4 * KERN_SLABALLOC.C - Kernel SLAB memory allocator
6 * Copyright (c) 2003,2004,2010 The DragonFly Project. All rights reserved.
8 * This code is derived from software contributed to The DragonFly Project
9 * by Matthew Dillon <dillon@backplane.com>
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in
19 * the documentation and/or other materials provided with the
21 * 3. Neither the name of The DragonFly Project nor the names of its
22 * contributors may be used to endorse or promote products derived
23 * from this software without specific, prior written permission.
25 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
26 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
27 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
28 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
29 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
30 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
31 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
32 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
33 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
34 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
35 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * This module implements a slab allocator drop-in replacement for the
41 * A slab allocator reserves a ZONE for each chunk size, then lays the
42 * chunks out in an array within the zone. Allocation and deallocation
43 * is nearly instantanious, and fragmentation/overhead losses are limited
44 * to a fixed worst-case amount.
46 * The downside of this slab implementation is in the chunk size
47 * multiplied by the number of zones. ~80 zones * 128K = 10MB of VM per cpu.
48 * In a kernel implementation all this memory will be physical so
49 * the zone size is adjusted downward on machines with less physical
50 * memory. The upside is that overhead is bounded... this is the *worst*
53 * Slab management is done on a per-cpu basis and no locking or mutexes
54 * are required, only a critical section. When one cpu frees memory
55 * belonging to another cpu's slab manager an asynchronous IPI message
56 * will be queued to execute the operation. In addition, both the
57 * high level slab allocator and the low level zone allocator optimize
58 * M_ZERO requests, and the slab allocator does not have to pre initialize
59 * the linked list of chunks.
61 * XXX Balancing is needed between cpus. Balance will be handled through
62 * asynchronous IPIs primarily by reassigning the z_Cpu ownership of chunks.
64 * XXX If we have to allocate a new zone and M_USE_RESERVE is set, use of
65 * the new zone should be restricted to M_USE_RESERVE requests only.
67 * Alloc Size Chunking Number of zones
77 * (if PAGE_SIZE is 4K the maximum zone allocation is 16383)
79 * Allocations >= ZoneLimit go directly to kmem.
81 * API REQUIREMENTS AND SIDE EFFECTS
83 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we
84 * have remained compatible with the following API requirements:
86 * + small power-of-2 sized allocations are power-of-2 aligned (kern_tty)
87 * + all power-of-2 sized allocations are power-of-2 aligned (twe)
88 * + malloc(0) is allowed and returns non-NULL (ahc driver)
89 * + ability to allocate arbitrarily large chunks of memory
94 #include <sys/param.h>
95 #include <sys/systm.h>
96 #include <sys/kernel.h>
97 #include <sys/slaballoc.h>
99 #include <sys/vmmeter.h>
100 #include <sys/lock.h>
101 #include <sys/thread.h>
102 #include <sys/globaldata.h>
103 #include <sys/sysctl.h>
107 #include <vm/vm_param.h>
108 #include <vm/vm_kern.h>
109 #include <vm/vm_extern.h>
110 #include <vm/vm_object.h>
112 #include <vm/vm_map.h>
113 #include <vm/vm_page.h>
114 #include <vm/vm_pageout.h>
116 #include <machine/cpu.h>
118 #include <sys/thread2.h>
120 #define btokup(z) (&pmap_kvtom((vm_offset_t)(z))->ku_pagecnt)
122 #define MEMORY_STRING "ptr=%p type=%p size=%lu flags=%04x"
123 #define MEMORY_ARGS void *ptr, void *type, unsigned long size, int flags
125 #if !defined(KTR_MEMORY)
126 #define KTR_MEMORY KTR_ALL
128 KTR_INFO_MASTER(memory);
129 KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin");
130 KTR_INFO(KTR_MEMORY, memory, malloc_end, 1, MEMORY_STRING, MEMORY_ARGS);
131 KTR_INFO(KTR_MEMORY, memory, free_zero, 2, MEMORY_STRING, MEMORY_ARGS);
132 KTR_INFO(KTR_MEMORY, memory, free_ovsz, 3, MEMORY_STRING, MEMORY_ARGS);
133 KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 4, MEMORY_STRING, MEMORY_ARGS);
134 KTR_INFO(KTR_MEMORY, memory, free_chunk, 5, MEMORY_STRING, MEMORY_ARGS);
136 KTR_INFO(KTR_MEMORY, memory, free_request, 6, MEMORY_STRING, MEMORY_ARGS);
137 KTR_INFO(KTR_MEMORY, memory, free_rem_beg, 7, MEMORY_STRING, MEMORY_ARGS);
138 KTR_INFO(KTR_MEMORY, memory, free_rem_end, 8, MEMORY_STRING, MEMORY_ARGS);
140 KTR_INFO(KTR_MEMORY, memory, free_beg, 9, "free begin");
141 KTR_INFO(KTR_MEMORY, memory, free_end, 10, "free end");
143 #define logmemory(name, ptr, type, size, flags) \
144 KTR_LOG(memory_ ## name, ptr, type, size, flags)
145 #define logmemory_quick(name) \
146 KTR_LOG(memory_ ## name)
149 * Fixed globals (not per-cpu)
152 static int ZoneLimit;
153 static int ZonePageCount;
154 static uintptr_t ZoneMask;
155 static int ZoneBigAlloc; /* in KB */
156 static int ZoneGenAlloc; /* in KB */
157 struct malloc_type *kmemstatistics; /* exported to vmstat */
158 static int32_t weirdary[16];
160 static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
161 static void kmem_slab_free(void *ptr, vm_size_t bytes);
163 #if defined(INVARIANTS)
164 static void chunk_mark_allocated(SLZone *z, void *chunk);
165 static void chunk_mark_free(SLZone *z, void *chunk);
167 #define chunk_mark_allocated(z, chunk)
168 #define chunk_mark_free(z, chunk)
172 * Misc constants. Note that allocations that are exact multiples of
173 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
174 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
176 #define MIN_CHUNK_SIZE 8 /* in bytes */
177 #define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
178 #define ZONE_RELS_THRESH 32 /* threshold number of zones */
179 #define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
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, "");
230 * Returns the kernel memory size limit for the purposes of initializing
231 * various subsystem caches. The smaller of available memory and the KVM
232 * memory space is returned.
234 * The size in megabytes is returned.
241 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
242 if (limsize > KvaSize)
244 return (limsize / (1024 * 1024));
248 kmeminit(void *dummy)
254 limsize = kmem_lim_size();
255 usesize = (int)(limsize * 1024); /* convert to KB */
258 * If the machine has a large KVM space and more than 8G of ram,
259 * double the zone release threshold to reduce SMP invalidations.
260 * If more than 16G of ram, do it again.
262 * The BIOS eats a little ram so add some slop. We want 8G worth of
263 * memory sticks to trigger the first adjustment.
265 if (ZoneRelsThresh == ZONE_RELS_THRESH) {
266 if (limsize >= 7 * 1024)
268 if (limsize >= 15 * 1024)
273 * Calculate the zone size. This typically calculates to
274 * ZALLOC_MAX_ZONE_SIZE
276 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
277 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
279 ZoneLimit = ZoneSize / 4;
280 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
281 ZoneLimit = ZALLOC_ZONE_LIMIT;
282 ZoneMask = ~(uintptr_t)(ZoneSize - 1);
283 ZonePageCount = ZoneSize / PAGE_SIZE;
285 for (i = 0; i < NELEM(weirdary); ++i)
286 weirdary[i] = WEIRD_ADDR;
288 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
291 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
295 * Initialize a malloc type tracking structure.
298 malloc_init(void *data)
300 struct malloc_type *type = data;
303 if (type->ks_magic != M_MAGIC)
304 panic("malloc type lacks magic");
306 if (type->ks_limit != 0)
309 if (vmstats.v_page_count == 0)
310 panic("malloc_init not allowed before vm init");
312 limsize = kmem_lim_size() * (1024 * 1024);
313 type->ks_limit = limsize / 10;
315 type->ks_next = kmemstatistics;
316 kmemstatistics = type;
320 malloc_uninit(void *data)
322 struct malloc_type *type = data;
323 struct malloc_type *t;
329 if (type->ks_magic != M_MAGIC)
330 panic("malloc type lacks magic");
332 if (vmstats.v_page_count == 0)
333 panic("malloc_uninit not allowed before vm init");
335 if (type->ks_limit == 0)
336 panic("malloc_uninit on uninitialized type");
339 /* Make sure that all pending kfree()s are finished. */
340 lwkt_synchronize_ipiqs("muninit");
345 * memuse is only correct in aggregation. Due to memory being allocated
346 * on one cpu and freed on another individual array entries may be
347 * negative or positive (canceling each other out).
349 for (i = ttl = 0; i < ncpus; ++i)
350 ttl += type->ks_memuse[i];
352 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
353 ttl, type->ks_shortdesc, i);
356 if (type == kmemstatistics) {
357 kmemstatistics = type->ks_next;
359 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
360 if (t->ks_next == type) {
361 t->ks_next = type->ks_next;
366 type->ks_next = NULL;
371 * Increase the kmalloc pool limit for the specified pool. No changes
372 * are the made if the pool would shrink.
375 kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
377 if (type->ks_limit == 0)
381 if (type->ks_limit < bytes)
382 type->ks_limit = bytes;
386 * Dynamically create a malloc pool. This function is a NOP if *typep is
390 kmalloc_create(struct malloc_type **typep, const char *descr)
392 struct malloc_type *type;
394 if (*typep == NULL) {
395 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
396 type->ks_magic = M_MAGIC;
397 type->ks_shortdesc = descr;
404 * Destroy a dynamically created malloc pool. This function is a NOP if
405 * the pool has already been destroyed.
408 kmalloc_destroy(struct malloc_type **typep)
410 if (*typep != NULL) {
411 malloc_uninit(*typep);
412 kfree(*typep, M_TEMP);
418 * Calculate the zone index for the allocation request size and set the
419 * allocation request size to that particular zone's chunk size.
422 zoneindex(unsigned long *bytes)
424 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
426 *bytes = n = (n + 7) & ~7;
427 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
430 *bytes = n = (n + 15) & ~15;
435 *bytes = n = (n + 31) & ~31;
439 *bytes = n = (n + 63) & ~63;
443 *bytes = n = (n + 127) & ~127;
444 return(n / 128 + 31);
447 *bytes = n = (n + 255) & ~255;
448 return(n / 256 + 39);
450 *bytes = n = (n + 511) & ~511;
451 return(n / 512 + 47);
453 #if ZALLOC_ZONE_LIMIT > 8192
455 *bytes = n = (n + 1023) & ~1023;
456 return(n / 1024 + 55);
459 #if ZALLOC_ZONE_LIMIT > 16384
461 *bytes = n = (n + 2047) & ~2047;
462 return(n / 2048 + 63);
465 panic("Unexpected byte count %d", n);
471 * Used to debug memory corruption issues. Record up to (typically 32)
472 * allocation sources for this zone (for a particular chunk size).
476 slab_record_source(SLZone *z, const char *file, int line)
479 int b = line & (SLAB_DEBUG_ENTRIES - 1);
483 if (z->z_Sources[i].file == file && z->z_Sources[i].line == line)
485 if (z->z_Sources[i].file == NULL)
487 i = (i + 1) & (SLAB_DEBUG_ENTRIES - 1);
489 z->z_Sources[i].file = file;
490 z->z_Sources[i].line = line;
496 * kmalloc() (SLAB ALLOCATOR)
498 * Allocate memory via the slab allocator. If the request is too large,
499 * or if it page-aligned beyond a certain size, we fall back to the
500 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
501 * &SlabMisc if you don't care.
503 * M_RNOWAIT - don't block.
504 * M_NULLOK - return NULL instead of blocking.
505 * M_ZERO - zero the returned memory.
506 * M_USE_RESERVE - allow greater drawdown of the free list
507 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
514 kmalloc_debug(unsigned long size, struct malloc_type *type, int flags,
515 const char *file, int line)
518 kmalloc(unsigned long size, struct malloc_type *type, int flags)
527 struct globaldata *gd;
533 logmemory_quick(malloc_beg);
538 * XXX silly to have this in the critical path.
540 if (type->ks_limit == 0) {
542 if (type->ks_limit == 0)
549 * Handle the case where the limit is reached. Panic if we can't return
550 * NULL. The original malloc code looped, but this tended to
551 * simply deadlock the computer.
553 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
554 * to determine if a more complete limit check should be done. The
555 * actual memory use is tracked via ks_memuse[cpu].
557 while (type->ks_loosememuse >= type->ks_limit) {
561 for (i = ttl = 0; i < ncpus; ++i)
562 ttl += type->ks_memuse[i];
563 type->ks_loosememuse = ttl; /* not MP synchronized */
564 if ((ssize_t)ttl < 0) /* deal with occassional race */
566 if (ttl >= type->ks_limit) {
567 if (flags & M_NULLOK) {
568 logmemory(malloc_end, NULL, type, size, flags);
571 panic("%s: malloc limit exceeded", type->ks_shortdesc);
576 * Handle the degenerate size == 0 case. Yes, this does happen.
577 * Return a special pointer. This is to maintain compatibility with
578 * the original malloc implementation. Certain devices, such as the
579 * adaptec driver, not only allocate 0 bytes, they check for NULL and
580 * also realloc() later on. Joy.
583 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
584 return(ZERO_LENGTH_PTR);
588 * Handle hysteresis from prior frees here in malloc(). We cannot
589 * safely manipulate the kernel_map in free() due to free() possibly
590 * being called via an IPI message or from sensitive interrupt code.
592 * NOTE: ku_pagecnt must be cleared before we free the slab or we
593 * might race another cpu allocating the kva and setting
596 while (slgd->NFreeZones > ZoneRelsThresh && (flags & M_RNOWAIT) == 0) {
598 if (slgd->NFreeZones > ZoneRelsThresh) { /* crit sect race */
602 slgd->FreeZones = z->z_Next;
606 kmem_slab_free(z, ZoneSize); /* may block */
607 atomic_add_int(&ZoneGenAlloc, -(int)ZoneSize / 1024);
613 * XXX handle oversized frees that were queued from kfree().
615 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) {
617 if ((z = slgd->FreeOvZones) != NULL) {
620 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
621 slgd->FreeOvZones = z->z_Next;
622 tsize = z->z_ChunkSize;
623 kmem_slab_free(z, tsize); /* may block */
624 atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024);
630 * Handle large allocations directly. There should not be very many of
631 * these so performance is not a big issue.
633 * The backend allocator is pretty nasty on a SMP system. Use the
634 * slab allocator for one and two page-sized chunks even though we lose
635 * some efficiency. XXX maybe fix mmio and the elf loader instead.
637 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
640 size = round_page(size);
641 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
643 logmemory(malloc_end, NULL, type, size, flags);
646 atomic_add_int(&ZoneBigAlloc, (int)size / 1024);
647 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
648 flags |= M_PASSIVE_ZERO;
650 *kup = size / PAGE_SIZE;
656 * Attempt to allocate out of an existing zone. First try the free list,
657 * then allocate out of unallocated space. If we find a good zone move
658 * it to the head of the list so later allocations find it quickly
659 * (we might have thousands of zones in the list).
661 * Note: zoneindex() will panic of size is too large.
663 zi = zoneindex(&size);
664 KKASSERT(zi < NZONES);
667 if ((z = slgd->ZoneAry[zi]) != NULL) {
669 * Locate a chunk - we have to have at least one. If this is the
670 * last chunk go ahead and do the work to retrieve chunks freed
671 * from remote cpus, and if the zone is still empty move it off
674 if (--z->z_NFree <= 0) {
675 KKASSERT(z->z_NFree == 0);
679 * WARNING! This code competes with other cpus. It is ok
680 * for us to not drain RChunks here but we might as well, and
681 * it is ok if more accumulate after we're done.
683 * Set RSignal before pulling rchunks off, indicating that we
684 * will be moving ourselves off of the ZoneAry. Remote ends will
685 * read RSignal before putting rchunks on thus interlocking
686 * their IPI signaling.
688 if (z->z_RChunks == NULL)
689 atomic_swap_int(&z->z_RSignal, 1);
691 while ((bchunk = z->z_RChunks) != NULL) {
693 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
694 *z->z_LChunksp = bchunk;
696 chunk_mark_free(z, bchunk);
697 z->z_LChunksp = &bchunk->c_Next;
698 bchunk = bchunk->c_Next;
706 * Remove from the zone list if no free chunks remain.
709 if (z->z_NFree == 0) {
710 slgd->ZoneAry[zi] = z->z_Next;
718 * Fast path, we have chunks available in z_LChunks.
720 chunk = z->z_LChunks;
722 chunk_mark_allocated(z, chunk);
723 z->z_LChunks = chunk->c_Next;
724 if (z->z_LChunks == NULL)
725 z->z_LChunksp = &z->z_LChunks;
727 slab_record_source(z, file, line);
733 * No chunks are available in LChunks, the free chunk MUST be
734 * in the never-before-used memory area, controlled by UIndex.
736 * The consequences are very serious if our zone got corrupted so
737 * we use an explicit panic rather than a KASSERT.
739 if (z->z_UIndex + 1 != z->z_NMax)
744 if (z->z_UIndex == z->z_UEndIndex)
745 panic("slaballoc: corrupted zone");
747 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
748 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
750 flags |= M_PASSIVE_ZERO;
752 chunk_mark_allocated(z, chunk);
754 slab_record_source(z, file, line);
760 * If all zones are exhausted we need to allocate a new zone for this
761 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
762 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
763 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
764 * we do not pre-zero it because we do not want to mess up the L1 cache.
766 * At least one subsystem, the tty code (see CROUND) expects power-of-2
767 * allocations to be power-of-2 aligned. We maintain compatibility by
768 * adjusting the base offset below.
774 if ((z = slgd->FreeZones) != NULL) {
775 slgd->FreeZones = z->z_Next;
777 bzero(z, sizeof(SLZone));
778 z->z_Flags |= SLZF_UNOTZEROD;
780 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
783 atomic_add_int(&ZoneGenAlloc, (int)ZoneSize / 1024);
787 * How big is the base structure?
789 #if defined(INVARIANTS)
791 * Make room for z_Bitmap. An exact calculation is somewhat more
792 * complicated so don't make an exact calculation.
794 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
795 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
797 off = sizeof(SLZone);
801 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
802 * Otherwise just 8-byte align the data.
804 if ((size | (size - 1)) + 1 == (size << 1))
805 off = (off + size - 1) & ~(size - 1);
807 off = (off + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
808 z->z_Magic = ZALLOC_SLAB_MAGIC;
810 z->z_NMax = (ZoneSize - off) / size;
811 z->z_NFree = z->z_NMax - 1;
812 z->z_BasePtr = (char *)z + off;
813 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
814 z->z_ChunkSize = size;
816 z->z_Cpu = gd->gd_cpuid;
817 z->z_LChunksp = &z->z_LChunks;
819 bcopy(z->z_Sources, z->z_AltSources, sizeof(z->z_Sources));
820 bzero(z->z_Sources, sizeof(z->z_Sources));
822 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
823 z->z_Next = slgd->ZoneAry[zi];
824 slgd->ZoneAry[zi] = z;
825 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
826 flags &= ~M_ZERO; /* already zero'd */
827 flags |= M_PASSIVE_ZERO;
830 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */
831 chunk_mark_allocated(z, chunk);
833 slab_record_source(z, file, line);
837 * Slide the base index for initial allocations out of the next
838 * zone we create so we do not over-weight the lower part of the
841 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
842 & (ZALLOC_MAX_ZONE_SIZE - 1);
846 ++type->ks_inuse[gd->gd_cpuid];
847 type->ks_memuse[gd->gd_cpuid] += size;
848 type->ks_loosememuse += size; /* not MP synchronized */
854 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
855 if (use_malloc_pattern) {
856 for (i = 0; i < size; i += sizeof(int)) {
857 *(int *)((char *)chunk + i) = -1;
860 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
863 logmemory(malloc_end, chunk, type, size, flags);
867 logmemory(malloc_end, NULL, type, size, flags);
872 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
874 * Generally speaking this routine is not called very often and we do
875 * not attempt to optimize it beyond reusing the same pointer if the
876 * new size fits within the chunking of the old pointer's zone.
880 krealloc_debug(void *ptr, unsigned long size,
881 struct malloc_type *type, int flags,
882 const char *file, int line)
885 krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
893 KKASSERT((flags & M_ZERO) == 0); /* not supported */
895 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
896 return(kmalloc_debug(size, type, flags, file, line));
903 * Handle oversized allocations. XXX we really should require that a
904 * size be passed to free() instead of this nonsense.
908 osize = *kup << PAGE_SHIFT;
909 if (osize == round_page(size))
911 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
913 bcopy(ptr, nptr, min(size, osize));
919 * Get the original allocation's zone. If the new request winds up
920 * using the same chunk size we do not have to do anything.
922 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
925 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
928 * Allocate memory for the new request size. Note that zoneindex has
929 * already adjusted the request size to the appropriate chunk size, which
930 * should optimize our bcopy(). Then copy and return the new pointer.
932 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
933 * necessary align the result.
935 * We can only zoneindex (to align size to the chunk size) if the new
936 * size is not too large.
938 if (size < ZoneLimit) {
940 if (z->z_ChunkSize == size)
943 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
945 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
951 * Return the kmalloc limit for this type, in bytes.
954 kmalloc_limit(struct malloc_type *type)
956 if (type->ks_limit == 0) {
958 if (type->ks_limit == 0)
962 return(type->ks_limit);
966 * Allocate a copy of the specified string.
968 * (MP SAFE) (MAY BLOCK)
972 kstrdup_debug(const char *str, struct malloc_type *type,
973 const char *file, int line)
976 kstrdup(const char *str, struct malloc_type *type)
979 int zlen; /* length inclusive of terminating NUL */
984 zlen = strlen(str) + 1;
985 nstr = kmalloc_debug(zlen, type, M_WAITOK, file, line);
986 bcopy(str, nstr, zlen);
992 * Notify our cpu that a remote cpu has freed some chunks in a zone that
993 * we own. RCount will be bumped so the memory should be good, but validate
998 kfree_remote(void *ptr)
1006 slgd = &mycpu->gd_slab;
1009 KKASSERT(*kup == -((int)mycpuid + 1));
1010 KKASSERT(z->z_RCount > 0);
1011 atomic_subtract_int(&z->z_RCount, 1);
1013 logmemory(free_rem_beg, z, NULL, 0L, 0);
1014 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1015 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
1019 * Indicate that we will no longer be off of the ZoneAry by
1026 * Atomically extract the bchunks list and then process it back
1027 * into the lchunks list. We want to append our bchunks to the
1028 * lchunks list and not prepend since we likely do not have
1029 * cache mastership of the related data (not that it helps since
1030 * we are using c_Next).
1032 while ((bchunk = z->z_RChunks) != NULL) {
1034 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
1035 *z->z_LChunksp = bchunk;
1037 chunk_mark_free(z, bchunk);
1038 z->z_LChunksp = &bchunk->c_Next;
1039 bchunk = bchunk->c_Next;
1045 if (z->z_NFree && nfree == 0) {
1046 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1047 slgd->ZoneAry[z->z_ZoneIndex] = z;
1051 * If the zone becomes totally free, and there are other zones we
1052 * can allocate from, move this zone to the FreeZones list. Since
1053 * this code can be called from an IPI callback, do *NOT* try to mess
1054 * with kernel_map here. Hysteresis will be performed at malloc() time.
1056 * Do not move the zone if there is an IPI inflight, otherwise MP
1057 * races can result in our free_remote code accessing a destroyed
1060 if (z->z_NFree == z->z_NMax &&
1061 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1067 for (pz = &slgd->ZoneAry[z->z_ZoneIndex];
1069 pz = &(*pz)->z_Next) {
1074 z->z_Next = slgd->FreeZones;
1075 slgd->FreeZones = z;
1080 logmemory(free_rem_end, z, bchunk, 0L, 0);
1086 * free (SLAB ALLOCATOR)
1088 * Free a memory block previously allocated by malloc. Note that we do not
1089 * attempt to update ks_loosememuse as MP races could prevent us from
1090 * checking memory limits in malloc.
1095 kfree(void *ptr, struct malloc_type *type)
1100 struct globaldata *gd;
1108 logmemory_quick(free_beg);
1110 slgd = &gd->gd_slab;
1113 panic("trying to free NULL pointer");
1116 * Handle special 0-byte allocations
1118 if (ptr == ZERO_LENGTH_PTR) {
1119 logmemory(free_zero, ptr, type, -1UL, 0);
1120 logmemory_quick(free_end);
1125 * Panic on bad malloc type
1127 if (type->ks_magic != M_MAGIC)
1128 panic("free: malloc type lacks magic");
1131 * Handle oversized allocations. XXX we really should require that a
1132 * size be passed to free() instead of this nonsense.
1134 * This code is never called via an ipi.
1138 size = *kup << PAGE_SHIFT;
1141 KKASSERT(sizeof(weirdary) <= size);
1142 bcopy(weirdary, ptr, sizeof(weirdary));
1145 * NOTE: For oversized allocations we do not record the
1146 * originating cpu. It gets freed on the cpu calling
1147 * kfree(). The statistics are in aggregate.
1149 * note: XXX we have still inherited the interrupts-can't-block
1150 * assumption. An interrupt thread does not bump
1151 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1152 * primarily until we can fix softupdate's assumptions about free().
1155 --type->ks_inuse[gd->gd_cpuid];
1156 type->ks_memuse[gd->gd_cpuid] -= size;
1157 if (mycpu->gd_intr_nesting_level ||
1158 (gd->gd_curthread->td_flags & TDF_INTTHREAD))
1160 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1162 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1163 z->z_Next = slgd->FreeOvZones;
1164 z->z_ChunkSize = size;
1165 slgd->FreeOvZones = z;
1169 logmemory(free_ovsz, ptr, type, size, 0);
1170 kmem_slab_free(ptr, size); /* may block */
1171 atomic_add_int(&ZoneBigAlloc, -(int)size / 1024);
1173 logmemory_quick(free_end);
1178 * Zone case. Figure out the zone based on the fact that it is
1181 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1184 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1187 * If we do not own the zone then use atomic ops to free to the
1188 * remote cpu linked list and notify the target zone using a
1191 * The target zone cannot be deallocated while we own a chunk of it,
1192 * so the zone header's storage is stable until the very moment
1193 * we adjust z_RChunks. After that we cannot safely dereference (z).
1195 * (no critical section needed)
1197 if (z->z_CpuGd != gd) {
1200 * Making these adjustments now allow us to avoid passing (type)
1201 * to the remote cpu. Note that ks_inuse/ks_memuse is being
1202 * adjusted on OUR cpu, not the zone cpu, but it should all still
1203 * sum up properly and cancel out.
1206 --type->ks_inuse[gd->gd_cpuid];
1207 type->ks_memuse[gd->gd_cpuid] -= z->z_ChunkSize;
1211 * WARNING! This code competes with other cpus. Once we
1212 * successfully link the chunk to RChunks the remote
1213 * cpu can rip z's storage out from under us.
1215 * Bumping RCount prevents z's storage from getting
1218 rsignal = z->z_RSignal;
1221 atomic_add_int(&z->z_RCount, 1);
1225 bchunk = z->z_RChunks;
1227 chunk->c_Next = bchunk;
1230 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1235 * We have to signal the remote cpu if our actions will cause
1236 * the remote zone to be placed back on ZoneAry so it can
1237 * move the zone back on.
1239 * We only need to deal with NULL->non-NULL RChunk transitions
1240 * and only if z_RSignal is set. We interlock by reading rsignal
1241 * before adding our chunk to RChunks. This should result in
1242 * virtually no IPI traffic.
1244 * We can use a passive IPI to reduce overhead even further.
1246 if (bchunk == NULL && rsignal) {
1247 logmemory(free_request, ptr, type, (unsigned long)z->z_ChunkSize, 0);
1248 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
1249 /* z can get ripped out from under us from this point on */
1250 } else if (rsignal) {
1251 atomic_subtract_int(&z->z_RCount, 1);
1252 /* z can get ripped out from under us from this point on */
1255 panic("Corrupt SLZone");
1257 logmemory_quick(free_end);
1264 logmemory(free_chunk, ptr, type, (unsigned long)z->z_ChunkSize, 0);
1268 chunk_mark_free(z, chunk);
1271 * Put weird data into the memory to detect modifications after freeing,
1272 * illegal pointer use after freeing (we should fault on the odd address),
1273 * and so forth. XXX needs more work, see the old malloc code.
1276 if (z->z_ChunkSize < sizeof(weirdary))
1277 bcopy(weirdary, chunk, z->z_ChunkSize);
1279 bcopy(weirdary, chunk, sizeof(weirdary));
1283 * Add this free non-zero'd chunk to a linked list for reuse. Add
1284 * to the front of the linked list so it is more likely to be
1285 * reallocated, since it is already in our L1 cache.
1288 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
1289 panic("BADFREE %p", chunk);
1291 chunk->c_Next = z->z_LChunks;
1292 z->z_LChunks = chunk;
1293 if (chunk->c_Next == NULL)
1294 z->z_LChunksp = &chunk->c_Next;
1297 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
1302 * Bump the number of free chunks. If it becomes non-zero the zone
1303 * must be added back onto the appropriate list.
1305 if (z->z_NFree++ == 0) {
1306 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1307 slgd->ZoneAry[z->z_ZoneIndex] = z;
1310 --type->ks_inuse[z->z_Cpu];
1311 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
1314 * If the zone becomes totally free, and there are other zones we
1315 * can allocate from, move this zone to the FreeZones list. Since
1316 * this code can be called from an IPI callback, do *NOT* try to mess
1317 * with kernel_map here. Hysteresis will be performed at malloc() time.
1319 if (z->z_NFree == z->z_NMax &&
1320 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1326 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
1330 z->z_Next = slgd->FreeZones;
1331 slgd->FreeZones = z;
1336 logmemory_quick(free_end);
1340 #if defined(INVARIANTS)
1343 * Helper routines for sanity checks
1347 chunk_mark_allocated(SLZone *z, void *chunk)
1349 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1352 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1353 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1354 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
1355 bitptr = &z->z_Bitmap[bitdex >> 5];
1357 KASSERT((*bitptr & (1 << bitdex)) == 0,
1358 ("memory chunk %p is already allocated!", chunk));
1359 *bitptr |= 1 << bitdex;
1364 chunk_mark_free(SLZone *z, void *chunk)
1366 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1369 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1370 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1371 ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
1372 bitptr = &z->z_Bitmap[bitdex >> 5];
1374 KASSERT((*bitptr & (1 << bitdex)) != 0,
1375 ("memory chunk %p is already free!", chunk));
1376 *bitptr &= ~(1 << bitdex);
1384 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1385 * specified alignment. M_* flags are expected in the flags field.
1387 * Alignment must be a multiple of PAGE_SIZE.
1389 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1390 * but when we move zalloc() over to use this function as its backend
1391 * we will have to switch to kreserve/krelease and call reserve(0)
1392 * after the new space is made available.
1394 * Interrupt code which has preempted other code is not allowed to
1395 * use PQ_CACHE pages. However, if an interrupt thread is run
1396 * non-preemptively or blocks and then runs non-preemptively, then
1397 * it is free to use PQ_CACHE pages. <--- may not apply any longer XXX
1400 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1404 int count, vmflags, base_vmflags;
1405 vm_page_t mbase = NULL;
1409 size = round_page(size);
1410 addr = vm_map_min(&kernel_map);
1412 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1414 vm_map_lock(&kernel_map);
1415 if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) {
1416 vm_map_unlock(&kernel_map);
1417 if ((flags & M_NULLOK) == 0)
1418 panic("kmem_slab_alloc(): kernel_map ran out of space!");
1419 vm_map_entry_release(count);
1425 * kernel_object maps 1:1 to kernel_map.
1427 vm_object_hold(&kernel_object);
1428 vm_object_reference_locked(&kernel_object);
1429 vm_map_insert(&kernel_map, &count,
1430 &kernel_object, addr, addr, addr + size,
1432 VM_PROT_ALL, VM_PROT_ALL,
1434 vm_object_drop(&kernel_object);
1435 vm_map_set_wired_quick(&kernel_map, addr, size, &count);
1436 vm_map_unlock(&kernel_map);
1442 base_vmflags |= VM_ALLOC_ZERO;
1443 if (flags & M_USE_RESERVE)
1444 base_vmflags |= VM_ALLOC_SYSTEM;
1445 if (flags & M_USE_INTERRUPT_RESERVE)
1446 base_vmflags |= VM_ALLOC_INTERRUPT;
1447 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1448 panic("kmem_slab_alloc: bad flags %08x (%p)",
1449 flags, ((int **)&size)[-1]);
1453 * Allocate the pages. Do not mess with the PG_ZERO flag or map
1454 * them yet. VM_ALLOC_NORMAL can only be set if we are not preempting.
1456 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1457 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1458 * implied in this case), though I'm not sure if we really need to
1461 vmflags = base_vmflags;
1462 if (flags & M_WAITOK) {
1463 if (td->td_preempted)
1464 vmflags |= VM_ALLOC_SYSTEM;
1466 vmflags |= VM_ALLOC_NORMAL;
1469 vm_object_hold(&kernel_object);
1470 for (i = 0; i < size; i += PAGE_SIZE) {
1471 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
1476 * If the allocation failed we either return NULL or we retry.
1478 * If M_WAITOK is specified we wait for more memory and retry.
1479 * If M_WAITOK is specified from a preemption we yield instead of
1480 * wait. Livelock will not occur because the interrupt thread
1481 * will not be preempting anyone the second time around after the
1485 if (flags & M_WAITOK) {
1486 if (td->td_preempted) {
1491 i -= PAGE_SIZE; /* retry */
1499 * Check and deal with an allocation failure
1504 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1505 /* page should already be busy */
1508 vm_map_lock(&kernel_map);
1509 vm_map_delete(&kernel_map, addr, addr + size, &count);
1510 vm_map_unlock(&kernel_map);
1511 vm_object_drop(&kernel_object);
1513 vm_map_entry_release(count);
1521 * NOTE: The VM pages are still busied. mbase points to the first one
1522 * but we have to iterate via vm_page_next()
1524 vm_object_drop(&kernel_object);
1528 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1535 * page should already be busy
1537 m->valid = VM_PAGE_BITS_ALL;
1539 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL | VM_PROT_NOSYNC, 1);
1540 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1541 bzero((char *)addr + i, PAGE_SIZE);
1542 vm_page_flag_clear(m, PG_ZERO);
1543 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1544 vm_page_flag_set(m, PG_REFERENCED);
1548 vm_object_hold(&kernel_object);
1549 m = vm_page_next(m);
1550 vm_object_drop(&kernel_object);
1553 vm_map_entry_release(count);
1554 return((void *)addr);
1561 kmem_slab_free(void *ptr, vm_size_t size)
1564 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);