]>
Commit | Line | Data |
---|---|---|
970d7e83 LB |
1 | #define JEMALLOC_ARENA_C_ |
2 | #include "jemalloc/internal/jemalloc_internal.h" | |
3 | ||
4 | /******************************************************************************/ | |
5 | /* Data. */ | |
6 | ||
54a0048b SL |
7 | purge_mode_t opt_purge = PURGE_DEFAULT; |
8 | const char *purge_mode_names[] = { | |
9 | "ratio", | |
10 | "decay", | |
11 | "N/A" | |
12 | }; | |
970d7e83 | 13 | ssize_t opt_lg_dirty_mult = LG_DIRTY_MULT_DEFAULT; |
54a0048b SL |
14 | static ssize_t lg_dirty_mult_default; |
15 | ssize_t opt_decay_time = DECAY_TIME_DEFAULT; | |
16 | static ssize_t decay_time_default; | |
970d7e83 | 17 | |
54a0048b | 18 | arena_bin_info_t arena_bin_info[NBINS]; |
7453a54e | 19 | |
54a0048b SL |
20 | size_t map_bias; |
21 | size_t map_misc_offset; | |
22 | size_t arena_maxrun; /* Max run size for arenas. */ | |
23 | size_t large_maxclass; /* Max large size class. */ | |
24 | size_t run_quantize_max; /* Max run_quantize_*() input. */ | |
25 | static size_t small_maxrun; /* Max run size for small size classes. */ | |
26 | static bool *small_run_tab; /* Valid small run page multiples. */ | |
27 | static size_t *run_quantize_floor_tab; /* run_quantize_floor() memoization. */ | |
28 | static size_t *run_quantize_ceil_tab; /* run_quantize_ceil() memoization. */ | |
29 | unsigned nlclasses; /* Number of large size classes. */ | |
30 | unsigned nhclasses; /* Number of huge size classes. */ | |
31 | static szind_t runs_avail_bias; /* Size index for first runs_avail tree. */ | |
32 | static szind_t runs_avail_nclasses; /* Number of runs_avail trees. */ | |
970d7e83 LB |
33 | |
34 | /******************************************************************************/ | |
1a4d82fc JJ |
35 | /* |
36 | * Function prototypes for static functions that are referenced prior to | |
37 | * definition. | |
38 | */ | |
39 | ||
54a0048b | 40 | static void arena_purge_to_limit(arena_t *arena, size_t ndirty_limit); |
970d7e83 | 41 | static void arena_run_dalloc(arena_t *arena, arena_run_t *run, bool dirty, |
54a0048b | 42 | bool cleaned, bool decommitted); |
970d7e83 LB |
43 | static void arena_dalloc_bin_run(arena_t *arena, arena_chunk_t *chunk, |
44 | arena_run_t *run, arena_bin_t *bin); | |
45 | static void arena_bin_lower_run(arena_t *arena, arena_chunk_t *chunk, | |
46 | arena_run_t *run, arena_bin_t *bin); | |
970d7e83 LB |
47 | |
48 | /******************************************************************************/ | |
49 | ||
1a4d82fc | 50 | JEMALLOC_INLINE_C size_t |
54a0048b | 51 | arena_miscelm_size_get(const arena_chunk_map_misc_t *miscelm) |
1a4d82fc | 52 | { |
54a0048b SL |
53 | arena_chunk_t *chunk; |
54 | size_t pageind, mapbits; | |
1a4d82fc | 55 | |
54a0048b SL |
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)); | |
1a4d82fc JJ |
60 | } |
61 | ||
54a0048b SL |
62 | JEMALLOC_INLINE_C int |
63 | arena_run_addr_comp(const arena_chunk_map_misc_t *a, | |
64 | const arena_chunk_map_misc_t *b) | |
970d7e83 | 65 | { |
1a4d82fc JJ |
66 | uintptr_t a_miscelm = (uintptr_t)a; |
67 | uintptr_t b_miscelm = (uintptr_t)b; | |
970d7e83 LB |
68 | |
69 | assert(a != NULL); | |
70 | assert(b != NULL); | |
71 | ||
1a4d82fc | 72 | return ((a_miscelm > b_miscelm) - (a_miscelm < b_miscelm)); |
970d7e83 LB |
73 | } |
74 | ||
75 | /* Generate red-black tree functions. */ | |
1a4d82fc | 76 | rb_gen(static UNUSED, arena_run_tree_, arena_run_tree_t, arena_chunk_map_misc_t, |
54a0048b | 77 | rb_link, arena_run_addr_comp) |
970d7e83 | 78 | |
54a0048b SL |
79 | static size_t |
80 | run_quantize_floor_compute(size_t size) | |
970d7e83 | 81 | { |
54a0048b | 82 | size_t qsize; |
970d7e83 | 83 | |
54a0048b SL |
84 | assert(size != 0); |
85 | assert(size == PAGE_CEILING(size)); | |
9cc50fc6 | 86 | |
54a0048b SL |
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 | ||
103 | static size_t | |
104 | run_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. | |
113 | * Quantized sizes comprise the union of run sizes that back small | |
114 | * region runs, and run sizes that back large regions with no explicit | |
115 | * alignment constraints. | |
116 | */ | |
117 | ||
118 | if (size > SMALL_MAXCLASS) { | |
119 | large_run_size_next = PAGE_CEILING(index2size(size2index(size - | |
120 | large_pad) + 1) + large_pad); | |
121 | } else | |
122 | large_run_size_next = SIZE_T_MAX; | |
123 | if (size >= small_maxrun) | |
124 | return (large_run_size_next); | |
125 | ||
126 | while (true) { | |
127 | size += PAGE; | |
128 | assert(size <= small_maxrun); | |
129 | if (small_run_tab[size >> LG_PAGE]) { | |
130 | if (large_run_size_next < size) | |
131 | return (large_run_size_next); | |
132 | return (size); | |
970d7e83 | 133 | } |
970d7e83 | 134 | } |
54a0048b SL |
135 | } |
136 | ||
137 | static size_t | |
138 | run_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 | |
160 | static size_t | |
161 | run_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); | |
970d7e83 | 168 | |
54a0048b SL |
169 | ret = run_quantize_floor_tab[(size >> LG_PAGE) - 1]; |
170 | assert(ret == run_quantize_floor_compute(size)); | |
970d7e83 LB |
171 | return (ret); |
172 | } | |
54a0048b SL |
173 | #ifdef JEMALLOC_JET |
174 | #undef run_quantize_floor | |
175 | #define run_quantize_floor JEMALLOC_N(run_quantize_floor) | |
176 | run_quantize_t *run_quantize_floor = JEMALLOC_N(run_quantize_floor_impl); | |
177 | #endif | |
970d7e83 | 178 | |
54a0048b SL |
179 | #ifdef JEMALLOC_JET |
180 | #undef run_quantize_ceil | |
181 | #define run_quantize_ceil JEMALLOC_N(run_quantize_ceil_impl) | |
182 | #endif | |
183 | static size_t | |
184 | run_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) | |
199 | run_quantize_t *run_quantize_ceil = JEMALLOC_N(run_quantize_ceil_impl); | |
200 | #endif | |
201 | ||
202 | static arena_run_tree_t * | |
203 | arena_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 | } | |
970d7e83 | 211 | |
1a4d82fc JJ |
212 | static void |
213 | arena_avail_insert(arena_t *arena, arena_chunk_t *chunk, size_t pageind, | |
214 | size_t npages) | |
970d7e83 | 215 | { |
54a0048b SL |
216 | szind_t ind = size2index(run_quantize_floor(arena_miscelm_size_get( |
217 | arena_miscelm_get(chunk, pageind)))); | |
1a4d82fc JJ |
218 | assert(npages == (arena_mapbits_unallocated_size_get(chunk, pageind) >> |
219 | LG_PAGE)); | |
54a0048b SL |
220 | arena_run_tree_insert(arena_runs_avail_get(arena, ind), |
221 | arena_miscelm_get(chunk, pageind)); | |
970d7e83 LB |
222 | } |
223 | ||
1a4d82fc JJ |
224 | static void |
225 | arena_avail_remove(arena_t *arena, arena_chunk_t *chunk, size_t pageind, | |
226 | size_t npages) | |
970d7e83 | 227 | { |
54a0048b SL |
228 | szind_t ind = size2index(run_quantize_floor(arena_miscelm_size_get( |
229 | arena_miscelm_get(chunk, pageind)))); | |
1a4d82fc JJ |
230 | assert(npages == (arena_mapbits_unallocated_size_get(chunk, pageind) >> |
231 | LG_PAGE)); | |
54a0048b SL |
232 | arena_run_tree_remove(arena_runs_avail_get(arena, ind), |
233 | arena_miscelm_get(chunk, pageind)); | |
970d7e83 LB |
234 | } |
235 | ||
236 | static void | |
54a0048b | 237 | arena_run_dirty_insert(arena_t *arena, arena_chunk_t *chunk, size_t pageind, |
1a4d82fc | 238 | size_t npages) |
970d7e83 | 239 | { |
1a4d82fc | 240 | arena_chunk_map_misc_t *miscelm = arena_miscelm_get(chunk, pageind); |
54a0048b | 241 | |
970d7e83 LB |
242 | assert(npages == (arena_mapbits_unallocated_size_get(chunk, pageind) >> |
243 | LG_PAGE)); | |
1a4d82fc JJ |
244 | assert(arena_mapbits_dirty_get(chunk, pageind) == CHUNK_MAP_DIRTY); |
245 | assert(arena_mapbits_dirty_get(chunk, pageind+npages-1) == | |
246 | CHUNK_MAP_DIRTY); | |
54a0048b SL |
247 | |
248 | qr_new(&miscelm->rd, rd_link); | |
249 | qr_meld(&arena->runs_dirty, &miscelm->rd, rd_link); | |
1a4d82fc | 250 | arena->ndirty += npages; |
970d7e83 LB |
251 | } |
252 | ||
253 | static void | |
54a0048b | 254 | arena_run_dirty_remove(arena_t *arena, arena_chunk_t *chunk, size_t pageind, |
1a4d82fc | 255 | size_t npages) |
970d7e83 | 256 | { |
1a4d82fc | 257 | arena_chunk_map_misc_t *miscelm = arena_miscelm_get(chunk, pageind); |
54a0048b | 258 | |
970d7e83 LB |
259 | assert(npages == (arena_mapbits_unallocated_size_get(chunk, pageind) >> |
260 | LG_PAGE)); | |
1a4d82fc JJ |
261 | assert(arena_mapbits_dirty_get(chunk, pageind) == CHUNK_MAP_DIRTY); |
262 | assert(arena_mapbits_dirty_get(chunk, pageind+npages-1) == | |
263 | CHUNK_MAP_DIRTY); | |
54a0048b SL |
264 | |
265 | qr_remove(&miscelm->rd, rd_link); | |
266 | assert(arena->ndirty >= npages); | |
1a4d82fc | 267 | arena->ndirty -= npages; |
970d7e83 LB |
268 | } |
269 | ||
54a0048b SL |
270 | static size_t |
271 | arena_chunk_dirty_npages(const extent_node_t *node) | |
272 | { | |
273 | ||
274 | return (extent_node_size_get(node) >> LG_PAGE); | |
275 | } | |
276 | ||
277 | void | |
278 | arena_chunk_cache_maybe_insert(arena_t *arena, extent_node_t *node, bool cache) | |
279 | { | |
280 | ||
281 | if (cache) { | |
282 | extent_node_dirty_linkage_init(node); | |
283 | extent_node_dirty_insert(node, &arena->runs_dirty, | |
284 | &arena->chunks_cache); | |
285 | arena->ndirty += arena_chunk_dirty_npages(node); | |
286 | } | |
287 | } | |
288 | ||
289 | void | |
290 | arena_chunk_cache_maybe_remove(arena_t *arena, extent_node_t *node, bool dirty) | |
291 | { | |
292 | ||
293 | if (dirty) { | |
294 | extent_node_dirty_remove(node); | |
295 | assert(arena->ndirty >= arena_chunk_dirty_npages(node)); | |
296 | arena->ndirty -= arena_chunk_dirty_npages(node); | |
297 | } | |
298 | } | |
299 | ||
300 | JEMALLOC_INLINE_C void * | |
970d7e83 LB |
301 | arena_run_reg_alloc(arena_run_t *run, arena_bin_info_t *bin_info) |
302 | { | |
303 | void *ret; | |
54a0048b | 304 | size_t regind; |
1a4d82fc JJ |
305 | arena_chunk_map_misc_t *miscelm; |
306 | void *rpages; | |
970d7e83 LB |
307 | |
308 | assert(run->nfree > 0); | |
1a4d82fc | 309 | assert(!bitmap_full(run->bitmap, &bin_info->bitmap_info)); |
970d7e83 | 310 | |
54a0048b | 311 | regind = (unsigned)bitmap_sfu(run->bitmap, &bin_info->bitmap_info); |
1a4d82fc JJ |
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 + | |
970d7e83 LB |
315 | (uintptr_t)(bin_info->reg_interval * regind)); |
316 | run->nfree--; | |
970d7e83 LB |
317 | return (ret); |
318 | } | |
319 | ||
54a0048b | 320 | JEMALLOC_INLINE_C void |
970d7e83 LB |
321 | arena_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); | |
54a0048b | 326 | szind_t binind = arena_ptr_small_binind_get(ptr, mapbits); |
970d7e83 | 327 | arena_bin_info_t *bin_info = &arena_bin_info[binind]; |
54a0048b | 328 | size_t regind = arena_run_regind(run, bin_info, ptr); |
970d7e83 LB |
329 | |
330 | assert(run->nfree < bin_info->nregs); | |
331 | /* Freeing an interior pointer can cause assertion failure. */ | |
1a4d82fc JJ |
332 | assert(((uintptr_t)ptr - |
333 | ((uintptr_t)arena_miscelm_to_rpages(arena_run_to_miscelm(run)) + | |
970d7e83 LB |
334 | (uintptr_t)bin_info->reg0_offset)) % |
335 | (uintptr_t)bin_info->reg_interval == 0); | |
1a4d82fc JJ |
336 | assert((uintptr_t)ptr >= |
337 | (uintptr_t)arena_miscelm_to_rpages(arena_run_to_miscelm(run)) + | |
970d7e83 LB |
338 | (uintptr_t)bin_info->reg0_offset); |
339 | /* Freeing an unallocated pointer can cause assertion failure. */ | |
1a4d82fc | 340 | assert(bitmap_get(run->bitmap, &bin_info->bitmap_info, regind)); |
970d7e83 | 341 | |
1a4d82fc | 342 | bitmap_unset(run->bitmap, &bin_info->bitmap_info, regind); |
970d7e83 LB |
343 | run->nfree++; |
344 | } | |
345 | ||
54a0048b | 346 | JEMALLOC_INLINE_C void |
970d7e83 LB |
347 | arena_run_zero(arena_chunk_t *chunk, size_t run_ind, size_t npages) |
348 | { | |
349 | ||
1a4d82fc JJ |
350 | JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED((void *)((uintptr_t)chunk + |
351 | (run_ind << LG_PAGE)), (npages << LG_PAGE)); | |
970d7e83 LB |
352 | memset((void *)((uintptr_t)chunk + (run_ind << LG_PAGE)), 0, |
353 | (npages << LG_PAGE)); | |
354 | } | |
355 | ||
54a0048b | 356 | JEMALLOC_INLINE_C void |
1a4d82fc JJ |
357 | arena_run_page_mark_zeroed(arena_chunk_t *chunk, size_t run_ind) |
358 | { | |
359 | ||
360 | JEMALLOC_VALGRIND_MAKE_MEM_DEFINED((void *)((uintptr_t)chunk + (run_ind | |
361 | << LG_PAGE)), PAGE); | |
362 | } | |
363 | ||
54a0048b | 364 | JEMALLOC_INLINE_C void |
970d7e83 LB |
365 | arena_run_page_validate_zeroed(arena_chunk_t *chunk, size_t run_ind) |
366 | { | |
367 | size_t i; | |
368 | UNUSED size_t *p = (size_t *)((uintptr_t)chunk + (run_ind << LG_PAGE)); | |
369 | ||
1a4d82fc | 370 | arena_run_page_mark_zeroed(chunk, run_ind); |
970d7e83 LB |
371 | for (i = 0; i < PAGE / sizeof(size_t); i++) |
372 | assert(p[i] == 0); | |
373 | } | |
374 | ||
375 | static void | |
54a0048b | 376 | arena_nactive_add(arena_t *arena, size_t add_pages) |
970d7e83 | 377 | { |
970d7e83 | 378 | |
1a4d82fc | 379 | if (config_stats) { |
54a0048b SL |
380 | size_t cactive_add = CHUNK_CEILING((arena->nactive + |
381 | add_pages) << LG_PAGE) - CHUNK_CEILING(arena->nactive << | |
1a4d82fc | 382 | LG_PAGE); |
54a0048b SL |
383 | if (cactive_add != 0) |
384 | stats_cactive_add(cactive_add); | |
385 | } | |
386 | arena->nactive += add_pages; | |
387 | } | |
388 | ||
389 | static void | |
390 | arena_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); | |
1a4d82fc | 398 | } |
54a0048b | 399 | arena->nactive -= sub_pages; |
1a4d82fc JJ |
400 | } |
401 | ||
402 | static void | |
403 | arena_run_split_remove(arena_t *arena, arena_chunk_t *chunk, size_t run_ind, | |
54a0048b | 404 | size_t flag_dirty, size_t flag_decommitted, size_t need_pages) |
1a4d82fc JJ |
405 | { |
406 | size_t total_pages, rem_pages; | |
970d7e83 | 407 | |
54a0048b SL |
408 | assert(flag_dirty == 0 || flag_decommitted == 0); |
409 | ||
970d7e83 LB |
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); | |
970d7e83 LB |
414 | assert(need_pages <= total_pages); |
415 | rem_pages = total_pages - need_pages; | |
416 | ||
1a4d82fc JJ |
417 | arena_avail_remove(arena, chunk, run_ind, total_pages); |
418 | if (flag_dirty != 0) | |
54a0048b SL |
419 | arena_run_dirty_remove(arena, chunk, run_ind, total_pages); |
420 | arena_nactive_add(arena, need_pages); | |
970d7e83 LB |
421 | |
422 | /* Keep track of trailing unused pages for later use. */ | |
423 | if (rem_pages > 0) { | |
54a0048b SL |
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, | |
429 | (rem_pages << LG_PAGE), flags | | |
430 | (arena_mapbits_unzeroed_get(chunk, run_ind+need_pages) & | |
431 | flag_unzeroed_mask)); | |
432 | arena_mapbits_unallocated_set(chunk, run_ind+total_pages-1, | |
433 | (rem_pages << LG_PAGE), flags | | |
434 | (arena_mapbits_unzeroed_get(chunk, run_ind+total_pages-1) & | |
435 | flag_unzeroed_mask)); | |
970d7e83 | 436 | if (flag_dirty != 0) { |
54a0048b | 437 | arena_run_dirty_insert(arena, chunk, run_ind+need_pages, |
1a4d82fc | 438 | rem_pages); |
970d7e83 | 439 | } |
1a4d82fc | 440 | arena_avail_insert(arena, chunk, run_ind+need_pages, rem_pages); |
970d7e83 | 441 | } |
1a4d82fc | 442 | } |
970d7e83 | 443 | |
54a0048b | 444 | static bool |
1a4d82fc JJ |
445 | arena_run_split_large_helper(arena_t *arena, arena_run_t *run, size_t size, |
446 | bool remove, bool zero) | |
447 | { | |
448 | arena_chunk_t *chunk; | |
449 | arena_chunk_map_misc_t *miscelm; | |
54a0048b SL |
450 | size_t flag_dirty, flag_decommitted, run_ind, need_pages; |
451 | size_t flag_unzeroed_mask; | |
1a4d82fc JJ |
452 | |
453 | chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run); | |
454 | miscelm = arena_run_to_miscelm(run); | |
455 | run_ind = arena_miscelm_to_pageind(miscelm); | |
456 | flag_dirty = arena_mapbits_dirty_get(chunk, run_ind); | |
54a0048b | 457 | flag_decommitted = arena_mapbits_decommitted_get(chunk, run_ind); |
1a4d82fc JJ |
458 | need_pages = (size >> LG_PAGE); |
459 | assert(need_pages > 0); | |
460 | ||
54a0048b SL |
461 | if (flag_decommitted != 0 && arena->chunk_hooks.commit(chunk, chunksize, |
462 | run_ind << LG_PAGE, size, arena->ind)) | |
463 | return (true); | |
464 | ||
1a4d82fc JJ |
465 | if (remove) { |
466 | arena_run_split_remove(arena, chunk, run_ind, flag_dirty, | |
54a0048b | 467 | flag_decommitted, need_pages); |
1a4d82fc JJ |
468 | } |
469 | ||
470 | if (zero) { | |
54a0048b SL |
471 | if (flag_decommitted != 0) { |
472 | /* The run is untouched, and therefore zeroed. */ | |
473 | JEMALLOC_VALGRIND_MAKE_MEM_DEFINED((void | |
474 | *)((uintptr_t)chunk + (run_ind << LG_PAGE)), | |
475 | (need_pages << LG_PAGE)); | |
476 | } else if (flag_dirty != 0) { | |
477 | /* The run is dirty, so all pages must be zeroed. */ | |
478 | arena_run_zero(chunk, run_ind, need_pages); | |
479 | } else { | |
1a4d82fc JJ |
480 | /* |
481 | * The run is clean, so some pages may be zeroed (i.e. | |
482 | * never before touched). | |
483 | */ | |
54a0048b | 484 | size_t i; |
1a4d82fc JJ |
485 | for (i = 0; i < need_pages; i++) { |
486 | if (arena_mapbits_unzeroed_get(chunk, run_ind+i) | |
487 | != 0) | |
488 | arena_run_zero(chunk, run_ind+i, 1); | |
489 | else if (config_debug) { | |
490 | arena_run_page_validate_zeroed(chunk, | |
491 | run_ind+i); | |
492 | } else { | |
493 | arena_run_page_mark_zeroed(chunk, | |
494 | run_ind+i); | |
970d7e83 | 495 | } |
970d7e83 LB |
496 | } |
497 | } | |
970d7e83 | 498 | } else { |
1a4d82fc JJ |
499 | JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED((void *)((uintptr_t)chunk + |
500 | (run_ind << LG_PAGE)), (need_pages << LG_PAGE)); | |
970d7e83 | 501 | } |
1a4d82fc JJ |
502 | |
503 | /* | |
504 | * Set the last element first, in case the run only contains one page | |
505 | * (i.e. both statements set the same element). | |
506 | */ | |
54a0048b SL |
507 | flag_unzeroed_mask = (flag_dirty | flag_decommitted) == 0 ? |
508 | CHUNK_MAP_UNZEROED : 0; | |
509 | arena_mapbits_large_set(chunk, run_ind+need_pages-1, 0, flag_dirty | | |
510 | (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk, | |
511 | run_ind+need_pages-1))); | |
512 | arena_mapbits_large_set(chunk, run_ind, size, flag_dirty | | |
513 | (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk, run_ind))); | |
514 | return (false); | |
1a4d82fc JJ |
515 | } |
516 | ||
54a0048b | 517 | static bool |
1a4d82fc JJ |
518 | arena_run_split_large(arena_t *arena, arena_run_t *run, size_t size, bool zero) |
519 | { | |
520 | ||
54a0048b | 521 | return (arena_run_split_large_helper(arena, run, size, true, zero)); |
1a4d82fc JJ |
522 | } |
523 | ||
54a0048b | 524 | static bool |
1a4d82fc JJ |
525 | arena_run_init_large(arena_t *arena, arena_run_t *run, size_t size, bool zero) |
526 | { | |
527 | ||
54a0048b | 528 | return (arena_run_split_large_helper(arena, run, size, false, zero)); |
1a4d82fc JJ |
529 | } |
530 | ||
54a0048b | 531 | static bool |
1a4d82fc | 532 | arena_run_split_small(arena_t *arena, arena_run_t *run, size_t size, |
54a0048b | 533 | szind_t binind) |
1a4d82fc JJ |
534 | { |
535 | arena_chunk_t *chunk; | |
536 | arena_chunk_map_misc_t *miscelm; | |
54a0048b | 537 | size_t flag_dirty, flag_decommitted, run_ind, need_pages, i; |
1a4d82fc JJ |
538 | |
539 | assert(binind != BININD_INVALID); | |
540 | ||
541 | chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run); | |
542 | miscelm = arena_run_to_miscelm(run); | |
543 | run_ind = arena_miscelm_to_pageind(miscelm); | |
544 | flag_dirty = arena_mapbits_dirty_get(chunk, run_ind); | |
54a0048b | 545 | flag_decommitted = arena_mapbits_decommitted_get(chunk, run_ind); |
1a4d82fc JJ |
546 | need_pages = (size >> LG_PAGE); |
547 | assert(need_pages > 0); | |
548 | ||
54a0048b SL |
549 | if (flag_decommitted != 0 && arena->chunk_hooks.commit(chunk, chunksize, |
550 | run_ind << LG_PAGE, size, arena->ind)) | |
551 | return (true); | |
9cc50fc6 | 552 | |
54a0048b SL |
553 | arena_run_split_remove(arena, chunk, run_ind, flag_dirty, |
554 | flag_decommitted, need_pages); | |
555 | ||
556 | for (i = 0; i < need_pages; i++) { | |
557 | size_t flag_unzeroed = arena_mapbits_unzeroed_get(chunk, | |
558 | run_ind+i); | |
559 | arena_mapbits_small_set(chunk, run_ind+i, i, binind, | |
560 | flag_unzeroed); | |
561 | if (config_debug && flag_dirty == 0 && flag_unzeroed == 0) | |
1a4d82fc JJ |
562 | arena_run_page_validate_zeroed(chunk, run_ind+i); |
563 | } | |
1a4d82fc JJ |
564 | JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED((void *)((uintptr_t)chunk + |
565 | (run_ind << LG_PAGE)), (need_pages << LG_PAGE)); | |
54a0048b | 566 | return (false); |
970d7e83 LB |
567 | } |
568 | ||
569 | static arena_chunk_t * | |
1a4d82fc | 570 | arena_chunk_init_spare(arena_t *arena) |
970d7e83 LB |
571 | { |
572 | arena_chunk_t *chunk; | |
970d7e83 | 573 | |
1a4d82fc | 574 | assert(arena->spare != NULL); |
970d7e83 | 575 | |
1a4d82fc JJ |
576 | chunk = arena->spare; |
577 | arena->spare = NULL; | |
970d7e83 | 578 | |
1a4d82fc JJ |
579 | assert(arena_mapbits_allocated_get(chunk, map_bias) == 0); |
580 | assert(arena_mapbits_allocated_get(chunk, chunk_npages-1) == 0); | |
581 | assert(arena_mapbits_unallocated_size_get(chunk, map_bias) == | |
54a0048b | 582 | arena_maxrun); |
1a4d82fc | 583 | assert(arena_mapbits_unallocated_size_get(chunk, chunk_npages-1) == |
54a0048b | 584 | arena_maxrun); |
1a4d82fc JJ |
585 | assert(arena_mapbits_dirty_get(chunk, map_bias) == |
586 | arena_mapbits_dirty_get(chunk, chunk_npages-1)); | |
970d7e83 | 587 | |
1a4d82fc JJ |
588 | return (chunk); |
589 | } | |
970d7e83 | 590 | |
54a0048b SL |
591 | static bool |
592 | arena_chunk_register(arena_t *arena, arena_chunk_t *chunk, bool zero) | |
593 | { | |
594 | ||
595 | /* | |
596 | * The extent node notion of "committed" doesn't directly apply to | |
597 | * arena chunks. Arbitrarily mark them as committed. The commit state | |
598 | * of runs is tracked individually, and upon chunk deallocation the | |
599 | * entire chunk is in a consistent commit state. | |
600 | */ | |
601 | extent_node_init(&chunk->node, arena, chunk, chunksize, zero, true); | |
602 | extent_node_achunk_set(&chunk->node, true); | |
603 | return (chunk_register(chunk, &chunk->node)); | |
604 | } | |
605 | ||
1a4d82fc | 606 | static arena_chunk_t * |
54a0048b SL |
607 | arena_chunk_alloc_internal_hard(arena_t *arena, chunk_hooks_t *chunk_hooks, |
608 | bool *zero, bool *commit) | |
1a4d82fc JJ |
609 | { |
610 | arena_chunk_t *chunk; | |
970d7e83 | 611 | |
1a4d82fc | 612 | malloc_mutex_unlock(&arena->lock); |
7453a54e | 613 | |
54a0048b SL |
614 | chunk = (arena_chunk_t *)chunk_alloc_wrapper(arena, chunk_hooks, NULL, |
615 | chunksize, chunksize, zero, commit); | |
616 | if (chunk != NULL && !*commit) { | |
617 | /* Commit header. */ | |
618 | if (chunk_hooks->commit(chunk, chunksize, 0, map_bias << | |
619 | LG_PAGE, arena->ind)) { | |
620 | chunk_dalloc_wrapper(arena, chunk_hooks, | |
621 | (void *)chunk, chunksize, *commit); | |
622 | chunk = NULL; | |
623 | } | |
624 | } | |
625 | if (chunk != NULL && arena_chunk_register(arena, chunk, *zero)) { | |
626 | if (!*commit) { | |
627 | /* Undo commit of header. */ | |
628 | chunk_hooks->decommit(chunk, chunksize, 0, map_bias << | |
629 | LG_PAGE, arena->ind); | |
630 | } | |
631 | chunk_dalloc_wrapper(arena, chunk_hooks, (void *)chunk, | |
632 | chunksize, *commit); | |
633 | chunk = NULL; | |
634 | } | |
635 | ||
636 | malloc_mutex_lock(&arena->lock); | |
1a4d82fc JJ |
637 | return (chunk); |
638 | } | |
639 | ||
54a0048b SL |
640 | static arena_chunk_t * |
641 | arena_chunk_alloc_internal(arena_t *arena, bool *zero, bool *commit) | |
1a4d82fc | 642 | { |
54a0048b SL |
643 | arena_chunk_t *chunk; |
644 | chunk_hooks_t chunk_hooks = CHUNK_HOOKS_INITIALIZER; | |
1a4d82fc | 645 | |
54a0048b SL |
646 | chunk = chunk_alloc_cache(arena, &chunk_hooks, NULL, chunksize, |
647 | chunksize, zero, true); | |
648 | if (chunk != NULL) { | |
649 | if (arena_chunk_register(arena, chunk, *zero)) { | |
650 | chunk_dalloc_cache(arena, &chunk_hooks, chunk, | |
651 | chunksize, true); | |
652 | return (NULL); | |
653 | } | |
654 | *commit = true; | |
655 | } | |
656 | if (chunk == NULL) { | |
657 | chunk = arena_chunk_alloc_internal_hard(arena, &chunk_hooks, | |
658 | zero, commit); | |
1a4d82fc | 659 | } |
1a4d82fc | 660 | |
54a0048b SL |
661 | if (config_stats && chunk != NULL) { |
662 | arena->stats.mapped += chunksize; | |
663 | arena->stats.metadata_mapped += (map_bias << LG_PAGE); | |
1a4d82fc JJ |
664 | } |
665 | ||
54a0048b | 666 | return (chunk); |
1a4d82fc JJ |
667 | } |
668 | ||
669 | static arena_chunk_t * | |
670 | arena_chunk_init_hard(arena_t *arena) | |
671 | { | |
672 | arena_chunk_t *chunk; | |
54a0048b SL |
673 | bool zero, commit; |
674 | size_t flag_unzeroed, flag_decommitted, i; | |
1a4d82fc JJ |
675 | |
676 | assert(arena->spare == NULL); | |
677 | ||
678 | zero = false; | |
54a0048b SL |
679 | commit = false; |
680 | chunk = arena_chunk_alloc_internal(arena, &zero, &commit); | |
1a4d82fc JJ |
681 | if (chunk == NULL) |
682 | return (NULL); | |
683 | ||
1a4d82fc JJ |
684 | /* |
685 | * Initialize the map to contain one maximal free untouched run. Mark | |
54a0048b SL |
686 | * the pages as zeroed if chunk_alloc() returned a zeroed or decommitted |
687 | * chunk. | |
1a4d82fc | 688 | */ |
54a0048b SL |
689 | flag_unzeroed = (zero || !commit) ? 0 : CHUNK_MAP_UNZEROED; |
690 | flag_decommitted = commit ? 0 : CHUNK_MAP_DECOMMITTED; | |
691 | arena_mapbits_unallocated_set(chunk, map_bias, arena_maxrun, | |
692 | flag_unzeroed | flag_decommitted); | |
1a4d82fc JJ |
693 | /* |
694 | * There is no need to initialize the internal page map entries unless | |
695 | * the chunk is not zeroed. | |
696 | */ | |
697 | if (!zero) { | |
698 | JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED( | |
699 | (void *)arena_bitselm_get(chunk, map_bias+1), | |
700 | (size_t)((uintptr_t) arena_bitselm_get(chunk, | |
701 | chunk_npages-1) - (uintptr_t)arena_bitselm_get(chunk, | |
702 | map_bias+1))); | |
703 | for (i = map_bias+1; i < chunk_npages-1; i++) | |
54a0048b | 704 | arena_mapbits_internal_set(chunk, i, flag_unzeroed); |
1a4d82fc JJ |
705 | } else { |
706 | JEMALLOC_VALGRIND_MAKE_MEM_DEFINED((void | |
707 | *)arena_bitselm_get(chunk, map_bias+1), (size_t)((uintptr_t) | |
708 | arena_bitselm_get(chunk, chunk_npages-1) - | |
709 | (uintptr_t)arena_bitselm_get(chunk, map_bias+1))); | |
710 | if (config_debug) { | |
970d7e83 LB |
711 | for (i = map_bias+1; i < chunk_npages-1; i++) { |
712 | assert(arena_mapbits_unzeroed_get(chunk, i) == | |
54a0048b | 713 | flag_unzeroed); |
970d7e83 LB |
714 | } |
715 | } | |
1a4d82fc | 716 | } |
54a0048b SL |
717 | arena_mapbits_unallocated_set(chunk, chunk_npages-1, arena_maxrun, |
718 | flag_unzeroed); | |
1a4d82fc JJ |
719 | |
720 | return (chunk); | |
721 | } | |
722 | ||
723 | static arena_chunk_t * | |
724 | arena_chunk_alloc(arena_t *arena) | |
725 | { | |
726 | arena_chunk_t *chunk; | |
727 | ||
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); | |
970d7e83 LB |
734 | } |
735 | ||
1a4d82fc | 736 | arena_avail_insert(arena, chunk, map_bias, chunk_npages-map_bias); |
970d7e83 LB |
737 | |
738 | return (chunk); | |
739 | } | |
740 | ||
1a4d82fc JJ |
741 | static void |
742 | arena_chunk_dalloc(arena_t *arena, arena_chunk_t *chunk) | |
970d7e83 | 743 | { |
1a4d82fc | 744 | |
970d7e83 LB |
745 | assert(arena_mapbits_allocated_get(chunk, map_bias) == 0); |
746 | assert(arena_mapbits_allocated_get(chunk, chunk_npages-1) == 0); | |
747 | assert(arena_mapbits_unallocated_size_get(chunk, map_bias) == | |
54a0048b | 748 | arena_maxrun); |
970d7e83 | 749 | assert(arena_mapbits_unallocated_size_get(chunk, chunk_npages-1) == |
54a0048b | 750 | arena_maxrun); |
970d7e83 LB |
751 | assert(arena_mapbits_dirty_get(chunk, map_bias) == |
752 | arena_mapbits_dirty_get(chunk, chunk_npages-1)); | |
54a0048b SL |
753 | assert(arena_mapbits_decommitted_get(chunk, map_bias) == |
754 | arena_mapbits_decommitted_get(chunk, chunk_npages-1)); | |
970d7e83 | 755 | |
54a0048b | 756 | /* Remove run from runs_avail, so that the arena does not use it. */ |
1a4d82fc | 757 | arena_avail_remove(arena, chunk, map_bias, chunk_npages-map_bias); |
970d7e83 LB |
758 | |
759 | if (arena->spare != NULL) { | |
760 | arena_chunk_t *spare = arena->spare; | |
54a0048b SL |
761 | chunk_hooks_t chunk_hooks = CHUNK_HOOKS_INITIALIZER; |
762 | bool committed; | |
970d7e83 LB |
763 | |
764 | arena->spare = chunk; | |
1a4d82fc | 765 | if (arena_mapbits_dirty_get(spare, map_bias) != 0) { |
54a0048b | 766 | arena_run_dirty_remove(arena, spare, map_bias, |
1a4d82fc JJ |
767 | chunk_npages-map_bias); |
768 | } | |
54a0048b SL |
769 | |
770 | chunk_deregister(spare, &spare->node); | |
771 | ||
772 | committed = (arena_mapbits_decommitted_get(spare, map_bias) == | |
773 | 0); | |
774 | if (!committed) { | |
775 | /* | |
776 | * Decommit the header. Mark the chunk as decommitted | |
777 | * even if header decommit fails, since treating a | |
778 | * partially committed chunk as committed has a high | |
779 | * potential for causing later access of decommitted | |
780 | * memory. | |
781 | */ | |
782 | chunk_hooks = chunk_hooks_get(arena); | |
783 | chunk_hooks.decommit(spare, chunksize, 0, map_bias << | |
784 | LG_PAGE, arena->ind); | |
785 | } | |
786 | ||
787 | chunk_dalloc_cache(arena, &chunk_hooks, (void *)spare, | |
788 | chunksize, committed); | |
789 | ||
790 | if (config_stats) { | |
791 | arena->stats.mapped -= chunksize; | |
792 | arena->stats.metadata_mapped -= (map_bias << LG_PAGE); | |
793 | } | |
970d7e83 LB |
794 | } else |
795 | arena->spare = chunk; | |
796 | } | |
797 | ||
54a0048b SL |
798 | static void |
799 | arena_huge_malloc_stats_update(arena_t *arena, size_t usize) | |
800 | { | |
801 | szind_t index = size2index(usize) - nlclasses - NBINS; | |
802 | ||
803 | cassert(config_stats); | |
804 | ||
805 | arena->stats.nmalloc_huge++; | |
806 | arena->stats.allocated_huge += usize; | |
807 | arena->stats.hstats[index].nmalloc++; | |
808 | arena->stats.hstats[index].curhchunks++; | |
809 | } | |
810 | ||
811 | static void | |
812 | arena_huge_malloc_stats_update_undo(arena_t *arena, size_t usize) | |
813 | { | |
814 | szind_t index = size2index(usize) - nlclasses - NBINS; | |
815 | ||
816 | cassert(config_stats); | |
817 | ||
818 | arena->stats.nmalloc_huge--; | |
819 | arena->stats.allocated_huge -= usize; | |
820 | arena->stats.hstats[index].nmalloc--; | |
821 | arena->stats.hstats[index].curhchunks--; | |
822 | } | |
823 | ||
824 | static void | |
825 | arena_huge_dalloc_stats_update(arena_t *arena, size_t usize) | |
826 | { | |
827 | szind_t index = size2index(usize) - nlclasses - NBINS; | |
828 | ||
829 | cassert(config_stats); | |
830 | ||
831 | arena->stats.ndalloc_huge++; | |
832 | arena->stats.allocated_huge -= usize; | |
833 | arena->stats.hstats[index].ndalloc++; | |
834 | arena->stats.hstats[index].curhchunks--; | |
835 | } | |
836 | ||
837 | static void | |
838 | arena_huge_dalloc_stats_update_undo(arena_t *arena, size_t usize) | |
839 | { | |
840 | szind_t index = size2index(usize) - nlclasses - NBINS; | |
841 | ||
842 | cassert(config_stats); | |
843 | ||
844 | arena->stats.ndalloc_huge--; | |
845 | arena->stats.allocated_huge += usize; | |
846 | arena->stats.hstats[index].ndalloc--; | |
847 | arena->stats.hstats[index].curhchunks++; | |
848 | } | |
849 | ||
850 | static void | |
851 | arena_huge_ralloc_stats_update(arena_t *arena, size_t oldsize, size_t usize) | |
852 | { | |
853 | ||
854 | arena_huge_dalloc_stats_update(arena, oldsize); | |
855 | arena_huge_malloc_stats_update(arena, usize); | |
856 | } | |
857 | ||
858 | static void | |
859 | arena_huge_ralloc_stats_update_undo(arena_t *arena, size_t oldsize, | |
860 | size_t usize) | |
861 | { | |
862 | ||
863 | arena_huge_dalloc_stats_update_undo(arena, oldsize); | |
864 | arena_huge_malloc_stats_update_undo(arena, usize); | |
865 | } | |
866 | ||
867 | extent_node_t * | |
868 | arena_node_alloc(arena_t *arena) | |
869 | { | |
870 | extent_node_t *node; | |
871 | ||
872 | malloc_mutex_lock(&arena->node_cache_mtx); | |
873 | node = ql_last(&arena->node_cache, ql_link); | |
874 | if (node == NULL) { | |
875 | malloc_mutex_unlock(&arena->node_cache_mtx); | |
876 | return (base_alloc(sizeof(extent_node_t))); | |
877 | } | |
878 | ql_tail_remove(&arena->node_cache, extent_node_t, ql_link); | |
879 | malloc_mutex_unlock(&arena->node_cache_mtx); | |
880 | return (node); | |
881 | } | |
882 | ||
883 | void | |
884 | arena_node_dalloc(arena_t *arena, extent_node_t *node) | |
885 | { | |
886 | ||
887 | malloc_mutex_lock(&arena->node_cache_mtx); | |
888 | ql_elm_new(node, ql_link); | |
889 | ql_tail_insert(&arena->node_cache, node, ql_link); | |
890 | malloc_mutex_unlock(&arena->node_cache_mtx); | |
891 | } | |
892 | ||
893 | static void * | |
894 | arena_chunk_alloc_huge_hard(arena_t *arena, chunk_hooks_t *chunk_hooks, | |
895 | size_t usize, size_t alignment, bool *zero, size_t csize) | |
896 | { | |
897 | void *ret; | |
898 | bool commit = true; | |
899 | ||
900 | ret = chunk_alloc_wrapper(arena, chunk_hooks, NULL, csize, alignment, | |
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 | ||
916 | void * | |
917 | arena_chunk_alloc_huge(arena_t *arena, size_t usize, size_t alignment, | |
918 | bool *zero) | |
919 | { | |
920 | void *ret; | |
921 | chunk_hooks_t chunk_hooks = CHUNK_HOOKS_INITIALIZER; | |
922 | size_t csize = CHUNK_CEILING(usize); | |
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 | ||
944 | void | |
945 | arena_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 | ||
962 | void | |
963 | arena_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 | ||
980 | void | |
981 | arena_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 | } | |
1002 | malloc_mutex_unlock(&arena->lock); | |
1003 | } | |
1004 | ||
1005 | static bool | |
1006 | arena_chunk_ralloc_huge_expand_hard(arena_t *arena, chunk_hooks_t *chunk_hooks, | |
1007 | void *chunk, size_t oldsize, size_t usize, bool *zero, void *nchunk, | |
1008 | size_t udiff, size_t cdiff) | |
1009 | { | |
1010 | bool err; | |
1011 | bool commit = true; | |
1012 | ||
1013 | err = (chunk_alloc_wrapper(arena, chunk_hooks, nchunk, cdiff, chunksize, | |
1014 | zero, &commit) == NULL); | |
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); | |
1032 | } | |
1033 | ||
1034 | bool | |
1035 | arena_chunk_ralloc_huge_expand(arena_t *arena, void *chunk, size_t oldsize, | |
1036 | size_t usize, bool *zero) | |
1037 | { | |
1038 | bool err; | |
1039 | chunk_hooks_t chunk_hooks = chunk_hooks_get(arena); | |
1040 | void *nchunk = (void *)((uintptr_t)chunk + CHUNK_CEILING(oldsize)); | |
1041 | size_t udiff = usize - oldsize; | |
1042 | size_t cdiff = CHUNK_CEILING(usize) - CHUNK_CEILING(oldsize); | |
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 | */ | |
9cc50fc6 | 1075 | static arena_run_t * |
54a0048b | 1076 | arena_run_first_best_fit(arena_t *arena, size_t size) |
9cc50fc6 | 1077 | { |
54a0048b | 1078 | szind_t ind, i; |
7453a54e | 1079 | |
54a0048b SL |
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); | |
9cc50fc6 | 1086 | } |
7453a54e SL |
1087 | |
1088 | return (NULL); | |
1a4d82fc JJ |
1089 | } |
1090 | ||
54a0048b SL |
1091 | static arena_run_t * |
1092 | arena_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)) | |
1097 | run = NULL; | |
1098 | } | |
1099 | return (run); | |
1100 | } | |
1101 | ||
1a4d82fc JJ |
1102 | static arena_run_t * |
1103 | arena_run_alloc_large(arena_t *arena, size_t size, bool zero) | |
970d7e83 | 1104 | { |
1a4d82fc | 1105 | arena_chunk_t *chunk; |
970d7e83 | 1106 | arena_run_t *run; |
970d7e83 | 1107 | |
54a0048b SL |
1108 | assert(size <= arena_maxrun); |
1109 | assert(size == PAGE_CEILING(size)); | |
970d7e83 | 1110 | |
1a4d82fc JJ |
1111 | /* Search the arena's chunks for the lowest best fit. */ |
1112 | run = arena_run_alloc_large_helper(arena, size, zero); | |
1113 | if (run != NULL) | |
1114 | return (run); | |
1115 | ||
1116 | /* | |
1117 | * No usable runs. Create a new chunk from which to allocate the run. | |
1118 | */ | |
1119 | chunk = arena_chunk_alloc(arena); | |
1120 | if (chunk != NULL) { | |
1121 | run = &arena_miscelm_get(chunk, map_bias)->run; | |
54a0048b SL |
1122 | if (arena_run_split_large(arena, run, size, zero)) |
1123 | run = NULL; | |
1a4d82fc JJ |
1124 | return (run); |
1125 | } | |
1126 | ||
1127 | /* | |
1128 | * arena_chunk_alloc() failed, but another thread may have made | |
1129 | * sufficient memory available while this one dropped arena->lock in | |
1130 | * arena_chunk_alloc(), so search one more time. | |
1131 | */ | |
1132 | return (arena_run_alloc_large_helper(arena, size, zero)); | |
1133 | } | |
1134 | ||
1135 | static arena_run_t * | |
54a0048b | 1136 | arena_run_alloc_small_helper(arena_t *arena, size_t size, szind_t binind) |
1a4d82fc | 1137 | { |
54a0048b SL |
1138 | arena_run_t *run = arena_run_first_best_fit(arena, size); |
1139 | if (run != NULL) { | |
1140 | if (arena_run_split_small(arena, run, size, binind)) | |
1141 | run = NULL; | |
970d7e83 | 1142 | } |
54a0048b | 1143 | return (run); |
970d7e83 LB |
1144 | } |
1145 | ||
1146 | static arena_run_t * | |
54a0048b | 1147 | arena_run_alloc_small(arena_t *arena, size_t size, szind_t binind) |
970d7e83 LB |
1148 | { |
1149 | arena_chunk_t *chunk; | |
1150 | arena_run_t *run; | |
1151 | ||
54a0048b SL |
1152 | assert(size <= arena_maxrun); |
1153 | assert(size == PAGE_CEILING(size)); | |
1a4d82fc | 1154 | assert(binind != BININD_INVALID); |
970d7e83 LB |
1155 | |
1156 | /* Search the arena's chunks for the lowest best fit. */ | |
1a4d82fc | 1157 | run = arena_run_alloc_small_helper(arena, size, binind); |
970d7e83 LB |
1158 | if (run != NULL) |
1159 | return (run); | |
1160 | ||
1161 | /* | |
1162 | * No usable runs. Create a new chunk from which to allocate the run. | |
1163 | */ | |
1164 | chunk = arena_chunk_alloc(arena); | |
1165 | if (chunk != NULL) { | |
1a4d82fc | 1166 | run = &arena_miscelm_get(chunk, map_bias)->run; |
54a0048b SL |
1167 | if (arena_run_split_small(arena, run, size, binind)) |
1168 | run = NULL; | |
970d7e83 LB |
1169 | return (run); |
1170 | } | |
1171 | ||
1172 | /* | |
1173 | * arena_chunk_alloc() failed, but another thread may have made | |
1174 | * sufficient memory available while this one dropped arena->lock in | |
1175 | * arena_chunk_alloc(), so search one more time. | |
1176 | */ | |
1a4d82fc | 1177 | return (arena_run_alloc_small_helper(arena, size, binind)); |
970d7e83 LB |
1178 | } |
1179 | ||
54a0048b SL |
1180 | static bool |
1181 | arena_lg_dirty_mult_valid(ssize_t lg_dirty_mult) | |
970d7e83 | 1182 | { |
970d7e83 | 1183 | |
54a0048b SL |
1184 | return (lg_dirty_mult >= -1 && lg_dirty_mult < (ssize_t)(sizeof(size_t) |
1185 | << 3)); | |
1186 | } | |
1187 | ||
1188 | ssize_t | |
1189 | arena_lg_dirty_mult_get(arena_t *arena) | |
1190 | { | |
1191 | ssize_t lg_dirty_mult; | |
1192 | ||
1193 | malloc_mutex_lock(&arena->lock); | |
1194 | lg_dirty_mult = arena->lg_dirty_mult; | |
1195 | malloc_mutex_unlock(&arena->lock); | |
1196 | ||
1197 | return (lg_dirty_mult); | |
1198 | } | |
1199 | ||
1200 | bool | |
1201 | arena_lg_dirty_mult_set(arena_t *arena, ssize_t lg_dirty_mult) | |
1202 | { | |
1203 | ||
1204 | if (!arena_lg_dirty_mult_valid(lg_dirty_mult)) | |
1205 | return (true); | |
1206 | ||
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 | ||
1215 | static void | |
1216 | arena_decay_deadline_init(arena_t *arena) | |
1217 | { | |
1218 | ||
1219 | assert(opt_purge == purge_mode_decay); | |
1220 | ||
970d7e83 | 1221 | /* |
54a0048b SL |
1222 | * Generate a new deadline that is uniformly random within the next |
1223 | * epoch after the current one. | |
970d7e83 | 1224 | */ |
54a0048b SL |
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 | } | |
7453a54e | 1235 | |
54a0048b SL |
1236 | static bool |
1237 | arena_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); | |
970d7e83 LB |
1243 | } |
1244 | ||
1a4d82fc | 1245 | static size_t |
54a0048b SL |
1246 | arena_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 | ||
1273 | static void | |
1274 | arena_decay_epoch_advance(arena_t *arena, const nstime_t *time) | |
970d7e83 | 1275 | { |
54a0048b SL |
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)); | |
1a4d82fc | 1282 | |
54a0048b SL |
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); | |
1a4d82fc JJ |
1314 | } |
1315 | ||
1316 | static size_t | |
54a0048b SL |
1317 | arena_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 | ||
1332 | static void | |
1333 | arena_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 | ||
1351 | static bool | |
1352 | arena_decay_time_valid(ssize_t decay_time) | |
1353 | { | |
1354 | ||
1355 | return (decay_time >= -1 && decay_time <= NSTIME_SEC_MAX); | |
1356 | } | |
1357 | ||
1358 | ssize_t | |
1359 | arena_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 | ||
1370 | bool | |
1371 | arena_decay_time_set(arena_t *arena, ssize_t decay_time) | |
1a4d82fc | 1372 | { |
970d7e83 | 1373 | |
54a0048b SL |
1374 | if (!arena_decay_time_valid(decay_time)) |
1375 | return (true); | |
1376 | ||
1377 | malloc_mutex_lock(&arena->lock); | |
970d7e83 | 1378 | /* |
54a0048b SL |
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. | |
970d7e83 | 1385 | */ |
54a0048b SL |
1386 | arena_decay_init(arena, decay_time); |
1387 | arena_maybe_purge(arena); | |
1388 | malloc_mutex_unlock(&arena->lock); | |
970d7e83 | 1389 | |
54a0048b SL |
1390 | return (false); |
1391 | } | |
1392 | ||
1393 | static void | |
1394 | arena_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 | ||
1421 | static void | |
1422 | arena_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 | ||
1456 | void | |
1457 | arena_maybe_purge(arena_t *arena) | |
1458 | { | |
1459 | ||
1460 | /* Don't recursively purge. */ | |
1461 | if (arena->purging) | |
1462 | return; | |
970d7e83 | 1463 | |
54a0048b SL |
1464 | if (opt_purge == purge_mode_ratio) |
1465 | arena_maybe_purge_ratio(arena); | |
1466 | else | |
1467 | arena_maybe_purge_decay(arena); | |
1a4d82fc | 1468 | } |
970d7e83 | 1469 | |
1a4d82fc | 1470 | static size_t |
54a0048b | 1471 | arena_dirty_count(arena_t *arena) |
1a4d82fc | 1472 | { |
54a0048b SL |
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), | |
1478 | chunkselm = qr_next(&arena->chunks_cache, cc_link); | |
1479 | rdelm != &arena->runs_dirty; rdelm = qr_next(rdelm, rd_link)) { | |
1480 | size_t npages; | |
1481 | ||
1482 | if (rdelm == &chunkselm->rd) { | |
1483 | npages = extent_node_size_get(chunkselm) >> LG_PAGE; | |
1484 | chunkselm = qr_next(chunkselm, cc_link); | |
1485 | } else { | |
1486 | arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE( | |
1487 | rdelm); | |
1488 | arena_chunk_map_misc_t *miscelm = | |
1489 | arena_rd_to_miscelm(rdelm); | |
1490 | size_t pageind = arena_miscelm_to_pageind(miscelm); | |
1491 | assert(arena_mapbits_allocated_get(chunk, pageind) == | |
1492 | 0); | |
1493 | assert(arena_mapbits_large_get(chunk, pageind) == 0); | |
1494 | assert(arena_mapbits_dirty_get(chunk, pageind) != 0); | |
1495 | npages = arena_mapbits_unallocated_size_get(chunk, | |
1496 | pageind) >> LG_PAGE; | |
1497 | } | |
1498 | ndirty += npages; | |
1499 | } | |
1500 | ||
1501 | return (ndirty); | |
1502 | } | |
1503 | ||
1504 | static size_t | |
1505 | arena_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; | |
1a4d82fc JJ |
1511 | size_t nstashed = 0; |
1512 | ||
54a0048b SL |
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); | |
9cc50fc6 | 1519 | |
54a0048b SL |
1520 | if (rdelm == &chunkselm->rd) { |
1521 | extent_node_t *chunkselm_next; | |
1522 | bool zero; | |
1523 | UNUSED void *chunk; | |
9cc50fc6 | 1524 | |
54a0048b SL |
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; | |
1a4d82fc | 1529 | |
54a0048b SL |
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. | |
1569 | */ | |
1570 | if (chunk == arena->spare) | |
1571 | arena_chunk_alloc(arena); | |
1572 | ||
1573 | /* Temporarily allocate the free dirty run. */ | |
1574 | arena_run_split_large(arena, run, run_size, false); | |
1575 | /* Stash. */ | |
1576 | if (false) | |
1577 | qr_new(rdelm, rd_link); /* Redundant. */ | |
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 | } | |
7453a54e | 1584 | |
54a0048b SL |
1585 | nstashed += npages; |
1586 | if (opt_purge == purge_mode_ratio && arena->ndirty - nstashed <= | |
1587 | ndirty_limit) | |
1a4d82fc | 1588 | break; |
970d7e83 | 1589 | } |
970d7e83 | 1590 | |
1a4d82fc JJ |
1591 | return (nstashed); |
1592 | } | |
1593 | ||
1594 | static size_t | |
54a0048b SL |
1595 | arena_purge_stashed(arena_t *arena, chunk_hooks_t *chunk_hooks, |
1596 | arena_runs_dirty_link_t *purge_runs_sentinel, | |
1597 | extent_node_t *purge_chunks_sentinel) | |
1a4d82fc JJ |
1598 | { |
1599 | size_t npurged, nmadvise; | |
54a0048b SL |
1600 | arena_runs_dirty_link_t *rdelm; |
1601 | extent_node_t *chunkselm; | |
1a4d82fc | 1602 | |
970d7e83 LB |
1603 | if (config_stats) |
1604 | nmadvise = 0; | |
1605 | npurged = 0; | |
1a4d82fc JJ |
1606 | |
1607 | malloc_mutex_unlock(&arena->lock); | |
54a0048b SL |
1608 | for (rdelm = qr_next(purge_runs_sentinel, rd_link), |
1609 | chunkselm = qr_next(purge_chunks_sentinel, cc_link); | |
1610 | rdelm != purge_runs_sentinel; rdelm = qr_next(rdelm, rd_link)) { | |
1611 | size_t npages; | |
1a4d82fc | 1612 | |
54a0048b SL |
1613 | if (rdelm == &chunkselm->rd) { |
1614 | /* | |
1615 | * Don't actually purge the chunk here because 1) | |
1616 | * chunkselm is embedded in the chunk and must remain | |
1617 | * valid, and 2) we deallocate the chunk in | |
1618 | * arena_unstash_purged(), where it is destroyed, | |
1619 | * decommitted, or purged, depending on chunk | |
1620 | * deallocation policy. | |
1621 | */ | |
1622 | size_t size = extent_node_size_get(chunkselm); | |
1623 | npages = size >> LG_PAGE; | |
1624 | chunkselm = qr_next(chunkselm, cc_link); | |
1625 | } else { | |
1626 | size_t pageind, run_size, flag_unzeroed, flags, i; | |
1627 | bool decommitted; | |
1628 | arena_chunk_t *chunk = | |
1629 | (arena_chunk_t *)CHUNK_ADDR2BASE(rdelm); | |
1630 | arena_chunk_map_misc_t *miscelm = | |
1631 | arena_rd_to_miscelm(rdelm); | |
1632 | pageind = arena_miscelm_to_pageind(miscelm); | |
1633 | run_size = arena_mapbits_large_size_get(chunk, pageind); | |
1634 | npages = run_size >> LG_PAGE; | |
1635 | ||
1636 | assert(pageind + npages <= chunk_npages); | |
1637 | assert(!arena_mapbits_decommitted_get(chunk, pageind)); | |
1638 | assert(!arena_mapbits_decommitted_get(chunk, | |
1639 | pageind+npages-1)); | |
1640 | decommitted = !chunk_hooks->decommit(chunk, chunksize, | |
1641 | pageind << LG_PAGE, npages << LG_PAGE, arena->ind); | |
1642 | if (decommitted) { | |
1643 | flag_unzeroed = 0; | |
1644 | flags = CHUNK_MAP_DECOMMITTED; | |
1645 | } else { | |
1646 | flag_unzeroed = chunk_purge_wrapper(arena, | |
1647 | chunk_hooks, chunk, chunksize, pageind << | |
1648 | LG_PAGE, run_size) ? CHUNK_MAP_UNZEROED : 0; | |
1649 | flags = flag_unzeroed; | |
1650 | } | |
1651 | arena_mapbits_large_set(chunk, pageind+npages-1, 0, | |
1652 | flags); | |
1653 | arena_mapbits_large_set(chunk, pageind, run_size, | |
1654 | flags); | |
7453a54e | 1655 | |
54a0048b SL |
1656 | /* |
1657 | * Set the unzeroed flag for internal pages, now that | |
1658 | * chunk_purge_wrapper() has returned whether the pages | |
1659 | * were zeroed as a side effect of purging. This chunk | |
1660 | * map modification is safe even though the arena mutex | |
1661 | * isn't currently owned by this thread, because the run | |
1662 | * is marked as allocated, thus protecting it from being | |
1663 | * modified by any other thread. As long as these | |
1664 | * writes don't perturb the first and last elements' | |
1665 | * CHUNK_MAP_ALLOCATED bits, behavior is well defined. | |
1666 | */ | |
1667 | for (i = 1; i < npages-1; i++) { | |
1668 | arena_mapbits_internal_set(chunk, pageind+i, | |
1669 | flag_unzeroed); | |
1670 | } | |
970d7e83 | 1671 | } |
1a4d82fc | 1672 | |
970d7e83 LB |
1673 | npurged += npages; |
1674 | if (config_stats) | |
1675 | nmadvise++; | |
1676 | } | |
1a4d82fc | 1677 | malloc_mutex_lock(&arena->lock); |
970d7e83 | 1678 | |
1a4d82fc JJ |
1679 | if (config_stats) { |
1680 | arena->stats.nmadvise += nmadvise; | |
1681 | arena->stats.purged += npurged; | |
970d7e83 LB |
1682 | } |
1683 | ||
1684 | return (npurged); | |
1685 | } | |
1686 | ||
1a4d82fc | 1687 | static void |
54a0048b SL |
1688 | arena_unstash_purged(arena_t *arena, chunk_hooks_t *chunk_hooks, |
1689 | arena_runs_dirty_link_t *purge_runs_sentinel, | |
1690 | extent_node_t *purge_chunks_sentinel) | |
1691 | { | |
1692 | arena_runs_dirty_link_t *rdelm, *rdelm_next; | |
1693 | extent_node_t *chunkselm; | |
1694 | ||
1695 | /* Deallocate chunks/runs. */ | |
1696 | for (rdelm = qr_next(purge_runs_sentinel, rd_link), | |
1697 | chunkselm = qr_next(purge_chunks_sentinel, cc_link); | |
1698 | rdelm != purge_runs_sentinel; rdelm = rdelm_next) { | |
1699 | rdelm_next = qr_next(rdelm, rd_link); | |
1700 | if (rdelm == &chunkselm->rd) { | |
1701 | extent_node_t *chunkselm_next = qr_next(chunkselm, | |
1702 | cc_link); | |
1703 | void *addr = extent_node_addr_get(chunkselm); | |
1704 | size_t size = extent_node_size_get(chunkselm); | |
1705 | bool zeroed = extent_node_zeroed_get(chunkselm); | |
1706 | bool committed = extent_node_committed_get(chunkselm); | |
1707 | extent_node_dirty_remove(chunkselm); | |
1708 | arena_node_dalloc(arena, chunkselm); | |
1709 | chunkselm = chunkselm_next; | |
1710 | chunk_dalloc_arena(arena, chunk_hooks, addr, size, | |
1711 | zeroed, committed); | |
1712 | } else { | |
1713 | arena_chunk_t *chunk = | |
1714 | (arena_chunk_t *)CHUNK_ADDR2BASE(rdelm); | |
1715 | arena_chunk_map_misc_t *miscelm = | |
1716 | arena_rd_to_miscelm(rdelm); | |
1717 | size_t pageind = arena_miscelm_to_pageind(miscelm); | |
1718 | bool decommitted = (arena_mapbits_decommitted_get(chunk, | |
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 | } | |
1a4d82fc | 1724 | } |
970d7e83 LB |
1725 | } |
1726 | ||
54a0048b SL |
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 | */ | |
1736 | static void | |
1737 | arena_purge_to_limit(arena_t *arena, size_t ndirty_limit) | |
970d7e83 | 1738 | { |
54a0048b SL |
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; | |
9cc50fc6 | 1743 | |
54a0048b SL |
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) { | |
1a4d82fc | 1751 | size_t ndirty = arena_dirty_count(arena); |
970d7e83 LB |
1752 | assert(ndirty == arena->ndirty); |
1753 | } | |
54a0048b SL |
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); | |
970d7e83 LB |
1769 | |
1770 | if (config_stats) | |
1771 | arena->stats.npurge++; | |
1772 | ||
54a0048b SL |
1773 | label_return: |
1774 | arena->purging = false; | |
970d7e83 LB |
1775 | } |
1776 | ||
1777 | void | |
54a0048b | 1778 | arena_purge(arena_t *arena, bool all) |
970d7e83 LB |
1779 | { |
1780 | ||
1781 | malloc_mutex_lock(&arena->lock); | |
54a0048b SL |
1782 | if (all) |
1783 | arena_purge_to_limit(arena, 0); | |
1784 | else | |
1785 | arena_maybe_purge(arena); | |
970d7e83 LB |
1786 | malloc_mutex_unlock(&arena->lock); |
1787 | } | |
1788 | ||
1789 | static void | |
1a4d82fc | 1790 | arena_run_coalesce(arena_t *arena, arena_chunk_t *chunk, size_t *p_size, |
54a0048b SL |
1791 | size_t *p_run_ind, size_t *p_run_pages, size_t flag_dirty, |
1792 | size_t flag_decommitted) | |
970d7e83 | 1793 | { |
1a4d82fc JJ |
1794 | size_t size = *p_size; |
1795 | size_t run_ind = *p_run_ind; | |
1796 | size_t run_pages = *p_run_pages; | |
970d7e83 LB |
1797 | |
1798 | /* Try to coalesce forward. */ | |
1799 | if (run_ind + run_pages < chunk_npages && | |
1800 | arena_mapbits_allocated_get(chunk, run_ind+run_pages) == 0 && | |
54a0048b SL |
1801 | arena_mapbits_dirty_get(chunk, run_ind+run_pages) == flag_dirty && |
1802 | arena_mapbits_decommitted_get(chunk, run_ind+run_pages) == | |
1803 | flag_decommitted) { | |
970d7e83 LB |
1804 | size_t nrun_size = arena_mapbits_unallocated_size_get(chunk, |
1805 | run_ind+run_pages); | |
1806 | size_t nrun_pages = nrun_size >> LG_PAGE; | |
1807 | ||
1808 | /* | |
1809 | * Remove successor from runs_avail; the coalesced run is | |
1810 | * inserted later. | |
1811 | */ | |
1812 | assert(arena_mapbits_unallocated_size_get(chunk, | |
1813 | run_ind+run_pages+nrun_pages-1) == nrun_size); | |
1814 | assert(arena_mapbits_dirty_get(chunk, | |
1815 | run_ind+run_pages+nrun_pages-1) == flag_dirty); | |
54a0048b SL |
1816 | assert(arena_mapbits_decommitted_get(chunk, |
1817 | run_ind+run_pages+nrun_pages-1) == flag_decommitted); | |
1a4d82fc JJ |
1818 | arena_avail_remove(arena, chunk, run_ind+run_pages, nrun_pages); |
1819 | ||
54a0048b SL |
1820 | /* |
1821 | * If the successor is dirty, remove it from the set of dirty | |
1822 | * pages. | |
1823 | */ | |
1a4d82fc | 1824 | if (flag_dirty != 0) { |
54a0048b | 1825 | arena_run_dirty_remove(arena, chunk, run_ind+run_pages, |
1a4d82fc JJ |
1826 | nrun_pages); |
1827 | } | |
970d7e83 LB |
1828 | |
1829 | size += nrun_size; | |
1830 | run_pages += nrun_pages; | |
1831 | ||
1832 | arena_mapbits_unallocated_size_set(chunk, run_ind, size); | |
1833 | arena_mapbits_unallocated_size_set(chunk, run_ind+run_pages-1, | |
1834 | size); | |
1835 | } | |
1836 | ||
1837 | /* Try to coalesce backward. */ | |
1a4d82fc JJ |
1838 | if (run_ind > map_bias && arena_mapbits_allocated_get(chunk, |
1839 | run_ind-1) == 0 && arena_mapbits_dirty_get(chunk, run_ind-1) == | |
54a0048b SL |
1840 | flag_dirty && arena_mapbits_decommitted_get(chunk, run_ind-1) == |
1841 | flag_decommitted) { | |
970d7e83 LB |
1842 | size_t prun_size = arena_mapbits_unallocated_size_get(chunk, |
1843 | run_ind-1); | |
1844 | size_t prun_pages = prun_size >> LG_PAGE; | |
1845 | ||
1846 | run_ind -= prun_pages; | |
1847 | ||
1848 | /* | |
1849 | * Remove predecessor from runs_avail; the coalesced run is | |
1850 | * inserted later. | |
1851 | */ | |
1852 | assert(arena_mapbits_unallocated_size_get(chunk, run_ind) == | |
1853 | prun_size); | |
1854 | assert(arena_mapbits_dirty_get(chunk, run_ind) == flag_dirty); | |
54a0048b SL |
1855 | assert(arena_mapbits_decommitted_get(chunk, run_ind) == |
1856 | flag_decommitted); | |
1a4d82fc JJ |
1857 | arena_avail_remove(arena, chunk, run_ind, prun_pages); |
1858 | ||
54a0048b SL |
1859 | /* |
1860 | * If the predecessor is dirty, remove it from the set of dirty | |
1861 | * pages. | |
1862 | */ | |
1863 | if (flag_dirty != 0) { | |
1864 | arena_run_dirty_remove(arena, chunk, run_ind, | |
1865 | prun_pages); | |
1866 | } | |
970d7e83 LB |
1867 | |
1868 | size += prun_size; | |
1869 | run_pages += prun_pages; | |
1870 | ||
1871 | arena_mapbits_unallocated_size_set(chunk, run_ind, size); | |
1872 | arena_mapbits_unallocated_size_set(chunk, run_ind+run_pages-1, | |
1873 | size); | |
1874 | } | |
1875 | ||
1a4d82fc JJ |
1876 | *p_size = size; |
1877 | *p_run_ind = run_ind; | |
1878 | *p_run_pages = run_pages; | |
1879 | } | |
1880 | ||
54a0048b SL |
1881 | static size_t |
1882 | arena_run_size_get(arena_t *arena, arena_chunk_t *chunk, arena_run_t *run, | |
1883 | size_t run_ind) | |
9cc50fc6 | 1884 | { |
54a0048b | 1885 | size_t size; |
9cc50fc6 | 1886 | |
9cc50fc6 SL |
1887 | assert(run_ind >= map_bias); |
1888 | assert(run_ind < chunk_npages); | |
54a0048b | 1889 | |
7453a54e SL |
1890 | if (arena_mapbits_large_get(chunk, run_ind) != 0) { |
1891 | size = arena_mapbits_large_size_get(chunk, run_ind); | |
54a0048b | 1892 | assert(size == PAGE || arena_mapbits_large_size_get(chunk, |
7453a54e SL |
1893 | run_ind+(size>>LG_PAGE)-1) == 0); |
1894 | } else { | |
54a0048b | 1895 | arena_bin_info_t *bin_info = &arena_bin_info[run->binind]; |
7453a54e SL |
1896 | size = bin_info->run_size; |
1897 | } | |
54a0048b SL |
1898 | |
1899 | return (size); | |
1900 | } | |
1901 | ||
1902 | static void | |
1903 | arena_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); | |
1a4d82fc | 1916 | run_pages = (size >> LG_PAGE); |
54a0048b | 1917 | arena_nactive_sub(arena, run_pages); |
1a4d82fc JJ |
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)); | |
54a0048b SL |
1926 | if (!cleaned && !decommitted && arena_mapbits_dirty_get(chunk, run_ind) |
1927 | != 0) | |
1a4d82fc JJ |
1928 | dirty = true; |
1929 | flag_dirty = dirty ? CHUNK_MAP_DIRTY : 0; | |
54a0048b | 1930 | flag_decommitted = decommitted ? CHUNK_MAP_DECOMMITTED : 0; |
1a4d82fc JJ |
1931 | |
1932 | /* Mark pages as unallocated in the chunk map. */ | |
54a0048b SL |
1933 | if (dirty || decommitted) { |
1934 | size_t flags = flag_dirty | flag_decommitted; | |
1935 | arena_mapbits_unallocated_set(chunk, run_ind, size, flags); | |
1a4d82fc | 1936 | arena_mapbits_unallocated_set(chunk, run_ind+run_pages-1, size, |
54a0048b | 1937 | flags); |
1a4d82fc JJ |
1938 | } else { |
1939 | arena_mapbits_unallocated_set(chunk, run_ind, size, | |
1940 | arena_mapbits_unzeroed_get(chunk, run_ind)); | |
1941 | arena_mapbits_unallocated_set(chunk, run_ind+run_pages-1, size, | |
1942 | arena_mapbits_unzeroed_get(chunk, run_ind+run_pages-1)); | |
1943 | } | |
1944 | ||
54a0048b SL |
1945 | arena_run_coalesce(arena, chunk, &size, &run_ind, &run_pages, |
1946 | flag_dirty, flag_decommitted); | |
1a4d82fc | 1947 | |
970d7e83 LB |
1948 | /* Insert into runs_avail, now that coalescing is complete. */ |
1949 | assert(arena_mapbits_unallocated_size_get(chunk, run_ind) == | |
1950 | arena_mapbits_unallocated_size_get(chunk, run_ind+run_pages-1)); | |
1951 | assert(arena_mapbits_dirty_get(chunk, run_ind) == | |
1952 | arena_mapbits_dirty_get(chunk, run_ind+run_pages-1)); | |
54a0048b SL |
1953 | assert(arena_mapbits_decommitted_get(chunk, run_ind) == |
1954 | arena_mapbits_decommitted_get(chunk, run_ind+run_pages-1)); | |
1a4d82fc JJ |
1955 | arena_avail_insert(arena, chunk, run_ind, run_pages); |
1956 | ||
1957 | if (dirty) | |
54a0048b | 1958 | arena_run_dirty_insert(arena, chunk, run_ind, run_pages); |
970d7e83 LB |
1959 | |
1960 | /* Deallocate chunk if it is now completely unused. */ | |
54a0048b | 1961 | if (size == arena_maxrun) { |
970d7e83 | 1962 | assert(run_ind == map_bias); |
54a0048b | 1963 | assert(run_pages == (arena_maxrun >> LG_PAGE)); |
1a4d82fc | 1964 | arena_chunk_dalloc(arena, chunk); |
970d7e83 LB |
1965 | } |
1966 | ||
1967 | /* | |
1968 | * It is okay to do dirty page processing here even if the chunk was | |
1969 | * deallocated above, since in that case it is the spare. Waiting | |
1970 | * until after possible chunk deallocation to do dirty processing | |
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 | ||
1978 | static void | |
1979 | arena_run_trim_head(arena_t *arena, arena_chunk_t *chunk, arena_run_t *run, | |
1980 | size_t oldsize, size_t newsize) | |
1981 | { | |
1a4d82fc JJ |
1982 | arena_chunk_map_misc_t *miscelm = arena_run_to_miscelm(run); |
1983 | size_t pageind = arena_miscelm_to_pageind(miscelm); | |
970d7e83 LB |
1984 | size_t head_npages = (oldsize - newsize) >> LG_PAGE; |
1985 | size_t flag_dirty = arena_mapbits_dirty_get(chunk, pageind); | |
54a0048b SL |
1986 | size_t flag_decommitted = arena_mapbits_decommitted_get(chunk, pageind); |
1987 | size_t flag_unzeroed_mask = (flag_dirty | flag_decommitted) == 0 ? | |
1988 | CHUNK_MAP_UNZEROED : 0; | |
970d7e83 LB |
1989 | |
1990 | assert(oldsize > newsize); | |
1991 | ||
1992 | /* | |
1993 | * Update the chunk map so that arena_run_dalloc() can treat the | |
1994 | * leading run as separately allocated. Set the last element of each | |
1995 | * run first, in case of single-page runs. | |
1996 | */ | |
1997 | assert(arena_mapbits_large_size_get(chunk, pageind) == oldsize); | |
54a0048b SL |
1998 | arena_mapbits_large_set(chunk, pageind+head_npages-1, 0, flag_dirty | |
1999 | (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk, | |
2000 | pageind+head_npages-1))); | |
2001 | arena_mapbits_large_set(chunk, pageind, oldsize-newsize, flag_dirty | | |
2002 | (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk, pageind))); | |
970d7e83 LB |
2003 | |
2004 | if (config_debug) { | |
2005 | UNUSED size_t tail_npages = newsize >> LG_PAGE; | |
2006 | assert(arena_mapbits_large_size_get(chunk, | |
2007 | pageind+head_npages+tail_npages-1) == 0); | |
2008 | assert(arena_mapbits_dirty_get(chunk, | |
2009 | pageind+head_npages+tail_npages-1) == flag_dirty); | |
2010 | } | |
2011 | arena_mapbits_large_set(chunk, pageind+head_npages, newsize, | |
54a0048b SL |
2012 | flag_dirty | (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk, |
2013 | pageind+head_npages))); | |
970d7e83 | 2014 | |
54a0048b | 2015 | arena_run_dalloc(arena, run, false, false, (flag_decommitted != 0)); |
970d7e83 LB |
2016 | } |
2017 | ||
2018 | static void | |
2019 | arena_run_trim_tail(arena_t *arena, arena_chunk_t *chunk, arena_run_t *run, | |
2020 | size_t oldsize, size_t newsize, bool dirty) | |
2021 | { | |
1a4d82fc JJ |
2022 | arena_chunk_map_misc_t *miscelm = arena_run_to_miscelm(run); |
2023 | size_t pageind = arena_miscelm_to_pageind(miscelm); | |
970d7e83 LB |
2024 | size_t head_npages = newsize >> LG_PAGE; |
2025 | size_t flag_dirty = arena_mapbits_dirty_get(chunk, pageind); | |
54a0048b SL |
2026 | size_t flag_decommitted = arena_mapbits_decommitted_get(chunk, pageind); |
2027 | size_t flag_unzeroed_mask = (flag_dirty | flag_decommitted) == 0 ? | |
2028 | CHUNK_MAP_UNZEROED : 0; | |
1a4d82fc JJ |
2029 | arena_chunk_map_misc_t *tail_miscelm; |
2030 | arena_run_t *tail_run; | |
970d7e83 LB |
2031 | |
2032 | assert(oldsize > newsize); | |
2033 | ||
2034 | /* | |
2035 | * Update the chunk map so that arena_run_dalloc() can treat the | |
2036 | * trailing run as separately allocated. Set the last element of each | |
2037 | * run first, in case of single-page runs. | |
2038 | */ | |
2039 | assert(arena_mapbits_large_size_get(chunk, pageind) == oldsize); | |
54a0048b SL |
2040 | arena_mapbits_large_set(chunk, pageind+head_npages-1, 0, flag_dirty | |
2041 | (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk, | |
2042 | pageind+head_npages-1))); | |
2043 | arena_mapbits_large_set(chunk, pageind, newsize, flag_dirty | | |
2044 | (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk, pageind))); | |
970d7e83 LB |
2045 | |
2046 | if (config_debug) { | |
2047 | UNUSED size_t tail_npages = (oldsize - newsize) >> LG_PAGE; | |
2048 | assert(arena_mapbits_large_size_get(chunk, | |
2049 | pageind+head_npages+tail_npages-1) == 0); | |
2050 | assert(arena_mapbits_dirty_get(chunk, | |
2051 | pageind+head_npages+tail_npages-1) == flag_dirty); | |
2052 | } | |
2053 | arena_mapbits_large_set(chunk, pageind+head_npages, oldsize-newsize, | |
54a0048b SL |
2054 | flag_dirty | (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk, |
2055 | pageind+head_npages))); | |
970d7e83 | 2056 | |
1a4d82fc JJ |
2057 | tail_miscelm = arena_miscelm_get(chunk, pageind + head_npages); |
2058 | tail_run = &tail_miscelm->run; | |
54a0048b SL |
2059 | arena_run_dalloc(arena, tail_run, dirty, false, (flag_decommitted != |
2060 | 0)); | |
970d7e83 LB |
2061 | } |
2062 | ||
2063 | static arena_run_t * | |
2064 | arena_bin_runs_first(arena_bin_t *bin) | |
2065 | { | |
1a4d82fc JJ |
2066 | arena_chunk_map_misc_t *miscelm = arena_run_tree_first(&bin->runs); |
2067 | if (miscelm != NULL) | |
2068 | return (&miscelm->run); | |
970d7e83 LB |
2069 | |
2070 | return (NULL); | |
2071 | } | |
2072 | ||
2073 | static void | |
2074 | arena_bin_runs_insert(arena_bin_t *bin, arena_run_t *run) | |
2075 | { | |
1a4d82fc | 2076 | arena_chunk_map_misc_t *miscelm = arena_run_to_miscelm(run); |
970d7e83 | 2077 | |
1a4d82fc | 2078 | assert(arena_run_tree_search(&bin->runs, miscelm) == NULL); |
970d7e83 | 2079 | |
1a4d82fc | 2080 | arena_run_tree_insert(&bin->runs, miscelm); |
970d7e83 LB |
2081 | } |
2082 | ||
2083 | static void | |
2084 | arena_bin_runs_remove(arena_bin_t *bin, arena_run_t *run) | |
2085 | { | |
1a4d82fc | 2086 | arena_chunk_map_misc_t *miscelm = arena_run_to_miscelm(run); |
970d7e83 | 2087 | |
1a4d82fc | 2088 | assert(arena_run_tree_search(&bin->runs, miscelm) != NULL); |
970d7e83 | 2089 | |
1a4d82fc | 2090 | arena_run_tree_remove(&bin->runs, miscelm); |
970d7e83 LB |
2091 | } |
2092 | ||
2093 | static arena_run_t * | |
2094 | arena_bin_nonfull_run_tryget(arena_bin_t *bin) | |
2095 | { | |
2096 | arena_run_t *run = arena_bin_runs_first(bin); | |
2097 | if (run != NULL) { | |
2098 | arena_bin_runs_remove(bin, run); | |
2099 | if (config_stats) | |
2100 | bin->stats.reruns++; | |
2101 | } | |
2102 | return (run); | |
2103 | } | |
2104 | ||
2105 | static arena_run_t * | |
2106 | arena_bin_nonfull_run_get(arena_t *arena, arena_bin_t *bin) | |
2107 | { | |
2108 | arena_run_t *run; | |
54a0048b | 2109 | szind_t binind; |
970d7e83 LB |
2110 | arena_bin_info_t *bin_info; |
2111 | ||
2112 | /* Look for a usable run. */ | |
2113 | run = arena_bin_nonfull_run_tryget(bin); | |
2114 | if (run != NULL) | |
2115 | return (run); | |
2116 | /* No existing runs have any space available. */ | |
2117 | ||
2118 | binind = arena_bin_index(arena, bin); | |
2119 | bin_info = &arena_bin_info[binind]; | |
2120 | ||
2121 | /* Allocate a new run. */ | |
2122 | malloc_mutex_unlock(&bin->lock); | |
2123 | /******************************/ | |
2124 | malloc_mutex_lock(&arena->lock); | |
1a4d82fc | 2125 | run = arena_run_alloc_small(arena, bin_info->run_size, binind); |
970d7e83 | 2126 | if (run != NULL) { |
970d7e83 | 2127 | /* Initialize run internals. */ |
54a0048b | 2128 | run->binind = binind; |
970d7e83 | 2129 | run->nfree = bin_info->nregs; |
1a4d82fc | 2130 | bitmap_init(run->bitmap, &bin_info->bitmap_info); |
970d7e83 LB |
2131 | } |
2132 | malloc_mutex_unlock(&arena->lock); | |
2133 | /********************************/ | |
2134 | malloc_mutex_lock(&bin->lock); | |
2135 | if (run != NULL) { | |
2136 | if (config_stats) { | |
2137 | bin->stats.nruns++; | |
2138 | bin->stats.curruns++; | |
2139 | } | |
2140 | return (run); | |
2141 | } | |
2142 | ||
2143 | /* | |
1a4d82fc | 2144 | * arena_run_alloc_small() failed, but another thread may have made |
970d7e83 LB |
2145 | * sufficient memory available while this one dropped bin->lock above, |
2146 | * so search one more time. | |
2147 | */ | |
2148 | run = arena_bin_nonfull_run_tryget(bin); | |
2149 | if (run != NULL) | |
2150 | return (run); | |
2151 | ||
2152 | return (NULL); | |
2153 | } | |
2154 | ||
2155 | /* Re-fill bin->runcur, then call arena_run_reg_alloc(). */ | |
2156 | static void * | |
2157 | arena_bin_malloc_hard(arena_t *arena, arena_bin_t *bin) | |
2158 | { | |
54a0048b | 2159 | szind_t binind; |
970d7e83 LB |
2160 | arena_bin_info_t *bin_info; |
2161 | arena_run_t *run; | |
2162 | ||
2163 | binind = arena_bin_index(arena, bin); | |
2164 | bin_info = &arena_bin_info[binind]; | |
2165 | bin->runcur = NULL; | |
2166 | run = arena_bin_nonfull_run_get(arena, bin); | |
2167 | if (bin->runcur != NULL && bin->runcur->nfree > 0) { | |
2168 | /* | |
2169 | * Another thread updated runcur while this one ran without the | |
2170 | * bin lock in arena_bin_nonfull_run_get(). | |
2171 | */ | |
54a0048b | 2172 | void *ret; |
970d7e83 LB |
2173 | assert(bin->runcur->nfree > 0); |
2174 | ret = arena_run_reg_alloc(bin->runcur, bin_info); | |
2175 | if (run != NULL) { | |
2176 | arena_chunk_t *chunk; | |
2177 | ||
2178 | /* | |
1a4d82fc JJ |
2179 | * arena_run_alloc_small() may have allocated run, or |
2180 | * it may have pulled run from the bin's run tree. | |
2181 | * Therefore it is unsafe to make any assumptions about | |
2182 | * how run has previously been used, and | |
2183 | * arena_bin_lower_run() must be called, as if a region | |
2184 | * were just deallocated from the run. | |
970d7e83 LB |
2185 | */ |
2186 | chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run); | |
2187 | if (run->nfree == bin_info->nregs) | |
2188 | arena_dalloc_bin_run(arena, chunk, run, bin); | |
2189 | else | |
2190 | arena_bin_lower_run(arena, chunk, run, bin); | |
2191 | } | |
2192 | return (ret); | |
2193 | } | |
2194 | ||
2195 | if (run == NULL) | |
2196 | return (NULL); | |
2197 | ||
2198 | bin->runcur = run; | |
2199 | ||
2200 | assert(bin->runcur->nfree > 0); | |
2201 | ||
2202 | return (arena_run_reg_alloc(bin->runcur, bin_info)); | |
2203 | } | |
2204 | ||
2205 | void | |
54a0048b SL |
2206 | arena_tcache_fill_small(tsd_t *tsd, arena_t *arena, tcache_bin_t *tbin, |
2207 | szind_t binind, uint64_t prof_accumbytes) | |
970d7e83 LB |
2208 | { |
2209 | unsigned i, nfill; | |
2210 | arena_bin_t *bin; | |
970d7e83 LB |
2211 | |
2212 | assert(tbin->ncached == 0); | |
2213 | ||
2214 | if (config_prof && arena_prof_accum(arena, prof_accumbytes)) | |
2215 | prof_idump(); | |
2216 | bin = &arena->bins[binind]; | |
2217 | malloc_mutex_lock(&bin->lock); | |
2218 | for (i = 0, nfill = (tcache_bin_info[binind].ncached_max >> | |
2219 | tbin->lg_fill_div); i < nfill; i++) { | |
54a0048b SL |
2220 | arena_run_t *run; |
2221 | void *ptr; | |
970d7e83 LB |
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); | |
1a4d82fc JJ |
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 | |
54a0048b | 2230 | * be moved just before tbin->avail before bailing out. |
1a4d82fc JJ |
2231 | */ |
2232 | if (i > 0) { | |
54a0048b | 2233 | memmove(tbin->avail - i, tbin->avail - nfill, |
1a4d82fc JJ |
2234 | i * sizeof(void *)); |
2235 | } | |
970d7e83 | 2236 | break; |
1a4d82fc | 2237 | } |
54a0048b | 2238 | if (config_fill && unlikely(opt_junk_alloc)) { |
970d7e83 LB |
2239 | arena_alloc_junk_small(ptr, &arena_bin_info[binind], |
2240 | true); | |
2241 | } | |
2242 | /* Insert such that low regions get used first. */ | |
54a0048b | 2243 | *(tbin->avail - nfill + i) = ptr; |
970d7e83 LB |
2244 | } |
2245 | if (config_stats) { | |
970d7e83 LB |
2246 | bin->stats.nmalloc += i; |
2247 | bin->stats.nrequests += tbin->tstats.nrequests; | |
54a0048b | 2248 | bin->stats.curregs += i; |
970d7e83 LB |
2249 | bin->stats.nfills++; |
2250 | tbin->tstats.nrequests = 0; | |
2251 | } | |
2252 | malloc_mutex_unlock(&bin->lock); | |
2253 | tbin->ncached = i; | |
54a0048b | 2254 | arena_decay_tick(tsd, arena); |
970d7e83 LB |
2255 | } |
2256 | ||
2257 | void | |
2258 | arena_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; | |
2263 | memset((void *)((uintptr_t)ptr - redzone_size), 0xa5, | |
2264 | redzone_size); | |
2265 | memset((void *)((uintptr_t)ptr + bin_info->reg_size), 0xa5, | |
2266 | redzone_size); | |
2267 | } else { | |
2268 | memset((void *)((uintptr_t)ptr - bin_info->redzone_size), 0xa5, | |
2269 | bin_info->reg_interval); | |
2270 | } | |
2271 | } | |
2272 | ||
1a4d82fc JJ |
2273 | #ifdef JEMALLOC_JET |
2274 | #undef arena_redzone_corruption | |
2275 | #define arena_redzone_corruption JEMALLOC_N(arena_redzone_corruption_impl) | |
2276 | #endif | |
2277 | static void | |
2278 | arena_redzone_corruption(void *ptr, size_t usize, bool after, | |
2279 | size_t offset, uint8_t byte) | |
2280 | { | |
2281 | ||
2282 | malloc_printf("<jemalloc>: Corrupt redzone %zu byte%s %s %p " | |
2283 | "(size %zu), byte=%#x\n", offset, (offset == 1) ? "" : "s", | |
2284 | after ? "after" : "before", ptr, usize, byte); | |
2285 | } | |
2286 | #ifdef JEMALLOC_JET | |
2287 | #undef arena_redzone_corruption | |
2288 | #define arena_redzone_corruption JEMALLOC_N(arena_redzone_corruption) | |
2289 | arena_redzone_corruption_t *arena_redzone_corruption = | |
2290 | JEMALLOC_N(arena_redzone_corruption_impl); | |
2291 | #endif | |
2292 | ||
2293 | static void | |
2294 | arena_redzones_validate(void *ptr, arena_bin_info_t *bin_info, bool reset) | |
970d7e83 | 2295 | { |
970d7e83 LB |
2296 | bool error = false; |
2297 | ||
54a0048b SL |
2298 | if (opt_junk_alloc) { |
2299 | size_t size = bin_info->reg_size; | |
2300 | size_t redzone_size = bin_info->redzone_size; | |
2301 | size_t i; | |
2302 | ||
2303 | for (i = 1; i <= redzone_size; i++) { | |
2304 | uint8_t *byte = (uint8_t *)((uintptr_t)ptr - i); | |
2305 | if (*byte != 0xa5) { | |
2306 | error = true; | |
2307 | arena_redzone_corruption(ptr, size, false, i, | |
2308 | *byte); | |
2309 | if (reset) | |
2310 | *byte = 0xa5; | |
2311 | } | |
970d7e83 | 2312 | } |
54a0048b SL |
2313 | for (i = 0; i < redzone_size; i++) { |
2314 | uint8_t *byte = (uint8_t *)((uintptr_t)ptr + size + i); | |
2315 | if (*byte != 0xa5) { | |
2316 | error = true; | |
2317 | arena_redzone_corruption(ptr, size, true, i, | |
2318 | *byte); | |
2319 | if (reset) | |
2320 | *byte = 0xa5; | |
2321 | } | |
970d7e83 LB |
2322 | } |
2323 | } | |
54a0048b | 2324 | |
970d7e83 LB |
2325 | if (opt_abort && error) |
2326 | abort(); | |
1a4d82fc | 2327 | } |
970d7e83 | 2328 | |
1a4d82fc JJ |
2329 | #ifdef JEMALLOC_JET |
2330 | #undef arena_dalloc_junk_small | |
2331 | #define arena_dalloc_junk_small JEMALLOC_N(arena_dalloc_junk_small_impl) | |
2332 | #endif | |
2333 | void | |
2334 | arena_dalloc_junk_small(void *ptr, arena_bin_info_t *bin_info) | |
2335 | { | |
2336 | size_t redzone_size = bin_info->redzone_size; | |
2337 | ||
2338 | arena_redzones_validate(ptr, bin_info, false); | |
970d7e83 LB |
2339 | memset((void *)((uintptr_t)ptr - redzone_size), 0x5a, |
2340 | bin_info->reg_interval); | |
2341 | } | |
1a4d82fc JJ |
2342 | #ifdef JEMALLOC_JET |
2343 | #undef arena_dalloc_junk_small | |
2344 | #define arena_dalloc_junk_small JEMALLOC_N(arena_dalloc_junk_small) | |
2345 | arena_dalloc_junk_small_t *arena_dalloc_junk_small = | |
2346 | JEMALLOC_N(arena_dalloc_junk_small_impl); | |
2347 | #endif | |
2348 | ||
2349 | void | |
2350 | arena_quarantine_junk_small(void *ptr, size_t usize) | |
2351 | { | |
54a0048b | 2352 | szind_t binind; |
1a4d82fc JJ |
2353 | arena_bin_info_t *bin_info; |
2354 | cassert(config_fill); | |
54a0048b | 2355 | assert(opt_junk_free); |
1a4d82fc JJ |
2356 | assert(opt_quarantine); |
2357 | assert(usize <= SMALL_MAXCLASS); | |
2358 | ||
54a0048b | 2359 | binind = size2index(usize); |
1a4d82fc JJ |
2360 | bin_info = &arena_bin_info[binind]; |
2361 | arena_redzones_validate(ptr, bin_info, true); | |
2362 | } | |
970d7e83 | 2363 | |
54a0048b SL |
2364 | static void * |
2365 | arena_malloc_small(tsd_t *tsd, arena_t *arena, szind_t binind, bool zero) | |
970d7e83 LB |
2366 | { |
2367 | void *ret; | |
2368 | arena_bin_t *bin; | |
54a0048b | 2369 | size_t usize; |
970d7e83 | 2370 | arena_run_t *run; |
970d7e83 | 2371 | |
970d7e83 LB |
2372 | assert(binind < NBINS); |
2373 | bin = &arena->bins[binind]; | |
54a0048b | 2374 | usize = index2size(binind); |
970d7e83 LB |
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) { | |
970d7e83 LB |
2388 | bin->stats.nmalloc++; |
2389 | bin->stats.nrequests++; | |
54a0048b | 2390 | bin->stats.curregs++; |
970d7e83 LB |
2391 | } |
2392 | malloc_mutex_unlock(&bin->lock); | |
54a0048b | 2393 | if (config_prof && !isthreaded && arena_prof_accum(arena, usize)) |
970d7e83 LB |
2394 | prof_idump(); |
2395 | ||
1a4d82fc | 2396 | if (!zero) { |
970d7e83 | 2397 | if (config_fill) { |
54a0048b | 2398 | if (unlikely(opt_junk_alloc)) { |
970d7e83 LB |
2399 | arena_alloc_junk_small(ret, |
2400 | &arena_bin_info[binind], false); | |
1a4d82fc | 2401 | } else if (unlikely(opt_zero)) |
54a0048b | 2402 | memset(ret, 0, usize); |
970d7e83 | 2403 | } |
54a0048b | 2404 | JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED(ret, usize); |
970d7e83 | 2405 | } else { |
54a0048b | 2406 | if (config_fill && unlikely(opt_junk_alloc)) { |
970d7e83 LB |
2407 | arena_alloc_junk_small(ret, &arena_bin_info[binind], |
2408 | true); | |
2409 | } | |
54a0048b SL |
2410 | JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED(ret, usize); |
2411 | memset(ret, 0, usize); | |
970d7e83 | 2412 | } |
970d7e83 | 2413 | |
54a0048b | 2414 | arena_decay_tick(tsd, arena); |
970d7e83 LB |
2415 | return (ret); |
2416 | } | |
2417 | ||
2418 | void * | |
54a0048b | 2419 | arena_malloc_large(tsd_t *tsd, arena_t *arena, szind_t binind, bool zero) |
970d7e83 LB |
2420 | { |
2421 | void *ret; | |
54a0048b SL |
2422 | size_t usize; |
2423 | uintptr_t random_offset; | |
1a4d82fc JJ |
2424 | arena_run_t *run; |
2425 | arena_chunk_map_misc_t *miscelm; | |
54a0048b | 2426 | UNUSED bool idump JEMALLOC_CC_SILENCE_INIT(false); |
970d7e83 LB |
2427 | |
2428 | /* Large allocation. */ | |
54a0048b | 2429 | usize = index2size(binind); |
970d7e83 | 2430 | malloc_mutex_lock(&arena->lock); |
54a0048b SL |
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); | |
1a4d82fc | 2444 | if (run == NULL) { |
970d7e83 LB |
2445 | malloc_mutex_unlock(&arena->lock); |
2446 | return (NULL); | |
2447 | } | |
1a4d82fc | 2448 | miscelm = arena_run_to_miscelm(run); |
54a0048b SL |
2449 | ret = (void *)((uintptr_t)arena_miscelm_to_rpages(miscelm) + |
2450 | random_offset); | |
970d7e83 | 2451 | if (config_stats) { |
54a0048b SL |
2452 | szind_t index = binind - NBINS; |
2453 | ||
970d7e83 LB |
2454 | arena->stats.nmalloc_large++; |
2455 | arena->stats.nrequests_large++; | |
54a0048b SL |
2456 | arena->stats.allocated_large += usize; |
2457 | arena->stats.lstats[index].nmalloc++; | |
2458 | arena->stats.lstats[index].nrequests++; | |
2459 | arena->stats.lstats[index].curruns++; | |
970d7e83 LB |
2460 | } |
2461 | if (config_prof) | |
54a0048b | 2462 | idump = arena_prof_accum_locked(arena, usize); |
970d7e83 LB |
2463 | malloc_mutex_unlock(&arena->lock); |
2464 | if (config_prof && idump) | |
2465 | prof_idump(); | |
2466 | ||
1a4d82fc | 2467 | if (!zero) { |
970d7e83 | 2468 | if (config_fill) { |
54a0048b SL |
2469 | if (unlikely(opt_junk_alloc)) |
2470 | memset(ret, 0xa5, usize); | |
1a4d82fc | 2471 | else if (unlikely(opt_zero)) |
54a0048b | 2472 | memset(ret, 0, usize); |
970d7e83 LB |
2473 | } |
2474 | } | |
2475 | ||
54a0048b | 2476 | arena_decay_tick(tsd, arena); |
970d7e83 LB |
2477 | return (ret); |
2478 | } | |
2479 | ||
7453a54e | 2480 | void * |
54a0048b SL |
2481 | arena_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. */ | |
2497 | static void * | |
2498 | arena_palloc_large(tsd_t *tsd, arena_t *arena, size_t usize, size_t alignment, | |
2499 | bool zero) | |
970d7e83 LB |
2500 | { |
2501 | void *ret; | |
2502 | size_t alloc_size, leadsize, trailsize; | |
2503 | arena_run_t *run; | |
2504 | arena_chunk_t *chunk; | |
1a4d82fc JJ |
2505 | arena_chunk_map_misc_t *miscelm; |
2506 | void *rpages; | |
970d7e83 | 2507 | |
54a0048b SL |
2508 | assert(usize == PAGE_CEILING(usize)); |
2509 | ||
2510 | arena = arena_choose(tsd, arena); | |
2511 | if (unlikely(arena == NULL)) | |
2512 | return (NULL); | |
970d7e83 LB |
2513 | |
2514 | alignment = PAGE_CEILING(alignment); | |
54a0048b | 2515 | alloc_size = usize + large_pad + alignment - PAGE; |
970d7e83 LB |
2516 | |
2517 | malloc_mutex_lock(&arena->lock); | |
1a4d82fc | 2518 | run = arena_run_alloc_large(arena, alloc_size, false); |
970d7e83 LB |
2519 | if (run == NULL) { |
2520 | malloc_mutex_unlock(&arena->lock); | |
2521 | return (NULL); | |
2522 | } | |
2523 | chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run); | |
1a4d82fc JJ |
2524 | miscelm = arena_run_to_miscelm(run); |
2525 | rpages = arena_miscelm_to_rpages(miscelm); | |
970d7e83 | 2526 | |
1a4d82fc JJ |
2527 | leadsize = ALIGNMENT_CEILING((uintptr_t)rpages, alignment) - |
2528 | (uintptr_t)rpages; | |
54a0048b SL |
2529 | assert(alloc_size >= leadsize + usize); |
2530 | trailsize = alloc_size - leadsize - usize - large_pad; | |
970d7e83 | 2531 | if (leadsize != 0) { |
1a4d82fc JJ |
2532 | arena_chunk_map_misc_t *head_miscelm = miscelm; |
2533 | arena_run_t *head_run = run; | |
2534 | ||
2535 | miscelm = arena_miscelm_get(chunk, | |
2536 | arena_miscelm_to_pageind(head_miscelm) + (leadsize >> | |
2537 | LG_PAGE)); | |
2538 | run = &miscelm->run; | |
2539 | ||
2540 | arena_run_trim_head(arena, chunk, head_run, alloc_size, | |
2541 | alloc_size - leadsize); | |
970d7e83 LB |
2542 | } |
2543 | if (trailsize != 0) { | |
54a0048b SL |
2544 | arena_run_trim_tail(arena, chunk, run, usize + large_pad + |
2545 | trailsize, usize + large_pad, false); | |
2546 | } | |
2547 | if (arena_run_init_large(arena, run, usize + large_pad, zero)) { | |
2548 | size_t run_ind = | |
2549 | arena_miscelm_to_pageind(arena_run_to_miscelm(run)); | |
2550 | bool dirty = (arena_mapbits_dirty_get(chunk, run_ind) != 0); | |
2551 | bool decommitted = (arena_mapbits_decommitted_get(chunk, | |
2552 | run_ind) != 0); | |
2553 | ||
2554 | assert(decommitted); /* Cause of OOM. */ | |
2555 | arena_run_dalloc(arena, run, dirty, false, decommitted); | |
2556 | malloc_mutex_unlock(&arena->lock); | |
2557 | return (NULL); | |
970d7e83 | 2558 | } |
1a4d82fc | 2559 | ret = arena_miscelm_to_rpages(miscelm); |
970d7e83 LB |
2560 | |
2561 | if (config_stats) { | |
54a0048b SL |
2562 | szind_t index = size2index(usize) - NBINS; |
2563 | ||
970d7e83 LB |
2564 | arena->stats.nmalloc_large++; |
2565 | arena->stats.nrequests_large++; | |
54a0048b SL |
2566 | arena->stats.allocated_large += usize; |
2567 | arena->stats.lstats[index].nmalloc++; | |
2568 | arena->stats.lstats[index].nrequests++; | |
2569 | arena->stats.lstats[index].curruns++; | |
970d7e83 LB |
2570 | } |
2571 | malloc_mutex_unlock(&arena->lock); | |
2572 | ||
1a4d82fc | 2573 | if (config_fill && !zero) { |
54a0048b SL |
2574 | if (unlikely(opt_junk_alloc)) |
2575 | memset(ret, 0xa5, usize); | |
1a4d82fc | 2576 | else if (unlikely(opt_zero)) |
54a0048b SL |
2577 | memset(ret, 0, usize); |
2578 | } | |
2579 | arena_decay_tick(tsd, arena); | |
2580 | return (ret); | |
2581 | } | |
2582 | ||
2583 | void * | |
2584 | arena_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); | |
2611 | else { | |
2612 | ret = huge_palloc(tsd, arena, usize, alignment, zero, | |
2613 | tcache); | |
2614 | } | |
970d7e83 LB |
2615 | } |
2616 | return (ret); | |
2617 | } | |
2618 | ||
2619 | void | |
2620 | arena_prof_promoted(const void *ptr, size_t size) | |
2621 | { | |
2622 | arena_chunk_t *chunk; | |
54a0048b SL |
2623 | size_t pageind; |
2624 | szind_t binind; | |
970d7e83 LB |
2625 | |
2626 | cassert(config_prof); | |
2627 | assert(ptr != NULL); | |
2628 | assert(CHUNK_ADDR2BASE(ptr) != ptr); | |
54a0048b SL |
2629 | assert(isalloc(ptr, false) == LARGE_MINCLASS); |
2630 | assert(isalloc(ptr, true) == LARGE_MINCLASS); | |
970d7e83 LB |
2631 | assert(size <= SMALL_MAXCLASS); |
2632 | ||
2633 | chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); | |
2634 | pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; | |
54a0048b | 2635 | binind = size2index(size); |
970d7e83 LB |
2636 | assert(binind < NBINS); |
2637 | arena_mapbits_large_binind_set(chunk, pageind, binind); | |
2638 | ||
54a0048b | 2639 | assert(isalloc(ptr, false) == LARGE_MINCLASS); |
970d7e83 LB |
2640 | assert(isalloc(ptr, true) == size); |
2641 | } | |
2642 | ||
2643 | static void | |
2644 | arena_dissociate_bin_run(arena_chunk_t *chunk, arena_run_t *run, | |
2645 | arena_bin_t *bin) | |
2646 | { | |
2647 | ||
2648 | /* Dissociate run from bin. */ | |
2649 | if (run == bin->runcur) | |
2650 | bin->runcur = NULL; | |
2651 | else { | |
54a0048b SL |
2652 | szind_t binind = arena_bin_index(extent_node_arena_get( |
2653 | &chunk->node), bin); | |
970d7e83 LB |
2654 | arena_bin_info_t *bin_info = &arena_bin_info[binind]; |
2655 | ||
2656 | if (bin_info->nregs != 1) { | |
2657 | /* | |
2658 | * This block's conditional is necessary because if the | |
2659 | * run only contains one region, then it never gets | |
2660 | * inserted into the non-full runs tree. | |
2661 | */ | |
2662 | arena_bin_runs_remove(bin, run); | |
2663 | } | |
2664 | } | |
2665 | } | |
2666 | ||
2667 | static void | |
2668 | arena_dalloc_bin_run(arena_t *arena, arena_chunk_t *chunk, arena_run_t *run, | |
2669 | arena_bin_t *bin) | |
2670 | { | |
970d7e83 LB |
2671 | |
2672 | assert(run != bin->runcur); | |
1a4d82fc JJ |
2673 | assert(arena_run_tree_search(&bin->runs, arena_run_to_miscelm(run)) == |
2674 | NULL); | |
970d7e83 | 2675 | |
970d7e83 LB |
2676 | malloc_mutex_unlock(&bin->lock); |
2677 | /******************************/ | |
970d7e83 | 2678 | malloc_mutex_lock(&arena->lock); |
54a0048b | 2679 | arena_run_dalloc(arena, run, true, false, false); |
970d7e83 LB |
2680 | malloc_mutex_unlock(&arena->lock); |
2681 | /****************************/ | |
2682 | malloc_mutex_lock(&bin->lock); | |
2683 | if (config_stats) | |
2684 | bin->stats.curruns--; | |
2685 | } | |
2686 | ||
2687 | static void | |
2688 | arena_bin_lower_run(arena_t *arena, arena_chunk_t *chunk, arena_run_t *run, | |
2689 | arena_bin_t *bin) | |
2690 | { | |
2691 | ||
2692 | /* | |
2693 | * Make sure that if bin->runcur is non-NULL, it refers to the lowest | |
2694 | * non-full run. It is okay to NULL runcur out rather than proactively | |
2695 | * keeping it pointing at the lowest non-full run. | |
2696 | */ | |
2697 | if ((uintptr_t)run < (uintptr_t)bin->runcur) { | |
2698 | /* Switch runcur. */ | |
2699 | if (bin->runcur->nfree > 0) | |
2700 | arena_bin_runs_insert(bin, bin->runcur); | |
2701 | bin->runcur = run; | |
2702 | if (config_stats) | |
2703 | bin->stats.reruns++; | |
2704 | } else | |
2705 | arena_bin_runs_insert(bin, run); | |
2706 | } | |
2707 | ||
54a0048b SL |
2708 | static void |
2709 | arena_dalloc_bin_locked_impl(arena_t *arena, arena_chunk_t *chunk, void *ptr, | |
2710 | arena_chunk_map_bits_t *bitselm, bool junked) | |
970d7e83 | 2711 | { |
1a4d82fc | 2712 | size_t pageind, rpages_ind; |
970d7e83 LB |
2713 | arena_run_t *run; |
2714 | arena_bin_t *bin; | |
2715 | arena_bin_info_t *bin_info; | |
54a0048b | 2716 | szind_t binind; |
970d7e83 LB |
2717 | |
2718 | pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; | |
1a4d82fc JJ |
2719 | rpages_ind = pageind - arena_mapbits_small_runind_get(chunk, pageind); |
2720 | run = &arena_miscelm_get(chunk, rpages_ind)->run; | |
54a0048b SL |
2721 | binind = run->binind; |
2722 | bin = &arena->bins[binind]; | |
970d7e83 | 2723 | bin_info = &arena_bin_info[binind]; |
970d7e83 | 2724 | |
54a0048b | 2725 | if (!junked && config_fill && unlikely(opt_junk_free)) |
970d7e83 LB |
2726 | arena_dalloc_junk_small(ptr, bin_info); |
2727 | ||
2728 | arena_run_reg_dalloc(run, ptr); | |
2729 | if (run->nfree == bin_info->nregs) { | |
2730 | arena_dissociate_bin_run(chunk, run, bin); | |
2731 | arena_dalloc_bin_run(arena, chunk, run, bin); | |
2732 | } else if (run->nfree == 1 && run != bin->runcur) | |
2733 | arena_bin_lower_run(arena, chunk, run, bin); | |
2734 | ||
2735 | if (config_stats) { | |
970d7e83 | 2736 | bin->stats.ndalloc++; |
54a0048b | 2737 | bin->stats.curregs--; |
970d7e83 LB |
2738 | } |
2739 | } | |
2740 | ||
54a0048b SL |
2741 | void |
2742 | arena_dalloc_bin_junked_locked(arena_t *arena, arena_chunk_t *chunk, void *ptr, | |
2743 | arena_chunk_map_bits_t *bitselm) | |
2744 | { | |
2745 | ||
2746 | arena_dalloc_bin_locked_impl(arena, chunk, ptr, bitselm, true); | |
2747 | } | |
2748 | ||
970d7e83 LB |
2749 | void |
2750 | arena_dalloc_bin(arena_t *arena, arena_chunk_t *chunk, void *ptr, | |
1a4d82fc | 2751 | size_t pageind, arena_chunk_map_bits_t *bitselm) |
970d7e83 LB |
2752 | { |
2753 | arena_run_t *run; | |
2754 | arena_bin_t *bin; | |
1a4d82fc | 2755 | size_t rpages_ind; |
970d7e83 | 2756 | |
1a4d82fc JJ |
2757 | rpages_ind = pageind - arena_mapbits_small_runind_get(chunk, pageind); |
2758 | run = &arena_miscelm_get(chunk, rpages_ind)->run; | |
54a0048b | 2759 | bin = &arena->bins[run->binind]; |
970d7e83 | 2760 | malloc_mutex_lock(&bin->lock); |
54a0048b | 2761 | arena_dalloc_bin_locked_impl(arena, chunk, ptr, bitselm, false); |
970d7e83 LB |
2762 | malloc_mutex_unlock(&bin->lock); |
2763 | } | |
2764 | ||
2765 | void | |
54a0048b | 2766 | arena_dalloc_small(tsd_t *tsd, arena_t *arena, arena_chunk_t *chunk, void *ptr, |
970d7e83 LB |
2767 | size_t pageind) |
2768 | { | |
1a4d82fc | 2769 | arena_chunk_map_bits_t *bitselm; |
970d7e83 LB |
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 | } | |
1a4d82fc JJ |
2776 | bitselm = arena_bitselm_get(chunk, pageind); |
2777 | arena_dalloc_bin(arena, chunk, ptr, pageind, bitselm); | |
54a0048b | 2778 | arena_decay_tick(tsd, arena); |
970d7e83 LB |
2779 | } |
2780 | ||
1a4d82fc JJ |
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 | |
54a0048b | 2785 | void |
1a4d82fc JJ |
2786 | arena_dalloc_junk_large(void *ptr, size_t usize) |
2787 | { | |
2788 | ||
54a0048b | 2789 | if (config_fill && unlikely(opt_junk_free)) |
1a4d82fc JJ |
2790 | memset(ptr, 0x5a, usize); |
2791 | } | |
2792 | #ifdef JEMALLOC_JET | |
2793 | #undef arena_dalloc_junk_large | |
2794 | #define arena_dalloc_junk_large JEMALLOC_N(arena_dalloc_junk_large) | |
2795 | arena_dalloc_junk_large_t *arena_dalloc_junk_large = | |
2796 | JEMALLOC_N(arena_dalloc_junk_large_impl); | |
2797 | #endif | |
2798 | ||
54a0048b SL |
2799 | static void |
2800 | arena_dalloc_large_locked_impl(arena_t *arena, arena_chunk_t *chunk, | |
2801 | void *ptr, bool junked) | |
970d7e83 | 2802 | { |
1a4d82fc JJ |
2803 | size_t pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; |
2804 | arena_chunk_map_misc_t *miscelm = arena_miscelm_get(chunk, pageind); | |
2805 | arena_run_t *run = &miscelm->run; | |
970d7e83 LB |
2806 | |
2807 | if (config_fill || config_stats) { | |
54a0048b SL |
2808 | size_t usize = arena_mapbits_large_size_get(chunk, pageind) - |
2809 | large_pad; | |
970d7e83 | 2810 | |
54a0048b SL |
2811 | if (!junked) |
2812 | arena_dalloc_junk_large(ptr, usize); | |
970d7e83 | 2813 | if (config_stats) { |
54a0048b SL |
2814 | szind_t index = size2index(usize) - NBINS; |
2815 | ||
970d7e83 | 2816 | arena->stats.ndalloc_large++; |
1a4d82fc | 2817 | arena->stats.allocated_large -= usize; |
54a0048b SL |
2818 | arena->stats.lstats[index].ndalloc++; |
2819 | arena->stats.lstats[index].curruns--; | |
970d7e83 LB |
2820 | } |
2821 | } | |
2822 | ||
54a0048b SL |
2823 | arena_run_dalloc(arena, run, true, false, false); |
2824 | } | |
2825 | ||
2826 | void | |
2827 | arena_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); | |
970d7e83 LB |
2832 | } |
2833 | ||
2834 | void | |
54a0048b | 2835 | arena_dalloc_large(tsd_t *tsd, arena_t *arena, arena_chunk_t *chunk, void *ptr) |
970d7e83 LB |
2836 | { |
2837 | ||
2838 | malloc_mutex_lock(&arena->lock); | |
54a0048b | 2839 | arena_dalloc_large_locked_impl(arena, chunk, ptr, false); |
970d7e83 | 2840 | malloc_mutex_unlock(&arena->lock); |
54a0048b | 2841 | arena_decay_tick(tsd, arena); |
970d7e83 LB |
2842 | } |
2843 | ||
2844 | static void | |
2845 | arena_ralloc_large_shrink(arena_t *arena, arena_chunk_t *chunk, void *ptr, | |
2846 | size_t oldsize, size_t size) | |
2847 | { | |
1a4d82fc JJ |
2848 | size_t pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; |
2849 | arena_chunk_map_misc_t *miscelm = arena_miscelm_get(chunk, pageind); | |
2850 | arena_run_t *run = &miscelm->run; | |
970d7e83 LB |
2851 | |
2852 | assert(size < oldsize); | |
2853 | ||
2854 | /* | |
2855 | * Shrink the run, and make trailing pages available for other | |
2856 | * allocations. | |
2857 | */ | |
2858 | malloc_mutex_lock(&arena->lock); | |
54a0048b SL |
2859 | arena_run_trim_tail(arena, chunk, run, oldsize + large_pad, size + |
2860 | large_pad, true); | |
970d7e83 | 2861 | if (config_stats) { |
54a0048b SL |
2862 | szind_t oldindex = size2index(oldsize) - NBINS; |
2863 | szind_t index = size2index(size) - NBINS; | |
2864 | ||
970d7e83 LB |
2865 | arena->stats.ndalloc_large++; |
2866 | arena->stats.allocated_large -= oldsize; | |
54a0048b SL |
2867 | arena->stats.lstats[oldindex].ndalloc++; |
2868 | arena->stats.lstats[oldindex].curruns--; | |
970d7e83 LB |
2869 | |
2870 | arena->stats.nmalloc_large++; | |
2871 | arena->stats.nrequests_large++; | |
2872 | arena->stats.allocated_large += size; | |
54a0048b SL |
2873 | arena->stats.lstats[index].nmalloc++; |
2874 | arena->stats.lstats[index].nrequests++; | |
2875 | arena->stats.lstats[index].curruns++; | |
970d7e83 LB |
2876 | } |
2877 | malloc_mutex_unlock(&arena->lock); | |
2878 | } | |
2879 | ||
2880 | static bool | |
2881 | arena_ralloc_large_grow(arena_t *arena, arena_chunk_t *chunk, void *ptr, | |
54a0048b | 2882 | size_t oldsize, size_t usize_min, size_t usize_max, bool zero) |
970d7e83 LB |
2883 | { |
2884 | size_t pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; | |
54a0048b | 2885 | size_t npages = (oldsize + large_pad) >> LG_PAGE; |
970d7e83 LB |
2886 | size_t followsize; |
2887 | ||
54a0048b SL |
2888 | assert(oldsize == arena_mapbits_large_size_get(chunk, pageind) - |
2889 | large_pad); | |
970d7e83 LB |
2890 | |
2891 | /* Try to extend the run. */ | |
970d7e83 | 2892 | malloc_mutex_lock(&arena->lock); |
54a0048b SL |
2893 | if (pageind+npages >= chunk_npages || arena_mapbits_allocated_get(chunk, |
2894 | pageind+npages) != 0) | |
2895 | goto label_fail; | |
2896 | followsize = arena_mapbits_unallocated_size_get(chunk, pageind+npages); | |
2897 | if (oldsize + followsize >= usize_min) { | |
970d7e83 LB |
2898 | /* |
2899 | * The next run is available and sufficiently large. Split the | |
2900 | * following run, then merge the first part with the existing | |
2901 | * allocation. | |
2902 | */ | |
54a0048b SL |
2903 | arena_run_t *run; |
2904 | size_t usize, splitsize, size, flag_dirty, flag_unzeroed_mask; | |
2905 | ||
2906 | usize = usize_max; | |
2907 | while (oldsize + followsize < usize) | |
2908 | usize = index2size(size2index(usize)-1); | |
2909 | assert(usize >= usize_min); | |
2910 | assert(usize >= oldsize); | |
2911 | splitsize = usize - oldsize; | |
2912 | if (splitsize == 0) | |
2913 | goto label_fail; | |
2914 | ||
2915 | run = &arena_miscelm_get(chunk, pageind+npages)->run; | |
2916 | if (arena_run_split_large(arena, run, splitsize, zero)) | |
2917 | goto label_fail; | |
2918 | ||
2919 | if (config_cache_oblivious && zero) { | |
2920 | /* | |
2921 | * Zero the trailing bytes of the original allocation's | |
2922 | * last page, since they are in an indeterminate state. | |
2923 | * There will always be trailing bytes, because ptr's | |
2924 | * offset from the beginning of the run is a multiple of | |
2925 | * CACHELINE in [0 .. PAGE). | |
2926 | */ | |
2927 | void *zbase = (void *)((uintptr_t)ptr + oldsize); | |
2928 | void *zpast = PAGE_ADDR2BASE((void *)((uintptr_t)zbase + | |
2929 | PAGE)); | |
2930 | size_t nzero = (uintptr_t)zpast - (uintptr_t)zbase; | |
2931 | assert(nzero > 0); | |
2932 | memset(zbase, 0, nzero); | |
2933 | } | |
970d7e83 LB |
2934 | |
2935 | size = oldsize + splitsize; | |
54a0048b | 2936 | npages = (size + large_pad) >> LG_PAGE; |
970d7e83 LB |
2937 | |
2938 | /* | |
2939 | * Mark the extended run as dirty if either portion of the run | |
2940 | * was dirty before allocation. This is rather pedantic, | |
2941 | * because there's not actually any sequence of events that | |
2942 | * could cause the resulting run to be passed to | |
2943 | * arena_run_dalloc() with the dirty argument set to false | |
2944 | * (which is when dirty flag consistency would really matter). | |
2945 | */ | |
2946 | flag_dirty = arena_mapbits_dirty_get(chunk, pageind) | | |
2947 | arena_mapbits_dirty_get(chunk, pageind+npages-1); | |
54a0048b SL |
2948 | flag_unzeroed_mask = flag_dirty == 0 ? CHUNK_MAP_UNZEROED : 0; |
2949 | arena_mapbits_large_set(chunk, pageind, size + large_pad, | |
2950 | flag_dirty | (flag_unzeroed_mask & | |
2951 | arena_mapbits_unzeroed_get(chunk, pageind))); | |
2952 | arena_mapbits_large_set(chunk, pageind+npages-1, 0, flag_dirty | | |
2953 | (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk, | |
2954 | pageind+npages-1))); | |
970d7e83 LB |
2955 | |
2956 | if (config_stats) { | |
54a0048b SL |
2957 | szind_t oldindex = size2index(oldsize) - NBINS; |
2958 | szind_t index = size2index(size) - NBINS; | |
2959 | ||
970d7e83 LB |
2960 | arena->stats.ndalloc_large++; |
2961 | arena->stats.allocated_large -= oldsize; | |
54a0048b SL |
2962 | arena->stats.lstats[oldindex].ndalloc++; |
2963 | arena->stats.lstats[oldindex].curruns--; | |
970d7e83 LB |
2964 | |
2965 | arena->stats.nmalloc_large++; | |
2966 | arena->stats.nrequests_large++; | |
2967 | arena->stats.allocated_large += size; | |
54a0048b SL |
2968 | arena->stats.lstats[index].nmalloc++; |
2969 | arena->stats.lstats[index].nrequests++; | |
2970 | arena->stats.lstats[index].curruns++; | |
970d7e83 LB |
2971 | } |
2972 | malloc_mutex_unlock(&arena->lock); | |
2973 | return (false); | |
2974 | } | |
54a0048b | 2975 | label_fail: |
970d7e83 | 2976 | malloc_mutex_unlock(&arena->lock); |
970d7e83 LB |
2977 | return (true); |
2978 | } | |
2979 | ||
1a4d82fc JJ |
2980 | #ifdef JEMALLOC_JET |
2981 | #undef arena_ralloc_junk_large | |
2982 | #define arena_ralloc_junk_large JEMALLOC_N(arena_ralloc_junk_large_impl) | |
2983 | #endif | |
2984 | static void | |
2985 | arena_ralloc_junk_large(void *ptr, size_t old_usize, size_t usize) | |
2986 | { | |
2987 | ||
54a0048b | 2988 | if (config_fill && unlikely(opt_junk_free)) { |
1a4d82fc JJ |
2989 | memset((void *)((uintptr_t)ptr + usize), 0x5a, |
2990 | old_usize - usize); | |
2991 | } | |
2992 | } | |
2993 | #ifdef JEMALLOC_JET | |
2994 | #undef arena_ralloc_junk_large | |
2995 | #define arena_ralloc_junk_large JEMALLOC_N(arena_ralloc_junk_large) | |
2996 | arena_ralloc_junk_large_t *arena_ralloc_junk_large = | |
2997 | JEMALLOC_N(arena_ralloc_junk_large_impl); | |
2998 | #endif | |
2999 | ||
970d7e83 LB |
3000 | /* |
3001 | * Try to resize a large allocation, in order to avoid copying. This will | |
3002 | * always fail if growing an object, and the following run is already in use. | |
3003 | */ | |
3004 | static bool | |
54a0048b SL |
3005 | arena_ralloc_large(void *ptr, size_t oldsize, size_t usize_min, |
3006 | size_t usize_max, bool zero) | |
970d7e83 | 3007 | { |
54a0048b SL |
3008 | arena_chunk_t *chunk; |
3009 | arena_t *arena; | |
970d7e83 | 3010 | |
54a0048b SL |
3011 | if (oldsize == usize_max) { |
3012 | /* Current size class is compatible and maximal. */ | |
970d7e83 | 3013 | return (false); |
54a0048b | 3014 | } |
7453a54e | 3015 | |
54a0048b SL |
3016 | chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); |
3017 | arena = extent_node_arena_get(&chunk->node); | |
3018 | ||
3019 | if (oldsize < usize_max) { | |
3020 | bool ret = arena_ralloc_large_grow(arena, chunk, ptr, oldsize, | |
3021 | usize_min, usize_max, zero); | |
3022 | if (config_fill && !ret && !zero) { | |
3023 | if (unlikely(opt_junk_alloc)) { | |
3024 | memset((void *)((uintptr_t)ptr + oldsize), 0xa5, | |
3025 | isalloc(ptr, config_prof) - oldsize); | |
3026 | } else if (unlikely(opt_zero)) { | |
3027 | memset((void *)((uintptr_t)ptr + oldsize), 0, | |
3028 | isalloc(ptr, config_prof) - oldsize); | |
970d7e83 | 3029 | } |
970d7e83 | 3030 | } |
54a0048b | 3031 | return (ret); |
970d7e83 | 3032 | } |
54a0048b SL |
3033 | |
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); | |
970d7e83 LB |
3039 | } |
3040 | ||
1a4d82fc | 3041 | bool |
54a0048b SL |
3042 | arena_ralloc_no_move(tsd_t *tsd, void *ptr, size_t oldsize, size_t size, |
3043 | size_t extra, bool zero) | |
970d7e83 | 3044 | { |
54a0048b | 3045 | size_t usize_min, usize_max; |
970d7e83 | 3046 | |
54a0048b SL |
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 | */ | |
970d7e83 | 3062 | if (oldsize <= SMALL_MAXCLASS) { |
54a0048b SL |
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); | |
970d7e83 | 3069 | } else { |
54a0048b SL |
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); | |
970d7e83 | 3075 | } |
54a0048b SL |
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)); | |
970d7e83 | 3083 | } |
54a0048b | 3084 | } |
9cc50fc6 | 3085 | |
54a0048b SL |
3086 | static void * |
3087 | arena_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)); | |
970d7e83 LB |
3098 | } |
3099 | ||
3100 | void * | |
1a4d82fc | 3101 | arena_ralloc(tsd_t *tsd, arena_t *arena, void *ptr, size_t oldsize, size_t size, |
54a0048b | 3102 | size_t alignment, bool zero, tcache_t *tcache) |
970d7e83 LB |
3103 | { |
3104 | void *ret; | |
54a0048b | 3105 | size_t usize; |
970d7e83 | 3106 | |
54a0048b SL |
3107 | usize = s2u(size); |
3108 | if (unlikely(usize == 0 || size > HUGE_MAXCLASS)) | |
3109 | return (NULL); | |
970d7e83 | 3110 | |
54a0048b SL |
3111 | if (likely(usize <= large_maxclass)) { |
3112 | size_t copysize; | |
970d7e83 | 3113 | |
54a0048b SL |
3114 | /* Try to avoid moving the allocation. */ |
3115 | if (!arena_ralloc_no_move(tsd, ptr, oldsize, usize, 0, zero)) | |
3116 | return (ptr); | |
970d7e83 | 3117 | |
54a0048b SL |
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, | |
3124 | zero, tcache); | |
970d7e83 LB |
3125 | if (ret == NULL) |
3126 | return (NULL); | |
970d7e83 | 3127 | |
54a0048b SL |
3128 | /* |
3129 | * Junk/zero-filling were already done by | |
3130 | * ipalloc()/arena_malloc(). | |
3131 | */ | |
970d7e83 | 3132 | |
54a0048b SL |
3133 | copysize = (usize < oldsize) ? usize : oldsize; |
3134 | JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED(ret, copysize); | |
3135 | memcpy(ret, ptr, copysize); | |
3136 | isqalloc(tsd, ptr, oldsize, tcache); | |
3137 | } else { | |
3138 | ret = huge_ralloc(tsd, arena, ptr, oldsize, usize, alignment, | |
3139 | zero, tcache); | |
3140 | } | |
970d7e83 LB |
3141 | return (ret); |
3142 | } | |
3143 | ||
3144 | dss_prec_t | |
3145 | arena_dss_prec_get(arena_t *arena) | |
3146 | { | |
3147 | dss_prec_t ret; | |
3148 | ||
3149 | malloc_mutex_lock(&arena->lock); | |
3150 | ret = arena->dss_prec; | |
3151 | malloc_mutex_unlock(&arena->lock); | |
3152 | return (ret); | |
3153 | } | |
3154 | ||
1a4d82fc | 3155 | bool |
970d7e83 LB |
3156 | arena_dss_prec_set(arena_t *arena, dss_prec_t dss_prec) |
3157 | { | |
3158 | ||
1a4d82fc JJ |
3159 | if (!have_dss) |
3160 | return (dss_prec != dss_prec_disabled); | |
970d7e83 LB |
3161 | malloc_mutex_lock(&arena->lock); |
3162 | arena->dss_prec = dss_prec; | |
3163 | malloc_mutex_unlock(&arena->lock); | |
1a4d82fc | 3164 | return (false); |
970d7e83 LB |
3165 | } |
3166 | ||
54a0048b SL |
3167 | ssize_t |
3168 | arena_lg_dirty_mult_default_get(void) | |
3169 | { | |
3170 | ||
3171 | return ((ssize_t)atomic_read_z((size_t *)&lg_dirty_mult_default)); | |
3172 | } | |
3173 | ||
3174 | bool | |
3175 | arena_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 | ||
3186 | ssize_t | |
3187 | arena_decay_time_default_get(void) | |
3188 | { | |
3189 | ||
3190 | return ((ssize_t)atomic_read_z((size_t *)&decay_time_default)); | |
3191 | } | |
3192 | ||
3193 | bool | |
3194 | arena_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 | ||
3205 | static void | |
3206 | arena_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 | ||
970d7e83 | 3219 | void |
54a0048b SL |
3220 | arena_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 | ||
3231 | void | |
3232 | arena_stats_merge(arena_t *arena, unsigned *nthreads, const char **dss, | |
3233 | ssize_t *lg_dirty_mult, ssize_t *decay_time, size_t *nactive, | |
7453a54e | 3234 | size_t *ndirty, arena_stats_t *astats, malloc_bin_stats_t *bstats, |
54a0048b | 3235 | malloc_large_stats_t *lstats, malloc_huge_stats_t *hstats) |
970d7e83 LB |
3236 | { |
3237 | unsigned i; | |
3238 | ||
54a0048b SL |
3239 | cassert(config_stats); |
3240 | ||
970d7e83 | 3241 | malloc_mutex_lock(&arena->lock); |
54a0048b SL |
3242 | arena_basic_stats_merge_locked(arena, nthreads, dss, lg_dirty_mult, |
3243 | decay_time, nactive, ndirty); | |
970d7e83 LB |
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; | |
54a0048b SL |
3249 | astats->metadata_mapped += arena->stats.metadata_mapped; |
3250 | astats->metadata_allocated += arena_metadata_allocated_get(arena); | |
970d7e83 LB |
3251 | astats->allocated_large += arena->stats.allocated_large; |
3252 | astats->nmalloc_large += arena->stats.nmalloc_large; | |
3253 | astats->ndalloc_large += arena->stats.ndalloc_large; | |
3254 | astats->nrequests_large += arena->stats.nrequests_large; | |
1a4d82fc JJ |
3255 | astats->allocated_huge += arena->stats.allocated_huge; |
3256 | astats->nmalloc_huge += arena->stats.nmalloc_huge; | |
3257 | astats->ndalloc_huge += arena->stats.ndalloc_huge; | |
970d7e83 LB |
3258 | |
3259 | for (i = 0; i < nlclasses; i++) { | |
3260 | lstats[i].nmalloc += arena->stats.lstats[i].nmalloc; | |
3261 | lstats[i].ndalloc += arena->stats.lstats[i].ndalloc; | |
3262 | lstats[i].nrequests += arena->stats.lstats[i].nrequests; | |
3263 | lstats[i].curruns += arena->stats.lstats[i].curruns; | |
3264 | } | |
54a0048b SL |
3265 | |
3266 | for (i = 0; i < nhclasses; i++) { | |
3267 | hstats[i].nmalloc += arena->stats.hstats[i].nmalloc; | |
3268 | hstats[i].ndalloc += arena->stats.hstats[i].ndalloc; | |
3269 | hstats[i].curhchunks += arena->stats.hstats[i].curhchunks; | |
3270 | } | |
970d7e83 LB |
3271 | malloc_mutex_unlock(&arena->lock); |
3272 | ||
3273 | for (i = 0; i < NBINS; i++) { | |
3274 | arena_bin_t *bin = &arena->bins[i]; | |
3275 | ||
3276 | malloc_mutex_lock(&bin->lock); | |
970d7e83 LB |
3277 | bstats[i].nmalloc += bin->stats.nmalloc; |
3278 | bstats[i].ndalloc += bin->stats.ndalloc; | |
3279 | bstats[i].nrequests += bin->stats.nrequests; | |
54a0048b | 3280 | bstats[i].curregs += bin->stats.curregs; |
970d7e83 LB |
3281 | if (config_tcache) { |
3282 | bstats[i].nfills += bin->stats.nfills; | |
3283 | bstats[i].nflushes += bin->stats.nflushes; | |
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 | ||
54a0048b SL |
3292 | unsigned |
3293 | arena_nthreads_get(arena_t *arena) | |
3294 | { | |
3295 | ||
3296 | return (atomic_read_u(&arena->nthreads)); | |
3297 | } | |
3298 | ||
3299 | void | |
3300 | arena_nthreads_inc(arena_t *arena) | |
3301 | { | |
3302 | ||
3303 | atomic_add_u(&arena->nthreads, 1); | |
3304 | } | |
3305 | ||
3306 | void | |
3307 | arena_nthreads_dec(arena_t *arena) | |
970d7e83 | 3308 | { |
54a0048b SL |
3309 | |
3310 | atomic_sub_u(&arena->nthreads, 1); | |
3311 | } | |
3312 | ||
3313 | arena_t * | |
3314 | arena_new(unsigned ind) | |
3315 | { | |
3316 | arena_t *arena; | |
3317 | size_t arena_size; | |
970d7e83 LB |
3318 | unsigned i; |
3319 | arena_bin_t *bin; | |
3320 | ||
54a0048b SL |
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 | ||
970d7e83 LB |
3337 | arena->ind = ind; |
3338 | arena->nthreads = 0; | |
970d7e83 | 3339 | if (malloc_mutex_init(&arena->lock)) |
54a0048b | 3340 | return (NULL); |
970d7e83 LB |
3341 | |
3342 | if (config_stats) { | |
3343 | memset(&arena->stats, 0, sizeof(arena_stats_t)); | |
54a0048b SL |
3344 | arena->stats.lstats = (malloc_large_stats_t *)((uintptr_t)arena |
3345 | + CACHELINE_CEILING(arena_size)); | |
970d7e83 LB |
3346 | memset(arena->stats.lstats, 0, nlclasses * |
3347 | sizeof(malloc_large_stats_t)); | |
54a0048b SL |
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)); | |
970d7e83 LB |
3353 | if (config_tcache) |
3354 | ql_new(&arena->tcache_ql); | |
3355 | } | |
3356 | ||
3357 | if (config_prof) | |
3358 | arena->prof_accumbytes = 0; | |
3359 | ||
54a0048b SL |
3360 | if (config_cache_oblivious) { |
3361 | /* | |
3362 | * A nondeterministic seed based on the address of arena reduces | |
3363 | * the likelihood of lockstep non-uniform cache index | |
3364 | * utilization among identical concurrent processes, but at the | |
3365 | * cost of test repeatability. For debug builds, instead use a | |
3366 | * deterministic seed. | |
3367 | */ | |
3368 | arena->offset_state = config_debug ? ind : | |
3369 | (uint64_t)(uintptr_t)arena; | |
3370 | } | |
3371 | ||
970d7e83 LB |
3372 | arena->dss_prec = chunk_dss_prec_get(); |
3373 | ||
970d7e83 LB |
3374 | arena->spare = NULL; |
3375 | ||
54a0048b SL |
3376 | arena->lg_dirty_mult = arena_lg_dirty_mult_default_get(); |
3377 | arena->purging = false; | |
970d7e83 LB |
3378 | arena->nactive = 0; |
3379 | arena->ndirty = 0; | |
970d7e83 | 3380 | |
54a0048b SL |
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); | |
3397 | if (malloc_mutex_init(&arena->chunks_mtx)) | |
3398 | return (NULL); | |
3399 | ql_new(&arena->node_cache); | |
3400 | if (malloc_mutex_init(&arena->node_cache_mtx)) | |
3401 | return (NULL); | |
3402 | ||
3403 | arena->chunk_hooks = chunk_hooks_default; | |
970d7e83 LB |
3404 | |
3405 | /* Initialize bins. */ | |
3406 | for (i = 0; i < NBINS; i++) { | |
3407 | bin = &arena->bins[i]; | |
3408 | if (malloc_mutex_init(&bin->lock)) | |
54a0048b | 3409 | return (NULL); |
970d7e83 LB |
3410 | bin->runcur = NULL; |
3411 | arena_run_tree_new(&bin->runs); | |
3412 | if (config_stats) | |
3413 | memset(&bin->stats, 0, sizeof(malloc_bin_stats_t)); | |
3414 | } | |
3415 | ||
54a0048b | 3416 | return (arena); |
970d7e83 LB |
3417 | } |
3418 | ||
3419 | /* | |
3420 | * Calculate bin_info->run_size such that it meets the following constraints: | |
3421 | * | |
54a0048b | 3422 | * *) bin_info->run_size <= arena_maxrun |
970d7e83 LB |
3423 | * *) bin_info->nregs <= RUN_MAXREGS |
3424 | * | |
1a4d82fc JJ |
3425 | * bin_info->nregs and bin_info->reg0_offset are also calculated here, since |
3426 | * these settings are all interdependent. | |
970d7e83 | 3427 | */ |
1a4d82fc JJ |
3428 | static void |
3429 | bin_info_run_size_calc(arena_bin_info_t *bin_info) | |
970d7e83 LB |
3430 | { |
3431 | size_t pad_size; | |
1a4d82fc JJ |
3432 | size_t try_run_size, perfect_run_size, actual_run_size; |
3433 | uint32_t try_nregs, perfect_nregs, actual_nregs; | |
970d7e83 LB |
3434 | |
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 | */ | |
1a4d82fc | 3442 | if (config_fill && unlikely(opt_redzone)) { |
54a0048b | 3443 | size_t align_min = ZU(1) << (ffs_zu(bin_info->reg_size) - 1); |
970d7e83 LB |
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 | /* | |
1a4d82fc JJ |
3459 | * Compute run size under ideal conditions (no redzones, no limit on run |
3460 | * size). | |
970d7e83 | 3461 | */ |
1a4d82fc | 3462 | try_run_size = PAGE; |
54a0048b | 3463 | try_nregs = (uint32_t)(try_run_size / bin_info->reg_size); |
970d7e83 | 3464 | do { |
1a4d82fc JJ |
3465 | perfect_run_size = try_run_size; |
3466 | perfect_nregs = try_nregs; | |
970d7e83 | 3467 | |
970d7e83 | 3468 | try_run_size += PAGE; |
54a0048b | 3469 | try_nregs = (uint32_t)(try_run_size / bin_info->reg_size); |
1a4d82fc JJ |
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; | |
54a0048b SL |
3474 | actual_nregs = (uint32_t)((actual_run_size - pad_size) / |
3475 | bin_info->reg_interval); | |
1a4d82fc JJ |
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; | |
54a0048b SL |
3487 | actual_nregs = (uint32_t)((actual_run_size - pad_size) / |
3488 | bin_info->reg_interval); | |
1a4d82fc JJ |
3489 | } |
3490 | ||
3491 | /* | |
3492 | * Make sure that the run will fit within an arena chunk. | |
3493 | */ | |
54a0048b | 3494 | while (actual_run_size > arena_maxrun) { |
1a4d82fc | 3495 | actual_run_size -= PAGE; |
54a0048b SL |
3496 | actual_nregs = (uint32_t)((actual_run_size - pad_size) / |
3497 | bin_info->reg_interval); | |
1a4d82fc JJ |
3498 | } |
3499 | assert(actual_nregs > 0); | |
54a0048b | 3500 | assert(actual_run_size == s2u(actual_run_size)); |
970d7e83 LB |
3501 | |
3502 | /* Copy final settings. */ | |
1a4d82fc JJ |
3503 | bin_info->run_size = actual_run_size; |
3504 | bin_info->nregs = actual_nregs; | |
54a0048b SL |
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; | |
970d7e83 LB |
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); | |
970d7e83 LB |
3513 | } |
3514 | ||
3515 | static void | |
3516 | bin_info_init(void) | |
3517 | { | |
3518 | arena_bin_info_t *bin_info; | |
970d7e83 | 3519 | |
54a0048b | 3520 | #define BIN_INFO_INIT_bin_yes(index, size) \ |
1a4d82fc | 3521 | bin_info = &arena_bin_info[index]; \ |
970d7e83 | 3522 | bin_info->reg_size = size; \ |
1a4d82fc | 3523 | bin_info_run_size_calc(bin_info); \ |
970d7e83 | 3524 | bitmap_info_init(&bin_info->bitmap_info, bin_info->nregs); |
1a4d82fc JJ |
3525 | #define BIN_INFO_INIT_bin_no(index, size) |
3526 | #define SC(index, lg_grp, lg_delta, ndelta, bin, lg_delta_lookup) \ | |
3527 | BIN_INFO_INIT_bin_##bin(index, (ZU(1)<<lg_grp) + (ZU(ndelta)<<lg_delta)) | |
970d7e83 | 3528 | SIZE_CLASSES |
1a4d82fc JJ |
3529 | #undef BIN_INFO_INIT_bin_yes |
3530 | #undef BIN_INFO_INIT_bin_no | |
3531 | #undef SC | |
970d7e83 LB |
3532 | } |
3533 | ||
54a0048b SL |
3534 | static bool |
3535 | small_run_size_init(void) | |
3536 | { | |
3537 | ||
3538 | assert(small_maxrun != 0); | |
3539 | ||
3540 | small_run_tab = (bool *)base_alloc(sizeof(bool) * (small_maxrun >> | |
3541 | LG_PAGE)); | |
3542 | if (small_run_tab == NULL) | |
3543 | return (true); | |
3544 | ||
3545 | #define TAB_INIT_bin_yes(index, size) { \ | |
3546 | arena_bin_info_t *bin_info = &arena_bin_info[index]; \ | |
3547 | small_run_tab[bin_info->run_size >> LG_PAGE] = true; \ | |
3548 | } | |
3549 | #define TAB_INIT_bin_no(index, size) | |
3550 | #define SC(index, lg_grp, lg_delta, ndelta, bin, lg_delta_lookup) \ | |
3551 | TAB_INIT_bin_##bin(index, (ZU(1)<<lg_grp) + (ZU(ndelta)<<lg_delta)) | |
3552 | SIZE_CLASSES | |
3553 | #undef TAB_INIT_bin_yes | |
3554 | #undef TAB_INIT_bin_no | |
3555 | #undef SC | |
3556 | ||
3557 | return (false); | |
3558 | } | |
3559 | ||
3560 | static bool | |
3561 | run_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 | ||
3589 | bool | |
970d7e83 LB |
3590 | arena_boot(void) |
3591 | { | |
970d7e83 LB |
3592 | unsigned i; |
3593 | ||
54a0048b SL |
3594 | arena_lg_dirty_mult_default_set(opt_lg_dirty_mult); |
3595 | arena_decay_time_default_set(opt_decay_time); | |
3596 | ||
970d7e83 LB |
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. | |
3604 | * 2) Refine map_bias based on (1) to omit the header pages in the page | |
3605 | * map. The resulting map_bias may be one too small. | |
3606 | * 3) Refine map_bias based on (2). The result will be >= the result | |
3607 | * from (2), and will always be correct. | |
3608 | */ | |
3609 | map_bias = 0; | |
3610 | for (i = 0; i < 3; i++) { | |
54a0048b | 3611 | size_t header_size = offsetof(arena_chunk_t, map_bits) + |
1a4d82fc JJ |
3612 | ((sizeof(arena_chunk_map_bits_t) + |
3613 | sizeof(arena_chunk_map_misc_t)) * (chunk_npages-map_bias)); | |
3614 | map_bias = (header_size + PAGE_MASK) >> LG_PAGE; | |
970d7e83 LB |
3615 | } |
3616 | assert(map_bias > 0); | |
3617 | ||
1a4d82fc JJ |
3618 | map_misc_offset = offsetof(arena_chunk_t, map_bits) + |
3619 | sizeof(arena_chunk_map_bits_t) * (chunk_npages-map_bias); | |
3620 | ||
54a0048b SL |
3621 | arena_maxrun = chunksize - (map_bias << LG_PAGE); |
3622 | assert(arena_maxrun > 0); | |
3623 | large_maxclass = index2size(size2index(chunksize)-1); | |
3624 | if (large_maxclass > arena_maxrun) { | |
3625 | /* | |
3626 | * For small chunk sizes it's possible for there to be fewer | |
3627 | * non-header pages available than are necessary to serve the | |
3628 | * size classes just below chunksize. | |
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; | |
970d7e83 LB |
3635 | |
3636 | bin_info_init(); | |
54a0048b SL |
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); | |
970d7e83 LB |
3646 | } |
3647 | ||
3648 | void | |
3649 | arena_prefork(arena_t *arena) | |
3650 | { | |
3651 | unsigned i; | |
3652 | ||
3653 | malloc_mutex_prefork(&arena->lock); | |
54a0048b SL |
3654 | malloc_mutex_prefork(&arena->huge_mtx); |
3655 | malloc_mutex_prefork(&arena->chunks_mtx); | |
3656 | malloc_mutex_prefork(&arena->node_cache_mtx); | |
970d7e83 LB |
3657 | for (i = 0; i < NBINS; i++) |
3658 | malloc_mutex_prefork(&arena->bins[i].lock); | |
3659 | } | |
3660 | ||
3661 | void | |
3662 | arena_postfork_parent(arena_t *arena) | |
3663 | { | |
3664 | unsigned i; | |
3665 | ||
3666 | for (i = 0; i < NBINS; i++) | |
3667 | malloc_mutex_postfork_parent(&arena->bins[i].lock); | |
54a0048b SL |
3668 | malloc_mutex_postfork_parent(&arena->node_cache_mtx); |
3669 | malloc_mutex_postfork_parent(&arena->chunks_mtx); | |
3670 | malloc_mutex_postfork_parent(&arena->huge_mtx); | |
970d7e83 LB |
3671 | malloc_mutex_postfork_parent(&arena->lock); |
3672 | } | |
3673 | ||
3674 | void | |
3675 | arena_postfork_child(arena_t *arena) | |
3676 | { | |
3677 | unsigned i; | |
3678 | ||
3679 | for (i = 0; i < NBINS; i++) | |
3680 | malloc_mutex_postfork_child(&arena->bins[i].lock); | |
54a0048b SL |
3681 | malloc_mutex_postfork_child(&arena->node_cache_mtx); |
3682 | malloc_mutex_postfork_child(&arena->chunks_mtx); | |
3683 | malloc_mutex_postfork_child(&arena->huge_mtx); | |
970d7e83 LB |
3684 | malloc_mutex_postfork_child(&arena->lock); |
3685 | } |