xref: /freebsd/contrib/jemalloc/src/arena.c (revision 640235e2c2ba32947f7c59d168437ffa1280f1e6)
1 #define	JEMALLOC_ARENA_C_
2 #include "jemalloc/internal/jemalloc_internal.h"
3 
4 /******************************************************************************/
5 /* Data. */
6 
7 purge_mode_t	opt_purge = PURGE_DEFAULT;
8 const char	*purge_mode_names[] = {
9 	"ratio",
10 	"decay",
11 	"N/A"
12 };
13 ssize_t		opt_lg_dirty_mult = LG_DIRTY_MULT_DEFAULT;
14 static ssize_t	lg_dirty_mult_default;
15 ssize_t		opt_decay_time = DECAY_TIME_DEFAULT;
16 static ssize_t	decay_time_default;
17 
18 arena_bin_info_t	arena_bin_info[NBINS];
19 
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. */
33 
34 /******************************************************************************/
35 /*
36  * Function prototypes for static functions that are referenced prior to
37  * definition.
38  */
39 
40 static void	arena_purge_to_limit(tsdn_t *tsdn, arena_t *arena,
41     size_t ndirty_limit);
42 static void	arena_run_dalloc(tsdn_t *tsdn, arena_t *arena, arena_run_t *run,
43     bool dirty, bool cleaned, bool decommitted);
44 static void	arena_dalloc_bin_run(tsdn_t *tsdn, arena_t *arena,
45     arena_chunk_t *chunk, arena_run_t *run, arena_bin_t *bin);
46 static void	arena_bin_lower_run(arena_t *arena, arena_chunk_t *chunk,
47     arena_run_t *run, arena_bin_t *bin);
48 
49 /******************************************************************************/
50 
51 JEMALLOC_INLINE_C size_t
52 arena_miscelm_size_get(const arena_chunk_map_misc_t *miscelm)
53 {
54 	arena_chunk_t *chunk;
55 	size_t pageind, mapbits;
56 
57 	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(miscelm);
58 	pageind = arena_miscelm_to_pageind(miscelm);
59 	mapbits = arena_mapbits_get(chunk, pageind);
60 	return (arena_mapbits_size_decode(mapbits));
61 }
62 
63 JEMALLOC_INLINE_C int
64 arena_run_addr_comp(const arena_chunk_map_misc_t *a,
65     const arena_chunk_map_misc_t *b)
66 {
67 	uintptr_t a_miscelm = (uintptr_t)a;
68 	uintptr_t b_miscelm = (uintptr_t)b;
69 
70 	assert(a != NULL);
71 	assert(b != NULL);
72 
73 	return ((a_miscelm > b_miscelm) - (a_miscelm < b_miscelm));
74 }
75 
76 /* Generate pairing heap functions. */
77 ph_gen(static UNUSED, arena_run_heap_, arena_run_heap_t, arena_chunk_map_misc_t,
78     ph_link, arena_run_addr_comp)
79 
80 static size_t
81 run_quantize_floor_compute(size_t size)
82 {
83 	size_t qsize;
84 
85 	assert(size != 0);
86 	assert(size == PAGE_CEILING(size));
87 
88 	/* Don't change sizes that are valid small run sizes. */
89 	if (size <= small_maxrun && small_run_tab[size >> LG_PAGE])
90 		return (size);
91 
92 	/*
93 	 * Round down to the nearest run size that can actually be requested
94 	 * during normal large allocation.  Add large_pad so that cache index
95 	 * randomization can offset the allocation from the page boundary.
96 	 */
97 	qsize = index2size(size2index(size - large_pad + 1) - 1) + large_pad;
98 	if (qsize <= SMALL_MAXCLASS + large_pad)
99 		return (run_quantize_floor_compute(size - large_pad));
100 	assert(qsize <= size);
101 	return (qsize);
102 }
103 
104 static size_t
105 run_quantize_ceil_compute_hard(size_t size)
106 {
107 	size_t large_run_size_next;
108 
109 	assert(size != 0);
110 	assert(size == PAGE_CEILING(size));
111 
112 	/*
113 	 * Return the next quantized size greater than the input size.
114 	 * Quantized sizes comprise the union of run sizes that back small
115 	 * region runs, and run sizes that back large regions with no explicit
116 	 * alignment constraints.
117 	 */
118 
119 	if (size > SMALL_MAXCLASS) {
120 		large_run_size_next = PAGE_CEILING(index2size(size2index(size -
121 		    large_pad) + 1) + large_pad);
122 	} else
123 		large_run_size_next = SIZE_T_MAX;
124 	if (size >= small_maxrun)
125 		return (large_run_size_next);
126 
127 	while (true) {
128 		size += PAGE;
129 		assert(size <= small_maxrun);
130 		if (small_run_tab[size >> LG_PAGE]) {
131 			if (large_run_size_next < size)
132 				return (large_run_size_next);
133 			return (size);
134 		}
135 	}
136 }
137 
138 static size_t
139 run_quantize_ceil_compute(size_t size)
140 {
141 	size_t qsize = run_quantize_floor_compute(size);
142 
143 	if (qsize < size) {
144 		/*
145 		 * Skip a quantization that may have an adequately large run,
146 		 * because under-sized runs may be mixed in.  This only happens
147 		 * when an unusual size is requested, i.e. for aligned
148 		 * allocation, and is just one of several places where linear
149 		 * search would potentially find sufficiently aligned available
150 		 * memory somewhere lower.
151 		 */
152 		qsize = run_quantize_ceil_compute_hard(qsize);
153 	}
154 	return (qsize);
155 }
156 
157 #ifdef JEMALLOC_JET
158 #undef run_quantize_floor
159 #define	run_quantize_floor JEMALLOC_N(n_run_quantize_floor)
160 #endif
161 static size_t
162 run_quantize_floor(size_t size)
163 {
164 	size_t ret;
165 
166 	assert(size > 0);
167 	assert(size <= run_quantize_max);
168 	assert((size & PAGE_MASK) == 0);
169 
170 	ret = run_quantize_floor_tab[(size >> LG_PAGE) - 1];
171 	assert(ret == run_quantize_floor_compute(size));
172 	return (ret);
173 }
174 #ifdef JEMALLOC_JET
175 #undef run_quantize_floor
176 #define	run_quantize_floor JEMALLOC_N(run_quantize_floor)
177 run_quantize_t *run_quantize_floor = JEMALLOC_N(n_run_quantize_floor);
178 #endif
179 
180 #ifdef JEMALLOC_JET
181 #undef run_quantize_ceil
182 #define	run_quantize_ceil JEMALLOC_N(n_run_quantize_ceil)
183 #endif
184 static size_t
185 run_quantize_ceil(size_t size)
186 {
187 	size_t ret;
188 
189 	assert(size > 0);
190 	assert(size <= run_quantize_max);
191 	assert((size & PAGE_MASK) == 0);
192 
193 	ret = run_quantize_ceil_tab[(size >> LG_PAGE) - 1];
194 	assert(ret == run_quantize_ceil_compute(size));
195 	return (ret);
196 }
197 #ifdef JEMALLOC_JET
198 #undef run_quantize_ceil
199 #define	run_quantize_ceil JEMALLOC_N(run_quantize_ceil)
200 run_quantize_t *run_quantize_ceil = JEMALLOC_N(n_run_quantize_ceil);
201 #endif
202 
203 static arena_run_heap_t *
204 arena_runs_avail_get(arena_t *arena, szind_t ind)
205 {
206 
207 	assert(ind >= runs_avail_bias);
208 	assert(ind - runs_avail_bias < runs_avail_nclasses);
209 
210 	return (&arena->runs_avail[ind - runs_avail_bias]);
211 }
212 
213 static void
214 arena_avail_insert(arena_t *arena, arena_chunk_t *chunk, size_t pageind,
215     size_t npages)
216 {
217 	szind_t ind = size2index(run_quantize_floor(arena_miscelm_size_get(
218 	    arena_miscelm_get_const(chunk, pageind))));
219 	assert(npages == (arena_mapbits_unallocated_size_get(chunk, pageind) >>
220 	    LG_PAGE));
221 	arena_run_heap_insert(arena_runs_avail_get(arena, ind),
222 	    arena_miscelm_get_mutable(chunk, pageind));
223 }
224 
225 static void
226 arena_avail_remove(arena_t *arena, arena_chunk_t *chunk, size_t pageind,
227     size_t npages)
228 {
229 	szind_t ind = size2index(run_quantize_floor(arena_miscelm_size_get(
230 	    arena_miscelm_get_const(chunk, pageind))));
231 	assert(npages == (arena_mapbits_unallocated_size_get(chunk, pageind) >>
232 	    LG_PAGE));
233 	arena_run_heap_remove(arena_runs_avail_get(arena, ind),
234 	    arena_miscelm_get_mutable(chunk, pageind));
235 }
236 
237 static void
238 arena_run_dirty_insert(arena_t *arena, arena_chunk_t *chunk, size_t pageind,
239     size_t npages)
240 {
241 	arena_chunk_map_misc_t *miscelm = arena_miscelm_get_mutable(chunk,
242 	    pageind);
243 
244 	assert(npages == (arena_mapbits_unallocated_size_get(chunk, pageind) >>
245 	    LG_PAGE));
246 	assert(arena_mapbits_dirty_get(chunk, pageind) == CHUNK_MAP_DIRTY);
247 	assert(arena_mapbits_dirty_get(chunk, pageind+npages-1) ==
248 	    CHUNK_MAP_DIRTY);
249 
250 	qr_new(&miscelm->rd, rd_link);
251 	qr_meld(&arena->runs_dirty, &miscelm->rd, rd_link);
252 	arena->ndirty += npages;
253 }
254 
255 static void
256 arena_run_dirty_remove(arena_t *arena, arena_chunk_t *chunk, size_t pageind,
257     size_t npages)
258 {
259 	arena_chunk_map_misc_t *miscelm = arena_miscelm_get_mutable(chunk,
260 	    pageind);
261 
262 	assert(npages == (arena_mapbits_unallocated_size_get(chunk, pageind) >>
263 	    LG_PAGE));
264 	assert(arena_mapbits_dirty_get(chunk, pageind) == CHUNK_MAP_DIRTY);
265 	assert(arena_mapbits_dirty_get(chunk, pageind+npages-1) ==
266 	    CHUNK_MAP_DIRTY);
267 
268 	qr_remove(&miscelm->rd, rd_link);
269 	assert(arena->ndirty >= npages);
270 	arena->ndirty -= npages;
271 }
272 
273 static size_t
274 arena_chunk_dirty_npages(const extent_node_t *node)
275 {
276 
277 	return (extent_node_size_get(node) >> LG_PAGE);
278 }
279 
280 void
281 arena_chunk_cache_maybe_insert(arena_t *arena, extent_node_t *node, bool cache)
282 {
283 
284 	if (cache) {
285 		extent_node_dirty_linkage_init(node);
286 		extent_node_dirty_insert(node, &arena->runs_dirty,
287 		    &arena->chunks_cache);
288 		arena->ndirty += arena_chunk_dirty_npages(node);
289 	}
290 }
291 
292 void
293 arena_chunk_cache_maybe_remove(arena_t *arena, extent_node_t *node, bool dirty)
294 {
295 
296 	if (dirty) {
297 		extent_node_dirty_remove(node);
298 		assert(arena->ndirty >= arena_chunk_dirty_npages(node));
299 		arena->ndirty -= arena_chunk_dirty_npages(node);
300 	}
301 }
302 
303 JEMALLOC_INLINE_C void *
304 arena_run_reg_alloc(arena_run_t *run, arena_bin_info_t *bin_info)
305 {
306 	void *ret;
307 	size_t regind;
308 	arena_chunk_map_misc_t *miscelm;
309 	void *rpages;
310 
311 	assert(run->nfree > 0);
312 	assert(!bitmap_full(run->bitmap, &bin_info->bitmap_info));
313 
314 	regind = (unsigned)bitmap_sfu(run->bitmap, &bin_info->bitmap_info);
315 	miscelm = arena_run_to_miscelm(run);
316 	rpages = arena_miscelm_to_rpages(miscelm);
317 	ret = (void *)((uintptr_t)rpages + (uintptr_t)bin_info->reg0_offset +
318 	    (uintptr_t)(bin_info->reg_interval * regind));
319 	run->nfree--;
320 	return (ret);
321 }
322 
323 JEMALLOC_INLINE_C void
324 arena_run_reg_dalloc(arena_run_t *run, void *ptr)
325 {
326 	arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run);
327 	size_t pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
328 	size_t mapbits = arena_mapbits_get(chunk, pageind);
329 	szind_t binind = arena_ptr_small_binind_get(ptr, mapbits);
330 	arena_bin_info_t *bin_info = &arena_bin_info[binind];
331 	size_t regind = arena_run_regind(run, bin_info, ptr);
332 
333 	assert(run->nfree < bin_info->nregs);
334 	/* Freeing an interior pointer can cause assertion failure. */
335 	assert(((uintptr_t)ptr -
336 	    ((uintptr_t)arena_miscelm_to_rpages(arena_run_to_miscelm(run)) +
337 	    (uintptr_t)bin_info->reg0_offset)) %
338 	    (uintptr_t)bin_info->reg_interval == 0);
339 	assert((uintptr_t)ptr >=
340 	    (uintptr_t)arena_miscelm_to_rpages(arena_run_to_miscelm(run)) +
341 	    (uintptr_t)bin_info->reg0_offset);
342 	/* Freeing an unallocated pointer can cause assertion failure. */
343 	assert(bitmap_get(run->bitmap, &bin_info->bitmap_info, regind));
344 
345 	bitmap_unset(run->bitmap, &bin_info->bitmap_info, regind);
346 	run->nfree++;
347 }
348 
349 JEMALLOC_INLINE_C void
350 arena_run_zero(arena_chunk_t *chunk, size_t run_ind, size_t npages)
351 {
352 
353 	JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED((void *)((uintptr_t)chunk +
354 	    (run_ind << LG_PAGE)), (npages << LG_PAGE));
355 	memset((void *)((uintptr_t)chunk + (run_ind << LG_PAGE)), 0,
356 	    (npages << LG_PAGE));
357 }
358 
359 JEMALLOC_INLINE_C void
360 arena_run_page_mark_zeroed(arena_chunk_t *chunk, size_t run_ind)
361 {
362 
363 	JEMALLOC_VALGRIND_MAKE_MEM_DEFINED((void *)((uintptr_t)chunk + (run_ind
364 	    << LG_PAGE)), PAGE);
365 }
366 
367 JEMALLOC_INLINE_C void
368 arena_run_page_validate_zeroed(arena_chunk_t *chunk, size_t run_ind)
369 {
370 	size_t i;
371 	UNUSED size_t *p = (size_t *)((uintptr_t)chunk + (run_ind << LG_PAGE));
372 
373 	arena_run_page_mark_zeroed(chunk, run_ind);
374 	for (i = 0; i < PAGE / sizeof(size_t); i++)
375 		assert(p[i] == 0);
376 }
377 
378 static void
379 arena_nactive_add(arena_t *arena, size_t add_pages)
380 {
381 
382 	if (config_stats) {
383 		size_t cactive_add = CHUNK_CEILING((arena->nactive +
384 		    add_pages) << LG_PAGE) - CHUNK_CEILING(arena->nactive <<
385 		    LG_PAGE);
386 		if (cactive_add != 0)
387 			stats_cactive_add(cactive_add);
388 	}
389 	arena->nactive += add_pages;
390 }
391 
392 static void
393 arena_nactive_sub(arena_t *arena, size_t sub_pages)
394 {
395 
396 	if (config_stats) {
397 		size_t cactive_sub = CHUNK_CEILING(arena->nactive << LG_PAGE) -
398 		    CHUNK_CEILING((arena->nactive - sub_pages) << LG_PAGE);
399 		if (cactive_sub != 0)
400 			stats_cactive_sub(cactive_sub);
401 	}
402 	arena->nactive -= sub_pages;
403 }
404 
405 static void
406 arena_run_split_remove(arena_t *arena, arena_chunk_t *chunk, size_t run_ind,
407     size_t flag_dirty, size_t flag_decommitted, size_t need_pages)
408 {
409 	size_t total_pages, rem_pages;
410 
411 	assert(flag_dirty == 0 || flag_decommitted == 0);
412 
413 	total_pages = arena_mapbits_unallocated_size_get(chunk, run_ind) >>
414 	    LG_PAGE;
415 	assert(arena_mapbits_dirty_get(chunk, run_ind+total_pages-1) ==
416 	    flag_dirty);
417 	assert(need_pages <= total_pages);
418 	rem_pages = total_pages - need_pages;
419 
420 	arena_avail_remove(arena, chunk, run_ind, total_pages);
421 	if (flag_dirty != 0)
422 		arena_run_dirty_remove(arena, chunk, run_ind, total_pages);
423 	arena_nactive_add(arena, need_pages);
424 
425 	/* Keep track of trailing unused pages for later use. */
426 	if (rem_pages > 0) {
427 		size_t flags = flag_dirty | flag_decommitted;
428 		size_t flag_unzeroed_mask = (flags == 0) ?  CHUNK_MAP_UNZEROED :
429 		    0;
430 
431 		arena_mapbits_unallocated_set(chunk, run_ind+need_pages,
432 		    (rem_pages << LG_PAGE), flags |
433 		    (arena_mapbits_unzeroed_get(chunk, run_ind+need_pages) &
434 		    flag_unzeroed_mask));
435 		arena_mapbits_unallocated_set(chunk, run_ind+total_pages-1,
436 		    (rem_pages << LG_PAGE), flags |
437 		    (arena_mapbits_unzeroed_get(chunk, run_ind+total_pages-1) &
438 		    flag_unzeroed_mask));
439 		if (flag_dirty != 0) {
440 			arena_run_dirty_insert(arena, chunk, run_ind+need_pages,
441 			    rem_pages);
442 		}
443 		arena_avail_insert(arena, chunk, run_ind+need_pages, rem_pages);
444 	}
445 }
446 
447 static bool
448 arena_run_split_large_helper(arena_t *arena, arena_run_t *run, size_t size,
449     bool remove, bool zero)
450 {
451 	arena_chunk_t *chunk;
452 	arena_chunk_map_misc_t *miscelm;
453 	size_t flag_dirty, flag_decommitted, run_ind, need_pages;
454 	size_t flag_unzeroed_mask;
455 
456 	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run);
457 	miscelm = arena_run_to_miscelm(run);
458 	run_ind = arena_miscelm_to_pageind(miscelm);
459 	flag_dirty = arena_mapbits_dirty_get(chunk, run_ind);
460 	flag_decommitted = arena_mapbits_decommitted_get(chunk, run_ind);
461 	need_pages = (size >> LG_PAGE);
462 	assert(need_pages > 0);
463 
464 	if (flag_decommitted != 0 && arena->chunk_hooks.commit(chunk, chunksize,
465 	    run_ind << LG_PAGE, size, arena->ind))
466 		return (true);
467 
468 	if (remove) {
469 		arena_run_split_remove(arena, chunk, run_ind, flag_dirty,
470 		    flag_decommitted, need_pages);
471 	}
472 
473 	if (zero) {
474 		if (flag_decommitted != 0) {
475 			/* The run is untouched, and therefore zeroed. */
476 			JEMALLOC_VALGRIND_MAKE_MEM_DEFINED((void
477 			    *)((uintptr_t)chunk + (run_ind << LG_PAGE)),
478 			    (need_pages << LG_PAGE));
479 		} else if (flag_dirty != 0) {
480 			/* The run is dirty, so all pages must be zeroed. */
481 			arena_run_zero(chunk, run_ind, need_pages);
482 		} else {
483 			/*
484 			 * The run is clean, so some pages may be zeroed (i.e.
485 			 * never before touched).
486 			 */
487 			size_t i;
488 			for (i = 0; i < need_pages; i++) {
489 				if (arena_mapbits_unzeroed_get(chunk, run_ind+i)
490 				    != 0)
491 					arena_run_zero(chunk, run_ind+i, 1);
492 				else if (config_debug) {
493 					arena_run_page_validate_zeroed(chunk,
494 					    run_ind+i);
495 				} else {
496 					arena_run_page_mark_zeroed(chunk,
497 					    run_ind+i);
498 				}
499 			}
500 		}
501 	} else {
502 		JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED((void *)((uintptr_t)chunk +
503 		    (run_ind << LG_PAGE)), (need_pages << LG_PAGE));
504 	}
505 
506 	/*
507 	 * Set the last element first, in case the run only contains one page
508 	 * (i.e. both statements set the same element).
509 	 */
510 	flag_unzeroed_mask = (flag_dirty | flag_decommitted) == 0 ?
511 	    CHUNK_MAP_UNZEROED : 0;
512 	arena_mapbits_large_set(chunk, run_ind+need_pages-1, 0, flag_dirty |
513 	    (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk,
514 	    run_ind+need_pages-1)));
515 	arena_mapbits_large_set(chunk, run_ind, size, flag_dirty |
516 	    (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk, run_ind)));
517 	return (false);
518 }
519 
520 static bool
521 arena_run_split_large(arena_t *arena, arena_run_t *run, size_t size, bool zero)
522 {
523 
524 	return (arena_run_split_large_helper(arena, run, size, true, zero));
525 }
526 
527 static bool
528 arena_run_init_large(arena_t *arena, arena_run_t *run, size_t size, bool zero)
529 {
530 
531 	return (arena_run_split_large_helper(arena, run, size, false, zero));
532 }
533 
534 static bool
535 arena_run_split_small(arena_t *arena, arena_run_t *run, size_t size,
536     szind_t binind)
537 {
538 	arena_chunk_t *chunk;
539 	arena_chunk_map_misc_t *miscelm;
540 	size_t flag_dirty, flag_decommitted, run_ind, need_pages, i;
541 
542 	assert(binind != BININD_INVALID);
543 
544 	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run);
545 	miscelm = arena_run_to_miscelm(run);
546 	run_ind = arena_miscelm_to_pageind(miscelm);
547 	flag_dirty = arena_mapbits_dirty_get(chunk, run_ind);
548 	flag_decommitted = arena_mapbits_decommitted_get(chunk, run_ind);
549 	need_pages = (size >> LG_PAGE);
550 	assert(need_pages > 0);
551 
552 	if (flag_decommitted != 0 && arena->chunk_hooks.commit(chunk, chunksize,
553 	    run_ind << LG_PAGE, size, arena->ind))
554 		return (true);
555 
556 	arena_run_split_remove(arena, chunk, run_ind, flag_dirty,
557 	    flag_decommitted, need_pages);
558 
559 	for (i = 0; i < need_pages; i++) {
560 		size_t flag_unzeroed = arena_mapbits_unzeroed_get(chunk,
561 		    run_ind+i);
562 		arena_mapbits_small_set(chunk, run_ind+i, i, binind,
563 		    flag_unzeroed);
564 		if (config_debug && flag_dirty == 0 && flag_unzeroed == 0)
565 			arena_run_page_validate_zeroed(chunk, run_ind+i);
566 	}
567 	JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED((void *)((uintptr_t)chunk +
568 	    (run_ind << LG_PAGE)), (need_pages << LG_PAGE));
569 	return (false);
570 }
571 
572 static arena_chunk_t *
573 arena_chunk_init_spare(arena_t *arena)
574 {
575 	arena_chunk_t *chunk;
576 
577 	assert(arena->spare != NULL);
578 
579 	chunk = arena->spare;
580 	arena->spare = NULL;
581 
582 	assert(arena_mapbits_allocated_get(chunk, map_bias) == 0);
583 	assert(arena_mapbits_allocated_get(chunk, chunk_npages-1) == 0);
584 	assert(arena_mapbits_unallocated_size_get(chunk, map_bias) ==
585 	    arena_maxrun);
586 	assert(arena_mapbits_unallocated_size_get(chunk, chunk_npages-1) ==
587 	    arena_maxrun);
588 	assert(arena_mapbits_dirty_get(chunk, map_bias) ==
589 	    arena_mapbits_dirty_get(chunk, chunk_npages-1));
590 
591 	return (chunk);
592 }
593 
594 static bool
595 arena_chunk_register(tsdn_t *tsdn, arena_t *arena, arena_chunk_t *chunk,
596     bool zero)
597 {
598 
599 	/*
600 	 * The extent node notion of "committed" doesn't directly apply to
601 	 * arena chunks.  Arbitrarily mark them as committed.  The commit state
602 	 * of runs is tracked individually, and upon chunk deallocation the
603 	 * entire chunk is in a consistent commit state.
604 	 */
605 	extent_node_init(&chunk->node, arena, chunk, chunksize, zero, true);
606 	extent_node_achunk_set(&chunk->node, true);
607 	return (chunk_register(tsdn, chunk, &chunk->node));
608 }
609 
610 static arena_chunk_t *
611 arena_chunk_alloc_internal_hard(tsdn_t *tsdn, arena_t *arena,
612     chunk_hooks_t *chunk_hooks, bool *zero, bool *commit)
613 {
614 	arena_chunk_t *chunk;
615 
616 	malloc_mutex_unlock(tsdn, &arena->lock);
617 
618 	chunk = (arena_chunk_t *)chunk_alloc_wrapper(tsdn, arena, chunk_hooks,
619 	    NULL, chunksize, chunksize, zero, commit);
620 	if (chunk != NULL && !*commit) {
621 		/* Commit header. */
622 		if (chunk_hooks->commit(chunk, chunksize, 0, map_bias <<
623 		    LG_PAGE, arena->ind)) {
624 			chunk_dalloc_wrapper(tsdn, arena, chunk_hooks,
625 			    (void *)chunk, chunksize, *zero, *commit);
626 			chunk = NULL;
627 		}
628 	}
629 	if (chunk != NULL && arena_chunk_register(tsdn, arena, chunk, *zero)) {
630 		if (!*commit) {
631 			/* Undo commit of header. */
632 			chunk_hooks->decommit(chunk, chunksize, 0, map_bias <<
633 			    LG_PAGE, arena->ind);
634 		}
635 		chunk_dalloc_wrapper(tsdn, arena, chunk_hooks, (void *)chunk,
636 		    chunksize, *zero, *commit);
637 		chunk = NULL;
638 	}
639 
640 	malloc_mutex_lock(tsdn, &arena->lock);
641 	return (chunk);
642 }
643 
644 static arena_chunk_t *
645 arena_chunk_alloc_internal(tsdn_t *tsdn, arena_t *arena, bool *zero,
646     bool *commit)
647 {
648 	arena_chunk_t *chunk;
649 	chunk_hooks_t chunk_hooks = CHUNK_HOOKS_INITIALIZER;
650 
651 	chunk = chunk_alloc_cache(tsdn, arena, &chunk_hooks, NULL, chunksize,
652 	    chunksize, zero, true);
653 	if (chunk != NULL) {
654 		if (arena_chunk_register(tsdn, arena, chunk, *zero)) {
655 			chunk_dalloc_cache(tsdn, arena, &chunk_hooks, chunk,
656 			    chunksize, true);
657 			return (NULL);
658 		}
659 		*commit = true;
660 	}
661 	if (chunk == NULL) {
662 		chunk = arena_chunk_alloc_internal_hard(tsdn, arena,
663 		    &chunk_hooks, zero, commit);
664 	}
665 
666 	if (config_stats && chunk != NULL) {
667 		arena->stats.mapped += chunksize;
668 		arena->stats.metadata_mapped += (map_bias << LG_PAGE);
669 	}
670 
671 	return (chunk);
672 }
673 
674 static arena_chunk_t *
675 arena_chunk_init_hard(tsdn_t *tsdn, arena_t *arena)
676 {
677 	arena_chunk_t *chunk;
678 	bool zero, commit;
679 	size_t flag_unzeroed, flag_decommitted, i;
680 
681 	assert(arena->spare == NULL);
682 
683 	zero = false;
684 	commit = false;
685 	chunk = arena_chunk_alloc_internal(tsdn, arena, &zero, &commit);
686 	if (chunk == NULL)
687 		return (NULL);
688 
689 	/*
690 	 * Initialize the map to contain one maximal free untouched run.  Mark
691 	 * the pages as zeroed if arena_chunk_alloc_internal() returned a zeroed
692 	 * or decommitted chunk.
693 	 */
694 	flag_unzeroed = (zero || !commit) ? 0 : CHUNK_MAP_UNZEROED;
695 	flag_decommitted = commit ? 0 : CHUNK_MAP_DECOMMITTED;
696 	arena_mapbits_unallocated_set(chunk, map_bias, arena_maxrun,
697 	    flag_unzeroed | flag_decommitted);
698 	/*
699 	 * There is no need to initialize the internal page map entries unless
700 	 * the chunk is not zeroed.
701 	 */
702 	if (!zero) {
703 		JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED(
704 		    (void *)arena_bitselm_get_const(chunk, map_bias+1),
705 		    (size_t)((uintptr_t)arena_bitselm_get_const(chunk,
706 		    chunk_npages-1) -
707 		    (uintptr_t)arena_bitselm_get_const(chunk, map_bias+1)));
708 		for (i = map_bias+1; i < chunk_npages-1; i++)
709 			arena_mapbits_internal_set(chunk, i, flag_unzeroed);
710 	} else {
711 		JEMALLOC_VALGRIND_MAKE_MEM_DEFINED((void
712 		    *)arena_bitselm_get_const(chunk, map_bias+1),
713 		    (size_t)((uintptr_t)arena_bitselm_get_const(chunk,
714 		    chunk_npages-1) -
715 		    (uintptr_t)arena_bitselm_get_const(chunk, map_bias+1)));
716 		if (config_debug) {
717 			for (i = map_bias+1; i < chunk_npages-1; i++) {
718 				assert(arena_mapbits_unzeroed_get(chunk, i) ==
719 				    flag_unzeroed);
720 			}
721 		}
722 	}
723 	arena_mapbits_unallocated_set(chunk, chunk_npages-1, arena_maxrun,
724 	    flag_unzeroed);
725 
726 	return (chunk);
727 }
728 
729 static arena_chunk_t *
730 arena_chunk_alloc(tsdn_t *tsdn, arena_t *arena)
731 {
732 	arena_chunk_t *chunk;
733 
734 	if (arena->spare != NULL)
735 		chunk = arena_chunk_init_spare(arena);
736 	else {
737 		chunk = arena_chunk_init_hard(tsdn, arena);
738 		if (chunk == NULL)
739 			return (NULL);
740 	}
741 
742 	ql_elm_new(&chunk->node, ql_link);
743 	ql_tail_insert(&arena->achunks, &chunk->node, ql_link);
744 	arena_avail_insert(arena, chunk, map_bias, chunk_npages-map_bias);
745 
746 	return (chunk);
747 }
748 
749 static void
750 arena_chunk_discard(tsdn_t *tsdn, arena_t *arena, arena_chunk_t *chunk)
751 {
752 	bool committed;
753 	chunk_hooks_t chunk_hooks = CHUNK_HOOKS_INITIALIZER;
754 
755 	chunk_deregister(chunk, &chunk->node);
756 
757 	committed = (arena_mapbits_decommitted_get(chunk, map_bias) == 0);
758 	if (!committed) {
759 		/*
760 		 * Decommit the header.  Mark the chunk as decommitted even if
761 		 * header decommit fails, since treating a partially committed
762 		 * chunk as committed has a high potential for causing later
763 		 * access of decommitted memory.
764 		 */
765 		chunk_hooks = chunk_hooks_get(tsdn, arena);
766 		chunk_hooks.decommit(chunk, chunksize, 0, map_bias << LG_PAGE,
767 		    arena->ind);
768 	}
769 
770 	chunk_dalloc_cache(tsdn, arena, &chunk_hooks, (void *)chunk, chunksize,
771 	    committed);
772 
773 	if (config_stats) {
774 		arena->stats.mapped -= chunksize;
775 		arena->stats.metadata_mapped -= (map_bias << LG_PAGE);
776 	}
777 }
778 
779 static void
780 arena_spare_discard(tsdn_t *tsdn, arena_t *arena, arena_chunk_t *spare)
781 {
782 
783 	assert(arena->spare != spare);
784 
785 	if (arena_mapbits_dirty_get(spare, map_bias) != 0) {
786 		arena_run_dirty_remove(arena, spare, map_bias,
787 		    chunk_npages-map_bias);
788 	}
789 
790 	arena_chunk_discard(tsdn, arena, spare);
791 }
792 
793 static void
794 arena_chunk_dalloc(tsdn_t *tsdn, arena_t *arena, arena_chunk_t *chunk)
795 {
796 	arena_chunk_t *spare;
797 
798 	assert(arena_mapbits_allocated_get(chunk, map_bias) == 0);
799 	assert(arena_mapbits_allocated_get(chunk, chunk_npages-1) == 0);
800 	assert(arena_mapbits_unallocated_size_get(chunk, map_bias) ==
801 	    arena_maxrun);
802 	assert(arena_mapbits_unallocated_size_get(chunk, chunk_npages-1) ==
803 	    arena_maxrun);
804 	assert(arena_mapbits_dirty_get(chunk, map_bias) ==
805 	    arena_mapbits_dirty_get(chunk, chunk_npages-1));
806 	assert(arena_mapbits_decommitted_get(chunk, map_bias) ==
807 	    arena_mapbits_decommitted_get(chunk, chunk_npages-1));
808 
809 	/* Remove run from runs_avail, so that the arena does not use it. */
810 	arena_avail_remove(arena, chunk, map_bias, chunk_npages-map_bias);
811 
812 	ql_remove(&arena->achunks, &chunk->node, ql_link);
813 	spare = arena->spare;
814 	arena->spare = chunk;
815 	if (spare != NULL)
816 		arena_spare_discard(tsdn, arena, spare);
817 }
818 
819 static void
820 arena_huge_malloc_stats_update(arena_t *arena, size_t usize)
821 {
822 	szind_t index = size2index(usize) - nlclasses - NBINS;
823 
824 	cassert(config_stats);
825 
826 	arena->stats.nmalloc_huge++;
827 	arena->stats.allocated_huge += usize;
828 	arena->stats.hstats[index].nmalloc++;
829 	arena->stats.hstats[index].curhchunks++;
830 }
831 
832 static void
833 arena_huge_malloc_stats_update_undo(arena_t *arena, size_t usize)
834 {
835 	szind_t index = size2index(usize) - nlclasses - NBINS;
836 
837 	cassert(config_stats);
838 
839 	arena->stats.nmalloc_huge--;
840 	arena->stats.allocated_huge -= usize;
841 	arena->stats.hstats[index].nmalloc--;
842 	arena->stats.hstats[index].curhchunks--;
843 }
844 
845 static void
846 arena_huge_dalloc_stats_update(arena_t *arena, size_t usize)
847 {
848 	szind_t index = size2index(usize) - nlclasses - NBINS;
849 
850 	cassert(config_stats);
851 
852 	arena->stats.ndalloc_huge++;
853 	arena->stats.allocated_huge -= usize;
854 	arena->stats.hstats[index].ndalloc++;
855 	arena->stats.hstats[index].curhchunks--;
856 }
857 
858 static void
859 arena_huge_reset_stats_cancel(arena_t *arena, size_t usize)
860 {
861 	szind_t index = size2index(usize) - nlclasses - NBINS;
862 
863 	cassert(config_stats);
864 
865 	arena->stats.ndalloc_huge++;
866 	arena->stats.hstats[index].ndalloc--;
867 }
868 
869 static void
870 arena_huge_dalloc_stats_update_undo(arena_t *arena, size_t usize)
871 {
872 	szind_t index = size2index(usize) - nlclasses - NBINS;
873 
874 	cassert(config_stats);
875 
876 	arena->stats.ndalloc_huge--;
877 	arena->stats.allocated_huge += usize;
878 	arena->stats.hstats[index].ndalloc--;
879 	arena->stats.hstats[index].curhchunks++;
880 }
881 
882 static void
883 arena_huge_ralloc_stats_update(arena_t *arena, size_t oldsize, size_t usize)
884 {
885 
886 	arena_huge_dalloc_stats_update(arena, oldsize);
887 	arena_huge_malloc_stats_update(arena, usize);
888 }
889 
890 static void
891 arena_huge_ralloc_stats_update_undo(arena_t *arena, size_t oldsize,
892     size_t usize)
893 {
894 
895 	arena_huge_dalloc_stats_update_undo(arena, oldsize);
896 	arena_huge_malloc_stats_update_undo(arena, usize);
897 }
898 
899 extent_node_t *
900 arena_node_alloc(tsdn_t *tsdn, arena_t *arena)
901 {
902 	extent_node_t *node;
903 
904 	malloc_mutex_lock(tsdn, &arena->node_cache_mtx);
905 	node = ql_last(&arena->node_cache, ql_link);
906 	if (node == NULL) {
907 		malloc_mutex_unlock(tsdn, &arena->node_cache_mtx);
908 		return (base_alloc(tsdn, sizeof(extent_node_t)));
909 	}
910 	ql_tail_remove(&arena->node_cache, extent_node_t, ql_link);
911 	malloc_mutex_unlock(tsdn, &arena->node_cache_mtx);
912 	return (node);
913 }
914 
915 void
916 arena_node_dalloc(tsdn_t *tsdn, arena_t *arena, extent_node_t *node)
917 {
918 
919 	malloc_mutex_lock(tsdn, &arena->node_cache_mtx);
920 	ql_elm_new(node, ql_link);
921 	ql_tail_insert(&arena->node_cache, node, ql_link);
922 	malloc_mutex_unlock(tsdn, &arena->node_cache_mtx);
923 }
924 
925 static void *
926 arena_chunk_alloc_huge_hard(tsdn_t *tsdn, arena_t *arena,
927     chunk_hooks_t *chunk_hooks, size_t usize, size_t alignment, bool *zero,
928     size_t csize)
929 {
930 	void *ret;
931 	bool commit = true;
932 
933 	ret = chunk_alloc_wrapper(tsdn, arena, chunk_hooks, NULL, csize,
934 	    alignment, zero, &commit);
935 	if (ret == NULL) {
936 		/* Revert optimistic stats updates. */
937 		malloc_mutex_lock(tsdn, &arena->lock);
938 		if (config_stats) {
939 			arena_huge_malloc_stats_update_undo(arena, usize);
940 			arena->stats.mapped -= usize;
941 		}
942 		arena_nactive_sub(arena, usize >> LG_PAGE);
943 		malloc_mutex_unlock(tsdn, &arena->lock);
944 	}
945 
946 	return (ret);
947 }
948 
949 void *
950 arena_chunk_alloc_huge(tsdn_t *tsdn, arena_t *arena, size_t usize,
951     size_t alignment, bool *zero)
952 {
953 	void *ret;
954 	chunk_hooks_t chunk_hooks = CHUNK_HOOKS_INITIALIZER;
955 	size_t csize = CHUNK_CEILING(usize);
956 
957 	malloc_mutex_lock(tsdn, &arena->lock);
958 
959 	/* Optimistically update stats. */
960 	if (config_stats) {
961 		arena_huge_malloc_stats_update(arena, usize);
962 		arena->stats.mapped += usize;
963 	}
964 	arena_nactive_add(arena, usize >> LG_PAGE);
965 
966 	ret = chunk_alloc_cache(tsdn, arena, &chunk_hooks, NULL, csize,
967 	    alignment, zero, true);
968 	malloc_mutex_unlock(tsdn, &arena->lock);
969 	if (ret == NULL) {
970 		ret = arena_chunk_alloc_huge_hard(tsdn, arena, &chunk_hooks,
971 		    usize, alignment, zero, csize);
972 	}
973 
974 	return (ret);
975 }
976 
977 void
978 arena_chunk_dalloc_huge(tsdn_t *tsdn, arena_t *arena, void *chunk, size_t usize)
979 {
980 	chunk_hooks_t chunk_hooks = CHUNK_HOOKS_INITIALIZER;
981 	size_t csize;
982 
983 	csize = CHUNK_CEILING(usize);
984 	malloc_mutex_lock(tsdn, &arena->lock);
985 	if (config_stats) {
986 		arena_huge_dalloc_stats_update(arena, usize);
987 		arena->stats.mapped -= usize;
988 	}
989 	arena_nactive_sub(arena, usize >> LG_PAGE);
990 
991 	chunk_dalloc_cache(tsdn, arena, &chunk_hooks, chunk, csize, true);
992 	malloc_mutex_unlock(tsdn, &arena->lock);
993 }
994 
995 void
996 arena_chunk_ralloc_huge_similar(tsdn_t *tsdn, arena_t *arena, void *chunk,
997     size_t oldsize, size_t usize)
998 {
999 
1000 	assert(CHUNK_CEILING(oldsize) == CHUNK_CEILING(usize));
1001 	assert(oldsize != usize);
1002 
1003 	malloc_mutex_lock(tsdn, &arena->lock);
1004 	if (config_stats)
1005 		arena_huge_ralloc_stats_update(arena, oldsize, usize);
1006 	if (oldsize < usize)
1007 		arena_nactive_add(arena, (usize - oldsize) >> LG_PAGE);
1008 	else
1009 		arena_nactive_sub(arena, (oldsize - usize) >> LG_PAGE);
1010 	malloc_mutex_unlock(tsdn, &arena->lock);
1011 }
1012 
1013 void
1014 arena_chunk_ralloc_huge_shrink(tsdn_t *tsdn, arena_t *arena, void *chunk,
1015     size_t oldsize, size_t usize)
1016 {
1017 	size_t udiff = oldsize - usize;
1018 	size_t cdiff = CHUNK_CEILING(oldsize) - CHUNK_CEILING(usize);
1019 
1020 	malloc_mutex_lock(tsdn, &arena->lock);
1021 	if (config_stats) {
1022 		arena_huge_ralloc_stats_update(arena, oldsize, usize);
1023 		if (cdiff != 0)
1024 			arena->stats.mapped -= cdiff;
1025 	}
1026 	arena_nactive_sub(arena, udiff >> LG_PAGE);
1027 
1028 	if (cdiff != 0) {
1029 		chunk_hooks_t chunk_hooks = CHUNK_HOOKS_INITIALIZER;
1030 		void *nchunk = (void *)((uintptr_t)chunk +
1031 		    CHUNK_CEILING(usize));
1032 
1033 		chunk_dalloc_cache(tsdn, arena, &chunk_hooks, nchunk, cdiff,
1034 		    true);
1035 	}
1036 	malloc_mutex_unlock(tsdn, &arena->lock);
1037 }
1038 
1039 static bool
1040 arena_chunk_ralloc_huge_expand_hard(tsdn_t *tsdn, arena_t *arena,
1041     chunk_hooks_t *chunk_hooks, void *chunk, size_t oldsize, size_t usize,
1042     bool *zero, void *nchunk, size_t udiff, size_t cdiff)
1043 {
1044 	bool err;
1045 	bool commit = true;
1046 
1047 	err = (chunk_alloc_wrapper(tsdn, arena, chunk_hooks, nchunk, cdiff,
1048 	    chunksize, zero, &commit) == NULL);
1049 	if (err) {
1050 		/* Revert optimistic stats updates. */
1051 		malloc_mutex_lock(tsdn, &arena->lock);
1052 		if (config_stats) {
1053 			arena_huge_ralloc_stats_update_undo(arena, oldsize,
1054 			    usize);
1055 			arena->stats.mapped -= cdiff;
1056 		}
1057 		arena_nactive_sub(arena, udiff >> LG_PAGE);
1058 		malloc_mutex_unlock(tsdn, &arena->lock);
1059 	} else if (chunk_hooks->merge(chunk, CHUNK_CEILING(oldsize), nchunk,
1060 	    cdiff, true, arena->ind)) {
1061 		chunk_dalloc_wrapper(tsdn, arena, chunk_hooks, nchunk, cdiff,
1062 		    *zero, true);
1063 		err = true;
1064 	}
1065 	return (err);
1066 }
1067 
1068 bool
1069 arena_chunk_ralloc_huge_expand(tsdn_t *tsdn, arena_t *arena, void *chunk,
1070     size_t oldsize, size_t usize, bool *zero)
1071 {
1072 	bool err;
1073 	chunk_hooks_t chunk_hooks = chunk_hooks_get(tsdn, arena);
1074 	void *nchunk = (void *)((uintptr_t)chunk + CHUNK_CEILING(oldsize));
1075 	size_t udiff = usize - oldsize;
1076 	size_t cdiff = CHUNK_CEILING(usize) - CHUNK_CEILING(oldsize);
1077 
1078 	malloc_mutex_lock(tsdn, &arena->lock);
1079 
1080 	/* Optimistically update stats. */
1081 	if (config_stats) {
1082 		arena_huge_ralloc_stats_update(arena, oldsize, usize);
1083 		arena->stats.mapped += cdiff;
1084 	}
1085 	arena_nactive_add(arena, udiff >> LG_PAGE);
1086 
1087 	err = (chunk_alloc_cache(tsdn, arena, &chunk_hooks, nchunk, cdiff,
1088 	    chunksize, zero, true) == NULL);
1089 	malloc_mutex_unlock(tsdn, &arena->lock);
1090 	if (err) {
1091 		err = arena_chunk_ralloc_huge_expand_hard(tsdn, arena,
1092 		    &chunk_hooks, chunk, oldsize, usize, zero, nchunk, udiff,
1093 		    cdiff);
1094 	} else if (chunk_hooks.merge(chunk, CHUNK_CEILING(oldsize), nchunk,
1095 	    cdiff, true, arena->ind)) {
1096 		chunk_dalloc_wrapper(tsdn, arena, &chunk_hooks, nchunk, cdiff,
1097 		    *zero, true);
1098 		err = true;
1099 	}
1100 
1101 	return (err);
1102 }
1103 
1104 /*
1105  * Do first-best-fit run selection, i.e. select the lowest run that best fits.
1106  * Run sizes are indexed, so not all candidate runs are necessarily exactly the
1107  * same size.
1108  */
1109 static arena_run_t *
1110 arena_run_first_best_fit(arena_t *arena, size_t size)
1111 {
1112 	szind_t ind, i;
1113 
1114 	ind = size2index(run_quantize_ceil(size));
1115 	for (i = ind; i < runs_avail_nclasses + runs_avail_bias; i++) {
1116 		arena_chunk_map_misc_t *miscelm = arena_run_heap_first(
1117 		    arena_runs_avail_get(arena, i));
1118 		if (miscelm != NULL)
1119 			return (&miscelm->run);
1120 	}
1121 
1122 	return (NULL);
1123 }
1124 
1125 static arena_run_t *
1126 arena_run_alloc_large_helper(arena_t *arena, size_t size, bool zero)
1127 {
1128 	arena_run_t *run = arena_run_first_best_fit(arena, s2u(size));
1129 	if (run != NULL) {
1130 		if (arena_run_split_large(arena, run, size, zero))
1131 			run = NULL;
1132 	}
1133 	return (run);
1134 }
1135 
1136 static arena_run_t *
1137 arena_run_alloc_large(tsdn_t *tsdn, arena_t *arena, size_t size, bool zero)
1138 {
1139 	arena_chunk_t *chunk;
1140 	arena_run_t *run;
1141 
1142 	assert(size <= arena_maxrun);
1143 	assert(size == PAGE_CEILING(size));
1144 
1145 	/* Search the arena's chunks for the lowest best fit. */
1146 	run = arena_run_alloc_large_helper(arena, size, zero);
1147 	if (run != NULL)
1148 		return (run);
1149 
1150 	/*
1151 	 * No usable runs.  Create a new chunk from which to allocate the run.
1152 	 */
1153 	chunk = arena_chunk_alloc(tsdn, arena);
1154 	if (chunk != NULL) {
1155 		run = &arena_miscelm_get_mutable(chunk, map_bias)->run;
1156 		if (arena_run_split_large(arena, run, size, zero))
1157 			run = NULL;
1158 		return (run);
1159 	}
1160 
1161 	/*
1162 	 * arena_chunk_alloc() failed, but another thread may have made
1163 	 * sufficient memory available while this one dropped arena->lock in
1164 	 * arena_chunk_alloc(), so search one more time.
1165 	 */
1166 	return (arena_run_alloc_large_helper(arena, size, zero));
1167 }
1168 
1169 static arena_run_t *
1170 arena_run_alloc_small_helper(arena_t *arena, size_t size, szind_t binind)
1171 {
1172 	arena_run_t *run = arena_run_first_best_fit(arena, size);
1173 	if (run != NULL) {
1174 		if (arena_run_split_small(arena, run, size, binind))
1175 			run = NULL;
1176 	}
1177 	return (run);
1178 }
1179 
1180 static arena_run_t *
1181 arena_run_alloc_small(tsdn_t *tsdn, arena_t *arena, size_t size, szind_t binind)
1182 {
1183 	arena_chunk_t *chunk;
1184 	arena_run_t *run;
1185 
1186 	assert(size <= arena_maxrun);
1187 	assert(size == PAGE_CEILING(size));
1188 	assert(binind != BININD_INVALID);
1189 
1190 	/* Search the arena's chunks for the lowest best fit. */
1191 	run = arena_run_alloc_small_helper(arena, size, binind);
1192 	if (run != NULL)
1193 		return (run);
1194 
1195 	/*
1196 	 * No usable runs.  Create a new chunk from which to allocate the run.
1197 	 */
1198 	chunk = arena_chunk_alloc(tsdn, arena);
1199 	if (chunk != NULL) {
1200 		run = &arena_miscelm_get_mutable(chunk, map_bias)->run;
1201 		if (arena_run_split_small(arena, run, size, binind))
1202 			run = NULL;
1203 		return (run);
1204 	}
1205 
1206 	/*
1207 	 * arena_chunk_alloc() failed, but another thread may have made
1208 	 * sufficient memory available while this one dropped arena->lock in
1209 	 * arena_chunk_alloc(), so search one more time.
1210 	 */
1211 	return (arena_run_alloc_small_helper(arena, size, binind));
1212 }
1213 
1214 static bool
1215 arena_lg_dirty_mult_valid(ssize_t lg_dirty_mult)
1216 {
1217 
1218 	return (lg_dirty_mult >= -1 && lg_dirty_mult < (ssize_t)(sizeof(size_t)
1219 	    << 3));
1220 }
1221 
1222 ssize_t
1223 arena_lg_dirty_mult_get(tsdn_t *tsdn, arena_t *arena)
1224 {
1225 	ssize_t lg_dirty_mult;
1226 
1227 	malloc_mutex_lock(tsdn, &arena->lock);
1228 	lg_dirty_mult = arena->lg_dirty_mult;
1229 	malloc_mutex_unlock(tsdn, &arena->lock);
1230 
1231 	return (lg_dirty_mult);
1232 }
1233 
1234 bool
1235 arena_lg_dirty_mult_set(tsdn_t *tsdn, arena_t *arena, ssize_t lg_dirty_mult)
1236 {
1237 
1238 	if (!arena_lg_dirty_mult_valid(lg_dirty_mult))
1239 		return (true);
1240 
1241 	malloc_mutex_lock(tsdn, &arena->lock);
1242 	arena->lg_dirty_mult = lg_dirty_mult;
1243 	arena_maybe_purge(tsdn, arena);
1244 	malloc_mutex_unlock(tsdn, &arena->lock);
1245 
1246 	return (false);
1247 }
1248 
1249 static void
1250 arena_decay_deadline_init(arena_t *arena)
1251 {
1252 
1253 	assert(opt_purge == purge_mode_decay);
1254 
1255 	/*
1256 	 * Generate a new deadline that is uniformly random within the next
1257 	 * epoch after the current one.
1258 	 */
1259 	nstime_copy(&arena->decay_deadline, &arena->decay_epoch);
1260 	nstime_add(&arena->decay_deadline, &arena->decay_interval);
1261 	if (arena->decay_time > 0) {
1262 		nstime_t jitter;
1263 
1264 		nstime_init(&jitter, prng_range(&arena->decay_jitter_state,
1265 		    nstime_ns(&arena->decay_interval)));
1266 		nstime_add(&arena->decay_deadline, &jitter);
1267 	}
1268 }
1269 
1270 static bool
1271 arena_decay_deadline_reached(const arena_t *arena, const nstime_t *time)
1272 {
1273 
1274 	assert(opt_purge == purge_mode_decay);
1275 
1276 	return (nstime_compare(&arena->decay_deadline, time) <= 0);
1277 }
1278 
1279 static size_t
1280 arena_decay_backlog_npages_limit(const arena_t *arena)
1281 {
1282 	static const uint64_t h_steps[] = {
1283 #define	STEP(step, h, x, y) \
1284 		h,
1285 		SMOOTHSTEP
1286 #undef STEP
1287 	};
1288 	uint64_t sum;
1289 	size_t npages_limit_backlog;
1290 	unsigned i;
1291 
1292 	assert(opt_purge == purge_mode_decay);
1293 
1294 	/*
1295 	 * For each element of decay_backlog, multiply by the corresponding
1296 	 * fixed-point smoothstep decay factor.  Sum the products, then divide
1297 	 * to round down to the nearest whole number of pages.
1298 	 */
1299 	sum = 0;
1300 	for (i = 0; i < SMOOTHSTEP_NSTEPS; i++)
1301 		sum += arena->decay_backlog[i] * h_steps[i];
1302 	npages_limit_backlog = (size_t)(sum >> SMOOTHSTEP_BFP);
1303 
1304 	return (npages_limit_backlog);
1305 }
1306 
1307 static void
1308 arena_decay_epoch_advance(arena_t *arena, const nstime_t *time)
1309 {
1310 	uint64_t nadvance_u64;
1311 	nstime_t delta;
1312 	size_t ndirty_delta;
1313 
1314 	assert(opt_purge == purge_mode_decay);
1315 	assert(arena_decay_deadline_reached(arena, time));
1316 
1317 	nstime_copy(&delta, time);
1318 	nstime_subtract(&delta, &arena->decay_epoch);
1319 	nadvance_u64 = nstime_divide(&delta, &arena->decay_interval);
1320 	assert(nadvance_u64 > 0);
1321 
1322 	/* Add nadvance_u64 decay intervals to epoch. */
1323 	nstime_copy(&delta, &arena->decay_interval);
1324 	nstime_imultiply(&delta, nadvance_u64);
1325 	nstime_add(&arena->decay_epoch, &delta);
1326 
1327 	/* Set a new deadline. */
1328 	arena_decay_deadline_init(arena);
1329 
1330 	/* Update the backlog. */
1331 	if (nadvance_u64 >= SMOOTHSTEP_NSTEPS) {
1332 		memset(arena->decay_backlog, 0, (SMOOTHSTEP_NSTEPS-1) *
1333 		    sizeof(size_t));
1334 	} else {
1335 		size_t nadvance_z = (size_t)nadvance_u64;
1336 
1337 		assert((uint64_t)nadvance_z == nadvance_u64);
1338 
1339 		memmove(arena->decay_backlog, &arena->decay_backlog[nadvance_z],
1340 		    (SMOOTHSTEP_NSTEPS - nadvance_z) * sizeof(size_t));
1341 		if (nadvance_z > 1) {
1342 			memset(&arena->decay_backlog[SMOOTHSTEP_NSTEPS -
1343 			    nadvance_z], 0, (nadvance_z-1) * sizeof(size_t));
1344 		}
1345 	}
1346 	ndirty_delta = (arena->ndirty > arena->decay_ndirty) ? arena->ndirty -
1347 	    arena->decay_ndirty : 0;
1348 	arena->decay_ndirty = arena->ndirty;
1349 	arena->decay_backlog[SMOOTHSTEP_NSTEPS-1] = ndirty_delta;
1350 	arena->decay_backlog_npages_limit =
1351 	    arena_decay_backlog_npages_limit(arena);
1352 }
1353 
1354 static size_t
1355 arena_decay_npages_limit(arena_t *arena)
1356 {
1357 	size_t npages_limit;
1358 
1359 	assert(opt_purge == purge_mode_decay);
1360 
1361 	npages_limit = arena->decay_backlog_npages_limit;
1362 
1363 	/* Add in any dirty pages created during the current epoch. */
1364 	if (arena->ndirty > arena->decay_ndirty)
1365 		npages_limit += arena->ndirty - arena->decay_ndirty;
1366 
1367 	return (npages_limit);
1368 }
1369 
1370 static void
1371 arena_decay_init(arena_t *arena, ssize_t decay_time)
1372 {
1373 
1374 	arena->decay_time = decay_time;
1375 	if (decay_time > 0) {
1376 		nstime_init2(&arena->decay_interval, decay_time, 0);
1377 		nstime_idivide(&arena->decay_interval, SMOOTHSTEP_NSTEPS);
1378 	}
1379 
1380 	nstime_init(&arena->decay_epoch, 0);
1381 	nstime_update(&arena->decay_epoch);
1382 	arena->decay_jitter_state = (uint64_t)(uintptr_t)arena;
1383 	arena_decay_deadline_init(arena);
1384 	arena->decay_ndirty = arena->ndirty;
1385 	arena->decay_backlog_npages_limit = 0;
1386 	memset(arena->decay_backlog, 0, SMOOTHSTEP_NSTEPS * sizeof(size_t));
1387 }
1388 
1389 static bool
1390 arena_decay_time_valid(ssize_t decay_time)
1391 {
1392 
1393 	if (decay_time < -1)
1394 		return (false);
1395 	if (decay_time == -1 || (uint64_t)decay_time <= NSTIME_SEC_MAX)
1396 		return (true);
1397 	return (false);
1398 }
1399 
1400 ssize_t
1401 arena_decay_time_get(tsdn_t *tsdn, arena_t *arena)
1402 {
1403 	ssize_t decay_time;
1404 
1405 	malloc_mutex_lock(tsdn, &arena->lock);
1406 	decay_time = arena->decay_time;
1407 	malloc_mutex_unlock(tsdn, &arena->lock);
1408 
1409 	return (decay_time);
1410 }
1411 
1412 bool
1413 arena_decay_time_set(tsdn_t *tsdn, arena_t *arena, ssize_t decay_time)
1414 {
1415 
1416 	if (!arena_decay_time_valid(decay_time))
1417 		return (true);
1418 
1419 	malloc_mutex_lock(tsdn, &arena->lock);
1420 	/*
1421 	 * Restart decay backlog from scratch, which may cause many dirty pages
1422 	 * to be immediately purged.  It would conceptually be possible to map
1423 	 * the old backlog onto the new backlog, but there is no justification
1424 	 * for such complexity since decay_time changes are intended to be
1425 	 * infrequent, either between the {-1, 0, >0} states, or a one-time
1426 	 * arbitrary change during initial arena configuration.
1427 	 */
1428 	arena_decay_init(arena, decay_time);
1429 	arena_maybe_purge(tsdn, arena);
1430 	malloc_mutex_unlock(tsdn, &arena->lock);
1431 
1432 	return (false);
1433 }
1434 
1435 static void
1436 arena_maybe_purge_ratio(tsdn_t *tsdn, arena_t *arena)
1437 {
1438 
1439 	assert(opt_purge == purge_mode_ratio);
1440 
1441 	/* Don't purge if the option is disabled. */
1442 	if (arena->lg_dirty_mult < 0)
1443 		return;
1444 
1445 	/*
1446 	 * Iterate, since preventing recursive purging could otherwise leave too
1447 	 * many dirty pages.
1448 	 */
1449 	while (true) {
1450 		size_t threshold = (arena->nactive >> arena->lg_dirty_mult);
1451 		if (threshold < chunk_npages)
1452 			threshold = chunk_npages;
1453 		/*
1454 		 * Don't purge unless the number of purgeable pages exceeds the
1455 		 * threshold.
1456 		 */
1457 		if (arena->ndirty <= threshold)
1458 			return;
1459 		arena_purge_to_limit(tsdn, arena, threshold);
1460 	}
1461 }
1462 
1463 static void
1464 arena_maybe_purge_decay(tsdn_t *tsdn, arena_t *arena)
1465 {
1466 	nstime_t time;
1467 	size_t ndirty_limit;
1468 
1469 	assert(opt_purge == purge_mode_decay);
1470 
1471 	/* Purge all or nothing if the option is disabled. */
1472 	if (arena->decay_time <= 0) {
1473 		if (arena->decay_time == 0)
1474 			arena_purge_to_limit(tsdn, arena, 0);
1475 		return;
1476 	}
1477 
1478 	nstime_copy(&time, &arena->decay_epoch);
1479 	if (unlikely(nstime_update(&time))) {
1480 		/* Time went backwards.  Force an epoch advance. */
1481 		nstime_copy(&time, &arena->decay_deadline);
1482 	}
1483 
1484 	if (arena_decay_deadline_reached(arena, &time))
1485 		arena_decay_epoch_advance(arena, &time);
1486 
1487 	ndirty_limit = arena_decay_npages_limit(arena);
1488 
1489 	/*
1490 	 * Don't try to purge unless the number of purgeable pages exceeds the
1491 	 * current limit.
1492 	 */
1493 	if (arena->ndirty <= ndirty_limit)
1494 		return;
1495 	arena_purge_to_limit(tsdn, arena, ndirty_limit);
1496 }
1497 
1498 void
1499 arena_maybe_purge(tsdn_t *tsdn, arena_t *arena)
1500 {
1501 
1502 	/* Don't recursively purge. */
1503 	if (arena->purging)
1504 		return;
1505 
1506 	if (opt_purge == purge_mode_ratio)
1507 		arena_maybe_purge_ratio(tsdn, arena);
1508 	else
1509 		arena_maybe_purge_decay(tsdn, arena);
1510 }
1511 
1512 static size_t
1513 arena_dirty_count(arena_t *arena)
1514 {
1515 	size_t ndirty = 0;
1516 	arena_runs_dirty_link_t *rdelm;
1517 	extent_node_t *chunkselm;
1518 
1519 	for (rdelm = qr_next(&arena->runs_dirty, rd_link),
1520 	    chunkselm = qr_next(&arena->chunks_cache, cc_link);
1521 	    rdelm != &arena->runs_dirty; rdelm = qr_next(rdelm, rd_link)) {
1522 		size_t npages;
1523 
1524 		if (rdelm == &chunkselm->rd) {
1525 			npages = extent_node_size_get(chunkselm) >> LG_PAGE;
1526 			chunkselm = qr_next(chunkselm, cc_link);
1527 		} else {
1528 			arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(
1529 			    rdelm);
1530 			arena_chunk_map_misc_t *miscelm =
1531 			    arena_rd_to_miscelm(rdelm);
1532 			size_t pageind = arena_miscelm_to_pageind(miscelm);
1533 			assert(arena_mapbits_allocated_get(chunk, pageind) ==
1534 			    0);
1535 			assert(arena_mapbits_large_get(chunk, pageind) == 0);
1536 			assert(arena_mapbits_dirty_get(chunk, pageind) != 0);
1537 			npages = arena_mapbits_unallocated_size_get(chunk,
1538 			    pageind) >> LG_PAGE;
1539 		}
1540 		ndirty += npages;
1541 	}
1542 
1543 	return (ndirty);
1544 }
1545 
1546 static size_t
1547 arena_stash_dirty(tsdn_t *tsdn, arena_t *arena, chunk_hooks_t *chunk_hooks,
1548     size_t ndirty_limit, arena_runs_dirty_link_t *purge_runs_sentinel,
1549     extent_node_t *purge_chunks_sentinel)
1550 {
1551 	arena_runs_dirty_link_t *rdelm, *rdelm_next;
1552 	extent_node_t *chunkselm;
1553 	size_t nstashed = 0;
1554 
1555 	/* Stash runs/chunks according to ndirty_limit. */
1556 	for (rdelm = qr_next(&arena->runs_dirty, rd_link),
1557 	    chunkselm = qr_next(&arena->chunks_cache, cc_link);
1558 	    rdelm != &arena->runs_dirty; rdelm = rdelm_next) {
1559 		size_t npages;
1560 		rdelm_next = qr_next(rdelm, rd_link);
1561 
1562 		if (rdelm == &chunkselm->rd) {
1563 			extent_node_t *chunkselm_next;
1564 			bool zero;
1565 			UNUSED void *chunk;
1566 
1567 			npages = extent_node_size_get(chunkselm) >> LG_PAGE;
1568 			if (opt_purge == purge_mode_decay && arena->ndirty -
1569 			    (nstashed + npages) < ndirty_limit)
1570 				break;
1571 
1572 			chunkselm_next = qr_next(chunkselm, cc_link);
1573 			/*
1574 			 * Allocate.  chunkselm remains valid due to the
1575 			 * dalloc_node=false argument to chunk_alloc_cache().
1576 			 */
1577 			zero = false;
1578 			chunk = chunk_alloc_cache(tsdn, arena, chunk_hooks,
1579 			    extent_node_addr_get(chunkselm),
1580 			    extent_node_size_get(chunkselm), chunksize, &zero,
1581 			    false);
1582 			assert(chunk == extent_node_addr_get(chunkselm));
1583 			assert(zero == extent_node_zeroed_get(chunkselm));
1584 			extent_node_dirty_insert(chunkselm, purge_runs_sentinel,
1585 			    purge_chunks_sentinel);
1586 			assert(npages == (extent_node_size_get(chunkselm) >>
1587 			    LG_PAGE));
1588 			chunkselm = chunkselm_next;
1589 		} else {
1590 			arena_chunk_t *chunk =
1591 			    (arena_chunk_t *)CHUNK_ADDR2BASE(rdelm);
1592 			arena_chunk_map_misc_t *miscelm =
1593 			    arena_rd_to_miscelm(rdelm);
1594 			size_t pageind = arena_miscelm_to_pageind(miscelm);
1595 			arena_run_t *run = &miscelm->run;
1596 			size_t run_size =
1597 			    arena_mapbits_unallocated_size_get(chunk, pageind);
1598 
1599 			npages = run_size >> LG_PAGE;
1600 			if (opt_purge == purge_mode_decay && arena->ndirty -
1601 			    (nstashed + npages) < ndirty_limit)
1602 				break;
1603 
1604 			assert(pageind + npages <= chunk_npages);
1605 			assert(arena_mapbits_dirty_get(chunk, pageind) ==
1606 			    arena_mapbits_dirty_get(chunk, pageind+npages-1));
1607 
1608 			/*
1609 			 * If purging the spare chunk's run, make it available
1610 			 * prior to allocation.
1611 			 */
1612 			if (chunk == arena->spare)
1613 				arena_chunk_alloc(tsdn, arena);
1614 
1615 			/* Temporarily allocate the free dirty run. */
1616 			arena_run_split_large(arena, run, run_size, false);
1617 			/* Stash. */
1618 			if (false)
1619 				qr_new(rdelm, rd_link); /* Redundant. */
1620 			else {
1621 				assert(qr_next(rdelm, rd_link) == rdelm);
1622 				assert(qr_prev(rdelm, rd_link) == rdelm);
1623 			}
1624 			qr_meld(purge_runs_sentinel, rdelm, rd_link);
1625 		}
1626 
1627 		nstashed += npages;
1628 		if (opt_purge == purge_mode_ratio && arena->ndirty - nstashed <=
1629 		    ndirty_limit)
1630 			break;
1631 	}
1632 
1633 	return (nstashed);
1634 }
1635 
1636 static size_t
1637 arena_purge_stashed(tsdn_t *tsdn, arena_t *arena, chunk_hooks_t *chunk_hooks,
1638     arena_runs_dirty_link_t *purge_runs_sentinel,
1639     extent_node_t *purge_chunks_sentinel)
1640 {
1641 	size_t npurged, nmadvise;
1642 	arena_runs_dirty_link_t *rdelm;
1643 	extent_node_t *chunkselm;
1644 
1645 	if (config_stats)
1646 		nmadvise = 0;
1647 	npurged = 0;
1648 
1649 	malloc_mutex_unlock(tsdn, &arena->lock);
1650 	for (rdelm = qr_next(purge_runs_sentinel, rd_link),
1651 	    chunkselm = qr_next(purge_chunks_sentinel, cc_link);
1652 	    rdelm != purge_runs_sentinel; rdelm = qr_next(rdelm, rd_link)) {
1653 		size_t npages;
1654 
1655 		if (rdelm == &chunkselm->rd) {
1656 			/*
1657 			 * Don't actually purge the chunk here because 1)
1658 			 * chunkselm is embedded in the chunk and must remain
1659 			 * valid, and 2) we deallocate the chunk in
1660 			 * arena_unstash_purged(), where it is destroyed,
1661 			 * decommitted, or purged, depending on chunk
1662 			 * deallocation policy.
1663 			 */
1664 			size_t size = extent_node_size_get(chunkselm);
1665 			npages = size >> LG_PAGE;
1666 			chunkselm = qr_next(chunkselm, cc_link);
1667 		} else {
1668 			size_t pageind, run_size, flag_unzeroed, flags, i;
1669 			bool decommitted;
1670 			arena_chunk_t *chunk =
1671 			    (arena_chunk_t *)CHUNK_ADDR2BASE(rdelm);
1672 			arena_chunk_map_misc_t *miscelm =
1673 			    arena_rd_to_miscelm(rdelm);
1674 			pageind = arena_miscelm_to_pageind(miscelm);
1675 			run_size = arena_mapbits_large_size_get(chunk, pageind);
1676 			npages = run_size >> LG_PAGE;
1677 
1678 			assert(pageind + npages <= chunk_npages);
1679 			assert(!arena_mapbits_decommitted_get(chunk, pageind));
1680 			assert(!arena_mapbits_decommitted_get(chunk,
1681 			    pageind+npages-1));
1682 			decommitted = !chunk_hooks->decommit(chunk, chunksize,
1683 			    pageind << LG_PAGE, npages << LG_PAGE, arena->ind);
1684 			if (decommitted) {
1685 				flag_unzeroed = 0;
1686 				flags = CHUNK_MAP_DECOMMITTED;
1687 			} else {
1688 				flag_unzeroed = chunk_purge_wrapper(tsdn, arena,
1689 				    chunk_hooks, chunk, chunksize, pageind <<
1690 				    LG_PAGE, run_size) ? CHUNK_MAP_UNZEROED : 0;
1691 				flags = flag_unzeroed;
1692 			}
1693 			arena_mapbits_large_set(chunk, pageind+npages-1, 0,
1694 			    flags);
1695 			arena_mapbits_large_set(chunk, pageind, run_size,
1696 			    flags);
1697 
1698 			/*
1699 			 * Set the unzeroed flag for internal pages, now that
1700 			 * chunk_purge_wrapper() has returned whether the pages
1701 			 * were zeroed as a side effect of purging.  This chunk
1702 			 * map modification is safe even though the arena mutex
1703 			 * isn't currently owned by this thread, because the run
1704 			 * is marked as allocated, thus protecting it from being
1705 			 * modified by any other thread.  As long as these
1706 			 * writes don't perturb the first and last elements'
1707 			 * CHUNK_MAP_ALLOCATED bits, behavior is well defined.
1708 			 */
1709 			for (i = 1; i < npages-1; i++) {
1710 				arena_mapbits_internal_set(chunk, pageind+i,
1711 				    flag_unzeroed);
1712 			}
1713 		}
1714 
1715 		npurged += npages;
1716 		if (config_stats)
1717 			nmadvise++;
1718 	}
1719 	malloc_mutex_lock(tsdn, &arena->lock);
1720 
1721 	if (config_stats) {
1722 		arena->stats.nmadvise += nmadvise;
1723 		arena->stats.purged += npurged;
1724 	}
1725 
1726 	return (npurged);
1727 }
1728 
1729 static void
1730 arena_unstash_purged(tsdn_t *tsdn, arena_t *arena, chunk_hooks_t *chunk_hooks,
1731     arena_runs_dirty_link_t *purge_runs_sentinel,
1732     extent_node_t *purge_chunks_sentinel)
1733 {
1734 	arena_runs_dirty_link_t *rdelm, *rdelm_next;
1735 	extent_node_t *chunkselm;
1736 
1737 	/* Deallocate chunks/runs. */
1738 	for (rdelm = qr_next(purge_runs_sentinel, rd_link),
1739 	    chunkselm = qr_next(purge_chunks_sentinel, cc_link);
1740 	    rdelm != purge_runs_sentinel; rdelm = rdelm_next) {
1741 		rdelm_next = qr_next(rdelm, rd_link);
1742 		if (rdelm == &chunkselm->rd) {
1743 			extent_node_t *chunkselm_next = qr_next(chunkselm,
1744 			    cc_link);
1745 			void *addr = extent_node_addr_get(chunkselm);
1746 			size_t size = extent_node_size_get(chunkselm);
1747 			bool zeroed = extent_node_zeroed_get(chunkselm);
1748 			bool committed = extent_node_committed_get(chunkselm);
1749 			extent_node_dirty_remove(chunkselm);
1750 			arena_node_dalloc(tsdn, arena, chunkselm);
1751 			chunkselm = chunkselm_next;
1752 			chunk_dalloc_wrapper(tsdn, arena, chunk_hooks, addr,
1753 			    size, zeroed, committed);
1754 		} else {
1755 			arena_chunk_t *chunk =
1756 			    (arena_chunk_t *)CHUNK_ADDR2BASE(rdelm);
1757 			arena_chunk_map_misc_t *miscelm =
1758 			    arena_rd_to_miscelm(rdelm);
1759 			size_t pageind = arena_miscelm_to_pageind(miscelm);
1760 			bool decommitted = (arena_mapbits_decommitted_get(chunk,
1761 			    pageind) != 0);
1762 			arena_run_t *run = &miscelm->run;
1763 			qr_remove(rdelm, rd_link);
1764 			arena_run_dalloc(tsdn, arena, run, false, true,
1765 			    decommitted);
1766 		}
1767 	}
1768 }
1769 
1770 /*
1771  * NB: ndirty_limit is interpreted differently depending on opt_purge:
1772  *   - purge_mode_ratio: Purge as few dirty run/chunks as possible to reach the
1773  *                       desired state:
1774  *                       (arena->ndirty <= ndirty_limit)
1775  *   - purge_mode_decay: Purge as many dirty runs/chunks as possible without
1776  *                       violating the invariant:
1777  *                       (arena->ndirty >= ndirty_limit)
1778  */
1779 static void
1780 arena_purge_to_limit(tsdn_t *tsdn, arena_t *arena, size_t ndirty_limit)
1781 {
1782 	chunk_hooks_t chunk_hooks = chunk_hooks_get(tsdn, arena);
1783 	size_t npurge, npurged;
1784 	arena_runs_dirty_link_t purge_runs_sentinel;
1785 	extent_node_t purge_chunks_sentinel;
1786 
1787 	arena->purging = true;
1788 
1789 	/*
1790 	 * Calls to arena_dirty_count() are disabled even for debug builds
1791 	 * because overhead grows nonlinearly as memory usage increases.
1792 	 */
1793 	if (false && config_debug) {
1794 		size_t ndirty = arena_dirty_count(arena);
1795 		assert(ndirty == arena->ndirty);
1796 	}
1797 	assert(opt_purge != purge_mode_ratio || (arena->nactive >>
1798 	    arena->lg_dirty_mult) < arena->ndirty || ndirty_limit == 0);
1799 
1800 	qr_new(&purge_runs_sentinel, rd_link);
1801 	extent_node_dirty_linkage_init(&purge_chunks_sentinel);
1802 
1803 	npurge = arena_stash_dirty(tsdn, arena, &chunk_hooks, ndirty_limit,
1804 	    &purge_runs_sentinel, &purge_chunks_sentinel);
1805 	if (npurge == 0)
1806 		goto label_return;
1807 	npurged = arena_purge_stashed(tsdn, arena, &chunk_hooks,
1808 	    &purge_runs_sentinel, &purge_chunks_sentinel);
1809 	assert(npurged == npurge);
1810 	arena_unstash_purged(tsdn, arena, &chunk_hooks, &purge_runs_sentinel,
1811 	    &purge_chunks_sentinel);
1812 
1813 	if (config_stats)
1814 		arena->stats.npurge++;
1815 
1816 label_return:
1817 	arena->purging = false;
1818 }
1819 
1820 void
1821 arena_purge(tsdn_t *tsdn, arena_t *arena, bool all)
1822 {
1823 
1824 	malloc_mutex_lock(tsdn, &arena->lock);
1825 	if (all)
1826 		arena_purge_to_limit(tsdn, arena, 0);
1827 	else
1828 		arena_maybe_purge(tsdn, arena);
1829 	malloc_mutex_unlock(tsdn, &arena->lock);
1830 }
1831 
1832 static void
1833 arena_achunk_prof_reset(tsd_t *tsd, arena_t *arena, arena_chunk_t *chunk)
1834 {
1835 	size_t pageind, npages;
1836 
1837 	cassert(config_prof);
1838 	assert(opt_prof);
1839 
1840 	/*
1841 	 * Iterate over the allocated runs and remove profiled allocations from
1842 	 * the sample set.
1843 	 */
1844 	for (pageind = map_bias; pageind < chunk_npages; pageind += npages) {
1845 		if (arena_mapbits_allocated_get(chunk, pageind) != 0) {
1846 			if (arena_mapbits_large_get(chunk, pageind) != 0) {
1847 				void *ptr = (void *)((uintptr_t)chunk + (pageind
1848 				    << LG_PAGE));
1849 				size_t usize = isalloc(tsd_tsdn(tsd), ptr,
1850 				    config_prof);
1851 
1852 				prof_free(tsd, ptr, usize);
1853 				npages = arena_mapbits_large_size_get(chunk,
1854 				    pageind) >> LG_PAGE;
1855 			} else {
1856 				/* Skip small run. */
1857 				size_t binind = arena_mapbits_binind_get(chunk,
1858 				    pageind);
1859 				arena_bin_info_t *bin_info =
1860 				    &arena_bin_info[binind];
1861 				npages = bin_info->run_size >> LG_PAGE;
1862 			}
1863 		} else {
1864 			/* Skip unallocated run. */
1865 			npages = arena_mapbits_unallocated_size_get(chunk,
1866 			    pageind) >> LG_PAGE;
1867 		}
1868 		assert(pageind + npages <= chunk_npages);
1869 	}
1870 }
1871 
1872 void
1873 arena_reset(tsd_t *tsd, arena_t *arena)
1874 {
1875 	unsigned i;
1876 	extent_node_t *node;
1877 
1878 	/*
1879 	 * Locking in this function is unintuitive.  The caller guarantees that
1880 	 * no concurrent operations are happening in this arena, but there are
1881 	 * still reasons that some locking is necessary:
1882 	 *
1883 	 * - Some of the functions in the transitive closure of calls assume
1884 	 *   appropriate locks are held, and in some cases these locks are
1885 	 *   temporarily dropped to avoid lock order reversal or deadlock due to
1886 	 *   reentry.
1887 	 * - mallctl("epoch", ...) may concurrently refresh stats.  While
1888 	 *   strictly speaking this is a "concurrent operation", disallowing
1889 	 *   stats refreshes would impose an inconvenient burden.
1890 	 */
1891 
1892 	/* Remove large allocations from prof sample set. */
1893 	if (config_prof && opt_prof) {
1894 		ql_foreach(node, &arena->achunks, ql_link) {
1895 			arena_achunk_prof_reset(tsd, arena,
1896 			    extent_node_addr_get(node));
1897 		}
1898 	}
1899 
1900 	/* Reset curruns for large size classes. */
1901 	if (config_stats) {
1902 		for (i = 0; i < nlclasses; i++)
1903 			arena->stats.lstats[i].curruns = 0;
1904 	}
1905 
1906 	/* Huge allocations. */
1907 	malloc_mutex_lock(tsd_tsdn(tsd), &arena->huge_mtx);
1908 	for (node = ql_last(&arena->huge, ql_link); node != NULL; node =
1909 	    ql_last(&arena->huge, ql_link)) {
1910 		void *ptr = extent_node_addr_get(node);
1911 		size_t usize;
1912 
1913 		malloc_mutex_unlock(tsd_tsdn(tsd), &arena->huge_mtx);
1914 		if (config_stats || (config_prof && opt_prof))
1915 			usize = isalloc(tsd_tsdn(tsd), ptr, config_prof);
1916 		/* Remove huge allocation from prof sample set. */
1917 		if (config_prof && opt_prof)
1918 			prof_free(tsd, ptr, usize);
1919 		huge_dalloc(tsd_tsdn(tsd), ptr);
1920 		malloc_mutex_lock(tsd_tsdn(tsd), &arena->huge_mtx);
1921 		/* Cancel out unwanted effects on stats. */
1922 		if (config_stats)
1923 			arena_huge_reset_stats_cancel(arena, usize);
1924 	}
1925 	malloc_mutex_unlock(tsd_tsdn(tsd), &arena->huge_mtx);
1926 
1927 	malloc_mutex_lock(tsd_tsdn(tsd), &arena->lock);
1928 
1929 	/* Bins. */
1930 	for (i = 0; i < NBINS; i++) {
1931 		arena_bin_t *bin = &arena->bins[i];
1932 		malloc_mutex_lock(tsd_tsdn(tsd), &bin->lock);
1933 		bin->runcur = NULL;
1934 		arena_run_heap_new(&bin->runs);
1935 		if (config_stats) {
1936 			bin->stats.curregs = 0;
1937 			bin->stats.curruns = 0;
1938 		}
1939 		malloc_mutex_unlock(tsd_tsdn(tsd), &bin->lock);
1940 	}
1941 
1942 	/*
1943 	 * Re-initialize runs_dirty such that the chunks_cache and runs_dirty
1944 	 * chains directly correspond.
1945 	 */
1946 	qr_new(&arena->runs_dirty, rd_link);
1947 	for (node = qr_next(&arena->chunks_cache, cc_link);
1948 	    node != &arena->chunks_cache; node = qr_next(node, cc_link)) {
1949 		qr_new(&node->rd, rd_link);
1950 		qr_meld(&arena->runs_dirty, &node->rd, rd_link);
1951 	}
1952 
1953 	/* Arena chunks. */
1954 	for (node = ql_last(&arena->achunks, ql_link); node != NULL; node =
1955 	    ql_last(&arena->achunks, ql_link)) {
1956 		ql_remove(&arena->achunks, node, ql_link);
1957 		arena_chunk_discard(tsd_tsdn(tsd), arena,
1958 		    extent_node_addr_get(node));
1959 	}
1960 
1961 	/* Spare. */
1962 	if (arena->spare != NULL) {
1963 		arena_chunk_discard(tsd_tsdn(tsd), arena, arena->spare);
1964 		arena->spare = NULL;
1965 	}
1966 
1967 	assert(!arena->purging);
1968 	arena->nactive = 0;
1969 
1970 	for(i = 0; i < runs_avail_nclasses; i++)
1971 		arena_run_heap_new(&arena->runs_avail[i]);
1972 
1973 	malloc_mutex_unlock(tsd_tsdn(tsd), &arena->lock);
1974 }
1975 
1976 static void
1977 arena_run_coalesce(arena_t *arena, arena_chunk_t *chunk, size_t *p_size,
1978     size_t *p_run_ind, size_t *p_run_pages, size_t flag_dirty,
1979     size_t flag_decommitted)
1980 {
1981 	size_t size = *p_size;
1982 	size_t run_ind = *p_run_ind;
1983 	size_t run_pages = *p_run_pages;
1984 
1985 	/* Try to coalesce forward. */
1986 	if (run_ind + run_pages < chunk_npages &&
1987 	    arena_mapbits_allocated_get(chunk, run_ind+run_pages) == 0 &&
1988 	    arena_mapbits_dirty_get(chunk, run_ind+run_pages) == flag_dirty &&
1989 	    arena_mapbits_decommitted_get(chunk, run_ind+run_pages) ==
1990 	    flag_decommitted) {
1991 		size_t nrun_size = arena_mapbits_unallocated_size_get(chunk,
1992 		    run_ind+run_pages);
1993 		size_t nrun_pages = nrun_size >> LG_PAGE;
1994 
1995 		/*
1996 		 * Remove successor from runs_avail; the coalesced run is
1997 		 * inserted later.
1998 		 */
1999 		assert(arena_mapbits_unallocated_size_get(chunk,
2000 		    run_ind+run_pages+nrun_pages-1) == nrun_size);
2001 		assert(arena_mapbits_dirty_get(chunk,
2002 		    run_ind+run_pages+nrun_pages-1) == flag_dirty);
2003 		assert(arena_mapbits_decommitted_get(chunk,
2004 		    run_ind+run_pages+nrun_pages-1) == flag_decommitted);
2005 		arena_avail_remove(arena, chunk, run_ind+run_pages, nrun_pages);
2006 
2007 		/*
2008 		 * If the successor is dirty, remove it from the set of dirty
2009 		 * pages.
2010 		 */
2011 		if (flag_dirty != 0) {
2012 			arena_run_dirty_remove(arena, chunk, run_ind+run_pages,
2013 			    nrun_pages);
2014 		}
2015 
2016 		size += nrun_size;
2017 		run_pages += nrun_pages;
2018 
2019 		arena_mapbits_unallocated_size_set(chunk, run_ind, size);
2020 		arena_mapbits_unallocated_size_set(chunk, run_ind+run_pages-1,
2021 		    size);
2022 	}
2023 
2024 	/* Try to coalesce backward. */
2025 	if (run_ind > map_bias && arena_mapbits_allocated_get(chunk,
2026 	    run_ind-1) == 0 && arena_mapbits_dirty_get(chunk, run_ind-1) ==
2027 	    flag_dirty && arena_mapbits_decommitted_get(chunk, run_ind-1) ==
2028 	    flag_decommitted) {
2029 		size_t prun_size = arena_mapbits_unallocated_size_get(chunk,
2030 		    run_ind-1);
2031 		size_t prun_pages = prun_size >> LG_PAGE;
2032 
2033 		run_ind -= prun_pages;
2034 
2035 		/*
2036 		 * Remove predecessor from runs_avail; the coalesced run is
2037 		 * inserted later.
2038 		 */
2039 		assert(arena_mapbits_unallocated_size_get(chunk, run_ind) ==
2040 		    prun_size);
2041 		assert(arena_mapbits_dirty_get(chunk, run_ind) == flag_dirty);
2042 		assert(arena_mapbits_decommitted_get(chunk, run_ind) ==
2043 		    flag_decommitted);
2044 		arena_avail_remove(arena, chunk, run_ind, prun_pages);
2045 
2046 		/*
2047 		 * If the predecessor is dirty, remove it from the set of dirty
2048 		 * pages.
2049 		 */
2050 		if (flag_dirty != 0) {
2051 			arena_run_dirty_remove(arena, chunk, run_ind,
2052 			    prun_pages);
2053 		}
2054 
2055 		size += prun_size;
2056 		run_pages += prun_pages;
2057 
2058 		arena_mapbits_unallocated_size_set(chunk, run_ind, size);
2059 		arena_mapbits_unallocated_size_set(chunk, run_ind+run_pages-1,
2060 		    size);
2061 	}
2062 
2063 	*p_size = size;
2064 	*p_run_ind = run_ind;
2065 	*p_run_pages = run_pages;
2066 }
2067 
2068 static size_t
2069 arena_run_size_get(arena_t *arena, arena_chunk_t *chunk, arena_run_t *run,
2070     size_t run_ind)
2071 {
2072 	size_t size;
2073 
2074 	assert(run_ind >= map_bias);
2075 	assert(run_ind < chunk_npages);
2076 
2077 	if (arena_mapbits_large_get(chunk, run_ind) != 0) {
2078 		size = arena_mapbits_large_size_get(chunk, run_ind);
2079 		assert(size == PAGE || arena_mapbits_large_size_get(chunk,
2080 		    run_ind+(size>>LG_PAGE)-1) == 0);
2081 	} else {
2082 		arena_bin_info_t *bin_info = &arena_bin_info[run->binind];
2083 		size = bin_info->run_size;
2084 	}
2085 
2086 	return (size);
2087 }
2088 
2089 static void
2090 arena_run_dalloc(tsdn_t *tsdn, arena_t *arena, arena_run_t *run, bool dirty,
2091     bool cleaned, bool decommitted)
2092 {
2093 	arena_chunk_t *chunk;
2094 	arena_chunk_map_misc_t *miscelm;
2095 	size_t size, run_ind, run_pages, flag_dirty, flag_decommitted;
2096 
2097 	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run);
2098 	miscelm = arena_run_to_miscelm(run);
2099 	run_ind = arena_miscelm_to_pageind(miscelm);
2100 	assert(run_ind >= map_bias);
2101 	assert(run_ind < chunk_npages);
2102 	size = arena_run_size_get(arena, chunk, run, run_ind);
2103 	run_pages = (size >> LG_PAGE);
2104 	arena_nactive_sub(arena, run_pages);
2105 
2106 	/*
2107 	 * The run is dirty if the caller claims to have dirtied it, as well as
2108 	 * if it was already dirty before being allocated and the caller
2109 	 * doesn't claim to have cleaned it.
2110 	 */
2111 	assert(arena_mapbits_dirty_get(chunk, run_ind) ==
2112 	    arena_mapbits_dirty_get(chunk, run_ind+run_pages-1));
2113 	if (!cleaned && !decommitted && arena_mapbits_dirty_get(chunk, run_ind)
2114 	    != 0)
2115 		dirty = true;
2116 	flag_dirty = dirty ? CHUNK_MAP_DIRTY : 0;
2117 	flag_decommitted = decommitted ? CHUNK_MAP_DECOMMITTED : 0;
2118 
2119 	/* Mark pages as unallocated in the chunk map. */
2120 	if (dirty || decommitted) {
2121 		size_t flags = flag_dirty | flag_decommitted;
2122 		arena_mapbits_unallocated_set(chunk, run_ind, size, flags);
2123 		arena_mapbits_unallocated_set(chunk, run_ind+run_pages-1, size,
2124 		    flags);
2125 	} else {
2126 		arena_mapbits_unallocated_set(chunk, run_ind, size,
2127 		    arena_mapbits_unzeroed_get(chunk, run_ind));
2128 		arena_mapbits_unallocated_set(chunk, run_ind+run_pages-1, size,
2129 		    arena_mapbits_unzeroed_get(chunk, run_ind+run_pages-1));
2130 	}
2131 
2132 	arena_run_coalesce(arena, chunk, &size, &run_ind, &run_pages,
2133 	    flag_dirty, flag_decommitted);
2134 
2135 	/* Insert into runs_avail, now that coalescing is complete. */
2136 	assert(arena_mapbits_unallocated_size_get(chunk, run_ind) ==
2137 	    arena_mapbits_unallocated_size_get(chunk, run_ind+run_pages-1));
2138 	assert(arena_mapbits_dirty_get(chunk, run_ind) ==
2139 	    arena_mapbits_dirty_get(chunk, run_ind+run_pages-1));
2140 	assert(arena_mapbits_decommitted_get(chunk, run_ind) ==
2141 	    arena_mapbits_decommitted_get(chunk, run_ind+run_pages-1));
2142 	arena_avail_insert(arena, chunk, run_ind, run_pages);
2143 
2144 	if (dirty)
2145 		arena_run_dirty_insert(arena, chunk, run_ind, run_pages);
2146 
2147 	/* Deallocate chunk if it is now completely unused. */
2148 	if (size == arena_maxrun) {
2149 		assert(run_ind == map_bias);
2150 		assert(run_pages == (arena_maxrun >> LG_PAGE));
2151 		arena_chunk_dalloc(tsdn, arena, chunk);
2152 	}
2153 
2154 	/*
2155 	 * It is okay to do dirty page processing here even if the chunk was
2156 	 * deallocated above, since in that case it is the spare.  Waiting
2157 	 * until after possible chunk deallocation to do dirty processing
2158 	 * allows for an old spare to be fully deallocated, thus decreasing the
2159 	 * chances of spuriously crossing the dirty page purging threshold.
2160 	 */
2161 	if (dirty)
2162 		arena_maybe_purge(tsdn, arena);
2163 }
2164 
2165 static void
2166 arena_run_trim_head(tsdn_t *tsdn, arena_t *arena, arena_chunk_t *chunk,
2167     arena_run_t *run, size_t oldsize, size_t newsize)
2168 {
2169 	arena_chunk_map_misc_t *miscelm = arena_run_to_miscelm(run);
2170 	size_t pageind = arena_miscelm_to_pageind(miscelm);
2171 	size_t head_npages = (oldsize - newsize) >> LG_PAGE;
2172 	size_t flag_dirty = arena_mapbits_dirty_get(chunk, pageind);
2173 	size_t flag_decommitted = arena_mapbits_decommitted_get(chunk, pageind);
2174 	size_t flag_unzeroed_mask = (flag_dirty | flag_decommitted) == 0 ?
2175 	    CHUNK_MAP_UNZEROED : 0;
2176 
2177 	assert(oldsize > newsize);
2178 
2179 	/*
2180 	 * Update the chunk map so that arena_run_dalloc() can treat the
2181 	 * leading run as separately allocated.  Set the last element of each
2182 	 * run first, in case of single-page runs.
2183 	 */
2184 	assert(arena_mapbits_large_size_get(chunk, pageind) == oldsize);
2185 	arena_mapbits_large_set(chunk, pageind+head_npages-1, 0, flag_dirty |
2186 	    (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk,
2187 	    pageind+head_npages-1)));
2188 	arena_mapbits_large_set(chunk, pageind, oldsize-newsize, flag_dirty |
2189 	    (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk, pageind)));
2190 
2191 	if (config_debug) {
2192 		UNUSED size_t tail_npages = newsize >> LG_PAGE;
2193 		assert(arena_mapbits_large_size_get(chunk,
2194 		    pageind+head_npages+tail_npages-1) == 0);
2195 		assert(arena_mapbits_dirty_get(chunk,
2196 		    pageind+head_npages+tail_npages-1) == flag_dirty);
2197 	}
2198 	arena_mapbits_large_set(chunk, pageind+head_npages, newsize,
2199 	    flag_dirty | (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk,
2200 	    pageind+head_npages)));
2201 
2202 	arena_run_dalloc(tsdn, arena, run, false, false, (flag_decommitted !=
2203 	    0));
2204 }
2205 
2206 static void
2207 arena_run_trim_tail(tsdn_t *tsdn, arena_t *arena, arena_chunk_t *chunk,
2208     arena_run_t *run, size_t oldsize, size_t newsize, bool dirty)
2209 {
2210 	arena_chunk_map_misc_t *miscelm = arena_run_to_miscelm(run);
2211 	size_t pageind = arena_miscelm_to_pageind(miscelm);
2212 	size_t head_npages = newsize >> LG_PAGE;
2213 	size_t flag_dirty = arena_mapbits_dirty_get(chunk, pageind);
2214 	size_t flag_decommitted = arena_mapbits_decommitted_get(chunk, pageind);
2215 	size_t flag_unzeroed_mask = (flag_dirty | flag_decommitted) == 0 ?
2216 	    CHUNK_MAP_UNZEROED : 0;
2217 	arena_chunk_map_misc_t *tail_miscelm;
2218 	arena_run_t *tail_run;
2219 
2220 	assert(oldsize > newsize);
2221 
2222 	/*
2223 	 * Update the chunk map so that arena_run_dalloc() can treat the
2224 	 * trailing run as separately allocated.  Set the last element of each
2225 	 * run first, in case of single-page runs.
2226 	 */
2227 	assert(arena_mapbits_large_size_get(chunk, pageind) == oldsize);
2228 	arena_mapbits_large_set(chunk, pageind+head_npages-1, 0, flag_dirty |
2229 	    (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk,
2230 	    pageind+head_npages-1)));
2231 	arena_mapbits_large_set(chunk, pageind, newsize, flag_dirty |
2232 	    (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk, pageind)));
2233 
2234 	if (config_debug) {
2235 		UNUSED size_t tail_npages = (oldsize - newsize) >> LG_PAGE;
2236 		assert(arena_mapbits_large_size_get(chunk,
2237 		    pageind+head_npages+tail_npages-1) == 0);
2238 		assert(arena_mapbits_dirty_get(chunk,
2239 		    pageind+head_npages+tail_npages-1) == flag_dirty);
2240 	}
2241 	arena_mapbits_large_set(chunk, pageind+head_npages, oldsize-newsize,
2242 	    flag_dirty | (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk,
2243 	    pageind+head_npages)));
2244 
2245 	tail_miscelm = arena_miscelm_get_mutable(chunk, pageind + head_npages);
2246 	tail_run = &tail_miscelm->run;
2247 	arena_run_dalloc(tsdn, arena, tail_run, dirty, false, (flag_decommitted
2248 	    != 0));
2249 }
2250 
2251 static void
2252 arena_bin_runs_insert(arena_bin_t *bin, arena_run_t *run)
2253 {
2254 	arena_chunk_map_misc_t *miscelm = arena_run_to_miscelm(run);
2255 
2256 	arena_run_heap_insert(&bin->runs, miscelm);
2257 }
2258 
2259 static arena_run_t *
2260 arena_bin_nonfull_run_tryget(arena_bin_t *bin)
2261 {
2262 	arena_chunk_map_misc_t *miscelm;
2263 
2264 	miscelm = arena_run_heap_remove_first(&bin->runs);
2265 	if (miscelm == NULL)
2266 		return (NULL);
2267 	if (config_stats)
2268 		bin->stats.reruns++;
2269 
2270 	return (&miscelm->run);
2271 }
2272 
2273 static arena_run_t *
2274 arena_bin_nonfull_run_get(tsdn_t *tsdn, arena_t *arena, arena_bin_t *bin)
2275 {
2276 	arena_run_t *run;
2277 	szind_t binind;
2278 	arena_bin_info_t *bin_info;
2279 
2280 	/* Look for a usable run. */
2281 	run = arena_bin_nonfull_run_tryget(bin);
2282 	if (run != NULL)
2283 		return (run);
2284 	/* No existing runs have any space available. */
2285 
2286 	binind = arena_bin_index(arena, bin);
2287 	bin_info = &arena_bin_info[binind];
2288 
2289 	/* Allocate a new run. */
2290 	malloc_mutex_unlock(tsdn, &bin->lock);
2291 	/******************************/
2292 	malloc_mutex_lock(tsdn, &arena->lock);
2293 	run = arena_run_alloc_small(tsdn, arena, bin_info->run_size, binind);
2294 	if (run != NULL) {
2295 		/* Initialize run internals. */
2296 		run->binind = binind;
2297 		run->nfree = bin_info->nregs;
2298 		bitmap_init(run->bitmap, &bin_info->bitmap_info);
2299 	}
2300 	malloc_mutex_unlock(tsdn, &arena->lock);
2301 	/********************************/
2302 	malloc_mutex_lock(tsdn, &bin->lock);
2303 	if (run != NULL) {
2304 		if (config_stats) {
2305 			bin->stats.nruns++;
2306 			bin->stats.curruns++;
2307 		}
2308 		return (run);
2309 	}
2310 
2311 	/*
2312 	 * arena_run_alloc_small() failed, but another thread may have made
2313 	 * sufficient memory available while this one dropped bin->lock above,
2314 	 * so search one more time.
2315 	 */
2316 	run = arena_bin_nonfull_run_tryget(bin);
2317 	if (run != NULL)
2318 		return (run);
2319 
2320 	return (NULL);
2321 }
2322 
2323 /* Re-fill bin->runcur, then call arena_run_reg_alloc(). */
2324 static void *
2325 arena_bin_malloc_hard(tsdn_t *tsdn, arena_t *arena, arena_bin_t *bin)
2326 {
2327 	szind_t binind;
2328 	arena_bin_info_t *bin_info;
2329 	arena_run_t *run;
2330 
2331 	binind = arena_bin_index(arena, bin);
2332 	bin_info = &arena_bin_info[binind];
2333 	bin->runcur = NULL;
2334 	run = arena_bin_nonfull_run_get(tsdn, arena, bin);
2335 	if (bin->runcur != NULL && bin->runcur->nfree > 0) {
2336 		/*
2337 		 * Another thread updated runcur while this one ran without the
2338 		 * bin lock in arena_bin_nonfull_run_get().
2339 		 */
2340 		void *ret;
2341 		assert(bin->runcur->nfree > 0);
2342 		ret = arena_run_reg_alloc(bin->runcur, bin_info);
2343 		if (run != NULL) {
2344 			arena_chunk_t *chunk;
2345 
2346 			/*
2347 			 * arena_run_alloc_small() may have allocated run, or
2348 			 * it may have pulled run from the bin's run tree.
2349 			 * Therefore it is unsafe to make any assumptions about
2350 			 * how run has previously been used, and
2351 			 * arena_bin_lower_run() must be called, as if a region
2352 			 * were just deallocated from the run.
2353 			 */
2354 			chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run);
2355 			if (run->nfree == bin_info->nregs) {
2356 				arena_dalloc_bin_run(tsdn, arena, chunk, run,
2357 				    bin);
2358 			} else
2359 				arena_bin_lower_run(arena, chunk, run, bin);
2360 		}
2361 		return (ret);
2362 	}
2363 
2364 	if (run == NULL)
2365 		return (NULL);
2366 
2367 	bin->runcur = run;
2368 
2369 	assert(bin->runcur->nfree > 0);
2370 
2371 	return (arena_run_reg_alloc(bin->runcur, bin_info));
2372 }
2373 
2374 void
2375 arena_tcache_fill_small(tsdn_t *tsdn, arena_t *arena, tcache_bin_t *tbin,
2376     szind_t binind, uint64_t prof_accumbytes)
2377 {
2378 	unsigned i, nfill;
2379 	arena_bin_t *bin;
2380 
2381 	assert(tbin->ncached == 0);
2382 
2383 	if (config_prof && arena_prof_accum(tsdn, arena, prof_accumbytes))
2384 		prof_idump(tsdn);
2385 	bin = &arena->bins[binind];
2386 	malloc_mutex_lock(tsdn, &bin->lock);
2387 	for (i = 0, nfill = (tcache_bin_info[binind].ncached_max >>
2388 	    tbin->lg_fill_div); i < nfill; i++) {
2389 		arena_run_t *run;
2390 		void *ptr;
2391 		if ((run = bin->runcur) != NULL && run->nfree > 0)
2392 			ptr = arena_run_reg_alloc(run, &arena_bin_info[binind]);
2393 		else
2394 			ptr = arena_bin_malloc_hard(tsdn, arena, bin);
2395 		if (ptr == NULL) {
2396 			/*
2397 			 * OOM.  tbin->avail isn't yet filled down to its first
2398 			 * element, so the successful allocations (if any) must
2399 			 * be moved just before tbin->avail before bailing out.
2400 			 */
2401 			if (i > 0) {
2402 				memmove(tbin->avail - i, tbin->avail - nfill,
2403 				    i * sizeof(void *));
2404 			}
2405 			break;
2406 		}
2407 		if (config_fill && unlikely(opt_junk_alloc)) {
2408 			arena_alloc_junk_small(ptr, &arena_bin_info[binind],
2409 			    true);
2410 		}
2411 		/* Insert such that low regions get used first. */
2412 		*(tbin->avail - nfill + i) = ptr;
2413 	}
2414 	if (config_stats) {
2415 		bin->stats.nmalloc += i;
2416 		bin->stats.nrequests += tbin->tstats.nrequests;
2417 		bin->stats.curregs += i;
2418 		bin->stats.nfills++;
2419 		tbin->tstats.nrequests = 0;
2420 	}
2421 	malloc_mutex_unlock(tsdn, &bin->lock);
2422 	tbin->ncached = i;
2423 	arena_decay_tick(tsdn, arena);
2424 }
2425 
2426 void
2427 arena_alloc_junk_small(void *ptr, arena_bin_info_t *bin_info, bool zero)
2428 {
2429 
2430 	size_t redzone_size = bin_info->redzone_size;
2431 
2432 	if (zero) {
2433 		memset((void *)((uintptr_t)ptr - redzone_size),
2434 		    JEMALLOC_ALLOC_JUNK, redzone_size);
2435 		memset((void *)((uintptr_t)ptr + bin_info->reg_size),
2436 		    JEMALLOC_ALLOC_JUNK, redzone_size);
2437 	} else {
2438 		memset((void *)((uintptr_t)ptr - redzone_size),
2439 		    JEMALLOC_ALLOC_JUNK, bin_info->reg_interval);
2440 	}
2441 }
2442 
2443 #ifdef JEMALLOC_JET
2444 #undef arena_redzone_corruption
2445 #define	arena_redzone_corruption JEMALLOC_N(n_arena_redzone_corruption)
2446 #endif
2447 static void
2448 arena_redzone_corruption(void *ptr, size_t usize, bool after,
2449     size_t offset, uint8_t byte)
2450 {
2451 
2452 	malloc_printf("<jemalloc>: Corrupt redzone %zu byte%s %s %p "
2453 	    "(size %zu), byte=%#x\n", offset, (offset == 1) ? "" : "s",
2454 	    after ? "after" : "before", ptr, usize, byte);
2455 }
2456 #ifdef JEMALLOC_JET
2457 #undef arena_redzone_corruption
2458 #define	arena_redzone_corruption JEMALLOC_N(arena_redzone_corruption)
2459 arena_redzone_corruption_t *arena_redzone_corruption =
2460     JEMALLOC_N(n_arena_redzone_corruption);
2461 #endif
2462 
2463 static void
2464 arena_redzones_validate(void *ptr, arena_bin_info_t *bin_info, bool reset)
2465 {
2466 	bool error = false;
2467 
2468 	if (opt_junk_alloc) {
2469 		size_t size = bin_info->reg_size;
2470 		size_t redzone_size = bin_info->redzone_size;
2471 		size_t i;
2472 
2473 		for (i = 1; i <= redzone_size; i++) {
2474 			uint8_t *byte = (uint8_t *)((uintptr_t)ptr - i);
2475 			if (*byte != JEMALLOC_ALLOC_JUNK) {
2476 				error = true;
2477 				arena_redzone_corruption(ptr, size, false, i,
2478 				    *byte);
2479 				if (reset)
2480 					*byte = JEMALLOC_ALLOC_JUNK;
2481 			}
2482 		}
2483 		for (i = 0; i < redzone_size; i++) {
2484 			uint8_t *byte = (uint8_t *)((uintptr_t)ptr + size + i);
2485 			if (*byte != JEMALLOC_ALLOC_JUNK) {
2486 				error = true;
2487 				arena_redzone_corruption(ptr, size, true, i,
2488 				    *byte);
2489 				if (reset)
2490 					*byte = JEMALLOC_ALLOC_JUNK;
2491 			}
2492 		}
2493 	}
2494 
2495 	if (opt_abort && error)
2496 		abort();
2497 }
2498 
2499 #ifdef JEMALLOC_JET
2500 #undef arena_dalloc_junk_small
2501 #define	arena_dalloc_junk_small JEMALLOC_N(n_arena_dalloc_junk_small)
2502 #endif
2503 void
2504 arena_dalloc_junk_small(void *ptr, arena_bin_info_t *bin_info)
2505 {
2506 	size_t redzone_size = bin_info->redzone_size;
2507 
2508 	arena_redzones_validate(ptr, bin_info, false);
2509 	memset((void *)((uintptr_t)ptr - redzone_size), JEMALLOC_FREE_JUNK,
2510 	    bin_info->reg_interval);
2511 }
2512 #ifdef JEMALLOC_JET
2513 #undef arena_dalloc_junk_small
2514 #define	arena_dalloc_junk_small JEMALLOC_N(arena_dalloc_junk_small)
2515 arena_dalloc_junk_small_t *arena_dalloc_junk_small =
2516     JEMALLOC_N(n_arena_dalloc_junk_small);
2517 #endif
2518 
2519 void
2520 arena_quarantine_junk_small(void *ptr, size_t usize)
2521 {
2522 	szind_t binind;
2523 	arena_bin_info_t *bin_info;
2524 	cassert(config_fill);
2525 	assert(opt_junk_free);
2526 	assert(opt_quarantine);
2527 	assert(usize <= SMALL_MAXCLASS);
2528 
2529 	binind = size2index(usize);
2530 	bin_info = &arena_bin_info[binind];
2531 	arena_redzones_validate(ptr, bin_info, true);
2532 }
2533 
2534 static void *
2535 arena_malloc_small(tsdn_t *tsdn, arena_t *arena, szind_t binind, bool zero)
2536 {
2537 	void *ret;
2538 	arena_bin_t *bin;
2539 	size_t usize;
2540 	arena_run_t *run;
2541 
2542 	assert(binind < NBINS);
2543 	bin = &arena->bins[binind];
2544 	usize = index2size(binind);
2545 
2546 	malloc_mutex_lock(tsdn, &bin->lock);
2547 	if ((run = bin->runcur) != NULL && run->nfree > 0)
2548 		ret = arena_run_reg_alloc(run, &arena_bin_info[binind]);
2549 	else
2550 		ret = arena_bin_malloc_hard(tsdn, arena, bin);
2551 
2552 	if (ret == NULL) {
2553 		malloc_mutex_unlock(tsdn, &bin->lock);
2554 		return (NULL);
2555 	}
2556 
2557 	if (config_stats) {
2558 		bin->stats.nmalloc++;
2559 		bin->stats.nrequests++;
2560 		bin->stats.curregs++;
2561 	}
2562 	malloc_mutex_unlock(tsdn, &bin->lock);
2563 	if (config_prof && !isthreaded && arena_prof_accum(tsdn, arena, usize))
2564 		prof_idump(tsdn);
2565 
2566 	if (!zero) {
2567 		if (config_fill) {
2568 			if (unlikely(opt_junk_alloc)) {
2569 				arena_alloc_junk_small(ret,
2570 				    &arena_bin_info[binind], false);
2571 			} else if (unlikely(opt_zero))
2572 				memset(ret, 0, usize);
2573 		}
2574 		JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED(ret, usize);
2575 	} else {
2576 		if (config_fill && unlikely(opt_junk_alloc)) {
2577 			arena_alloc_junk_small(ret, &arena_bin_info[binind],
2578 			    true);
2579 		}
2580 		JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED(ret, usize);
2581 		memset(ret, 0, usize);
2582 	}
2583 
2584 	arena_decay_tick(tsdn, arena);
2585 	return (ret);
2586 }
2587 
2588 void *
2589 arena_malloc_large(tsdn_t *tsdn, arena_t *arena, szind_t binind, bool zero)
2590 {
2591 	void *ret;
2592 	size_t usize;
2593 	uintptr_t random_offset;
2594 	arena_run_t *run;
2595 	arena_chunk_map_misc_t *miscelm;
2596 	UNUSED bool idump JEMALLOC_CC_SILENCE_INIT(false);
2597 
2598 	/* Large allocation. */
2599 	usize = index2size(binind);
2600 	malloc_mutex_lock(tsdn, &arena->lock);
2601 	if (config_cache_oblivious) {
2602 		uint64_t r;
2603 
2604 		/*
2605 		 * Compute a uniformly distributed offset within the first page
2606 		 * that is a multiple of the cacheline size, e.g. [0 .. 63) * 64
2607 		 * for 4 KiB pages and 64-byte cachelines.
2608 		 */
2609 		r = prng_lg_range(&arena->offset_state, LG_PAGE - LG_CACHELINE);
2610 		random_offset = ((uintptr_t)r) << LG_CACHELINE;
2611 	} else
2612 		random_offset = 0;
2613 	run = arena_run_alloc_large(tsdn, arena, usize + large_pad, zero);
2614 	if (run == NULL) {
2615 		malloc_mutex_unlock(tsdn, &arena->lock);
2616 		return (NULL);
2617 	}
2618 	miscelm = arena_run_to_miscelm(run);
2619 	ret = (void *)((uintptr_t)arena_miscelm_to_rpages(miscelm) +
2620 	    random_offset);
2621 	if (config_stats) {
2622 		szind_t index = binind - NBINS;
2623 
2624 		arena->stats.nmalloc_large++;
2625 		arena->stats.nrequests_large++;
2626 		arena->stats.allocated_large += usize;
2627 		arena->stats.lstats[index].nmalloc++;
2628 		arena->stats.lstats[index].nrequests++;
2629 		arena->stats.lstats[index].curruns++;
2630 	}
2631 	if (config_prof)
2632 		idump = arena_prof_accum_locked(arena, usize);
2633 	malloc_mutex_unlock(tsdn, &arena->lock);
2634 	if (config_prof && idump)
2635 		prof_idump(tsdn);
2636 
2637 	if (!zero) {
2638 		if (config_fill) {
2639 			if (unlikely(opt_junk_alloc))
2640 				memset(ret, JEMALLOC_ALLOC_JUNK, usize);
2641 			else if (unlikely(opt_zero))
2642 				memset(ret, 0, usize);
2643 		}
2644 	}
2645 
2646 	arena_decay_tick(tsdn, arena);
2647 	return (ret);
2648 }
2649 
2650 void *
2651 arena_malloc_hard(tsdn_t *tsdn, arena_t *arena, size_t size, szind_t ind,
2652     bool zero)
2653 {
2654 
2655 	assert(!tsdn_null(tsdn) || arena != NULL);
2656 
2657 	if (likely(!tsdn_null(tsdn)))
2658 		arena = arena_choose(tsdn_tsd(tsdn), arena);
2659 	if (unlikely(arena == NULL))
2660 		return (NULL);
2661 
2662 	if (likely(size <= SMALL_MAXCLASS))
2663 		return (arena_malloc_small(tsdn, arena, ind, zero));
2664 	if (likely(size <= large_maxclass))
2665 		return (arena_malloc_large(tsdn, arena, ind, zero));
2666 	return (huge_malloc(tsdn, arena, index2size(ind), zero));
2667 }
2668 
2669 /* Only handles large allocations that require more than page alignment. */
2670 static void *
2671 arena_palloc_large(tsdn_t *tsdn, arena_t *arena, size_t usize, size_t alignment,
2672     bool zero)
2673 {
2674 	void *ret;
2675 	size_t alloc_size, leadsize, trailsize;
2676 	arena_run_t *run;
2677 	arena_chunk_t *chunk;
2678 	arena_chunk_map_misc_t *miscelm;
2679 	void *rpages;
2680 
2681 	assert(!tsdn_null(tsdn) || arena != NULL);
2682 	assert(usize == PAGE_CEILING(usize));
2683 
2684 	if (likely(!tsdn_null(tsdn)))
2685 		arena = arena_choose(tsdn_tsd(tsdn), arena);
2686 	if (unlikely(arena == NULL))
2687 		return (NULL);
2688 
2689 	alignment = PAGE_CEILING(alignment);
2690 	alloc_size = usize + large_pad + alignment - PAGE;
2691 
2692 	malloc_mutex_lock(tsdn, &arena->lock);
2693 	run = arena_run_alloc_large(tsdn, arena, alloc_size, false);
2694 	if (run == NULL) {
2695 		malloc_mutex_unlock(tsdn, &arena->lock);
2696 		return (NULL);
2697 	}
2698 	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run);
2699 	miscelm = arena_run_to_miscelm(run);
2700 	rpages = arena_miscelm_to_rpages(miscelm);
2701 
2702 	leadsize = ALIGNMENT_CEILING((uintptr_t)rpages, alignment) -
2703 	    (uintptr_t)rpages;
2704 	assert(alloc_size >= leadsize + usize);
2705 	trailsize = alloc_size - leadsize - usize - large_pad;
2706 	if (leadsize != 0) {
2707 		arena_chunk_map_misc_t *head_miscelm = miscelm;
2708 		arena_run_t *head_run = run;
2709 
2710 		miscelm = arena_miscelm_get_mutable(chunk,
2711 		    arena_miscelm_to_pageind(head_miscelm) + (leadsize >>
2712 		    LG_PAGE));
2713 		run = &miscelm->run;
2714 
2715 		arena_run_trim_head(tsdn, arena, chunk, head_run, alloc_size,
2716 		    alloc_size - leadsize);
2717 	}
2718 	if (trailsize != 0) {
2719 		arena_run_trim_tail(tsdn, arena, chunk, run, usize + large_pad +
2720 		    trailsize, usize + large_pad, false);
2721 	}
2722 	if (arena_run_init_large(arena, run, usize + large_pad, zero)) {
2723 		size_t run_ind =
2724 		    arena_miscelm_to_pageind(arena_run_to_miscelm(run));
2725 		bool dirty = (arena_mapbits_dirty_get(chunk, run_ind) != 0);
2726 		bool decommitted = (arena_mapbits_decommitted_get(chunk,
2727 		    run_ind) != 0);
2728 
2729 		assert(decommitted); /* Cause of OOM. */
2730 		arena_run_dalloc(tsdn, arena, run, dirty, false, decommitted);
2731 		malloc_mutex_unlock(tsdn, &arena->lock);
2732 		return (NULL);
2733 	}
2734 	ret = arena_miscelm_to_rpages(miscelm);
2735 
2736 	if (config_stats) {
2737 		szind_t index = size2index(usize) - NBINS;
2738 
2739 		arena->stats.nmalloc_large++;
2740 		arena->stats.nrequests_large++;
2741 		arena->stats.allocated_large += usize;
2742 		arena->stats.lstats[index].nmalloc++;
2743 		arena->stats.lstats[index].nrequests++;
2744 		arena->stats.lstats[index].curruns++;
2745 	}
2746 	malloc_mutex_unlock(tsdn, &arena->lock);
2747 
2748 	if (config_fill && !zero) {
2749 		if (unlikely(opt_junk_alloc))
2750 			memset(ret, JEMALLOC_ALLOC_JUNK, usize);
2751 		else if (unlikely(opt_zero))
2752 			memset(ret, 0, usize);
2753 	}
2754 	arena_decay_tick(tsdn, arena);
2755 	return (ret);
2756 }
2757 
2758 void *
2759 arena_palloc(tsdn_t *tsdn, arena_t *arena, size_t usize, size_t alignment,
2760     bool zero, tcache_t *tcache)
2761 {
2762 	void *ret;
2763 
2764 	if (usize <= SMALL_MAXCLASS && (alignment < PAGE || (alignment == PAGE
2765 	    && (usize & PAGE_MASK) == 0))) {
2766 		/* Small; alignment doesn't require special run placement. */
2767 		ret = arena_malloc(tsdn, arena, usize, size2index(usize), zero,
2768 		    tcache, true);
2769 	} else if (usize <= large_maxclass && alignment <= PAGE) {
2770 		/*
2771 		 * Large; alignment doesn't require special run placement.
2772 		 * However, the cached pointer may be at a random offset from
2773 		 * the base of the run, so do some bit manipulation to retrieve
2774 		 * the base.
2775 		 */
2776 		ret = arena_malloc(tsdn, arena, usize, size2index(usize), zero,
2777 		    tcache, true);
2778 		if (config_cache_oblivious)
2779 			ret = (void *)((uintptr_t)ret & ~PAGE_MASK);
2780 	} else {
2781 		if (likely(usize <= large_maxclass)) {
2782 			ret = arena_palloc_large(tsdn, arena, usize, alignment,
2783 			    zero);
2784 		} else if (likely(alignment <= chunksize))
2785 			ret = huge_malloc(tsdn, arena, usize, zero);
2786 		else {
2787 			ret = huge_palloc(tsdn, arena, usize, alignment, zero);
2788 		}
2789 	}
2790 	return (ret);
2791 }
2792 
2793 void
2794 arena_prof_promoted(tsdn_t *tsdn, const void *ptr, size_t size)
2795 {
2796 	arena_chunk_t *chunk;
2797 	size_t pageind;
2798 	szind_t binind;
2799 
2800 	cassert(config_prof);
2801 	assert(ptr != NULL);
2802 	assert(CHUNK_ADDR2BASE(ptr) != ptr);
2803 	assert(isalloc(tsdn, ptr, false) == LARGE_MINCLASS);
2804 	assert(isalloc(tsdn, ptr, true) == LARGE_MINCLASS);
2805 	assert(size <= SMALL_MAXCLASS);
2806 
2807 	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
2808 	pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
2809 	binind = size2index(size);
2810 	assert(binind < NBINS);
2811 	arena_mapbits_large_binind_set(chunk, pageind, binind);
2812 
2813 	assert(isalloc(tsdn, ptr, false) == LARGE_MINCLASS);
2814 	assert(isalloc(tsdn, ptr, true) == size);
2815 }
2816 
2817 static void
2818 arena_dissociate_bin_run(arena_chunk_t *chunk, arena_run_t *run,
2819     arena_bin_t *bin)
2820 {
2821 
2822 	/* Dissociate run from bin. */
2823 	if (run == bin->runcur)
2824 		bin->runcur = NULL;
2825 	else {
2826 		szind_t binind = arena_bin_index(extent_node_arena_get(
2827 		    &chunk->node), bin);
2828 		arena_bin_info_t *bin_info = &arena_bin_info[binind];
2829 
2830 		/*
2831 		 * The following block's conditional is necessary because if the
2832 		 * run only contains one region, then it never gets inserted
2833 		 * into the non-full runs tree.
2834 		 */
2835 		if (bin_info->nregs != 1) {
2836 			arena_chunk_map_misc_t *miscelm =
2837 			    arena_run_to_miscelm(run);
2838 
2839 			arena_run_heap_remove(&bin->runs, miscelm);
2840 		}
2841 	}
2842 }
2843 
2844 static void
2845 arena_dalloc_bin_run(tsdn_t *tsdn, arena_t *arena, arena_chunk_t *chunk,
2846     arena_run_t *run, arena_bin_t *bin)
2847 {
2848 
2849 	assert(run != bin->runcur);
2850 
2851 	malloc_mutex_unlock(tsdn, &bin->lock);
2852 	/******************************/
2853 	malloc_mutex_lock(tsdn, &arena->lock);
2854 	arena_run_dalloc(tsdn, arena, run, true, false, false);
2855 	malloc_mutex_unlock(tsdn, &arena->lock);
2856 	/****************************/
2857 	malloc_mutex_lock(tsdn, &bin->lock);
2858 	if (config_stats)
2859 		bin->stats.curruns--;
2860 }
2861 
2862 static void
2863 arena_bin_lower_run(arena_t *arena, arena_chunk_t *chunk, arena_run_t *run,
2864     arena_bin_t *bin)
2865 {
2866 
2867 	/*
2868 	 * Make sure that if bin->runcur is non-NULL, it refers to the lowest
2869 	 * non-full run.  It is okay to NULL runcur out rather than proactively
2870 	 * keeping it pointing at the lowest non-full run.
2871 	 */
2872 	if ((uintptr_t)run < (uintptr_t)bin->runcur) {
2873 		/* Switch runcur. */
2874 		if (bin->runcur->nfree > 0)
2875 			arena_bin_runs_insert(bin, bin->runcur);
2876 		bin->runcur = run;
2877 		if (config_stats)
2878 			bin->stats.reruns++;
2879 	} else
2880 		arena_bin_runs_insert(bin, run);
2881 }
2882 
2883 static void
2884 arena_dalloc_bin_locked_impl(tsdn_t *tsdn, arena_t *arena, arena_chunk_t *chunk,
2885     void *ptr, arena_chunk_map_bits_t *bitselm, bool junked)
2886 {
2887 	size_t pageind, rpages_ind;
2888 	arena_run_t *run;
2889 	arena_bin_t *bin;
2890 	arena_bin_info_t *bin_info;
2891 	szind_t binind;
2892 
2893 	pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
2894 	rpages_ind = pageind - arena_mapbits_small_runind_get(chunk, pageind);
2895 	run = &arena_miscelm_get_mutable(chunk, rpages_ind)->run;
2896 	binind = run->binind;
2897 	bin = &arena->bins[binind];
2898 	bin_info = &arena_bin_info[binind];
2899 
2900 	if (!junked && config_fill && unlikely(opt_junk_free))
2901 		arena_dalloc_junk_small(ptr, bin_info);
2902 
2903 	arena_run_reg_dalloc(run, ptr);
2904 	if (run->nfree == bin_info->nregs) {
2905 		arena_dissociate_bin_run(chunk, run, bin);
2906 		arena_dalloc_bin_run(tsdn, arena, chunk, run, bin);
2907 	} else if (run->nfree == 1 && run != bin->runcur)
2908 		arena_bin_lower_run(arena, chunk, run, bin);
2909 
2910 	if (config_stats) {
2911 		bin->stats.ndalloc++;
2912 		bin->stats.curregs--;
2913 	}
2914 }
2915 
2916 void
2917 arena_dalloc_bin_junked_locked(tsdn_t *tsdn, arena_t *arena,
2918     arena_chunk_t *chunk, void *ptr, arena_chunk_map_bits_t *bitselm)
2919 {
2920 
2921 	arena_dalloc_bin_locked_impl(tsdn, arena, chunk, ptr, bitselm, true);
2922 }
2923 
2924 void
2925 arena_dalloc_bin(tsdn_t *tsdn, arena_t *arena, arena_chunk_t *chunk, void *ptr,
2926     size_t pageind, arena_chunk_map_bits_t *bitselm)
2927 {
2928 	arena_run_t *run;
2929 	arena_bin_t *bin;
2930 	size_t rpages_ind;
2931 
2932 	rpages_ind = pageind - arena_mapbits_small_runind_get(chunk, pageind);
2933 	run = &arena_miscelm_get_mutable(chunk, rpages_ind)->run;
2934 	bin = &arena->bins[run->binind];
2935 	malloc_mutex_lock(tsdn, &bin->lock);
2936 	arena_dalloc_bin_locked_impl(tsdn, arena, chunk, ptr, bitselm, false);
2937 	malloc_mutex_unlock(tsdn, &bin->lock);
2938 }
2939 
2940 void
2941 arena_dalloc_small(tsdn_t *tsdn, arena_t *arena, arena_chunk_t *chunk,
2942     void *ptr, size_t pageind)
2943 {
2944 	arena_chunk_map_bits_t *bitselm;
2945 
2946 	if (config_debug) {
2947 		/* arena_ptr_small_binind_get() does extra sanity checking. */
2948 		assert(arena_ptr_small_binind_get(ptr, arena_mapbits_get(chunk,
2949 		    pageind)) != BININD_INVALID);
2950 	}
2951 	bitselm = arena_bitselm_get_mutable(chunk, pageind);
2952 	arena_dalloc_bin(tsdn, arena, chunk, ptr, pageind, bitselm);
2953 	arena_decay_tick(tsdn, arena);
2954 }
2955 
2956 #ifdef JEMALLOC_JET
2957 #undef arena_dalloc_junk_large
2958 #define	arena_dalloc_junk_large JEMALLOC_N(n_arena_dalloc_junk_large)
2959 #endif
2960 void
2961 arena_dalloc_junk_large(void *ptr, size_t usize)
2962 {
2963 
2964 	if (config_fill && unlikely(opt_junk_free))
2965 		memset(ptr, JEMALLOC_FREE_JUNK, usize);
2966 }
2967 #ifdef JEMALLOC_JET
2968 #undef arena_dalloc_junk_large
2969 #define	arena_dalloc_junk_large JEMALLOC_N(arena_dalloc_junk_large)
2970 arena_dalloc_junk_large_t *arena_dalloc_junk_large =
2971     JEMALLOC_N(n_arena_dalloc_junk_large);
2972 #endif
2973 
2974 static void
2975 arena_dalloc_large_locked_impl(tsdn_t *tsdn, arena_t *arena,
2976     arena_chunk_t *chunk, void *ptr, bool junked)
2977 {
2978 	size_t pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
2979 	arena_chunk_map_misc_t *miscelm = arena_miscelm_get_mutable(chunk,
2980 	    pageind);
2981 	arena_run_t *run = &miscelm->run;
2982 
2983 	if (config_fill || config_stats) {
2984 		size_t usize = arena_mapbits_large_size_get(chunk, pageind) -
2985 		    large_pad;
2986 
2987 		if (!junked)
2988 			arena_dalloc_junk_large(ptr, usize);
2989 		if (config_stats) {
2990 			szind_t index = size2index(usize) - NBINS;
2991 
2992 			arena->stats.ndalloc_large++;
2993 			arena->stats.allocated_large -= usize;
2994 			arena->stats.lstats[index].ndalloc++;
2995 			arena->stats.lstats[index].curruns--;
2996 		}
2997 	}
2998 
2999 	arena_run_dalloc(tsdn, arena, run, true, false, false);
3000 }
3001 
3002 void
3003 arena_dalloc_large_junked_locked(tsdn_t *tsdn, arena_t *arena,
3004     arena_chunk_t *chunk, void *ptr)
3005 {
3006 
3007 	arena_dalloc_large_locked_impl(tsdn, arena, chunk, ptr, true);
3008 }
3009 
3010 void
3011 arena_dalloc_large(tsdn_t *tsdn, arena_t *arena, arena_chunk_t *chunk,
3012     void *ptr)
3013 {
3014 
3015 	malloc_mutex_lock(tsdn, &arena->lock);
3016 	arena_dalloc_large_locked_impl(tsdn, arena, chunk, ptr, false);
3017 	malloc_mutex_unlock(tsdn, &arena->lock);
3018 	arena_decay_tick(tsdn, arena);
3019 }
3020 
3021 static void
3022 arena_ralloc_large_shrink(tsdn_t *tsdn, arena_t *arena, arena_chunk_t *chunk,
3023     void *ptr, size_t oldsize, size_t size)
3024 {
3025 	size_t pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
3026 	arena_chunk_map_misc_t *miscelm = arena_miscelm_get_mutable(chunk,
3027 	    pageind);
3028 	arena_run_t *run = &miscelm->run;
3029 
3030 	assert(size < oldsize);
3031 
3032 	/*
3033 	 * Shrink the run, and make trailing pages available for other
3034 	 * allocations.
3035 	 */
3036 	malloc_mutex_lock(tsdn, &arena->lock);
3037 	arena_run_trim_tail(tsdn, arena, chunk, run, oldsize + large_pad, size +
3038 	    large_pad, true);
3039 	if (config_stats) {
3040 		szind_t oldindex = size2index(oldsize) - NBINS;
3041 		szind_t index = size2index(size) - NBINS;
3042 
3043 		arena->stats.ndalloc_large++;
3044 		arena->stats.allocated_large -= oldsize;
3045 		arena->stats.lstats[oldindex].ndalloc++;
3046 		arena->stats.lstats[oldindex].curruns--;
3047 
3048 		arena->stats.nmalloc_large++;
3049 		arena->stats.nrequests_large++;
3050 		arena->stats.allocated_large += size;
3051 		arena->stats.lstats[index].nmalloc++;
3052 		arena->stats.lstats[index].nrequests++;
3053 		arena->stats.lstats[index].curruns++;
3054 	}
3055 	malloc_mutex_unlock(tsdn, &arena->lock);
3056 }
3057 
3058 static bool
3059 arena_ralloc_large_grow(tsdn_t *tsdn, arena_t *arena, arena_chunk_t *chunk,
3060     void *ptr, size_t oldsize, size_t usize_min, size_t usize_max, bool zero)
3061 {
3062 	size_t pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
3063 	size_t npages = (oldsize + large_pad) >> LG_PAGE;
3064 	size_t followsize;
3065 
3066 	assert(oldsize == arena_mapbits_large_size_get(chunk, pageind) -
3067 	    large_pad);
3068 
3069 	/* Try to extend the run. */
3070 	malloc_mutex_lock(tsdn, &arena->lock);
3071 	if (pageind+npages >= chunk_npages || arena_mapbits_allocated_get(chunk,
3072 	    pageind+npages) != 0)
3073 		goto label_fail;
3074 	followsize = arena_mapbits_unallocated_size_get(chunk, pageind+npages);
3075 	if (oldsize + followsize >= usize_min) {
3076 		/*
3077 		 * The next run is available and sufficiently large.  Split the
3078 		 * following run, then merge the first part with the existing
3079 		 * allocation.
3080 		 */
3081 		arena_run_t *run;
3082 		size_t usize, splitsize, size, flag_dirty, flag_unzeroed_mask;
3083 
3084 		usize = usize_max;
3085 		while (oldsize + followsize < usize)
3086 			usize = index2size(size2index(usize)-1);
3087 		assert(usize >= usize_min);
3088 		assert(usize >= oldsize);
3089 		splitsize = usize - oldsize;
3090 		if (splitsize == 0)
3091 			goto label_fail;
3092 
3093 		run = &arena_miscelm_get_mutable(chunk, pageind+npages)->run;
3094 		if (arena_run_split_large(arena, run, splitsize, zero))
3095 			goto label_fail;
3096 
3097 		if (config_cache_oblivious && zero) {
3098 			/*
3099 			 * Zero the trailing bytes of the original allocation's
3100 			 * last page, since they are in an indeterminate state.
3101 			 * There will always be trailing bytes, because ptr's
3102 			 * offset from the beginning of the run is a multiple of
3103 			 * CACHELINE in [0 .. PAGE).
3104 			 */
3105 			void *zbase = (void *)((uintptr_t)ptr + oldsize);
3106 			void *zpast = PAGE_ADDR2BASE((void *)((uintptr_t)zbase +
3107 			    PAGE));
3108 			size_t nzero = (uintptr_t)zpast - (uintptr_t)zbase;
3109 			assert(nzero > 0);
3110 			memset(zbase, 0, nzero);
3111 		}
3112 
3113 		size = oldsize + splitsize;
3114 		npages = (size + large_pad) >> LG_PAGE;
3115 
3116 		/*
3117 		 * Mark the extended run as dirty if either portion of the run
3118 		 * was dirty before allocation.  This is rather pedantic,
3119 		 * because there's not actually any sequence of events that
3120 		 * could cause the resulting run to be passed to
3121 		 * arena_run_dalloc() with the dirty argument set to false
3122 		 * (which is when dirty flag consistency would really matter).
3123 		 */
3124 		flag_dirty = arena_mapbits_dirty_get(chunk, pageind) |
3125 		    arena_mapbits_dirty_get(chunk, pageind+npages-1);
3126 		flag_unzeroed_mask = flag_dirty == 0 ? CHUNK_MAP_UNZEROED : 0;
3127 		arena_mapbits_large_set(chunk, pageind, size + large_pad,
3128 		    flag_dirty | (flag_unzeroed_mask &
3129 		    arena_mapbits_unzeroed_get(chunk, pageind)));
3130 		arena_mapbits_large_set(chunk, pageind+npages-1, 0, flag_dirty |
3131 		    (flag_unzeroed_mask & arena_mapbits_unzeroed_get(chunk,
3132 		    pageind+npages-1)));
3133 
3134 		if (config_stats) {
3135 			szind_t oldindex = size2index(oldsize) - NBINS;
3136 			szind_t index = size2index(size) - NBINS;
3137 
3138 			arena->stats.ndalloc_large++;
3139 			arena->stats.allocated_large -= oldsize;
3140 			arena->stats.lstats[oldindex].ndalloc++;
3141 			arena->stats.lstats[oldindex].curruns--;
3142 
3143 			arena->stats.nmalloc_large++;
3144 			arena->stats.nrequests_large++;
3145 			arena->stats.allocated_large += size;
3146 			arena->stats.lstats[index].nmalloc++;
3147 			arena->stats.lstats[index].nrequests++;
3148 			arena->stats.lstats[index].curruns++;
3149 		}
3150 		malloc_mutex_unlock(tsdn, &arena->lock);
3151 		return (false);
3152 	}
3153 label_fail:
3154 	malloc_mutex_unlock(tsdn, &arena->lock);
3155 	return (true);
3156 }
3157 
3158 #ifdef JEMALLOC_JET
3159 #undef arena_ralloc_junk_large
3160 #define	arena_ralloc_junk_large JEMALLOC_N(n_arena_ralloc_junk_large)
3161 #endif
3162 static void
3163 arena_ralloc_junk_large(void *ptr, size_t old_usize, size_t usize)
3164 {
3165 
3166 	if (config_fill && unlikely(opt_junk_free)) {
3167 		memset((void *)((uintptr_t)ptr + usize), JEMALLOC_FREE_JUNK,
3168 		    old_usize - usize);
3169 	}
3170 }
3171 #ifdef JEMALLOC_JET
3172 #undef arena_ralloc_junk_large
3173 #define	arena_ralloc_junk_large JEMALLOC_N(arena_ralloc_junk_large)
3174 arena_ralloc_junk_large_t *arena_ralloc_junk_large =
3175     JEMALLOC_N(n_arena_ralloc_junk_large);
3176 #endif
3177 
3178 /*
3179  * Try to resize a large allocation, in order to avoid copying.  This will
3180  * always fail if growing an object, and the following run is already in use.
3181  */
3182 static bool
3183 arena_ralloc_large(tsdn_t *tsdn, void *ptr, size_t oldsize, size_t usize_min,
3184     size_t usize_max, bool zero)
3185 {
3186 	arena_chunk_t *chunk;
3187 	arena_t *arena;
3188 
3189 	if (oldsize == usize_max) {
3190 		/* Current size class is compatible and maximal. */
3191 		return (false);
3192 	}
3193 
3194 	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
3195 	arena = extent_node_arena_get(&chunk->node);
3196 
3197 	if (oldsize < usize_max) {
3198 		bool ret = arena_ralloc_large_grow(tsdn, arena, chunk, ptr,
3199 		    oldsize, usize_min, usize_max, zero);
3200 		if (config_fill && !ret && !zero) {
3201 			if (unlikely(opt_junk_alloc)) {
3202 				memset((void *)((uintptr_t)ptr + oldsize),
3203 				    JEMALLOC_ALLOC_JUNK,
3204 				    isalloc(tsdn, ptr, config_prof) - oldsize);
3205 			} else if (unlikely(opt_zero)) {
3206 				memset((void *)((uintptr_t)ptr + oldsize), 0,
3207 				    isalloc(tsdn, ptr, config_prof) - oldsize);
3208 			}
3209 		}
3210 		return (ret);
3211 	}
3212 
3213 	assert(oldsize > usize_max);
3214 	/* Fill before shrinking in order avoid a race. */
3215 	arena_ralloc_junk_large(ptr, oldsize, usize_max);
3216 	arena_ralloc_large_shrink(tsdn, arena, chunk, ptr, oldsize, usize_max);
3217 	return (false);
3218 }
3219 
3220 bool
3221 arena_ralloc_no_move(tsdn_t *tsdn, void *ptr, size_t oldsize, size_t size,
3222     size_t extra, bool zero)
3223 {
3224 	size_t usize_min, usize_max;
3225 
3226 	/* Calls with non-zero extra had to clamp extra. */
3227 	assert(extra == 0 || size + extra <= HUGE_MAXCLASS);
3228 
3229 	if (unlikely(size > HUGE_MAXCLASS))
3230 		return (true);
3231 
3232 	usize_min = s2u(size);
3233 	usize_max = s2u(size + extra);
3234 	if (likely(oldsize <= large_maxclass && usize_min <= large_maxclass)) {
3235 		arena_chunk_t *chunk;
3236 
3237 		/*
3238 		 * Avoid moving the allocation if the size class can be left the
3239 		 * same.
3240 		 */
3241 		if (oldsize <= SMALL_MAXCLASS) {
3242 			assert(arena_bin_info[size2index(oldsize)].reg_size ==
3243 			    oldsize);
3244 			if ((usize_max > SMALL_MAXCLASS ||
3245 			    size2index(usize_max) != size2index(oldsize)) &&
3246 			    (size > oldsize || usize_max < oldsize))
3247 				return (true);
3248 		} else {
3249 			if (usize_max <= SMALL_MAXCLASS)
3250 				return (true);
3251 			if (arena_ralloc_large(tsdn, ptr, oldsize, usize_min,
3252 			    usize_max, zero))
3253 				return (true);
3254 		}
3255 
3256 		chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
3257 		arena_decay_tick(tsdn, extent_node_arena_get(&chunk->node));
3258 		return (false);
3259 	} else {
3260 		return (huge_ralloc_no_move(tsdn, ptr, oldsize, usize_min,
3261 		    usize_max, zero));
3262 	}
3263 }
3264 
3265 static void *
3266 arena_ralloc_move_helper(tsdn_t *tsdn, arena_t *arena, size_t usize,
3267     size_t alignment, bool zero, tcache_t *tcache)
3268 {
3269 
3270 	if (alignment == 0)
3271 		return (arena_malloc(tsdn, arena, usize, size2index(usize),
3272 		    zero, tcache, true));
3273 	usize = sa2u(usize, alignment);
3274 	if (unlikely(usize == 0 || usize > HUGE_MAXCLASS))
3275 		return (NULL);
3276 	return (ipalloct(tsdn, usize, alignment, zero, tcache, arena));
3277 }
3278 
3279 void *
3280 arena_ralloc(tsd_t *tsd, arena_t *arena, void *ptr, size_t oldsize, size_t size,
3281     size_t alignment, bool zero, tcache_t *tcache)
3282 {
3283 	void *ret;
3284 	size_t usize;
3285 
3286 	usize = s2u(size);
3287 	if (unlikely(usize == 0 || size > HUGE_MAXCLASS))
3288 		return (NULL);
3289 
3290 	if (likely(usize <= large_maxclass)) {
3291 		size_t copysize;
3292 
3293 		/* Try to avoid moving the allocation. */
3294 		if (!arena_ralloc_no_move(tsd_tsdn(tsd), ptr, oldsize, usize, 0,
3295 		    zero))
3296 			return (ptr);
3297 
3298 		/*
3299 		 * size and oldsize are different enough that we need to move
3300 		 * the object.  In that case, fall back to allocating new space
3301 		 * and copying.
3302 		 */
3303 		ret = arena_ralloc_move_helper(tsd_tsdn(tsd), arena, usize,
3304 		    alignment, zero, tcache);
3305 		if (ret == NULL)
3306 			return (NULL);
3307 
3308 		/*
3309 		 * Junk/zero-filling were already done by
3310 		 * ipalloc()/arena_malloc().
3311 		 */
3312 
3313 		copysize = (usize < oldsize) ? usize : oldsize;
3314 		JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED(ret, copysize);
3315 		memcpy(ret, ptr, copysize);
3316 		isqalloc(tsd, ptr, oldsize, tcache, true);
3317 	} else {
3318 		ret = huge_ralloc(tsd, arena, ptr, oldsize, usize, alignment,
3319 		    zero, tcache);
3320 	}
3321 	return (ret);
3322 }
3323 
3324 dss_prec_t
3325 arena_dss_prec_get(tsdn_t *tsdn, arena_t *arena)
3326 {
3327 	dss_prec_t ret;
3328 
3329 	malloc_mutex_lock(tsdn, &arena->lock);
3330 	ret = arena->dss_prec;
3331 	malloc_mutex_unlock(tsdn, &arena->lock);
3332 	return (ret);
3333 }
3334 
3335 bool
3336 arena_dss_prec_set(tsdn_t *tsdn, arena_t *arena, dss_prec_t dss_prec)
3337 {
3338 
3339 	if (!have_dss)
3340 		return (dss_prec != dss_prec_disabled);
3341 	malloc_mutex_lock(tsdn, &arena->lock);
3342 	arena->dss_prec = dss_prec;
3343 	malloc_mutex_unlock(tsdn, &arena->lock);
3344 	return (false);
3345 }
3346 
3347 ssize_t
3348 arena_lg_dirty_mult_default_get(void)
3349 {
3350 
3351 	return ((ssize_t)atomic_read_z((size_t *)&lg_dirty_mult_default));
3352 }
3353 
3354 bool
3355 arena_lg_dirty_mult_default_set(ssize_t lg_dirty_mult)
3356 {
3357 
3358 	if (opt_purge != purge_mode_ratio)
3359 		return (true);
3360 	if (!arena_lg_dirty_mult_valid(lg_dirty_mult))
3361 		return (true);
3362 	atomic_write_z((size_t *)&lg_dirty_mult_default, (size_t)lg_dirty_mult);
3363 	return (false);
3364 }
3365 
3366 ssize_t
3367 arena_decay_time_default_get(void)
3368 {
3369 
3370 	return ((ssize_t)atomic_read_z((size_t *)&decay_time_default));
3371 }
3372 
3373 bool
3374 arena_decay_time_default_set(ssize_t decay_time)
3375 {
3376 
3377 	if (opt_purge != purge_mode_decay)
3378 		return (true);
3379 	if (!arena_decay_time_valid(decay_time))
3380 		return (true);
3381 	atomic_write_z((size_t *)&decay_time_default, (size_t)decay_time);
3382 	return (false);
3383 }
3384 
3385 static void
3386 arena_basic_stats_merge_locked(arena_t *arena, unsigned *nthreads,
3387     const char **dss, ssize_t *lg_dirty_mult, ssize_t *decay_time,
3388     size_t *nactive, size_t *ndirty)
3389 {
3390 
3391 	*nthreads += arena_nthreads_get(arena, false);
3392 	*dss = dss_prec_names[arena->dss_prec];
3393 	*lg_dirty_mult = arena->lg_dirty_mult;
3394 	*decay_time = arena->decay_time;
3395 	*nactive += arena->nactive;
3396 	*ndirty += arena->ndirty;
3397 }
3398 
3399 void
3400 arena_basic_stats_merge(tsdn_t *tsdn, arena_t *arena, unsigned *nthreads,
3401     const char **dss, ssize_t *lg_dirty_mult, ssize_t *decay_time,
3402     size_t *nactive, size_t *ndirty)
3403 {
3404 
3405 	malloc_mutex_lock(tsdn, &arena->lock);
3406 	arena_basic_stats_merge_locked(arena, nthreads, dss, lg_dirty_mult,
3407 	    decay_time, nactive, ndirty);
3408 	malloc_mutex_unlock(tsdn, &arena->lock);
3409 }
3410 
3411 void
3412 arena_stats_merge(tsdn_t *tsdn, arena_t *arena, unsigned *nthreads,
3413     const char **dss, ssize_t *lg_dirty_mult, ssize_t *decay_time,
3414     size_t *nactive, size_t *ndirty, arena_stats_t *astats,
3415     malloc_bin_stats_t *bstats, malloc_large_stats_t *lstats,
3416     malloc_huge_stats_t *hstats)
3417 {
3418 	unsigned i;
3419 
3420 	cassert(config_stats);
3421 
3422 	malloc_mutex_lock(tsdn, &arena->lock);
3423 	arena_basic_stats_merge_locked(arena, nthreads, dss, lg_dirty_mult,
3424 	    decay_time, nactive, ndirty);
3425 
3426 	astats->mapped += arena->stats.mapped;
3427 	astats->retained += arena->stats.retained;
3428 	astats->npurge += arena->stats.npurge;
3429 	astats->nmadvise += arena->stats.nmadvise;
3430 	astats->purged += arena->stats.purged;
3431 	astats->metadata_mapped += arena->stats.metadata_mapped;
3432 	astats->metadata_allocated += arena_metadata_allocated_get(arena);
3433 	astats->allocated_large += arena->stats.allocated_large;
3434 	astats->nmalloc_large += arena->stats.nmalloc_large;
3435 	astats->ndalloc_large += arena->stats.ndalloc_large;
3436 	astats->nrequests_large += arena->stats.nrequests_large;
3437 	astats->allocated_huge += arena->stats.allocated_huge;
3438 	astats->nmalloc_huge += arena->stats.nmalloc_huge;
3439 	astats->ndalloc_huge += arena->stats.ndalloc_huge;
3440 
3441 	for (i = 0; i < nlclasses; i++) {
3442 		lstats[i].nmalloc += arena->stats.lstats[i].nmalloc;
3443 		lstats[i].ndalloc += arena->stats.lstats[i].ndalloc;
3444 		lstats[i].nrequests += arena->stats.lstats[i].nrequests;
3445 		lstats[i].curruns += arena->stats.lstats[i].curruns;
3446 	}
3447 
3448 	for (i = 0; i < nhclasses; i++) {
3449 		hstats[i].nmalloc += arena->stats.hstats[i].nmalloc;
3450 		hstats[i].ndalloc += arena->stats.hstats[i].ndalloc;
3451 		hstats[i].curhchunks += arena->stats.hstats[i].curhchunks;
3452 	}
3453 	malloc_mutex_unlock(tsdn, &arena->lock);
3454 
3455 	for (i = 0; i < NBINS; i++) {
3456 		arena_bin_t *bin = &arena->bins[i];
3457 
3458 		malloc_mutex_lock(tsdn, &bin->lock);
3459 		bstats[i].nmalloc += bin->stats.nmalloc;
3460 		bstats[i].ndalloc += bin->stats.ndalloc;
3461 		bstats[i].nrequests += bin->stats.nrequests;
3462 		bstats[i].curregs += bin->stats.curregs;
3463 		if (config_tcache) {
3464 			bstats[i].nfills += bin->stats.nfills;
3465 			bstats[i].nflushes += bin->stats.nflushes;
3466 		}
3467 		bstats[i].nruns += bin->stats.nruns;
3468 		bstats[i].reruns += bin->stats.reruns;
3469 		bstats[i].curruns += bin->stats.curruns;
3470 		malloc_mutex_unlock(tsdn, &bin->lock);
3471 	}
3472 }
3473 
3474 unsigned
3475 arena_nthreads_get(arena_t *arena, bool internal)
3476 {
3477 
3478 	return (atomic_read_u(&arena->nthreads[internal]));
3479 }
3480 
3481 void
3482 arena_nthreads_inc(arena_t *arena, bool internal)
3483 {
3484 
3485 	atomic_add_u(&arena->nthreads[internal], 1);
3486 }
3487 
3488 void
3489 arena_nthreads_dec(arena_t *arena, bool internal)
3490 {
3491 
3492 	atomic_sub_u(&arena->nthreads[internal], 1);
3493 }
3494 
3495 arena_t *
3496 arena_new(tsdn_t *tsdn, unsigned ind)
3497 {
3498 	arena_t *arena;
3499 	size_t arena_size;
3500 	unsigned i;
3501 
3502 	/* Compute arena size to incorporate sufficient runs_avail elements. */
3503 	arena_size = offsetof(arena_t, runs_avail) + (sizeof(arena_run_heap_t) *
3504 	    runs_avail_nclasses);
3505 	/*
3506 	 * Allocate arena, arena->lstats, and arena->hstats contiguously, mainly
3507 	 * because there is no way to clean up if base_alloc() OOMs.
3508 	 */
3509 	if (config_stats) {
3510 		arena = (arena_t *)base_alloc(tsdn,
3511 		    CACHELINE_CEILING(arena_size) + QUANTUM_CEILING(nlclasses *
3512 		    sizeof(malloc_large_stats_t) + nhclasses) *
3513 		    sizeof(malloc_huge_stats_t));
3514 	} else
3515 		arena = (arena_t *)base_alloc(tsdn, arena_size);
3516 	if (arena == NULL)
3517 		return (NULL);
3518 
3519 	arena->ind = ind;
3520 	arena->nthreads[0] = arena->nthreads[1] = 0;
3521 	if (malloc_mutex_init(&arena->lock, "arena", WITNESS_RANK_ARENA))
3522 		return (NULL);
3523 
3524 	if (config_stats) {
3525 		memset(&arena->stats, 0, sizeof(arena_stats_t));
3526 		arena->stats.lstats = (malloc_large_stats_t *)((uintptr_t)arena
3527 		    + CACHELINE_CEILING(arena_size));
3528 		memset(arena->stats.lstats, 0, nlclasses *
3529 		    sizeof(malloc_large_stats_t));
3530 		arena->stats.hstats = (malloc_huge_stats_t *)((uintptr_t)arena
3531 		    + CACHELINE_CEILING(arena_size) +
3532 		    QUANTUM_CEILING(nlclasses * sizeof(malloc_large_stats_t)));
3533 		memset(arena->stats.hstats, 0, nhclasses *
3534 		    sizeof(malloc_huge_stats_t));
3535 		if (config_tcache)
3536 			ql_new(&arena->tcache_ql);
3537 	}
3538 
3539 	if (config_prof)
3540 		arena->prof_accumbytes = 0;
3541 
3542 	if (config_cache_oblivious) {
3543 		/*
3544 		 * A nondeterministic seed based on the address of arena reduces
3545 		 * the likelihood of lockstep non-uniform cache index
3546 		 * utilization among identical concurrent processes, but at the
3547 		 * cost of test repeatability.  For debug builds, instead use a
3548 		 * deterministic seed.
3549 		 */
3550 		arena->offset_state = config_debug ? ind :
3551 		    (uint64_t)(uintptr_t)arena;
3552 	}
3553 
3554 	arena->dss_prec = chunk_dss_prec_get(tsdn);
3555 
3556 	ql_new(&arena->achunks);
3557 
3558 	arena->spare = NULL;
3559 
3560 	arena->lg_dirty_mult = arena_lg_dirty_mult_default_get();
3561 	arena->purging = false;
3562 	arena->nactive = 0;
3563 	arena->ndirty = 0;
3564 
3565 	for(i = 0; i < runs_avail_nclasses; i++)
3566 		arena_run_heap_new(&arena->runs_avail[i]);
3567 	qr_new(&arena->runs_dirty, rd_link);
3568 	qr_new(&arena->chunks_cache, cc_link);
3569 
3570 	if (opt_purge == purge_mode_decay)
3571 		arena_decay_init(arena, arena_decay_time_default_get());
3572 
3573 	ql_new(&arena->huge);
3574 	if (malloc_mutex_init(&arena->huge_mtx, "arena_huge",
3575 	    WITNESS_RANK_ARENA_HUGE))
3576 		return (NULL);
3577 
3578 	extent_tree_szad_new(&arena->chunks_szad_cached);
3579 	extent_tree_ad_new(&arena->chunks_ad_cached);
3580 	extent_tree_szad_new(&arena->chunks_szad_retained);
3581 	extent_tree_ad_new(&arena->chunks_ad_retained);
3582 	if (malloc_mutex_init(&arena->chunks_mtx, "arena_chunks",
3583 	    WITNESS_RANK_ARENA_CHUNKS))
3584 		return (NULL);
3585 	ql_new(&arena->node_cache);
3586 	if (malloc_mutex_init(&arena->node_cache_mtx, "arena_node_cache",
3587 	    WITNESS_RANK_ARENA_NODE_CACHE))
3588 		return (NULL);
3589 
3590 	arena->chunk_hooks = chunk_hooks_default;
3591 
3592 	/* Initialize bins. */
3593 	for (i = 0; i < NBINS; i++) {
3594 		arena_bin_t *bin = &arena->bins[i];
3595 		if (malloc_mutex_init(&bin->lock, "arena_bin",
3596 		    WITNESS_RANK_ARENA_BIN))
3597 			return (NULL);
3598 		bin->runcur = NULL;
3599 		arena_run_heap_new(&bin->runs);
3600 		if (config_stats)
3601 			memset(&bin->stats, 0, sizeof(malloc_bin_stats_t));
3602 	}
3603 
3604 	return (arena);
3605 }
3606 
3607 /*
3608  * Calculate bin_info->run_size such that it meets the following constraints:
3609  *
3610  *   *) bin_info->run_size <= arena_maxrun
3611  *   *) bin_info->nregs <= RUN_MAXREGS
3612  *
3613  * bin_info->nregs and bin_info->reg0_offset are also calculated here, since
3614  * these settings are all interdependent.
3615  */
3616 static void
3617 bin_info_run_size_calc(arena_bin_info_t *bin_info)
3618 {
3619 	size_t pad_size;
3620 	size_t try_run_size, perfect_run_size, actual_run_size;
3621 	uint32_t try_nregs, perfect_nregs, actual_nregs;
3622 
3623 	/*
3624 	 * Determine redzone size based on minimum alignment and minimum
3625 	 * redzone size.  Add padding to the end of the run if it is needed to
3626 	 * align the regions.  The padding allows each redzone to be half the
3627 	 * minimum alignment; without the padding, each redzone would have to
3628 	 * be twice as large in order to maintain alignment.
3629 	 */
3630 	if (config_fill && unlikely(opt_redzone)) {
3631 		size_t align_min = ZU(1) << (ffs_zu(bin_info->reg_size) - 1);
3632 		if (align_min <= REDZONE_MINSIZE) {
3633 			bin_info->redzone_size = REDZONE_MINSIZE;
3634 			pad_size = 0;
3635 		} else {
3636 			bin_info->redzone_size = align_min >> 1;
3637 			pad_size = bin_info->redzone_size;
3638 		}
3639 	} else {
3640 		bin_info->redzone_size = 0;
3641 		pad_size = 0;
3642 	}
3643 	bin_info->reg_interval = bin_info->reg_size +
3644 	    (bin_info->redzone_size << 1);
3645 
3646 	/*
3647 	 * Compute run size under ideal conditions (no redzones, no limit on run
3648 	 * size).
3649 	 */
3650 	try_run_size = PAGE;
3651 	try_nregs = (uint32_t)(try_run_size / bin_info->reg_size);
3652 	do {
3653 		perfect_run_size = try_run_size;
3654 		perfect_nregs = try_nregs;
3655 
3656 		try_run_size += PAGE;
3657 		try_nregs = (uint32_t)(try_run_size / bin_info->reg_size);
3658 	} while (perfect_run_size != perfect_nregs * bin_info->reg_size);
3659 	assert(perfect_nregs <= RUN_MAXREGS);
3660 
3661 	actual_run_size = perfect_run_size;
3662 	actual_nregs = (uint32_t)((actual_run_size - pad_size) /
3663 	    bin_info->reg_interval);
3664 
3665 	/*
3666 	 * Redzones can require enough padding that not even a single region can
3667 	 * fit within the number of pages that would normally be dedicated to a
3668 	 * run for this size class.  Increase the run size until at least one
3669 	 * region fits.
3670 	 */
3671 	while (actual_nregs == 0) {
3672 		assert(config_fill && unlikely(opt_redzone));
3673 
3674 		actual_run_size += PAGE;
3675 		actual_nregs = (uint32_t)((actual_run_size - pad_size) /
3676 		    bin_info->reg_interval);
3677 	}
3678 
3679 	/*
3680 	 * Make sure that the run will fit within an arena chunk.
3681 	 */
3682 	while (actual_run_size > arena_maxrun) {
3683 		actual_run_size -= PAGE;
3684 		actual_nregs = (uint32_t)((actual_run_size - pad_size) /
3685 		    bin_info->reg_interval);
3686 	}
3687 	assert(actual_nregs > 0);
3688 	assert(actual_run_size == s2u(actual_run_size));
3689 
3690 	/* Copy final settings. */
3691 	bin_info->run_size = actual_run_size;
3692 	bin_info->nregs = actual_nregs;
3693 	bin_info->reg0_offset = (uint32_t)(actual_run_size - (actual_nregs *
3694 	    bin_info->reg_interval) - pad_size + bin_info->redzone_size);
3695 
3696 	if (actual_run_size > small_maxrun)
3697 		small_maxrun = actual_run_size;
3698 
3699 	assert(bin_info->reg0_offset - bin_info->redzone_size + (bin_info->nregs
3700 	    * bin_info->reg_interval) + pad_size == bin_info->run_size);
3701 }
3702 
3703 static void
3704 bin_info_init(void)
3705 {
3706 	arena_bin_info_t *bin_info;
3707 
3708 #define	BIN_INFO_INIT_bin_yes(index, size)				\
3709 	bin_info = &arena_bin_info[index];				\
3710 	bin_info->reg_size = size;					\
3711 	bin_info_run_size_calc(bin_info);				\
3712 	bitmap_info_init(&bin_info->bitmap_info, bin_info->nregs);
3713 #define	BIN_INFO_INIT_bin_no(index, size)
3714 #define	SC(index, lg_grp, lg_delta, ndelta, bin, lg_delta_lookup)	\
3715 	BIN_INFO_INIT_bin_##bin(index, (ZU(1)<<lg_grp) + (ZU(ndelta)<<lg_delta))
3716 	SIZE_CLASSES
3717 #undef BIN_INFO_INIT_bin_yes
3718 #undef BIN_INFO_INIT_bin_no
3719 #undef SC
3720 }
3721 
3722 static bool
3723 small_run_size_init(void)
3724 {
3725 
3726 	assert(small_maxrun != 0);
3727 
3728 	small_run_tab = (bool *)base_alloc(NULL, sizeof(bool) * (small_maxrun >>
3729 	    LG_PAGE));
3730 	if (small_run_tab == NULL)
3731 		return (true);
3732 
3733 #define	TAB_INIT_bin_yes(index, size) {					\
3734 		arena_bin_info_t *bin_info = &arena_bin_info[index];	\
3735 		small_run_tab[bin_info->run_size >> LG_PAGE] = true;	\
3736 	}
3737 #define	TAB_INIT_bin_no(index, size)
3738 #define	SC(index, lg_grp, lg_delta, ndelta, bin, lg_delta_lookup)	\
3739 	TAB_INIT_bin_##bin(index, (ZU(1)<<lg_grp) + (ZU(ndelta)<<lg_delta))
3740 	SIZE_CLASSES
3741 #undef TAB_INIT_bin_yes
3742 #undef TAB_INIT_bin_no
3743 #undef SC
3744 
3745 	return (false);
3746 }
3747 
3748 static bool
3749 run_quantize_init(void)
3750 {
3751 	unsigned i;
3752 
3753 	run_quantize_max = chunksize + large_pad;
3754 
3755 	run_quantize_floor_tab = (size_t *)base_alloc(NULL, sizeof(size_t) *
3756 	    (run_quantize_max >> LG_PAGE));
3757 	if (run_quantize_floor_tab == NULL)
3758 		return (true);
3759 
3760 	run_quantize_ceil_tab = (size_t *)base_alloc(NULL, sizeof(size_t) *
3761 	    (run_quantize_max >> LG_PAGE));
3762 	if (run_quantize_ceil_tab == NULL)
3763 		return (true);
3764 
3765 	for (i = 1; i <= run_quantize_max >> LG_PAGE; i++) {
3766 		size_t run_size = i << LG_PAGE;
3767 
3768 		run_quantize_floor_tab[i-1] =
3769 		    run_quantize_floor_compute(run_size);
3770 		run_quantize_ceil_tab[i-1] =
3771 		    run_quantize_ceil_compute(run_size);
3772 	}
3773 
3774 	return (false);
3775 }
3776 
3777 bool
3778 arena_boot(void)
3779 {
3780 	unsigned i;
3781 
3782 	arena_lg_dirty_mult_default_set(opt_lg_dirty_mult);
3783 	arena_decay_time_default_set(opt_decay_time);
3784 
3785 	/*
3786 	 * Compute the header size such that it is large enough to contain the
3787 	 * page map.  The page map is biased to omit entries for the header
3788 	 * itself, so some iteration is necessary to compute the map bias.
3789 	 *
3790 	 * 1) Compute safe header_size and map_bias values that include enough
3791 	 *    space for an unbiased page map.
3792 	 * 2) Refine map_bias based on (1) to omit the header pages in the page
3793 	 *    map.  The resulting map_bias may be one too small.
3794 	 * 3) Refine map_bias based on (2).  The result will be >= the result
3795 	 *    from (2), and will always be correct.
3796 	 */
3797 	map_bias = 0;
3798 	for (i = 0; i < 3; i++) {
3799 		size_t header_size = offsetof(arena_chunk_t, map_bits) +
3800 		    ((sizeof(arena_chunk_map_bits_t) +
3801 		    sizeof(arena_chunk_map_misc_t)) * (chunk_npages-map_bias));
3802 		map_bias = (header_size + PAGE_MASK) >> LG_PAGE;
3803 	}
3804 	assert(map_bias > 0);
3805 
3806 	map_misc_offset = offsetof(arena_chunk_t, map_bits) +
3807 	    sizeof(arena_chunk_map_bits_t) * (chunk_npages-map_bias);
3808 
3809 	arena_maxrun = chunksize - (map_bias << LG_PAGE);
3810 	assert(arena_maxrun > 0);
3811 	large_maxclass = index2size(size2index(chunksize)-1);
3812 	if (large_maxclass > arena_maxrun) {
3813 		/*
3814 		 * For small chunk sizes it's possible for there to be fewer
3815 		 * non-header pages available than are necessary to serve the
3816 		 * size classes just below chunksize.
3817 		 */
3818 		large_maxclass = arena_maxrun;
3819 	}
3820 	assert(large_maxclass > 0);
3821 	nlclasses = size2index(large_maxclass) - size2index(SMALL_MAXCLASS);
3822 	nhclasses = NSIZES - nlclasses - NBINS;
3823 
3824 	bin_info_init();
3825 	if (small_run_size_init())
3826 		return (true);
3827 	if (run_quantize_init())
3828 		return (true);
3829 
3830 	runs_avail_bias = size2index(PAGE);
3831 	runs_avail_nclasses = size2index(run_quantize_max)+1 - runs_avail_bias;
3832 
3833 	return (false);
3834 }
3835 
3836 void
3837 arena_prefork0(tsdn_t *tsdn, arena_t *arena)
3838 {
3839 
3840 	malloc_mutex_prefork(tsdn, &arena->lock);
3841 }
3842 
3843 void
3844 arena_prefork1(tsdn_t *tsdn, arena_t *arena)
3845 {
3846 
3847 	malloc_mutex_prefork(tsdn, &arena->chunks_mtx);
3848 }
3849 
3850 void
3851 arena_prefork2(tsdn_t *tsdn, arena_t *arena)
3852 {
3853 
3854 	malloc_mutex_prefork(tsdn, &arena->node_cache_mtx);
3855 }
3856 
3857 void
3858 arena_prefork3(tsdn_t *tsdn, arena_t *arena)
3859 {
3860 	unsigned i;
3861 
3862 	for (i = 0; i < NBINS; i++)
3863 		malloc_mutex_prefork(tsdn, &arena->bins[i].lock);
3864 	malloc_mutex_prefork(tsdn, &arena->huge_mtx);
3865 }
3866 
3867 void
3868 arena_postfork_parent(tsdn_t *tsdn, arena_t *arena)
3869 {
3870 	unsigned i;
3871 
3872 	malloc_mutex_postfork_parent(tsdn, &arena->huge_mtx);
3873 	for (i = 0; i < NBINS; i++)
3874 		malloc_mutex_postfork_parent(tsdn, &arena->bins[i].lock);
3875 	malloc_mutex_postfork_parent(tsdn, &arena->node_cache_mtx);
3876 	malloc_mutex_postfork_parent(tsdn, &arena->chunks_mtx);
3877 	malloc_mutex_postfork_parent(tsdn, &arena->lock);
3878 }
3879 
3880 void
3881 arena_postfork_child(tsdn_t *tsdn, arena_t *arena)
3882 {
3883 	unsigned i;
3884 
3885 	malloc_mutex_postfork_child(tsdn, &arena->huge_mtx);
3886 	for (i = 0; i < NBINS; i++)
3887 		malloc_mutex_postfork_child(tsdn, &arena->bins[i].lock);
3888 	malloc_mutex_postfork_child(tsdn, &arena->node_cache_mtx);
3889 	malloc_mutex_postfork_child(tsdn, &arena->chunks_mtx);
3890 	malloc_mutex_postfork_child(tsdn, &arena->lock);
3891 }
3892