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 arysize(ary) (sizeof(ary)/sizeof((ary)[0]))
122 #define btokup(z) (&pmap_kvtom((vm_offset_t)(z))->ku_pagecnt)
124 #define MEMORY_STRING "ptr=%p type=%p size=%d flags=%04x"
125 #define MEMORY_ARG_SIZE (sizeof(void *) * 2 + sizeof(unsigned long) + \
128 #if !defined(KTR_MEMORY)
129 #define KTR_MEMORY KTR_ALL
131 KTR_INFO_MASTER(memory);
132 KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin", 0);
133 KTR_INFO(KTR_MEMORY, memory, malloc_end, 1, MEMORY_STRING, MEMORY_ARG_SIZE);
134 KTR_INFO(KTR_MEMORY, memory, free_zero, 2, MEMORY_STRING, MEMORY_ARG_SIZE);
135 KTR_INFO(KTR_MEMORY, memory, free_ovsz, 3, MEMORY_STRING, MEMORY_ARG_SIZE);
136 KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 4, MEMORY_STRING, MEMORY_ARG_SIZE);
137 KTR_INFO(KTR_MEMORY, memory, free_chunk, 5, MEMORY_STRING, MEMORY_ARG_SIZE);
139 KTR_INFO(KTR_MEMORY, memory, free_request, 6, MEMORY_STRING, MEMORY_ARG_SIZE);
140 KTR_INFO(KTR_MEMORY, memory, free_rem_beg, 7, MEMORY_STRING, MEMORY_ARG_SIZE);
141 KTR_INFO(KTR_MEMORY, memory, free_rem_end, 8, MEMORY_STRING, MEMORY_ARG_SIZE);
143 KTR_INFO(KTR_MEMORY, memory, free_beg, 9, "free begin", 0);
144 KTR_INFO(KTR_MEMORY, memory, free_end, 10, "free end", 0);
146 #define logmemory(name, ptr, type, size, flags) \
147 KTR_LOG(memory_ ## name, ptr, type, size, flags)
148 #define logmemory_quick(name) \
149 KTR_LOG(memory_ ## name)
152 * Fixed globals (not per-cpu)
155 static int ZoneLimit;
156 static int ZonePageCount;
157 static uintptr_t ZoneMask;
158 static int ZoneBigAlloc; /* in KB */
159 static int ZoneGenAlloc; /* in KB */
160 struct malloc_type *kmemstatistics; /* exported to vmstat */
161 static int32_t weirdary[16];
163 static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
164 static void kmem_slab_free(void *ptr, vm_size_t bytes);
166 #if defined(INVARIANTS)
167 static void chunk_mark_allocated(SLZone *z, void *chunk);
168 static void chunk_mark_free(SLZone *z, void *chunk);
170 #define chunk_mark_allocated(z, chunk)
171 #define chunk_mark_free(z, chunk)
175 * Misc constants. Note that allocations that are exact multiples of
176 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
177 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
179 #define MIN_CHUNK_SIZE 8 /* in bytes */
180 #define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
181 #define ZONE_RELS_THRESH 2 /* threshold number of zones */
182 #define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
185 * The WEIRD_ADDR is used as known text to copy into free objects to
186 * try to create deterministic failure cases if the data is accessed after
189 #define WEIRD_ADDR 0xdeadc0de
190 #define MAX_COPY sizeof(weirdary)
191 #define ZERO_LENGTH_PTR ((void *)-8)
194 * Misc global malloc buckets
197 MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
198 MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
199 MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
201 MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
202 MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
205 * Initialize the slab memory allocator. We have to choose a zone size based
206 * on available physical memory. We choose a zone side which is approximately
207 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
208 * 128K. The zone size is limited to the bounds set in slaballoc.h
209 * (typically 32K min, 128K max).
211 static void kmeminit(void *dummy);
215 SYSINIT(kmem, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL)
219 * If enabled any memory allocated without M_ZERO is initialized to -1.
221 static int use_malloc_pattern;
222 SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
223 &use_malloc_pattern, 0, "");
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, "");
230 kmeminit(void *dummy)
236 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
237 if (limsize > KvaSize)
240 usesize = (int)(limsize / 1024); /* convert to KB */
242 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
243 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
245 ZoneLimit = ZoneSize / 4;
246 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
247 ZoneLimit = ZALLOC_ZONE_LIMIT;
248 ZoneMask = ~(uintptr_t)(ZoneSize - 1);
249 ZonePageCount = ZoneSize / PAGE_SIZE;
251 for (i = 0; i < arysize(weirdary); ++i)
252 weirdary[i] = WEIRD_ADDR;
254 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
257 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
261 * Initialize a malloc type tracking structure.
264 malloc_init(void *data)
266 struct malloc_type *type = data;
269 if (type->ks_magic != M_MAGIC)
270 panic("malloc type lacks magic");
272 if (type->ks_limit != 0)
275 if (vmstats.v_page_count == 0)
276 panic("malloc_init not allowed before vm init");
278 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
279 if (limsize > KvaSize)
281 type->ks_limit = limsize / 10;
283 type->ks_next = kmemstatistics;
284 kmemstatistics = type;
288 malloc_uninit(void *data)
290 struct malloc_type *type = data;
291 struct malloc_type *t;
297 if (type->ks_magic != M_MAGIC)
298 panic("malloc type lacks magic");
300 if (vmstats.v_page_count == 0)
301 panic("malloc_uninit not allowed before vm init");
303 if (type->ks_limit == 0)
304 panic("malloc_uninit on uninitialized type");
307 /* Make sure that all pending kfree()s are finished. */
308 lwkt_synchronize_ipiqs("muninit");
313 * memuse is only correct in aggregation. Due to memory being allocated
314 * on one cpu and freed on another individual array entries may be
315 * negative or positive (canceling each other out).
317 for (i = ttl = 0; i < ncpus; ++i)
318 ttl += type->ks_memuse[i];
320 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
321 ttl, type->ks_shortdesc, i);
324 if (type == kmemstatistics) {
325 kmemstatistics = type->ks_next;
327 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
328 if (t->ks_next == type) {
329 t->ks_next = type->ks_next;
334 type->ks_next = NULL;
339 * Increase the kmalloc pool limit for the specified pool. No changes
340 * are the made if the pool would shrink.
343 kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
345 if (type->ks_limit == 0)
349 if (type->ks_limit < bytes)
350 type->ks_limit = bytes;
354 * Dynamically create a malloc pool. This function is a NOP if *typep is
358 kmalloc_create(struct malloc_type **typep, const char *descr)
360 struct malloc_type *type;
362 if (*typep == NULL) {
363 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
364 type->ks_magic = M_MAGIC;
365 type->ks_shortdesc = descr;
372 * Destroy a dynamically created malloc pool. This function is a NOP if
373 * the pool has already been destroyed.
376 kmalloc_destroy(struct malloc_type **typep)
378 if (*typep != NULL) {
379 malloc_uninit(*typep);
380 kfree(*typep, M_TEMP);
386 * Calculate the zone index for the allocation request size and set the
387 * allocation request size to that particular zone's chunk size.
390 zoneindex(unsigned long *bytes)
392 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
394 *bytes = n = (n + 7) & ~7;
395 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
398 *bytes = n = (n + 15) & ~15;
403 *bytes = n = (n + 31) & ~31;
407 *bytes = n = (n + 63) & ~63;
411 *bytes = n = (n + 127) & ~127;
412 return(n / 128 + 31);
415 *bytes = n = (n + 255) & ~255;
416 return(n / 256 + 39);
418 *bytes = n = (n + 511) & ~511;
419 return(n / 512 + 47);
421 #if ZALLOC_ZONE_LIMIT > 8192
423 *bytes = n = (n + 1023) & ~1023;
424 return(n / 1024 + 55);
427 #if ZALLOC_ZONE_LIMIT > 16384
429 *bytes = n = (n + 2047) & ~2047;
430 return(n / 2048 + 63);
433 panic("Unexpected byte count %d", n);
438 * kmalloc() (SLAB ALLOCATOR)
440 * Allocate memory via the slab allocator. If the request is too large,
441 * or if it page-aligned beyond a certain size, we fall back to the
442 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
443 * &SlabMisc if you don't care.
445 * M_RNOWAIT - don't block.
446 * M_NULLOK - return NULL instead of blocking.
447 * M_ZERO - zero the returned memory.
448 * M_USE_RESERVE - allow greater drawdown of the free list
449 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
454 kmalloc(unsigned long size, struct malloc_type *type, int flags)
462 struct globaldata *gd;
468 logmemory_quick(malloc_beg);
473 * XXX silly to have this in the critical path.
475 if (type->ks_limit == 0) {
477 if (type->ks_limit == 0)
484 * Handle the case where the limit is reached. Panic if we can't return
485 * NULL. The original malloc code looped, but this tended to
486 * simply deadlock the computer.
488 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
489 * to determine if a more complete limit check should be done. The
490 * actual memory use is tracked via ks_memuse[cpu].
492 while (type->ks_loosememuse >= type->ks_limit) {
496 for (i = ttl = 0; i < ncpus; ++i)
497 ttl += type->ks_memuse[i];
498 type->ks_loosememuse = ttl; /* not MP synchronized */
499 if ((ssize_t)ttl < 0) /* deal with occassional race */
501 if (ttl >= type->ks_limit) {
502 if (flags & M_NULLOK) {
503 logmemory(malloc_end, NULL, type, size, flags);
506 panic("%s: malloc limit exceeded", type->ks_shortdesc);
511 * Handle the degenerate size == 0 case. Yes, this does happen.
512 * Return a special pointer. This is to maintain compatibility with
513 * the original malloc implementation. Certain devices, such as the
514 * adaptec driver, not only allocate 0 bytes, they check for NULL and
515 * also realloc() later on. Joy.
518 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
519 return(ZERO_LENGTH_PTR);
523 * Handle hysteresis from prior frees here in malloc(). We cannot
524 * safely manipulate the kernel_map in free() due to free() possibly
525 * being called via an IPI message or from sensitive interrupt code.
527 * NOTE: ku_pagecnt must be cleared before we free the slab or we
528 * might race another cpu allocating the kva and setting
531 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_RNOWAIT) == 0) {
533 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */
537 slgd->FreeZones = z->z_Next;
541 kmem_slab_free(z, ZoneSize); /* may block */
542 atomic_add_int(&ZoneGenAlloc, -(int)ZoneSize / 1024);
548 * XXX handle oversized frees that were queued from kfree().
550 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) {
552 if ((z = slgd->FreeOvZones) != NULL) {
555 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
556 slgd->FreeOvZones = z->z_Next;
557 tsize = z->z_ChunkSize;
558 kmem_slab_free(z, tsize); /* may block */
559 atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024);
565 * Handle large allocations directly. There should not be very many of
566 * these so performance is not a big issue.
568 * The backend allocator is pretty nasty on a SMP system. Use the
569 * slab allocator for one and two page-sized chunks even though we lose
570 * some efficiency. XXX maybe fix mmio and the elf loader instead.
572 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
575 size = round_page(size);
576 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
578 logmemory(malloc_end, NULL, type, size, flags);
581 atomic_add_int(&ZoneBigAlloc, (int)size / 1024);
582 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
583 flags |= M_PASSIVE_ZERO;
585 *kup = size / PAGE_SIZE;
591 * Attempt to allocate out of an existing zone. First try the free list,
592 * then allocate out of unallocated space. If we find a good zone move
593 * it to the head of the list so later allocations find it quickly
594 * (we might have thousands of zones in the list).
596 * Note: zoneindex() will panic of size is too large.
598 zi = zoneindex(&size);
599 KKASSERT(zi < NZONES);
602 if ((z = slgd->ZoneAry[zi]) != NULL) {
604 * Locate a chunk - we have to have at least one. If this is the
605 * last chunk go ahead and do the work to retrieve chunks freed
606 * from remote cpus, and if the zone is still empty move it off
609 if (--z->z_NFree <= 0) {
610 KKASSERT(z->z_NFree == 0);
614 * WARNING! This code competes with other cpus. It is ok
615 * for us to not drain RChunks here but we might as well, and
616 * it is ok if more accumulate after we're done.
618 * Set RSignal before pulling rchunks off, indicating that we
619 * will be moving ourselves off of the ZoneAry. Remote ends will
620 * read RSignal before putting rchunks on thus interlocking
621 * their IPI signaling.
623 if (z->z_RChunks == NULL)
624 atomic_swap_int(&z->z_RSignal, 1);
626 while ((bchunk = z->z_RChunks) != NULL) {
628 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
629 *z->z_LChunksp = bchunk;
631 chunk_mark_free(z, bchunk);
632 z->z_LChunksp = &bchunk->c_Next;
633 bchunk = bchunk->c_Next;
641 * Remove from the zone list if no free chunks remain.
644 if (z->z_NFree == 0) {
645 slgd->ZoneAry[zi] = z->z_Next;
653 * Fast path, we have chunks available in z_LChunks.
655 chunk = z->z_LChunks;
657 chunk_mark_allocated(z, chunk);
658 z->z_LChunks = chunk->c_Next;
659 if (z->z_LChunks == NULL)
660 z->z_LChunksp = &z->z_LChunks;
665 * No chunks are available in LChunks, the free chunk MUST be
666 * in the never-before-used memory area, controlled by UIndex.
668 * The consequences are very serious if our zone got corrupted so
669 * we use an explicit panic rather than a KASSERT.
671 if (z->z_UIndex + 1 != z->z_NMax)
676 if (z->z_UIndex == z->z_UEndIndex)
677 panic("slaballoc: corrupted zone");
679 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
680 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
682 flags |= M_PASSIVE_ZERO;
684 chunk_mark_allocated(z, chunk);
689 * If all zones are exhausted we need to allocate a new zone for this
690 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
691 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
692 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
693 * we do not pre-zero it because we do not want to mess up the L1 cache.
695 * At least one subsystem, the tty code (see CROUND) expects power-of-2
696 * allocations to be power-of-2 aligned. We maintain compatibility by
697 * adjusting the base offset below.
703 if ((z = slgd->FreeZones) != NULL) {
704 slgd->FreeZones = z->z_Next;
706 bzero(z, sizeof(SLZone));
707 z->z_Flags |= SLZF_UNOTZEROD;
709 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
712 atomic_add_int(&ZoneGenAlloc, (int)ZoneSize / 1024);
716 * How big is the base structure?
718 #if defined(INVARIANTS)
720 * Make room for z_Bitmap. An exact calculation is somewhat more
721 * complicated so don't make an exact calculation.
723 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
724 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
726 off = sizeof(SLZone);
730 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
731 * Otherwise just 8-byte align the data.
733 if ((size | (size - 1)) + 1 == (size << 1))
734 off = (off + size - 1) & ~(size - 1);
736 off = (off + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
737 z->z_Magic = ZALLOC_SLAB_MAGIC;
739 z->z_NMax = (ZoneSize - off) / size;
740 z->z_NFree = z->z_NMax - 1;
741 z->z_BasePtr = (char *)z + off;
742 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
743 z->z_ChunkSize = size;
745 z->z_Cpu = gd->gd_cpuid;
746 z->z_LChunksp = &z->z_LChunks;
747 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
748 z->z_Next = slgd->ZoneAry[zi];
749 slgd->ZoneAry[zi] = z;
750 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
751 flags &= ~M_ZERO; /* already zero'd */
752 flags |= M_PASSIVE_ZERO;
755 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */
756 chunk_mark_allocated(z, chunk);
759 * Slide the base index for initial allocations out of the next
760 * zone we create so we do not over-weight the lower part of the
763 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
764 & (ZALLOC_MAX_ZONE_SIZE - 1);
768 ++type->ks_inuse[gd->gd_cpuid];
769 type->ks_memuse[gd->gd_cpuid] += size;
770 type->ks_loosememuse += size; /* not MP synchronized */
776 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
777 if (use_malloc_pattern) {
778 for (i = 0; i < size; i += sizeof(int)) {
779 *(int *)((char *)chunk + i) = -1;
782 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
785 logmemory(malloc_end, chunk, type, size, flags);
789 logmemory(malloc_end, NULL, type, size, flags);
794 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
796 * Generally speaking this routine is not called very often and we do
797 * not attempt to optimize it beyond reusing the same pointer if the
798 * new size fits within the chunking of the old pointer's zone.
801 krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
808 KKASSERT((flags & M_ZERO) == 0); /* not supported */
810 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
811 return(kmalloc(size, type, flags));
818 * Handle oversized allocations. XXX we really should require that a
819 * size be passed to free() instead of this nonsense.
823 osize = *kup << PAGE_SHIFT;
824 if (osize == round_page(size))
826 if ((nptr = kmalloc(size, type, flags)) == NULL)
828 bcopy(ptr, nptr, min(size, osize));
834 * Get the original allocation's zone. If the new request winds up
835 * using the same chunk size we do not have to do anything.
837 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
840 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
843 * Allocate memory for the new request size. Note that zoneindex has
844 * already adjusted the request size to the appropriate chunk size, which
845 * should optimize our bcopy(). Then copy and return the new pointer.
847 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
848 * necessary align the result.
850 * We can only zoneindex (to align size to the chunk size) if the new
851 * size is not too large.
853 if (size < ZoneLimit) {
855 if (z->z_ChunkSize == size)
858 if ((nptr = kmalloc(size, type, flags)) == NULL)
860 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
866 * Return the kmalloc limit for this type, in bytes.
869 kmalloc_limit(struct malloc_type *type)
871 if (type->ks_limit == 0) {
873 if (type->ks_limit == 0)
877 return(type->ks_limit);
881 * Allocate a copy of the specified string.
883 * (MP SAFE) (MAY BLOCK)
886 kstrdup(const char *str, struct malloc_type *type)
888 int zlen; /* length inclusive of terminating NUL */
893 zlen = strlen(str) + 1;
894 nstr = kmalloc(zlen, type, M_WAITOK);
895 bcopy(str, nstr, zlen);
901 * Notify our cpu that a remote cpu has freed some chunks in a zone that
902 * we own. RCount will be bumped so the memory should be good, but validate
907 kfree_remote(void *ptr)
915 slgd = &mycpu->gd_slab;
918 KKASSERT(*kup == -((int)mycpuid + 1));
919 KKASSERT(z->z_RCount > 0);
920 atomic_subtract_int(&z->z_RCount, 1);
922 logmemory(free_rem_beg, z, NULL, 0, 0);
923 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
924 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
928 * Indicate that we will no longer be off of the ZoneAry by
935 * Atomically extract the bchunks list and then process it back
936 * into the lchunks list. We want to append our bchunks to the
937 * lchunks list and not prepend since we likely do not have
938 * cache mastership of the related data (not that it helps since
939 * we are using c_Next).
941 while ((bchunk = z->z_RChunks) != NULL) {
943 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
944 *z->z_LChunksp = bchunk;
946 chunk_mark_free(z, bchunk);
947 z->z_LChunksp = &bchunk->c_Next;
948 bchunk = bchunk->c_Next;
954 if (z->z_NFree && nfree == 0) {
955 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
956 slgd->ZoneAry[z->z_ZoneIndex] = z;
960 * If the zone becomes totally free, and there are other zones we
961 * can allocate from, move this zone to the FreeZones list. Since
962 * this code can be called from an IPI callback, do *NOT* try to mess
963 * with kernel_map here. Hysteresis will be performed at malloc() time.
965 * Do not move the zone if there is an IPI inflight, otherwise MP
966 * races can result in our free_remote code accessing a destroyed
969 if (z->z_NFree == z->z_NMax &&
970 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
976 for (pz = &slgd->ZoneAry[z->z_ZoneIndex];
978 pz = &(*pz)->z_Next) {
983 z->z_Next = slgd->FreeZones;
989 logmemory(free_rem_end, z, bchunk, 0, 0);
995 * free (SLAB ALLOCATOR)
997 * Free a memory block previously allocated by malloc. Note that we do not
998 * attempt to update ks_loosememuse as MP races could prevent us from
999 * checking memory limits in malloc.
1004 kfree(void *ptr, struct malloc_type *type)
1009 struct globaldata *gd;
1017 logmemory_quick(free_beg);
1019 slgd = &gd->gd_slab;
1022 panic("trying to free NULL pointer");
1025 * Handle special 0-byte allocations
1027 if (ptr == ZERO_LENGTH_PTR) {
1028 logmemory(free_zero, ptr, type, -1, 0);
1029 logmemory_quick(free_end);
1034 * Panic on bad malloc type
1036 if (type->ks_magic != M_MAGIC)
1037 panic("free: malloc type lacks magic");
1040 * Handle oversized allocations. XXX we really should require that a
1041 * size be passed to free() instead of this nonsense.
1043 * This code is never called via an ipi.
1047 size = *kup << PAGE_SHIFT;
1050 KKASSERT(sizeof(weirdary) <= size);
1051 bcopy(weirdary, ptr, sizeof(weirdary));
1054 * NOTE: For oversized allocations we do not record the
1055 * originating cpu. It gets freed on the cpu calling
1056 * kfree(). The statistics are in aggregate.
1058 * note: XXX we have still inherited the interrupts-can't-block
1059 * assumption. An interrupt thread does not bump
1060 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1061 * primarily until we can fix softupdate's assumptions about free().
1064 --type->ks_inuse[gd->gd_cpuid];
1065 type->ks_memuse[gd->gd_cpuid] -= size;
1066 if (mycpu->gd_intr_nesting_level ||
1067 (gd->gd_curthread->td_flags & TDF_INTTHREAD))
1069 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1071 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1072 z->z_Next = slgd->FreeOvZones;
1073 z->z_ChunkSize = size;
1074 slgd->FreeOvZones = z;
1078 logmemory(free_ovsz, ptr, type, size, 0);
1079 kmem_slab_free(ptr, size); /* may block */
1080 atomic_add_int(&ZoneBigAlloc, -(int)size / 1024);
1082 logmemory_quick(free_end);
1087 * Zone case. Figure out the zone based on the fact that it is
1090 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1093 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1096 * If we do not own the zone then use atomic ops to free to the
1097 * remote cpu linked list and notify the target zone using a
1100 * The target zone cannot be deallocated while we own a chunk of it,
1101 * so the zone header's storage is stable until the very moment
1102 * we adjust z_RChunks. After that we cannot safely dereference (z).
1104 * (no critical section needed)
1106 if (z->z_CpuGd != gd) {
1109 * Making these adjustments now allow us to avoid passing (type)
1110 * to the remote cpu. Note that ks_inuse/ks_memuse is being
1111 * adjusted on OUR cpu, not the zone cpu, but it should all still
1112 * sum up properly and cancel out.
1115 --type->ks_inuse[gd->gd_cpuid];
1116 type->ks_memuse[gd->gd_cpuid] -= z->z_ChunkSize;
1120 * WARNING! This code competes with other cpus. Once we
1121 * successfully link the chunk to RChunks the remote
1122 * cpu can rip z's storage out from under us.
1124 * Bumping RCount prevents z's storage from getting
1127 rsignal = z->z_RSignal;
1130 atomic_add_int(&z->z_RCount, 1);
1134 bchunk = z->z_RChunks;
1136 chunk->c_Next = bchunk;
1139 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1144 * We have to signal the remote cpu if our actions will cause
1145 * the remote zone to be placed back on ZoneAry so it can
1146 * move the zone back on.
1148 * We only need to deal with NULL->non-NULL RChunk transitions
1149 * and only if z_RSignal is set. We interlock by reading rsignal
1150 * before adding our chunk to RChunks. This should result in
1151 * virtually no IPI traffic.
1153 * We can use a passive IPI to reduce overhead even further.
1155 if (bchunk == NULL && rsignal) {
1156 logmemory(free_request, ptr, type, z->z_ChunkSize, 0);
1157 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
1158 /* z can get ripped out from under us from this point on */
1159 } else if (rsignal) {
1160 atomic_subtract_int(&z->z_RCount, 1);
1161 /* z can get ripped out from under us from this point on */
1164 panic("Corrupt SLZone");
1166 logmemory_quick(free_end);
1173 logmemory(free_chunk, ptr, type, z->z_ChunkSize, 0);
1177 chunk_mark_free(z, chunk);
1180 * Put weird data into the memory to detect modifications after freeing,
1181 * illegal pointer use after freeing (we should fault on the odd address),
1182 * and so forth. XXX needs more work, see the old malloc code.
1185 if (z->z_ChunkSize < sizeof(weirdary))
1186 bcopy(weirdary, chunk, z->z_ChunkSize);
1188 bcopy(weirdary, chunk, sizeof(weirdary));
1192 * Add this free non-zero'd chunk to a linked list for reuse. Add
1193 * to the front of the linked list so it is more likely to be
1194 * reallocated, since it is already in our L1 cache.
1197 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
1198 panic("BADFREE %p", chunk);
1200 chunk->c_Next = z->z_LChunks;
1201 z->z_LChunks = chunk;
1202 if (chunk->c_Next == NULL)
1203 z->z_LChunksp = &chunk->c_Next;
1206 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
1211 * Bump the number of free chunks. If it becomes non-zero the zone
1212 * must be added back onto the appropriate list.
1214 if (z->z_NFree++ == 0) {
1215 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1216 slgd->ZoneAry[z->z_ZoneIndex] = z;
1219 --type->ks_inuse[z->z_Cpu];
1220 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
1223 * If the zone becomes totally free, and there are other zones we
1224 * can allocate from, move this zone to the FreeZones list. Since
1225 * this code can be called from an IPI callback, do *NOT* try to mess
1226 * with kernel_map here. Hysteresis will be performed at malloc() time.
1228 if (z->z_NFree == z->z_NMax &&
1229 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1235 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
1239 z->z_Next = slgd->FreeZones;
1240 slgd->FreeZones = z;
1245 logmemory_quick(free_end);
1249 #if defined(INVARIANTS)
1252 * Helper routines for sanity checks
1256 chunk_mark_allocated(SLZone *z, void *chunk)
1258 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1261 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1262 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1263 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
1264 bitptr = &z->z_Bitmap[bitdex >> 5];
1266 KASSERT((*bitptr & (1 << bitdex)) == 0,
1267 ("memory chunk %p is already allocated!", chunk));
1268 *bitptr |= 1 << bitdex;
1273 chunk_mark_free(SLZone *z, void *chunk)
1275 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1278 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1279 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1280 ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
1281 bitptr = &z->z_Bitmap[bitdex >> 5];
1283 KASSERT((*bitptr & (1 << bitdex)) != 0,
1284 ("memory chunk %p is already free!", chunk));
1285 *bitptr &= ~(1 << bitdex);
1293 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1294 * specified alignment. M_* flags are expected in the flags field.
1296 * Alignment must be a multiple of PAGE_SIZE.
1298 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1299 * but when we move zalloc() over to use this function as its backend
1300 * we will have to switch to kreserve/krelease and call reserve(0)
1301 * after the new space is made available.
1303 * Interrupt code which has preempted other code is not allowed to
1304 * use PQ_CACHE pages. However, if an interrupt thread is run
1305 * non-preemptively or blocks and then runs non-preemptively, then
1306 * it is free to use PQ_CACHE pages.
1309 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1313 int count, vmflags, base_vmflags;
1314 vm_page_t mp[ZALLOC_MAX_ZONE_SIZE / PAGE_SIZE];
1317 size = round_page(size);
1318 addr = vm_map_min(&kernel_map);
1321 * Reserve properly aligned space from kernel_map. RNOWAIT allocations
1324 if (flags & M_RNOWAIT) {
1325 if (lwkt_trytoken(&vm_token) == 0)
1328 lwkt_gettoken(&vm_token);
1330 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1332 vm_map_lock(&kernel_map);
1333 if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) {
1334 vm_map_unlock(&kernel_map);
1335 if ((flags & M_NULLOK) == 0)
1336 panic("kmem_slab_alloc(): kernel_map ran out of space!");
1337 vm_map_entry_release(count);
1339 lwkt_reltoken(&vm_token);
1344 * kernel_object maps 1:1 to kernel_map.
1346 vm_object_reference(&kernel_object);
1347 vm_map_insert(&kernel_map, &count,
1348 &kernel_object, addr, addr, addr + size,
1350 VM_PROT_ALL, VM_PROT_ALL,
1357 base_vmflags |= VM_ALLOC_ZERO;
1358 if (flags & M_USE_RESERVE)
1359 base_vmflags |= VM_ALLOC_SYSTEM;
1360 if (flags & M_USE_INTERRUPT_RESERVE)
1361 base_vmflags |= VM_ALLOC_INTERRUPT;
1362 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1363 panic("kmem_slab_alloc: bad flags %08x (%p)",
1364 flags, ((int **)&size)[-1]);
1369 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
1371 for (i = 0; i < size; i += PAGE_SIZE) {
1375 * VM_ALLOC_NORMAL can only be set if we are not preempting.
1377 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1378 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1379 * implied in this case), though I'm not sure if we really need to
1382 vmflags = base_vmflags;
1383 if (flags & M_WAITOK) {
1384 if (td->td_preempted)
1385 vmflags |= VM_ALLOC_SYSTEM;
1387 vmflags |= VM_ALLOC_NORMAL;
1390 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
1391 if ((i / PAGE_SIZE) < (sizeof(mp) / sizeof(mp[0])))
1392 mp[i / PAGE_SIZE] = m;
1395 * If the allocation failed we either return NULL or we retry.
1397 * If M_WAITOK is specified we wait for more memory and retry.
1398 * If M_WAITOK is specified from a preemption we yield instead of
1399 * wait. Livelock will not occur because the interrupt thread
1400 * will not be preempting anyone the second time around after the
1404 if (flags & M_WAITOK) {
1405 if (td->td_preempted) {
1406 vm_map_unlock(&kernel_map);
1408 vm_map_lock(&kernel_map);
1410 vm_map_unlock(&kernel_map);
1412 vm_map_lock(&kernel_map);
1414 i -= PAGE_SIZE; /* retry */
1419 * We were unable to recover, cleanup and return NULL
1421 * (vm_token already held)
1425 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1426 /* page should already be busy */
1429 vm_map_delete(&kernel_map, addr, addr + size, &count);
1430 vm_map_unlock(&kernel_map);
1431 vm_map_entry_release(count);
1433 lwkt_reltoken(&vm_token);
1441 * Mark the map entry as non-pageable using a routine that allows us to
1442 * populate the underlying pages.
1444 * The pages were busied by the allocations above.
1446 vm_map_set_wired_quick(&kernel_map, addr, size, &count);
1450 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1452 for (i = 0; i < size; i += PAGE_SIZE) {
1455 if ((i / PAGE_SIZE) < (sizeof(mp) / sizeof(mp[0])))
1456 m = mp[i / PAGE_SIZE];
1458 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1459 m->valid = VM_PAGE_BITS_ALL;
1460 /* page should already be busy */
1463 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
1464 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1465 bzero((char *)addr + i, PAGE_SIZE);
1466 vm_page_flag_clear(m, PG_ZERO);
1467 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1468 vm_page_flag_set(m, PG_REFERENCED);
1470 vm_map_unlock(&kernel_map);
1471 vm_map_entry_release(count);
1472 lwkt_reltoken(&vm_token);
1473 return((void *)addr);
1480 kmem_slab_free(void *ptr, vm_size_t size)
1483 lwkt_gettoken(&vm_token);
1484 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
1485 lwkt_reltoken(&vm_token);