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 * Alignment properties:
82 * - All power-of-2 sized allocations are power-of-2 aligned.
83 * - Allocations with M_POWEROF2 are power-of-2 aligned on the nearest
84 * power-of-2 round up of 'size'.
85 * - Non-power-of-2 sized allocations are zone chunk size aligned (see the
86 * above table 'Chunking' column).
88 * API REQUIREMENTS AND SIDE EFFECTS
90 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we
91 * have remained compatible with the following API requirements:
93 * + malloc(0) is allowed and returns non-NULL (ahc driver)
94 * + ability to allocate arbitrarily large chunks of memory
99 #include <sys/param.h>
100 #include <sys/systm.h>
101 #include <sys/kernel.h>
102 #include <sys/slaballoc.h>
103 #include <sys/mbuf.h>
104 #include <sys/vmmeter.h>
105 #include <sys/lock.h>
106 #include <sys/thread.h>
107 #include <sys/globaldata.h>
108 #include <sys/sysctl.h>
112 #include <vm/vm_param.h>
113 #include <vm/vm_kern.h>
114 #include <vm/vm_extern.h>
115 #include <vm/vm_object.h>
117 #include <vm/vm_map.h>
118 #include <vm/vm_page.h>
119 #include <vm/vm_pageout.h>
121 #include <machine/cpu.h>
123 #include <sys/thread2.h>
124 #include <vm/vm_page2.h>
126 #define btokup(z) (&pmap_kvtom((vm_offset_t)(z))->ku_pagecnt)
128 #define MEMORY_STRING "ptr=%p type=%p size=%lu flags=%04x"
129 #define MEMORY_ARGS void *ptr, void *type, unsigned long size, int flags
131 #if !defined(KTR_MEMORY)
132 #define KTR_MEMORY KTR_ALL
134 KTR_INFO_MASTER(memory);
135 KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin");
136 KTR_INFO(KTR_MEMORY, memory, malloc_end, 1, MEMORY_STRING, MEMORY_ARGS);
137 KTR_INFO(KTR_MEMORY, memory, free_zero, 2, MEMORY_STRING, MEMORY_ARGS);
138 KTR_INFO(KTR_MEMORY, memory, free_ovsz, 3, MEMORY_STRING, MEMORY_ARGS);
139 KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 4, MEMORY_STRING, MEMORY_ARGS);
140 KTR_INFO(KTR_MEMORY, memory, free_chunk, 5, MEMORY_STRING, MEMORY_ARGS);
141 KTR_INFO(KTR_MEMORY, memory, free_request, 6, MEMORY_STRING, MEMORY_ARGS);
142 KTR_INFO(KTR_MEMORY, memory, free_rem_beg, 7, MEMORY_STRING, MEMORY_ARGS);
143 KTR_INFO(KTR_MEMORY, memory, free_rem_end, 8, MEMORY_STRING, MEMORY_ARGS);
144 KTR_INFO(KTR_MEMORY, memory, free_beg, 9, "free begin");
145 KTR_INFO(KTR_MEMORY, memory, free_end, 10, "free end");
147 #define logmemory(name, ptr, type, size, flags) \
148 KTR_LOG(memory_ ## name, ptr, type, size, flags)
149 #define logmemory_quick(name) \
150 KTR_LOG(memory_ ## name)
153 * Fixed globals (not per-cpu)
156 static int ZoneLimit;
157 static int ZonePageCount;
158 static uintptr_t ZoneMask;
159 static int ZoneBigAlloc; /* in KB */
160 static int ZoneGenAlloc; /* in KB */
161 struct malloc_type *kmemstatistics; /* exported to vmstat */
162 static int32_t weirdary[16];
164 static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
165 static void kmem_slab_free(void *ptr, vm_size_t bytes);
167 #if defined(INVARIANTS)
168 static void chunk_mark_allocated(SLZone *z, void *chunk);
169 static void chunk_mark_free(SLZone *z, void *chunk);
171 #define chunk_mark_allocated(z, chunk)
172 #define chunk_mark_free(z, chunk)
176 * Misc constants. Note that allocations that are exact multiples of
177 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
179 #define ZONE_RELS_THRESH 32 /* threshold number of zones */
182 * The WEIRD_ADDR is used as known text to copy into free objects to
183 * try to create deterministic failure cases if the data is accessed after
186 #define WEIRD_ADDR 0xdeadc0de
187 #define MAX_COPY sizeof(weirdary)
188 #define ZERO_LENGTH_PTR ((void *)-8)
191 * Misc global malloc buckets
194 MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
195 MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
196 MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
198 MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
199 MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
202 * Initialize the slab memory allocator. We have to choose a zone size based
203 * on available physical memory. We choose a zone side which is approximately
204 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
205 * 128K. The zone size is limited to the bounds set in slaballoc.h
206 * (typically 32K min, 128K max).
208 static void kmeminit(void *dummy);
212 SYSINIT(kmem, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL);
216 * If enabled any memory allocated without M_ZERO is initialized to -1.
218 static int use_malloc_pattern;
219 SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
220 &use_malloc_pattern, 0,
221 "Initialize memory to -1 if M_ZERO not specified");
224 static int ZoneRelsThresh = ZONE_RELS_THRESH;
225 SYSCTL_INT(_kern, OID_AUTO, zone_big_alloc, CTLFLAG_RD, &ZoneBigAlloc, 0, "");
226 SYSCTL_INT(_kern, OID_AUTO, zone_gen_alloc, CTLFLAG_RD, &ZoneGenAlloc, 0, "");
227 SYSCTL_INT(_kern, OID_AUTO, zone_cache, CTLFLAG_RW, &ZoneRelsThresh, 0, "");
228 static long SlabsAllocated;
229 static long SlabsFreed;
230 SYSCTL_LONG(_kern, OID_AUTO, slabs_allocated, CTLFLAG_RD,
231 &SlabsAllocated, 0, "");
232 SYSCTL_LONG(_kern, OID_AUTO, slabs_freed, CTLFLAG_RD,
234 static int SlabFreeToTail;
235 SYSCTL_INT(_kern, OID_AUTO, slab_freetotail, CTLFLAG_RW,
236 &SlabFreeToTail, 0, "");
239 * Returns the kernel memory size limit for the purposes of initializing
240 * various subsystem caches. The smaller of available memory and the KVM
241 * memory space is returned.
243 * The size in megabytes is returned.
250 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
251 if (limsize > KvaSize)
253 return (limsize / (1024 * 1024));
257 kmeminit(void *dummy)
263 limsize = kmem_lim_size();
264 usesize = (int)(limsize * 1024); /* convert to KB */
267 * If the machine has a large KVM space and more than 8G of ram,
268 * double the zone release threshold to reduce SMP invalidations.
269 * If more than 16G of ram, do it again.
271 * The BIOS eats a little ram so add some slop. We want 8G worth of
272 * memory sticks to trigger the first adjustment.
274 if (ZoneRelsThresh == ZONE_RELS_THRESH) {
275 if (limsize >= 7 * 1024)
277 if (limsize >= 15 * 1024)
282 * Calculate the zone size. This typically calculates to
283 * ZALLOC_MAX_ZONE_SIZE
285 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
286 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
288 ZoneLimit = ZoneSize / 4;
289 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
290 ZoneLimit = ZALLOC_ZONE_LIMIT;
291 ZoneMask = ~(uintptr_t)(ZoneSize - 1);
292 ZonePageCount = ZoneSize / PAGE_SIZE;
294 for (i = 0; i < NELEM(weirdary); ++i)
295 weirdary[i] = WEIRD_ADDR;
297 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
300 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
304 * (low level) Initialize slab-related elements in the globaldata structure.
306 * Occurs after kmeminit().
309 slab_gdinit(globaldata_t gd)
315 for (i = 0; i < NZONES; ++i)
316 TAILQ_INIT(&slgd->ZoneAry[i]);
317 TAILQ_INIT(&slgd->FreeZones);
318 TAILQ_INIT(&slgd->FreeOvZones);
322 * Initialize a malloc type tracking structure.
325 malloc_init(void *data)
327 struct malloc_type *type = data;
330 if (type->ks_magic != M_MAGIC)
331 panic("malloc type lacks magic");
333 if (type->ks_limit != 0)
336 if (vmstats.v_page_count == 0)
337 panic("malloc_init not allowed before vm init");
339 limsize = kmem_lim_size() * (1024 * 1024);
340 type->ks_limit = limsize / 10;
342 type->ks_next = kmemstatistics;
343 kmemstatistics = type;
347 malloc_uninit(void *data)
349 struct malloc_type *type = data;
350 struct malloc_type *t;
356 if (type->ks_magic != M_MAGIC)
357 panic("malloc type lacks magic");
359 if (vmstats.v_page_count == 0)
360 panic("malloc_uninit not allowed before vm init");
362 if (type->ks_limit == 0)
363 panic("malloc_uninit on uninitialized type");
365 /* Make sure that all pending kfree()s are finished. */
366 lwkt_synchronize_ipiqs("muninit");
370 * memuse is only correct in aggregation. Due to memory being allocated
371 * on one cpu and freed on another individual array entries may be
372 * negative or positive (canceling each other out).
374 for (i = ttl = 0; i < ncpus; ++i)
375 ttl += type->ks_memuse[i];
377 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
378 ttl, type->ks_shortdesc, i);
381 if (type == kmemstatistics) {
382 kmemstatistics = type->ks_next;
384 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
385 if (t->ks_next == type) {
386 t->ks_next = type->ks_next;
391 type->ks_next = NULL;
396 * Increase the kmalloc pool limit for the specified pool. No changes
397 * are the made if the pool would shrink.
400 kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
402 if (type->ks_limit == 0)
406 if (type->ks_limit < bytes)
407 type->ks_limit = bytes;
411 * Dynamically create a malloc pool. This function is a NOP if *typep is
415 kmalloc_create(struct malloc_type **typep, const char *descr)
417 struct malloc_type *type;
419 if (*typep == NULL) {
420 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
421 type->ks_magic = M_MAGIC;
422 type->ks_shortdesc = descr;
429 * Destroy a dynamically created malloc pool. This function is a NOP if
430 * the pool has already been destroyed.
433 kmalloc_destroy(struct malloc_type **typep)
435 if (*typep != NULL) {
436 malloc_uninit(*typep);
437 kfree(*typep, M_TEMP);
443 * Calculate the zone index for the allocation request size and set the
444 * allocation request size to that particular zone's chunk size.
447 zoneindex(unsigned long *bytes, unsigned long *align)
449 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
451 *bytes = n = (n + 7) & ~7;
453 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
456 *bytes = n = (n + 15) & ~15;
462 *bytes = n = (n + 31) & ~31;
467 *bytes = n = (n + 63) & ~63;
472 *bytes = n = (n + 127) & ~127;
474 return(n / 128 + 31);
477 *bytes = n = (n + 255) & ~255;
479 return(n / 256 + 39);
481 *bytes = n = (n + 511) & ~511;
483 return(n / 512 + 47);
485 #if ZALLOC_ZONE_LIMIT > 8192
487 *bytes = n = (n + 1023) & ~1023;
489 return(n / 1024 + 55);
492 #if ZALLOC_ZONE_LIMIT > 16384
494 *bytes = n = (n + 2047) & ~2047;
496 return(n / 2048 + 63);
499 panic("Unexpected byte count %d", n);
505 clean_zone_rchunks(SLZone *z)
509 while ((bchunk = z->z_RChunks) != NULL) {
511 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
512 *z->z_LChunksp = bchunk;
514 chunk_mark_free(z, bchunk);
515 z->z_LChunksp = &bchunk->c_Next;
516 bchunk = bchunk->c_Next;
526 * If the zone becomes totally free and is not the only zone listed for a
527 * chunk size we move it to the FreeZones list. We always leave at least
528 * one zone per chunk size listed, even if it is freeable.
530 * Do not move the zone if there is an IPI in_flight (z_RCount != 0),
531 * otherwise MP races can result in our free_remote code accessing a
532 * destroyed zone. The remote end interlocks z_RCount with z_RChunks
533 * so one has to test both z_NFree and z_RCount.
535 * Since this code can be called from an IPI callback, do *NOT* try to mess
536 * with kernel_map here. Hysteresis will be performed at kmalloc() time.
540 check_zone_free(SLGlobalData *slgd, SLZone *z)
544 znext = TAILQ_NEXT(z, z_Entry);
545 if (z->z_NFree == z->z_NMax && z->z_RCount == 0 &&
546 (TAILQ_FIRST(&slgd->ZoneAry[z->z_ZoneIndex]) != z || znext)
550 TAILQ_REMOVE(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
553 TAILQ_INSERT_HEAD(&slgd->FreeZones, z, z_Entry);
563 * Used to debug memory corruption issues. Record up to (typically 32)
564 * allocation sources for this zone (for a particular chunk size).
568 slab_record_source(SLZone *z, const char *file, int line)
571 int b = line & (SLAB_DEBUG_ENTRIES - 1);
575 if (z->z_Sources[i].file == file && z->z_Sources[i].line == line)
577 if (z->z_Sources[i].file == NULL)
579 i = (i + 1) & (SLAB_DEBUG_ENTRIES - 1);
581 z->z_Sources[i].file = file;
582 z->z_Sources[i].line = line;
587 static __inline unsigned long
588 powerof2_size(unsigned long size)
592 if (size == 0 || powerof2(size))
600 * kmalloc() (SLAB ALLOCATOR)
602 * Allocate memory via the slab allocator. If the request is too large,
603 * or if it page-aligned beyond a certain size, we fall back to the
604 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
605 * &SlabMisc if you don't care.
607 * M_RNOWAIT - don't block.
608 * M_NULLOK - return NULL instead of blocking.
609 * M_ZERO - zero the returned memory.
610 * M_USE_RESERVE - allow greater drawdown of the free list
611 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
612 * M_POWEROF2 - roundup size to the nearest power of 2
619 kmalloc_debug(unsigned long size, struct malloc_type *type, int flags,
620 const char *file, int line)
623 kmalloc(unsigned long size, struct malloc_type *type, int flags)
629 struct globaldata *gd;
636 logmemory_quick(malloc_beg);
641 * XXX silly to have this in the critical path.
643 if (type->ks_limit == 0) {
650 if (flags & M_POWEROF2)
651 size = powerof2_size(size);
654 * Handle the case where the limit is reached. Panic if we can't return
655 * NULL. The original malloc code looped, but this tended to
656 * simply deadlock the computer.
658 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
659 * to determine if a more complete limit check should be done. The
660 * actual memory use is tracked via ks_memuse[cpu].
662 while (type->ks_loosememuse >= type->ks_limit) {
666 for (i = ttl = 0; i < ncpus; ++i)
667 ttl += type->ks_memuse[i];
668 type->ks_loosememuse = ttl; /* not MP synchronized */
669 if ((ssize_t)ttl < 0) /* deal with occassional race */
671 if (ttl >= type->ks_limit) {
672 if (flags & M_NULLOK) {
673 logmemory(malloc_end, NULL, type, size, flags);
676 panic("%s: malloc limit exceeded", type->ks_shortdesc);
681 * Handle the degenerate size == 0 case. Yes, this does happen.
682 * Return a special pointer. This is to maintain compatibility with
683 * the original malloc implementation. Certain devices, such as the
684 * adaptec driver, not only allocate 0 bytes, they check for NULL and
685 * also realloc() later on. Joy.
688 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
689 return(ZERO_LENGTH_PTR);
693 * Handle hysteresis from prior frees here in malloc(). We cannot
694 * safely manipulate the kernel_map in free() due to free() possibly
695 * being called via an IPI message or from sensitive interrupt code.
697 * NOTE: ku_pagecnt must be cleared before we free the slab or we
698 * might race another cpu allocating the kva and setting
701 while (slgd->NFreeZones > ZoneRelsThresh && (flags & M_RNOWAIT) == 0) {
703 if (slgd->NFreeZones > ZoneRelsThresh) { /* crit sect race */
706 z = TAILQ_LAST(&slgd->FreeZones, SLZoneList);
708 TAILQ_REMOVE(&slgd->FreeZones, z, z_Entry);
712 kmem_slab_free(z, ZoneSize); /* may block */
713 atomic_add_int(&ZoneGenAlloc, -ZoneSize / 1024);
719 * XXX handle oversized frees that were queued from kfree().
721 while (TAILQ_FIRST(&slgd->FreeOvZones) && (flags & M_RNOWAIT) == 0) {
723 if ((z = TAILQ_LAST(&slgd->FreeOvZones, SLZoneList)) != NULL) {
726 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
727 TAILQ_REMOVE(&slgd->FreeOvZones, z, z_Entry);
728 tsize = z->z_ChunkSize;
729 kmem_slab_free(z, tsize); /* may block */
730 atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024);
736 * Handle large allocations directly. There should not be very many of
737 * these so performance is not a big issue.
739 * The backend allocator is pretty nasty on a SMP system. Use the
740 * slab allocator for one and two page-sized chunks even though we lose
741 * some efficiency. XXX maybe fix mmio and the elf loader instead.
743 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
746 size = round_page(size);
747 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
749 logmemory(malloc_end, NULL, type, size, flags);
752 atomic_add_int(&ZoneBigAlloc, (int)size / 1024);
753 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
754 flags |= M_PASSIVE_ZERO;
756 *kup = size / PAGE_SIZE;
762 * Attempt to allocate out of an existing zone. First try the free list,
763 * then allocate out of unallocated space. If we find a good zone move
764 * it to the head of the list so later allocations find it quickly
765 * (we might have thousands of zones in the list).
767 * Note: zoneindex() will panic of size is too large.
769 zi = zoneindex(&size, &align);
770 KKASSERT(zi < NZONES);
773 if ((z = TAILQ_LAST(&slgd->ZoneAry[zi], SLZoneList)) != NULL) {
775 * Locate a chunk - we have to have at least one. If this is the
776 * last chunk go ahead and do the work to retrieve chunks freed
777 * from remote cpus, and if the zone is still empty move it off
780 if (--z->z_NFree <= 0) {
781 KKASSERT(z->z_NFree == 0);
784 * WARNING! This code competes with other cpus. It is ok
785 * for us to not drain RChunks here but we might as well, and
786 * it is ok if more accumulate after we're done.
788 * Set RSignal before pulling rchunks off, indicating that we
789 * will be moving ourselves off of the ZoneAry. Remote ends will
790 * read RSignal before putting rchunks on thus interlocking
791 * their IPI signaling.
793 if (z->z_RChunks == NULL)
794 atomic_swap_int(&z->z_RSignal, 1);
796 clean_zone_rchunks(z);
799 * Remove from the zone list if no free chunks remain.
802 if (z->z_NFree == 0) {
803 TAILQ_REMOVE(&slgd->ZoneAry[zi], z, z_Entry);
810 * Fast path, we have chunks available in z_LChunks.
812 chunk = z->z_LChunks;
814 chunk_mark_allocated(z, chunk);
815 z->z_LChunks = chunk->c_Next;
816 if (z->z_LChunks == NULL)
817 z->z_LChunksp = &z->z_LChunks;
819 slab_record_source(z, file, line);
825 * No chunks are available in LChunks, the free chunk MUST be
826 * in the never-before-used memory area, controlled by UIndex.
828 * The consequences are very serious if our zone got corrupted so
829 * we use an explicit panic rather than a KASSERT.
831 if (z->z_UIndex + 1 != z->z_NMax)
836 if (z->z_UIndex == z->z_UEndIndex)
837 panic("slaballoc: corrupted zone");
839 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
840 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
842 flags |= M_PASSIVE_ZERO;
844 chunk_mark_allocated(z, chunk);
846 slab_record_source(z, file, line);
852 * If all zones are exhausted we need to allocate a new zone for this
853 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
854 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
855 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
856 * we do not pre-zero it because we do not want to mess up the L1 cache.
858 * At least one subsystem, the tty code (see CROUND) expects power-of-2
859 * allocations to be power-of-2 aligned. We maintain compatibility by
860 * adjusting the base offset below.
866 if ((z = TAILQ_FIRST(&slgd->FreeZones)) != NULL) {
867 TAILQ_REMOVE(&slgd->FreeZones, z, z_Entry);
869 bzero(z, sizeof(SLZone));
870 z->z_Flags |= SLZF_UNOTZEROD;
872 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
875 atomic_add_int(&ZoneGenAlloc, ZoneSize / 1024);
879 * How big is the base structure?
881 #if defined(INVARIANTS)
883 * Make room for z_Bitmap. An exact calculation is somewhat more
884 * complicated so don't make an exact calculation.
886 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
887 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
889 off = sizeof(SLZone);
893 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
894 * Otherwise properly align the data according to the chunk size.
898 off = roundup2(off, align);
900 z->z_Magic = ZALLOC_SLAB_MAGIC;
902 z->z_NMax = (ZoneSize - off) / size;
903 z->z_NFree = z->z_NMax - 1;
904 z->z_BasePtr = (char *)z + off;
905 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
906 z->z_ChunkSize = size;
908 z->z_Cpu = gd->gd_cpuid;
909 z->z_LChunksp = &z->z_LChunks;
911 bcopy(z->z_Sources, z->z_AltSources, sizeof(z->z_Sources));
912 bzero(z->z_Sources, sizeof(z->z_Sources));
914 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
915 TAILQ_INSERT_HEAD(&slgd->ZoneAry[zi], z, z_Entry);
916 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
917 flags &= ~M_ZERO; /* already zero'd */
918 flags |= M_PASSIVE_ZERO;
921 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */
922 chunk_mark_allocated(z, chunk);
924 slab_record_source(z, file, line);
928 * Slide the base index for initial allocations out of the next
929 * zone we create so we do not over-weight the lower part of the
932 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
933 & (ZALLOC_MAX_ZONE_SIZE - 1);
937 ++type->ks_inuse[gd->gd_cpuid];
938 type->ks_memuse[gd->gd_cpuid] += size;
939 type->ks_loosememuse += size; /* not MP synchronized */
945 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
946 if (use_malloc_pattern) {
947 for (i = 0; i < size; i += sizeof(int)) {
948 *(int *)((char *)chunk + i) = -1;
951 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
954 logmemory(malloc_end, chunk, type, size, flags);
958 logmemory(malloc_end, NULL, type, size, flags);
963 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
965 * Generally speaking this routine is not called very often and we do
966 * not attempt to optimize it beyond reusing the same pointer if the
967 * new size fits within the chunking of the old pointer's zone.
971 krealloc_debug(void *ptr, unsigned long size,
972 struct malloc_type *type, int flags,
973 const char *file, int line)
976 krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
985 KKASSERT((flags & M_ZERO) == 0); /* not supported */
987 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
988 return(kmalloc_debug(size, type, flags, file, line));
995 * Handle oversized allocations. XXX we really should require that a
996 * size be passed to free() instead of this nonsense.
1000 osize = *kup << PAGE_SHIFT;
1001 if (osize == round_page(size))
1003 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
1005 bcopy(ptr, nptr, min(size, osize));
1011 * Get the original allocation's zone. If the new request winds up
1012 * using the same chunk size we do not have to do anything.
1014 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1017 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1020 * Allocate memory for the new request size. Note that zoneindex has
1021 * already adjusted the request size to the appropriate chunk size, which
1022 * should optimize our bcopy(). Then copy and return the new pointer.
1024 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
1025 * necessary align the result.
1027 * We can only zoneindex (to align size to the chunk size) if the new
1028 * size is not too large.
1030 if (size < ZoneLimit) {
1031 zoneindex(&size, &align);
1032 if (z->z_ChunkSize == size)
1035 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
1037 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
1043 * Return the kmalloc limit for this type, in bytes.
1046 kmalloc_limit(struct malloc_type *type)
1048 if (type->ks_limit == 0) {
1050 if (type->ks_limit == 0)
1054 return(type->ks_limit);
1058 * Allocate a copy of the specified string.
1060 * (MP SAFE) (MAY BLOCK)
1064 kstrdup_debug(const char *str, struct malloc_type *type,
1065 const char *file, int line)
1068 kstrdup(const char *str, struct malloc_type *type)
1071 int zlen; /* length inclusive of terminating NUL */
1076 zlen = strlen(str) + 1;
1077 nstr = kmalloc_debug(zlen, type, M_WAITOK, file, line);
1078 bcopy(str, nstr, zlen);
1083 * Notify our cpu that a remote cpu has freed some chunks in a zone that
1084 * we own. RCount will be bumped so the memory should be good, but validate
1085 * that it really is.
1089 kfree_remote(void *ptr)
1096 slgd = &mycpu->gd_slab;
1099 KKASSERT(*kup == -((int)mycpuid + 1));
1100 KKASSERT(z->z_RCount > 0);
1101 atomic_subtract_int(&z->z_RCount, 1);
1103 logmemory(free_rem_beg, z, NULL, 0L, 0);
1104 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1105 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
1109 * Indicate that we will no longer be off of the ZoneAry by
1116 * Atomically extract the bchunks list and then process it back
1117 * into the lchunks list. We want to append our bchunks to the
1118 * lchunks list and not prepend since we likely do not have
1119 * cache mastership of the related data (not that it helps since
1120 * we are using c_Next).
1122 clean_zone_rchunks(z);
1123 if (z->z_NFree && nfree == 0) {
1124 TAILQ_INSERT_HEAD(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1128 * If the zone becomes totally free and is not the only zone listed for a
1129 * chunk size we move it to the FreeZones list. We always leave at least
1130 * one zone per chunk size listed, even if it is freeable.
1132 * Since this code can be called from an IPI callback, do *NOT* try to
1133 * mess with kernel_map here. Hysteresis will be performed at malloc()
1136 * Do not move the zone if there is an IPI in_flight (z_RCount != 0),
1137 * otherwise MP races can result in our free_remote code accessing a
1138 * destroyed zone. The remote end interlocks z_RCount with z_RChunks
1139 * so one has to test both z_NFree and z_RCount.
1141 if (z->z_NFree == z->z_NMax && z->z_RCount == 0 &&
1142 (TAILQ_FIRST(&slgd->ZoneAry[z->z_ZoneIndex]) != z ||
1143 TAILQ_NEXT(z, z_Entry))
1147 TAILQ_REMOVE(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1149 TAILQ_INSERT_HEAD(&slgd->FreeZones, z, z_Entry);
1154 logmemory(free_rem_end, z, NULL, 0L, 0);
1158 * free (SLAB ALLOCATOR)
1160 * Free a memory block previously allocated by malloc. Note that we do not
1161 * attempt to update ks_loosememuse as MP races could prevent us from
1162 * checking memory limits in malloc.
1167 kfree(void *ptr, struct malloc_type *type)
1172 struct globaldata *gd;
1178 logmemory_quick(free_beg);
1180 slgd = &gd->gd_slab;
1183 panic("trying to free NULL pointer");
1186 * Handle special 0-byte allocations
1188 if (ptr == ZERO_LENGTH_PTR) {
1189 logmemory(free_zero, ptr, type, -1UL, 0);
1190 logmemory_quick(free_end);
1195 * Panic on bad malloc type
1197 if (type->ks_magic != M_MAGIC)
1198 panic("free: malloc type lacks magic");
1201 * Handle oversized allocations. XXX we really should require that a
1202 * size be passed to free() instead of this nonsense.
1204 * This code is never called via an ipi.
1208 size = *kup << PAGE_SHIFT;
1211 KKASSERT(sizeof(weirdary) <= size);
1212 bcopy(weirdary, ptr, sizeof(weirdary));
1215 * NOTE: For oversized allocations we do not record the
1216 * originating cpu. It gets freed on the cpu calling
1217 * kfree(). The statistics are in aggregate.
1219 * note: XXX we have still inherited the interrupts-can't-block
1220 * assumption. An interrupt thread does not bump
1221 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1222 * primarily until we can fix softupdate's assumptions about free().
1225 --type->ks_inuse[gd->gd_cpuid];
1226 type->ks_memuse[gd->gd_cpuid] -= size;
1227 if (mycpu->gd_intr_nesting_level ||
1228 (gd->gd_curthread->td_flags & TDF_INTTHREAD))
1230 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1232 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1233 z->z_ChunkSize = size;
1235 TAILQ_INSERT_HEAD(&slgd->FreeOvZones, z, z_Entry);
1239 logmemory(free_ovsz, ptr, type, size, 0);
1240 kmem_slab_free(ptr, size); /* may block */
1241 atomic_add_int(&ZoneBigAlloc, -(int)size / 1024);
1243 logmemory_quick(free_end);
1248 * Zone case. Figure out the zone based on the fact that it is
1251 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1254 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1257 * If we do not own the zone then use atomic ops to free to the
1258 * remote cpu linked list and notify the target zone using a
1261 * The target zone cannot be deallocated while we own a chunk of it,
1262 * so the zone header's storage is stable until the very moment
1263 * we adjust z_RChunks. After that we cannot safely dereference (z).
1265 * (no critical section needed)
1267 if (z->z_CpuGd != gd) {
1269 * Making these adjustments now allow us to avoid passing (type)
1270 * to the remote cpu. Note that ks_inuse/ks_memuse is being
1271 * adjusted on OUR cpu, not the zone cpu, but it should all still
1272 * sum up properly and cancel out.
1275 --type->ks_inuse[gd->gd_cpuid];
1276 type->ks_memuse[gd->gd_cpuid] -= z->z_ChunkSize;
1280 * WARNING! This code competes with other cpus. Once we
1281 * successfully link the chunk to RChunks the remote
1282 * cpu can rip z's storage out from under us.
1284 * Bumping RCount prevents z's storage from getting
1287 rsignal = z->z_RSignal;
1290 atomic_add_int(&z->z_RCount, 1);
1294 bchunk = z->z_RChunks;
1296 chunk->c_Next = bchunk;
1299 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1304 * We have to signal the remote cpu if our actions will cause
1305 * the remote zone to be placed back on ZoneAry so it can
1306 * move the zone back on.
1308 * We only need to deal with NULL->non-NULL RChunk transitions
1309 * and only if z_RSignal is set. We interlock by reading rsignal
1310 * before adding our chunk to RChunks. This should result in
1311 * virtually no IPI traffic.
1313 * We can use a passive IPI to reduce overhead even further.
1315 if (bchunk == NULL && rsignal) {
1316 logmemory(free_request, ptr, type,
1317 (unsigned long)z->z_ChunkSize, 0);
1318 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
1319 /* z can get ripped out from under us from this point on */
1320 } else if (rsignal) {
1321 atomic_subtract_int(&z->z_RCount, 1);
1322 /* z can get ripped out from under us from this point on */
1324 logmemory_quick(free_end);
1331 logmemory(free_chunk, ptr, type, (unsigned long)z->z_ChunkSize, 0);
1335 chunk_mark_free(z, chunk);
1338 * Put weird data into the memory to detect modifications after freeing,
1339 * illegal pointer use after freeing (we should fault on the odd address),
1340 * and so forth. XXX needs more work, see the old malloc code.
1343 if (z->z_ChunkSize < sizeof(weirdary))
1344 bcopy(weirdary, chunk, z->z_ChunkSize);
1346 bcopy(weirdary, chunk, sizeof(weirdary));
1350 * Add this free non-zero'd chunk to a linked list for reuse. Add
1351 * to the front of the linked list so it is more likely to be
1352 * reallocated, since it is already in our L1 cache.
1355 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
1356 panic("BADFREE %p", chunk);
1358 chunk->c_Next = z->z_LChunks;
1359 z->z_LChunks = chunk;
1360 if (chunk->c_Next == NULL)
1361 z->z_LChunksp = &chunk->c_Next;
1364 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
1369 * Bump the number of free chunks. If it becomes non-zero the zone
1370 * must be added back onto the appropriate list. A fully allocated
1371 * zone that sees its first free is considered 'mature' and is placed
1372 * at the head, giving the system time to potentially free the remaining
1373 * entries even while other allocations are going on and making the zone
1376 if (z->z_NFree++ == 0) {
1378 TAILQ_INSERT_TAIL(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1380 TAILQ_INSERT_HEAD(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1383 --type->ks_inuse[z->z_Cpu];
1384 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
1386 check_zone_free(slgd, z);
1387 logmemory_quick(free_end);
1392 * Cleanup slabs which are hanging around due to RChunks or which are wholely
1393 * free and can be moved to the free list if not moved by other means.
1395 * Called once every 10 seconds on all cpus.
1400 SLGlobalData *slgd = &mycpu->gd_slab;
1405 for (i = 0; i < NZONES; ++i) {
1406 if ((z = TAILQ_FIRST(&slgd->ZoneAry[i])) == NULL)
1414 * Shift all RChunks to the end of the LChunks list. This is
1415 * an O(1) operation.
1417 * Then free the zone if possible.
1419 clean_zone_rchunks(z);
1420 z = check_zone_free(slgd, z);
1426 #if defined(INVARIANTS)
1429 * Helper routines for sanity checks
1433 chunk_mark_allocated(SLZone *z, void *chunk)
1435 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1438 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1439 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1440 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
1441 bitptr = &z->z_Bitmap[bitdex >> 5];
1443 KASSERT((*bitptr & (1 << bitdex)) == 0,
1444 ("memory chunk %p is already allocated!", chunk));
1445 *bitptr |= 1 << bitdex;
1450 chunk_mark_free(SLZone *z, void *chunk)
1452 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1455 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1456 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1457 ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
1458 bitptr = &z->z_Bitmap[bitdex >> 5];
1460 KASSERT((*bitptr & (1 << bitdex)) != 0,
1461 ("memory chunk %p is already free!", chunk));
1462 *bitptr &= ~(1 << bitdex);
1470 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1471 * specified alignment. M_* flags are expected in the flags field.
1473 * Alignment must be a multiple of PAGE_SIZE.
1475 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1476 * but when we move zalloc() over to use this function as its backend
1477 * we will have to switch to kreserve/krelease and call reserve(0)
1478 * after the new space is made available.
1480 * Interrupt code which has preempted other code is not allowed to
1481 * use PQ_CACHE pages. However, if an interrupt thread is run
1482 * non-preemptively or blocks and then runs non-preemptively, then
1483 * it is free to use PQ_CACHE pages. <--- may not apply any longer XXX
1486 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1490 int count, vmflags, base_vmflags;
1491 vm_page_t mbase = NULL;
1495 size = round_page(size);
1496 addr = vm_map_min(&kernel_map);
1498 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1500 vm_map_lock(&kernel_map);
1501 if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) {
1502 vm_map_unlock(&kernel_map);
1503 if ((flags & M_NULLOK) == 0)
1504 panic("kmem_slab_alloc(): kernel_map ran out of space!");
1505 vm_map_entry_release(count);
1511 * kernel_object maps 1:1 to kernel_map.
1513 vm_object_hold(&kernel_object);
1514 vm_object_reference_locked(&kernel_object);
1515 vm_map_insert(&kernel_map, &count,
1516 &kernel_object, NULL,
1517 addr, addr, addr + size,
1519 VM_PROT_ALL, VM_PROT_ALL,
1521 vm_object_drop(&kernel_object);
1522 vm_map_set_wired_quick(&kernel_map, addr, size, &count);
1523 vm_map_unlock(&kernel_map);
1529 base_vmflags |= VM_ALLOC_ZERO;
1530 if (flags & M_USE_RESERVE)
1531 base_vmflags |= VM_ALLOC_SYSTEM;
1532 if (flags & M_USE_INTERRUPT_RESERVE)
1533 base_vmflags |= VM_ALLOC_INTERRUPT;
1534 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1535 panic("kmem_slab_alloc: bad flags %08x (%p)",
1536 flags, ((int **)&size)[-1]);
1540 * Allocate the pages. Do not mess with the PG_ZERO flag or map
1541 * them yet. VM_ALLOC_NORMAL can only be set if we are not preempting.
1543 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1544 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1545 * implied in this case), though I'm not sure if we really need to
1548 vmflags = base_vmflags;
1549 if (flags & M_WAITOK) {
1550 if (td->td_preempted)
1551 vmflags |= VM_ALLOC_SYSTEM;
1553 vmflags |= VM_ALLOC_NORMAL;
1556 vm_object_hold(&kernel_object);
1557 for (i = 0; i < size; i += PAGE_SIZE) {
1558 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
1563 * If the allocation failed we either return NULL or we retry.
1565 * If M_WAITOK is specified we wait for more memory and retry.
1566 * If M_WAITOK is specified from a preemption we yield instead of
1567 * wait. Livelock will not occur because the interrupt thread
1568 * will not be preempting anyone the second time around after the
1572 if (flags & M_WAITOK) {
1573 if (td->td_preempted) {
1578 i -= PAGE_SIZE; /* retry */
1586 * Check and deal with an allocation failure
1591 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1592 /* page should already be busy */
1595 vm_map_lock(&kernel_map);
1596 vm_map_delete(&kernel_map, addr, addr + size, &count);
1597 vm_map_unlock(&kernel_map);
1598 vm_object_drop(&kernel_object);
1600 vm_map_entry_release(count);
1608 * NOTE: The VM pages are still busied. mbase points to the first one
1609 * but we have to iterate via vm_page_next()
1611 vm_object_drop(&kernel_object);
1615 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1622 * page should already be busy
1624 m->valid = VM_PAGE_BITS_ALL;
1626 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL | VM_PROT_NOSYNC,
1628 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1629 bzero((char *)addr + i, PAGE_SIZE);
1630 vm_page_flag_clear(m, PG_ZERO);
1631 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1632 vm_page_flag_set(m, PG_REFERENCED);
1636 vm_object_hold(&kernel_object);
1637 m = vm_page_next(m);
1638 vm_object_drop(&kernel_object);
1641 vm_map_entry_release(count);
1642 atomic_add_long(&SlabsAllocated, 1);
1643 return((void *)addr);
1650 kmem_slab_free(void *ptr, vm_size_t size)
1653 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
1654 atomic_add_long(&SlabsFreed, 1);
1659 kmalloc_cachealign(unsigned long size_alloc, struct malloc_type *type,
1662 #if (__VM_CACHELINE_SIZE == 32)
1663 #define CAN_CACHEALIGN(sz) ((sz) >= 256)
1664 #elif (__VM_CACHELINE_SIZE == 64)
1665 #define CAN_CACHEALIGN(sz) ((sz) >= 512)
1666 #elif (__VM_CACHELINE_SIZE == 128)
1667 #define CAN_CACHEALIGN(sz) ((sz) >= 1024)
1669 #error "unsupported cacheline size"
1674 if (size_alloc < __VM_CACHELINE_SIZE)
1675 size_alloc = __VM_CACHELINE_SIZE;
1676 else if (!CAN_CACHEALIGN(size_alloc))
1677 flags |= M_POWEROF2;
1679 ret = kmalloc(size_alloc, type, flags);
1680 KASSERT(((uintptr_t)ret & (__VM_CACHELINE_SIZE - 1)) == 0,
1681 ("%p(%lu) not cacheline %d aligned",
1682 ret, size_alloc, __VM_CACHELINE_SIZE));
1685 #undef CAN_CACHEALIGN