2 * NMALLOC.C - New Malloc (ported from kernel slab allocator)
4 * Copyright (c) 2003,2004,2009,2010 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> and by
8 * Venkatesh Srinivas <me@endeavour.zapto.org>.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in
18 * the documentation and/or other materials provided with the
20 * 3. Neither the name of The DragonFly Project nor the names of its
21 * contributors may be used to endorse or promote products derived
22 * from this software without specific, prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
25 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
26 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
27 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
28 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
29 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
30 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
31 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
32 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
33 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
34 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37 * $Id: nmalloc.c,v 1.37 2010/07/23 08:20:35 vsrinivas Exp $
40 * This module implements a slab allocator drop-in replacement for the
43 * A slab allocator reserves a ZONE for each chunk size, then lays the
44 * chunks out in an array within the zone. Allocation and deallocation
45 * is nearly instantaneous, and overhead losses are limited to a fixed
48 * The slab allocator does not have to pre-initialize the list of
49 * free chunks for each zone, and the underlying VM will not be
50 * touched at all beyond the zone header until an actual allocation
53 * Slab management and locking is done on a per-zone basis.
55 * Alloc Size Chunking Number of zones
66 * Allocations >= ZoneLimit (16K) go directly to mmap and a hash table
67 * is used to locate for free. One and Two-page allocations use the
68 * zone mechanic to avoid excessive mmap()/munmap() calls.
70 * API FEATURES AND SIDE EFFECTS
72 * + power-of-2 sized allocations up to a page will be power-of-2 aligned.
73 * Above that power-of-2 sized allocations are page-aligned. Non
74 * power-of-2 sized allocations are aligned the same as the chunk
75 * size for their zone.
76 * + malloc(0) returns a special non-NULL value
77 * + ability to allocate arbitrarily large chunks of memory
78 * + realloc will reuse the passed pointer if possible, within the
79 * limitations of the zone chunking.
81 * Multithreaded enhancements for small allocations introduced August 2010.
82 * These are in the spirit of 'libumem'. See:
83 * Bonwick, J.; Adams, J. (2001). "Magazines and Vmem: Extending the
84 * slab allocator to many CPUs and arbitrary resources". In Proc. 2001
85 * USENIX Technical Conference. USENIX Association.
87 * Oversized allocations employ the BIGCACHE mechanic whereby large
88 * allocations may be handed significantly larger buffers, allowing them
89 * to avoid mmap/munmap operations even through significant realloc()s.
90 * The excess space is only trimmed if too many large allocations have been
91 * given this treatment.
95 * The value of the environment variable MALLOC_OPTIONS is a character string
96 * containing various flags to tune nmalloc.
98 * 'U' / ['u'] Generate / do not generate utrace entries for ktrace(1)
99 * This will generate utrace events for all malloc,
100 * realloc, and free calls. There are tools (mtrplay) to
101 * replay and allocation pattern or to graph heap structure
102 * (mtrgraph) which can interpret these logs.
103 * 'Z' / ['z'] Zero out / do not zero all allocations.
104 * Each new byte of memory allocated by malloc, realloc, or
105 * reallocf will be initialized to 0. This is intended for
106 * debugging and will affect performance negatively.
107 * 'H' / ['h'] Pass a hint to the kernel about pages unused by the
108 * allocation functions.
111 /* cc -shared -fPIC -g -O -I/usr/src/lib/libc/include -o nmalloc.so nmalloc.c */
113 #include "libc_private.h"
115 #include <sys/param.h>
116 #include <sys/types.h>
117 #include <sys/mman.h>
118 #include <sys/queue.h>
120 #include <sys/ktrace.h>
131 #include <machine/atomic.h>
133 #include "spinlock.h"
134 #include "un-namespace.h"
138 * Linked list of large allocations
140 typedef struct bigalloc {
141 struct bigalloc *next; /* hash link */
142 void *base; /* base pointer */
143 u_long active; /* bytes active */
144 u_long bytes; /* bytes allocated */
148 * Note that any allocations which are exact multiples of PAGE_SIZE, or
149 * which are >= ZALLOC_ZONE_LIMIT, will fall through to the kmem subsystem.
151 #define ZALLOC_ZONE_LIMIT (16 * 1024) /* max slab-managed alloc */
152 #define ZALLOC_MIN_ZONE_SIZE (32 * 1024) /* minimum zone size */
153 #define ZALLOC_MAX_ZONE_SIZE (128 * 1024) /* maximum zone size */
154 #define ZALLOC_ZONE_SIZE (64 * 1024)
155 #define ZALLOC_SLAB_MAGIC 0x736c6162 /* magic sanity */
156 #define ZALLOC_SLAB_SLIDE 20 /* L1-cache skip */
158 #if ZALLOC_ZONE_LIMIT == 16384
160 #elif ZALLOC_ZONE_LIMIT == 32768
163 #error "I couldn't figure out NZONES"
167 * Chunk structure for free elements
169 typedef struct slchunk {
170 struct slchunk *c_Next;
174 * The IN-BAND zone header is placed at the beginning of each zone.
178 typedef struct slzone {
179 int32_t z_Magic; /* magic number for sanity check */
180 int z_NFree; /* total free chunks / ualloc space */
181 struct slzone *z_Next; /* ZoneAry[] link if z_NFree non-zero */
182 int z_NMax; /* maximum free chunks */
183 char *z_BasePtr; /* pointer to start of chunk array */
184 int z_UIndex; /* current initial allocation index */
185 int z_UEndIndex; /* last (first) allocation index */
186 int z_ChunkSize; /* chunk size for validation */
187 int z_FirstFreePg; /* chunk list on a page-by-page basis */
190 struct slchunk *z_PageAry[ZALLOC_ZONE_SIZE / PAGE_SIZE];
193 typedef struct slglobaldata {
195 slzone_t ZoneAry[NZONES];/* linked list of zones NFree > 0 */
199 #define SLZF_UNOTZEROD 0x0001
201 #define FASTSLABREALLOC 0x02
204 * Misc constants. Note that allocations that are exact multiples of
205 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
206 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
208 #define MIN_CHUNK_SIZE 8 /* in bytes */
209 #define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
210 #define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
213 * WARNING: A limited number of spinlocks are available, BIGXSIZE should
214 * not be larger then 64.
216 #define BIGHSHIFT 10 /* bigalloc hash table */
217 #define BIGHSIZE (1 << BIGHSHIFT)
218 #define BIGHMASK (BIGHSIZE - 1)
219 #define BIGXSIZE (BIGHSIZE / 16) /* bigalloc lock table */
220 #define BIGXMASK (BIGXSIZE - 1)
223 * BIGCACHE caches oversized allocations. Note that a linear search is
224 * performed, so do not make the cache too large.
226 * BIGCACHE will garbage-collect excess space when the excess exceeds the
227 * specified value. A relatively large number should be used here because
228 * garbage collection is expensive.
231 #define BIGCACHE_MASK (BIGCACHE - 1)
232 #define BIGCACHE_LIMIT (1024 * 1024) /* size limit */
233 #define BIGCACHE_EXCESS (16 * 1024 * 1024) /* garbage collect */
235 #define SAFLAG_ZERO 0x0001
236 #define SAFLAG_PASSIVE 0x0002
242 #define arysize(ary) (sizeof(ary)/sizeof((ary)[0]))
244 #define MASSERT(exp) do { if (__predict_false(!(exp))) \
245 _mpanic("assertion: %s in %s", \
253 #define M_MAX_ROUNDS 64
254 #define M_ZONE_ROUNDS 64
255 #define M_LOW_ROUNDS 32
256 #define M_INIT_ROUNDS 8
257 #define M_BURST_FACTOR 8
258 #define M_BURST_NSCALE 2
260 #define M_BURST 0x0001
261 #define M_BURST_EARLY 0x0002
264 SLIST_ENTRY(magazine) nextmagazine;
267 int capacity; /* Max rounds in this magazine */
268 int rounds; /* Current number of free rounds */
269 int burst_factor; /* Number of blocks to prefill with */
270 int low_factor; /* Free till low_factor from full mag */
271 void *objects[M_MAX_ROUNDS];
274 SLIST_HEAD(magazinelist, magazine);
276 static spinlock_t zone_mag_lock;
277 static spinlock_t depot_spinlock;
278 static struct magazine zone_magazine = {
279 .flags = M_BURST | M_BURST_EARLY,
280 .capacity = M_ZONE_ROUNDS,
282 .burst_factor = M_BURST_FACTOR,
283 .low_factor = M_LOW_ROUNDS
286 #define MAGAZINE_FULL(mp) (mp->rounds == mp->capacity)
287 #define MAGAZINE_NOTFULL(mp) (mp->rounds < mp->capacity)
288 #define MAGAZINE_EMPTY(mp) (mp->rounds == 0)
289 #define MAGAZINE_NOTEMPTY(mp) (mp->rounds != 0)
292 * Each thread will have a pair of magazines per size-class (NZONES)
293 * The loaded magazine will support immediate allocations, the previous
294 * magazine will either be full or empty and can be swapped at need
296 typedef struct magazine_pair {
297 struct magazine *loaded;
298 struct magazine *prev;
301 /* A depot is a collection of magazines for a single zone. */
302 typedef struct magazine_depot {
303 struct magazinelist full;
304 struct magazinelist empty;
308 typedef struct thr_mags {
309 magazine_pair mags[NZONES];
310 struct magazine *newmag;
315 * With this attribute set, do not require a function call for accessing
316 * this variable when the code is compiled -fPIC.
318 * Must be empty for libc_rtld (similar to __thread).
321 #define TLS_ATTRIBUTE
323 #define TLS_ATTRIBUTE __attribute__ ((tls_model ("initial-exec")))
326 static __thread thr_mags thread_mags TLS_ATTRIBUTE;
327 static pthread_key_t thread_mags_key;
328 static pthread_once_t thread_mags_once = PTHREAD_ONCE_INIT;
329 static magazine_depot depots[NZONES];
332 * Fixed globals (not per-cpu)
334 static const int ZoneSize = ZALLOC_ZONE_SIZE;
335 static const int ZoneLimit = ZALLOC_ZONE_LIMIT;
336 static const int ZonePageCount = ZALLOC_ZONE_SIZE / PAGE_SIZE;
337 static const int ZoneMask = ZALLOC_ZONE_SIZE - 1;
339 static int opt_madvise = 0;
340 static int opt_utrace = 0;
341 static int g_malloc_flags = 0;
342 static struct slglobaldata SLGlobalData;
343 static bigalloc_t bigalloc_array[BIGHSIZE];
344 static spinlock_t bigspin_array[BIGXSIZE];
345 static volatile void *bigcache_array[BIGCACHE]; /* atomic swap */
346 static volatile size_t bigcache_size_array[BIGCACHE]; /* SMP races ok */
347 static volatile int bigcache_index; /* SMP races ok */
348 static int malloc_panic;
349 static size_t excess_alloc; /* excess big allocs */
351 static void *_slaballoc(size_t size, int flags);
352 static void *_slabrealloc(void *ptr, size_t size);
353 static void _slabfree(void *ptr, int, bigalloc_t *);
354 static void *_vmem_alloc(size_t bytes, size_t align, int flags);
355 static void _vmem_free(void *ptr, size_t bytes);
356 static void *magazine_alloc(struct magazine *, int *);
357 static int magazine_free(struct magazine *, void *);
358 static void *mtmagazine_alloc(int zi);
359 static int mtmagazine_free(int zi, void *);
360 static void mtmagazine_init(void);
361 static void mtmagazine_destructor(void *);
362 static slzone_t zone_alloc(int flags);
363 static void zone_free(void *z);
364 static void _mpanic(const char *ctl, ...) __printflike(1, 2);
365 static void malloc_init(void) __constructor(101);
367 struct nmalloc_utrace {
373 #define UTRACE(a, b, c) \
375 struct nmalloc_utrace ut = { \
380 utrace(&ut, sizeof(ut)); \
386 const char *p = NULL;
388 if (issetugid() == 0)
389 p = getenv("MALLOC_OPTIONS");
391 for (; p != NULL && *p != '\0'; p++) {
393 case 'u': opt_utrace = 0; break;
394 case 'U': opt_utrace = 1; break;
395 case 'h': opt_madvise = 0; break;
396 case 'H': opt_madvise = 1; break;
397 case 'z': g_malloc_flags = 0; break;
398 case 'Z': g_malloc_flags = SAFLAG_ZERO; break;
404 UTRACE((void *) -1, 0, NULL);
408 * We have to install a handler for nmalloc thread teardowns when
409 * the thread is created. We cannot delay this because destructors in
410 * sophisticated userland programs can call malloc() for the first time
411 * during their thread exit.
413 * This routine is called directly from pthreads.
416 _nmalloc_thr_init(void)
418 static int init_once;
422 * Disallow mtmagazine operations until the mtmagazine is
428 if (init_once == 0) {
430 pthread_once(&thread_mags_once, mtmagazine_init);
432 pthread_setspecific(thread_mags_key, tp);
437 _nmalloc_thr_prepfork(void)
440 _SPINLOCK(&zone_mag_lock);
441 _SPINLOCK(&depot_spinlock);
446 _nmalloc_thr_parentfork(void)
449 _SPINUNLOCK(&depot_spinlock);
450 _SPINUNLOCK(&zone_mag_lock);
455 _nmalloc_thr_childfork(void)
458 _SPINUNLOCK(&depot_spinlock);
459 _SPINUNLOCK(&zone_mag_lock);
467 slgd_lock(slglobaldata_t slgd)
470 _SPINLOCK(&slgd->Spinlock);
474 slgd_unlock(slglobaldata_t slgd)
477 _SPINUNLOCK(&slgd->Spinlock);
481 depot_lock(magazine_depot *dp)
484 _SPINLOCK(&depot_spinlock);
487 _SPINLOCK(&dp->lock);
492 depot_unlock(magazine_depot *dp)
495 _SPINUNLOCK(&depot_spinlock);
498 _SPINUNLOCK(&dp->lock);
503 zone_magazine_lock(void)
506 _SPINLOCK(&zone_mag_lock);
510 zone_magazine_unlock(void)
513 _SPINUNLOCK(&zone_mag_lock);
517 swap_mags(magazine_pair *mp)
519 struct magazine *tmp;
521 mp->loaded = mp->prev;
526 * bigalloc hashing and locking support.
528 * Return an unmasked hash code for the passed pointer.
531 _bigalloc_hash(void *ptr)
535 hv = ((int)(intptr_t)ptr >> PAGE_SHIFT) ^
536 ((int)(intptr_t)ptr >> (PAGE_SHIFT + BIGHSHIFT));
542 * Lock the hash chain and return a pointer to its base for the specified
545 static __inline bigalloc_t *
546 bigalloc_lock(void *ptr)
548 int hv = _bigalloc_hash(ptr);
551 bigp = &bigalloc_array[hv & BIGHMASK];
553 _SPINLOCK(&bigspin_array[hv & BIGXMASK]);
558 * Lock the hash chain and return a pointer to its base for the specified
561 * BUT, if the hash chain is empty, just return NULL and do not bother
564 static __inline bigalloc_t *
565 bigalloc_check_and_lock(void *ptr)
567 int hv = _bigalloc_hash(ptr);
570 bigp = &bigalloc_array[hv & BIGHMASK];
574 _SPINLOCK(&bigspin_array[hv & BIGXMASK]);
580 bigalloc_unlock(void *ptr)
585 hv = _bigalloc_hash(ptr);
586 _SPINUNLOCK(&bigspin_array[hv & BIGXMASK]);
591 * Find a bigcache entry that might work for the allocation. SMP races are
592 * ok here except for the swap (that is, it is ok if bigcache_size_array[i]
593 * is wrong or if a NULL or too-small big is returned).
595 * Generally speaking it is ok to find a large entry even if the bytes
596 * requested are relatively small (but still oversized), because we really
597 * don't know *what* the application is going to do with the buffer.
601 bigcache_find_alloc(size_t bytes)
603 bigalloc_t big = NULL;
607 for (i = 0; i < BIGCACHE; ++i) {
608 test = bigcache_size_array[i];
610 bigcache_size_array[i] = 0;
611 big = atomic_swap_ptr(&bigcache_array[i], NULL);
619 * Free a bigcache entry, possibly returning one that the caller really must
620 * free. This is used to cache recent oversized memory blocks. Only
621 * big blocks smaller than BIGCACHE_LIMIT will be cached this way, so try
622 * to collect the biggest ones we can that are under the limit.
626 bigcache_find_free(bigalloc_t big)
632 b = ++bigcache_index;
633 for (i = 0; i < BIGCACHE; ++i) {
634 j = (b + i) & BIGCACHE_MASK;
635 if (bigcache_size_array[j] < big->bytes) {
636 bigcache_size_array[j] = big->bytes;
637 big = atomic_swap_ptr(&bigcache_array[j], big);
646 handle_excess_big(void)
652 if (excess_alloc <= BIGCACHE_EXCESS)
655 for (i = 0; i < BIGHSIZE; ++i) {
656 bigp = &bigalloc_array[i];
660 _SPINLOCK(&bigspin_array[i & BIGXMASK]);
661 for (big = *bigp; big; big = big->next) {
662 if (big->active < big->bytes) {
663 MASSERT((big->active & PAGE_MASK) == 0);
664 MASSERT((big->bytes & PAGE_MASK) == 0);
665 munmap((char *)big->base + big->active,
666 big->bytes - big->active);
667 atomic_add_long(&excess_alloc,
668 big->active - big->bytes);
669 big->bytes = big->active;
673 _SPINUNLOCK(&bigspin_array[i & BIGXMASK]);
678 * Calculate the zone index for the allocation request size and set the
679 * allocation request size to that particular zone's chunk size.
682 zoneindex(size_t *bytes, size_t *chunking)
684 size_t n = (unsigned int)*bytes; /* unsigned for shift opt */
687 * This used to be 8-byte chunks and 16 zones for n < 128.
688 * However some instructions may require 16-byte alignment
689 * (aka SIMD) and programs might not request an aligned size
690 * (aka GCC-7), so change this as follows:
692 * 0-15 bytes 8-byte alignment in two zones (0-1)
693 * 16-127 bytes 16-byte alignment in four zones (3-10)
694 * zone index 2 and 11-15 are currently unused.
697 *bytes = n = (n + 7) & ~7;
699 return(n / 8 - 1); /* 8 byte chunks, 2 zones */
700 /* zones 0,1, zone 2 is unused */
703 *bytes = n = (n + 15) & ~15;
705 return(n / 16 + 2); /* 16 byte chunks, 8 zones */
706 /* zones 3-10, zones 11-15 unused */
709 *bytes = n = (n + 15) & ~15;
715 *bytes = n = (n + 31) & ~31;
720 *bytes = n = (n + 63) & ~63;
725 *bytes = n = (n + 127) & ~127;
727 return(n / 128 + 31);
730 *bytes = n = (n + 255) & ~255;
732 return(n / 256 + 39);
734 *bytes = n = (n + 511) & ~511;
736 return(n / 512 + 47);
738 #if ZALLOC_ZONE_LIMIT > 8192
740 *bytes = n = (n + 1023) & ~1023;
742 return(n / 1024 + 55);
745 #if ZALLOC_ZONE_LIMIT > 16384
747 *bytes = n = (n + 2047) & ~2047;
749 return(n / 2048 + 63);
752 _mpanic("Unexpected byte count %zu", n);
757 * malloc() - call internal slab allocator
760 __malloc(size_t size)
764 ptr = _slaballoc(size, 0);
768 UTRACE(0, size, ptr);
772 #define MUL_NO_OVERFLOW (1UL << (sizeof(size_t) * 4))
775 * calloc() - call internal slab allocator
778 __calloc(size_t number, size_t size)
782 if ((number >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&
783 number > 0 && SIZE_MAX / number < size) {
788 ptr = _slaballoc(number * size, SAFLAG_ZERO);
792 UTRACE(0, number * size, ptr);
797 * realloc() (SLAB ALLOCATOR)
799 * We do not attempt to optimize this routine beyond reusing the same
800 * pointer if the new size fits within the chunking of the old pointer's
804 __realloc(void *ptr, size_t size)
807 ret = _slabrealloc(ptr, size);
811 UTRACE(ptr, size, ret);
818 * Allocate (size) bytes with a alignment of (alignment), where (alignment)
819 * is a power of 2 >= sizeof(void *).
821 * The slab allocator will allocate on power-of-2 boundaries up to
822 * at least PAGE_SIZE. We use the zoneindex mechanic to find a
823 * zone matching the requirements, and _vmem_alloc() otherwise.
826 __posix_memalign(void **memptr, size_t alignment, size_t size)
834 * OpenGroup spec issue 6 checks
836 if ((alignment | (alignment - 1)) + 1 != (alignment << 1)) {
840 if (alignment < sizeof(void *)) {
846 * Our zone mechanism guarantees same-sized alignment for any
847 * power-of-2 allocation. If size is a power-of-2 and reasonable
848 * we can just call _slaballoc() and be done. We round size up
849 * to the nearest alignment boundary to improve our odds of
850 * it becoming a power-of-2 if it wasn't before.
852 if (size <= alignment)
855 size = (size + alignment - 1) & ~(size_t)(alignment - 1);
858 * If we have overflown above when rounding to the nearest alignment
859 * boundary, just return ENOMEM, size should be == N * sizeof(void *).
864 if (size < PAGE_SIZE && (size | (size - 1)) + 1 == (size << 1)) {
865 *memptr = _slaballoc(size, 0);
866 return(*memptr ? 0 : ENOMEM);
870 * Otherwise locate a zone with a chunking that matches
871 * the requested alignment, within reason. Consider two cases:
873 * (1) A 1K allocation on a 32-byte alignment. The first zoneindex
874 * we find will be the best fit because the chunking will be
875 * greater or equal to the alignment.
877 * (2) A 513 allocation on a 256-byte alignment. In this case
878 * the first zoneindex we find will be for 576 byte allocations
879 * with a chunking of 64, which is not sufficient. To fix this
880 * we simply find the nearest power-of-2 >= size and use the
881 * same side-effect of _slaballoc() which guarantees
882 * same-alignment on a power-of-2 allocation.
884 if (size < PAGE_SIZE) {
885 zi = zoneindex(&size, &chunking);
886 if (chunking >= alignment) {
887 *memptr = _slaballoc(size, 0);
888 return(*memptr ? 0 : ENOMEM);
894 while (alignment < size)
896 *memptr = _slaballoc(alignment, 0);
897 return(*memptr ? 0 : ENOMEM);
901 * If the slab allocator cannot handle it use vmem_alloc().
903 * Alignment must be adjusted up to at least PAGE_SIZE in this case.
905 if (alignment < PAGE_SIZE)
906 alignment = PAGE_SIZE;
907 if (size < alignment)
909 size = (size + PAGE_MASK) & ~(size_t)PAGE_MASK;
910 *memptr = _vmem_alloc(size, alignment, 0);
914 big = _slaballoc(sizeof(struct bigalloc), 0);
916 _vmem_free(*memptr, size);
920 bigp = bigalloc_lock(*memptr);
923 big->bytes = size; /* no excess */
926 bigalloc_unlock(*memptr);
932 * free() (SLAB ALLOCATOR) - do the obvious
938 _slabfree(ptr, 0, NULL);
942 * _slaballoc() (SLAB ALLOCATOR)
944 * Allocate memory via the slab allocator. If the request is too large,
945 * or if it page-aligned beyond a certain size, we fall back to the
949 _slaballoc(size_t size, int flags)
960 * Handle the degenerate size == 0 case. Yes, this does happen.
961 * Return a special pointer. This is to maintain compatibility with
962 * the original malloc implementation. Certain devices, such as the
963 * adaptec driver, not only allocate 0 bytes, they check for NULL and
964 * also realloc() later on. Joy.
969 /* Capture global flags */
970 flags |= g_malloc_flags;
973 * Handle large allocations directly. There should not be very many
974 * of these so performance is not a big issue.
976 * The backend allocator is pretty nasty on a SMP system. Use the
977 * slab allocator for one and two page-sized chunks even though we
978 * lose some efficiency.
980 if (size >= ZoneLimit ||
981 ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
986 * Page-align and cache-color in case of virtually indexed
987 * physically tagged L1 caches (aka SandyBridge). No sweat
988 * otherwise, so just do it.
990 * (don't count as excess).
992 size = (size + PAGE_MASK) & ~(size_t)PAGE_MASK;
993 if ((size & (PAGE_SIZE * 2 - 1)) == 0)
997 * Try to reuse a cached big block to avoid mmap'ing. If it
998 * turns out not to fit our requirements we throw it away
999 * and allocate normally.
1002 if (size <= BIGCACHE_LIMIT) {
1003 big = bigcache_find_alloc(size);
1004 if (big && big->bytes < size) {
1005 _slabfree(big->base, FASTSLABREALLOC, &big);
1011 if (flags & SAFLAG_ZERO)
1014 chunk = _vmem_alloc(size, PAGE_SIZE, flags);
1018 big = _slaballoc(sizeof(struct bigalloc), 0);
1020 _vmem_free(chunk, size);
1028 bigp = bigalloc_lock(chunk);
1029 if (big->active < big->bytes) {
1030 atomic_add_long(&excess_alloc,
1031 big->bytes - big->active);
1035 bigalloc_unlock(chunk);
1036 handle_excess_big();
1041 /* Compute allocation zone; zoneindex will panic on excessive sizes */
1042 zi = zoneindex(&size, &chunking);
1043 MASSERT(zi < NZONES);
1045 obj = mtmagazine_alloc(zi);
1047 if (flags & SAFLAG_ZERO)
1052 slgd = &SLGlobalData;
1056 * Attempt to allocate out of an existing zone. If all zones are
1057 * exhausted pull one off the free list or allocate a new one.
1059 if ((z = slgd->ZoneAry[zi]) == NULL) {
1060 z = zone_alloc(flags);
1065 * How big is the base structure?
1067 off = sizeof(struct slzone);
1070 * Align the storage in the zone based on the chunking.
1072 * Guarantee power-of-2 alignment for power-of-2-sized
1073 * chunks. Otherwise align based on the chunking size
1074 * (typically 8 or 16 bytes for small allocations).
1076 * NOTE: Allocations >= ZoneLimit are governed by the
1077 * bigalloc code and typically only guarantee page-alignment.
1079 * Set initial conditions for UIndex near the zone header
1080 * to reduce unecessary page faults, vs semi-randomization
1081 * to improve L1 cache saturation.
1083 if ((size | (size - 1)) + 1 == (size << 1))
1084 off = roundup2(off, size);
1086 off = roundup2(off, chunking);
1087 z->z_Magic = ZALLOC_SLAB_MAGIC;
1088 z->z_ZoneIndex = zi;
1089 z->z_NMax = (ZoneSize - off) / size;
1090 z->z_NFree = z->z_NMax;
1091 z->z_BasePtr = (char *)z + off;
1092 z->z_UIndex = z->z_UEndIndex = 0;
1093 z->z_ChunkSize = size;
1094 z->z_FirstFreePg = ZonePageCount;
1095 z->z_Next = slgd->ZoneAry[zi];
1096 slgd->ZoneAry[zi] = z;
1097 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
1098 flags &= ~SAFLAG_ZERO; /* already zero'd */
1099 flags |= SAFLAG_PASSIVE;
1103 * Slide the base index for initial allocations out of the
1104 * next zone we create so we do not over-weight the lower
1105 * part of the cpu memory caches.
1107 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
1108 & (ZALLOC_MAX_ZONE_SIZE - 1);
1112 * Ok, we have a zone from which at least one chunk is available.
1114 * Remove us from the ZoneAry[] when we become empty
1116 MASSERT(z->z_NFree > 0);
1118 if (--z->z_NFree == 0) {
1119 slgd->ZoneAry[zi] = z->z_Next;
1124 * Locate a chunk in a free page. This attempts to localize
1125 * reallocations into earlier pages without us having to sort
1126 * the chunk list. A chunk may still overlap a page boundary.
1128 while (z->z_FirstFreePg < ZonePageCount) {
1129 if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) {
1130 MASSERT((uintptr_t)chunk & ZoneMask);
1131 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next;
1138 * No chunks are available but NFree said we had some memory,
1139 * so it must be available in the never-before-used-memory
1140 * area governed by UIndex. The consequences are very
1141 * serious if our zone got corrupted so we use an explicit
1142 * panic rather then a KASSERT.
1144 chunk = (slchunk_t)(z->z_BasePtr + z->z_UIndex * size);
1146 if (++z->z_UIndex == z->z_NMax)
1148 if (z->z_UIndex == z->z_UEndIndex) {
1149 if (z->z_NFree != 0)
1150 _mpanic("slaballoc: corrupted zone");
1153 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
1154 flags &= ~SAFLAG_ZERO;
1155 flags |= SAFLAG_PASSIVE;
1160 if (flags & SAFLAG_ZERO)
1169 * Reallocate memory within the chunk
1172 _slabrealloc(void *ptr, size_t size)
1180 return(_slaballoc(size, 0));
1187 * Handle oversized allocations.
1189 if ((bigp = bigalloc_check_and_lock(ptr)) != NULL) {
1193 while ((big = *bigp) != NULL) {
1194 if (big->base == ptr) {
1195 size = (size + PAGE_MASK) & ~(size_t)PAGE_MASK;
1196 bigbytes = big->bytes;
1199 * If it already fits determine if it makes
1200 * sense to shrink/reallocate. Try to optimize
1201 * programs which stupidly make incremental
1202 * reallocations larger or smaller by scaling
1203 * the allocation. Also deal with potential
1206 if (size >= (bigbytes >> 1) &&
1208 if (big->active != size) {
1209 atomic_add_long(&excess_alloc,
1214 bigalloc_unlock(ptr);
1219 * For large reallocations, allocate more space
1220 * than we need to try to avoid excessive
1221 * reallocations later on.
1223 chunking = size + (size >> 3);
1224 chunking = (chunking + PAGE_MASK) &
1228 * Try to allocate adjacently in case the
1229 * program is idiotically realloc()ing a
1230 * huge memory block just slightly bigger.
1231 * (llvm's llc tends to do this a lot).
1233 * (MAP_TRYFIXED forces mmap to fail if there
1234 * is already something at the address).
1236 if (chunking > bigbytes) {
1238 int errno_save = errno;
1240 addr = mmap((char *)ptr + bigbytes,
1241 chunking - bigbytes,
1242 PROT_READ|PROT_WRITE,
1243 MAP_PRIVATE|MAP_ANON|
1247 if (addr == (char *)ptr + bigbytes) {
1248 atomic_add_long(&excess_alloc,
1253 big->bytes = chunking;
1255 bigalloc_unlock(ptr);
1259 MASSERT((void *)addr == MAP_FAILED);
1263 * Failed, unlink big and allocate fresh.
1264 * (note that we have to leave (big) intact
1265 * in case the slaballoc fails).
1268 bigalloc_unlock(ptr);
1269 if ((nptr = _slaballoc(size, 0)) == NULL) {
1271 bigp = bigalloc_lock(ptr);
1274 bigalloc_unlock(ptr);
1277 if (size > bigbytes)
1279 bcopy(ptr, nptr, size);
1280 atomic_add_long(&excess_alloc, big->active -
1282 _slabfree(ptr, FASTSLABREALLOC, &big);
1288 bigalloc_unlock(ptr);
1289 handle_excess_big();
1293 * Get the original allocation's zone. If the new request winds
1294 * up using the same chunk size we do not have to do anything.
1296 * NOTE: We don't have to lock the globaldata here, the fields we
1297 * access here will not change at least as long as we have control
1298 * over the allocation.
1300 z = (slzone_t)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
1301 MASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1304 * Use zoneindex() to chunk-align the new size, as long as the
1305 * new size is not too large.
1307 if (size < ZoneLimit) {
1308 zoneindex(&size, &chunking);
1309 if (z->z_ChunkSize == size) {
1315 * Allocate memory for the new request size and copy as appropriate.
1317 if ((nptr = _slaballoc(size, 0)) != NULL) {
1318 if (size > z->z_ChunkSize)
1319 size = z->z_ChunkSize;
1320 bcopy(ptr, nptr, size);
1321 _slabfree(ptr, 0, NULL);
1328 * free (SLAB ALLOCATOR)
1330 * Free a memory block previously allocated by malloc. Note that we do not
1331 * attempt to uplodate ks_loosememuse as MP races could prevent us from
1332 * checking memory limits in malloc.
1335 * FASTSLABREALLOC Fast call from realloc, *rbigp already
1341 _slabfree(void *ptr, int flags, bigalloc_t *rbigp)
1347 slglobaldata_t slgd;
1352 /* Fast realloc path for big allocations */
1353 if (flags & FASTSLABREALLOC) {
1355 goto fastslabrealloc;
1359 * Handle NULL frees and special 0-byte allocations
1365 * Handle oversized allocations.
1367 if ((bigp = bigalloc_check_and_lock(ptr)) != NULL) {
1368 while ((big = *bigp) != NULL) {
1369 if (big->base == ptr) {
1371 atomic_add_long(&excess_alloc, big->active -
1373 bigalloc_unlock(ptr);
1376 * Try to stash the block we are freeing,
1377 * potentially receiving another block in
1378 * return which must be freed.
1381 if (big->bytes <= BIGCACHE_LIMIT) {
1382 big = bigcache_find_free(big);
1386 ptr = big->base; /* reload */
1388 _slabfree(big, 0, NULL);
1389 _vmem_free(ptr, size);
1394 bigalloc_unlock(ptr);
1395 handle_excess_big();
1399 * Zone case. Figure out the zone based on the fact that it is
1402 z = (slzone_t)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
1403 MASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1405 size = z->z_ChunkSize;
1406 zi = z->z_ZoneIndex;
1408 if (g_malloc_flags & SAFLAG_ZERO)
1411 if (mtmagazine_free(zi, ptr) == 0)
1414 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT;
1416 slgd = &SLGlobalData;
1420 * Add this free non-zero'd chunk to a linked list for reuse, adjust
1423 chunk->c_Next = z->z_PageAry[pgno];
1424 z->z_PageAry[pgno] = chunk;
1425 if (z->z_FirstFreePg > pgno)
1426 z->z_FirstFreePg = pgno;
1429 * Bump the number of free chunks. If it becomes non-zero the zone
1430 * must be added back onto the appropriate list.
1432 if (z->z_NFree++ == 0) {
1433 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1434 slgd->ZoneAry[z->z_ZoneIndex] = z;
1438 * If the zone becomes totally free then release it.
1440 if (z->z_NFree == z->z_NMax) {
1443 pz = &slgd->ZoneAry[z->z_ZoneIndex];
1445 pz = &(*pz)->z_Next;
1450 /* slgd lock released */
1457 * Allocate and return a magazine. NULL is returned and *burst is adjusted
1458 * if the magazine is empty.
1460 static __inline void *
1461 magazine_alloc(struct magazine *mp, int *burst)
1467 if (MAGAZINE_NOTEMPTY(mp)) {
1468 obj = mp->objects[--mp->rounds];
1473 * Return burst factor to caller along with NULL
1475 if ((mp->flags & M_BURST) && (burst != NULL)) {
1476 *burst = mp->burst_factor;
1478 /* Reduce burst factor by NSCALE; if it hits 1, disable BURST */
1479 if ((mp->flags & M_BURST) && (mp->flags & M_BURST_EARLY) &&
1481 mp->burst_factor -= M_BURST_NSCALE;
1482 if (mp->burst_factor <= 1) {
1483 mp->burst_factor = 1;
1484 mp->flags &= ~(M_BURST);
1485 mp->flags &= ~(M_BURST_EARLY);
1492 magazine_free(struct magazine *mp, void *p)
1494 if (mp != NULL && MAGAZINE_NOTFULL(mp)) {
1495 mp->objects[mp->rounds++] = p;
1503 mtmagazine_alloc(int zi)
1506 struct magazine *mp, *emptymag;
1511 * Do not try to access per-thread magazines while the mtmagazine
1512 * is being initialized or destroyed.
1519 * Primary per-thread allocation loop
1523 * If the loaded magazine has rounds, allocate and return
1525 mp = tp->mags[zi].loaded;
1526 obj = magazine_alloc(mp, NULL);
1531 * If the prev magazine is full, swap with the loaded
1532 * magazine and retry.
1534 mp = tp->mags[zi].prev;
1535 if (mp && MAGAZINE_FULL(mp)) {
1536 MASSERT(mp->rounds != 0);
1537 swap_mags(&tp->mags[zi]); /* prev now empty */
1542 * Try to get a full magazine from the depot. Cycle
1543 * through depot(full)->loaded->prev->depot(empty).
1544 * Retry if a full magazine was available from the depot.
1546 * Return NULL (caller will fall through) if no magazines
1547 * can be found anywhere.
1551 emptymag = tp->mags[zi].prev;
1553 SLIST_INSERT_HEAD(&d->empty, emptymag, nextmagazine);
1554 tp->mags[zi].prev = tp->mags[zi].loaded;
1555 mp = SLIST_FIRST(&d->full); /* loaded magazine */
1556 tp->mags[zi].loaded = mp;
1558 SLIST_REMOVE_HEAD(&d->full, nextmagazine);
1559 MASSERT(MAGAZINE_NOTEMPTY(mp));
1571 mtmagazine_free(int zi, void *ptr)
1574 struct magazine *mp, *loadedmag;
1579 * Do not try to access per-thread magazines while the mtmagazine
1580 * is being initialized or destroyed.
1587 * Primary per-thread freeing loop
1591 * Make sure a new magazine is available in case we have
1592 * to use it. Staging the newmag allows us to avoid
1593 * some locking/reentrancy complexity.
1595 * Temporarily disable the per-thread caches for this
1596 * allocation to avoid reentrancy and/or to avoid a
1597 * stack overflow if the [zi] happens to be the same that
1598 * would be used to allocate the new magazine.
1600 if (tp->newmag == NULL) {
1602 tp->newmag = _slaballoc(sizeof(struct magazine),
1605 if (tp->newmag == NULL) {
1612 * If the loaded magazine has space, free directly to it
1614 rc = magazine_free(tp->mags[zi].loaded, ptr);
1619 * If the prev magazine is empty, swap with the loaded
1620 * magazine and retry.
1622 mp = tp->mags[zi].prev;
1623 if (mp && MAGAZINE_EMPTY(mp)) {
1624 MASSERT(mp->rounds == 0);
1625 swap_mags(&tp->mags[zi]); /* prev now full */
1630 * Try to get an empty magazine from the depot. Cycle
1631 * through depot(empty)->loaded->prev->depot(full).
1632 * Retry if an empty magazine was available from the depot.
1637 if ((loadedmag = tp->mags[zi].prev) != NULL)
1638 SLIST_INSERT_HEAD(&d->full, loadedmag, nextmagazine);
1639 tp->mags[zi].prev = tp->mags[zi].loaded;
1640 mp = SLIST_FIRST(&d->empty);
1642 tp->mags[zi].loaded = mp;
1643 SLIST_REMOVE_HEAD(&d->empty, nextmagazine);
1644 MASSERT(MAGAZINE_NOTFULL(mp));
1648 mp->capacity = M_MAX_ROUNDS;
1651 tp->mags[zi].loaded = mp;
1660 mtmagazine_init(void)
1664 error = pthread_key_create(&thread_mags_key, mtmagazine_destructor);
1670 * This function is only used by the thread exit destructor
1673 mtmagazine_drain(struct magazine *mp)
1677 while (MAGAZINE_NOTEMPTY(mp)) {
1678 obj = magazine_alloc(mp, NULL);
1679 _slabfree(obj, 0, NULL);
1684 * mtmagazine_destructor()
1686 * When a thread exits, we reclaim all its resources; all its magazines are
1687 * drained and the structures are freed.
1689 * WARNING! The destructor can be called multiple times if the larger user
1690 * program has its own destructors which run after ours which
1691 * allocate or free memory.
1694 mtmagazine_destructor(void *thrp)
1696 thr_mags *tp = thrp;
1697 struct magazine *mp;
1701 * Prevent further use of mtmagazines while we are destructing
1702 * them, as well as for any destructors which are run after us
1703 * prior to the thread actually being destroyed.
1707 for (i = 0; i < NZONES; i++) {
1708 mp = tp->mags[i].loaded;
1709 tp->mags[i].loaded = NULL;
1711 if (MAGAZINE_NOTEMPTY(mp))
1712 mtmagazine_drain(mp);
1713 _slabfree(mp, 0, NULL);
1716 mp = tp->mags[i].prev;
1717 tp->mags[i].prev = NULL;
1719 if (MAGAZINE_NOTEMPTY(mp))
1720 mtmagazine_drain(mp);
1721 _slabfree(mp, 0, NULL);
1728 _slabfree(mp, 0, NULL);
1735 * Attempt to allocate a zone from the zone magazine; the zone magazine has
1736 * M_BURST_EARLY enabled, so honor the burst request from the magazine.
1739 zone_alloc(int flags)
1741 slglobaldata_t slgd = &SLGlobalData;
1746 zone_magazine_lock();
1749 z = magazine_alloc(&zone_magazine, &burst);
1750 if (z == NULL && burst == 1) {
1751 zone_magazine_unlock();
1752 z = _vmem_alloc(ZoneSize * burst, ZoneSize, flags);
1753 } else if (z == NULL) {
1754 z = _vmem_alloc(ZoneSize * burst, ZoneSize, flags);
1756 for (i = 1; i < burst; i++) {
1757 j = magazine_free(&zone_magazine,
1758 (char *) z + (ZoneSize * i));
1762 zone_magazine_unlock();
1764 z->z_Flags |= SLZF_UNOTZEROD;
1765 zone_magazine_unlock();
1774 * Release a zone and unlock the slgd lock.
1779 slglobaldata_t slgd = &SLGlobalData;
1780 void *excess[M_ZONE_ROUNDS - M_LOW_ROUNDS] = {};
1783 zone_magazine_lock();
1786 bzero(z, sizeof(struct slzone));
1789 madvise(z, ZoneSize, MADV_FREE);
1791 i = magazine_free(&zone_magazine, z);
1794 * If we failed to free, collect excess magazines; release the zone
1795 * magazine lock, and then free to the system via _vmem_free. Re-enable
1796 * BURST mode for the magazine.
1799 j = zone_magazine.rounds - zone_magazine.low_factor;
1800 for (i = 0; i < j; i++) {
1801 excess[i] = magazine_alloc(&zone_magazine, NULL);
1802 MASSERT(excess[i] != NULL);
1805 zone_magazine_unlock();
1807 for (i = 0; i < j; i++)
1808 _vmem_free(excess[i], ZoneSize);
1810 _vmem_free(z, ZoneSize);
1812 zone_magazine_unlock();
1819 * Directly map memory in PAGE_SIZE'd chunks with the specified
1822 * Alignment must be a multiple of PAGE_SIZE.
1824 * Size must be >= alignment.
1827 _vmem_alloc(size_t size, size_t align, int flags)
1834 * Map anonymous private memory.
1836 addr = mmap(NULL, size, PROT_READ|PROT_WRITE,
1837 MAP_PRIVATE|MAP_ANON, -1, 0);
1838 if (addr == MAP_FAILED)
1842 * Check alignment. The misaligned offset is also the excess
1843 * amount. If misaligned unmap the excess so we have a chance of
1844 * mapping at the next alignment point and recursively try again.
1846 * BBBBBBBBBBB BBBBBBBBBBB BBBBBBBBBBB block alignment
1847 * aaaaaaaaa aaaaaaaaaaa aa mis-aligned allocation
1848 * xxxxxxxxx final excess calculation
1849 * ^ returned address
1851 excess = (uintptr_t)addr & (align - 1);
1854 excess = align - excess;
1857 munmap(save + excess, size - excess);
1858 addr = _vmem_alloc(size, align, flags);
1859 munmap(save, excess);
1861 return((void *)addr);
1867 * Free a chunk of memory allocated with _vmem_alloc()
1870 _vmem_free(void *ptr, size_t size)
1876 * Panic on fatal conditions
1879 _mpanic(const char *ctl, ...)
1883 if (malloc_panic == 0) {
1886 vfprintf(stderr, ctl, va);
1887 fprintf(stderr, "\n");
1894 __weak_reference(__malloc, malloc);
1895 __weak_reference(__calloc, calloc);
1896 __weak_reference(__posix_memalign, posix_memalign);
1897 __weak_reference(__realloc, realloc);
1898 __weak_reference(__free, free);