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.
80 * (n * PAGE_SIZE, n > 2) allocations go directly to kmem.
82 * Alignment properties:
83 * - All power-of-2 sized allocations are power-of-2 aligned.
84 * - Allocations with M_POWEROF2 are power-of-2 aligned on the nearest
85 * power-of-2 round up of 'size'.
86 * - Non-power-of-2 sized allocations are zone chunk size aligned (see the
87 * above table 'Chunking' column).
89 * API REQUIREMENTS AND SIDE EFFECTS
91 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we
92 * have remained compatible with the following API requirements:
94 * + malloc(0) is allowed and returns non-NULL (ahc driver)
95 * + ability to allocate arbitrarily large chunks of memory
100 #include <sys/param.h>
101 #include <sys/systm.h>
102 #include <sys/kernel.h>
103 #include <sys/slaballoc.h>
104 #include <sys/mbuf.h>
105 #include <sys/vmmeter.h>
106 #include <sys/lock.h>
107 #include <sys/thread.h>
108 #include <sys/globaldata.h>
109 #include <sys/sysctl.h>
113 #include <vm/vm_param.h>
114 #include <vm/vm_kern.h>
115 #include <vm/vm_extern.h>
116 #include <vm/vm_object.h>
118 #include <vm/vm_map.h>
119 #include <vm/vm_page.h>
120 #include <vm/vm_pageout.h>
122 #include <machine/cpu.h>
124 #include <sys/thread2.h>
125 #include <vm/vm_page2.h>
127 #define btokup(z) (&pmap_kvtom((vm_offset_t)(z))->ku_pagecnt)
129 #define MEMORY_STRING "ptr=%p type=%p size=%lu flags=%04x"
130 #define MEMORY_ARGS void *ptr, void *type, unsigned long size, int flags
132 #if !defined(KTR_MEMORY)
133 #define KTR_MEMORY KTR_ALL
135 KTR_INFO_MASTER(memory
);
136 KTR_INFO(KTR_MEMORY
, memory
, malloc_beg
, 0, "malloc begin");
137 KTR_INFO(KTR_MEMORY
, memory
, malloc_end
, 1, MEMORY_STRING
, MEMORY_ARGS
);
138 KTR_INFO(KTR_MEMORY
, memory
, free_zero
, 2, MEMORY_STRING
, MEMORY_ARGS
);
139 KTR_INFO(KTR_MEMORY
, memory
, free_ovsz
, 3, MEMORY_STRING
, MEMORY_ARGS
);
140 KTR_INFO(KTR_MEMORY
, memory
, free_ovsz_delayed
, 4, MEMORY_STRING
, MEMORY_ARGS
);
141 KTR_INFO(KTR_MEMORY
, memory
, free_chunk
, 5, MEMORY_STRING
, MEMORY_ARGS
);
142 KTR_INFO(KTR_MEMORY
, memory
, free_request
, 6, MEMORY_STRING
, MEMORY_ARGS
);
143 KTR_INFO(KTR_MEMORY
, memory
, free_rem_beg
, 7, MEMORY_STRING
, MEMORY_ARGS
);
144 KTR_INFO(KTR_MEMORY
, memory
, free_rem_end
, 8, MEMORY_STRING
, MEMORY_ARGS
);
145 KTR_INFO(KTR_MEMORY
, memory
, free_beg
, 9, "free begin");
146 KTR_INFO(KTR_MEMORY
, memory
, free_end
, 10, "free end");
148 #define logmemory(name, ptr, type, size, flags) \
149 KTR_LOG(memory_ ## name, ptr, type, size, flags)
150 #define logmemory_quick(name) \
151 KTR_LOG(memory_ ## name)
154 * Fixed globals (not per-cpu)
157 static int ZoneLimit
;
158 static int ZonePageCount
;
159 static uintptr_t ZoneMask
;
160 static int ZoneBigAlloc
; /* in KB */
161 static int ZoneGenAlloc
; /* in KB */
162 struct malloc_type
*kmemstatistics
; /* exported to vmstat */
164 static int32_t weirdary
[16];
167 static void *kmem_slab_alloc(vm_size_t bytes
, vm_offset_t align
, int flags
);
168 static void kmem_slab_free(void *ptr
, vm_size_t bytes
);
170 #if defined(INVARIANTS)
171 static void chunk_mark_allocated(SLZone
*z
, void *chunk
);
172 static void chunk_mark_free(SLZone
*z
, void *chunk
);
174 #define chunk_mark_allocated(z, chunk)
175 #define chunk_mark_free(z, chunk)
179 * Misc constants. Note that allocations that are exact multiples of
180 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
182 #define ZONE_RELS_THRESH 32 /* threshold number of zones */
186 * The WEIRD_ADDR is used as known text to copy into free objects to
187 * try to create deterministic failure cases if the data is accessed after
190 #define WEIRD_ADDR 0xdeadc0de
192 #define ZERO_LENGTH_PTR ((void *)-8)
195 * Misc global malloc buckets
198 MALLOC_DEFINE(M_CACHE
, "cache", "Various Dynamically allocated caches");
199 MALLOC_DEFINE(M_DEVBUF
, "devbuf", "device driver memory");
200 MALLOC_DEFINE(M_TEMP
, "temp", "misc temporary data buffers");
201 MALLOC_DEFINE(M_DRM
, "m_drm", "DRM memory allocations");
203 MALLOC_DEFINE(M_IP6OPT
, "ip6opt", "IPv6 options");
204 MALLOC_DEFINE(M_IP6NDP
, "ip6ndp", "IPv6 Neighbor Discovery");
207 * Initialize the slab memory allocator. We have to choose a zone size based
208 * on available physical memory. We choose a zone side which is approximately
209 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
210 * 128K. The zone size is limited to the bounds set in slaballoc.h
211 * (typically 32K min, 128K max).
213 static void kmeminit(void *dummy
);
217 SYSINIT(kmem
, SI_BOOT1_ALLOCATOR
, SI_ORDER_FIRST
, kmeminit
, NULL
);
221 * If enabled any memory allocated without M_ZERO is initialized to -1.
223 static int use_malloc_pattern
;
224 SYSCTL_INT(_debug
, OID_AUTO
, use_malloc_pattern
, CTLFLAG_RW
,
225 &use_malloc_pattern
, 0,
226 "Initialize memory to -1 if M_ZERO not specified");
229 static int ZoneRelsThresh
= ZONE_RELS_THRESH
;
230 SYSCTL_INT(_kern
, OID_AUTO
, zone_big_alloc
, CTLFLAG_RD
, &ZoneBigAlloc
, 0, "");
231 SYSCTL_INT(_kern
, OID_AUTO
, zone_gen_alloc
, CTLFLAG_RD
, &ZoneGenAlloc
, 0, "");
232 SYSCTL_INT(_kern
, OID_AUTO
, zone_cache
, CTLFLAG_RW
, &ZoneRelsThresh
, 0, "");
233 static long SlabsAllocated
;
234 static long SlabsFreed
;
235 SYSCTL_LONG(_kern
, OID_AUTO
, slabs_allocated
, CTLFLAG_RD
,
236 &SlabsAllocated
, 0, "");
237 SYSCTL_LONG(_kern
, OID_AUTO
, slabs_freed
, CTLFLAG_RD
,
239 static int SlabFreeToTail
;
240 SYSCTL_INT(_kern
, OID_AUTO
, slab_freetotail
, CTLFLAG_RW
,
241 &SlabFreeToTail
, 0, "");
243 static struct spinlock kmemstat_spin
=
244 SPINLOCK_INITIALIZER(&kmemstat_spin
, "malinit");
247 * Returns the kernel memory size limit for the purposes of initializing
248 * various subsystem caches. The smaller of available memory and the KVM
249 * memory space is returned.
251 * The size in megabytes is returned.
258 limsize
= (size_t)vmstats
.v_page_count
* PAGE_SIZE
;
259 if (limsize
> KvaSize
)
261 return (limsize
/ (1024 * 1024));
265 kmeminit(void *dummy
)
273 limsize
= kmem_lim_size();
274 usesize
= (int)(limsize
* 1024); /* convert to KB */
277 * If the machine has a large KVM space and more than 8G of ram,
278 * double the zone release threshold to reduce SMP invalidations.
279 * If more than 16G of ram, do it again.
281 * The BIOS eats a little ram so add some slop. We want 8G worth of
282 * memory sticks to trigger the first adjustment.
284 if (ZoneRelsThresh
== ZONE_RELS_THRESH
) {
285 if (limsize
>= 7 * 1024)
287 if (limsize
>= 15 * 1024)
292 * Calculate the zone size. This typically calculates to
293 * ZALLOC_MAX_ZONE_SIZE
295 ZoneSize
= ZALLOC_MIN_ZONE_SIZE
;
296 while (ZoneSize
< ZALLOC_MAX_ZONE_SIZE
&& (ZoneSize
<< 1) < usesize
)
298 ZoneLimit
= ZoneSize
/ 4;
299 if (ZoneLimit
> ZALLOC_ZONE_LIMIT
)
300 ZoneLimit
= ZALLOC_ZONE_LIMIT
;
301 ZoneMask
= ~(uintptr_t)(ZoneSize
- 1);
302 ZonePageCount
= ZoneSize
/ PAGE_SIZE
;
305 for (i
= 0; i
< NELEM(weirdary
); ++i
)
306 weirdary
[i
] = WEIRD_ADDR
;
309 ZeroPage
= kmem_slab_alloc(PAGE_SIZE
, PAGE_SIZE
, M_WAITOK
|M_ZERO
);
312 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize
/ 1024);
316 * (low level) Initialize slab-related elements in the globaldata structure.
318 * Occurs after kmeminit().
321 slab_gdinit(globaldata_t gd
)
327 for (i
= 0; i
< NZONES
; ++i
)
328 TAILQ_INIT(&slgd
->ZoneAry
[i
]);
329 TAILQ_INIT(&slgd
->FreeZones
);
330 TAILQ_INIT(&slgd
->FreeOvZones
);
334 * Initialize a malloc type tracking structure.
337 malloc_init(void *data
)
339 struct malloc_type
*type
= data
;
342 if (type
->ks_magic
!= M_MAGIC
)
343 panic("malloc type lacks magic");
345 if (type
->ks_limit
!= 0)
348 if (vmstats
.v_page_count
== 0)
349 panic("malloc_init not allowed before vm init");
351 limsize
= kmem_lim_size() * (1024 * 1024);
352 type
->ks_limit
= limsize
/ 10;
354 spin_lock(&kmemstat_spin
);
355 type
->ks_next
= kmemstatistics
;
356 kmemstatistics
= type
;
357 spin_unlock(&kmemstat_spin
);
361 malloc_uninit(void *data
)
363 struct malloc_type
*type
= data
;
364 struct malloc_type
*t
;
370 if (type
->ks_magic
!= M_MAGIC
)
371 panic("malloc type lacks magic");
373 if (vmstats
.v_page_count
== 0)
374 panic("malloc_uninit not allowed before vm init");
376 if (type
->ks_limit
== 0)
377 panic("malloc_uninit on uninitialized type");
379 /* Make sure that all pending kfree()s are finished. */
380 lwkt_synchronize_ipiqs("muninit");
384 * memuse is only correct in aggregation. Due to memory being allocated
385 * on one cpu and freed on another individual array entries may be
386 * negative or positive (canceling each other out).
388 for (i
= ttl
= 0; i
< ncpus
; ++i
)
389 ttl
+= type
->ks_use
[i
].memuse
;
391 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
392 ttl
, type
->ks_shortdesc
, i
);
395 spin_lock(&kmemstat_spin
);
396 if (type
== kmemstatistics
) {
397 kmemstatistics
= type
->ks_next
;
399 for (t
= kmemstatistics
; t
->ks_next
!= NULL
; t
= t
->ks_next
) {
400 if (t
->ks_next
== type
) {
401 t
->ks_next
= type
->ks_next
;
406 type
->ks_next
= NULL
;
408 spin_unlock(&kmemstat_spin
);
412 * Increase the kmalloc pool limit for the specified pool. No changes
413 * are the made if the pool would shrink.
416 kmalloc_raise_limit(struct malloc_type
*type
, size_t bytes
)
418 if (type
->ks_limit
== 0)
422 if (type
->ks_limit
< bytes
)
423 type
->ks_limit
= bytes
;
427 kmalloc_set_unlimited(struct malloc_type
*type
)
429 type
->ks_limit
= kmem_lim_size() * (1024 * 1024);
433 * Dynamically create a malloc pool. This function is a NOP if *typep is
437 kmalloc_create(struct malloc_type
**typep
, const char *descr
)
439 struct malloc_type
*type
;
441 if (*typep
== NULL
) {
442 type
= kmalloc(sizeof(*type
), M_TEMP
, M_WAITOK
| M_ZERO
);
443 type
->ks_magic
= M_MAGIC
;
444 type
->ks_shortdesc
= descr
;
451 * Destroy a dynamically created malloc pool. This function is a NOP if
452 * the pool has already been destroyed.
455 kmalloc_destroy(struct malloc_type
**typep
)
457 if (*typep
!= NULL
) {
458 malloc_uninit(*typep
);
459 kfree(*typep
, M_TEMP
);
465 * Calculate the zone index for the allocation request size and set the
466 * allocation request size to that particular zone's chunk size.
469 zoneindex(unsigned long *bytes
, unsigned long *align
)
471 unsigned int n
= (unsigned int)*bytes
; /* unsigned for shift opt */
474 *bytes
= n
= (n
+ 7) & ~7;
476 return(n
/ 8 - 1); /* 8 byte chunks, 16 zones */
479 *bytes
= n
= (n
+ 15) & ~15;
485 *bytes
= n
= (n
+ 31) & ~31;
490 *bytes
= n
= (n
+ 63) & ~63;
495 *bytes
= n
= (n
+ 127) & ~127;
497 return(n
/ 128 + 31);
500 *bytes
= n
= (n
+ 255) & ~255;
502 return(n
/ 256 + 39);
504 *bytes
= n
= (n
+ 511) & ~511;
506 return(n
/ 512 + 47);
508 #if ZALLOC_ZONE_LIMIT > 8192
510 *bytes
= n
= (n
+ 1023) & ~1023;
512 return(n
/ 1024 + 55);
515 #if ZALLOC_ZONE_LIMIT > 16384
517 *bytes
= n
= (n
+ 2047) & ~2047;
519 return(n
/ 2048 + 63);
522 panic("Unexpected byte count %d", n
);
527 clean_zone_rchunks(SLZone
*z
)
531 while ((bchunk
= z
->z_RChunks
) != NULL
) {
533 if (atomic_cmpset_ptr(&z
->z_RChunks
, bchunk
, NULL
)) {
534 *z
->z_LChunksp
= bchunk
;
536 chunk_mark_free(z
, bchunk
);
537 z
->z_LChunksp
= &bchunk
->c_Next
;
538 bchunk
= bchunk
->c_Next
;
548 * If the zone becomes totally free and is not the only zone listed for a
549 * chunk size we move it to the FreeZones list. We always leave at least
550 * one zone per chunk size listed, even if it is freeable.
552 * Do not move the zone if there is an IPI in_flight (z_RCount != 0),
553 * otherwise MP races can result in our free_remote code accessing a
554 * destroyed zone. The remote end interlocks z_RCount with z_RChunks
555 * so one has to test both z_NFree and z_RCount.
557 * Since this code can be called from an IPI callback, do *NOT* try to mess
558 * with kernel_map here. Hysteresis will be performed at kmalloc() time.
560 static __inline SLZone
*
561 check_zone_free(SLGlobalData
*slgd
, SLZone
*z
)
565 znext
= TAILQ_NEXT(z
, z_Entry
);
566 if (z
->z_NFree
== z
->z_NMax
&& z
->z_RCount
== 0 &&
567 (TAILQ_FIRST(&slgd
->ZoneAry
[z
->z_ZoneIndex
]) != z
|| znext
)) {
570 TAILQ_REMOVE(&slgd
->ZoneAry
[z
->z_ZoneIndex
], z
, z_Entry
);
573 TAILQ_INSERT_HEAD(&slgd
->FreeZones
, z
, z_Entry
);
583 * Used to debug memory corruption issues. Record up to (typically 32)
584 * allocation sources for this zone (for a particular chunk size).
588 slab_record_source(SLZone
*z
, const char *file
, int line
)
591 int b
= line
& (SLAB_DEBUG_ENTRIES
- 1);
595 if (z
->z_Sources
[i
].file
== file
&& z
->z_Sources
[i
].line
== line
)
597 if (z
->z_Sources
[i
].file
== NULL
)
599 i
= (i
+ 1) & (SLAB_DEBUG_ENTRIES
- 1);
601 z
->z_Sources
[i
].file
= file
;
602 z
->z_Sources
[i
].line
= line
;
607 static __inline
unsigned long
608 powerof2_size(unsigned long size
)
612 if (size
== 0 || powerof2(size
))
620 * kmalloc() (SLAB ALLOCATOR)
622 * Allocate memory via the slab allocator. If the request is too large,
623 * or if it page-aligned beyond a certain size, we fall back to the
624 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
625 * &SlabMisc if you don't care.
627 * M_RNOWAIT - don't block.
628 * M_NULLOK - return NULL instead of blocking.
629 * M_ZERO - zero the returned memory.
630 * M_USE_RESERVE - allow greater drawdown of the free list
631 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
632 * M_POWEROF2 - roundup size to the nearest power of 2
637 /* don't let kmalloc macro mess up function declaration */
642 kmalloc_debug(unsigned long size
, struct malloc_type
*type
, int flags
,
643 const char *file
, int line
)
646 kmalloc(unsigned long size
, struct malloc_type
*type
, int flags
)
652 struct globaldata
*gd
;
659 logmemory_quick(malloc_beg
);
664 * XXX silly to have this in the critical path.
666 if (type
->ks_limit
== 0) {
671 ++type
->ks_use
[gd
->gd_cpuid
].calls
;
673 if (flags
& M_POWEROF2
)
674 size
= powerof2_size(size
);
677 * Handle the case where the limit is reached. Panic if we can't return
678 * NULL. The original malloc code looped, but this tended to
679 * simply deadlock the computer.
681 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
682 * to determine if a more complete limit check should be done. The
683 * actual memory use is tracked via ks_use[cpu].memuse.
685 while (type
->ks_loosememuse
>= type
->ks_limit
) {
689 for (i
= ttl
= 0; i
< ncpus
; ++i
)
690 ttl
+= type
->ks_use
[i
].memuse
;
691 type
->ks_loosememuse
= ttl
; /* not MP synchronized */
692 if ((ssize_t
)ttl
< 0) /* deal with occassional race */
694 if (ttl
>= type
->ks_limit
) {
695 if (flags
& M_NULLOK
) {
696 logmemory(malloc_end
, NULL
, type
, size
, flags
);
699 panic("%s: malloc limit exceeded", type
->ks_shortdesc
);
704 * Handle the degenerate size == 0 case. Yes, this does happen.
705 * Return a special pointer. This is to maintain compatibility with
706 * the original malloc implementation. Certain devices, such as the
707 * adaptec driver, not only allocate 0 bytes, they check for NULL and
708 * also realloc() later on. Joy.
711 logmemory(malloc_end
, ZERO_LENGTH_PTR
, type
, size
, flags
);
712 return(ZERO_LENGTH_PTR
);
716 * Handle hysteresis from prior frees here in malloc(). We cannot
717 * safely manipulate the kernel_map in free() due to free() possibly
718 * being called via an IPI message or from sensitive interrupt code.
720 * NOTE: ku_pagecnt must be cleared before we free the slab or we
721 * might race another cpu allocating the kva and setting
724 while (slgd
->NFreeZones
> ZoneRelsThresh
&& (flags
& M_RNOWAIT
) == 0) {
726 if (slgd
->NFreeZones
> ZoneRelsThresh
) { /* crit sect race */
729 z
= TAILQ_LAST(&slgd
->FreeZones
, SLZoneList
);
731 TAILQ_REMOVE(&slgd
->FreeZones
, z
, z_Entry
);
735 kmem_slab_free(z
, ZoneSize
); /* may block */
736 atomic_add_int(&ZoneGenAlloc
, -ZoneSize
/ 1024);
742 * XXX handle oversized frees that were queued from kfree().
744 while (TAILQ_FIRST(&slgd
->FreeOvZones
) && (flags
& M_RNOWAIT
) == 0) {
746 if ((z
= TAILQ_LAST(&slgd
->FreeOvZones
, SLZoneList
)) != NULL
) {
749 KKASSERT(z
->z_Magic
== ZALLOC_OVSZ_MAGIC
);
750 TAILQ_REMOVE(&slgd
->FreeOvZones
, z
, z_Entry
);
751 tsize
= z
->z_ChunkSize
;
752 kmem_slab_free(z
, tsize
); /* may block */
753 atomic_add_int(&ZoneBigAlloc
, -(int)tsize
/ 1024);
759 * Handle large allocations directly. There should not be very many of
760 * these so performance is not a big issue.
762 * The backend allocator is pretty nasty on a SMP system. Use the
763 * slab allocator for one and two page-sized chunks even though we lose
764 * some efficiency. XXX maybe fix mmio and the elf loader instead.
766 if (size
>= ZoneLimit
|| ((size
& PAGE_MASK
) == 0 && size
> PAGE_SIZE
*2)) {
769 size
= round_page(size
);
770 chunk
= kmem_slab_alloc(size
, PAGE_SIZE
, flags
);
772 logmemory(malloc_end
, NULL
, type
, size
, flags
);
775 atomic_add_int(&ZoneBigAlloc
, (int)size
/ 1024);
776 flags
&= ~M_ZERO
; /* result already zero'd if M_ZERO was set */
777 flags
|= M_PASSIVE_ZERO
;
779 *kup
= size
/ PAGE_SIZE
;
785 * Attempt to allocate out of an existing zone. First try the free list,
786 * then allocate out of unallocated space. If we find a good zone move
787 * it to the head of the list so later allocations find it quickly
788 * (we might have thousands of zones in the list).
790 * Note: zoneindex() will panic of size is too large.
792 zi
= zoneindex(&size
, &align
);
793 KKASSERT(zi
< NZONES
);
796 if ((z
= TAILQ_LAST(&slgd
->ZoneAry
[zi
], SLZoneList
)) != NULL
) {
798 * Locate a chunk - we have to have at least one. If this is the
799 * last chunk go ahead and do the work to retrieve chunks freed
800 * from remote cpus, and if the zone is still empty move it off
803 if (--z
->z_NFree
<= 0) {
804 KKASSERT(z
->z_NFree
== 0);
807 * WARNING! This code competes with other cpus. It is ok
808 * for us to not drain RChunks here but we might as well, and
809 * it is ok if more accumulate after we're done.
811 * Set RSignal before pulling rchunks off, indicating that we
812 * will be moving ourselves off of the ZoneAry. Remote ends will
813 * read RSignal before putting rchunks on thus interlocking
814 * their IPI signaling.
816 if (z
->z_RChunks
== NULL
)
817 atomic_swap_int(&z
->z_RSignal
, 1);
819 clean_zone_rchunks(z
);
822 * Remove from the zone list if no free chunks remain.
825 if (z
->z_NFree
== 0) {
826 TAILQ_REMOVE(&slgd
->ZoneAry
[zi
], z
, z_Entry
);
833 * Fast path, we have chunks available in z_LChunks.
835 chunk
= z
->z_LChunks
;
837 chunk_mark_allocated(z
, chunk
);
838 z
->z_LChunks
= chunk
->c_Next
;
839 if (z
->z_LChunks
== NULL
)
840 z
->z_LChunksp
= &z
->z_LChunks
;
842 slab_record_source(z
, file
, line
);
848 * No chunks are available in LChunks, the free chunk MUST be
849 * in the never-before-used memory area, controlled by UIndex.
851 * The consequences are very serious if our zone got corrupted so
852 * we use an explicit panic rather than a KASSERT.
854 if (z
->z_UIndex
+ 1 != z
->z_NMax
)
859 if (z
->z_UIndex
== z
->z_UEndIndex
)
860 panic("slaballoc: corrupted zone");
862 chunk
= (SLChunk
*)(z
->z_BasePtr
+ z
->z_UIndex
* size
);
863 if ((z
->z_Flags
& SLZF_UNOTZEROD
) == 0) {
865 flags
|= M_PASSIVE_ZERO
;
867 chunk_mark_allocated(z
, chunk
);
869 slab_record_source(z
, file
, line
);
875 * If all zones are exhausted we need to allocate a new zone for this
876 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
877 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
878 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
879 * we do not pre-zero it because we do not want to mess up the L1 cache.
881 * At least one subsystem, the tty code (see CROUND) expects power-of-2
882 * allocations to be power-of-2 aligned. We maintain compatibility by
883 * adjusting the base offset below.
889 if ((z
= TAILQ_FIRST(&slgd
->FreeZones
)) != NULL
) {
890 TAILQ_REMOVE(&slgd
->FreeZones
, z
, z_Entry
);
892 bzero(z
, sizeof(SLZone
));
893 z
->z_Flags
|= SLZF_UNOTZEROD
;
895 z
= kmem_slab_alloc(ZoneSize
, ZoneSize
, flags
|M_ZERO
);
898 atomic_add_int(&ZoneGenAlloc
, ZoneSize
/ 1024);
902 * How big is the base structure?
904 #if defined(INVARIANTS)
906 * Make room for z_Bitmap. An exact calculation is somewhat more
907 * complicated so don't make an exact calculation.
909 off
= offsetof(SLZone
, z_Bitmap
[(ZoneSize
/ size
+ 31) / 32]);
910 bzero(z
->z_Bitmap
, (ZoneSize
/ size
+ 31) / 8);
912 off
= sizeof(SLZone
);
916 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
917 * Otherwise properly align the data according to the chunk size.
921 off
= roundup2(off
, align
);
923 z
->z_Magic
= ZALLOC_SLAB_MAGIC
;
925 z
->z_NMax
= (ZoneSize
- off
) / size
;
926 z
->z_NFree
= z
->z_NMax
- 1;
927 z
->z_BasePtr
= (char *)z
+ off
;
928 z
->z_UIndex
= z
->z_UEndIndex
= slgd
->JunkIndex
% z
->z_NMax
;
929 z
->z_ChunkSize
= size
;
931 z
->z_Cpu
= gd
->gd_cpuid
;
932 z
->z_LChunksp
= &z
->z_LChunks
;
934 bcopy(z
->z_Sources
, z
->z_AltSources
, sizeof(z
->z_Sources
));
935 bzero(z
->z_Sources
, sizeof(z
->z_Sources
));
937 chunk
= (SLChunk
*)(z
->z_BasePtr
+ z
->z_UIndex
* size
);
938 TAILQ_INSERT_HEAD(&slgd
->ZoneAry
[zi
], z
, z_Entry
);
939 if ((z
->z_Flags
& SLZF_UNOTZEROD
) == 0) {
940 flags
&= ~M_ZERO
; /* already zero'd */
941 flags
|= M_PASSIVE_ZERO
;
944 *kup
= -(z
->z_Cpu
+ 1); /* -1 to -(N+1) */
945 chunk_mark_allocated(z
, chunk
);
947 slab_record_source(z
, file
, line
);
951 * Slide the base index for initial allocations out of the next
952 * zone we create so we do not over-weight the lower part of the
955 slgd
->JunkIndex
= (slgd
->JunkIndex
+ ZALLOC_SLAB_SLIDE
)
956 & (ZALLOC_MAX_ZONE_SIZE
- 1);
960 ++type
->ks_use
[gd
->gd_cpuid
].inuse
;
961 type
->ks_use
[gd
->gd_cpuid
].memuse
+= size
;
962 type
->ks_use
[gd
->gd_cpuid
].loosememuse
+= size
;
963 if (type
->ks_use
[gd
->gd_cpuid
].loosememuse
>= ZoneSize
) {
964 /* not MP synchronized */
965 type
->ks_loosememuse
+= type
->ks_use
[gd
->gd_cpuid
].loosememuse
;
966 type
->ks_use
[gd
->gd_cpuid
].loosememuse
= 0;
973 else if ((flags
& (M_ZERO
|M_PASSIVE_ZERO
)) == 0) {
974 if (use_malloc_pattern
) {
975 for (i
= 0; i
< size
; i
+= sizeof(int)) {
976 *(int *)((char *)chunk
+ i
) = -1;
979 chunk
->c_Next
= (void *)-1; /* avoid accidental double-free check */
982 logmemory(malloc_end
, chunk
, type
, size
, flags
);
986 logmemory(malloc_end
, NULL
, type
, size
, flags
);
991 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
993 * Generally speaking this routine is not called very often and we do
994 * not attempt to optimize it beyond reusing the same pointer if the
995 * new size fits within the chunking of the old pointer's zone.
999 krealloc_debug(void *ptr
, unsigned long size
,
1000 struct malloc_type
*type
, int flags
,
1001 const char *file
, int line
)
1004 krealloc(void *ptr
, unsigned long size
, struct malloc_type
*type
, int flags
)
1007 unsigned long osize
;
1008 unsigned long align
;
1013 KKASSERT((flags
& M_ZERO
) == 0); /* not supported */
1015 if (ptr
== NULL
|| ptr
== ZERO_LENGTH_PTR
)
1016 return(kmalloc_debug(size
, type
, flags
, file
, line
));
1023 * Handle oversized allocations. XXX we really should require that a
1024 * size be passed to free() instead of this nonsense.
1028 osize
= *kup
<< PAGE_SHIFT
;
1029 if (osize
== round_page(size
))
1031 if ((nptr
= kmalloc_debug(size
, type
, flags
, file
, line
)) == NULL
)
1033 bcopy(ptr
, nptr
, min(size
, osize
));
1039 * Get the original allocation's zone. If the new request winds up
1040 * using the same chunk size we do not have to do anything.
1042 z
= (SLZone
*)((uintptr_t)ptr
& ZoneMask
);
1045 KKASSERT(z
->z_Magic
== ZALLOC_SLAB_MAGIC
);
1048 * Allocate memory for the new request size. Note that zoneindex has
1049 * already adjusted the request size to the appropriate chunk size, which
1050 * should optimize our bcopy(). Then copy and return the new pointer.
1052 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
1053 * necessary align the result.
1055 * We can only zoneindex (to align size to the chunk size) if the new
1056 * size is not too large.
1058 if (size
< ZoneLimit
) {
1059 zoneindex(&size
, &align
);
1060 if (z
->z_ChunkSize
== size
)
1063 if ((nptr
= kmalloc_debug(size
, type
, flags
, file
, line
)) == NULL
)
1065 bcopy(ptr
, nptr
, min(size
, z
->z_ChunkSize
));
1071 * Return the kmalloc limit for this type, in bytes.
1074 kmalloc_limit(struct malloc_type
*type
)
1076 if (type
->ks_limit
== 0) {
1078 if (type
->ks_limit
== 0)
1082 return(type
->ks_limit
);
1086 * Allocate a copy of the specified string.
1088 * (MP SAFE) (MAY BLOCK)
1092 kstrdup_debug(const char *str
, struct malloc_type
*type
,
1093 const char *file
, int line
)
1096 kstrdup(const char *str
, struct malloc_type
*type
)
1099 int zlen
; /* length inclusive of terminating NUL */
1104 zlen
= strlen(str
) + 1;
1105 nstr
= kmalloc_debug(zlen
, type
, M_WAITOK
, file
, line
);
1106 bcopy(str
, nstr
, zlen
);
1112 kstrndup_debug(const char *str
, size_t maxlen
, struct malloc_type
*type
,
1113 const char *file
, int line
)
1116 kstrndup(const char *str
, size_t maxlen
, struct malloc_type
*type
)
1119 int zlen
; /* length inclusive of terminating NUL */
1124 zlen
= strnlen(str
, maxlen
) + 1;
1125 nstr
= kmalloc_debug(zlen
, type
, M_WAITOK
, file
, line
);
1126 bcopy(str
, nstr
, zlen
);
1127 nstr
[zlen
- 1] = '\0';
1132 * Notify our cpu that a remote cpu has freed some chunks in a zone that
1133 * we own. RCount will be bumped so the memory should be good, but validate
1134 * that it really is.
1137 kfree_remote(void *ptr
)
1144 slgd
= &mycpu
->gd_slab
;
1147 KKASSERT(*kup
== -((int)mycpuid
+ 1));
1148 KKASSERT(z
->z_RCount
> 0);
1149 atomic_subtract_int(&z
->z_RCount
, 1);
1151 logmemory(free_rem_beg
, z
, NULL
, 0L, 0);
1152 KKASSERT(z
->z_Magic
== ZALLOC_SLAB_MAGIC
);
1153 KKASSERT(z
->z_Cpu
== mycpu
->gd_cpuid
);
1157 * Indicate that we will no longer be off of the ZoneAry by
1164 * Atomically extract the bchunks list and then process it back
1165 * into the lchunks list. We want to append our bchunks to the
1166 * lchunks list and not prepend since we likely do not have
1167 * cache mastership of the related data (not that it helps since
1168 * we are using c_Next).
1170 clean_zone_rchunks(z
);
1171 if (z
->z_NFree
&& nfree
== 0) {
1172 TAILQ_INSERT_HEAD(&slgd
->ZoneAry
[z
->z_ZoneIndex
], z
, z_Entry
);
1175 check_zone_free(slgd
, z
);
1176 logmemory(free_rem_end
, z
, NULL
, 0L, 0);
1180 * free (SLAB ALLOCATOR)
1182 * Free a memory block previously allocated by malloc.
1184 * Note: We do not attempt to update ks_loosememuse as MP races could
1185 * prevent us from checking memory limits in malloc. YYY we may
1186 * consider updating ks_cpu.loosememuse.
1191 kfree(void *ptr
, struct malloc_type
*type
)
1196 struct globaldata
*gd
;
1202 logmemory_quick(free_beg
);
1204 slgd
= &gd
->gd_slab
;
1207 panic("trying to free NULL pointer");
1210 * Handle special 0-byte allocations
1212 if (ptr
== ZERO_LENGTH_PTR
) {
1213 logmemory(free_zero
, ptr
, type
, -1UL, 0);
1214 logmemory_quick(free_end
);
1219 * Panic on bad malloc type
1221 if (type
->ks_magic
!= M_MAGIC
)
1222 panic("free: malloc type lacks magic");
1225 * Handle oversized allocations. XXX we really should require that a
1226 * size be passed to free() instead of this nonsense.
1228 * This code is never called via an ipi.
1232 size
= *kup
<< PAGE_SHIFT
;
1235 KKASSERT(sizeof(weirdary
) <= size
);
1236 bcopy(weirdary
, ptr
, sizeof(weirdary
));
1239 * NOTE: For oversized allocations we do not record the
1240 * originating cpu. It gets freed on the cpu calling
1241 * kfree(). The statistics are in aggregate.
1243 * note: XXX we have still inherited the interrupts-can't-block
1244 * assumption. An interrupt thread does not bump
1245 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1246 * primarily until we can fix softupdate's assumptions about free().
1249 --type
->ks_use
[gd
->gd_cpuid
].inuse
;
1250 type
->ks_use
[gd
->gd_cpuid
].memuse
-= size
;
1251 if (mycpu
->gd_intr_nesting_level
||
1252 (gd
->gd_curthread
->td_flags
& TDF_INTTHREAD
)) {
1253 logmemory(free_ovsz_delayed
, ptr
, type
, size
, 0);
1255 z
->z_Magic
= ZALLOC_OVSZ_MAGIC
;
1256 z
->z_ChunkSize
= size
;
1258 TAILQ_INSERT_HEAD(&slgd
->FreeOvZones
, z
, z_Entry
);
1262 logmemory(free_ovsz
, ptr
, type
, size
, 0);
1263 kmem_slab_free(ptr
, size
); /* may block */
1264 atomic_add_int(&ZoneBigAlloc
, -(int)size
/ 1024);
1266 logmemory_quick(free_end
);
1271 * Zone case. Figure out the zone based on the fact that it is
1274 z
= (SLZone
*)((uintptr_t)ptr
& ZoneMask
);
1277 KKASSERT(z
->z_Magic
== ZALLOC_SLAB_MAGIC
);
1280 * If we do not own the zone then use atomic ops to free to the
1281 * remote cpu linked list and notify the target zone using a
1284 * The target zone cannot be deallocated while we own a chunk of it,
1285 * so the zone header's storage is stable until the very moment
1286 * we adjust z_RChunks. After that we cannot safely dereference (z).
1288 * (no critical section needed)
1290 if (z
->z_CpuGd
!= gd
) {
1292 * Making these adjustments now allow us to avoid passing (type)
1293 * to the remote cpu. Note that inuse/memuse is being
1294 * adjusted on OUR cpu, not the zone cpu, but it should all still
1295 * sum up properly and cancel out.
1298 --type
->ks_use
[gd
->gd_cpuid
].inuse
;
1299 type
->ks_use
[gd
->gd_cpuid
].memuse
-= z
->z_ChunkSize
;
1303 * WARNING! This code competes with other cpus. Once we
1304 * successfully link the chunk to RChunks the remote
1305 * cpu can rip z's storage out from under us.
1307 * Bumping RCount prevents z's storage from getting
1310 rsignal
= z
->z_RSignal
;
1313 atomic_add_int(&z
->z_RCount
, 1);
1317 bchunk
= z
->z_RChunks
;
1319 chunk
->c_Next
= bchunk
;
1322 if (atomic_cmpset_ptr(&z
->z_RChunks
, bchunk
, chunk
))
1327 * We have to signal the remote cpu if our actions will cause
1328 * the remote zone to be placed back on ZoneAry so it can
1329 * move the zone back on.
1331 * We only need to deal with NULL->non-NULL RChunk transitions
1332 * and only if z_RSignal is set. We interlock by reading rsignal
1333 * before adding our chunk to RChunks. This should result in
1334 * virtually no IPI traffic.
1336 * We can use a passive IPI to reduce overhead even further.
1338 if (bchunk
== NULL
&& rsignal
) {
1339 logmemory(free_request
, ptr
, type
,
1340 (unsigned long)z
->z_ChunkSize
, 0);
1341 lwkt_send_ipiq_passive(z
->z_CpuGd
, kfree_remote
, z
);
1342 /* z can get ripped out from under us from this point on */
1343 } else if (rsignal
) {
1344 atomic_subtract_int(&z
->z_RCount
, 1);
1345 /* z can get ripped out from under us from this point on */
1347 logmemory_quick(free_end
);
1354 logmemory(free_chunk
, ptr
, type
, (unsigned long)z
->z_ChunkSize
, 0);
1358 chunk_mark_free(z
, chunk
);
1361 * Put weird data into the memory to detect modifications after freeing,
1362 * illegal pointer use after freeing (we should fault on the odd address),
1363 * and so forth. XXX needs more work, see the old malloc code.
1366 if (z
->z_ChunkSize
< sizeof(weirdary
))
1367 bcopy(weirdary
, chunk
, z
->z_ChunkSize
);
1369 bcopy(weirdary
, chunk
, sizeof(weirdary
));
1373 * Add this free non-zero'd chunk to a linked list for reuse. Add
1374 * to the front of the linked list so it is more likely to be
1375 * reallocated, since it is already in our L1 cache.
1378 if ((vm_offset_t
)chunk
< KvaStart
|| (vm_offset_t
)chunk
>= KvaEnd
)
1379 panic("BADFREE %p", chunk
);
1381 chunk
->c_Next
= z
->z_LChunks
;
1382 z
->z_LChunks
= chunk
;
1383 if (chunk
->c_Next
== NULL
)
1384 z
->z_LChunksp
= &chunk
->c_Next
;
1387 if (chunk
->c_Next
&& (vm_offset_t
)chunk
->c_Next
< KvaStart
)
1392 * Bump the number of free chunks. If it becomes non-zero the zone
1393 * must be added back onto the appropriate list. A fully allocated
1394 * zone that sees its first free is considered 'mature' and is placed
1395 * at the head, giving the system time to potentially free the remaining
1396 * entries even while other allocations are going on and making the zone
1399 if (z
->z_NFree
++ == 0) {
1401 TAILQ_INSERT_TAIL(&slgd
->ZoneAry
[z
->z_ZoneIndex
], z
, z_Entry
);
1403 TAILQ_INSERT_HEAD(&slgd
->ZoneAry
[z
->z_ZoneIndex
], z
, z_Entry
);
1406 --type
->ks_use
[gd
->gd_cpuid
].inuse
;
1407 type
->ks_use
[gd
->gd_cpuid
].memuse
-= z
->z_ChunkSize
;
1409 check_zone_free(slgd
, z
);
1410 logmemory_quick(free_end
);
1415 * Cleanup slabs which are hanging around due to RChunks or which are wholely
1416 * free and can be moved to the free list if not moved by other means.
1418 * Called once every 10 seconds on all cpus.
1423 SLGlobalData
*slgd
= &mycpu
->gd_slab
;
1428 for (i
= 0; i
< NZONES
; ++i
) {
1429 if ((z
= TAILQ_FIRST(&slgd
->ZoneAry
[i
])) == NULL
)
1437 * Shift all RChunks to the end of the LChunks list. This is
1438 * an O(1) operation.
1440 * Then free the zone if possible.
1442 clean_zone_rchunks(z
);
1443 z
= check_zone_free(slgd
, z
);
1449 #if defined(INVARIANTS)
1452 * Helper routines for sanity checks
1455 chunk_mark_allocated(SLZone
*z
, void *chunk
)
1457 int bitdex
= ((char *)chunk
- (char *)z
->z_BasePtr
) / z
->z_ChunkSize
;
1460 KKASSERT((((intptr_t)chunk
^ (intptr_t)z
) & ZoneMask
) == 0);
1461 KASSERT(bitdex
>= 0 && bitdex
< z
->z_NMax
,
1462 ("memory chunk %p bit index %d is illegal", chunk
, bitdex
));
1463 bitptr
= &z
->z_Bitmap
[bitdex
>> 5];
1465 KASSERT((*bitptr
& (1 << bitdex
)) == 0,
1466 ("memory chunk %p is already allocated!", chunk
));
1467 *bitptr
|= 1 << bitdex
;
1471 chunk_mark_free(SLZone
*z
, void *chunk
)
1473 int bitdex
= ((char *)chunk
- (char *)z
->z_BasePtr
) / z
->z_ChunkSize
;
1476 KKASSERT((((intptr_t)chunk
^ (intptr_t)z
) & ZoneMask
) == 0);
1477 KASSERT(bitdex
>= 0 && bitdex
< z
->z_NMax
,
1478 ("memory chunk %p bit index %d is illegal!", chunk
, bitdex
));
1479 bitptr
= &z
->z_Bitmap
[bitdex
>> 5];
1481 KASSERT((*bitptr
& (1 << bitdex
)) != 0,
1482 ("memory chunk %p is already free!", chunk
));
1483 *bitptr
&= ~(1 << bitdex
);
1491 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1492 * specified alignment. M_* flags are expected in the flags field.
1494 * Alignment must be a multiple of PAGE_SIZE.
1496 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1497 * but when we move zalloc() over to use this function as its backend
1498 * we will have to switch to kreserve/krelease and call reserve(0)
1499 * after the new space is made available.
1501 * Interrupt code which has preempted other code is not allowed to
1502 * use PQ_CACHE pages. However, if an interrupt thread is run
1503 * non-preemptively or blocks and then runs non-preemptively, then
1504 * it is free to use PQ_CACHE pages. <--- may not apply any longer XXX
1507 kmem_slab_alloc(vm_size_t size
, vm_offset_t align
, int flags
)
1511 int count
, vmflags
, base_vmflags
;
1512 vm_page_t mbase
= NULL
;
1516 size
= round_page(size
);
1517 addr
= vm_map_min(&kernel_map
);
1519 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1521 vm_map_lock(&kernel_map
);
1522 if (vm_map_findspace(&kernel_map
, addr
, size
, align
, 0, &addr
)) {
1523 vm_map_unlock(&kernel_map
);
1524 if ((flags
& M_NULLOK
) == 0)
1525 panic("kmem_slab_alloc(): kernel_map ran out of space!");
1526 vm_map_entry_release(count
);
1532 * kernel_object maps 1:1 to kernel_map.
1534 vm_object_hold(&kernel_object
);
1535 vm_object_reference_locked(&kernel_object
);
1536 vm_map_insert(&kernel_map
, &count
,
1537 &kernel_object
, NULL
,
1538 addr
, addr
, addr
+ size
,
1541 VM_PROT_ALL
, VM_PROT_ALL
, 0);
1542 vm_object_drop(&kernel_object
);
1543 vm_map_set_wired_quick(&kernel_map
, addr
, size
, &count
);
1544 vm_map_unlock(&kernel_map
);
1550 base_vmflags
|= VM_ALLOC_ZERO
;
1551 if (flags
& M_USE_RESERVE
)
1552 base_vmflags
|= VM_ALLOC_SYSTEM
;
1553 if (flags
& M_USE_INTERRUPT_RESERVE
)
1554 base_vmflags
|= VM_ALLOC_INTERRUPT
;
1555 if ((flags
& (M_RNOWAIT
|M_WAITOK
)) == 0) {
1556 panic("kmem_slab_alloc: bad flags %08x (%p)",
1557 flags
, ((int **)&size
)[-1]);
1561 * Allocate the pages. Do not map them yet. VM_ALLOC_NORMAL can only
1562 * be set if we are not preempting.
1564 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1565 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1566 * implied in this case), though I'm not sure if we really need to
1569 vmflags
= base_vmflags
;
1570 if (flags
& M_WAITOK
) {
1571 if (td
->td_preempted
)
1572 vmflags
|= VM_ALLOC_SYSTEM
;
1574 vmflags
|= VM_ALLOC_NORMAL
;
1577 vm_object_hold(&kernel_object
);
1578 for (i
= 0; i
< size
; i
+= PAGE_SIZE
) {
1579 m
= vm_page_alloc(&kernel_object
, OFF_TO_IDX(addr
+ i
), vmflags
);
1584 * If the allocation failed we either return NULL or we retry.
1586 * If M_WAITOK is specified we wait for more memory and retry.
1587 * If M_WAITOK is specified from a preemption we yield instead of
1588 * wait. Livelock will not occur because the interrupt thread
1589 * will not be preempting anyone the second time around after the
1593 if (flags
& M_WAITOK
) {
1594 if (td
->td_preempted
) {
1599 i
-= PAGE_SIZE
; /* retry */
1607 * Check and deal with an allocation failure
1612 m
= vm_page_lookup(&kernel_object
, OFF_TO_IDX(addr
+ i
));
1613 /* page should already be busy */
1616 vm_map_lock(&kernel_map
);
1617 vm_map_delete(&kernel_map
, addr
, addr
+ size
, &count
);
1618 vm_map_unlock(&kernel_map
);
1619 vm_object_drop(&kernel_object
);
1621 vm_map_entry_release(count
);
1629 * NOTE: The VM pages are still busied. mbase points to the first one
1630 * but we have to iterate via vm_page_next()
1632 vm_object_drop(&kernel_object
);
1636 * Enter the pages into the pmap and deal with M_ZERO.
1643 * page should already be busy
1645 m
->valid
= VM_PAGE_BITS_ALL
;
1647 pmap_enter(&kernel_pmap
, addr
+ i
, m
,
1648 VM_PROT_ALL
| VM_PROT_NOSYNC
, 1, NULL
);
1650 pagezero((char *)addr
+ i
);
1651 KKASSERT(m
->flags
& (PG_WRITEABLE
| PG_MAPPED
));
1652 vm_page_flag_set(m
, PG_REFERENCED
);
1656 vm_object_hold(&kernel_object
);
1657 m
= vm_page_next(m
);
1658 vm_object_drop(&kernel_object
);
1661 vm_map_entry_release(count
);
1662 atomic_add_long(&SlabsAllocated
, 1);
1663 return((void *)addr
);
1670 kmem_slab_free(void *ptr
, vm_size_t size
)
1673 vm_map_remove(&kernel_map
, (vm_offset_t
)ptr
, (vm_offset_t
)ptr
+ size
);
1674 atomic_add_long(&SlabsFreed
, 1);
1679 kmalloc_cachealign(unsigned long size_alloc
, struct malloc_type
*type
,
1682 #if (__VM_CACHELINE_SIZE == 32)
1683 #define CAN_CACHEALIGN(sz) ((sz) >= 256)
1684 #elif (__VM_CACHELINE_SIZE == 64)
1685 #define CAN_CACHEALIGN(sz) ((sz) >= 512)
1686 #elif (__VM_CACHELINE_SIZE == 128)
1687 #define CAN_CACHEALIGN(sz) ((sz) >= 1024)
1689 #error "unsupported cacheline size"
1694 if (size_alloc
< __VM_CACHELINE_SIZE
)
1695 size_alloc
= __VM_CACHELINE_SIZE
;
1696 else if (!CAN_CACHEALIGN(size_alloc
))
1697 flags
|= M_POWEROF2
;
1700 ret
= kmalloc_debug(size_alloc
, type
, flags
, __FILE__
, __LINE__
);
1702 ret
= kmalloc(size_alloc
, type
, flags
);
1704 KASSERT(((uintptr_t)ret
& (__VM_CACHELINE_SIZE
- 1)) == 0,
1705 ("%p(%lu) not cacheline %d aligned",
1706 ret
, size_alloc
, __VM_CACHELINE_SIZE
));
1709 #undef CAN_CACHEALIGN