2 * KERN_SLABALLOC.C - Kernel SLAB memory allocator
4 * Copyright (c) 2003,2004 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
36 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.44 2006/12/23 00:35:04 swildner Exp $
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 arysize(ary) (sizeof(ary)/sizeof((ary)[0]))
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, 0, MEMORY_STRING, MEMORY_ARG_SIZE);
131 KTR_INFO(KTR_MEMORY, memory, free_zero, 1, MEMORY_STRING, MEMORY_ARG_SIZE);
132 KTR_INFO(KTR_MEMORY, memory, free_ovsz, 2, MEMORY_STRING, MEMORY_ARG_SIZE);
133 KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 3, MEMORY_STRING, MEMORY_ARG_SIZE);
134 KTR_INFO(KTR_MEMORY, memory, free_chunk, 4, MEMORY_STRING, MEMORY_ARG_SIZE);
136 KTR_INFO(KTR_MEMORY, memory, free_request, 5, MEMORY_STRING, MEMORY_ARG_SIZE);
137 KTR_INFO(KTR_MEMORY, memory, free_remote, 6, MEMORY_STRING, MEMORY_ARG_SIZE);
139 KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin", 0);
140 KTR_INFO(KTR_MEMORY, memory, free_beg, 0, "free begin", 0);
141 KTR_INFO(KTR_MEMORY, memory, free_end, 0, "free end", 0);
143 #define logmemory(name, ptr, type, size, flags) \
144 KTR_LOG(memory_ ## name, ptr, type, size, flags)
145 #define logmemory_quick(name) \
146 KTR_LOG(memory_ ## name)
149 * Fixed globals (not per-cpu)
152 static int ZoneLimit;
153 static int ZonePageCount;
155 static struct malloc_type *kmemstatistics;
156 static struct kmemusage *kmemusage;
157 static int32_t weirdary[16];
159 static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
160 static void kmem_slab_free(void *ptr, vm_size_t bytes);
161 #if defined(INVARIANTS)
162 static void chunk_mark_allocated(SLZone *z, void *chunk);
163 static void chunk_mark_free(SLZone *z, void *chunk);
167 * Misc constants. Note that allocations that are exact multiples of
168 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
169 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
171 #define MIN_CHUNK_SIZE 8 /* in bytes */
172 #define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
173 #define ZONE_RELS_THRESH 2 /* threshold number of zones */
174 #define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
177 * The WEIRD_ADDR is used as known text to copy into free objects to
178 * try to create deterministic failure cases if the data is accessed after
181 #define WEIRD_ADDR 0xdeadc0de
182 #define MAX_COPY sizeof(weirdary)
183 #define ZERO_LENGTH_PTR ((void *)-8)
186 * Misc global malloc buckets
189 MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
190 MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
191 MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
193 MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
194 MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
197 * Initialize the slab memory allocator. We have to choose a zone size based
198 * on available physical memory. We choose a zone side which is approximately
199 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
200 * 128K. The zone size is limited to the bounds set in slaballoc.h
201 * (typically 32K min, 128K max).
203 static void kmeminit(void *dummy);
205 SYSINIT(kmem, SI_SUB_KMEM, SI_ORDER_FIRST, kmeminit, NULL)
209 * If enabled any memory allocated without M_ZERO is initialized to -1.
211 static int use_malloc_pattern;
212 SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
213 &use_malloc_pattern, 0, "");
217 kmeminit(void *dummy)
224 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
225 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS)
226 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
228 usesize = (int)(limsize / 1024); /* convert to KB */
230 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
231 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
233 ZoneLimit = ZoneSize / 4;
234 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
235 ZoneLimit = ZALLOC_ZONE_LIMIT;
236 ZoneMask = ZoneSize - 1;
237 ZonePageCount = ZoneSize / PAGE_SIZE;
239 npg = (VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE;
240 kmemusage = kmem_slab_alloc(npg * sizeof(struct kmemusage), PAGE_SIZE, M_WAITOK|M_ZERO);
242 for (i = 0; i < arysize(weirdary); ++i)
243 weirdary[i] = WEIRD_ADDR;
246 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
250 * Initialize a malloc type tracking structure.
253 malloc_init(void *data)
255 struct malloc_type *type = data;
258 if (type->ks_magic != M_MAGIC)
259 panic("malloc type lacks magic");
261 if (type->ks_limit != 0)
264 if (vmstats.v_page_count == 0)
265 panic("malloc_init not allowed before vm init");
267 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
268 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS)
269 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
270 type->ks_limit = limsize / 10;
272 type->ks_next = kmemstatistics;
273 kmemstatistics = type;
277 malloc_uninit(void *data)
279 struct malloc_type *type = data;
280 struct malloc_type *t;
286 if (type->ks_magic != M_MAGIC)
287 panic("malloc type lacks magic");
289 if (vmstats.v_page_count == 0)
290 panic("malloc_uninit not allowed before vm init");
292 if (type->ks_limit == 0)
293 panic("malloc_uninit on uninitialized type");
297 * memuse is only correct in aggregation. Due to memory being allocated
298 * on one cpu and freed on another individual array entries may be
299 * negative or positive (canceling each other out).
301 for (i = ttl = 0; i < ncpus; ++i)
302 ttl += type->ks_memuse[i];
304 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
305 ttl, type->ks_shortdesc, i);
308 if (type == kmemstatistics) {
309 kmemstatistics = type->ks_next;
311 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
312 if (t->ks_next == type) {
313 t->ks_next = type->ks_next;
318 type->ks_next = NULL;
323 * Calculate the zone index for the allocation request size and set the
324 * allocation request size to that particular zone's chunk size.
327 zoneindex(unsigned long *bytes)
329 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
331 *bytes = n = (n + 7) & ~7;
332 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
335 *bytes = n = (n + 15) & ~15;
340 *bytes = n = (n + 31) & ~31;
344 *bytes = n = (n + 63) & ~63;
348 *bytes = n = (n + 127) & ~127;
349 return(n / 128 + 31);
352 *bytes = n = (n + 255) & ~255;
353 return(n / 256 + 39);
355 *bytes = n = (n + 511) & ~511;
356 return(n / 512 + 47);
358 #if ZALLOC_ZONE_LIMIT > 8192
360 *bytes = n = (n + 1023) & ~1023;
361 return(n / 1024 + 55);
364 #if ZALLOC_ZONE_LIMIT > 16384
366 *bytes = n = (n + 2047) & ~2047;
367 return(n / 2048 + 63);
370 panic("Unexpected byte count %d", n);
375 * malloc() (SLAB ALLOCATOR)
377 * Allocate memory via the slab allocator. If the request is too large,
378 * or if it page-aligned beyond a certain size, we fall back to the
379 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
380 * &SlabMisc if you don't care.
382 * M_RNOWAIT - don't block.
383 * M_NULLOK - return NULL instead of blocking.
384 * M_ZERO - zero the returned memory.
385 * M_USE_RESERVE - allow greater drawdown of the free list
386 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
392 kmalloc(unsigned long size, struct malloc_type *type, int flags)
397 struct globaldata *gd;
403 logmemory_quick(malloc_beg);
408 * XXX silly to have this in the critical path.
410 if (type->ks_limit == 0) {
412 if (type->ks_limit == 0)
419 * Handle the case where the limit is reached. Panic if we can't return
420 * NULL. The original malloc code looped, but this tended to
421 * simply deadlock the computer.
423 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
424 * to determine if a more complete limit check should be done. The
425 * actual memory use is tracked via ks_memuse[cpu].
427 while (type->ks_loosememuse >= type->ks_limit) {
431 for (i = ttl = 0; i < ncpus; ++i)
432 ttl += type->ks_memuse[i];
433 type->ks_loosememuse = ttl; /* not MP synchronized */
434 if (ttl >= type->ks_limit) {
435 if (flags & M_NULLOK) {
436 logmemory(malloc, NULL, type, size, flags);
439 panic("%s: malloc limit exceeded", type->ks_shortdesc);
444 * Handle the degenerate size == 0 case. Yes, this does happen.
445 * Return a special pointer. This is to maintain compatibility with
446 * the original malloc implementation. Certain devices, such as the
447 * adaptec driver, not only allocate 0 bytes, they check for NULL and
448 * also realloc() later on. Joy.
451 logmemory(malloc, ZERO_LENGTH_PTR, type, size, flags);
452 return(ZERO_LENGTH_PTR);
456 * Handle hysteresis from prior frees here in malloc(). We cannot
457 * safely manipulate the kernel_map in free() due to free() possibly
458 * being called via an IPI message or from sensitive interrupt code.
460 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_RNOWAIT) == 0) {
462 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */
464 slgd->FreeZones = z->z_Next;
466 kmem_slab_free(z, ZoneSize); /* may block */
471 * XXX handle oversized frees that were queued from free().
473 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) {
475 if ((z = slgd->FreeOvZones) != NULL) {
476 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
477 slgd->FreeOvZones = z->z_Next;
478 kmem_slab_free(z, z->z_ChunkSize); /* may block */
484 * Handle large allocations directly. There should not be very many of
485 * these so performance is not a big issue.
487 * Guarentee page alignment for allocations in multiples of PAGE_SIZE
489 if (size >= ZoneLimit || (size & PAGE_MASK) == 0) {
490 struct kmemusage *kup;
492 size = round_page(size);
493 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
495 logmemory(malloc, NULL, type, size, flags);
498 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
499 flags |= M_PASSIVE_ZERO;
501 kup->ku_pagecnt = size / PAGE_SIZE;
502 kup->ku_cpu = gd->gd_cpuid;
508 * Attempt to allocate out of an existing zone. First try the free list,
509 * then allocate out of unallocated space. If we find a good zone move
510 * it to the head of the list so later allocations find it quickly
511 * (we might have thousands of zones in the list).
513 * Note: zoneindex() will panic of size is too large.
515 zi = zoneindex(&size);
516 KKASSERT(zi < NZONES);
518 if ((z = slgd->ZoneAry[zi]) != NULL) {
519 KKASSERT(z->z_NFree > 0);
522 * Remove us from the ZoneAry[] when we become empty
524 if (--z->z_NFree == 0) {
525 slgd->ZoneAry[zi] = z->z_Next;
530 * Locate a chunk in a free page. This attempts to localize
531 * reallocations into earlier pages without us having to sort
532 * the chunk list. A chunk may still overlap a page boundary.
534 while (z->z_FirstFreePg < ZonePageCount) {
535 if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) {
538 * Diagnostic: c_Next is not total garbage.
540 KKASSERT(chunk->c_Next == NULL ||
541 ((intptr_t)chunk->c_Next & IN_SAME_PAGE_MASK) ==
542 ((intptr_t)chunk & IN_SAME_PAGE_MASK));
545 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
546 panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount);
547 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
548 panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount);
549 chunk_mark_allocated(z, chunk);
551 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next;
558 * No chunks are available but NFree said we had some memory, so
559 * it must be available in the never-before-used-memory area
560 * governed by UIndex. The consequences are very serious if our zone
561 * got corrupted so we use an explicit panic rather then a KASSERT.
563 if (z->z_UIndex + 1 != z->z_NMax)
564 z->z_UIndex = z->z_UIndex + 1;
567 if (z->z_UIndex == z->z_UEndIndex)
568 panic("slaballoc: corrupted zone");
569 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
570 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
572 flags |= M_PASSIVE_ZERO;
574 #if defined(INVARIANTS)
575 chunk_mark_allocated(z, chunk);
581 * If all zones are exhausted we need to allocate a new zone for this
582 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
583 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
584 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
585 * we do not pre-zero it because we do not want to mess up the L1 cache.
587 * At least one subsystem, the tty code (see CROUND) expects power-of-2
588 * allocations to be power-of-2 aligned. We maintain compatibility by
589 * adjusting the base offset below.
594 if ((z = slgd->FreeZones) != NULL) {
595 slgd->FreeZones = z->z_Next;
597 bzero(z, sizeof(SLZone));
598 z->z_Flags |= SLZF_UNOTZEROD;
600 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
606 * How big is the base structure?
608 #if defined(INVARIANTS)
610 * Make room for z_Bitmap. An exact calculation is somewhat more
611 * complicated so don't make an exact calculation.
613 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
614 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
616 off = sizeof(SLZone);
620 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
621 * Otherwise just 8-byte align the data.
623 if ((size | (size - 1)) + 1 == (size << 1))
624 off = (off + size - 1) & ~(size - 1);
626 off = (off + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
627 z->z_Magic = ZALLOC_SLAB_MAGIC;
629 z->z_NMax = (ZoneSize - off) / size;
630 z->z_NFree = z->z_NMax - 1;
631 z->z_BasePtr = (char *)z + off;
632 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
633 z->z_ChunkSize = size;
634 z->z_FirstFreePg = ZonePageCount;
636 z->z_Cpu = gd->gd_cpuid;
637 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
638 z->z_Next = slgd->ZoneAry[zi];
639 slgd->ZoneAry[zi] = z;
640 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
641 flags &= ~M_ZERO; /* already zero'd */
642 flags |= M_PASSIVE_ZERO;
644 #if defined(INVARIANTS)
645 chunk_mark_allocated(z, chunk);
649 * Slide the base index for initial allocations out of the next
650 * zone we create so we do not over-weight the lower part of the
653 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
654 & (ZALLOC_MAX_ZONE_SIZE - 1);
657 ++type->ks_inuse[gd->gd_cpuid];
658 type->ks_memuse[gd->gd_cpuid] += size;
659 type->ks_loosememuse += size; /* not MP synchronized */
664 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
665 if (use_malloc_pattern) {
666 for (i = 0; i < size; i += sizeof(int)) {
667 *(int *)((char *)chunk + i) = -1;
670 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
673 logmemory(malloc, chunk, type, size, flags);
677 logmemory(malloc, NULL, type, size, flags);
682 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
684 * Generally speaking this routine is not called very often and we do
685 * not attempt to optimize it beyond reusing the same pointer if the
686 * new size fits within the chunking of the old pointer's zone.
689 krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
695 KKASSERT((flags & M_ZERO) == 0); /* not supported */
697 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
698 return(kmalloc(size, type, flags));
705 * Handle oversized allocations. XXX we really should require that a
706 * size be passed to free() instead of this nonsense.
709 struct kmemusage *kup;
712 if (kup->ku_pagecnt) {
713 osize = kup->ku_pagecnt << PAGE_SHIFT;
714 if (osize == round_page(size))
716 if ((nptr = kmalloc(size, type, flags)) == NULL)
718 bcopy(ptr, nptr, min(size, osize));
725 * Get the original allocation's zone. If the new request winds up
726 * using the same chunk size we do not have to do anything.
728 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
729 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
732 if (z->z_ChunkSize == size)
736 * Allocate memory for the new request size. Note that zoneindex has
737 * already adjusted the request size to the appropriate chunk size, which
738 * should optimize our bcopy(). Then copy and return the new pointer.
740 if ((nptr = kmalloc(size, type, flags)) == NULL)
742 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
748 * Allocate a copy of the specified string.
750 * (MP SAFE) (MAY BLOCK)
753 kstrdup(const char *str, struct malloc_type *type)
755 int zlen; /* length inclusive of terminating NUL */
760 zlen = strlen(str) + 1;
761 nstr = kmalloc(zlen, type, M_WAITOK);
762 bcopy(str, nstr, zlen);
768 * free() (SLAB ALLOCATOR)
770 * Free the specified chunk of memory.
774 free_remote(void *ptr)
776 logmemory(free_remote, ptr, *(struct malloc_type **)ptr, -1, 0);
777 kfree(ptr, *(struct malloc_type **)ptr);
783 * free (SLAB ALLOCATOR)
785 * Free a memory block previously allocated by malloc. Note that we do not
786 * attempt to uplodate ks_loosememuse as MP races could prevent us from
787 * checking memory limits in malloc.
792 kfree(void *ptr, struct malloc_type *type)
797 struct globaldata *gd;
800 logmemory_quick(free_beg);
805 panic("trying to free NULL pointer");
808 * Handle special 0-byte allocations
810 if (ptr == ZERO_LENGTH_PTR) {
811 logmemory(free_zero, ptr, type, -1, 0);
812 logmemory_quick(free_end);
817 * Handle oversized allocations. XXX we really should require that a
818 * size be passed to free() instead of this nonsense.
820 * This code is never called via an ipi.
823 struct kmemusage *kup;
827 if (kup->ku_pagecnt) {
828 size = kup->ku_pagecnt << PAGE_SHIFT;
831 KKASSERT(sizeof(weirdary) <= size);
832 bcopy(weirdary, ptr, sizeof(weirdary));
835 * note: we always adjust our cpu's slot, not the originating
836 * cpu (kup->ku_cpuid). The statistics are in aggregate.
838 * note: XXX we have still inherited the interrupts-can't-block
839 * assumption. An interrupt thread does not bump
840 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
841 * primarily until we can fix softupdate's assumptions about free().
844 --type->ks_inuse[gd->gd_cpuid];
845 type->ks_memuse[gd->gd_cpuid] -= size;
846 if (mycpu->gd_intr_nesting_level || (gd->gd_curthread->td_flags & TDF_INTTHREAD)) {
847 logmemory(free_ovsz_delayed, ptr, type, size, 0);
849 z->z_Magic = ZALLOC_OVSZ_MAGIC;
850 z->z_Next = slgd->FreeOvZones;
851 z->z_ChunkSize = size;
852 slgd->FreeOvZones = z;
856 logmemory(free_ovsz, ptr, type, size, 0);
857 kmem_slab_free(ptr, size); /* may block */
859 logmemory_quick(free_end);
865 * Zone case. Figure out the zone based on the fact that it is
868 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
869 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
872 * If we do not own the zone then forward the request to the
873 * cpu that does. Since the timing is non-critical, a passive
876 if (z->z_CpuGd != gd) {
877 *(struct malloc_type **)ptr = type;
879 logmemory(free_request, ptr, type, z->z_ChunkSize, 0);
880 lwkt_send_ipiq_passive(z->z_CpuGd, free_remote, ptr);
882 panic("Corrupt SLZone");
884 logmemory_quick(free_end);
888 logmemory(free_chunk, ptr, type, z->z_ChunkSize, 0);
890 if (type->ks_magic != M_MAGIC)
891 panic("free: malloc type lacks magic");
894 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT;
899 * Attempt to detect a double-free. To reduce overhead we only check
900 * if there appears to be link pointer at the base of the data.
902 if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) {
904 for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) {
906 panic("Double free at %p", chunk);
909 chunk_mark_free(z, chunk);
913 * Put weird data into the memory to detect modifications after freeing,
914 * illegal pointer use after freeing (we should fault on the odd address),
915 * and so forth. XXX needs more work, see the old malloc code.
918 if (z->z_ChunkSize < sizeof(weirdary))
919 bcopy(weirdary, chunk, z->z_ChunkSize);
921 bcopy(weirdary, chunk, sizeof(weirdary));
925 * Add this free non-zero'd chunk to a linked list for reuse, adjust
929 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
930 panic("BADFREE %p", chunk);
932 chunk->c_Next = z->z_PageAry[pgno];
933 z->z_PageAry[pgno] = chunk;
935 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
938 if (z->z_FirstFreePg > pgno)
939 z->z_FirstFreePg = pgno;
942 * Bump the number of free chunks. If it becomes non-zero the zone
943 * must be added back onto the appropriate list.
945 if (z->z_NFree++ == 0) {
946 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
947 slgd->ZoneAry[z->z_ZoneIndex] = z;
950 --type->ks_inuse[z->z_Cpu];
951 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
954 * If the zone becomes totally free, and there are other zones we
955 * can allocate from, move this zone to the FreeZones list. Since
956 * this code can be called from an IPI callback, do *NOT* try to mess
957 * with kernel_map here. Hysteresis will be performed at malloc() time.
959 if (z->z_NFree == z->z_NMax &&
960 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z)
964 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
968 z->z_Next = slgd->FreeZones;
972 logmemory_quick(free_end);
976 #if defined(INVARIANTS)
978 * Helper routines for sanity checks
982 chunk_mark_allocated(SLZone *z, void *chunk)
984 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
987 KASSERT(bitdex >= 0 && bitdex < z->z_NMax, ("memory chunk %p bit index %d is illegal", chunk, bitdex));
988 bitptr = &z->z_Bitmap[bitdex >> 5];
990 KASSERT((*bitptr & (1 << bitdex)) == 0, ("memory chunk %p is already allocated!", chunk));
991 *bitptr |= 1 << bitdex;
996 chunk_mark_free(SLZone *z, void *chunk)
998 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1001 KASSERT(bitdex >= 0 && bitdex < z->z_NMax, ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
1002 bitptr = &z->z_Bitmap[bitdex >> 5];
1004 KASSERT((*bitptr & (1 << bitdex)) != 0, ("memory chunk %p is already free!", chunk));
1005 *bitptr &= ~(1 << bitdex);
1013 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1014 * specified alignment. M_* flags are expected in the flags field.
1016 * Alignment must be a multiple of PAGE_SIZE.
1018 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1019 * but when we move zalloc() over to use this function as its backend
1020 * we will have to switch to kreserve/krelease and call reserve(0)
1021 * after the new space is made available.
1023 * Interrupt code which has preempted other code is not allowed to
1024 * use PQ_CACHE pages. However, if an interrupt thread is run
1025 * non-preemptively or blocks and then runs non-preemptively, then
1026 * it is free to use PQ_CACHE pages.
1028 * This routine will currently obtain the BGL.
1030 * MPALMOSTSAFE - acquires mplock
1033 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1038 int count, vmflags, base_vmflags;
1040 vm_map_t map = kernel_map;
1042 size = round_page(size);
1043 addr = vm_map_min(map);
1046 * Reserve properly aligned space from kernel_map. RNOWAIT allocations
1049 if (flags & M_RNOWAIT) {
1050 if (try_mplock() == 0)
1055 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1058 if (vm_map_findspace(map, vm_map_min(map), size, align, &addr)) {
1060 if ((flags & M_NULLOK) == 0)
1061 panic("kmem_slab_alloc(): kernel_map ran out of space!");
1063 vm_map_entry_release(count);
1067 offset = addr - VM_MIN_KERNEL_ADDRESS;
1068 vm_object_reference(kernel_object);
1069 vm_map_insert(map, &count,
1070 kernel_object, offset, addr, addr + size,
1072 VM_PROT_ALL, VM_PROT_ALL,
1079 base_vmflags |= VM_ALLOC_ZERO;
1080 if (flags & M_USE_RESERVE)
1081 base_vmflags |= VM_ALLOC_SYSTEM;
1082 if (flags & M_USE_INTERRUPT_RESERVE)
1083 base_vmflags |= VM_ALLOC_INTERRUPT;
1084 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0)
1085 panic("kmem_slab_alloc: bad flags %08x (%p)", flags, ((int **)&size)[-1]);
1089 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
1091 for (i = 0; i < size; i += PAGE_SIZE) {
1093 vm_pindex_t idx = OFF_TO_IDX(offset + i);
1096 * VM_ALLOC_NORMAL can only be set if we are not preempting.
1098 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1099 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1100 * implied in this case), though I'm sure if we really need to do
1103 vmflags = base_vmflags;
1104 if (flags & M_WAITOK) {
1105 if (td->td_preempted)
1106 vmflags |= VM_ALLOC_SYSTEM;
1108 vmflags |= VM_ALLOC_NORMAL;
1111 m = vm_page_alloc(kernel_object, idx, vmflags);
1114 * If the allocation failed we either return NULL or we retry.
1116 * If M_WAITOK is specified we wait for more memory and retry.
1117 * If M_WAITOK is specified from a preemption we yield instead of
1118 * wait. Livelock will not occur because the interrupt thread
1119 * will not be preempting anyone the second time around after the
1123 if (flags & M_WAITOK) {
1124 if (td->td_preempted) {
1133 i -= PAGE_SIZE; /* retry */
1138 * We were unable to recover, cleanup and return NULL
1142 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
1145 vm_map_delete(map, addr, addr + size, &count);
1148 vm_map_entry_release(count);
1157 * Mark the map entry as non-pageable using a routine that allows us to
1158 * populate the underlying pages.
1160 vm_map_set_wired_quick(map, addr, size, &count);
1164 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1166 for (i = 0; i < size; i += PAGE_SIZE) {
1169 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
1170 m->valid = VM_PAGE_BITS_ALL;
1173 pmap_enter(kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
1174 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1175 bzero((char *)addr + i, PAGE_SIZE);
1176 vm_page_flag_clear(m, PG_ZERO);
1177 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_REFERENCED);
1180 vm_map_entry_release(count);
1182 return((void *)addr);
1188 * MPALMOSTSAFE - acquires mplock
1191 kmem_slab_free(void *ptr, vm_size_t size)
1195 vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);