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, &SlabsAllocated, 0, "");
231 SYSCTL_LONG(_kern, OID_AUTO, slabs_freed, CTLFLAG_RD, &SlabsFreed, 0, "");
234 * Returns the kernel memory size limit for the purposes of initializing
235 * various subsystem caches. The smaller of available memory and the KVM
236 * memory space is returned.
238 * The size in megabytes is returned.
245 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
246 if (limsize > KvaSize)
248 return (limsize / (1024 * 1024));
252 kmeminit(void *dummy)
258 limsize = kmem_lim_size();
259 usesize = (int)(limsize * 1024); /* convert to KB */
262 * If the machine has a large KVM space and more than 8G of ram,
263 * double the zone release threshold to reduce SMP invalidations.
264 * If more than 16G of ram, do it again.
266 * The BIOS eats a little ram so add some slop. We want 8G worth of
267 * memory sticks to trigger the first adjustment.
269 if (ZoneRelsThresh == ZONE_RELS_THRESH) {
270 if (limsize >= 7 * 1024)
272 if (limsize >= 15 * 1024)
277 * Calculate the zone size. This typically calculates to
278 * ZALLOC_MAX_ZONE_SIZE
280 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
281 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
283 ZoneLimit = ZoneSize / 4;
284 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
285 ZoneLimit = ZALLOC_ZONE_LIMIT;
286 ZoneMask = ~(uintptr_t)(ZoneSize - 1);
287 ZonePageCount = ZoneSize / PAGE_SIZE;
289 for (i = 0; i < NELEM(weirdary); ++i)
290 weirdary[i] = WEIRD_ADDR;
292 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
295 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
299 * Initialize a malloc type tracking structure.
302 malloc_init(void *data)
304 struct malloc_type *type = data;
307 if (type->ks_magic != M_MAGIC)
308 panic("malloc type lacks magic");
310 if (type->ks_limit != 0)
313 if (vmstats.v_page_count == 0)
314 panic("malloc_init not allowed before vm init");
316 limsize = kmem_lim_size() * (1024 * 1024);
317 type->ks_limit = limsize / 10;
319 type->ks_next = kmemstatistics;
320 kmemstatistics = type;
324 malloc_uninit(void *data)
326 struct malloc_type *type = data;
327 struct malloc_type *t;
333 if (type->ks_magic != M_MAGIC)
334 panic("malloc type lacks magic");
336 if (vmstats.v_page_count == 0)
337 panic("malloc_uninit not allowed before vm init");
339 if (type->ks_limit == 0)
340 panic("malloc_uninit on uninitialized type");
342 /* Make sure that all pending kfree()s are finished. */
343 lwkt_synchronize_ipiqs("muninit");
347 * memuse is only correct in aggregation. Due to memory being allocated
348 * on one cpu and freed on another individual array entries may be
349 * negative or positive (canceling each other out).
351 for (i = ttl = 0; i < ncpus; ++i)
352 ttl += type->ks_memuse[i];
354 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
355 ttl, type->ks_shortdesc, i);
358 if (type == kmemstatistics) {
359 kmemstatistics = type->ks_next;
361 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
362 if (t->ks_next == type) {
363 t->ks_next = type->ks_next;
368 type->ks_next = NULL;
373 * Increase the kmalloc pool limit for the specified pool. No changes
374 * are the made if the pool would shrink.
377 kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
379 if (type->ks_limit == 0)
383 if (type->ks_limit < bytes)
384 type->ks_limit = bytes;
388 * Dynamically create a malloc pool. This function is a NOP if *typep is
392 kmalloc_create(struct malloc_type **typep, const char *descr)
394 struct malloc_type *type;
396 if (*typep == NULL) {
397 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
398 type->ks_magic = M_MAGIC;
399 type->ks_shortdesc = descr;
406 * Destroy a dynamically created malloc pool. This function is a NOP if
407 * the pool has already been destroyed.
410 kmalloc_destroy(struct malloc_type **typep)
412 if (*typep != NULL) {
413 malloc_uninit(*typep);
414 kfree(*typep, M_TEMP);
420 * Calculate the zone index for the allocation request size and set the
421 * allocation request size to that particular zone's chunk size.
424 zoneindex(unsigned long *bytes, unsigned long *align)
426 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
428 *bytes = n = (n + 7) & ~7;
430 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
433 *bytes = n = (n + 15) & ~15;
439 *bytes = n = (n + 31) & ~31;
444 *bytes = n = (n + 63) & ~63;
449 *bytes = n = (n + 127) & ~127;
451 return(n / 128 + 31);
454 *bytes = n = (n + 255) & ~255;
456 return(n / 256 + 39);
458 *bytes = n = (n + 511) & ~511;
460 return(n / 512 + 47);
462 #if ZALLOC_ZONE_LIMIT > 8192
464 *bytes = n = (n + 1023) & ~1023;
466 return(n / 1024 + 55);
469 #if ZALLOC_ZONE_LIMIT > 16384
471 *bytes = n = (n + 2047) & ~2047;
473 return(n / 2048 + 63);
476 panic("Unexpected byte count %d", n);
482 clean_zone_rchunks(SLZone *z)
486 while ((bchunk = z->z_RChunks) != NULL) {
488 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
489 *z->z_LChunksp = bchunk;
491 chunk_mark_free(z, bchunk);
492 z->z_LChunksp = &bchunk->c_Next;
493 bchunk = bchunk->c_Next;
503 * If the zone becomes totally free, and there are other zones we
504 * can allocate from, move this zone to the FreeZones list. Since
505 * this code can be called from an IPI callback, do *NOT* try to mess
506 * with kernel_map here. Hysteresis will be performed at malloc() time.
510 check_zone_free(SLGlobalData *slgd, SLZone *z)
512 if (z->z_NFree == z->z_NMax &&
513 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
519 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
523 z->z_Next = slgd->FreeZones;
537 * Used to debug memory corruption issues. Record up to (typically 32)
538 * allocation sources for this zone (for a particular chunk size).
542 slab_record_source(SLZone *z, const char *file, int line)
545 int b = line & (SLAB_DEBUG_ENTRIES - 1);
549 if (z->z_Sources[i].file == file && z->z_Sources[i].line == line)
551 if (z->z_Sources[i].file == NULL)
553 i = (i + 1) & (SLAB_DEBUG_ENTRIES - 1);
555 z->z_Sources[i].file = file;
556 z->z_Sources[i].line = line;
561 static __inline unsigned long
562 powerof2_size(unsigned long size)
566 if (size == 0 || powerof2(size))
574 * kmalloc() (SLAB ALLOCATOR)
576 * Allocate memory via the slab allocator. If the request is too large,
577 * or if it page-aligned beyond a certain size, we fall back to the
578 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
579 * &SlabMisc if you don't care.
581 * M_RNOWAIT - don't block.
582 * M_NULLOK - return NULL instead of blocking.
583 * M_ZERO - zero the returned memory.
584 * M_USE_RESERVE - allow greater drawdown of the free list
585 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
586 * M_POWEROF2 - roundup size to the nearest power of 2
593 kmalloc_debug(unsigned long size, struct malloc_type *type, int flags,
594 const char *file, int line)
597 kmalloc(unsigned long size, struct malloc_type *type, int flags)
603 struct globaldata *gd;
610 logmemory_quick(malloc_beg);
615 * XXX silly to have this in the critical path.
617 if (type->ks_limit == 0) {
624 if (flags & M_POWEROF2)
625 size = powerof2_size(size);
628 * Handle the case where the limit is reached. Panic if we can't return
629 * NULL. The original malloc code looped, but this tended to
630 * simply deadlock the computer.
632 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
633 * to determine if a more complete limit check should be done. The
634 * actual memory use is tracked via ks_memuse[cpu].
636 while (type->ks_loosememuse >= type->ks_limit) {
640 for (i = ttl = 0; i < ncpus; ++i)
641 ttl += type->ks_memuse[i];
642 type->ks_loosememuse = ttl; /* not MP synchronized */
643 if ((ssize_t)ttl < 0) /* deal with occassional race */
645 if (ttl >= type->ks_limit) {
646 if (flags & M_NULLOK) {
647 logmemory(malloc_end, NULL, type, size, flags);
650 panic("%s: malloc limit exceeded", type->ks_shortdesc);
655 * Handle the degenerate size == 0 case. Yes, this does happen.
656 * Return a special pointer. This is to maintain compatibility with
657 * the original malloc implementation. Certain devices, such as the
658 * adaptec driver, not only allocate 0 bytes, they check for NULL and
659 * also realloc() later on. Joy.
662 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
663 return(ZERO_LENGTH_PTR);
667 * Handle hysteresis from prior frees here in malloc(). We cannot
668 * safely manipulate the kernel_map in free() due to free() possibly
669 * being called via an IPI message or from sensitive interrupt code.
671 * NOTE: ku_pagecnt must be cleared before we free the slab or we
672 * might race another cpu allocating the kva and setting
675 while (slgd->NFreeZones > ZoneRelsThresh && (flags & M_RNOWAIT) == 0) {
677 if (slgd->NFreeZones > ZoneRelsThresh) { /* crit sect race */
681 slgd->FreeZones = z->z_Next;
685 kmem_slab_free(z, ZoneSize); /* may block */
686 atomic_add_int(&ZoneGenAlloc, -ZoneSize / 1024);
692 * XXX handle oversized frees that were queued from kfree().
694 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) {
696 if ((z = slgd->FreeOvZones) != NULL) {
699 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
700 slgd->FreeOvZones = z->z_Next;
701 tsize = z->z_ChunkSize;
702 kmem_slab_free(z, tsize); /* may block */
703 atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024);
709 * Handle large allocations directly. There should not be very many of
710 * these so performance is not a big issue.
712 * The backend allocator is pretty nasty on a SMP system. Use the
713 * slab allocator for one and two page-sized chunks even though we lose
714 * some efficiency. XXX maybe fix mmio and the elf loader instead.
716 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
719 size = round_page(size);
720 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
722 logmemory(malloc_end, NULL, type, size, flags);
725 atomic_add_int(&ZoneBigAlloc, (int)size / 1024);
726 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
727 flags |= M_PASSIVE_ZERO;
729 *kup = size / PAGE_SIZE;
735 * Attempt to allocate out of an existing zone. First try the free list,
736 * then allocate out of unallocated space. If we find a good zone move
737 * it to the head of the list so later allocations find it quickly
738 * (we might have thousands of zones in the list).
740 * Note: zoneindex() will panic of size is too large.
742 zi = zoneindex(&size, &align);
743 KKASSERT(zi < NZONES);
746 if ((z = slgd->ZoneAry[zi]) != NULL) {
748 * Locate a chunk - we have to have at least one. If this is the
749 * last chunk go ahead and do the work to retrieve chunks freed
750 * from remote cpus, and if the zone is still empty move it off
753 if (--z->z_NFree <= 0) {
754 KKASSERT(z->z_NFree == 0);
757 * WARNING! This code competes with other cpus. It is ok
758 * for us to not drain RChunks here but we might as well, and
759 * it is ok if more accumulate after we're done.
761 * Set RSignal before pulling rchunks off, indicating that we
762 * will be moving ourselves off of the ZoneAry. Remote ends will
763 * read RSignal before putting rchunks on thus interlocking
764 * their IPI signaling.
766 if (z->z_RChunks == NULL)
767 atomic_swap_int(&z->z_RSignal, 1);
769 clean_zone_rchunks(z);
772 * Remove from the zone list if no free chunks remain.
775 if (z->z_NFree == 0) {
776 slgd->ZoneAry[zi] = z->z_Next;
784 * Fast path, we have chunks available in z_LChunks.
786 chunk = z->z_LChunks;
788 chunk_mark_allocated(z, chunk);
789 z->z_LChunks = chunk->c_Next;
790 if (z->z_LChunks == NULL)
791 z->z_LChunksp = &z->z_LChunks;
793 slab_record_source(z, file, line);
799 * No chunks are available in LChunks, the free chunk MUST be
800 * in the never-before-used memory area, controlled by UIndex.
802 * The consequences are very serious if our zone got corrupted so
803 * we use an explicit panic rather than a KASSERT.
805 if (z->z_UIndex + 1 != z->z_NMax)
810 if (z->z_UIndex == z->z_UEndIndex)
811 panic("slaballoc: corrupted zone");
813 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
814 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
816 flags |= M_PASSIVE_ZERO;
818 chunk_mark_allocated(z, chunk);
820 slab_record_source(z, file, line);
826 * If all zones are exhausted we need to allocate a new zone for this
827 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
828 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
829 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
830 * we do not pre-zero it because we do not want to mess up the L1 cache.
832 * At least one subsystem, the tty code (see CROUND) expects power-of-2
833 * allocations to be power-of-2 aligned. We maintain compatibility by
834 * adjusting the base offset below.
840 if ((z = slgd->FreeZones) != NULL) {
841 slgd->FreeZones = z->z_Next;
843 bzero(z, sizeof(SLZone));
844 z->z_Flags |= SLZF_UNOTZEROD;
846 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
849 atomic_add_int(&ZoneGenAlloc, ZoneSize / 1024);
853 * How big is the base structure?
855 #if defined(INVARIANTS)
857 * Make room for z_Bitmap. An exact calculation is somewhat more
858 * complicated so don't make an exact calculation.
860 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
861 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
863 off = sizeof(SLZone);
867 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
868 * Otherwise properly align the data according to the chunk size.
872 off = (off + align - 1) & ~(align - 1);
874 z->z_Magic = ZALLOC_SLAB_MAGIC;
876 z->z_NMax = (ZoneSize - off) / size;
877 z->z_NFree = z->z_NMax - 1;
878 z->z_BasePtr = (char *)z + off;
879 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
880 z->z_ChunkSize = size;
882 z->z_Cpu = gd->gd_cpuid;
883 z->z_LChunksp = &z->z_LChunks;
885 bcopy(z->z_Sources, z->z_AltSources, sizeof(z->z_Sources));
886 bzero(z->z_Sources, sizeof(z->z_Sources));
888 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
889 z->z_Next = slgd->ZoneAry[zi];
890 slgd->ZoneAry[zi] = z;
891 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
892 flags &= ~M_ZERO; /* already zero'd */
893 flags |= M_PASSIVE_ZERO;
896 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */
897 chunk_mark_allocated(z, chunk);
899 slab_record_source(z, file, line);
903 * Slide the base index for initial allocations out of the next
904 * zone we create so we do not over-weight the lower part of the
907 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
908 & (ZALLOC_MAX_ZONE_SIZE - 1);
912 ++type->ks_inuse[gd->gd_cpuid];
913 type->ks_memuse[gd->gd_cpuid] += size;
914 type->ks_loosememuse += size; /* not MP synchronized */
920 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
921 if (use_malloc_pattern) {
922 for (i = 0; i < size; i += sizeof(int)) {
923 *(int *)((char *)chunk + i) = -1;
926 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
929 logmemory(malloc_end, chunk, type, size, flags);
933 logmemory(malloc_end, NULL, type, size, flags);
938 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
940 * Generally speaking this routine is not called very often and we do
941 * not attempt to optimize it beyond reusing the same pointer if the
942 * new size fits within the chunking of the old pointer's zone.
946 krealloc_debug(void *ptr, unsigned long size,
947 struct malloc_type *type, int flags,
948 const char *file, int line)
951 krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
960 KKASSERT((flags & M_ZERO) == 0); /* not supported */
962 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
963 return(kmalloc_debug(size, type, flags, file, line));
970 * Handle oversized allocations. XXX we really should require that a
971 * size be passed to free() instead of this nonsense.
975 osize = *kup << PAGE_SHIFT;
976 if (osize == round_page(size))
978 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
980 bcopy(ptr, nptr, min(size, osize));
986 * Get the original allocation's zone. If the new request winds up
987 * using the same chunk size we do not have to do anything.
989 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
992 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
995 * Allocate memory for the new request size. Note that zoneindex has
996 * already adjusted the request size to the appropriate chunk size, which
997 * should optimize our bcopy(). Then copy and return the new pointer.
999 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
1000 * necessary align the result.
1002 * We can only zoneindex (to align size to the chunk size) if the new
1003 * size is not too large.
1005 if (size < ZoneLimit) {
1006 zoneindex(&size, &align);
1007 if (z->z_ChunkSize == size)
1010 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
1012 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
1018 * Return the kmalloc limit for this type, in bytes.
1021 kmalloc_limit(struct malloc_type *type)
1023 if (type->ks_limit == 0) {
1025 if (type->ks_limit == 0)
1029 return(type->ks_limit);
1033 * Allocate a copy of the specified string.
1035 * (MP SAFE) (MAY BLOCK)
1039 kstrdup_debug(const char *str, struct malloc_type *type,
1040 const char *file, int line)
1043 kstrdup(const char *str, struct malloc_type *type)
1046 int zlen; /* length inclusive of terminating NUL */
1051 zlen = strlen(str) + 1;
1052 nstr = kmalloc_debug(zlen, type, M_WAITOK, file, line);
1053 bcopy(str, nstr, zlen);
1058 * Notify our cpu that a remote cpu has freed some chunks in a zone that
1059 * we own. RCount will be bumped so the memory should be good, but validate
1060 * that it really is.
1064 kfree_remote(void *ptr)
1071 slgd = &mycpu->gd_slab;
1074 KKASSERT(*kup == -((int)mycpuid + 1));
1075 KKASSERT(z->z_RCount > 0);
1076 atomic_subtract_int(&z->z_RCount, 1);
1078 logmemory(free_rem_beg, z, NULL, 0L, 0);
1079 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1080 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
1084 * Indicate that we will no longer be off of the ZoneAry by
1091 * Atomically extract the bchunks list and then process it back
1092 * into the lchunks list. We want to append our bchunks to the
1093 * lchunks list and not prepend since we likely do not have
1094 * cache mastership of the related data (not that it helps since
1095 * we are using c_Next).
1097 clean_zone_rchunks(z);
1098 if (z->z_NFree && nfree == 0) {
1099 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1100 slgd->ZoneAry[z->z_ZoneIndex] = z;
1104 * If the zone becomes totally free, and there are other zones we
1105 * can allocate from, move this zone to the FreeZones list. Since
1106 * this code can be called from an IPI callback, do *NOT* try to mess
1107 * with kernel_map here. Hysteresis will be performed at malloc() time.
1109 * Do not move the zone if there is an IPI inflight, otherwise MP
1110 * races can result in our free_remote code accessing a destroyed
1113 if (z->z_NFree == z->z_NMax &&
1114 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1120 for (pz = &slgd->ZoneAry[z->z_ZoneIndex];
1122 pz = &(*pz)->z_Next) {
1127 z->z_Next = slgd->FreeZones;
1128 slgd->FreeZones = z;
1133 logmemory(free_rem_end, z, NULL, 0L, 0);
1137 * free (SLAB ALLOCATOR)
1139 * Free a memory block previously allocated by malloc. Note that we do not
1140 * attempt to update ks_loosememuse as MP races could prevent us from
1141 * checking memory limits in malloc.
1146 kfree(void *ptr, struct malloc_type *type)
1151 struct globaldata *gd;
1157 logmemory_quick(free_beg);
1159 slgd = &gd->gd_slab;
1162 panic("trying to free NULL pointer");
1165 * Handle special 0-byte allocations
1167 if (ptr == ZERO_LENGTH_PTR) {
1168 logmemory(free_zero, ptr, type, -1UL, 0);
1169 logmemory_quick(free_end);
1174 * Panic on bad malloc type
1176 if (type->ks_magic != M_MAGIC)
1177 panic("free: malloc type lacks magic");
1180 * Handle oversized allocations. XXX we really should require that a
1181 * size be passed to free() instead of this nonsense.
1183 * This code is never called via an ipi.
1187 size = *kup << PAGE_SHIFT;
1190 KKASSERT(sizeof(weirdary) <= size);
1191 bcopy(weirdary, ptr, sizeof(weirdary));
1194 * NOTE: For oversized allocations we do not record the
1195 * originating cpu. It gets freed on the cpu calling
1196 * kfree(). The statistics are in aggregate.
1198 * note: XXX we have still inherited the interrupts-can't-block
1199 * assumption. An interrupt thread does not bump
1200 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1201 * primarily until we can fix softupdate's assumptions about free().
1204 --type->ks_inuse[gd->gd_cpuid];
1205 type->ks_memuse[gd->gd_cpuid] -= size;
1206 if (mycpu->gd_intr_nesting_level ||
1207 (gd->gd_curthread->td_flags & TDF_INTTHREAD))
1209 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1211 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1212 z->z_Next = slgd->FreeOvZones;
1213 z->z_ChunkSize = size;
1214 slgd->FreeOvZones = z;
1218 logmemory(free_ovsz, ptr, type, size, 0);
1219 kmem_slab_free(ptr, size); /* may block */
1220 atomic_add_int(&ZoneBigAlloc, -(int)size / 1024);
1222 logmemory_quick(free_end);
1227 * Zone case. Figure out the zone based on the fact that it is
1230 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1233 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1236 * If we do not own the zone then use atomic ops to free to the
1237 * remote cpu linked list and notify the target zone using a
1240 * The target zone cannot be deallocated while we own a chunk of it,
1241 * so the zone header's storage is stable until the very moment
1242 * we adjust z_RChunks. After that we cannot safely dereference (z).
1244 * (no critical section needed)
1246 if (z->z_CpuGd != gd) {
1248 * Making these adjustments now allow us to avoid passing (type)
1249 * to the remote cpu. Note that ks_inuse/ks_memuse is being
1250 * adjusted on OUR cpu, not the zone cpu, but it should all still
1251 * sum up properly and cancel out.
1254 --type->ks_inuse[gd->gd_cpuid];
1255 type->ks_memuse[gd->gd_cpuid] -= z->z_ChunkSize;
1259 * WARNING! This code competes with other cpus. Once we
1260 * successfully link the chunk to RChunks the remote
1261 * cpu can rip z's storage out from under us.
1263 * Bumping RCount prevents z's storage from getting
1266 rsignal = z->z_RSignal;
1269 atomic_add_int(&z->z_RCount, 1);
1273 bchunk = z->z_RChunks;
1275 chunk->c_Next = bchunk;
1278 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1283 * We have to signal the remote cpu if our actions will cause
1284 * the remote zone to be placed back on ZoneAry so it can
1285 * move the zone back on.
1287 * We only need to deal with NULL->non-NULL RChunk transitions
1288 * and only if z_RSignal is set. We interlock by reading rsignal
1289 * before adding our chunk to RChunks. This should result in
1290 * virtually no IPI traffic.
1292 * We can use a passive IPI to reduce overhead even further.
1294 if (bchunk == NULL && rsignal) {
1295 logmemory(free_request, ptr, type,
1296 (unsigned long)z->z_ChunkSize, 0);
1297 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
1298 /* z can get ripped out from under us from this point on */
1299 } else if (rsignal) {
1300 atomic_subtract_int(&z->z_RCount, 1);
1301 /* z can get ripped out from under us from this point on */
1303 logmemory_quick(free_end);
1310 logmemory(free_chunk, ptr, type, (unsigned long)z->z_ChunkSize, 0);
1314 chunk_mark_free(z, chunk);
1317 * Put weird data into the memory to detect modifications after freeing,
1318 * illegal pointer use after freeing (we should fault on the odd address),
1319 * and so forth. XXX needs more work, see the old malloc code.
1322 if (z->z_ChunkSize < sizeof(weirdary))
1323 bcopy(weirdary, chunk, z->z_ChunkSize);
1325 bcopy(weirdary, chunk, sizeof(weirdary));
1329 * Add this free non-zero'd chunk to a linked list for reuse. Add
1330 * to the front of the linked list so it is more likely to be
1331 * reallocated, since it is already in our L1 cache.
1334 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
1335 panic("BADFREE %p", chunk);
1337 chunk->c_Next = z->z_LChunks;
1338 z->z_LChunks = chunk;
1339 if (chunk->c_Next == NULL)
1340 z->z_LChunksp = &chunk->c_Next;
1343 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
1348 * Bump the number of free chunks. If it becomes non-zero the zone
1349 * must be added back onto the appropriate list.
1351 if (z->z_NFree++ == 0) {
1352 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1353 slgd->ZoneAry[z->z_ZoneIndex] = z;
1356 --type->ks_inuse[z->z_Cpu];
1357 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
1359 check_zone_free(slgd, z);
1360 logmemory_quick(free_end);
1365 * Cleanup slabs which are hanging around due to RChunks. Called once every
1366 * 10 seconds on all cpus.
1371 SLGlobalData *slgd = &mycpu->gd_slab;
1376 for (i = 0; i < NZONES; ++i) {
1377 if ((z = slgd->ZoneAry[i]) == NULL)
1382 * Scan zones starting with the second zone in each list.
1386 * Shift all RChunks to the end of the LChunks list. This is
1387 * an O(1) operation.
1389 clean_zone_rchunks(z);
1390 z = check_zone_free(slgd, z);
1396 #if defined(INVARIANTS)
1399 * Helper routines for sanity checks
1403 chunk_mark_allocated(SLZone *z, void *chunk)
1405 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1408 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1409 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1410 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
1411 bitptr = &z->z_Bitmap[bitdex >> 5];
1413 KASSERT((*bitptr & (1 << bitdex)) == 0,
1414 ("memory chunk %p is already allocated!", chunk));
1415 *bitptr |= 1 << bitdex;
1420 chunk_mark_free(SLZone *z, void *chunk)
1422 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1425 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1426 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1427 ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
1428 bitptr = &z->z_Bitmap[bitdex >> 5];
1430 KASSERT((*bitptr & (1 << bitdex)) != 0,
1431 ("memory chunk %p is already free!", chunk));
1432 *bitptr &= ~(1 << bitdex);
1440 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1441 * specified alignment. M_* flags are expected in the flags field.
1443 * Alignment must be a multiple of PAGE_SIZE.
1445 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1446 * but when we move zalloc() over to use this function as its backend
1447 * we will have to switch to kreserve/krelease and call reserve(0)
1448 * after the new space is made available.
1450 * Interrupt code which has preempted other code is not allowed to
1451 * use PQ_CACHE pages. However, if an interrupt thread is run
1452 * non-preemptively or blocks and then runs non-preemptively, then
1453 * it is free to use PQ_CACHE pages. <--- may not apply any longer XXX
1456 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1460 int count, vmflags, base_vmflags;
1461 vm_page_t mbase = NULL;
1465 size = round_page(size);
1466 addr = vm_map_min(&kernel_map);
1468 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1470 vm_map_lock(&kernel_map);
1471 if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) {
1472 vm_map_unlock(&kernel_map);
1473 if ((flags & M_NULLOK) == 0)
1474 panic("kmem_slab_alloc(): kernel_map ran out of space!");
1475 vm_map_entry_release(count);
1481 * kernel_object maps 1:1 to kernel_map.
1483 vm_object_hold(&kernel_object);
1484 vm_object_reference_locked(&kernel_object);
1485 vm_map_insert(&kernel_map, &count,
1486 &kernel_object, NULL,
1487 addr, addr, addr + size,
1489 VM_PROT_ALL, VM_PROT_ALL,
1491 vm_object_drop(&kernel_object);
1492 vm_map_set_wired_quick(&kernel_map, addr, size, &count);
1493 vm_map_unlock(&kernel_map);
1499 base_vmflags |= VM_ALLOC_ZERO;
1500 if (flags & M_USE_RESERVE)
1501 base_vmflags |= VM_ALLOC_SYSTEM;
1502 if (flags & M_USE_INTERRUPT_RESERVE)
1503 base_vmflags |= VM_ALLOC_INTERRUPT;
1504 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1505 panic("kmem_slab_alloc: bad flags %08x (%p)",
1506 flags, ((int **)&size)[-1]);
1510 * Allocate the pages. Do not mess with the PG_ZERO flag or map
1511 * them yet. VM_ALLOC_NORMAL can only be set if we are not preempting.
1513 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1514 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1515 * implied in this case), though I'm not sure if we really need to
1518 vmflags = base_vmflags;
1519 if (flags & M_WAITOK) {
1520 if (td->td_preempted)
1521 vmflags |= VM_ALLOC_SYSTEM;
1523 vmflags |= VM_ALLOC_NORMAL;
1526 vm_object_hold(&kernel_object);
1527 for (i = 0; i < size; i += PAGE_SIZE) {
1528 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
1533 * If the allocation failed we either return NULL or we retry.
1535 * If M_WAITOK is specified we wait for more memory and retry.
1536 * If M_WAITOK is specified from a preemption we yield instead of
1537 * wait. Livelock will not occur because the interrupt thread
1538 * will not be preempting anyone the second time around after the
1542 if (flags & M_WAITOK) {
1543 if (td->td_preempted) {
1548 i -= PAGE_SIZE; /* retry */
1556 * Check and deal with an allocation failure
1561 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1562 /* page should already be busy */
1565 vm_map_lock(&kernel_map);
1566 vm_map_delete(&kernel_map, addr, addr + size, &count);
1567 vm_map_unlock(&kernel_map);
1568 vm_object_drop(&kernel_object);
1570 vm_map_entry_release(count);
1578 * NOTE: The VM pages are still busied. mbase points to the first one
1579 * but we have to iterate via vm_page_next()
1581 vm_object_drop(&kernel_object);
1585 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1592 * page should already be busy
1594 m->valid = VM_PAGE_BITS_ALL;
1596 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL | VM_PROT_NOSYNC,
1598 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1599 bzero((char *)addr + i, PAGE_SIZE);
1600 vm_page_flag_clear(m, PG_ZERO);
1601 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1602 vm_page_flag_set(m, PG_REFERENCED);
1606 vm_object_hold(&kernel_object);
1607 m = vm_page_next(m);
1608 vm_object_drop(&kernel_object);
1611 vm_map_entry_release(count);
1612 atomic_add_long(&SlabsAllocated, 1);
1613 return((void *)addr);
1620 kmem_slab_free(void *ptr, vm_size_t size)
1623 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
1624 atomic_add_long(&SlabsFreed, 1);
1629 kmalloc_cachealign(unsigned long size_alloc, struct malloc_type *type,
1632 #if (__VM_CACHELINE_SIZE == 32)
1633 #define CAN_CACHEALIGN(sz) ((sz) >= 256)
1634 #elif (__VM_CACHELINE_SIZE == 64)
1635 #define CAN_CACHEALIGN(sz) ((sz) >= 512)
1636 #elif (__VM_CACHELINE_SIZE == 128)
1637 #define CAN_CACHEALIGN(sz) ((sz) >= 1024)
1639 #error "unsupported cacheline size"
1644 if (size_alloc < __VM_CACHELINE_SIZE)
1645 size_alloc = __VM_CACHELINE_SIZE;
1646 else if (!CAN_CACHEALIGN(size_alloc))
1647 flags |= M_POWEROF2;
1649 ret = kmalloc(size_alloc, type, flags);
1650 KASSERT(((uintptr_t)ret & (__VM_CACHELINE_SIZE - 1)) == 0,
1651 ("%p(%lu) not cacheline %d aligned",
1652 ret, size_alloc, __VM_CACHELINE_SIZE));
1655 #undef CAN_CACHEALIGN