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");
197 MALLOC_DEFINE(M_DRM
, "m_drm", "DRM memory allocations");
199 MALLOC_DEFINE(M_IP6OPT
, "ip6opt", "IPv6 options");
200 MALLOC_DEFINE(M_IP6NDP
, "ip6ndp", "IPv6 Neighbor Discovery");
203 * Initialize the slab memory allocator. We have to choose a zone size based
204 * on available physical memory. We choose a zone side which is approximately
205 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
206 * 128K. The zone size is limited to the bounds set in slaballoc.h
207 * (typically 32K min, 128K max).
209 static void kmeminit(void *dummy
);
213 SYSINIT(kmem
, SI_BOOT1_ALLOCATOR
, SI_ORDER_FIRST
, kmeminit
, NULL
);
217 * If enabled any memory allocated without M_ZERO is initialized to -1.
219 static int use_malloc_pattern
;
220 SYSCTL_INT(_debug
, OID_AUTO
, use_malloc_pattern
, CTLFLAG_RW
,
221 &use_malloc_pattern
, 0,
222 "Initialize memory to -1 if M_ZERO not specified");
225 static int ZoneRelsThresh
= ZONE_RELS_THRESH
;
226 SYSCTL_INT(_kern
, OID_AUTO
, zone_big_alloc
, CTLFLAG_RD
, &ZoneBigAlloc
, 0, "");
227 SYSCTL_INT(_kern
, OID_AUTO
, zone_gen_alloc
, CTLFLAG_RD
, &ZoneGenAlloc
, 0, "");
228 SYSCTL_INT(_kern
, OID_AUTO
, zone_cache
, CTLFLAG_RW
, &ZoneRelsThresh
, 0, "");
229 static long SlabsAllocated
;
230 static long SlabsFreed
;
231 SYSCTL_LONG(_kern
, OID_AUTO
, slabs_allocated
, CTLFLAG_RD
,
232 &SlabsAllocated
, 0, "");
233 SYSCTL_LONG(_kern
, OID_AUTO
, slabs_freed
, CTLFLAG_RD
,
235 static int SlabFreeToTail
;
236 SYSCTL_INT(_kern
, OID_AUTO
, slab_freetotail
, CTLFLAG_RW
,
237 &SlabFreeToTail
, 0, "");
240 * Returns the kernel memory size limit for the purposes of initializing
241 * various subsystem caches. The smaller of available memory and the KVM
242 * memory space is returned.
244 * The size in megabytes is returned.
251 limsize
= (size_t)vmstats
.v_page_count
* PAGE_SIZE
;
252 if (limsize
> KvaSize
)
254 return (limsize
/ (1024 * 1024));
258 kmeminit(void *dummy
)
264 limsize
= kmem_lim_size();
265 usesize
= (int)(limsize
* 1024); /* convert to KB */
268 * If the machine has a large KVM space and more than 8G of ram,
269 * double the zone release threshold to reduce SMP invalidations.
270 * If more than 16G of ram, do it again.
272 * The BIOS eats a little ram so add some slop. We want 8G worth of
273 * memory sticks to trigger the first adjustment.
275 if (ZoneRelsThresh
== ZONE_RELS_THRESH
) {
276 if (limsize
>= 7 * 1024)
278 if (limsize
>= 15 * 1024)
283 * Calculate the zone size. This typically calculates to
284 * ZALLOC_MAX_ZONE_SIZE
286 ZoneSize
= ZALLOC_MIN_ZONE_SIZE
;
287 while (ZoneSize
< ZALLOC_MAX_ZONE_SIZE
&& (ZoneSize
<< 1) < usesize
)
289 ZoneLimit
= ZoneSize
/ 4;
290 if (ZoneLimit
> ZALLOC_ZONE_LIMIT
)
291 ZoneLimit
= ZALLOC_ZONE_LIMIT
;
292 ZoneMask
= ~(uintptr_t)(ZoneSize
- 1);
293 ZonePageCount
= ZoneSize
/ PAGE_SIZE
;
295 for (i
= 0; i
< NELEM(weirdary
); ++i
)
296 weirdary
[i
] = WEIRD_ADDR
;
298 ZeroPage
= kmem_slab_alloc(PAGE_SIZE
, PAGE_SIZE
, M_WAITOK
|M_ZERO
);
301 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize
/ 1024);
305 * (low level) Initialize slab-related elements in the globaldata structure.
307 * Occurs after kmeminit().
310 slab_gdinit(globaldata_t gd
)
316 for (i
= 0; i
< NZONES
; ++i
)
317 TAILQ_INIT(&slgd
->ZoneAry
[i
]);
318 TAILQ_INIT(&slgd
->FreeZones
);
319 TAILQ_INIT(&slgd
->FreeOvZones
);
323 * Initialize a malloc type tracking structure.
326 malloc_init(void *data
)
328 struct malloc_type
*type
= data
;
331 if (type
->ks_magic
!= M_MAGIC
)
332 panic("malloc type lacks magic");
334 if (type
->ks_limit
!= 0)
337 if (vmstats
.v_page_count
== 0)
338 panic("malloc_init not allowed before vm init");
340 limsize
= kmem_lim_size() * (1024 * 1024);
341 type
->ks_limit
= limsize
/ 10;
343 type
->ks_next
= kmemstatistics
;
344 kmemstatistics
= type
;
348 malloc_uninit(void *data
)
350 struct malloc_type
*type
= data
;
351 struct malloc_type
*t
;
357 if (type
->ks_magic
!= M_MAGIC
)
358 panic("malloc type lacks magic");
360 if (vmstats
.v_page_count
== 0)
361 panic("malloc_uninit not allowed before vm init");
363 if (type
->ks_limit
== 0)
364 panic("malloc_uninit on uninitialized type");
366 /* Make sure that all pending kfree()s are finished. */
367 lwkt_synchronize_ipiqs("muninit");
371 * memuse is only correct in aggregation. Due to memory being allocated
372 * on one cpu and freed on another individual array entries may be
373 * negative or positive (canceling each other out).
375 for (i
= ttl
= 0; i
< ncpus
; ++i
)
376 ttl
+= type
->ks_use
[i
].memuse
;
378 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
379 ttl
, type
->ks_shortdesc
, i
);
382 if (type
== kmemstatistics
) {
383 kmemstatistics
= type
->ks_next
;
385 for (t
= kmemstatistics
; t
->ks_next
!= NULL
; t
= t
->ks_next
) {
386 if (t
->ks_next
== type
) {
387 t
->ks_next
= type
->ks_next
;
392 type
->ks_next
= NULL
;
397 * Increase the kmalloc pool limit for the specified pool. No changes
398 * are the made if the pool would shrink.
401 kmalloc_raise_limit(struct malloc_type
*type
, size_t bytes
)
403 if (type
->ks_limit
== 0)
407 if (type
->ks_limit
< bytes
)
408 type
->ks_limit
= bytes
;
412 kmalloc_set_unlimited(struct malloc_type
*type
)
414 type
->ks_limit
= kmem_lim_size() * (1024 * 1024);
418 * Dynamically create a malloc pool. This function is a NOP if *typep is
422 kmalloc_create(struct malloc_type
**typep
, const char *descr
)
424 struct malloc_type
*type
;
426 if (*typep
== NULL
) {
427 type
= kmalloc(sizeof(*type
), M_TEMP
, M_WAITOK
| M_ZERO
);
428 type
->ks_magic
= M_MAGIC
;
429 type
->ks_shortdesc
= descr
;
436 * Destroy a dynamically created malloc pool. This function is a NOP if
437 * the pool has already been destroyed.
440 kmalloc_destroy(struct malloc_type
**typep
)
442 if (*typep
!= NULL
) {
443 malloc_uninit(*typep
);
444 kfree(*typep
, M_TEMP
);
450 * Calculate the zone index for the allocation request size and set the
451 * allocation request size to that particular zone's chunk size.
454 zoneindex(unsigned long *bytes
, unsigned long *align
)
456 unsigned int n
= (unsigned int)*bytes
; /* unsigned for shift opt */
458 *bytes
= n
= (n
+ 7) & ~7;
460 return(n
/ 8 - 1); /* 8 byte chunks, 16 zones */
463 *bytes
= n
= (n
+ 15) & ~15;
469 *bytes
= n
= (n
+ 31) & ~31;
474 *bytes
= n
= (n
+ 63) & ~63;
479 *bytes
= n
= (n
+ 127) & ~127;
481 return(n
/ 128 + 31);
484 *bytes
= n
= (n
+ 255) & ~255;
486 return(n
/ 256 + 39);
488 *bytes
= n
= (n
+ 511) & ~511;
490 return(n
/ 512 + 47);
492 #if ZALLOC_ZONE_LIMIT > 8192
494 *bytes
= n
= (n
+ 1023) & ~1023;
496 return(n
/ 1024 + 55);
499 #if ZALLOC_ZONE_LIMIT > 16384
501 *bytes
= n
= (n
+ 2047) & ~2047;
503 return(n
/ 2048 + 63);
506 panic("Unexpected byte count %d", n
);
512 clean_zone_rchunks(SLZone
*z
)
516 while ((bchunk
= z
->z_RChunks
) != NULL
) {
518 if (atomic_cmpset_ptr(&z
->z_RChunks
, bchunk
, NULL
)) {
519 *z
->z_LChunksp
= bchunk
;
521 chunk_mark_free(z
, bchunk
);
522 z
->z_LChunksp
= &bchunk
->c_Next
;
523 bchunk
= bchunk
->c_Next
;
533 * If the zone becomes totally free and is not the only zone listed for a
534 * chunk size we move it to the FreeZones list. We always leave at least
535 * one zone per chunk size listed, even if it is freeable.
537 * Do not move the zone if there is an IPI in_flight (z_RCount != 0),
538 * otherwise MP races can result in our free_remote code accessing a
539 * destroyed zone. The remote end interlocks z_RCount with z_RChunks
540 * so one has to test both z_NFree and z_RCount.
542 * Since this code can be called from an IPI callback, do *NOT* try to mess
543 * with kernel_map here. Hysteresis will be performed at kmalloc() time.
547 check_zone_free(SLGlobalData
*slgd
, SLZone
*z
)
551 znext
= TAILQ_NEXT(z
, z_Entry
);
552 if (z
->z_NFree
== z
->z_NMax
&& z
->z_RCount
== 0 &&
553 (TAILQ_FIRST(&slgd
->ZoneAry
[z
->z_ZoneIndex
]) != z
|| znext
)
557 TAILQ_REMOVE(&slgd
->ZoneAry
[z
->z_ZoneIndex
], z
, z_Entry
);
560 TAILQ_INSERT_HEAD(&slgd
->FreeZones
, z
, z_Entry
);
570 * Used to debug memory corruption issues. Record up to (typically 32)
571 * allocation sources for this zone (for a particular chunk size).
575 slab_record_source(SLZone
*z
, const char *file
, int line
)
578 int b
= line
& (SLAB_DEBUG_ENTRIES
- 1);
582 if (z
->z_Sources
[i
].file
== file
&& z
->z_Sources
[i
].line
== line
)
584 if (z
->z_Sources
[i
].file
== NULL
)
586 i
= (i
+ 1) & (SLAB_DEBUG_ENTRIES
- 1);
588 z
->z_Sources
[i
].file
= file
;
589 z
->z_Sources
[i
].line
= line
;
594 static __inline
unsigned long
595 powerof2_size(unsigned long size
)
599 if (size
== 0 || powerof2(size
))
607 * kmalloc() (SLAB ALLOCATOR)
609 * Allocate memory via the slab allocator. If the request is too large,
610 * or if it page-aligned beyond a certain size, we fall back to the
611 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
612 * &SlabMisc if you don't care.
614 * M_RNOWAIT - don't block.
615 * M_NULLOK - return NULL instead of blocking.
616 * M_ZERO - zero the returned memory.
617 * M_USE_RESERVE - allow greater drawdown of the free list
618 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
619 * M_POWEROF2 - roundup size to the nearest power of 2
626 kmalloc_debug(unsigned long size
, struct malloc_type
*type
, int flags
,
627 const char *file
, int line
)
630 kmalloc(unsigned long size
, struct malloc_type
*type
, int flags
)
636 struct globaldata
*gd
;
643 logmemory_quick(malloc_beg
);
648 * XXX silly to have this in the critical path.
650 if (type
->ks_limit
== 0) {
657 if (flags
& M_POWEROF2
)
658 size
= powerof2_size(size
);
661 * Handle the case where the limit is reached. Panic if we can't return
662 * NULL. The original malloc code looped, but this tended to
663 * simply deadlock the computer.
665 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
666 * to determine if a more complete limit check should be done. The
667 * actual memory use is tracked via ks_use[cpu].memuse.
669 while (type
->ks_loosememuse
>= type
->ks_limit
) {
673 for (i
= ttl
= 0; i
< ncpus
; ++i
)
674 ttl
+= type
->ks_use
[i
].memuse
;
675 type
->ks_loosememuse
= ttl
; /* not MP synchronized */
676 if ((ssize_t
)ttl
< 0) /* deal with occassional race */
678 if (ttl
>= type
->ks_limit
) {
679 if (flags
& M_NULLOK
) {
680 logmemory(malloc_end
, NULL
, type
, size
, flags
);
683 panic("%s: malloc limit exceeded", type
->ks_shortdesc
);
688 * Handle the degenerate size == 0 case. Yes, this does happen.
689 * Return a special pointer. This is to maintain compatibility with
690 * the original malloc implementation. Certain devices, such as the
691 * adaptec driver, not only allocate 0 bytes, they check for NULL and
692 * also realloc() later on. Joy.
695 logmemory(malloc_end
, ZERO_LENGTH_PTR
, type
, size
, flags
);
696 return(ZERO_LENGTH_PTR
);
700 * Handle hysteresis from prior frees here in malloc(). We cannot
701 * safely manipulate the kernel_map in free() due to free() possibly
702 * being called via an IPI message or from sensitive interrupt code.
704 * NOTE: ku_pagecnt must be cleared before we free the slab or we
705 * might race another cpu allocating the kva and setting
708 while (slgd
->NFreeZones
> ZoneRelsThresh
&& (flags
& M_RNOWAIT
) == 0) {
710 if (slgd
->NFreeZones
> ZoneRelsThresh
) { /* crit sect race */
713 z
= TAILQ_LAST(&slgd
->FreeZones
, SLZoneList
);
715 TAILQ_REMOVE(&slgd
->FreeZones
, z
, z_Entry
);
719 kmem_slab_free(z
, ZoneSize
); /* may block */
720 atomic_add_int(&ZoneGenAlloc
, -ZoneSize
/ 1024);
726 * XXX handle oversized frees that were queued from kfree().
728 while (TAILQ_FIRST(&slgd
->FreeOvZones
) && (flags
& M_RNOWAIT
) == 0) {
730 if ((z
= TAILQ_LAST(&slgd
->FreeOvZones
, SLZoneList
)) != NULL
) {
733 KKASSERT(z
->z_Magic
== ZALLOC_OVSZ_MAGIC
);
734 TAILQ_REMOVE(&slgd
->FreeOvZones
, z
, z_Entry
);
735 tsize
= z
->z_ChunkSize
;
736 kmem_slab_free(z
, tsize
); /* may block */
737 atomic_add_int(&ZoneBigAlloc
, -(int)tsize
/ 1024);
743 * Handle large allocations directly. There should not be very many of
744 * these so performance is not a big issue.
746 * The backend allocator is pretty nasty on a SMP system. Use the
747 * slab allocator for one and two page-sized chunks even though we lose
748 * some efficiency. XXX maybe fix mmio and the elf loader instead.
750 if (size
>= ZoneLimit
|| ((size
& PAGE_MASK
) == 0 && size
> PAGE_SIZE
*2)) {
753 size
= round_page(size
);
754 chunk
= kmem_slab_alloc(size
, PAGE_SIZE
, flags
);
756 logmemory(malloc_end
, NULL
, type
, size
, flags
);
759 atomic_add_int(&ZoneBigAlloc
, (int)size
/ 1024);
760 flags
&= ~M_ZERO
; /* result already zero'd if M_ZERO was set */
761 flags
|= M_PASSIVE_ZERO
;
763 *kup
= size
/ PAGE_SIZE
;
769 * Attempt to allocate out of an existing zone. First try the free list,
770 * then allocate out of unallocated space. If we find a good zone move
771 * it to the head of the list so later allocations find it quickly
772 * (we might have thousands of zones in the list).
774 * Note: zoneindex() will panic of size is too large.
776 zi
= zoneindex(&size
, &align
);
777 KKASSERT(zi
< NZONES
);
780 if ((z
= TAILQ_LAST(&slgd
->ZoneAry
[zi
], SLZoneList
)) != NULL
) {
782 * Locate a chunk - we have to have at least one. If this is the
783 * last chunk go ahead and do the work to retrieve chunks freed
784 * from remote cpus, and if the zone is still empty move it off
787 if (--z
->z_NFree
<= 0) {
788 KKASSERT(z
->z_NFree
== 0);
791 * WARNING! This code competes with other cpus. It is ok
792 * for us to not drain RChunks here but we might as well, and
793 * it is ok if more accumulate after we're done.
795 * Set RSignal before pulling rchunks off, indicating that we
796 * will be moving ourselves off of the ZoneAry. Remote ends will
797 * read RSignal before putting rchunks on thus interlocking
798 * their IPI signaling.
800 if (z
->z_RChunks
== NULL
)
801 atomic_swap_int(&z
->z_RSignal
, 1);
803 clean_zone_rchunks(z
);
806 * Remove from the zone list if no free chunks remain.
809 if (z
->z_NFree
== 0) {
810 TAILQ_REMOVE(&slgd
->ZoneAry
[zi
], z
, z_Entry
);
817 * Fast path, we have chunks available in z_LChunks.
819 chunk
= z
->z_LChunks
;
821 chunk_mark_allocated(z
, chunk
);
822 z
->z_LChunks
= chunk
->c_Next
;
823 if (z
->z_LChunks
== NULL
)
824 z
->z_LChunksp
= &z
->z_LChunks
;
826 slab_record_source(z
, file
, line
);
832 * No chunks are available in LChunks, the free chunk MUST be
833 * in the never-before-used memory area, controlled by UIndex.
835 * The consequences are very serious if our zone got corrupted so
836 * we use an explicit panic rather than a KASSERT.
838 if (z
->z_UIndex
+ 1 != z
->z_NMax
)
843 if (z
->z_UIndex
== z
->z_UEndIndex
)
844 panic("slaballoc: corrupted zone");
846 chunk
= (SLChunk
*)(z
->z_BasePtr
+ z
->z_UIndex
* size
);
847 if ((z
->z_Flags
& SLZF_UNOTZEROD
) == 0) {
849 flags
|= M_PASSIVE_ZERO
;
851 chunk_mark_allocated(z
, chunk
);
853 slab_record_source(z
, file
, line
);
859 * If all zones are exhausted we need to allocate a new zone for this
860 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
861 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
862 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
863 * we do not pre-zero it because we do not want to mess up the L1 cache.
865 * At least one subsystem, the tty code (see CROUND) expects power-of-2
866 * allocations to be power-of-2 aligned. We maintain compatibility by
867 * adjusting the base offset below.
873 if ((z
= TAILQ_FIRST(&slgd
->FreeZones
)) != NULL
) {
874 TAILQ_REMOVE(&slgd
->FreeZones
, z
, z_Entry
);
876 bzero(z
, sizeof(SLZone
));
877 z
->z_Flags
|= SLZF_UNOTZEROD
;
879 z
= kmem_slab_alloc(ZoneSize
, ZoneSize
, flags
|M_ZERO
);
882 atomic_add_int(&ZoneGenAlloc
, ZoneSize
/ 1024);
886 * How big is the base structure?
888 #if defined(INVARIANTS)
890 * Make room for z_Bitmap. An exact calculation is somewhat more
891 * complicated so don't make an exact calculation.
893 off
= offsetof(SLZone
, z_Bitmap
[(ZoneSize
/ size
+ 31) / 32]);
894 bzero(z
->z_Bitmap
, (ZoneSize
/ size
+ 31) / 8);
896 off
= sizeof(SLZone
);
900 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
901 * Otherwise properly align the data according to the chunk size.
905 off
= roundup2(off
, align
);
907 z
->z_Magic
= ZALLOC_SLAB_MAGIC
;
909 z
->z_NMax
= (ZoneSize
- off
) / size
;
910 z
->z_NFree
= z
->z_NMax
- 1;
911 z
->z_BasePtr
= (char *)z
+ off
;
912 z
->z_UIndex
= z
->z_UEndIndex
= slgd
->JunkIndex
% z
->z_NMax
;
913 z
->z_ChunkSize
= size
;
915 z
->z_Cpu
= gd
->gd_cpuid
;
916 z
->z_LChunksp
= &z
->z_LChunks
;
918 bcopy(z
->z_Sources
, z
->z_AltSources
, sizeof(z
->z_Sources
));
919 bzero(z
->z_Sources
, sizeof(z
->z_Sources
));
921 chunk
= (SLChunk
*)(z
->z_BasePtr
+ z
->z_UIndex
* size
);
922 TAILQ_INSERT_HEAD(&slgd
->ZoneAry
[zi
], z
, z_Entry
);
923 if ((z
->z_Flags
& SLZF_UNOTZEROD
) == 0) {
924 flags
&= ~M_ZERO
; /* already zero'd */
925 flags
|= M_PASSIVE_ZERO
;
928 *kup
= -(z
->z_Cpu
+ 1); /* -1 to -(N+1) */
929 chunk_mark_allocated(z
, chunk
);
931 slab_record_source(z
, file
, line
);
935 * Slide the base index for initial allocations out of the next
936 * zone we create so we do not over-weight the lower part of the
939 slgd
->JunkIndex
= (slgd
->JunkIndex
+ ZALLOC_SLAB_SLIDE
)
940 & (ZALLOC_MAX_ZONE_SIZE
- 1);
944 ++type
->ks_use
[gd
->gd_cpuid
].inuse
;
945 type
->ks_use
[gd
->gd_cpuid
].memuse
+= size
;
946 type
->ks_loosememuse
+= size
; /* not MP synchronized */
952 else if ((flags
& (M_ZERO
|M_PASSIVE_ZERO
)) == 0) {
953 if (use_malloc_pattern
) {
954 for (i
= 0; i
< size
; i
+= sizeof(int)) {
955 *(int *)((char *)chunk
+ i
) = -1;
958 chunk
->c_Next
= (void *)-1; /* avoid accidental double-free check */
961 logmemory(malloc_end
, chunk
, type
, size
, flags
);
965 logmemory(malloc_end
, NULL
, type
, size
, flags
);
970 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
972 * Generally speaking this routine is not called very often and we do
973 * not attempt to optimize it beyond reusing the same pointer if the
974 * new size fits within the chunking of the old pointer's zone.
978 krealloc_debug(void *ptr
, unsigned long size
,
979 struct malloc_type
*type
, int flags
,
980 const char *file
, int line
)
983 krealloc(void *ptr
, unsigned long size
, struct malloc_type
*type
, int flags
)
992 KKASSERT((flags
& M_ZERO
) == 0); /* not supported */
994 if (ptr
== NULL
|| ptr
== ZERO_LENGTH_PTR
)
995 return(kmalloc_debug(size
, type
, flags
, file
, line
));
1002 * Handle oversized allocations. XXX we really should require that a
1003 * size be passed to free() instead of this nonsense.
1007 osize
= *kup
<< PAGE_SHIFT
;
1008 if (osize
== round_page(size
))
1010 if ((nptr
= kmalloc_debug(size
, type
, flags
, file
, line
)) == NULL
)
1012 bcopy(ptr
, nptr
, min(size
, osize
));
1018 * Get the original allocation's zone. If the new request winds up
1019 * using the same chunk size we do not have to do anything.
1021 z
= (SLZone
*)((uintptr_t)ptr
& ZoneMask
);
1024 KKASSERT(z
->z_Magic
== ZALLOC_SLAB_MAGIC
);
1027 * Allocate memory for the new request size. Note that zoneindex has
1028 * already adjusted the request size to the appropriate chunk size, which
1029 * should optimize our bcopy(). Then copy and return the new pointer.
1031 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
1032 * necessary align the result.
1034 * We can only zoneindex (to align size to the chunk size) if the new
1035 * size is not too large.
1037 if (size
< ZoneLimit
) {
1038 zoneindex(&size
, &align
);
1039 if (z
->z_ChunkSize
== size
)
1042 if ((nptr
= kmalloc_debug(size
, type
, flags
, file
, line
)) == NULL
)
1044 bcopy(ptr
, nptr
, min(size
, z
->z_ChunkSize
));
1050 * Return the kmalloc limit for this type, in bytes.
1053 kmalloc_limit(struct malloc_type
*type
)
1055 if (type
->ks_limit
== 0) {
1057 if (type
->ks_limit
== 0)
1061 return(type
->ks_limit
);
1065 * Allocate a copy of the specified string.
1067 * (MP SAFE) (MAY BLOCK)
1071 kstrdup_debug(const char *str
, struct malloc_type
*type
,
1072 const char *file
, int line
)
1075 kstrdup(const char *str
, struct malloc_type
*type
)
1078 int zlen
; /* length inclusive of terminating NUL */
1083 zlen
= strlen(str
) + 1;
1084 nstr
= kmalloc_debug(zlen
, type
, M_WAITOK
, file
, line
);
1085 bcopy(str
, nstr
, zlen
);
1091 kstrndup_debug(const char *str
, size_t maxlen
, struct malloc_type
*type
,
1092 const char *file
, int line
)
1095 kstrndup(const char *str
, size_t maxlen
, struct malloc_type
*type
)
1098 int zlen
; /* length inclusive of terminating NUL */
1103 zlen
= strnlen(str
, maxlen
) + 1;
1104 nstr
= kmalloc_debug(zlen
, type
, M_WAITOK
, file
, line
);
1105 bcopy(str
, nstr
, zlen
);
1106 nstr
[zlen
- 1] = '\0';
1111 * Notify our cpu that a remote cpu has freed some chunks in a zone that
1112 * we own. RCount will be bumped so the memory should be good, but validate
1113 * that it really is.
1117 kfree_remote(void *ptr
)
1124 slgd
= &mycpu
->gd_slab
;
1127 KKASSERT(*kup
== -((int)mycpuid
+ 1));
1128 KKASSERT(z
->z_RCount
> 0);
1129 atomic_subtract_int(&z
->z_RCount
, 1);
1131 logmemory(free_rem_beg
, z
, NULL
, 0L, 0);
1132 KKASSERT(z
->z_Magic
== ZALLOC_SLAB_MAGIC
);
1133 KKASSERT(z
->z_Cpu
== mycpu
->gd_cpuid
);
1137 * Indicate that we will no longer be off of the ZoneAry by
1144 * Atomically extract the bchunks list and then process it back
1145 * into the lchunks list. We want to append our bchunks to the
1146 * lchunks list and not prepend since we likely do not have
1147 * cache mastership of the related data (not that it helps since
1148 * we are using c_Next).
1150 clean_zone_rchunks(z
);
1151 if (z
->z_NFree
&& nfree
== 0) {
1152 TAILQ_INSERT_HEAD(&slgd
->ZoneAry
[z
->z_ZoneIndex
], z
, z_Entry
);
1156 * If the zone becomes totally free and is not the only zone listed for a
1157 * chunk size we move it to the FreeZones list. We always leave at least
1158 * one zone per chunk size listed, even if it is freeable.
1160 * Since this code can be called from an IPI callback, do *NOT* try to
1161 * mess with kernel_map here. Hysteresis will be performed at malloc()
1164 * Do not move the zone if there is an IPI in_flight (z_RCount != 0),
1165 * otherwise MP races can result in our free_remote code accessing a
1166 * destroyed zone. The remote end interlocks z_RCount with z_RChunks
1167 * so one has to test both z_NFree and z_RCount.
1169 if (z
->z_NFree
== z
->z_NMax
&& z
->z_RCount
== 0 &&
1170 (TAILQ_FIRST(&slgd
->ZoneAry
[z
->z_ZoneIndex
]) != z
||
1171 TAILQ_NEXT(z
, z_Entry
))
1175 TAILQ_REMOVE(&slgd
->ZoneAry
[z
->z_ZoneIndex
], z
, z_Entry
);
1177 TAILQ_INSERT_HEAD(&slgd
->FreeZones
, z
, z_Entry
);
1182 logmemory(free_rem_end
, z
, NULL
, 0L, 0);
1186 * free (SLAB ALLOCATOR)
1188 * Free a memory block previously allocated by malloc. Note that we do not
1189 * attempt to update ks_loosememuse as MP races could prevent us from
1190 * checking memory limits in malloc.
1195 kfree(void *ptr
, struct malloc_type
*type
)
1200 struct globaldata
*gd
;
1206 logmemory_quick(free_beg
);
1208 slgd
= &gd
->gd_slab
;
1211 panic("trying to free NULL pointer");
1214 * Handle special 0-byte allocations
1216 if (ptr
== ZERO_LENGTH_PTR
) {
1217 logmemory(free_zero
, ptr
, type
, -1UL, 0);
1218 logmemory_quick(free_end
);
1223 * Panic on bad malloc type
1225 if (type
->ks_magic
!= M_MAGIC
)
1226 panic("free: malloc type lacks magic");
1229 * Handle oversized allocations. XXX we really should require that a
1230 * size be passed to free() instead of this nonsense.
1232 * This code is never called via an ipi.
1236 size
= *kup
<< PAGE_SHIFT
;
1239 KKASSERT(sizeof(weirdary
) <= size
);
1240 bcopy(weirdary
, ptr
, sizeof(weirdary
));
1243 * NOTE: For oversized allocations we do not record the
1244 * originating cpu. It gets freed on the cpu calling
1245 * kfree(). The statistics are in aggregate.
1247 * note: XXX we have still inherited the interrupts-can't-block
1248 * assumption. An interrupt thread does not bump
1249 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1250 * primarily until we can fix softupdate's assumptions about free().
1253 --type
->ks_use
[gd
->gd_cpuid
].inuse
;
1254 type
->ks_use
[gd
->gd_cpuid
].memuse
-= size
;
1255 if (mycpu
->gd_intr_nesting_level
||
1256 (gd
->gd_curthread
->td_flags
& TDF_INTTHREAD
))
1258 logmemory(free_ovsz_delayed
, ptr
, type
, size
, 0);
1260 z
->z_Magic
= ZALLOC_OVSZ_MAGIC
;
1261 z
->z_ChunkSize
= size
;
1263 TAILQ_INSERT_HEAD(&slgd
->FreeOvZones
, z
, z_Entry
);
1267 logmemory(free_ovsz
, ptr
, type
, size
, 0);
1268 kmem_slab_free(ptr
, size
); /* may block */
1269 atomic_add_int(&ZoneBigAlloc
, -(int)size
/ 1024);
1271 logmemory_quick(free_end
);
1276 * Zone case. Figure out the zone based on the fact that it is
1279 z
= (SLZone
*)((uintptr_t)ptr
& ZoneMask
);
1282 KKASSERT(z
->z_Magic
== ZALLOC_SLAB_MAGIC
);
1285 * If we do not own the zone then use atomic ops to free to the
1286 * remote cpu linked list and notify the target zone using a
1289 * The target zone cannot be deallocated while we own a chunk of it,
1290 * so the zone header's storage is stable until the very moment
1291 * we adjust z_RChunks. After that we cannot safely dereference (z).
1293 * (no critical section needed)
1295 if (z
->z_CpuGd
!= gd
) {
1297 * Making these adjustments now allow us to avoid passing (type)
1298 * to the remote cpu. Note that inuse/memuse is being
1299 * adjusted on OUR cpu, not the zone cpu, but it should all still
1300 * sum up properly and cancel out.
1303 --type
->ks_use
[gd
->gd_cpuid
].inuse
;
1304 type
->ks_use
[gd
->gd_cpuid
].memuse
-= z
->z_ChunkSize
;
1308 * WARNING! This code competes with other cpus. Once we
1309 * successfully link the chunk to RChunks the remote
1310 * cpu can rip z's storage out from under us.
1312 * Bumping RCount prevents z's storage from getting
1315 rsignal
= z
->z_RSignal
;
1318 atomic_add_int(&z
->z_RCount
, 1);
1322 bchunk
= z
->z_RChunks
;
1324 chunk
->c_Next
= bchunk
;
1327 if (atomic_cmpset_ptr(&z
->z_RChunks
, bchunk
, chunk
))
1332 * We have to signal the remote cpu if our actions will cause
1333 * the remote zone to be placed back on ZoneAry so it can
1334 * move the zone back on.
1336 * We only need to deal with NULL->non-NULL RChunk transitions
1337 * and only if z_RSignal is set. We interlock by reading rsignal
1338 * before adding our chunk to RChunks. This should result in
1339 * virtually no IPI traffic.
1341 * We can use a passive IPI to reduce overhead even further.
1343 if (bchunk
== NULL
&& rsignal
) {
1344 logmemory(free_request
, ptr
, type
,
1345 (unsigned long)z
->z_ChunkSize
, 0);
1346 lwkt_send_ipiq_passive(z
->z_CpuGd
, kfree_remote
, z
);
1347 /* z can get ripped out from under us from this point on */
1348 } else if (rsignal
) {
1349 atomic_subtract_int(&z
->z_RCount
, 1);
1350 /* z can get ripped out from under us from this point on */
1352 logmemory_quick(free_end
);
1359 logmemory(free_chunk
, ptr
, type
, (unsigned long)z
->z_ChunkSize
, 0);
1363 chunk_mark_free(z
, chunk
);
1366 * Put weird data into the memory to detect modifications after freeing,
1367 * illegal pointer use after freeing (we should fault on the odd address),
1368 * and so forth. XXX needs more work, see the old malloc code.
1371 if (z
->z_ChunkSize
< sizeof(weirdary
))
1372 bcopy(weirdary
, chunk
, z
->z_ChunkSize
);
1374 bcopy(weirdary
, chunk
, sizeof(weirdary
));
1378 * Add this free non-zero'd chunk to a linked list for reuse. Add
1379 * to the front of the linked list so it is more likely to be
1380 * reallocated, since it is already in our L1 cache.
1383 if ((vm_offset_t
)chunk
< KvaStart
|| (vm_offset_t
)chunk
>= KvaEnd
)
1384 panic("BADFREE %p", chunk
);
1386 chunk
->c_Next
= z
->z_LChunks
;
1387 z
->z_LChunks
= chunk
;
1388 if (chunk
->c_Next
== NULL
)
1389 z
->z_LChunksp
= &chunk
->c_Next
;
1392 if (chunk
->c_Next
&& (vm_offset_t
)chunk
->c_Next
< KvaStart
)
1397 * Bump the number of free chunks. If it becomes non-zero the zone
1398 * must be added back onto the appropriate list. A fully allocated
1399 * zone that sees its first free is considered 'mature' and is placed
1400 * at the head, giving the system time to potentially free the remaining
1401 * entries even while other allocations are going on and making the zone
1404 if (z
->z_NFree
++ == 0) {
1406 TAILQ_INSERT_TAIL(&slgd
->ZoneAry
[z
->z_ZoneIndex
], z
, z_Entry
);
1408 TAILQ_INSERT_HEAD(&slgd
->ZoneAry
[z
->z_ZoneIndex
], z
, z_Entry
);
1411 --type
->ks_use
[z
->z_Cpu
].inuse
;
1412 type
->ks_use
[z
->z_Cpu
].memuse
-= z
->z_ChunkSize
;
1414 check_zone_free(slgd
, z
);
1415 logmemory_quick(free_end
);
1420 * Cleanup slabs which are hanging around due to RChunks or which are wholely
1421 * free and can be moved to the free list if not moved by other means.
1423 * Called once every 10 seconds on all cpus.
1428 SLGlobalData
*slgd
= &mycpu
->gd_slab
;
1433 for (i
= 0; i
< NZONES
; ++i
) {
1434 if ((z
= TAILQ_FIRST(&slgd
->ZoneAry
[i
])) == NULL
)
1442 * Shift all RChunks to the end of the LChunks list. This is
1443 * an O(1) operation.
1445 * Then free the zone if possible.
1447 clean_zone_rchunks(z
);
1448 z
= check_zone_free(slgd
, z
);
1454 #if defined(INVARIANTS)
1457 * Helper routines for sanity checks
1461 chunk_mark_allocated(SLZone
*z
, void *chunk
)
1463 int bitdex
= ((char *)chunk
- (char *)z
->z_BasePtr
) / z
->z_ChunkSize
;
1466 KKASSERT((((intptr_t)chunk
^ (intptr_t)z
) & ZoneMask
) == 0);
1467 KASSERT(bitdex
>= 0 && bitdex
< z
->z_NMax
,
1468 ("memory chunk %p bit index %d is illegal", chunk
, bitdex
));
1469 bitptr
= &z
->z_Bitmap
[bitdex
>> 5];
1471 KASSERT((*bitptr
& (1 << bitdex
)) == 0,
1472 ("memory chunk %p is already allocated!", chunk
));
1473 *bitptr
|= 1 << bitdex
;
1478 chunk_mark_free(SLZone
*z
, void *chunk
)
1480 int bitdex
= ((char *)chunk
- (char *)z
->z_BasePtr
) / z
->z_ChunkSize
;
1483 KKASSERT((((intptr_t)chunk
^ (intptr_t)z
) & ZoneMask
) == 0);
1484 KASSERT(bitdex
>= 0 && bitdex
< z
->z_NMax
,
1485 ("memory chunk %p bit index %d is illegal!", chunk
, bitdex
));
1486 bitptr
= &z
->z_Bitmap
[bitdex
>> 5];
1488 KASSERT((*bitptr
& (1 << bitdex
)) != 0,
1489 ("memory chunk %p is already free!", chunk
));
1490 *bitptr
&= ~(1 << bitdex
);
1498 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1499 * specified alignment. M_* flags are expected in the flags field.
1501 * Alignment must be a multiple of PAGE_SIZE.
1503 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1504 * but when we move zalloc() over to use this function as its backend
1505 * we will have to switch to kreserve/krelease and call reserve(0)
1506 * after the new space is made available.
1508 * Interrupt code which has preempted other code is not allowed to
1509 * use PQ_CACHE pages. However, if an interrupt thread is run
1510 * non-preemptively or blocks and then runs non-preemptively, then
1511 * it is free to use PQ_CACHE pages. <--- may not apply any longer XXX
1514 kmem_slab_alloc(vm_size_t size
, vm_offset_t align
, int flags
)
1518 int count
, vmflags
, base_vmflags
;
1519 vm_page_t mbase
= NULL
;
1523 size
= round_page(size
);
1524 addr
= vm_map_min(&kernel_map
);
1526 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1528 vm_map_lock(&kernel_map
);
1529 if (vm_map_findspace(&kernel_map
, addr
, size
, align
, 0, &addr
)) {
1530 vm_map_unlock(&kernel_map
);
1531 if ((flags
& M_NULLOK
) == 0)
1532 panic("kmem_slab_alloc(): kernel_map ran out of space!");
1533 vm_map_entry_release(count
);
1539 * kernel_object maps 1:1 to kernel_map.
1541 vm_object_hold(&kernel_object
);
1542 vm_object_reference_locked(&kernel_object
);
1543 vm_map_insert(&kernel_map
, &count
,
1544 &kernel_object
, NULL
,
1545 addr
, addr
, addr
+ size
,
1548 VM_PROT_ALL
, VM_PROT_ALL
, 0);
1549 vm_object_drop(&kernel_object
);
1550 vm_map_set_wired_quick(&kernel_map
, addr
, size
, &count
);
1551 vm_map_unlock(&kernel_map
);
1557 base_vmflags
|= VM_ALLOC_ZERO
;
1558 if (flags
& M_USE_RESERVE
)
1559 base_vmflags
|= VM_ALLOC_SYSTEM
;
1560 if (flags
& M_USE_INTERRUPT_RESERVE
)
1561 base_vmflags
|= VM_ALLOC_INTERRUPT
;
1562 if ((flags
& (M_RNOWAIT
|M_WAITOK
)) == 0) {
1563 panic("kmem_slab_alloc: bad flags %08x (%p)",
1564 flags
, ((int **)&size
)[-1]);
1568 * Allocate the pages. Do not map them yet. VM_ALLOC_NORMAL can only
1569 * be set if we are not preempting.
1571 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1572 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1573 * implied in this case), though I'm not sure if we really need to
1576 vmflags
= base_vmflags
;
1577 if (flags
& M_WAITOK
) {
1578 if (td
->td_preempted
)
1579 vmflags
|= VM_ALLOC_SYSTEM
;
1581 vmflags
|= VM_ALLOC_NORMAL
;
1584 vm_object_hold(&kernel_object
);
1585 for (i
= 0; i
< size
; i
+= PAGE_SIZE
) {
1586 m
= vm_page_alloc(&kernel_object
, OFF_TO_IDX(addr
+ i
), vmflags
);
1591 * If the allocation failed we either return NULL or we retry.
1593 * If M_WAITOK is specified we wait for more memory and retry.
1594 * If M_WAITOK is specified from a preemption we yield instead of
1595 * wait. Livelock will not occur because the interrupt thread
1596 * will not be preempting anyone the second time around after the
1600 if (flags
& M_WAITOK
) {
1601 if (td
->td_preempted
) {
1606 i
-= PAGE_SIZE
; /* retry */
1614 * Check and deal with an allocation failure
1619 m
= vm_page_lookup(&kernel_object
, OFF_TO_IDX(addr
+ i
));
1620 /* page should already be busy */
1623 vm_map_lock(&kernel_map
);
1624 vm_map_delete(&kernel_map
, addr
, addr
+ size
, &count
);
1625 vm_map_unlock(&kernel_map
);
1626 vm_object_drop(&kernel_object
);
1628 vm_map_entry_release(count
);
1636 * NOTE: The VM pages are still busied. mbase points to the first one
1637 * but we have to iterate via vm_page_next()
1639 vm_object_drop(&kernel_object
);
1643 * Enter the pages into the pmap and deal with M_ZERO.
1650 * page should already be busy
1652 m
->valid
= VM_PAGE_BITS_ALL
;
1654 pmap_enter(&kernel_pmap
, addr
+ i
, m
, VM_PROT_ALL
| VM_PROT_NOSYNC
,
1657 pagezero((char *)addr
+ i
);
1658 KKASSERT(m
->flags
& (PG_WRITEABLE
| PG_MAPPED
));
1659 vm_page_flag_set(m
, PG_REFERENCED
);
1663 vm_object_hold(&kernel_object
);
1664 m
= vm_page_next(m
);
1665 vm_object_drop(&kernel_object
);
1668 vm_map_entry_release(count
);
1669 atomic_add_long(&SlabsAllocated
, 1);
1670 return((void *)addr
);
1677 kmem_slab_free(void *ptr
, vm_size_t size
)
1680 vm_map_remove(&kernel_map
, (vm_offset_t
)ptr
, (vm_offset_t
)ptr
+ size
);
1681 atomic_add_long(&SlabsFreed
, 1);
1686 kmalloc_cachealign(unsigned long size_alloc
, struct malloc_type
*type
,
1689 #if (__VM_CACHELINE_SIZE == 32)
1690 #define CAN_CACHEALIGN(sz) ((sz) >= 256)
1691 #elif (__VM_CACHELINE_SIZE == 64)
1692 #define CAN_CACHEALIGN(sz) ((sz) >= 512)
1693 #elif (__VM_CACHELINE_SIZE == 128)
1694 #define CAN_CACHEALIGN(sz) ((sz) >= 1024)
1696 #error "unsupported cacheline size"
1701 if (size_alloc
< __VM_CACHELINE_SIZE
)
1702 size_alloc
= __VM_CACHELINE_SIZE
;
1703 else if (!CAN_CACHEALIGN(size_alloc
))
1704 flags
|= M_POWEROF2
;
1706 ret
= kmalloc(size_alloc
, type
, flags
);
1707 KASSERT(((uintptr_t)ret
& (__VM_CACHELINE_SIZE
- 1)) == 0,
1708 ("%p(%lu) not cacheline %d aligned",
1709 ret
, size_alloc
, __VM_CACHELINE_SIZE
));
1712 #undef CAN_CACHEALIGN