache_t *tcache)
3103{
3104 void *ret;
3105 size_t usize;
3106
3107 usize = s2u(size);
3108 if (unlikely(usize == 0 || size > HUGE_MAXCLASS))
3109 return (NULL);
3110
3111 if (likely(usize <= large_maxclass)) {
3112 size_t copysize;
3113
3114 /* Try to avoid moving the allocation. */
3115 if (!arena_ralloc_no_move(tsd, ptr, oldsize, usize, 0, zero))
3116 return (ptr);
3117
3118 /*
3119 * size and oldsize are different enough that we need to move
3120 * the object. In that case, fall back to allocating new space
3121 * and copying.
3122 */
3123 ret = arena_ralloc_move_helper(tsd, arena, usize, alignment,
--- 46 unchanged lines hidden (view full) ---
3170
3171 return ((ssize_t)atomic_read_z((size_t *)&lg_dirty_mult_default));
3172}
3173
3174bool
3175arena_lg_dirty_mult_default_set(ssize_t lg_dirty_mult)
3176{
3177
3178 if (opt_purge != purge_mode_ratio)
3179 return (true);
3180 if (!arena_lg_dirty_mult_valid(lg_dirty_mult))
3181 return (true);
3182 atomic_write_z((size_t *)&lg_dirty_mult_default, (size_t)lg_dirty_mult);
3183 return (false);
3184}
3185
3186ssize_t
3187arena_decay_time_default_get(void)
3188{
3189
3190 return ((ssize_t)atomic_read_z((size_t *)&decay_time_default));
3191}
3192
3193bool
3194arena_decay_time_default_set(ssize_t decay_time)
3195{
3196
3197 if (opt_purge != purge_mode_decay)
3198 return (true);
3199 if (!arena_decay_time_valid(decay_time))
3200 return (true);
3201 atomic_write_z((size_t *)&decay_time_default, (size_t)decay_time);
3202 return (false);
3203}
3204
3205static void
3206arena_basic_stats_merge_locked(arena_t *arena, unsigned *nthreads,
3207 const char **dss, ssize_t *lg_dirty_mult, ssize_t *decay_time,
3208 size_t *nactive, size_t *ndirty)
3209{
3210
3211 *nthreads += arena_nthreads_get(arena);
3212 *dss = dss_prec_names[arena->dss_prec];
3213 *lg_dirty_mult = arena->lg_dirty_mult;
3214 *decay_time = arena->decay_time;
3215 *nactive += arena->nactive;
3216 *ndirty += arena->ndirty;
3217}
3218
3219void
3220arena_basic_stats_merge(arena_t *arena, unsigned *nthreads, const char **dss,
3221 ssize_t *lg_dirty_mult, ssize_t *decay_time, size_t *nactive,
3222 size_t *ndirty)
3223{
3224
3225 malloc_mutex_lock(&arena->lock);
3226 arena_basic_stats_merge_locked(arena, nthreads, dss, lg_dirty_mult,
3227 decay_time, nactive, ndirty);
3228 malloc_mutex_unlock(&arena->lock);
3229}
3230
3231void
3232arena_stats_merge(arena_t *arena, unsigned *nthreads, const char **dss,
3233 ssize_t *lg_dirty_mult, ssize_t *decay_time, size_t *nactive,
3234 size_t *ndirty, arena_stats_t *astats, malloc_bin_stats_t *bstats,
3235 malloc_large_stats_t *lstats, malloc_huge_stats_t *hstats)
3236{
3237 unsigned i;
3238
3239 cassert(config_stats);
3240
3241 malloc_mutex_lock(&arena->lock);
3242 arena_basic_stats_merge_locked(arena, nthreads, dss, lg_dirty_mult,
3243 decay_time, nactive, ndirty);
3244
3245 astats->mapped += arena->stats.mapped;
3246 astats->npurge += arena->stats.npurge;
3247 astats->nmadvise += arena->stats.nmadvise;
3248 astats->purged += arena->stats.purged;
3249 astats->metadata_mapped += arena->stats.metadata_mapped;
3250 astats->metadata_allocated += arena_metadata_allocated_get(arena);
3251 astats->allocated_large += arena->stats.allocated_large;
3252 astats->nmalloc_large += arena->stats.nmalloc_large;
--- 31 unchanged lines hidden (view full) ---
3284 }
3285 bstats[i].nruns += bin->stats.nruns;
3286 bstats[i].reruns += bin->stats.reruns;
3287 bstats[i].curruns += bin->stats.curruns;
3288 malloc_mutex_unlock(&bin->lock);
3289 }
3290}
3291
3292unsigned
3293arena_nthreads_get(arena_t *arena)
3294{
3295
3296 return (atomic_read_u(&arena->nthreads));
3297}
3298
3299void
3300arena_nthreads_inc(arena_t *arena)
3301{
3302
3303 atomic_add_u(&arena->nthreads, 1);
3304}
3305
3306void
3307arena_nthreads_dec(arena_t *arena)
3308{
3309
3310 atomic_sub_u(&arena->nthreads, 1);
3311}
3312
3313arena_t *
3314arena_new(unsigned ind)
3315{
3316 arena_t *arena;
3317 size_t arena_size;
3318 unsigned i;
3319 arena_bin_t *bin;
3320
3321 /* Compute arena size to incorporate sufficient runs_avail elements. */
3322 arena_size = offsetof(arena_t, runs_avail) + (sizeof(arena_run_tree_t) *
3323 runs_avail_nclasses);
3324 /*
3325 * Allocate arena, arena->lstats, and arena->hstats contiguously, mainly
3326 * because there is no way to clean up if base_alloc() OOMs.
3327 */
3328 if (config_stats) {
3329 arena = (arena_t *)base_alloc(CACHELINE_CEILING(arena_size) +
3330 QUANTUM_CEILING(nlclasses * sizeof(malloc_large_stats_t) +
3331 nhclasses) * sizeof(malloc_huge_stats_t));
3332 } else
3333 arena = (arena_t *)base_alloc(arena_size);
3334 if (arena == NULL)
3335 return (NULL);
3336
3337 arena->ind = ind;
3338 arena->nthreads = 0;
3339 if (malloc_mutex_init(&arena->lock))
3340 return (NULL);
3341
3342 if (config_stats) {
3343 memset(&arena->stats, 0, sizeof(arena_stats_t));
3344 arena->stats.lstats = (malloc_large_stats_t *)((uintptr_t)arena
3345 + CACHELINE_CEILING(arena_size));
3346 memset(arena->stats.lstats, 0, nlclasses *
3347 sizeof(malloc_large_stats_t));
3348 arena->stats.hstats = (malloc_huge_stats_t *)((uintptr_t)arena
3349 + CACHELINE_CEILING(arena_size) +
3350 QUANTUM_CEILING(nlclasses * sizeof(malloc_large_stats_t)));
3351 memset(arena->stats.hstats, 0, nhclasses *
3352 sizeof(malloc_huge_stats_t));
3353 if (config_tcache)
3354 ql_new(&arena->tcache_ql);
3355 }
3356
3357 if (config_prof)
--- 15 unchanged lines hidden (view full) ---
3373
3374 arena->spare = NULL;
3375
3376 arena->lg_dirty_mult = arena_lg_dirty_mult_default_get();
3377 arena->purging = false;
3378 arena->nactive = 0;
3379 arena->ndirty = 0;
3380
3381 for(i = 0; i < runs_avail_nclasses; i++)
3382 arena_run_tree_new(&arena->runs_avail[i]);
3383 qr_new(&arena->runs_dirty, rd_link);
3384 qr_new(&arena->chunks_cache, cc_link);
3385
3386 if (opt_purge == purge_mode_decay)
3387 arena_decay_init(arena, arena_decay_time_default_get());
3388
3389 ql_new(&arena->huge);
3390 if (malloc_mutex_init(&arena->huge_mtx))
3391 return (NULL);
3392
3393 extent_tree_szad_new(&arena->chunks_szad_cached);
3394 extent_tree_ad_new(&arena->chunks_ad_cached);
3395 extent_tree_szad_new(&arena->chunks_szad_retained);
3396 extent_tree_ad_new(&arena->chunks_ad_retained);
--- 38 unchanged lines hidden (view full) ---
3435 /*
3436 * Determine redzone size based on minimum alignment and minimum
3437 * redzone size. Add padding to the end of the run if it is needed to
3438 * align the regions. The padding allows each redzone to be half the
3439 * minimum alignment; without the padding, each redzone would have to
3440 * be twice as large in order to maintain alignment.
3441 */
3442 if (config_fill && unlikely(opt_redzone)) {
3443 size_t align_min = ZU(1) << (ffs_zu(bin_info->reg_size) - 1);
3444 if (align_min <= REDZONE_MINSIZE) {
3445 bin_info->redzone_size = REDZONE_MINSIZE;
3446 pad_size = 0;
3447 } else {
3448 bin_info->redzone_size = align_min >> 1;
3449 pad_size = bin_info->redzone_size;
3450 }
3451 } else {
3452 bin_info->redzone_size = 0;
3453 pad_size = 0;
3454 }
3455 bin_info->reg_interval = bin_info->reg_size +
3456 (bin_info->redzone_size << 1);
3457
3458 /*
3459 * Compute run size under ideal conditions (no redzones, no limit on run
3460 * size).
3461 */
3462 try_run_size = PAGE;
3463 try_nregs = (uint32_t)(try_run_size / bin_info->reg_size);
3464 do {
3465 perfect_run_size = try_run_size;
3466 perfect_nregs = try_nregs;
3467
3468 try_run_size += PAGE;
3469 try_nregs = (uint32_t)(try_run_size / bin_info->reg_size);
3470 } while (perfect_run_size != perfect_nregs * bin_info->reg_size);
3471 assert(perfect_nregs <= RUN_MAXREGS);
3472
3473 actual_run_size = perfect_run_size;
3474 actual_nregs = (uint32_t)((actual_run_size - pad_size) /
3475 bin_info->reg_interval);
3476
3477 /*
3478 * Redzones can require enough padding that not even a single region can
3479 * fit within the number of pages that would normally be dedicated to a
3480 * run for this size class. Increase the run size until at least one
3481 * region fits.
3482 */
3483 while (actual_nregs == 0) {
3484 assert(config_fill && unlikely(opt_redzone));
3485
3486 actual_run_size += PAGE;
3487 actual_nregs = (uint32_t)((actual_run_size - pad_size) /
3488 bin_info->reg_interval);
3489 }
3490
3491 /*
3492 * Make sure that the run will fit within an arena chunk.
3493 */
3494 while (actual_run_size > arena_maxrun) {
3495 actual_run_size -= PAGE;
3496 actual_nregs = (uint32_t)((actual_run_size - pad_size) /
3497 bin_info->reg_interval);
3498 }
3499 assert(actual_nregs > 0);
3500 assert(actual_run_size == s2u(actual_run_size));
3501
3502 /* Copy final settings. */
3503 bin_info->run_size = actual_run_size;
3504 bin_info->nregs = actual_nregs;
3505 bin_info->reg0_offset = (uint32_t)(actual_run_size - (actual_nregs *
3506 bin_info->reg_interval) - pad_size + bin_info->redzone_size);
3507
3508 if (actual_run_size > small_maxrun)
3509 small_maxrun = actual_run_size;
3510
3511 assert(bin_info->reg0_offset - bin_info->redzone_size + (bin_info->nregs
3512 * bin_info->reg_interval) + pad_size == bin_info->run_size);
3513}
3514
--- 37 unchanged lines hidden (view full) ---
3552 SIZE_CLASSES
3553#undef TAB_INIT_bin_yes
3554#undef TAB_INIT_bin_no
3555#undef SC
3556
3557 return (false);
3558}
3559
3560static bool
3561run_quantize_init(void)
3562{
3563 unsigned i;
3564
3565 run_quantize_max = chunksize + large_pad;
3566
3567 run_quantize_floor_tab = (size_t *)base_alloc(sizeof(size_t) *
3568 (run_quantize_max >> LG_PAGE));
3569 if (run_quantize_floor_tab == NULL)
3570 return (true);
3571
3572 run_quantize_ceil_tab = (size_t *)base_alloc(sizeof(size_t) *
3573 (run_quantize_max >> LG_PAGE));
3574 if (run_quantize_ceil_tab == NULL)
3575 return (true);
3576
3577 for (i = 1; i <= run_quantize_max >> LG_PAGE; i++) {
3578 size_t run_size = i << LG_PAGE;
3579
3580 run_quantize_floor_tab[i-1] =
3581 run_quantize_floor_compute(run_size);
3582 run_quantize_ceil_tab[i-1] =
3583 run_quantize_ceil_compute(run_size);
3584 }
3585
3586 return (false);
3587}
3588
3589bool
3590arena_boot(void)
3591{
3592 unsigned i;
3593
3594 arena_lg_dirty_mult_default_set(opt_lg_dirty_mult);
3595 arena_decay_time_default_set(opt_decay_time);
3596
3597 /*
3598 * Compute the header size such that it is large enough to contain the
3599 * page map. The page map is biased to omit entries for the header
3600 * itself, so some iteration is necessary to compute the map bias.
3601 *
3602 * 1) Compute safe header_size and map_bias values that include enough
3603 * space for an unbiased page map.
--- 25 unchanged lines hidden (view full) ---
3629 */
3630 large_maxclass = arena_maxrun;
3631 }
3632 assert(large_maxclass > 0);
3633 nlclasses = size2index(large_maxclass) - size2index(SMALL_MAXCLASS);
3634 nhclasses = NSIZES - nlclasses - NBINS;
3635
3636 bin_info_init();
3637 if (small_run_size_init())
3638 return (true);
3639 if (run_quantize_init())
3640 return (true);
3641
3642 runs_avail_bias = size2index(PAGE);
3643 runs_avail_nclasses = size2index(run_quantize_max)+1 - runs_avail_bias;
3644
3645 return (false);
3646}
3647
3648void
3649arena_prefork(arena_t *arena)
3650{
3651 unsigned i;
3652
3653 malloc_mutex_prefork(&arena->lock);
--- 32 unchanged lines hidden ---