2 * DMALLOC.C - Dillon's malloc
4 * Copyright (c) 2011,2017 The DragonFly Project. All rights reserved.
6 * This code is derived from software contributed to The DragonFly Project
7 * by Matthew Dillon <dillon@backplane.com>.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37 * This module implements a modified slab allocator as a drop-in replacement
38 * for the libc malloc(). The slab algorithm has been adjusted to support
39 * dynamic sizing of slabs which effectively allows slabs to be used for
40 * allocations of any size. Because of this we neither have a small-block
41 * allocator or a big-block allocator and the code paths are simplified.
43 * To support dynamic slab sizing available user virtual memory is broken
44 * down into ~1024 regions. Each region has fixed slab size whos value is
45 * set when the region is opened up for use. The free() path simply applies
46 * a mask based on the region to the pointer to acquire the base of the
47 * governing slab structure.
49 * Regions[NREGIONS] (1024)
51 * Slab management and locking is done on a per-zone basis.
53 * Alloc Size Chunking Number of zones
64 * ... continues forever ... 4 zones
66 * For a 2^63 memory space each doubling >= 64K is broken down into
67 * 4 chunking zones, so we support 88 + (48 * 4) = 280 zones.
69 * API FEATURES AND SIDE EFFECTS
71 * + power-of-2 sized allocations up to a page will be power-of-2 aligned.
72 * Above that power-of-2 sized allocations are page-aligned. Non
73 * power-of-2 sized allocations are aligned the same as the chunk
74 * size for their zone.
75 * + ability to allocate arbitrarily large chunks of memory
76 * + realloc will reuse the passed pointer if possible, within the
77 * limitations of the zone chunking.
79 * On top of the slab allocator we also implement a 16-entry-per-thread
80 * magazine cache for allocations <= NOMSLABSIZE.
84 * + [better] garbage collection
85 * + better initial sizing.
89 * The value of the environment variable MALLOC_OPTIONS is a character string
90 * containing various flags to tune nmalloc. Upper case letters enabled
91 * or increase the feature, lower case disables or decreases the feature.
93 * U Enable UTRACE for all operations, observable with ktrace.
94 * Diasbled by default.
96 * Z Zero out allocations, otherwise allocations (except for
97 * calloc) will contain garbage.
98 * Disabled by default.
100 * H Pass a hint with madvise() about unused pages.
101 * Disabled by default.
102 * Not currently implemented.
104 * F Disable local per-thread caching.
105 * Disabled by default.
107 * C Increase (decrease) how much excess cache to retain.
108 * Set to 4 by default.
111 /* cc -shared -fPIC -g -O -I/usr/src/lib/libc/include -o dmalloc.so dmalloc.c */
113 #ifndef STANDALONE_DEBUG
114 #include "libc_private.h"
117 #include <sys/param.h>
118 #include <sys/types.h>
119 #include <sys/mman.h>
120 #include <sys/queue.h>
122 #include <sys/ktrace.h>
135 #include <machine/atomic.h>
136 #include <machine/cpufunc.h>
138 #ifdef STANDALONE_DEBUG
139 void _nmalloc_thr_init(void);
141 #include "spinlock.h"
142 #include "un-namespace.h"
145 #ifndef MAP_SIZEALIGN
146 #define MAP_SIZEALIGN 0
149 #if SSIZE_MAX == 0x7FFFFFFF
151 #define UVM_BITS 32 /* worst case */
154 #define UVM_BITS 48 /* worst case XXX */
157 #if LONG_MAX == 0x7FFFFFFF
159 #define LONG_BITS_SHIFT 5
162 #define LONG_BITS_SHIFT 6
165 #define LOCKEDPTR ((void *)(intptr_t)-1)
170 #define NREGIONS_BITS 10
171 #define NREGIONS (1 << NREGIONS_BITS)
172 #define NREGIONS_MASK (NREGIONS - 1)
173 #define NREGIONS_SHIFT (UVM_BITS - NREGIONS_BITS)
174 #define NREGIONS_SIZE (1LU << NREGIONS_SHIFT)
176 typedef struct region *region_t;
177 typedef struct slglobaldata *slglobaldata_t;
178 typedef struct slab *slab_t;
182 slab_t slab; /* conditional out of band slab */
185 static struct region Regions[NREGIONS];
188 * Number of chunking zones available
190 #define CHUNKFACTOR 8
192 #define NZONES (16 + 9 * CHUNKFACTOR + 16 * CHUNKFACTOR)
194 #define NZONES (16 + 9 * CHUNKFACTOR + 48 * CHUNKFACTOR)
197 static int MaxChunks[NZONES];
199 #define NDEPOTS 8 /* must be power of 2 */
202 * Maximum number of chunks per slab, governed by the allocation bitmap in
203 * each slab. The maximum is reduced for large chunk sizes.
205 #define MAXCHUNKS (LONG_BITS * LONG_BITS)
206 #define MAXCHUNKS_BITS (LONG_BITS_SHIFT * LONG_BITS_SHIFT)
207 #define LITSLABSIZE (32 * 1024)
208 #define NOMSLABSIZE (2 * 1024 * 1024)
209 #define BIGSLABSIZE (128 * 1024 * 1024)
211 #define ZALLOC_SLAB_MAGIC 0x736c6162 /* magic sanity */
213 TAILQ_HEAD(slab_list, slab);
219 struct slab *next; /* slabs with available space */
220 TAILQ_ENTRY(slab) entry;
221 int32_t magic; /* magic number for sanity check */
222 u_int navail; /* number of free elements available */
224 u_int free_bit; /* free hint bitno */
225 u_int free_index; /* free hint index */
226 u_long bitmap[LONG_BITS]; /* free chunks */
227 size_t slab_size; /* size of entire slab */
228 size_t chunk_size; /* chunk size for validation */
230 enum { UNKNOWN, AVAIL, EMPTY, FULL } state;
232 region_t region; /* related region */
233 char *chunks; /* chunk base */
234 slglobaldata_t slgd; /* localized to thread else NULL */
238 * per-thread data + global depot
240 * NOTE: The magazine shortcut is only used for per-thread data.
242 #define NMAGSHORTCUT 16
244 struct slglobaldata {
245 spinlock_t lock; /* only used by slglobaldepot */
253 void *mag_shortcut[NMAGSHORTCUT];
255 struct slab_list full_zones; /* via entry */
261 #define SLAB_ZEROD 0x0001
264 * Misc constants. Note that allocations that are exact multiples of
265 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
266 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
268 #define MIN_CHUNK_SIZE 8 /* in bytes */
269 #define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
271 #define SAFLAG_ZERO 0x00000001
274 * The WEIRD_ADDR is used as known text to copy into free objects to
275 * try to create deterministic failure cases if the data is accessed after
278 * WARNING: A limited number of spinlocks are available, BIGXSIZE should
279 * not be larger then 64.
282 #define WEIRD_ADDR 0xdeadc0de
289 #define MASSERT(exp) do { if (__predict_false(!(exp))) \
290 _mpanic("assertion: %s in %s", \
295 * With this attribute set, do not require a function call for accessing
296 * this variable when the code is compiled -fPIC.
298 * Must be empty for libc_rtld (similar to __thread)
300 #if defined(__LIBC_RTLD)
301 #define TLS_ATTRIBUTE
303 #define TLS_ATTRIBUTE __attribute__ ((tls_model ("initial-exec")));
306 static __thread struct slglobaldata slglobal TLS_ATTRIBUTE;
307 static pthread_key_t thread_malloc_key;
308 static pthread_once_t thread_malloc_once = PTHREAD_ONCE_INIT;
309 static struct slglobaldata slglobaldepot;
311 static int opt_madvise = 0;
312 static int opt_free = 0;
313 static int opt_cache = 4;
314 static int opt_utrace = 0;
315 static int g_malloc_flags = 0;
316 static int malloc_panic;
319 static const int32_t weirdary[16] = {
320 WEIRD_ADDR, WEIRD_ADDR, WEIRD_ADDR, WEIRD_ADDR,
321 WEIRD_ADDR, WEIRD_ADDR, WEIRD_ADDR, WEIRD_ADDR,
322 WEIRD_ADDR, WEIRD_ADDR, WEIRD_ADDR, WEIRD_ADDR,
323 WEIRD_ADDR, WEIRD_ADDR, WEIRD_ADDR, WEIRD_ADDR
327 static void *memalloc(size_t size, int flags);
328 static void *memrealloc(void *ptr, size_t size);
329 static void memfree(void *ptr, int);
330 static slab_t slaballoc(int zi, size_t chunking, size_t chunk_size);
331 static void slabfree(slab_t slab);
332 static void slabterm(slglobaldata_t slgd, slab_t slab);
333 static void *_vmem_alloc(int ri, size_t slab_size);
334 static void _vmem_free(void *ptr, size_t slab_size);
335 static void _mpanic(const char *ctl, ...) __printflike(1, 2);
336 #ifndef STANDALONE_DEBUG
337 static void malloc_init(void) __constructor(101);
339 static void malloc_init(void) __constructor(101);
343 struct nmalloc_utrace {
349 #define UTRACE(a, b, c) \
351 struct nmalloc_utrace ut = { \
356 utrace(&ut, sizeof(ut)); \
361 * If enabled any memory allocated without M_ZERO is initialized to -1.
363 static int use_malloc_pattern;
369 const char *p = NULL;
371 TAILQ_INIT(&slglobal.full_zones);
373 Regions[0].mask = -1; /* disallow activity in lowest region */
375 if (issetugid() == 0)
376 p = getenv("MALLOC_OPTIONS");
378 for (; p != NULL && *p != '\0'; p++) {
409 g_malloc_flags = SAFLAG_ZERO;
416 UTRACE((void *) -1, 0, NULL);
420 * We have to install a handler for nmalloc thread teardowns when
421 * the thread is created. We cannot delay this because destructors in
422 * sophisticated userland programs can call malloc() for the first time
423 * during their thread exit.
425 * This routine is called directly from pthreads.
427 static void _nmalloc_thr_init_once(void);
428 static void _nmalloc_thr_destructor(void *thrp);
431 _nmalloc_thr_init(void)
435 TAILQ_INIT(&slglobal.full_zones);
443 pthread_once(&thread_malloc_once, _nmalloc_thr_init_once);
445 pthread_setspecific(thread_malloc_key, &slglobal);
450 _nmalloc_thr_prepfork(void)
453 _SPINLOCK(&slglobaldepot.lock);
457 _nmalloc_thr_parentfork(void)
460 _SPINUNLOCK(&slglobaldepot.lock);
464 _nmalloc_thr_childfork(void)
467 _SPINUNLOCK(&slglobaldepot.lock);
474 _nmalloc_thr_init_once(void)
478 error = pthread_key_create(&thread_malloc_key, _nmalloc_thr_destructor);
484 * Called for each thread undergoing exit
486 * Move all of the thread's slabs into a depot.
489 _nmalloc_thr_destructor(void *thrp)
491 slglobaldata_t slgd = thrp;
492 struct zoneinfo *zinfo;
500 for (i = 0; i <= slgd->biggest_index; i++) {
501 zinfo = &slgd->zone[i];
503 while ((j = zinfo->mag_index) > 0) {
505 ptr = zinfo->mag_shortcut[j];
506 zinfo->mag_shortcut[j] = NULL; /* SAFETY */
507 zinfo->mag_index = j;
511 while ((slab = zinfo->empty_base) != NULL) {
512 zinfo->empty_base = slab->next;
513 --zinfo->empty_count;
514 slabterm(slgd, slab);
517 while ((slab = zinfo->avail_base) != NULL) {
518 zinfo->avail_base = slab->next;
519 --zinfo->avail_count;
520 slabterm(slgd, slab);
523 while ((slab = TAILQ_FIRST(&slgd->full_zones)) != NULL) {
524 TAILQ_REMOVE(&slgd->full_zones, slab, entry);
525 slabterm(slgd, slab);
531 * Calculate the zone index for the allocation request size and set the
532 * allocation request size to that particular zone's chunk size.
534 * Minimum alignment is 16 bytes for allocations >= 16 bytes to conform
535 * with malloc requirements for intel/amd.
538 zoneindex(size_t *bytes, size_t *chunking)
540 size_t n = (size_t)*bytes;
547 *bytes = n = (n + 7) & ~7;
549 return(n / 8 - 1); /* 8 byte chunks, 2 zones */
551 *bytes = n = (n + 15) & ~15;
553 return(n / 16 + 2); /* 16 byte chunks, 8 zones */
558 c = x / (CHUNKFACTOR * 2);
562 c = x / (CHUNKFACTOR * 2);
563 i = 16 + CHUNKFACTOR * 5; /* 256->512,1024,2048,4096,8192 */
570 _mpanic("slaballoc: byte value too high");
572 *bytes = n = roundup2(n, c);
574 return (i + n / c - CHUNKFACTOR);
576 *bytes = n = (n + c - 1) & ~(c - 1);
581 *bytes = n = (n + 15) & ~15;
583 return(n / (CHUNKINGLO*2) + CHUNKINGLO*1 - 1);
587 *bytes = n = (n + 31) & ~31;
589 return(n / (CHUNKINGLO*4) + CHUNKINGLO*2 - 1);
592 *bytes = n = (n + 63) & ~63;
594 return(n / (CHUNKINGLO*8) + CHUNKINGLO*3 - 1);
597 *bytes = n = (n + 127) & ~127;
599 return(n / (CHUNKINGLO*16) + CHUNKINGLO*4 - 1);
602 *bytes = n = (n + 255) & ~255;
604 return(n / (CHUNKINGLO*32) + CHUNKINGLO*5 - 1);
606 *bytes = n = (n + 511) & ~511;
608 return(n / (CHUNKINGLO*64) + CHUNKINGLO*6 - 1);
611 *bytes = n = (n + 1023) & ~1023;
613 return(n / (CHUNKINGLO*128) + CHUNKINGLO*7 - 1);
615 if (n < 32768) { /* 16384-32767 */
616 *bytes = n = (n + 2047) & ~2047;
618 return(n / (CHUNKINGLO*256) + CHUNKINGLO*8 - 1);
621 *bytes = n = (n + 4095) & ~4095; /* 32768-65535 */
623 return(n / (CHUNKINGLO*512) + CHUNKINGLO*9 - 1);
628 i = CHUNKINGLO*10 - 1;
635 _mpanic("slaballoc: byte value too high");
637 *bytes = n = (n + c - 1) & ~(c - 1);
644 * malloc() - call internal slab allocator
647 __malloc(size_t size)
651 ptr = memalloc(size, 0);
655 UTRACE(0, size, ptr);
660 * calloc() - call internal slab allocator
663 __calloc(size_t number, size_t size)
667 ptr = memalloc(number * size, SAFLAG_ZERO);
671 UTRACE(0, number * size, ptr);
676 * realloc() (SLAB ALLOCATOR)
678 * We do not attempt to optimize this routine beyond reusing the same
679 * pointer if the new size fits within the chunking of the old pointer's
683 __realloc(void *ptr, size_t size)
688 ret = memalloc(size, 0);
690 ret = memrealloc(ptr, size);
694 UTRACE(ptr, size, ret);
701 * Allocate (size) bytes with a alignment of (alignment), where (alignment)
702 * is a power of 2 >= sizeof(void *).
704 * The slab allocator will allocate on power-of-2 boundaries up to at least
705 * PAGE_SIZE. Otherwise we use the zoneindex mechanic to find a zone
706 * matching the requirements.
709 __posix_memalign(void **memptr, size_t alignment, size_t size)
712 * OpenGroup spec issue 6 checks
714 if ((alignment | (alignment - 1)) + 1 != (alignment << 1)) {
718 if (alignment < sizeof(void *)) {
724 * XXX for now just find the nearest power of 2 >= size and also
725 * >= alignment and allocate that.
727 while (alignment < size) {
730 _mpanic("posix_memalign: byte value too high");
732 *memptr = memalloc(alignment, 0);
733 return(*memptr ? 0 : ENOMEM);
737 * free() (SLAB ALLOCATOR) - do the obvious
749 * memalloc() (SLAB ALLOCATOR)
751 * Allocate memory via the slab allocator.
754 memalloc(size_t size, int flags)
757 struct zoneinfo *zinfo;
771 * If 0 bytes is requested we have to return a unique pointer, allocate
777 /* Capture global flags */
778 flags |= g_malloc_flags;
780 /* Compute allocation zone; zoneindex will panic on excessive sizes */
781 zi = zoneindex(&size, &chunking);
782 MASSERT(zi < NZONES);
787 * Try magazine shortcut first
790 zinfo = &slgd->zone[zi];
792 if ((j = zinfo->mag_index) != 0) {
793 zinfo->mag_index = --j;
794 obj = zinfo->mag_shortcut[j];
795 zinfo->mag_shortcut[j] = NULL; /* SAFETY */
796 if (flags & SAFLAG_ZERO)
802 * Locate a slab with available space. If no slabs are available
803 * back-off to the empty list and if we still come up dry allocate
804 * a new slab (which will try the depot first).
807 if ((slab = zinfo->avail_base) == NULL) {
808 if ((slab = zinfo->empty_base) == NULL) {
812 slab = slaballoc(zi, chunking, size);
815 slab->next = zinfo->avail_base;
816 zinfo->avail_base = slab;
817 ++zinfo->avail_count;
819 if (slgd->biggest_index < zi)
820 slgd->biggest_index = zi;
824 * Pulled from empty list
826 zinfo->empty_base = slab->next;
827 slab->next = zinfo->avail_base;
828 zinfo->avail_base = slab;
829 ++zinfo->avail_count;
831 --zinfo->empty_count;
836 * Allocate a chunk out of the slab. HOT PATH
838 * Only the thread owning the slab can allocate out of it.
840 * NOTE: The last bit in the bitmap is always marked allocated so
841 * we cannot overflow here.
843 bno = slab->free_bit;
844 bmi = slab->free_index;
845 bmp = &slab->bitmap[bmi];
846 if (*bmp & (1LU << bno)) {
847 atomic_clear_long(bmp, 1LU << bno);
848 obj = slab->chunks + ((bmi << LONG_BITS_SHIFT) + bno) * size;
849 slab->free_bit = (bno + 1) & (LONG_BITS - 1);
850 atomic_add_int(&slab->navail, -1);
851 if (flags & SAFLAG_ZERO)
857 * Allocate a chunk out of a slab. COLD PATH
859 if (slab->navail == 0) {
860 zinfo->avail_base = slab->next;
861 --zinfo->avail_count;
863 TAILQ_INSERT_TAIL(&slgd->full_zones, slab, entry);
867 while (bmi < LONG_BITS) {
868 bmp = &slab->bitmap[bmi];
871 atomic_clear_long(bmp, 1LU << bno);
872 obj = slab->chunks + ((bmi << LONG_BITS_SHIFT) + bno) *
874 slab->free_index = bmi;
875 slab->free_bit = (bno + 1) & (LONG_BITS - 1);
876 atomic_add_int(&slab->navail, -1);
877 if (flags & SAFLAG_ZERO)
884 while (bmi < LONG_BITS) {
885 bmp = &slab->bitmap[bmi];
888 atomic_clear_long(bmp, 1LU << bno);
889 obj = slab->chunks + ((bmi << LONG_BITS_SHIFT) + bno) *
891 slab->free_index = bmi;
892 slab->free_bit = (bno + 1) & (LONG_BITS - 1);
893 atomic_add_int(&slab->navail, -1);
894 if (flags & SAFLAG_ZERO)
900 _mpanic("slaballoc: corrupted zone: navail %d", slab->navail);
906 * Reallocate memory within the chunk
909 memrealloc(void *ptr, size_t nsize)
918 * If 0 bytes is requested we have to return a unique pointer, allocate
924 /* Capture global flags */
925 flags |= g_malloc_flags;
928 * Locate the zone by looking up the dynamic slab size mask based
929 * on the memory region the allocation resides in.
931 region = &Regions[((uintptr_t)ptr >> NREGIONS_SHIFT) & NREGIONS_MASK];
932 if ((slab = region->slab) == NULL)
933 slab = (void *)((uintptr_t)ptr & region->mask);
934 MASSERT(slab->magic == ZALLOC_SLAB_MAGIC);
935 osize = slab->chunk_size;
936 if (nsize <= osize) {
937 if (osize < 32 || nsize >= osize / 2) {
939 if ((flags & SAFLAG_ZERO) && nsize < osize)
940 bzero(obj + nsize, osize - nsize);
946 * Otherwise resize the object
948 obj = memalloc(nsize, 0);
952 bcopy(ptr, obj, nsize);
959 * free (SLAB ALLOCATOR)
961 * Free a memory block previously allocated by malloc.
966 memfree(void *ptr, int flags)
979 * Locate the zone by looking up the dynamic slab size mask based
980 * on the memory region the allocation resides in.
982 * WARNING! The slab may be owned by another thread!
984 region = &Regions[((uintptr_t)ptr >> NREGIONS_SHIFT) & NREGIONS_MASK];
985 if ((slab = region->slab) == NULL)
986 slab = (void *)((uintptr_t)ptr & region->mask);
987 MASSERT(slab != NULL);
988 MASSERT(slab->magic == ZALLOC_SLAB_MAGIC);
992 * Put weird data into the memory to detect modifications after
993 * freeing, illegal pointer use after freeing (we should fault on
994 * the odd address), and so forth.
996 if (slab->chunk_size < sizeof(weirdary))
997 bcopy(weirdary, ptr, slab->chunk_size);
999 bcopy(weirdary, ptr, sizeof(weirdary));
1004 * Use mag_shortcut[] when possible
1006 if (slgd->masked == 0 && slab->chunk_size <= NOMSLABSIZE) {
1007 struct zoneinfo *zinfo;
1009 zinfo = &slgd->zone[slab->zone_index];
1010 j = zinfo->mag_index;
1011 if (j < NMAGSHORTCUT) {
1012 zinfo->mag_shortcut[j] = ptr;
1013 zinfo->mag_index = j + 1;
1019 * Free to slab and increment navail. We can delay incrementing
1020 * navail to prevent the slab from being destroyed out from under
1021 * us while we do other optimizations.
1023 bno = ((uintptr_t)ptr - (uintptr_t)slab->chunks) / slab->chunk_size;
1024 bmi = bno >> LONG_BITS_SHIFT;
1025 bno &= (LONG_BITS - 1);
1026 bmp = &slab->bitmap[bmi];
1028 MASSERT(bmi >= 0 && bmi < slab->nmax);
1029 MASSERT((*bmp & (1LU << bno)) == 0);
1030 atomic_set_long(bmp, 1LU << bno);
1032 if (slab->slgd == slgd) {
1034 * We can only do the following if we own the slab. Note
1035 * that navail can be incremented by any thread even if
1038 struct zoneinfo *zinfo;
1040 atomic_add_int(&slab->navail, 1);
1041 if (slab->free_index > bmi) {
1042 slab->free_index = bmi;
1043 slab->free_bit = bno;
1044 } else if (slab->free_index == bmi &&
1045 slab->free_bit > bno) {
1046 slab->free_bit = bno;
1048 zinfo = &slgd->zone[slab->zone_index];
1051 * Freeing an object from a full slab makes it less than
1052 * full. The slab must be moved to the available list.
1054 * If the available list has too many slabs, release some
1057 if (slab->state == FULL) {
1058 TAILQ_REMOVE(&slgd->full_zones, slab, entry);
1059 slab->state = AVAIL;
1060 stmp = zinfo->avail_base;
1062 zinfo->avail_base = slab;
1063 ++zinfo->avail_count;
1064 while (zinfo->avail_count > opt_cache) {
1065 slab = zinfo->avail_base;
1066 zinfo->avail_base = slab->next;
1067 --zinfo->avail_count;
1068 slabterm(slgd, slab);
1074 * If the slab becomes completely empty dispose of it in
1075 * some manner. By default each thread caches up to 4
1076 * empty slabs. Only small slabs are cached.
1078 if (slab->navail == slab->nmax && slab->state == AVAIL) {
1080 * Remove slab from available queue
1082 slabp = &zinfo->avail_base;
1083 while ((stmp = *slabp) != slab)
1084 slabp = &stmp->next;
1085 *slabp = slab->next;
1086 --zinfo->avail_count;
1088 if (opt_free || opt_cache == 0) {
1090 * If local caching is disabled cache the
1091 * slab in the depot (or free it).
1093 slabterm(slgd, slab);
1094 } else if (slab->slab_size > BIGSLABSIZE) {
1096 * We do not try to retain large slabs
1097 * in per-thread caches.
1099 slabterm(slgd, slab);
1100 } else if (zinfo->empty_count > opt_cache) {
1102 * We have too many slabs cached, but
1103 * instead of freeing this one free
1104 * an empty slab that's been idle longer.
1106 * (empty_count does not change)
1108 stmp = zinfo->empty_base;
1109 slab->state = EMPTY;
1110 slab->next = stmp->next;
1111 zinfo->empty_base = slab;
1112 slabterm(slgd, stmp);
1115 * Cache the empty slab in our thread local
1118 ++zinfo->empty_count;
1119 slab->state = EMPTY;
1120 slab->next = zinfo->empty_base;
1121 zinfo->empty_base = slab;
1124 } else if (slab->slgd == NULL && slab->navail + 1 == slab->nmax) {
1125 slglobaldata_t sldepot;
1128 * If freeing to a slab owned by the global depot, and
1129 * the slab becomes completely EMPTY, try to move it to
1132 sldepot = &slglobaldepot;
1134 _SPINLOCK(&sldepot->lock);
1135 if (slab->slgd == NULL && slab->navail + 1 == slab->nmax) {
1136 struct zoneinfo *zinfo;
1139 * Move the slab to the empty list
1141 MASSERT(slab->state == AVAIL);
1142 atomic_add_int(&slab->navail, 1);
1143 zinfo = &sldepot->zone[slab->zone_index];
1144 slabp = &zinfo->avail_base;
1145 while (slab != *slabp)
1146 slabp = &(*slabp)->next;
1147 *slabp = slab->next;
1148 --zinfo->avail_count;
1151 * Clean out excessive empty entries from the
1154 slab->state = EMPTY;
1155 slab->next = zinfo->empty_base;
1156 zinfo->empty_base = slab;
1157 ++zinfo->empty_count;
1158 while (zinfo->empty_count > opt_cache) {
1159 slab = zinfo->empty_base;
1160 zinfo->empty_base = slab->next;
1161 --zinfo->empty_count;
1162 slab->state = UNKNOWN;
1164 _SPINUNLOCK(&sldepot->lock);
1167 _SPINLOCK(&sldepot->lock);
1170 atomic_add_int(&slab->navail, 1);
1173 _SPINUNLOCK(&sldepot->lock);
1176 * We can't act on the slab other than by adjusting navail
1177 * (and the bitmap which we did in the common code at the
1180 atomic_add_int(&slab->navail, 1);
1187 * Allocate a new slab holding objects of size chunk_size.
1190 slaballoc(int zi, size_t chunking, size_t chunk_size)
1192 slglobaldata_t slgd;
1193 slglobaldata_t sldepot;
1194 struct zoneinfo *zinfo;
1201 uintptr_t chunk_offset;
1212 * First look in the depot. Any given zone in the depot may be
1213 * locked by being set to -1. We have to do this instead of simply
1214 * removing the entire chain because removing the entire chain can
1215 * cause racing threads to allocate local slabs for large objects,
1216 * resulting in a large VSZ.
1219 sldepot = &slglobaldepot;
1220 zinfo = &sldepot->zone[zi];
1222 if (zinfo->avail_base) {
1224 _SPINLOCK(&sldepot->lock);
1225 slab = zinfo->avail_base;
1227 MASSERT(slab->slgd == NULL);
1229 zinfo->avail_base = slab->next;
1230 --zinfo->avail_count;
1232 _SPINUNLOCK(&sldepot->lock);
1236 _SPINUNLOCK(&sldepot->lock);
1240 * Nothing in the depot, allocate a new slab by locating or assigning
1241 * a region and then using the system virtual memory allocator.
1246 * Calculate the start of the data chunks relative to the start
1247 * of the slab. If chunk_size is a power of 2 we guarantee
1248 * power of 2 alignment. If it is not we guarantee alignment
1249 * to the chunk size.
1251 if ((chunk_size ^ (chunk_size - 1)) == (chunk_size << 1) - 1) {
1253 chunk_offset = roundup2(sizeof(*slab), chunk_size);
1256 chunk_offset = sizeof(*slab) + chunking - 1;
1257 chunk_offset -= chunk_offset % chunking;
1261 * Calculate a reasonable number of chunks for the slab.
1263 * Once initialized the MaxChunks[] array can only ever be
1264 * reinitialized to the same value.
1266 maxchunks = MaxChunks[zi];
1267 if (maxchunks == 0) {
1269 * First calculate how many chunks would fit in 1/1024
1270 * available memory. This is around 2MB on a 32 bit
1271 * system and 128G on a 64-bit (48-bits available) system.
1273 maxchunks = (ssize_t)(NREGIONS_SIZE - chunk_offset) /
1274 (ssize_t)chunk_size;
1279 * A slab cannot handle more than MAXCHUNKS chunks, but
1280 * limit us to approximately MAXCHUNKS / 2 here because
1281 * we may have to expand maxchunks when we calculate the
1282 * actual power-of-2 slab.
1284 if (maxchunks > MAXCHUNKS / 2)
1285 maxchunks = MAXCHUNKS / 2;
1288 * Try to limit the slabs to BIGSLABSIZE (~128MB). Larger
1289 * slabs will be created if the allocation does not fit.
1291 if (chunk_offset + chunk_size * maxchunks > BIGSLABSIZE) {
1292 tmpchunks = (ssize_t)(BIGSLABSIZE - chunk_offset) /
1293 (ssize_t)chunk_size;
1296 if (maxchunks > tmpchunks)
1297 maxchunks = tmpchunks;
1301 * If the slab calculates to greater than 2MB see if we
1302 * can cut it down to ~2MB. This controls VSZ but has
1303 * no effect on run-time size or performance.
1305 * This is very important in case you core dump and also
1306 * important to reduce unnecessary region allocations.
1308 if (chunk_offset + chunk_size * maxchunks > NOMSLABSIZE) {
1309 tmpchunks = (ssize_t)(NOMSLABSIZE - chunk_offset) /
1310 (ssize_t)chunk_size;
1313 if (maxchunks > tmpchunks)
1314 maxchunks = tmpchunks;
1318 * If the slab calculates to greater than 128K see if we
1319 * can cut it down to ~128K while still maintaining a
1320 * reasonably large number of chunks in each slab. This
1321 * controls VSZ but has no effect on run-time size or
1324 * This is very important in case you core dump and also
1325 * important to reduce unnecessary region allocations.
1327 if (chunk_offset + chunk_size * maxchunks > LITSLABSIZE) {
1328 tmpchunks = (ssize_t)(LITSLABSIZE - chunk_offset) /
1329 (ssize_t)chunk_size;
1332 if (maxchunks > tmpchunks)
1333 maxchunks = tmpchunks;
1336 MaxChunks[zi] = maxchunks;
1338 MASSERT(maxchunks > 0 && maxchunks <= MAXCHUNKS);
1341 * Calculate the actual slab size. maxchunks will be recalculated
1344 slab_desire = chunk_offset + chunk_size * maxchunks;
1345 slab_size = 8 * MAXCHUNKS;
1346 power = 3 + MAXCHUNKS_BITS;
1347 while (slab_size < slab_desire) {
1353 * Do a quick recalculation based on the actual slab size but not
1354 * yet dealing with whether the slab header is in-band or out-of-band.
1355 * The purpose here is to see if we can reasonably reduce slab_size
1356 * to a power of 4 to allow more chunk sizes to use the same slab
1359 if ((power & 1) && slab_size > 32768) {
1360 maxchunks = (slab_size - chunk_offset) / chunk_size;
1361 if (maxchunks >= MAXCHUNKS / 8) {
1366 if ((power & 2) && slab_size > 32768 * 4) {
1367 maxchunks = (slab_size - chunk_offset) / chunk_size;
1368 if (maxchunks >= MAXCHUNKS / 4) {
1374 * This case occurs when the slab_size is larger than 1/1024 available
1377 nswath = slab_size / NREGIONS_SIZE;
1378 if (nswath > NREGIONS)
1383 * Try to allocate from our current best region for this zi
1385 region_mask = ~(slab_size - 1);
1386 ri = slgd->zone[zi].best_region;
1387 if (Regions[ri].mask == region_mask) {
1388 if ((slab = _vmem_alloc(ri, slab_size)) != NULL)
1393 * Try to find an existing region to allocate from. The normal
1394 * case will be for allocations that are less than 1/1024 available
1395 * UVM, which fit into a single Regions[] entry.
1397 while (slab_size <= NREGIONS_SIZE) {
1399 for (ri = 0; ri < NREGIONS; ++ri) {
1400 if (rx < 0 && Regions[ri].mask == 0)
1402 if (Regions[ri].mask == region_mask) {
1403 slab = _vmem_alloc(ri, slab_size);
1405 slgd->zone[zi].best_region = ri;
1415 * This can fail, retry either way
1417 atomic_cmpset_ptr((void **)&Regions[rx].mask,
1419 (void *)region_mask);
1424 for (ri = 0; ri < NREGIONS; ri += nswath) {
1425 if (Regions[ri].mask == region_mask) {
1426 slab = _vmem_alloc(ri, slab_size);
1428 slgd->zone[zi].best_region = ri;
1433 for (j = nswath - 1; j >= 0; --j) {
1434 if (Regions[ri+j].mask != 0)
1443 * We found a candidate, try to allocate it backwards so
1444 * another thread racing a slaballoc() does not see the
1445 * mask in the base index position until we are done.
1447 * We can safely zero-out any partial allocations because
1448 * the mask is only accessed from the base index. Any other
1449 * threads racing us will fail prior to us clearing the mask.
1453 for (j = nswath - 1; j >= 0; --j) {
1454 if (!atomic_cmpset_ptr((void **)&Regions[rx+j].mask,
1455 NULL, (void *)region_mask)) {
1456 while (++j < nswath)
1457 Regions[rx+j].mask = 0;
1465 * Fill in the new slab in region ri. If the slab_size completely
1466 * fills one or more region slots we move the slab structure out of
1467 * band which should optimize the chunking (particularly for a power
1471 region = &Regions[ri];
1472 MASSERT(region->slab == NULL);
1473 if (slab_size >= NREGIONS_SIZE) {
1475 slab = memalloc(sizeof(*slab), 0);
1476 bzero(slab, sizeof(*slab));
1477 slab->chunks = save;
1478 for (j = 0; j < nswath; ++j)
1479 region[j].slab = slab;
1482 bzero(slab, sizeof(*slab));
1483 slab->chunks = (char *)slab + chunk_offset;
1487 * Calculate the start of the chunks memory and recalculate the
1488 * actual number of chunks the slab can hold.
1490 maxchunks = (slab_size - chunk_offset) / chunk_size;
1491 if (maxchunks > MAXCHUNKS)
1492 maxchunks = MAXCHUNKS;
1495 * And fill in the rest
1497 slab->magic = ZALLOC_SLAB_MAGIC;
1498 slab->navail = maxchunks;
1499 slab->nmax = maxchunks;
1500 slab->slab_size = slab_size;
1501 slab->chunk_size = chunk_size;
1502 slab->zone_index = zi;
1504 slab->state = UNKNOWN;
1505 slab->region = region;
1507 for (ri = 0; ri < maxchunks; ri += LONG_BITS) {
1508 if (ri + LONG_BITS <= maxchunks)
1509 slab->bitmap[ri >> LONG_BITS_SHIFT] = ULONG_MAX;
1511 slab->bitmap[ri >> LONG_BITS_SHIFT] =
1512 (1LU << (maxchunks - ri)) - 1;
1521 slabfree(slab_t slab)
1526 if (slab->region->slab == slab) {
1530 nswath = slab->slab_size / NREGIONS_SIZE;
1531 for (j = 0; j < nswath; ++j)
1532 slab->region[j].slab = NULL;
1534 _vmem_free(slab->chunks, slab->slab_size);
1541 _vmem_free(slab, slab->slab_size);
1546 * Terminate a slab's use in the current thread. The slab may still have
1547 * outstanding allocations and thus not be deallocatable.
1550 slabterm(slglobaldata_t slgd, slab_t slab)
1552 slglobaldata_t sldepot;
1553 struct zoneinfo *zinfo;
1554 int zi = slab->zone_index;
1558 sldepot = &slglobaldepot;
1559 zinfo = &sldepot->zone[zi];
1562 * Move the slab to the avail list or the empty list.
1565 _SPINLOCK(&sldepot->lock);
1566 if (slab->navail == slab->nmax) {
1567 slab->state = EMPTY;
1568 slab->next = zinfo->empty_base;
1569 zinfo->empty_base = slab;
1570 ++zinfo->empty_count;
1572 slab->state = AVAIL;
1573 slab->next = zinfo->avail_base;
1574 zinfo->avail_base = slab;
1575 ++zinfo->avail_count;
1579 * Clean extra slabs out of the empty list
1581 while (zinfo->empty_count > opt_cache) {
1582 slab = zinfo->empty_base;
1583 zinfo->empty_base = slab->next;
1584 --zinfo->empty_count;
1585 slab->state = UNKNOWN;
1587 _SPINUNLOCK(&sldepot->lock);
1590 _SPINLOCK(&sldepot->lock);
1593 _SPINUNLOCK(&sldepot->lock);
1599 * Directly map memory in PAGE_SIZE'd chunks with the specified
1602 * Alignment must be a multiple of PAGE_SIZE.
1604 * Size must be >= alignment.
1607 _vmem_alloc(int ri, size_t slab_size)
1609 char *baddr = (void *)((uintptr_t)ri << NREGIONS_SHIFT);
1615 if (slab_size < NREGIONS_SIZE)
1616 eaddr = baddr + NREGIONS_SIZE;
1618 eaddr = baddr + slab_size;
1621 * This usually just works but might not.
1623 addr = mmap(baddr, slab_size, PROT_READ|PROT_WRITE,
1624 MAP_PRIVATE | MAP_ANON | MAP_SIZEALIGN, -1, 0);
1625 if (addr == MAP_FAILED) {
1628 if (addr < baddr || addr + slab_size > eaddr) {
1629 munmap(addr, slab_size);
1634 * Check alignment. The misaligned offset is also the excess
1635 * amount. If misaligned unmap the excess so we have a chance of
1636 * mapping at the next alignment point and recursively try again.
1638 * BBBBBBBBBBB BBBBBBBBBBB BBBBBBBBBBB block alignment
1639 * aaaaaaaaa aaaaaaaaaaa aa mis-aligned allocation
1640 * xxxxxxxxx final excess calculation
1641 * ^ returned address
1643 excess = (uintptr_t)addr & (slab_size - 1);
1645 excess = slab_size - excess;
1648 munmap(save + excess, slab_size - excess);
1649 addr = _vmem_alloc(ri, slab_size);
1650 munmap(save, excess);
1653 if (addr < baddr || addr + slab_size > eaddr) {
1654 munmap(addr, slab_size);
1657 excess = (uintptr_t)addr & (slab_size - 1);
1665 * Free a chunk of memory allocated with _vmem_alloc()
1668 _vmem_free(void *ptr, size_t size)
1674 * Panic on fatal conditions
1677 _mpanic(const char *ctl, ...)
1681 if (malloc_panic == 0) {
1684 vfprintf(stderr, ctl, va);
1685 fprintf(stderr, "\n");
1692 __weak_reference(__malloc, malloc);
1693 __weak_reference(__calloc, calloc);
1694 __weak_reference(__posix_memalign, posix_memalign);
1695 __weak_reference(__realloc, realloc);
1696 __weak_reference(__free, free);