2 * KERN_SLABALLOC.C - Kernel SLAB memory allocator
4 * Copyright (c) 2003,2004,2010-2019 The DragonFly Project.
7 * This code is derived from software contributed to The DragonFly Project
8 * by Matthew Dillon <dillon@backplane.com>
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in
18 * the documentation and/or other materials provided with the
20 * 3. Neither the name of The DragonFly Project nor the names of its
21 * contributors may be used to endorse or promote products derived
22 * from this software without specific, prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
25 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
26 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
27 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
28 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
29 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
30 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
31 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
32 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
33 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
34 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37 * This module implements a slab allocator drop-in replacement for the
40 * A slab allocator reserves a ZONE for each chunk size, then lays the
41 * chunks out in an array within the zone. Allocation and deallocation
42 * is nearly instantanious, and fragmentation/overhead losses are limited
43 * to a fixed worst-case amount.
45 * The downside of this slab implementation is in the chunk size
46 * multiplied by the number of zones. ~80 zones * 128K = 10MB of VM per cpu.
47 * In a kernel implementation all this memory will be physical so
48 * the zone size is adjusted downward on machines with less physical
49 * memory. The upside is that overhead is bounded... this is the *worst*
52 * Slab management is done on a per-cpu basis and no locking or mutexes
53 * are required, only a critical section. When one cpu frees memory
54 * belonging to another cpu's slab manager an asynchronous IPI message
55 * will be queued to execute the operation. In addition, both the
56 * high level slab allocator and the low level zone allocator optimize
57 * M_ZERO requests, and the slab allocator does not have to pre initialize
58 * the linked list of chunks.
60 * XXX Balancing is needed between cpus. Balance will be handled through
61 * asynchronous IPIs primarily by reassigning the z_Cpu ownership of chunks.
63 * XXX If we have to allocate a new zone and M_USE_RESERVE is set, use of
64 * the new zone should be restricted to M_USE_RESERVE requests only.
66 * Alloc Size Chunking Number of zones
76 * (if PAGE_SIZE is 4K the maximum zone allocation is 16383)
78 * Allocations >= ZoneLimit go directly to kmem.
79 * (n * PAGE_SIZE, n > 2) allocations 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 #if (__VM_CACHELINE_SIZE == 32)
127 #define CAN_CACHEALIGN(sz) ((sz) >= 256)
128 #elif (__VM_CACHELINE_SIZE == 64)
129 #define CAN_CACHEALIGN(sz) ((sz) >= 512)
130 #elif (__VM_CACHELINE_SIZE == 128)
131 #define CAN_CACHEALIGN(sz) ((sz) >= 1024)
133 #error "unsupported cacheline size"
136 #define btokup(z) (&pmap_kvtom((vm_offset_t)(z))->ku_pagecnt)
138 #define MEMORY_STRING "ptr=%p type=%p size=%lu flags=%04x"
139 #define MEMORY_ARGS void *ptr, void *type, unsigned long size, int flags
141 #if !defined(KTR_MEMORY)
142 #define KTR_MEMORY KTR_ALL
144 KTR_INFO_MASTER(memory);
145 KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin");
146 KTR_INFO(KTR_MEMORY, memory, malloc_end, 1, MEMORY_STRING, MEMORY_ARGS);
147 KTR_INFO(KTR_MEMORY, memory, free_zero, 2, MEMORY_STRING, MEMORY_ARGS);
148 KTR_INFO(KTR_MEMORY, memory, free_ovsz, 3, MEMORY_STRING, MEMORY_ARGS);
149 KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 4, MEMORY_STRING, MEMORY_ARGS);
150 KTR_INFO(KTR_MEMORY, memory, free_chunk, 5, MEMORY_STRING, MEMORY_ARGS);
151 KTR_INFO(KTR_MEMORY, memory, free_request, 6, MEMORY_STRING, MEMORY_ARGS);
152 KTR_INFO(KTR_MEMORY, memory, free_rem_beg, 7, MEMORY_STRING, MEMORY_ARGS);
153 KTR_INFO(KTR_MEMORY, memory, free_rem_end, 8, MEMORY_STRING, MEMORY_ARGS);
154 KTR_INFO(KTR_MEMORY, memory, free_beg, 9, "free begin");
155 KTR_INFO(KTR_MEMORY, memory, free_end, 10, "free end");
157 #define logmemory(name, ptr, type, size, flags) \
158 KTR_LOG(memory_ ## name, ptr, type, size, flags)
159 #define logmemory_quick(name) \
160 KTR_LOG(memory_ ## name)
163 * Fixed globals (not per-cpu)
166 static int ZoneLimit;
167 static int ZonePageCount;
168 static uintptr_t ZoneMask;
169 static int ZoneBigAlloc; /* in KB */
170 static int ZoneGenAlloc; /* in KB */
171 struct malloc_type *kmemstatistics; /* exported to vmstat */
173 static int32_t weirdary[16];
176 static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
177 static void kmem_slab_free(void *ptr, vm_size_t bytes);
179 #if defined(INVARIANTS)
180 static void chunk_mark_allocated(SLZone *z, void *chunk);
181 static void chunk_mark_free(SLZone *z, void *chunk);
183 #define chunk_mark_allocated(z, chunk)
184 #define chunk_mark_free(z, chunk)
188 * Misc constants. Note that allocations that are exact multiples of
189 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
191 #define ZONE_RELS_THRESH 32 /* threshold number of zones */
195 * The WEIRD_ADDR is used as known text to copy into free objects to
196 * try to create deterministic failure cases if the data is accessed after
199 #define WEIRD_ADDR 0xdeadc0de
201 #define ZERO_LENGTH_PTR ((void *)-8)
204 * Misc global malloc buckets
207 MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
208 MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
209 MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
210 MALLOC_DEFINE(M_DRM, "m_drm", "DRM memory allocations");
212 MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
213 MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
216 * Initialize the slab memory allocator. We have to choose a zone size based
217 * on available physical memory. We choose a zone side which is approximately
218 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
219 * 128K. The zone size is limited to the bounds set in slaballoc.h
220 * (typically 32K min, 128K max).
222 static void kmeminit(void *dummy);
226 SYSINIT(kmem, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL);
230 * If enabled any memory allocated without M_ZERO is initialized to -1.
232 static int use_malloc_pattern;
233 SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
234 &use_malloc_pattern, 0,
235 "Initialize memory to -1 if M_ZERO not specified");
238 static int ZoneRelsThresh = ZONE_RELS_THRESH;
239 SYSCTL_INT(_kern, OID_AUTO, zone_big_alloc, CTLFLAG_RD, &ZoneBigAlloc, 0, "");
240 SYSCTL_INT(_kern, OID_AUTO, zone_gen_alloc, CTLFLAG_RD, &ZoneGenAlloc, 0, "");
241 SYSCTL_INT(_kern, OID_AUTO, zone_cache, CTLFLAG_RW, &ZoneRelsThresh, 0, "");
242 static long SlabsAllocated;
243 static long SlabsFreed;
244 SYSCTL_LONG(_kern, OID_AUTO, slabs_allocated, CTLFLAG_RD,
245 &SlabsAllocated, 0, "");
246 SYSCTL_LONG(_kern, OID_AUTO, slabs_freed, CTLFLAG_RD,
248 static int SlabFreeToTail;
249 SYSCTL_INT(_kern, OID_AUTO, slab_freetotail, CTLFLAG_RW,
250 &SlabFreeToTail, 0, "");
252 static struct spinlock kmemstat_spin =
253 SPINLOCK_INITIALIZER(&kmemstat_spin, "malinit");
256 * Returns the kernel memory size limit for the purposes of initializing
257 * various subsystem caches. The smaller of available memory and the KVM
258 * memory space is returned.
260 * The size in megabytes is returned.
267 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
268 if (limsize > KvaSize)
270 return (limsize / (1024 * 1024));
274 kmeminit(void *dummy)
282 limsize = kmem_lim_size();
283 usesize = (int)(limsize * 1024); /* convert to KB */
286 * If the machine has a large KVM space and more than 8G of ram,
287 * double the zone release threshold to reduce SMP invalidations.
288 * If more than 16G of ram, do it again.
290 * The BIOS eats a little ram so add some slop. We want 8G worth of
291 * memory sticks to trigger the first adjustment.
293 if (ZoneRelsThresh == ZONE_RELS_THRESH) {
294 if (limsize >= 7 * 1024)
296 if (limsize >= 15 * 1024)
301 * Calculate the zone size. This typically calculates to
302 * ZALLOC_MAX_ZONE_SIZE
304 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
305 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
307 ZoneLimit = ZoneSize / 4;
308 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
309 ZoneLimit = ZALLOC_ZONE_LIMIT;
310 ZoneMask = ~(uintptr_t)(ZoneSize - 1);
311 ZonePageCount = ZoneSize / PAGE_SIZE;
314 for (i = 0; i < NELEM(weirdary); ++i)
315 weirdary[i] = WEIRD_ADDR;
318 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
321 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
325 * (low level) Initialize slab-related elements in the globaldata structure.
327 * Occurs after kmeminit().
330 slab_gdinit(globaldata_t gd)
336 for (i = 0; i < NZONES; ++i)
337 TAILQ_INIT(&slgd->ZoneAry[i]);
338 TAILQ_INIT(&slgd->FreeZones);
339 TAILQ_INIT(&slgd->FreeOvZones);
343 * Initialize a malloc type tracking structure.
346 malloc_init(void *data)
348 struct malloc_type *type = data;
351 if (type->ks_magic != M_MAGIC)
352 panic("malloc type lacks magic");
354 if (type->ks_limit != 0)
357 if (vmstats.v_page_count == 0)
358 panic("malloc_init not allowed before vm init");
360 limsize = kmem_lim_size() * (1024 * 1024);
361 type->ks_limit = limsize / 10;
363 spin_lock(&kmemstat_spin);
364 type->ks_next = kmemstatistics;
365 kmemstatistics = type;
366 spin_unlock(&kmemstat_spin);
370 malloc_uninit(void *data)
372 struct malloc_type *type = data;
373 struct malloc_type *t;
379 if (type->ks_magic != M_MAGIC)
380 panic("malloc type lacks magic");
382 if (vmstats.v_page_count == 0)
383 panic("malloc_uninit not allowed before vm init");
385 if (type->ks_limit == 0)
386 panic("malloc_uninit on uninitialized type");
388 /* Make sure that all pending kfree()s are finished. */
389 lwkt_synchronize_ipiqs("muninit");
393 * memuse is only correct in aggregation. Due to memory being allocated
394 * on one cpu and freed on another individual array entries may be
395 * negative or positive (canceling each other out).
397 for (i = ttl = 0; i < ncpus; ++i)
398 ttl += type->ks_use[i].memuse;
400 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
401 ttl, type->ks_shortdesc, i);
404 spin_lock(&kmemstat_spin);
405 if (type == kmemstatistics) {
406 kmemstatistics = type->ks_next;
408 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
409 if (t->ks_next == type) {
410 t->ks_next = type->ks_next;
415 type->ks_next = NULL;
417 spin_unlock(&kmemstat_spin);
421 * Increase the kmalloc pool limit for the specified pool. No changes
422 * are the made if the pool would shrink.
425 kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
427 if (type->ks_limit == 0)
431 if (type->ks_limit < bytes)
432 type->ks_limit = bytes;
436 kmalloc_set_unlimited(struct malloc_type *type)
438 type->ks_limit = kmem_lim_size() * (1024 * 1024);
442 * Dynamically create a malloc pool. This function is a NOP if *typep is
446 kmalloc_create(struct malloc_type **typep, const char *descr)
448 struct malloc_type *type;
450 if (*typep == NULL) {
451 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
452 type->ks_magic = M_MAGIC;
453 type->ks_shortdesc = descr;
460 * Destroy a dynamically created malloc pool. This function is a NOP if
461 * the pool has already been destroyed.
464 kmalloc_destroy(struct malloc_type **typep)
466 if (*typep != NULL) {
467 malloc_uninit(*typep);
468 kfree(*typep, M_TEMP);
474 * Calculate the zone index for the allocation request size and set the
475 * allocation request size to that particular zone's chunk size.
478 zoneindex(unsigned long *bytes, unsigned long *align)
480 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
483 *bytes = n = (n + 7) & ~7;
485 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
488 *bytes = n = (n + 15) & ~15;
494 *bytes = n = (n + 31) & ~31;
499 *bytes = n = (n + 63) & ~63;
504 *bytes = n = (n + 127) & ~127;
506 return(n / 128 + 31);
509 *bytes = n = (n + 255) & ~255;
511 return(n / 256 + 39);
513 *bytes = n = (n + 511) & ~511;
515 return(n / 512 + 47);
517 #if ZALLOC_ZONE_LIMIT > 8192
519 *bytes = n = (n + 1023) & ~1023;
521 return(n / 1024 + 55);
524 #if ZALLOC_ZONE_LIMIT > 16384
526 *bytes = n = (n + 2047) & ~2047;
528 return(n / 2048 + 63);
531 panic("Unexpected byte count %d", n);
536 clean_zone_rchunks(SLZone *z)
540 while ((bchunk = z->z_RChunks) != NULL) {
542 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
543 *z->z_LChunksp = bchunk;
545 chunk_mark_free(z, bchunk);
546 z->z_LChunksp = &bchunk->c_Next;
547 bchunk = bchunk->c_Next;
557 * If the zone becomes totally free and is not the only zone listed for a
558 * chunk size we move it to the FreeZones list. We always leave at least
559 * one zone per chunk size listed, even if it is freeable.
561 * Do not move the zone if there is an IPI in_flight (z_RCount != 0),
562 * otherwise MP races can result in our free_remote code accessing a
563 * destroyed zone. The remote end interlocks z_RCount with z_RChunks
564 * so one has to test both z_NFree and z_RCount.
566 * Since this code can be called from an IPI callback, do *NOT* try to mess
567 * with kernel_map here. Hysteresis will be performed at kmalloc() time.
569 static __inline SLZone *
570 check_zone_free(SLGlobalData *slgd, SLZone *z)
574 znext = TAILQ_NEXT(z, z_Entry);
575 if (z->z_NFree == z->z_NMax && z->z_RCount == 0 &&
576 (TAILQ_FIRST(&slgd->ZoneAry[z->z_ZoneIndex]) != z || znext)) {
579 TAILQ_REMOVE(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
582 TAILQ_INSERT_HEAD(&slgd->FreeZones, z, z_Entry);
592 * Used to debug memory corruption issues. Record up to (typically 32)
593 * allocation sources for this zone (for a particular chunk size).
597 slab_record_source(SLZone *z, const char *file, int line)
600 int b = line & (SLAB_DEBUG_ENTRIES - 1);
604 if (z->z_Sources[i].file == file && z->z_Sources[i].line == line)
606 if (z->z_Sources[i].file == NULL)
608 i = (i + 1) & (SLAB_DEBUG_ENTRIES - 1);
610 z->z_Sources[i].file = file;
611 z->z_Sources[i].line = line;
616 static __inline unsigned long
617 powerof2_size(unsigned long size)
621 if (size == 0 || powerof2(size))
629 * kmalloc() (SLAB ALLOCATOR)
631 * Allocate memory via the slab allocator. If the request is too large,
632 * or if it page-aligned beyond a certain size, we fall back to the
633 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
634 * &SlabMisc if you don't care.
636 * M_RNOWAIT - don't block.
637 * M_NULLOK - return NULL instead of blocking.
638 * M_ZERO - zero the returned memory.
639 * M_USE_RESERVE - allow greater drawdown of the free list
640 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
641 * M_POWEROF2 - roundup size to the nearest power of 2
646 /* don't let kmalloc macro mess up function declaration */
651 kmalloc_debug(unsigned long size, struct malloc_type *type, int flags,
652 const char *file, int line)
655 kmalloc(unsigned long size, struct malloc_type *type, int flags)
661 struct globaldata *gd;
668 logmemory_quick(malloc_beg);
673 * XXX silly to have this in the critical path.
675 if (type->ks_limit == 0) {
680 ++type->ks_use[gd->gd_cpuid].calls;
683 * Flagged for cache-alignment
685 if (flags & M_CACHEALIGN) {
686 if (size < __VM_CACHELINE_SIZE)
687 size = __VM_CACHELINE_SIZE;
688 else if (!CAN_CACHEALIGN(size))
693 * Flagged to force nearest power-of-2 (higher or same)
695 if (flags & M_POWEROF2)
696 size = powerof2_size(size);
699 * Handle the case where the limit is reached. Panic if we can't return
700 * NULL. The original malloc code looped, but this tended to
701 * simply deadlock the computer.
703 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
704 * to determine if a more complete limit check should be done. The
705 * actual memory use is tracked via ks_use[cpu].memuse.
707 while (type->ks_loosememuse >= type->ks_limit) {
711 for (i = ttl = 0; i < ncpus; ++i)
712 ttl += type->ks_use[i].memuse;
713 type->ks_loosememuse = ttl; /* not MP synchronized */
714 if ((ssize_t)ttl < 0) /* deal with occassional race */
716 if (ttl >= type->ks_limit) {
717 if (flags & M_NULLOK) {
718 logmemory(malloc_end, NULL, type, size, flags);
721 panic("%s: malloc limit exceeded", type->ks_shortdesc);
726 * Handle the degenerate size == 0 case. Yes, this does happen.
727 * Return a special pointer. This is to maintain compatibility with
728 * the original malloc implementation. Certain devices, such as the
729 * adaptec driver, not only allocate 0 bytes, they check for NULL and
730 * also realloc() later on. Joy.
733 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
734 return(ZERO_LENGTH_PTR);
738 * Handle hysteresis from prior frees here in malloc(). We cannot
739 * safely manipulate the kernel_map in free() due to free() possibly
740 * being called via an IPI message or from sensitive interrupt code.
742 * NOTE: ku_pagecnt must be cleared before we free the slab or we
743 * might race another cpu allocating the kva and setting
746 while (slgd->NFreeZones > ZoneRelsThresh && (flags & M_RNOWAIT) == 0) {
748 if (slgd->NFreeZones > ZoneRelsThresh) { /* crit sect race */
751 z = TAILQ_LAST(&slgd->FreeZones, SLZoneList);
753 TAILQ_REMOVE(&slgd->FreeZones, z, z_Entry);
757 kmem_slab_free(z, ZoneSize); /* may block */
758 atomic_add_int(&ZoneGenAlloc, -ZoneSize / 1024);
764 * XXX handle oversized frees that were queued from kfree().
766 while (TAILQ_FIRST(&slgd->FreeOvZones) && (flags & M_RNOWAIT) == 0) {
768 if ((z = TAILQ_LAST(&slgd->FreeOvZones, SLZoneList)) != NULL) {
771 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
772 TAILQ_REMOVE(&slgd->FreeOvZones, z, z_Entry);
773 tsize = z->z_ChunkSize;
774 kmem_slab_free(z, tsize); /* may block */
775 atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024);
781 * Handle large allocations directly. There should not be very many of
782 * these so performance is not a big issue.
784 * The backend allocator is pretty nasty on a SMP system. Use the
785 * slab allocator for one and two page-sized chunks even though we lose
786 * some efficiency. XXX maybe fix mmio and the elf loader instead.
788 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
791 size = round_page(size);
792 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
794 logmemory(malloc_end, NULL, type, size, flags);
797 atomic_add_int(&ZoneBigAlloc, (int)size / 1024);
798 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
799 flags |= M_PASSIVE_ZERO;
801 *kup = size / PAGE_SIZE;
807 * Attempt to allocate out of an existing zone. First try the free list,
808 * then allocate out of unallocated space. If we find a good zone move
809 * it to the head of the list so later allocations find it quickly
810 * (we might have thousands of zones in the list).
812 * Note: zoneindex() will panic of size is too large.
814 zi = zoneindex(&size, &align);
815 KKASSERT(zi < NZONES);
818 if ((z = TAILQ_LAST(&slgd->ZoneAry[zi], SLZoneList)) != NULL) {
820 * Locate a chunk - we have to have at least one. If this is the
821 * last chunk go ahead and do the work to retrieve chunks freed
822 * from remote cpus, and if the zone is still empty move it off
825 if (--z->z_NFree <= 0) {
826 KKASSERT(z->z_NFree == 0);
829 * WARNING! This code competes with other cpus. It is ok
830 * for us to not drain RChunks here but we might as well, and
831 * it is ok if more accumulate after we're done.
833 * Set RSignal before pulling rchunks off, indicating that we
834 * will be moving ourselves off of the ZoneAry. Remote ends will
835 * read RSignal before putting rchunks on thus interlocking
836 * their IPI signaling.
838 if (z->z_RChunks == NULL)
839 atomic_swap_int(&z->z_RSignal, 1);
841 clean_zone_rchunks(z);
844 * Remove from the zone list if no free chunks remain.
847 if (z->z_NFree == 0) {
848 TAILQ_REMOVE(&slgd->ZoneAry[zi], z, z_Entry);
855 * Fast path, we have chunks available in z_LChunks.
857 chunk = z->z_LChunks;
859 chunk_mark_allocated(z, chunk);
860 z->z_LChunks = chunk->c_Next;
861 if (z->z_LChunks == NULL)
862 z->z_LChunksp = &z->z_LChunks;
864 slab_record_source(z, file, line);
870 * No chunks are available in LChunks, the free chunk MUST be
871 * in the never-before-used memory area, controlled by UIndex.
873 * The consequences are very serious if our zone got corrupted so
874 * we use an explicit panic rather than a KASSERT.
876 if (z->z_UIndex + 1 != z->z_NMax)
881 if (z->z_UIndex == z->z_UEndIndex)
882 panic("slaballoc: corrupted zone");
884 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
885 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
887 flags |= M_PASSIVE_ZERO;
889 chunk_mark_allocated(z, chunk);
891 slab_record_source(z, file, line);
897 * If all zones are exhausted we need to allocate a new zone for this
898 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
899 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
900 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
901 * we do not pre-zero it because we do not want to mess up the L1 cache.
903 * At least one subsystem, the tty code (see CROUND) expects power-of-2
904 * allocations to be power-of-2 aligned. We maintain compatibility by
905 * adjusting the base offset below.
911 if ((z = TAILQ_FIRST(&slgd->FreeZones)) != NULL) {
912 TAILQ_REMOVE(&slgd->FreeZones, z, z_Entry);
914 bzero(z, sizeof(SLZone));
915 z->z_Flags |= SLZF_UNOTZEROD;
917 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
920 atomic_add_int(&ZoneGenAlloc, ZoneSize / 1024);
924 * How big is the base structure?
926 #if defined(INVARIANTS)
928 * Make room for z_Bitmap. An exact calculation is somewhat more
929 * complicated so don't make an exact calculation.
931 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
932 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
934 off = sizeof(SLZone);
938 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
939 * Otherwise properly align the data according to the chunk size.
943 off = roundup2(off, align);
945 z->z_Magic = ZALLOC_SLAB_MAGIC;
947 z->z_NMax = (ZoneSize - off) / size;
948 z->z_NFree = z->z_NMax - 1;
949 z->z_BasePtr = (char *)z + off;
950 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
951 z->z_ChunkSize = size;
953 z->z_Cpu = gd->gd_cpuid;
954 z->z_LChunksp = &z->z_LChunks;
956 bcopy(z->z_Sources, z->z_AltSources, sizeof(z->z_Sources));
957 bzero(z->z_Sources, sizeof(z->z_Sources));
959 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
960 TAILQ_INSERT_HEAD(&slgd->ZoneAry[zi], z, z_Entry);
961 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
962 flags &= ~M_ZERO; /* already zero'd */
963 flags |= M_PASSIVE_ZERO;
966 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */
967 chunk_mark_allocated(z, chunk);
969 slab_record_source(z, file, line);
973 * Slide the base index for initial allocations out of the next
974 * zone we create so we do not over-weight the lower part of the
977 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
978 & (ZALLOC_MAX_ZONE_SIZE - 1);
982 ++type->ks_use[gd->gd_cpuid].inuse;
983 type->ks_use[gd->gd_cpuid].memuse += size;
984 type->ks_use[gd->gd_cpuid].loosememuse += size;
985 if (type->ks_use[gd->gd_cpuid].loosememuse >= ZoneSize) {
986 /* not MP synchronized */
987 type->ks_loosememuse += type->ks_use[gd->gd_cpuid].loosememuse;
988 type->ks_use[gd->gd_cpuid].loosememuse = 0;
995 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
996 if (use_malloc_pattern) {
997 for (i = 0; i < size; i += sizeof(int)) {
998 *(int *)((char *)chunk + i) = -1;
1001 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
1004 logmemory(malloc_end, chunk, type, size, flags);
1008 logmemory(malloc_end, NULL, type, size, flags);
1013 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
1015 * Generally speaking this routine is not called very often and we do
1016 * not attempt to optimize it beyond reusing the same pointer if the
1017 * new size fits within the chunking of the old pointer's zone.
1021 krealloc_debug(void *ptr, unsigned long size,
1022 struct malloc_type *type, int flags,
1023 const char *file, int line)
1026 krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
1029 unsigned long osize;
1030 unsigned long align;
1035 KKASSERT((flags & M_ZERO) == 0); /* not supported */
1037 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
1038 return(kmalloc_debug(size, type, flags, file, line));
1045 * Handle oversized allocations. XXX we really should require that a
1046 * size be passed to free() instead of this nonsense.
1050 osize = *kup << PAGE_SHIFT;
1051 if (osize == round_page(size))
1053 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
1055 bcopy(ptr, nptr, min(size, osize));
1061 * Get the original allocation's zone. If the new request winds up
1062 * using the same chunk size we do not have to do anything.
1064 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1067 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1070 * Allocate memory for the new request size. Note that zoneindex has
1071 * already adjusted the request size to the appropriate chunk size, which
1072 * should optimize our bcopy(). Then copy and return the new pointer.
1074 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
1075 * necessary align the result.
1077 * We can only zoneindex (to align size to the chunk size) if the new
1078 * size is not too large.
1080 if (size < ZoneLimit) {
1081 zoneindex(&size, &align);
1082 if (z->z_ChunkSize == size)
1085 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
1087 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
1093 * Return the kmalloc limit for this type, in bytes.
1096 kmalloc_limit(struct malloc_type *type)
1098 if (type->ks_limit == 0) {
1100 if (type->ks_limit == 0)
1104 return(type->ks_limit);
1108 * Allocate a copy of the specified string.
1110 * (MP SAFE) (MAY BLOCK)
1114 kstrdup_debug(const char *str, struct malloc_type *type,
1115 const char *file, int line)
1118 kstrdup(const char *str, struct malloc_type *type)
1121 int zlen; /* length inclusive of terminating NUL */
1126 zlen = strlen(str) + 1;
1127 nstr = kmalloc_debug(zlen, type, M_WAITOK, file, line);
1128 bcopy(str, nstr, zlen);
1134 kstrndup_debug(const char *str, size_t maxlen, struct malloc_type *type,
1135 const char *file, int line)
1138 kstrndup(const char *str, size_t maxlen, struct malloc_type *type)
1141 int zlen; /* length inclusive of terminating NUL */
1146 zlen = strnlen(str, maxlen) + 1;
1147 nstr = kmalloc_debug(zlen, type, M_WAITOK, file, line);
1148 bcopy(str, nstr, zlen);
1149 nstr[zlen - 1] = '\0';
1154 * Notify our cpu that a remote cpu has freed some chunks in a zone that
1155 * we own. RCount will be bumped so the memory should be good, but validate
1156 * that it really is.
1159 kfree_remote(void *ptr)
1166 slgd = &mycpu->gd_slab;
1169 KKASSERT(*kup == -((int)mycpuid + 1));
1170 KKASSERT(z->z_RCount > 0);
1171 atomic_subtract_int(&z->z_RCount, 1);
1173 logmemory(free_rem_beg, z, NULL, 0L, 0);
1174 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1175 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
1179 * Indicate that we will no longer be off of the ZoneAry by
1186 * Atomically extract the bchunks list and then process it back
1187 * into the lchunks list. We want to append our bchunks to the
1188 * lchunks list and not prepend since we likely do not have
1189 * cache mastership of the related data (not that it helps since
1190 * we are using c_Next).
1192 clean_zone_rchunks(z);
1193 if (z->z_NFree && nfree == 0) {
1194 TAILQ_INSERT_HEAD(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1197 check_zone_free(slgd, z);
1198 logmemory(free_rem_end, z, NULL, 0L, 0);
1202 * free (SLAB ALLOCATOR)
1204 * Free a memory block previously allocated by malloc.
1206 * Note: We do not attempt to update ks_loosememuse as MP races could
1207 * prevent us from checking memory limits in malloc. YYY we may
1208 * consider updating ks_cpu.loosememuse.
1213 kfree(void *ptr, struct malloc_type *type)
1218 struct globaldata *gd;
1224 logmemory_quick(free_beg);
1226 slgd = &gd->gd_slab;
1229 panic("trying to free NULL pointer");
1232 * Handle special 0-byte allocations
1234 if (ptr == ZERO_LENGTH_PTR) {
1235 logmemory(free_zero, ptr, type, -1UL, 0);
1236 logmemory_quick(free_end);
1241 * Panic on bad malloc type
1243 if (type->ks_magic != M_MAGIC)
1244 panic("free: malloc type lacks magic");
1247 * Handle oversized allocations. XXX we really should require that a
1248 * size be passed to free() instead of this nonsense.
1250 * This code is never called via an ipi.
1254 size = *kup << PAGE_SHIFT;
1257 KKASSERT(sizeof(weirdary) <= size);
1258 bcopy(weirdary, ptr, sizeof(weirdary));
1261 * NOTE: For oversized allocations we do not record the
1262 * originating cpu. It gets freed on the cpu calling
1263 * kfree(). The statistics are in aggregate.
1265 * note: XXX we have still inherited the interrupts-can't-block
1266 * assumption. An interrupt thread does not bump
1267 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1268 * primarily until we can fix softupdate's assumptions about free().
1271 --type->ks_use[gd->gd_cpuid].inuse;
1272 type->ks_use[gd->gd_cpuid].memuse -= size;
1273 if (mycpu->gd_intr_nesting_level ||
1274 (gd->gd_curthread->td_flags & TDF_INTTHREAD)) {
1275 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1277 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1278 z->z_ChunkSize = size;
1280 TAILQ_INSERT_HEAD(&slgd->FreeOvZones, z, z_Entry);
1284 logmemory(free_ovsz, ptr, type, size, 0);
1285 kmem_slab_free(ptr, size); /* may block */
1286 atomic_add_int(&ZoneBigAlloc, -(int)size / 1024);
1288 logmemory_quick(free_end);
1293 * Zone case. Figure out the zone based on the fact that it is
1296 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1299 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1302 * If we do not own the zone then use atomic ops to free to the
1303 * remote cpu linked list and notify the target zone using a
1306 * The target zone cannot be deallocated while we own a chunk of it,
1307 * so the zone header's storage is stable until the very moment
1308 * we adjust z_RChunks. After that we cannot safely dereference (z).
1310 * (no critical section needed)
1312 if (z->z_CpuGd != gd) {
1314 * Making these adjustments now allow us to avoid passing (type)
1315 * to the remote cpu. Note that inuse/memuse is being
1316 * adjusted on OUR cpu, not the zone cpu, but it should all still
1317 * sum up properly and cancel out.
1320 --type->ks_use[gd->gd_cpuid].inuse;
1321 type->ks_use[gd->gd_cpuid].memuse -= z->z_ChunkSize;
1325 * WARNING! This code competes with other cpus. Once we
1326 * successfully link the chunk to RChunks the remote
1327 * cpu can rip z's storage out from under us.
1329 * Bumping RCount prevents z's storage from getting
1332 rsignal = z->z_RSignal;
1335 atomic_add_int(&z->z_RCount, 1);
1339 bchunk = z->z_RChunks;
1341 chunk->c_Next = bchunk;
1344 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1349 * We have to signal the remote cpu if our actions will cause
1350 * the remote zone to be placed back on ZoneAry so it can
1351 * move the zone back on.
1353 * We only need to deal with NULL->non-NULL RChunk transitions
1354 * and only if z_RSignal is set. We interlock by reading rsignal
1355 * before adding our chunk to RChunks. This should result in
1356 * virtually no IPI traffic.
1358 * We can use a passive IPI to reduce overhead even further.
1360 if (bchunk == NULL && rsignal) {
1361 logmemory(free_request, ptr, type,
1362 (unsigned long)z->z_ChunkSize, 0);
1363 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
1364 /* z can get ripped out from under us from this point on */
1365 } else if (rsignal) {
1366 atomic_subtract_int(&z->z_RCount, 1);
1367 /* z can get ripped out from under us from this point on */
1369 logmemory_quick(free_end);
1376 logmemory(free_chunk, ptr, type, (unsigned long)z->z_ChunkSize, 0);
1380 chunk_mark_free(z, chunk);
1383 * Put weird data into the memory to detect modifications after freeing,
1384 * illegal pointer use after freeing (we should fault on the odd address),
1385 * and so forth. XXX needs more work, see the old malloc code.
1388 if (z->z_ChunkSize < sizeof(weirdary))
1389 bcopy(weirdary, chunk, z->z_ChunkSize);
1391 bcopy(weirdary, chunk, sizeof(weirdary));
1395 * Add this free non-zero'd chunk to a linked list for reuse. Add
1396 * to the front of the linked list so it is more likely to be
1397 * reallocated, since it is already in our L1 cache.
1400 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
1401 panic("BADFREE %p", chunk);
1403 chunk->c_Next = z->z_LChunks;
1404 z->z_LChunks = chunk;
1405 if (chunk->c_Next == NULL)
1406 z->z_LChunksp = &chunk->c_Next;
1409 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
1414 * Bump the number of free chunks. If it becomes non-zero the zone
1415 * must be added back onto the appropriate list. A fully allocated
1416 * zone that sees its first free is considered 'mature' and is placed
1417 * at the head, giving the system time to potentially free the remaining
1418 * entries even while other allocations are going on and making the zone
1421 if (z->z_NFree++ == 0) {
1423 TAILQ_INSERT_TAIL(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1425 TAILQ_INSERT_HEAD(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1428 --type->ks_use[gd->gd_cpuid].inuse;
1429 type->ks_use[gd->gd_cpuid].memuse -= z->z_ChunkSize;
1431 check_zone_free(slgd, z);
1432 logmemory_quick(free_end);
1437 * Cleanup slabs which are hanging around due to RChunks or which are wholely
1438 * free and can be moved to the free list if not moved by other means.
1440 * Called once every 10 seconds on all cpus.
1445 SLGlobalData *slgd = &mycpu->gd_slab;
1450 for (i = 0; i < NZONES; ++i) {
1451 if ((z = TAILQ_FIRST(&slgd->ZoneAry[i])) == NULL)
1459 * Shift all RChunks to the end of the LChunks list. This is
1460 * an O(1) operation.
1462 * Then free the zone if possible.
1464 clean_zone_rchunks(z);
1465 z = check_zone_free(slgd, z);
1471 #if defined(INVARIANTS)
1474 * Helper routines for sanity checks
1477 chunk_mark_allocated(SLZone *z, void *chunk)
1479 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1482 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1483 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1484 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
1485 bitptr = &z->z_Bitmap[bitdex >> 5];
1487 KASSERT((*bitptr & (1 << bitdex)) == 0,
1488 ("memory chunk %p is already allocated!", chunk));
1489 *bitptr |= 1 << bitdex;
1493 chunk_mark_free(SLZone *z, void *chunk)
1495 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1498 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1499 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1500 ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
1501 bitptr = &z->z_Bitmap[bitdex >> 5];
1503 KASSERT((*bitptr & (1 << bitdex)) != 0,
1504 ("memory chunk %p is already free!", chunk));
1505 *bitptr &= ~(1 << bitdex);
1513 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1514 * specified alignment. M_* flags are expected in the flags field.
1516 * Alignment must be a multiple of PAGE_SIZE.
1518 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1519 * but when we move zalloc() over to use this function as its backend
1520 * we will have to switch to kreserve/krelease and call reserve(0)
1521 * after the new space is made available.
1523 * Interrupt code which has preempted other code is not allowed to
1524 * use PQ_CACHE pages. However, if an interrupt thread is run
1525 * non-preemptively or blocks and then runs non-preemptively, then
1526 * it is free to use PQ_CACHE pages. <--- may not apply any longer XXX
1529 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1533 int count, vmflags, base_vmflags;
1534 vm_page_t mbase = NULL;
1538 size = round_page(size);
1539 addr = vm_map_min(&kernel_map);
1541 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1543 vm_map_lock(&kernel_map);
1544 if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) {
1545 vm_map_unlock(&kernel_map);
1546 if ((flags & M_NULLOK) == 0)
1547 panic("kmem_slab_alloc(): kernel_map ran out of space!");
1548 vm_map_entry_release(count);
1554 * kernel_object maps 1:1 to kernel_map.
1556 vm_object_hold(&kernel_object);
1557 vm_object_reference_locked(&kernel_object);
1558 vm_map_insert(&kernel_map, &count,
1559 &kernel_object, NULL,
1560 addr, addr, addr + size,
1563 VM_PROT_ALL, VM_PROT_ALL, 0);
1564 vm_object_drop(&kernel_object);
1565 vm_map_set_wired_quick(&kernel_map, addr, size, &count);
1566 vm_map_unlock(&kernel_map);
1572 base_vmflags |= VM_ALLOC_ZERO;
1573 if (flags & M_USE_RESERVE)
1574 base_vmflags |= VM_ALLOC_SYSTEM;
1575 if (flags & M_USE_INTERRUPT_RESERVE)
1576 base_vmflags |= VM_ALLOC_INTERRUPT;
1577 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1578 panic("kmem_slab_alloc: bad flags %08x (%p)",
1579 flags, ((int **)&size)[-1]);
1583 * Allocate the pages. Do not map them yet. VM_ALLOC_NORMAL can only
1584 * be set if we are not preempting.
1586 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1587 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1588 * implied in this case), though I'm not sure if we really need to
1591 vmflags = base_vmflags;
1592 if (flags & M_WAITOK) {
1593 if (td->td_preempted)
1594 vmflags |= VM_ALLOC_SYSTEM;
1596 vmflags |= VM_ALLOC_NORMAL;
1599 vm_object_hold(&kernel_object);
1600 for (i = 0; i < size; i += PAGE_SIZE) {
1601 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
1606 * If the allocation failed we either return NULL or we retry.
1608 * If M_WAITOK is specified we wait for more memory and retry.
1609 * If M_WAITOK is specified from a preemption we yield instead of
1610 * wait. Livelock will not occur because the interrupt thread
1611 * will not be preempting anyone the second time around after the
1615 if (flags & M_WAITOK) {
1616 if (td->td_preempted) {
1621 i -= PAGE_SIZE; /* retry */
1629 * Check and deal with an allocation failure
1634 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1635 /* page should already be busy */
1638 vm_map_lock(&kernel_map);
1639 vm_map_delete(&kernel_map, addr, addr + size, &count);
1640 vm_map_unlock(&kernel_map);
1641 vm_object_drop(&kernel_object);
1643 vm_map_entry_release(count);
1651 * NOTE: The VM pages are still busied. mbase points to the first one
1652 * but we have to iterate via vm_page_next()
1654 vm_object_drop(&kernel_object);
1658 * Enter the pages into the pmap and deal with M_ZERO.
1665 * page should already be busy
1667 m->valid = VM_PAGE_BITS_ALL;
1669 pmap_enter(&kernel_pmap, addr + i, m,
1670 VM_PROT_ALL | VM_PROT_NOSYNC, 1, NULL);
1672 pagezero((char *)addr + i);
1673 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1674 vm_page_flag_set(m, PG_REFERENCED);
1678 vm_object_hold(&kernel_object);
1679 m = vm_page_next(m);
1680 vm_object_drop(&kernel_object);
1683 vm_map_entry_release(count);
1684 atomic_add_long(&SlabsAllocated, 1);
1685 return((void *)addr);
1692 kmem_slab_free(void *ptr, vm_size_t size)
1695 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
1696 atomic_add_long(&SlabsFreed, 1);