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 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.55 2008/10/22 01:42:17 dillon 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 instantanious, and fragmentation/overhead losses are limited
46 * to a fixed worst-case amount.
48 * The downside of this slab implementation is in the chunk size
49 * multiplied by the number of zones. ~80 zones * 128K = 10MB of VM per cpu.
50 * In a kernel implementation all this memory will be physical so
51 * the zone size is adjusted downward on machines with less physical
52 * memory. The upside is that overhead is bounded... this is the *worst*
55 * Slab management is done on a per-cpu basis and no locking or mutexes
56 * are required, only a critical section. When one cpu frees memory
57 * belonging to another cpu's slab manager an asynchronous IPI message
58 * will be queued to execute the operation. In addition, both the
59 * high level slab allocator and the low level zone allocator optimize
60 * M_ZERO requests, and the slab allocator does not have to pre initialize
61 * the linked list of chunks.
63 * XXX Balancing is needed between cpus. Balance will be handled through
64 * asynchronous IPIs primarily by reassigning the z_Cpu ownership of chunks.
66 * XXX If we have to allocate a new zone and M_USE_RESERVE is set, use of
67 * the new zone should be restricted to M_USE_RESERVE requests only.
69 * Alloc Size Chunking Number of zones
79 * (if PAGE_SIZE is 4K the maximum zone allocation is 16383)
81 * Allocations >= ZoneLimit go directly to kmem.
83 * API REQUIREMENTS AND SIDE EFFECTS
85 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we
86 * have remained compatible with the following API requirements:
88 * + small power-of-2 sized allocations are power-of-2 aligned (kern_tty)
89 * + all power-of-2 sized allocations are power-of-2 aligned (twe)
90 * + malloc(0) is allowed and returns non-NULL (ahc driver)
91 * + ability to allocate arbitrarily large chunks of memory
96 #include <sys/param.h>
97 #include <sys/systm.h>
98 #include <sys/kernel.h>
99 #include <sys/slaballoc.h>
100 #include <sys/mbuf.h>
101 #include <sys/vmmeter.h>
102 #include <sys/lock.h>
103 #include <sys/thread.h>
104 #include <sys/globaldata.h>
105 #include <sys/sysctl.h>
109 #include <vm/vm_param.h>
110 #include <vm/vm_kern.h>
111 #include <vm/vm_extern.h>
112 #include <vm/vm_object.h>
114 #include <vm/vm_map.h>
115 #include <vm/vm_page.h>
116 #include <vm/vm_pageout.h>
118 #include <machine/cpu.h>
120 #include <sys/thread2.h>
122 #define arysize(ary) (sizeof(ary)/sizeof((ary)[0]))
124 #define btokup(z) (&pmap_kvtom((vm_offset_t)(z))->ku_pagecnt)
126 #define MEMORY_STRING "ptr=%p type=%p size=%d flags=%04x"
127 #define MEMORY_ARG_SIZE (sizeof(void *) * 2 + sizeof(unsigned long) + \
130 #if !defined(KTR_MEMORY)
131 #define KTR_MEMORY KTR_ALL
133 KTR_INFO_MASTER(memory);
134 KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin", 0);
135 KTR_INFO(KTR_MEMORY, memory, malloc_end, 1, MEMORY_STRING, MEMORY_ARG_SIZE);
136 KTR_INFO(KTR_MEMORY, memory, free_zero, 2, MEMORY_STRING, MEMORY_ARG_SIZE);
137 KTR_INFO(KTR_MEMORY, memory, free_ovsz, 3, MEMORY_STRING, MEMORY_ARG_SIZE);
138 KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 4, MEMORY_STRING, MEMORY_ARG_SIZE);
139 KTR_INFO(KTR_MEMORY, memory, free_chunk, 5, MEMORY_STRING, MEMORY_ARG_SIZE);
141 KTR_INFO(KTR_MEMORY, memory, free_request, 6, MEMORY_STRING, MEMORY_ARG_SIZE);
142 KTR_INFO(KTR_MEMORY, memory, free_rem_beg, 7, MEMORY_STRING, MEMORY_ARG_SIZE);
143 KTR_INFO(KTR_MEMORY, memory, free_rem_end, 8, MEMORY_STRING, MEMORY_ARG_SIZE);
145 KTR_INFO(KTR_MEMORY, memory, free_beg, 9, "free begin", 0);
146 KTR_INFO(KTR_MEMORY, memory, free_end, 10, "free end", 0);
148 #define logmemory(name, ptr, type, size, flags) \
149 KTR_LOG(memory_ ## name, ptr, type, size, flags)
150 #define logmemory_quick(name) \
151 KTR_LOG(memory_ ## name)
154 * Fixed globals (not per-cpu)
157 static int ZoneLimit;
158 static int ZonePageCount;
159 static uintptr_t ZoneMask;
160 static int ZoneBigAlloc; /* in KB */
161 static int ZoneGenAlloc; /* in KB */
162 struct malloc_type *kmemstatistics; /* exported to vmstat */
163 static int32_t weirdary[16];
165 static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
166 static void kmem_slab_free(void *ptr, vm_size_t bytes);
168 #if defined(INVARIANTS)
169 static void chunk_mark_allocated(SLZone *z, void *chunk);
170 static void chunk_mark_free(SLZone *z, void *chunk);
172 #define chunk_mark_allocated(z, chunk)
173 #define chunk_mark_free(z, chunk)
177 * Misc constants. Note that allocations that are exact multiples of
178 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
179 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
181 #define MIN_CHUNK_SIZE 8 /* in bytes */
182 #define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
183 #define ZONE_RELS_THRESH 2 /* threshold number of zones */
184 #define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
187 * The WEIRD_ADDR is used as known text to copy into free objects to
188 * try to create deterministic failure cases if the data is accessed after
191 #define WEIRD_ADDR 0xdeadc0de
192 #define MAX_COPY sizeof(weirdary)
193 #define ZERO_LENGTH_PTR ((void *)-8)
196 * Misc global malloc buckets
199 MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
200 MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
201 MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
203 MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
204 MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
207 * Initialize the slab memory allocator. We have to choose a zone size based
208 * on available physical memory. We choose a zone side which is approximately
209 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
210 * 128K. The zone size is limited to the bounds set in slaballoc.h
211 * (typically 32K min, 128K max).
213 static void kmeminit(void *dummy);
217 SYSINIT(kmem, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL)
221 * If enabled any memory allocated without M_ZERO is initialized to -1.
223 static int use_malloc_pattern;
224 SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
225 &use_malloc_pattern, 0, "");
228 SYSCTL_INT(_kern, OID_AUTO, zone_big_alloc, CTLFLAG_RD, &ZoneBigAlloc, 0, "");
229 SYSCTL_INT(_kern, OID_AUTO, zone_gen_alloc, CTLFLAG_RD, &ZoneGenAlloc, 0, "");
232 kmeminit(void *dummy)
238 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
239 if (limsize > KvaSize)
242 usesize = (int)(limsize / 1024); /* convert to KB */
244 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
245 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
247 ZoneLimit = ZoneSize / 4;
248 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
249 ZoneLimit = ZALLOC_ZONE_LIMIT;
250 ZoneMask = ~(uintptr_t)(ZoneSize - 1);
251 ZonePageCount = ZoneSize / PAGE_SIZE;
253 for (i = 0; i < arysize(weirdary); ++i)
254 weirdary[i] = WEIRD_ADDR;
256 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
259 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
263 * Initialize a malloc type tracking structure.
266 malloc_init(void *data)
268 struct malloc_type *type = data;
271 if (type->ks_magic != M_MAGIC)
272 panic("malloc type lacks magic");
274 if (type->ks_limit != 0)
277 if (vmstats.v_page_count == 0)
278 panic("malloc_init not allowed before vm init");
280 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
281 if (limsize > KvaSize)
283 type->ks_limit = limsize / 10;
285 type->ks_next = kmemstatistics;
286 kmemstatistics = type;
290 malloc_uninit(void *data)
292 struct malloc_type *type = data;
293 struct malloc_type *t;
299 if (type->ks_magic != M_MAGIC)
300 panic("malloc type lacks magic");
302 if (vmstats.v_page_count == 0)
303 panic("malloc_uninit not allowed before vm init");
305 if (type->ks_limit == 0)
306 panic("malloc_uninit on uninitialized type");
309 /* Make sure that all pending kfree()s are finished. */
310 lwkt_synchronize_ipiqs("muninit");
315 * memuse is only correct in aggregation. Due to memory being allocated
316 * on one cpu and freed on another individual array entries may be
317 * negative or positive (canceling each other out).
319 for (i = ttl = 0; i < ncpus; ++i)
320 ttl += type->ks_memuse[i];
322 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
323 ttl, type->ks_shortdesc, i);
326 if (type == kmemstatistics) {
327 kmemstatistics = type->ks_next;
329 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
330 if (t->ks_next == type) {
331 t->ks_next = type->ks_next;
336 type->ks_next = NULL;
341 * Increase the kmalloc pool limit for the specified pool. No changes
342 * are the made if the pool would shrink.
345 kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
347 if (type->ks_limit == 0)
351 if (type->ks_limit < bytes)
352 type->ks_limit = bytes;
356 * Dynamically create a malloc pool. This function is a NOP if *typep is
360 kmalloc_create(struct malloc_type **typep, const char *descr)
362 struct malloc_type *type;
364 if (*typep == NULL) {
365 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
366 type->ks_magic = M_MAGIC;
367 type->ks_shortdesc = descr;
374 * Destroy a dynamically created malloc pool. This function is a NOP if
375 * the pool has already been destroyed.
378 kmalloc_destroy(struct malloc_type **typep)
380 if (*typep != NULL) {
381 malloc_uninit(*typep);
382 kfree(*typep, M_TEMP);
388 * Calculate the zone index for the allocation request size and set the
389 * allocation request size to that particular zone's chunk size.
392 zoneindex(unsigned long *bytes)
394 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
396 *bytes = n = (n + 7) & ~7;
397 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
400 *bytes = n = (n + 15) & ~15;
405 *bytes = n = (n + 31) & ~31;
409 *bytes = n = (n + 63) & ~63;
413 *bytes = n = (n + 127) & ~127;
414 return(n / 128 + 31);
417 *bytes = n = (n + 255) & ~255;
418 return(n / 256 + 39);
420 *bytes = n = (n + 511) & ~511;
421 return(n / 512 + 47);
423 #if ZALLOC_ZONE_LIMIT > 8192
425 *bytes = n = (n + 1023) & ~1023;
426 return(n / 1024 + 55);
429 #if ZALLOC_ZONE_LIMIT > 16384
431 *bytes = n = (n + 2047) & ~2047;
432 return(n / 2048 + 63);
435 panic("Unexpected byte count %d", n);
440 * kmalloc() (SLAB ALLOCATOR)
442 * Allocate memory via the slab allocator. If the request is too large,
443 * or if it page-aligned beyond a certain size, we fall back to the
444 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
445 * &SlabMisc if you don't care.
447 * M_RNOWAIT - don't block.
448 * M_NULLOK - return NULL instead of blocking.
449 * M_ZERO - zero the returned memory.
450 * M_USE_RESERVE - allow greater drawdown of the free list
451 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
456 kmalloc(unsigned long size, struct malloc_type *type, int flags)
464 struct globaldata *gd;
470 logmemory_quick(malloc_beg);
475 * XXX silly to have this in the critical path.
477 if (type->ks_limit == 0) {
479 if (type->ks_limit == 0)
486 * Handle the case where the limit is reached. Panic if we can't return
487 * NULL. The original malloc code looped, but this tended to
488 * simply deadlock the computer.
490 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
491 * to determine if a more complete limit check should be done. The
492 * actual memory use is tracked via ks_memuse[cpu].
494 while (type->ks_loosememuse >= type->ks_limit) {
498 for (i = ttl = 0; i < ncpus; ++i)
499 ttl += type->ks_memuse[i];
500 type->ks_loosememuse = ttl; /* not MP synchronized */
501 if ((ssize_t)ttl < 0) /* deal with occassional race */
503 if (ttl >= type->ks_limit) {
504 if (flags & M_NULLOK) {
505 logmemory(malloc_end, NULL, type, size, flags);
508 panic("%s: malloc limit exceeded", type->ks_shortdesc);
513 * Handle the degenerate size == 0 case. Yes, this does happen.
514 * Return a special pointer. This is to maintain compatibility with
515 * the original malloc implementation. Certain devices, such as the
516 * adaptec driver, not only allocate 0 bytes, they check for NULL and
517 * also realloc() later on. Joy.
520 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
521 return(ZERO_LENGTH_PTR);
525 * Handle hysteresis from prior frees here in malloc(). We cannot
526 * safely manipulate the kernel_map in free() due to free() possibly
527 * being called via an IPI message or from sensitive interrupt code.
529 * NOTE: ku_pagecnt must be cleared before we free the slab or we
530 * might race another cpu allocating the kva and setting
533 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_RNOWAIT) == 0) {
535 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */
539 slgd->FreeZones = z->z_Next;
543 kmem_slab_free(z, ZoneSize); /* may block */
544 atomic_add_int(&ZoneGenAlloc, -(int)ZoneSize / 1024);
550 * XXX handle oversized frees that were queued from kfree().
552 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) {
554 if ((z = slgd->FreeOvZones) != NULL) {
557 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
558 slgd->FreeOvZones = z->z_Next;
559 tsize = z->z_ChunkSize;
560 kmem_slab_free(z, tsize); /* may block */
561 atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024);
567 * Handle large allocations directly. There should not be very many of
568 * these so performance is not a big issue.
570 * The backend allocator is pretty nasty on a SMP system. Use the
571 * slab allocator for one and two page-sized chunks even though we lose
572 * some efficiency. XXX maybe fix mmio and the elf loader instead.
574 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
577 size = round_page(size);
578 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
580 logmemory(malloc_end, NULL, type, size, flags);
583 atomic_add_int(&ZoneBigAlloc, (int)size / 1024);
584 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
585 flags |= M_PASSIVE_ZERO;
587 *kup = size / PAGE_SIZE;
593 * Attempt to allocate out of an existing zone. First try the free list,
594 * then allocate out of unallocated space. If we find a good zone move
595 * it to the head of the list so later allocations find it quickly
596 * (we might have thousands of zones in the list).
598 * Note: zoneindex() will panic of size is too large.
600 zi = zoneindex(&size);
601 KKASSERT(zi < NZONES);
604 if ((z = slgd->ZoneAry[zi]) != NULL) {
606 * Locate a chunk - we have to have at least one. If this is the
607 * last chunk go ahead and do the work to retrieve chunks freed
608 * from remote cpus, and if the zone is still empty move it off
611 if (--z->z_NFree <= 0) {
612 KKASSERT(z->z_NFree == 0);
616 * WARNING! This code competes with other cpus. It is ok
617 * for us to not drain RChunks here but we might as well, and
618 * it is ok if more accumulate after we're done.
620 * Set RSignal before pulling rchunks off, indicating that we
621 * will be moving ourselves off of the ZoneAry. Remote ends will
622 * read RSignal before putting rchunks on thus interlocking
623 * their IPI signaling.
625 if (z->z_RChunks == NULL)
626 atomic_swap_int(&z->z_RSignal, 1);
628 while ((bchunk = z->z_RChunks) != NULL) {
630 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
631 *z->z_LChunksp = bchunk;
633 chunk_mark_free(z, bchunk);
634 z->z_LChunksp = &bchunk->c_Next;
635 bchunk = bchunk->c_Next;
643 * Remove from the zone list if no free chunks remain.
646 if (z->z_NFree == 0) {
647 slgd->ZoneAry[zi] = z->z_Next;
655 * Fast path, we have chunks available in z_LChunks.
657 chunk = z->z_LChunks;
659 chunk_mark_allocated(z, chunk);
660 z->z_LChunks = chunk->c_Next;
661 if (z->z_LChunks == NULL)
662 z->z_LChunksp = &z->z_LChunks;
667 * No chunks are available in LChunks, the free chunk MUST be
668 * in the never-before-used memory area, controlled by UIndex.
670 * The consequences are very serious if our zone got corrupted so
671 * we use an explicit panic rather than a KASSERT.
673 if (z->z_UIndex + 1 != z->z_NMax)
678 if (z->z_UIndex == z->z_UEndIndex)
679 panic("slaballoc: corrupted zone");
681 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
682 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
684 flags |= M_PASSIVE_ZERO;
686 chunk_mark_allocated(z, chunk);
691 * If all zones are exhausted we need to allocate a new zone for this
692 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
693 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
694 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
695 * we do not pre-zero it because we do not want to mess up the L1 cache.
697 * At least one subsystem, the tty code (see CROUND) expects power-of-2
698 * allocations to be power-of-2 aligned. We maintain compatibility by
699 * adjusting the base offset below.
705 if ((z = slgd->FreeZones) != NULL) {
706 slgd->FreeZones = z->z_Next;
708 bzero(z, sizeof(SLZone));
709 z->z_Flags |= SLZF_UNOTZEROD;
711 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
714 atomic_add_int(&ZoneGenAlloc, (int)ZoneSize / 1024);
718 * How big is the base structure?
720 #if defined(INVARIANTS)
722 * Make room for z_Bitmap. An exact calculation is somewhat more
723 * complicated so don't make an exact calculation.
725 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
726 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
728 off = sizeof(SLZone);
732 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
733 * Otherwise just 8-byte align the data.
735 if ((size | (size - 1)) + 1 == (size << 1))
736 off = (off + size - 1) & ~(size - 1);
738 off = (off + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
739 z->z_Magic = ZALLOC_SLAB_MAGIC;
741 z->z_NMax = (ZoneSize - off) / size;
742 z->z_NFree = z->z_NMax - 1;
743 z->z_BasePtr = (char *)z + off;
744 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
745 z->z_ChunkSize = size;
747 z->z_Cpu = gd->gd_cpuid;
748 z->z_LChunksp = &z->z_LChunks;
749 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
750 z->z_Next = slgd->ZoneAry[zi];
751 slgd->ZoneAry[zi] = z;
752 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
753 flags &= ~M_ZERO; /* already zero'd */
754 flags |= M_PASSIVE_ZERO;
757 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */
758 chunk_mark_allocated(z, chunk);
761 * Slide the base index for initial allocations out of the next
762 * zone we create so we do not over-weight the lower part of the
765 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
766 & (ZALLOC_MAX_ZONE_SIZE - 1);
770 ++type->ks_inuse[gd->gd_cpuid];
771 type->ks_memuse[gd->gd_cpuid] += size;
772 type->ks_loosememuse += size; /* not MP synchronized */
778 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
779 if (use_malloc_pattern) {
780 for (i = 0; i < size; i += sizeof(int)) {
781 *(int *)((char *)chunk + i) = -1;
784 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
787 logmemory(malloc_end, chunk, type, size, flags);
791 logmemory(malloc_end, NULL, type, size, flags);
796 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
798 * Generally speaking this routine is not called very often and we do
799 * not attempt to optimize it beyond reusing the same pointer if the
800 * new size fits within the chunking of the old pointer's zone.
803 krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
810 KKASSERT((flags & M_ZERO) == 0); /* not supported */
812 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
813 return(kmalloc(size, type, flags));
820 * Handle oversized allocations. XXX we really should require that a
821 * size be passed to free() instead of this nonsense.
825 osize = *kup << PAGE_SHIFT;
826 if (osize == round_page(size))
828 if ((nptr = kmalloc(size, type, flags)) == NULL)
830 bcopy(ptr, nptr, min(size, osize));
836 * Get the original allocation's zone. If the new request winds up
837 * using the same chunk size we do not have to do anything.
839 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
842 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
845 * Allocate memory for the new request size. Note that zoneindex has
846 * already adjusted the request size to the appropriate chunk size, which
847 * should optimize our bcopy(). Then copy and return the new pointer.
849 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
850 * necessary align the result.
852 * We can only zoneindex (to align size to the chunk size) if the new
853 * size is not too large.
855 if (size < ZoneLimit) {
857 if (z->z_ChunkSize == size)
860 if ((nptr = kmalloc(size, type, flags)) == NULL)
862 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
868 * Return the kmalloc limit for this type, in bytes.
871 kmalloc_limit(struct malloc_type *type)
873 if (type->ks_limit == 0) {
875 if (type->ks_limit == 0)
879 return(type->ks_limit);
883 * Allocate a copy of the specified string.
885 * (MP SAFE) (MAY BLOCK)
888 kstrdup(const char *str, struct malloc_type *type)
890 int zlen; /* length inclusive of terminating NUL */
895 zlen = strlen(str) + 1;
896 nstr = kmalloc(zlen, type, M_WAITOK);
897 bcopy(str, nstr, zlen);
903 * Notify our cpu that a remote cpu has freed some chunks in a zone that
904 * we own. RCount will be bumped so the memory should be good, but validate
909 kfree_remote(void *ptr)
917 slgd = &mycpu->gd_slab;
920 KKASSERT(*kup == -((int)mycpuid + 1));
921 KKASSERT(z->z_RCount > 0);
922 atomic_subtract_int(&z->z_RCount, 1);
924 logmemory(free_rem_beg, z, NULL, 0, 0);
925 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
926 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
930 * Indicate that we will no longer be off of the ZoneAry by
937 * Atomically extract the bchunks list and then process it back
938 * into the lchunks list. We want to append our bchunks to the
939 * lchunks list and not prepend since we likely do not have
940 * cache mastership of the related data (not that it helps since
941 * we are using c_Next).
943 while ((bchunk = z->z_RChunks) != NULL) {
945 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
946 *z->z_LChunksp = bchunk;
948 chunk_mark_free(z, bchunk);
949 z->z_LChunksp = &bchunk->c_Next;
950 bchunk = bchunk->c_Next;
956 if (z->z_NFree && nfree == 0) {
957 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
958 slgd->ZoneAry[z->z_ZoneIndex] = z;
962 * If the zone becomes totally free, and there are other zones we
963 * can allocate from, move this zone to the FreeZones list. Since
964 * this code can be called from an IPI callback, do *NOT* try to mess
965 * with kernel_map here. Hysteresis will be performed at malloc() time.
967 * Do not move the zone if there is an IPI inflight, otherwise MP
968 * races can result in our free_remote code accessing a destroyed
971 if (z->z_NFree == z->z_NMax &&
972 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
978 for (pz = &slgd->ZoneAry[z->z_ZoneIndex];
980 pz = &(*pz)->z_Next) {
985 z->z_Next = slgd->FreeZones;
991 logmemory(free_rem_end, z, bchunk, 0, 0);
997 * free (SLAB ALLOCATOR)
999 * Free a memory block previously allocated by malloc. Note that we do not
1000 * attempt to update ks_loosememuse as MP races could prevent us from
1001 * checking memory limits in malloc.
1006 kfree(void *ptr, struct malloc_type *type)
1011 struct globaldata *gd;
1019 logmemory_quick(free_beg);
1021 slgd = &gd->gd_slab;
1024 panic("trying to free NULL pointer");
1027 * Handle special 0-byte allocations
1029 if (ptr == ZERO_LENGTH_PTR) {
1030 logmemory(free_zero, ptr, type, -1, 0);
1031 logmemory_quick(free_end);
1036 * Panic on bad malloc type
1038 if (type->ks_magic != M_MAGIC)
1039 panic("free: malloc type lacks magic");
1042 * Handle oversized allocations. XXX we really should require that a
1043 * size be passed to free() instead of this nonsense.
1045 * This code is never called via an ipi.
1049 size = *kup << PAGE_SHIFT;
1052 KKASSERT(sizeof(weirdary) <= size);
1053 bcopy(weirdary, ptr, sizeof(weirdary));
1056 * NOTE: For oversized allocations we do not record the
1057 * originating cpu. It gets freed on the cpu calling
1058 * kfree(). The statistics are in aggregate.
1060 * note: XXX we have still inherited the interrupts-can't-block
1061 * assumption. An interrupt thread does not bump
1062 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1063 * primarily until we can fix softupdate's assumptions about free().
1066 --type->ks_inuse[gd->gd_cpuid];
1067 type->ks_memuse[gd->gd_cpuid] -= size;
1068 if (mycpu->gd_intr_nesting_level ||
1069 (gd->gd_curthread->td_flags & TDF_INTTHREAD))
1071 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1073 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1074 z->z_Next = slgd->FreeOvZones;
1075 z->z_ChunkSize = size;
1076 slgd->FreeOvZones = z;
1080 logmemory(free_ovsz, ptr, type, size, 0);
1081 kmem_slab_free(ptr, size); /* may block */
1082 atomic_add_int(&ZoneBigAlloc, -(int)size / 1024);
1084 logmemory_quick(free_end);
1089 * Zone case. Figure out the zone based on the fact that it is
1092 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1095 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1098 * If we do not own the zone then use atomic ops to free to the
1099 * remote cpu linked list and notify the target zone using a
1102 * The target zone cannot be deallocated while we own a chunk of it,
1103 * so the zone header's storage is stable until the very moment
1104 * we adjust z_RChunks. After that we cannot safely dereference (z).
1106 * (no critical section needed)
1108 if (z->z_CpuGd != gd) {
1111 * Making these adjustments now allow us to avoid passing (type)
1112 * to the remote cpu. Note that ks_inuse/ks_memuse is being
1113 * adjusted on OUR cpu, not the zone cpu, but it should all still
1114 * sum up properly and cancel out.
1117 --type->ks_inuse[gd->gd_cpuid];
1118 type->ks_memuse[gd->gd_cpuid] -= z->z_ChunkSize;
1122 * WARNING! This code competes with other cpus. Once we
1123 * successfully link the chunk to RChunks the remote
1124 * cpu can rip z's storage out from under us.
1126 * Bumping RCount prevents z's storage from getting
1129 rsignal = z->z_RSignal;
1132 atomic_add_int(&z->z_RCount, 1);
1136 bchunk = z->z_RChunks;
1138 chunk->c_Next = bchunk;
1141 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1146 * We have to signal the remote cpu if our actions will cause
1147 * the remote zone to be placed back on ZoneAry so it can
1148 * move the zone back on.
1150 * We only need to deal with NULL->non-NULL RChunk transitions
1151 * and only if z_RSignal is set. We interlock by reading rsignal
1152 * before adding our chunk to RChunks. This should result in
1153 * virtually no IPI traffic.
1155 * We can use a passive IPI to reduce overhead even further.
1157 if (bchunk == NULL && rsignal) {
1158 logmemory(free_request, ptr, type, z->z_ChunkSize, 0);
1159 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
1160 /* z can get ripped out from under us from this point on */
1161 } else if (rsignal) {
1162 atomic_subtract_int(&z->z_RCount, 1);
1163 /* z can get ripped out from under us from this point on */
1166 panic("Corrupt SLZone");
1168 logmemory_quick(free_end);
1175 logmemory(free_chunk, ptr, type, z->z_ChunkSize, 0);
1179 chunk_mark_free(z, chunk);
1182 * Put weird data into the memory to detect modifications after freeing,
1183 * illegal pointer use after freeing (we should fault on the odd address),
1184 * and so forth. XXX needs more work, see the old malloc code.
1187 if (z->z_ChunkSize < sizeof(weirdary))
1188 bcopy(weirdary, chunk, z->z_ChunkSize);
1190 bcopy(weirdary, chunk, sizeof(weirdary));
1194 * Add this free non-zero'd chunk to a linked list for reuse. Add
1195 * to the front of the linked list so it is more likely to be
1196 * reallocated, since it is already in our L1 cache.
1199 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
1200 panic("BADFREE %p", chunk);
1202 chunk->c_Next = z->z_LChunks;
1203 z->z_LChunks = chunk;
1204 if (chunk->c_Next == NULL)
1205 z->z_LChunksp = &chunk->c_Next;
1208 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
1213 * Bump the number of free chunks. If it becomes non-zero the zone
1214 * must be added back onto the appropriate list.
1216 if (z->z_NFree++ == 0) {
1217 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1218 slgd->ZoneAry[z->z_ZoneIndex] = z;
1221 --type->ks_inuse[z->z_Cpu];
1222 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
1225 * If the zone becomes totally free, and there are other zones we
1226 * can allocate from, move this zone to the FreeZones list. Since
1227 * this code can be called from an IPI callback, do *NOT* try to mess
1228 * with kernel_map here. Hysteresis will be performed at malloc() time.
1230 if (z->z_NFree == z->z_NMax &&
1231 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1237 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
1241 z->z_Next = slgd->FreeZones;
1242 slgd->FreeZones = z;
1247 logmemory_quick(free_end);
1251 #if defined(INVARIANTS)
1254 * Helper routines for sanity checks
1258 chunk_mark_allocated(SLZone *z, void *chunk)
1260 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1263 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1264 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1265 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
1266 bitptr = &z->z_Bitmap[bitdex >> 5];
1268 KASSERT((*bitptr & (1 << bitdex)) == 0,
1269 ("memory chunk %p is already allocated!", chunk));
1270 *bitptr |= 1 << bitdex;
1275 chunk_mark_free(SLZone *z, void *chunk)
1277 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1280 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1281 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1282 ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
1283 bitptr = &z->z_Bitmap[bitdex >> 5];
1285 KASSERT((*bitptr & (1 << bitdex)) != 0,
1286 ("memory chunk %p is already free!", chunk));
1287 *bitptr &= ~(1 << bitdex);
1295 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1296 * specified alignment. M_* flags are expected in the flags field.
1298 * Alignment must be a multiple of PAGE_SIZE.
1300 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1301 * but when we move zalloc() over to use this function as its backend
1302 * we will have to switch to kreserve/krelease and call reserve(0)
1303 * after the new space is made available.
1305 * Interrupt code which has preempted other code is not allowed to
1306 * use PQ_CACHE pages. However, if an interrupt thread is run
1307 * non-preemptively or blocks and then runs non-preemptively, then
1308 * it is free to use PQ_CACHE pages.
1311 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1315 int count, vmflags, base_vmflags;
1318 size = round_page(size);
1319 addr = vm_map_min(&kernel_map);
1322 * Reserve properly aligned space from kernel_map. RNOWAIT allocations
1325 if (flags & M_RNOWAIT) {
1326 if (lwkt_trytoken(&vm_token) == 0)
1329 lwkt_gettoken(&vm_token);
1331 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1333 vm_map_lock(&kernel_map);
1334 if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) {
1335 vm_map_unlock(&kernel_map);
1336 if ((flags & M_NULLOK) == 0)
1337 panic("kmem_slab_alloc(): kernel_map ran out of space!");
1338 vm_map_entry_release(count);
1340 lwkt_reltoken(&vm_token);
1345 * kernel_object maps 1:1 to kernel_map.
1347 vm_object_reference(&kernel_object);
1348 vm_map_insert(&kernel_map, &count,
1349 &kernel_object, addr, addr, addr + size,
1351 VM_PROT_ALL, VM_PROT_ALL,
1358 base_vmflags |= VM_ALLOC_ZERO;
1359 if (flags & M_USE_RESERVE)
1360 base_vmflags |= VM_ALLOC_SYSTEM;
1361 if (flags & M_USE_INTERRUPT_RESERVE)
1362 base_vmflags |= VM_ALLOC_INTERRUPT;
1363 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1364 panic("kmem_slab_alloc: bad flags %08x (%p)",
1365 flags, ((int **)&size)[-1]);
1370 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
1372 for (i = 0; i < size; i += PAGE_SIZE) {
1376 * VM_ALLOC_NORMAL can only be set if we are not preempting.
1378 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1379 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1380 * implied in this case), though I'm not sure if we really need to
1383 vmflags = base_vmflags;
1384 if (flags & M_WAITOK) {
1385 if (td->td_preempted)
1386 vmflags |= VM_ALLOC_SYSTEM;
1388 vmflags |= VM_ALLOC_NORMAL;
1391 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
1394 * If the allocation failed we either return NULL or we retry.
1396 * If M_WAITOK is specified we wait for more memory and retry.
1397 * If M_WAITOK is specified from a preemption we yield instead of
1398 * wait. Livelock will not occur because the interrupt thread
1399 * will not be preempting anyone the second time around after the
1403 if (flags & M_WAITOK) {
1404 if (td->td_preempted) {
1405 vm_map_unlock(&kernel_map);
1407 vm_map_lock(&kernel_map);
1409 vm_map_unlock(&kernel_map);
1411 vm_map_lock(&kernel_map);
1413 i -= PAGE_SIZE; /* retry */
1418 * We were unable to recover, cleanup and return NULL
1420 * (vm_token already held)
1424 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1425 /* page should already be busy */
1428 vm_map_delete(&kernel_map, addr, addr + size, &count);
1429 vm_map_unlock(&kernel_map);
1430 vm_map_entry_release(count);
1432 lwkt_reltoken(&vm_token);
1440 * Mark the map entry as non-pageable using a routine that allows us to
1441 * populate the underlying pages.
1443 * The pages were busied by the allocations above.
1445 vm_map_set_wired_quick(&kernel_map, addr, size, &count);
1449 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1451 lwkt_gettoken(&vm_token);
1452 for (i = 0; i < size; i += PAGE_SIZE) {
1455 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1456 m->valid = VM_PAGE_BITS_ALL;
1457 /* page should already be busy */
1460 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
1461 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1462 bzero((char *)addr + i, PAGE_SIZE);
1463 vm_page_flag_clear(m, PG_ZERO);
1464 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1465 vm_page_flag_set(m, PG_REFERENCED);
1467 lwkt_reltoken(&vm_token);
1468 vm_map_unlock(&kernel_map);
1469 vm_map_entry_release(count);
1470 lwkt_reltoken(&vm_token);
1471 return((void *)addr);
1478 kmem_slab_free(void *ptr, vm_size_t size)
1481 lwkt_gettoken(&vm_token);
1482 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
1483 lwkt_reltoken(&vm_token);