4 * KERN_SLABALLOC.C - Kernel SLAB memory allocator
6 * Copyright (c) 2003,2004,2010 The DragonFly Project. All rights reserved.
8 * This code is derived from software contributed to The DragonFly Project
9 * by Matthew Dillon <dillon@backplane.com>
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in
19 * the documentation and/or other materials provided with the
21 * 3. Neither the name of The DragonFly Project nor the names of its
22 * contributors may be used to endorse or promote products derived
23 * from this software without specific, prior written permission.
25 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
26 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
27 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
28 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
29 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
30 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
31 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
32 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
33 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
34 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
35 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * This module implements a slab allocator drop-in replacement for the
41 * A slab allocator reserves a ZONE for each chunk size, then lays the
42 * chunks out in an array within the zone. Allocation and deallocation
43 * is nearly instantanious, and fragmentation/overhead losses are limited
44 * to a fixed worst-case amount.
46 * The downside of this slab implementation is in the chunk size
47 * multiplied by the number of zones. ~80 zones * 128K = 10MB of VM per cpu.
48 * In a kernel implementation all this memory will be physical so
49 * the zone size is adjusted downward on machines with less physical
50 * memory. The upside is that overhead is bounded... this is the *worst*
53 * Slab management is done on a per-cpu basis and no locking or mutexes
54 * are required, only a critical section. When one cpu frees memory
55 * belonging to another cpu's slab manager an asynchronous IPI message
56 * will be queued to execute the operation. In addition, both the
57 * high level slab allocator and the low level zone allocator optimize
58 * M_ZERO requests, and the slab allocator does not have to pre initialize
59 * the linked list of chunks.
61 * XXX Balancing is needed between cpus. Balance will be handled through
62 * asynchronous IPIs primarily by reassigning the z_Cpu ownership of chunks.
64 * XXX If we have to allocate a new zone and M_USE_RESERVE is set, use of
65 * the new zone should be restricted to M_USE_RESERVE requests only.
67 * Alloc Size Chunking Number of zones
77 * (if PAGE_SIZE is 4K the maximum zone allocation is 16383)
79 * Allocations >= ZoneLimit go directly to kmem.
81 * API REQUIREMENTS AND SIDE EFFECTS
83 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we
84 * have remained compatible with the following API requirements:
86 * + small power-of-2 sized allocations are power-of-2 aligned (kern_tty)
87 * + all power-of-2 sized allocations are power-of-2 aligned (twe)
88 * + malloc(0) is allowed and returns non-NULL (ahc driver)
89 * + ability to allocate arbitrarily large chunks of memory
94 #include <sys/param.h>
95 #include <sys/systm.h>
96 #include <sys/kernel.h>
97 #include <sys/slaballoc.h>
99 #include <sys/vmmeter.h>
100 #include <sys/lock.h>
101 #include <sys/thread.h>
102 #include <sys/globaldata.h>
103 #include <sys/sysctl.h>
107 #include <vm/vm_param.h>
108 #include <vm/vm_kern.h>
109 #include <vm/vm_extern.h>
110 #include <vm/vm_object.h>
112 #include <vm/vm_map.h>
113 #include <vm/vm_page.h>
114 #include <vm/vm_pageout.h>
116 #include <machine/cpu.h>
118 #include <sys/thread2.h>
120 #define btokup(z) (&pmap_kvtom((vm_offset_t)(z))->ku_pagecnt)
122 #define MEMORY_STRING "ptr=%p type=%p size=%d flags=%04x"
123 #define MEMORY_ARG_SIZE (sizeof(void *) * 2 + sizeof(unsigned long) + \
126 #if !defined(KTR_MEMORY)
127 #define KTR_MEMORY KTR_ALL
129 KTR_INFO_MASTER(memory);
130 KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin", 0);
131 KTR_INFO(KTR_MEMORY, memory, malloc_end, 1, MEMORY_STRING, MEMORY_ARG_SIZE);
132 KTR_INFO(KTR_MEMORY, memory, free_zero, 2, MEMORY_STRING, MEMORY_ARG_SIZE);
133 KTR_INFO(KTR_MEMORY, memory, free_ovsz, 3, MEMORY_STRING, MEMORY_ARG_SIZE);
134 KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 4, MEMORY_STRING, MEMORY_ARG_SIZE);
135 KTR_INFO(KTR_MEMORY, memory, free_chunk, 5, MEMORY_STRING, MEMORY_ARG_SIZE);
137 KTR_INFO(KTR_MEMORY, memory, free_request, 6, MEMORY_STRING, MEMORY_ARG_SIZE);
138 KTR_INFO(KTR_MEMORY, memory, free_rem_beg, 7, MEMORY_STRING, MEMORY_ARG_SIZE);
139 KTR_INFO(KTR_MEMORY, memory, free_rem_end, 8, MEMORY_STRING, MEMORY_ARG_SIZE);
141 KTR_INFO(KTR_MEMORY, memory, free_beg, 9, "free begin", 0);
142 KTR_INFO(KTR_MEMORY, memory, free_end, 10, "free end", 0);
144 #define logmemory(name, ptr, type, size, flags) \
145 KTR_LOG(memory_ ## name, ptr, type, size, flags)
146 #define logmemory_quick(name) \
147 KTR_LOG(memory_ ## name)
150 * Fixed globals (not per-cpu)
153 static int ZoneLimit;
154 static int ZonePageCount;
155 static uintptr_t ZoneMask;
156 static int ZoneBigAlloc; /* in KB */
157 static int ZoneGenAlloc; /* in KB */
158 struct malloc_type *kmemstatistics; /* exported to vmstat */
159 static int32_t weirdary[16];
161 static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
162 static void kmem_slab_free(void *ptr, vm_size_t bytes);
164 #if defined(INVARIANTS)
165 static void chunk_mark_allocated(SLZone *z, void *chunk);
166 static void chunk_mark_free(SLZone *z, void *chunk);
168 #define chunk_mark_allocated(z, chunk)
169 #define chunk_mark_free(z, chunk)
173 * Misc constants. Note that allocations that are exact multiples of
174 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
175 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
177 #define MIN_CHUNK_SIZE 8 /* in bytes */
178 #define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
179 #define ZONE_RELS_THRESH 32 /* threshold number of zones */
180 #define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
183 * The WEIRD_ADDR is used as known text to copy into free objects to
184 * try to create deterministic failure cases if the data is accessed after
187 #define WEIRD_ADDR 0xdeadc0de
188 #define MAX_COPY sizeof(weirdary)
189 #define ZERO_LENGTH_PTR ((void *)-8)
192 * Misc global malloc buckets
195 MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
196 MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
197 MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
199 MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
200 MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
203 * Initialize the slab memory allocator. We have to choose a zone size based
204 * on available physical memory. We choose a zone side which is approximately
205 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
206 * 128K. The zone size is limited to the bounds set in slaballoc.h
207 * (typically 32K min, 128K max).
209 static void kmeminit(void *dummy);
213 SYSINIT(kmem, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL)
217 * If enabled any memory allocated without M_ZERO is initialized to -1.
219 static int use_malloc_pattern;
220 SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
221 &use_malloc_pattern, 0,
222 "Initialize memory to -1 if M_ZERO not specified");
225 static int ZoneRelsThresh = ZONE_RELS_THRESH;
226 SYSCTL_INT(_kern, OID_AUTO, zone_big_alloc, CTLFLAG_RD, &ZoneBigAlloc, 0, "");
227 SYSCTL_INT(_kern, OID_AUTO, zone_gen_alloc, CTLFLAG_RD, &ZoneGenAlloc, 0, "");
228 SYSCTL_INT(_kern, OID_AUTO, zone_cache, CTLFLAG_RW, &ZoneRelsThresh, 0, "");
231 kmeminit(void *dummy)
237 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
238 if (limsize > KvaSize)
241 usesize = (int)(limsize / 1024); /* convert to KB */
243 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
244 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
246 ZoneLimit = ZoneSize / 4;
247 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
248 ZoneLimit = ZALLOC_ZONE_LIMIT;
249 ZoneMask = ~(uintptr_t)(ZoneSize - 1);
250 ZonePageCount = ZoneSize / PAGE_SIZE;
252 for (i = 0; i < NELEM(weirdary); ++i)
253 weirdary[i] = WEIRD_ADDR;
255 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
258 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
262 * Initialize a malloc type tracking structure.
265 malloc_init(void *data)
267 struct malloc_type *type = data;
270 if (type->ks_magic != M_MAGIC)
271 panic("malloc type lacks magic");
273 if (type->ks_limit != 0)
276 if (vmstats.v_page_count == 0)
277 panic("malloc_init not allowed before vm init");
279 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
280 if (limsize > KvaSize)
282 type->ks_limit = limsize / 10;
284 type->ks_next = kmemstatistics;
285 kmemstatistics = type;
289 malloc_uninit(void *data)
291 struct malloc_type *type = data;
292 struct malloc_type *t;
298 if (type->ks_magic != M_MAGIC)
299 panic("malloc type lacks magic");
301 if (vmstats.v_page_count == 0)
302 panic("malloc_uninit not allowed before vm init");
304 if (type->ks_limit == 0)
305 panic("malloc_uninit on uninitialized type");
308 /* Make sure that all pending kfree()s are finished. */
309 lwkt_synchronize_ipiqs("muninit");
314 * memuse is only correct in aggregation. Due to memory being allocated
315 * on one cpu and freed on another individual array entries may be
316 * negative or positive (canceling each other out).
318 for (i = ttl = 0; i < ncpus; ++i)
319 ttl += type->ks_memuse[i];
321 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
322 ttl, type->ks_shortdesc, i);
325 if (type == kmemstatistics) {
326 kmemstatistics = type->ks_next;
328 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
329 if (t->ks_next == type) {
330 t->ks_next = type->ks_next;
335 type->ks_next = NULL;
340 * Increase the kmalloc pool limit for the specified pool. No changes
341 * are the made if the pool would shrink.
344 kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
346 if (type->ks_limit == 0)
350 if (type->ks_limit < bytes)
351 type->ks_limit = bytes;
355 * Dynamically create a malloc pool. This function is a NOP if *typep is
359 kmalloc_create(struct malloc_type **typep, const char *descr)
361 struct malloc_type *type;
363 if (*typep == NULL) {
364 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
365 type->ks_magic = M_MAGIC;
366 type->ks_shortdesc = descr;
373 * Destroy a dynamically created malloc pool. This function is a NOP if
374 * the pool has already been destroyed.
377 kmalloc_destroy(struct malloc_type **typep)
379 if (*typep != NULL) {
380 malloc_uninit(*typep);
381 kfree(*typep, M_TEMP);
387 * Calculate the zone index for the allocation request size and set the
388 * allocation request size to that particular zone's chunk size.
391 zoneindex(unsigned long *bytes)
393 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
395 *bytes = n = (n + 7) & ~7;
396 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
399 *bytes = n = (n + 15) & ~15;
404 *bytes = n = (n + 31) & ~31;
408 *bytes = n = (n + 63) & ~63;
412 *bytes = n = (n + 127) & ~127;
413 return(n / 128 + 31);
416 *bytes = n = (n + 255) & ~255;
417 return(n / 256 + 39);
419 *bytes = n = (n + 511) & ~511;
420 return(n / 512 + 47);
422 #if ZALLOC_ZONE_LIMIT > 8192
424 *bytes = n = (n + 1023) & ~1023;
425 return(n / 1024 + 55);
428 #if ZALLOC_ZONE_LIMIT > 16384
430 *bytes = n = (n + 2047) & ~2047;
431 return(n / 2048 + 63);
434 panic("Unexpected byte count %d", n);
440 * Used to debug memory corruption issues. Record up to (typically 32)
441 * allocation sources for this zone (for a particular chunk size).
445 slab_record_source(SLZone *z, const char *file, int line)
448 int b = line & (SLAB_DEBUG_ENTRIES - 1);
452 if (z->z_Sources[i].file == file && z->z_Sources[i].line == line)
454 if (z->z_Sources[i].file == NULL)
456 i = (i + 1) & (SLAB_DEBUG_ENTRIES - 1);
458 z->z_Sources[i].file = file;
459 z->z_Sources[i].line = line;
465 * kmalloc() (SLAB ALLOCATOR)
467 * Allocate memory via the slab allocator. If the request is too large,
468 * or if it page-aligned beyond a certain size, we fall back to the
469 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
470 * &SlabMisc if you don't care.
472 * M_RNOWAIT - don't block.
473 * M_NULLOK - return NULL instead of blocking.
474 * M_ZERO - zero the returned memory.
475 * M_USE_RESERVE - allow greater drawdown of the free list
476 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
483 kmalloc_debug(unsigned long size, struct malloc_type *type, int flags,
484 const char *file, int line)
487 kmalloc(unsigned long size, struct malloc_type *type, int flags)
496 struct globaldata *gd;
502 logmemory_quick(malloc_beg);
507 * XXX silly to have this in the critical path.
509 if (type->ks_limit == 0) {
511 if (type->ks_limit == 0)
518 * Handle the case where the limit is reached. Panic if we can't return
519 * NULL. The original malloc code looped, but this tended to
520 * simply deadlock the computer.
522 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
523 * to determine if a more complete limit check should be done. The
524 * actual memory use is tracked via ks_memuse[cpu].
526 while (type->ks_loosememuse >= type->ks_limit) {
530 for (i = ttl = 0; i < ncpus; ++i)
531 ttl += type->ks_memuse[i];
532 type->ks_loosememuse = ttl; /* not MP synchronized */
533 if ((ssize_t)ttl < 0) /* deal with occassional race */
535 if (ttl >= type->ks_limit) {
536 if (flags & M_NULLOK) {
537 logmemory(malloc_end, NULL, type, size, flags);
540 panic("%s: malloc limit exceeded", type->ks_shortdesc);
545 * Handle the degenerate size == 0 case. Yes, this does happen.
546 * Return a special pointer. This is to maintain compatibility with
547 * the original malloc implementation. Certain devices, such as the
548 * adaptec driver, not only allocate 0 bytes, they check for NULL and
549 * also realloc() later on. Joy.
552 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
553 return(ZERO_LENGTH_PTR);
557 * Handle hysteresis from prior frees here in malloc(). We cannot
558 * safely manipulate the kernel_map in free() due to free() possibly
559 * being called via an IPI message or from sensitive interrupt code.
561 * NOTE: ku_pagecnt must be cleared before we free the slab or we
562 * might race another cpu allocating the kva and setting
565 while (slgd->NFreeZones > ZoneRelsThresh && (flags & M_RNOWAIT) == 0) {
567 if (slgd->NFreeZones > ZoneRelsThresh) { /* crit sect race */
571 slgd->FreeZones = z->z_Next;
575 kmem_slab_free(z, ZoneSize); /* may block */
576 atomic_add_int(&ZoneGenAlloc, -(int)ZoneSize / 1024);
582 * XXX handle oversized frees that were queued from kfree().
584 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) {
586 if ((z = slgd->FreeOvZones) != NULL) {
589 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
590 slgd->FreeOvZones = z->z_Next;
591 tsize = z->z_ChunkSize;
592 kmem_slab_free(z, tsize); /* may block */
593 atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024);
599 * Handle large allocations directly. There should not be very many of
600 * these so performance is not a big issue.
602 * The backend allocator is pretty nasty on a SMP system. Use the
603 * slab allocator for one and two page-sized chunks even though we lose
604 * some efficiency. XXX maybe fix mmio and the elf loader instead.
606 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
609 size = round_page(size);
610 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
612 logmemory(malloc_end, NULL, type, size, flags);
615 atomic_add_int(&ZoneBigAlloc, (int)size / 1024);
616 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
617 flags |= M_PASSIVE_ZERO;
619 *kup = size / PAGE_SIZE;
625 * Attempt to allocate out of an existing zone. First try the free list,
626 * then allocate out of unallocated space. If we find a good zone move
627 * it to the head of the list so later allocations find it quickly
628 * (we might have thousands of zones in the list).
630 * Note: zoneindex() will panic of size is too large.
632 zi = zoneindex(&size);
633 KKASSERT(zi < NZONES);
636 if ((z = slgd->ZoneAry[zi]) != NULL) {
638 * Locate a chunk - we have to have at least one. If this is the
639 * last chunk go ahead and do the work to retrieve chunks freed
640 * from remote cpus, and if the zone is still empty move it off
643 if (--z->z_NFree <= 0) {
644 KKASSERT(z->z_NFree == 0);
648 * WARNING! This code competes with other cpus. It is ok
649 * for us to not drain RChunks here but we might as well, and
650 * it is ok if more accumulate after we're done.
652 * Set RSignal before pulling rchunks off, indicating that we
653 * will be moving ourselves off of the ZoneAry. Remote ends will
654 * read RSignal before putting rchunks on thus interlocking
655 * their IPI signaling.
657 if (z->z_RChunks == NULL)
658 atomic_swap_int(&z->z_RSignal, 1);
660 while ((bchunk = z->z_RChunks) != NULL) {
662 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
663 *z->z_LChunksp = bchunk;
665 chunk_mark_free(z, bchunk);
666 z->z_LChunksp = &bchunk->c_Next;
667 bchunk = bchunk->c_Next;
675 * Remove from the zone list if no free chunks remain.
678 if (z->z_NFree == 0) {
679 slgd->ZoneAry[zi] = z->z_Next;
687 * Fast path, we have chunks available in z_LChunks.
689 chunk = z->z_LChunks;
691 chunk_mark_allocated(z, chunk);
692 z->z_LChunks = chunk->c_Next;
693 if (z->z_LChunks == NULL)
694 z->z_LChunksp = &z->z_LChunks;
696 slab_record_source(z, file, line);
702 * No chunks are available in LChunks, the free chunk MUST be
703 * in the never-before-used memory area, controlled by UIndex.
705 * The consequences are very serious if our zone got corrupted so
706 * we use an explicit panic rather than a KASSERT.
708 if (z->z_UIndex + 1 != z->z_NMax)
713 if (z->z_UIndex == z->z_UEndIndex)
714 panic("slaballoc: corrupted zone");
716 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
717 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
719 flags |= M_PASSIVE_ZERO;
721 chunk_mark_allocated(z, chunk);
723 slab_record_source(z, file, line);
729 * If all zones are exhausted we need to allocate a new zone for this
730 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
731 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
732 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
733 * we do not pre-zero it because we do not want to mess up the L1 cache.
735 * At least one subsystem, the tty code (see CROUND) expects power-of-2
736 * allocations to be power-of-2 aligned. We maintain compatibility by
737 * adjusting the base offset below.
743 if ((z = slgd->FreeZones) != NULL) {
744 slgd->FreeZones = z->z_Next;
746 bzero(z, sizeof(SLZone));
747 z->z_Flags |= SLZF_UNOTZEROD;
749 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
752 atomic_add_int(&ZoneGenAlloc, (int)ZoneSize / 1024);
756 * How big is the base structure?
758 #if defined(INVARIANTS)
760 * Make room for z_Bitmap. An exact calculation is somewhat more
761 * complicated so don't make an exact calculation.
763 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
764 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
766 off = sizeof(SLZone);
770 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
771 * Otherwise just 8-byte align the data.
773 if ((size | (size - 1)) + 1 == (size << 1))
774 off = (off + size - 1) & ~(size - 1);
776 off = (off + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
777 z->z_Magic = ZALLOC_SLAB_MAGIC;
779 z->z_NMax = (ZoneSize - off) / size;
780 z->z_NFree = z->z_NMax - 1;
781 z->z_BasePtr = (char *)z + off;
782 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
783 z->z_ChunkSize = size;
785 z->z_Cpu = gd->gd_cpuid;
786 z->z_LChunksp = &z->z_LChunks;
788 bcopy(z->z_Sources, z->z_AltSources, sizeof(z->z_Sources));
789 bzero(z->z_Sources, sizeof(z->z_Sources));
791 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
792 z->z_Next = slgd->ZoneAry[zi];
793 slgd->ZoneAry[zi] = z;
794 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
795 flags &= ~M_ZERO; /* already zero'd */
796 flags |= M_PASSIVE_ZERO;
799 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */
800 chunk_mark_allocated(z, chunk);
802 slab_record_source(z, file, line);
806 * Slide the base index for initial allocations out of the next
807 * zone we create so we do not over-weight the lower part of the
810 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
811 & (ZALLOC_MAX_ZONE_SIZE - 1);
815 ++type->ks_inuse[gd->gd_cpuid];
816 type->ks_memuse[gd->gd_cpuid] += size;
817 type->ks_loosememuse += size; /* not MP synchronized */
823 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
824 if (use_malloc_pattern) {
825 for (i = 0; i < size; i += sizeof(int)) {
826 *(int *)((char *)chunk + i) = -1;
829 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
832 logmemory(malloc_end, chunk, type, size, flags);
836 logmemory(malloc_end, NULL, type, size, flags);
841 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
843 * Generally speaking this routine is not called very often and we do
844 * not attempt to optimize it beyond reusing the same pointer if the
845 * new size fits within the chunking of the old pointer's zone.
849 krealloc_debug(void *ptr, unsigned long size,
850 struct malloc_type *type, int flags,
851 const char *file, int line)
854 krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
862 KKASSERT((flags & M_ZERO) == 0); /* not supported */
864 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
865 return(kmalloc_debug(size, type, flags, file, line));
872 * Handle oversized allocations. XXX we really should require that a
873 * size be passed to free() instead of this nonsense.
877 osize = *kup << PAGE_SHIFT;
878 if (osize == round_page(size))
880 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
882 bcopy(ptr, nptr, min(size, osize));
888 * Get the original allocation's zone. If the new request winds up
889 * using the same chunk size we do not have to do anything.
891 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
894 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
897 * Allocate memory for the new request size. Note that zoneindex has
898 * already adjusted the request size to the appropriate chunk size, which
899 * should optimize our bcopy(). Then copy and return the new pointer.
901 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
902 * necessary align the result.
904 * We can only zoneindex (to align size to the chunk size) if the new
905 * size is not too large.
907 if (size < ZoneLimit) {
909 if (z->z_ChunkSize == size)
912 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
914 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
920 * Return the kmalloc limit for this type, in bytes.
923 kmalloc_limit(struct malloc_type *type)
925 if (type->ks_limit == 0) {
927 if (type->ks_limit == 0)
931 return(type->ks_limit);
935 * Allocate a copy of the specified string.
937 * (MP SAFE) (MAY BLOCK)
941 kstrdup_debug(const char *str, struct malloc_type *type,
942 const char *file, int line)
945 kstrdup(const char *str, struct malloc_type *type)
948 int zlen; /* length inclusive of terminating NUL */
953 zlen = strlen(str) + 1;
954 nstr = kmalloc_debug(zlen, type, M_WAITOK, file, line);
955 bcopy(str, nstr, zlen);
961 * Notify our cpu that a remote cpu has freed some chunks in a zone that
962 * we own. RCount will be bumped so the memory should be good, but validate
967 kfree_remote(void *ptr)
975 slgd = &mycpu->gd_slab;
978 KKASSERT(*kup == -((int)mycpuid + 1));
979 KKASSERT(z->z_RCount > 0);
980 atomic_subtract_int(&z->z_RCount, 1);
982 logmemory(free_rem_beg, z, NULL, 0, 0);
983 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
984 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
988 * Indicate that we will no longer be off of the ZoneAry by
995 * Atomically extract the bchunks list and then process it back
996 * into the lchunks list. We want to append our bchunks to the
997 * lchunks list and not prepend since we likely do not have
998 * cache mastership of the related data (not that it helps since
999 * we are using c_Next).
1001 while ((bchunk = z->z_RChunks) != NULL) {
1003 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
1004 *z->z_LChunksp = bchunk;
1006 chunk_mark_free(z, bchunk);
1007 z->z_LChunksp = &bchunk->c_Next;
1008 bchunk = bchunk->c_Next;
1014 if (z->z_NFree && nfree == 0) {
1015 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1016 slgd->ZoneAry[z->z_ZoneIndex] = z;
1020 * If the zone becomes totally free, and there are other zones we
1021 * can allocate from, move this zone to the FreeZones list. Since
1022 * this code can be called from an IPI callback, do *NOT* try to mess
1023 * with kernel_map here. Hysteresis will be performed at malloc() time.
1025 * Do not move the zone if there is an IPI inflight, otherwise MP
1026 * races can result in our free_remote code accessing a destroyed
1029 if (z->z_NFree == z->z_NMax &&
1030 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1036 for (pz = &slgd->ZoneAry[z->z_ZoneIndex];
1038 pz = &(*pz)->z_Next) {
1043 z->z_Next = slgd->FreeZones;
1044 slgd->FreeZones = z;
1049 logmemory(free_rem_end, z, bchunk, 0, 0);
1055 * free (SLAB ALLOCATOR)
1057 * Free a memory block previously allocated by malloc. Note that we do not
1058 * attempt to update ks_loosememuse as MP races could prevent us from
1059 * checking memory limits in malloc.
1064 kfree(void *ptr, struct malloc_type *type)
1069 struct globaldata *gd;
1077 logmemory_quick(free_beg);
1079 slgd = &gd->gd_slab;
1082 panic("trying to free NULL pointer");
1085 * Handle special 0-byte allocations
1087 if (ptr == ZERO_LENGTH_PTR) {
1088 logmemory(free_zero, ptr, type, -1, 0);
1089 logmemory_quick(free_end);
1094 * Panic on bad malloc type
1096 if (type->ks_magic != M_MAGIC)
1097 panic("free: malloc type lacks magic");
1100 * Handle oversized allocations. XXX we really should require that a
1101 * size be passed to free() instead of this nonsense.
1103 * This code is never called via an ipi.
1107 size = *kup << PAGE_SHIFT;
1110 KKASSERT(sizeof(weirdary) <= size);
1111 bcopy(weirdary, ptr, sizeof(weirdary));
1114 * NOTE: For oversized allocations we do not record the
1115 * originating cpu. It gets freed on the cpu calling
1116 * kfree(). The statistics are in aggregate.
1118 * note: XXX we have still inherited the interrupts-can't-block
1119 * assumption. An interrupt thread does not bump
1120 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1121 * primarily until we can fix softupdate's assumptions about free().
1124 --type->ks_inuse[gd->gd_cpuid];
1125 type->ks_memuse[gd->gd_cpuid] -= size;
1126 if (mycpu->gd_intr_nesting_level ||
1127 (gd->gd_curthread->td_flags & TDF_INTTHREAD))
1129 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1131 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1132 z->z_Next = slgd->FreeOvZones;
1133 z->z_ChunkSize = size;
1134 slgd->FreeOvZones = z;
1138 logmemory(free_ovsz, ptr, type, size, 0);
1139 kmem_slab_free(ptr, size); /* may block */
1140 atomic_add_int(&ZoneBigAlloc, -(int)size / 1024);
1142 logmemory_quick(free_end);
1147 * Zone case. Figure out the zone based on the fact that it is
1150 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1153 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1156 * If we do not own the zone then use atomic ops to free to the
1157 * remote cpu linked list and notify the target zone using a
1160 * The target zone cannot be deallocated while we own a chunk of it,
1161 * so the zone header's storage is stable until the very moment
1162 * we adjust z_RChunks. After that we cannot safely dereference (z).
1164 * (no critical section needed)
1166 if (z->z_CpuGd != gd) {
1169 * Making these adjustments now allow us to avoid passing (type)
1170 * to the remote cpu. Note that ks_inuse/ks_memuse is being
1171 * adjusted on OUR cpu, not the zone cpu, but it should all still
1172 * sum up properly and cancel out.
1175 --type->ks_inuse[gd->gd_cpuid];
1176 type->ks_memuse[gd->gd_cpuid] -= z->z_ChunkSize;
1180 * WARNING! This code competes with other cpus. Once we
1181 * successfully link the chunk to RChunks the remote
1182 * cpu can rip z's storage out from under us.
1184 * Bumping RCount prevents z's storage from getting
1187 rsignal = z->z_RSignal;
1190 atomic_add_int(&z->z_RCount, 1);
1194 bchunk = z->z_RChunks;
1196 chunk->c_Next = bchunk;
1199 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1204 * We have to signal the remote cpu if our actions will cause
1205 * the remote zone to be placed back on ZoneAry so it can
1206 * move the zone back on.
1208 * We only need to deal with NULL->non-NULL RChunk transitions
1209 * and only if z_RSignal is set. We interlock by reading rsignal
1210 * before adding our chunk to RChunks. This should result in
1211 * virtually no IPI traffic.
1213 * We can use a passive IPI to reduce overhead even further.
1215 if (bchunk == NULL && rsignal) {
1216 logmemory(free_request, ptr, type, z->z_ChunkSize, 0);
1217 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
1218 /* z can get ripped out from under us from this point on */
1219 } else if (rsignal) {
1220 atomic_subtract_int(&z->z_RCount, 1);
1221 /* z can get ripped out from under us from this point on */
1224 panic("Corrupt SLZone");
1226 logmemory_quick(free_end);
1233 logmemory(free_chunk, ptr, type, z->z_ChunkSize, 0);
1237 chunk_mark_free(z, chunk);
1240 * Put weird data into the memory to detect modifications after freeing,
1241 * illegal pointer use after freeing (we should fault on the odd address),
1242 * and so forth. XXX needs more work, see the old malloc code.
1245 if (z->z_ChunkSize < sizeof(weirdary))
1246 bcopy(weirdary, chunk, z->z_ChunkSize);
1248 bcopy(weirdary, chunk, sizeof(weirdary));
1252 * Add this free non-zero'd chunk to a linked list for reuse. Add
1253 * to the front of the linked list so it is more likely to be
1254 * reallocated, since it is already in our L1 cache.
1257 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
1258 panic("BADFREE %p", chunk);
1260 chunk->c_Next = z->z_LChunks;
1261 z->z_LChunks = chunk;
1262 if (chunk->c_Next == NULL)
1263 z->z_LChunksp = &chunk->c_Next;
1266 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
1271 * Bump the number of free chunks. If it becomes non-zero the zone
1272 * must be added back onto the appropriate list.
1274 if (z->z_NFree++ == 0) {
1275 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1276 slgd->ZoneAry[z->z_ZoneIndex] = z;
1279 --type->ks_inuse[z->z_Cpu];
1280 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
1283 * If the zone becomes totally free, and there are other zones we
1284 * can allocate from, move this zone to the FreeZones list. Since
1285 * this code can be called from an IPI callback, do *NOT* try to mess
1286 * with kernel_map here. Hysteresis will be performed at malloc() time.
1288 if (z->z_NFree == z->z_NMax &&
1289 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1295 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
1299 z->z_Next = slgd->FreeZones;
1300 slgd->FreeZones = z;
1305 logmemory_quick(free_end);
1309 #if defined(INVARIANTS)
1312 * Helper routines for sanity checks
1316 chunk_mark_allocated(SLZone *z, void *chunk)
1318 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1321 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1322 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1323 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
1324 bitptr = &z->z_Bitmap[bitdex >> 5];
1326 KASSERT((*bitptr & (1 << bitdex)) == 0,
1327 ("memory chunk %p is already allocated!", chunk));
1328 *bitptr |= 1 << bitdex;
1333 chunk_mark_free(SLZone *z, void *chunk)
1335 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1338 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1339 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1340 ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
1341 bitptr = &z->z_Bitmap[bitdex >> 5];
1343 KASSERT((*bitptr & (1 << bitdex)) != 0,
1344 ("memory chunk %p is already free!", chunk));
1345 *bitptr &= ~(1 << bitdex);
1353 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1354 * specified alignment. M_* flags are expected in the flags field.
1356 * Alignment must be a multiple of PAGE_SIZE.
1358 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1359 * but when we move zalloc() over to use this function as its backend
1360 * we will have to switch to kreserve/krelease and call reserve(0)
1361 * after the new space is made available.
1363 * Interrupt code which has preempted other code is not allowed to
1364 * use PQ_CACHE pages. However, if an interrupt thread is run
1365 * non-preemptively or blocks and then runs non-preemptively, then
1366 * it is free to use PQ_CACHE pages.
1369 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1373 int count, vmflags, base_vmflags;
1374 vm_page_t mp[ZALLOC_MAX_ZONE_SIZE / PAGE_SIZE];
1377 size = round_page(size);
1378 addr = vm_map_min(&kernel_map);
1381 * Reserve properly aligned space from kernel_map. RNOWAIT allocations
1384 if (flags & M_RNOWAIT) {
1385 if (lwkt_trytoken(&vm_token) == 0)
1388 lwkt_gettoken(&vm_token);
1390 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1392 vm_map_lock(&kernel_map);
1393 if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) {
1394 vm_map_unlock(&kernel_map);
1395 if ((flags & M_NULLOK) == 0)
1396 panic("kmem_slab_alloc(): kernel_map ran out of space!");
1397 vm_map_entry_release(count);
1399 lwkt_reltoken(&vm_token);
1404 * kernel_object maps 1:1 to kernel_map.
1406 vm_object_reference(&kernel_object);
1407 vm_map_insert(&kernel_map, &count,
1408 &kernel_object, addr, addr, addr + size,
1410 VM_PROT_ALL, VM_PROT_ALL,
1417 base_vmflags |= VM_ALLOC_ZERO;
1418 if (flags & M_USE_RESERVE)
1419 base_vmflags |= VM_ALLOC_SYSTEM;
1420 if (flags & M_USE_INTERRUPT_RESERVE)
1421 base_vmflags |= VM_ALLOC_INTERRUPT;
1422 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1423 panic("kmem_slab_alloc: bad flags %08x (%p)",
1424 flags, ((int **)&size)[-1]);
1429 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
1431 for (i = 0; i < size; i += PAGE_SIZE) {
1435 * VM_ALLOC_NORMAL can only be set if we are not preempting.
1437 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1438 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1439 * implied in this case), though I'm not sure if we really need to
1442 vmflags = base_vmflags;
1443 if (flags & M_WAITOK) {
1444 if (td->td_preempted)
1445 vmflags |= VM_ALLOC_SYSTEM;
1447 vmflags |= VM_ALLOC_NORMAL;
1450 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
1451 if (i / PAGE_SIZE < NELEM(mp))
1452 mp[i / PAGE_SIZE] = m;
1455 * If the allocation failed we either return NULL or we retry.
1457 * If M_WAITOK is specified we wait for more memory and retry.
1458 * If M_WAITOK is specified from a preemption we yield instead of
1459 * wait. Livelock will not occur because the interrupt thread
1460 * will not be preempting anyone the second time around after the
1464 if (flags & M_WAITOK) {
1465 if (td->td_preempted) {
1466 vm_map_unlock(&kernel_map);
1468 vm_map_lock(&kernel_map);
1470 vm_map_unlock(&kernel_map);
1472 vm_map_lock(&kernel_map);
1474 i -= PAGE_SIZE; /* retry */
1479 * We were unable to recover, cleanup and return NULL
1481 * (vm_token already held)
1485 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1486 /* page should already be busy */
1489 vm_map_delete(&kernel_map, addr, addr + size, &count);
1490 vm_map_unlock(&kernel_map);
1491 vm_map_entry_release(count);
1493 lwkt_reltoken(&vm_token);
1501 * Mark the map entry as non-pageable using a routine that allows us to
1502 * populate the underlying pages.
1504 * The pages were busied by the allocations above.
1506 vm_map_set_wired_quick(&kernel_map, addr, size, &count);
1510 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1512 for (i = 0; i < size; i += PAGE_SIZE) {
1515 if (i / PAGE_SIZE < NELEM(mp))
1516 m = mp[i / PAGE_SIZE];
1518 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1519 m->valid = VM_PAGE_BITS_ALL;
1520 /* page should already be busy */
1522 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
1523 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1524 bzero((char *)addr + i, PAGE_SIZE);
1525 vm_page_flag_clear(m, PG_ZERO);
1526 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1527 vm_page_flag_set(m, PG_REFERENCED);
1530 vm_map_unlock(&kernel_map);
1531 vm_map_entry_release(count);
1532 lwkt_reltoken(&vm_token);
1533 return((void *)addr);
1540 kmem_slab_free(void *ptr, vm_size_t size)
1543 lwkt_gettoken(&vm_token);
1544 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
1545 lwkt_reltoken(&vm_token);