2 * KERN_SLABALLOC.C - Kernel SLAB memory allocator (MP SAFE)
4 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
6 * This code is derived from software contributed to The DragonFly Project
7 * by Matthew Dillon <dillon@backplane.com>
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.37 2005/12/07 04:49:54 dillon Exp $
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 arysize(ary) (sizeof(ary)/sizeof((ary)[0]))
122 #define MEMORY_STRING "ptr=%p type=%p size=%d flags=%04x"
123 #define MEMORY_ARG_SIZE (sizeof(void *) * 2 + sizeof(unsigned long) + \
126 #if !defined(KTR_MEMORY)
127 #define KTR_MEMORY KTR_ALL
129 KTR_INFO_MASTER(memory
);
130 KTR_INFO(KTR_MEMORY
, memory
, malloc
, 0, MEMORY_STRING
, MEMORY_ARG_SIZE
);
131 KTR_INFO(KTR_MEMORY
, memory
, free_zero
, 1, MEMORY_STRING
, MEMORY_ARG_SIZE
);
132 KTR_INFO(KTR_MEMORY
, memory
, free_ovsz
, 2, MEMORY_STRING
, MEMORY_ARG_SIZE
);
133 KTR_INFO(KTR_MEMORY
, memory
, free_ovsz_delayed
, 3, MEMORY_STRING
, MEMORY_ARG_SIZE
);
134 KTR_INFO(KTR_MEMORY
, memory
, free_chunk
, 4, MEMORY_STRING
, MEMORY_ARG_SIZE
);
136 KTR_INFO(KTR_MEMORY
, memory
, free_request
, 5, MEMORY_STRING
, MEMORY_ARG_SIZE
);
137 KTR_INFO(KTR_MEMORY
, memory
, free_remote
, 6, MEMORY_STRING
, MEMORY_ARG_SIZE
);
139 KTR_INFO(KTR_MEMORY
, memory
, malloc_beg
, 0, "malloc begin", 0);
140 KTR_INFO(KTR_MEMORY
, memory
, free_beg
, 0, "free begin", 0);
141 KTR_INFO(KTR_MEMORY
, memory
, free_end
, 0, "free end", 0);
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
;
155 static struct malloc_type
*kmemstatistics
;
156 static struct kmemusage
*kmemusage
;
157 static int32_t weirdary
[16];
159 static void *kmem_slab_alloc(vm_size_t bytes
, vm_offset_t align
, int flags
);
160 static void kmem_slab_free(void *ptr
, vm_size_t bytes
);
161 #if defined(INVARIANTS)
162 static void chunk_mark_allocated(SLZone
*z
, void *chunk
);
163 static void chunk_mark_free(SLZone
*z
, void *chunk
);
167 * Misc constants. Note that allocations that are exact multiples of
168 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
169 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
171 #define MIN_CHUNK_SIZE 8 /* in bytes */
172 #define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
173 #define ZONE_RELS_THRESH 2 /* threshold number of zones */
174 #define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
177 * The WEIRD_ADDR is used as known text to copy into free objects to
178 * try to create deterministic failure cases if the data is accessed after
181 #define WEIRD_ADDR 0xdeadc0de
182 #define MAX_COPY sizeof(weirdary)
183 #define ZERO_LENGTH_PTR ((void *)-8)
186 * Misc global malloc buckets
189 MALLOC_DEFINE(M_CACHE
, "cache", "Various Dynamically allocated caches");
190 MALLOC_DEFINE(M_DEVBUF
, "devbuf", "device driver memory");
191 MALLOC_DEFINE(M_TEMP
, "temp", "misc temporary data buffers");
193 MALLOC_DEFINE(M_IP6OPT
, "ip6opt", "IPv6 options");
194 MALLOC_DEFINE(M_IP6NDP
, "ip6ndp", "IPv6 Neighbor Discovery");
197 * Initialize the slab memory allocator. We have to choose a zone size based
198 * on available physical memory. We choose a zone side which is approximately
199 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
200 * 128K. The zone size is limited to the bounds set in slaballoc.h
201 * (typically 32K min, 128K max).
203 static void kmeminit(void *dummy
);
205 SYSINIT(kmem
, SI_SUB_KMEM
, SI_ORDER_FIRST
, kmeminit
, NULL
)
209 * If enabled any memory allocated without M_ZERO is initialized to -1.
211 static int use_malloc_pattern
;
212 SYSCTL_INT(_debug
, OID_AUTO
, use_malloc_pattern
, CTLFLAG_RW
,
213 &use_malloc_pattern
, 0, "");
217 kmeminit(void *dummy
)
224 limsize
= (vm_poff_t
)vmstats
.v_page_count
* PAGE_SIZE
;
225 if (limsize
> VM_MAX_KERNEL_ADDRESS
- VM_MIN_KERNEL_ADDRESS
)
226 limsize
= VM_MAX_KERNEL_ADDRESS
- VM_MIN_KERNEL_ADDRESS
;
228 usesize
= (int)(limsize
/ 1024); /* convert to KB */
230 ZoneSize
= ZALLOC_MIN_ZONE_SIZE
;
231 while (ZoneSize
< ZALLOC_MAX_ZONE_SIZE
&& (ZoneSize
<< 1) < usesize
)
233 ZoneLimit
= ZoneSize
/ 4;
234 if (ZoneLimit
> ZALLOC_ZONE_LIMIT
)
235 ZoneLimit
= ZALLOC_ZONE_LIMIT
;
236 ZoneMask
= ZoneSize
- 1;
237 ZonePageCount
= ZoneSize
/ PAGE_SIZE
;
239 npg
= (VM_MAX_KERNEL_ADDRESS
- VM_MIN_KERNEL_ADDRESS
) / PAGE_SIZE
;
240 kmemusage
= kmem_slab_alloc(npg
* sizeof(struct kmemusage
), PAGE_SIZE
, M_WAITOK
|M_ZERO
);
242 for (i
= 0; i
< arysize(weirdary
); ++i
)
243 weirdary
[i
] = WEIRD_ADDR
;
246 printf("Slab ZoneSize set to %dKB\n", ZoneSize
/ 1024);
250 * Initialize a malloc type tracking structure.
253 malloc_init(void *data
)
255 struct malloc_type
*type
= data
;
258 if (type
->ks_magic
!= M_MAGIC
)
259 panic("malloc type lacks magic");
261 if (type
->ks_limit
!= 0)
264 if (vmstats
.v_page_count
== 0)
265 panic("malloc_init not allowed before vm init");
267 limsize
= (vm_poff_t
)vmstats
.v_page_count
* PAGE_SIZE
;
268 if (limsize
> VM_MAX_KERNEL_ADDRESS
- VM_MIN_KERNEL_ADDRESS
)
269 limsize
= VM_MAX_KERNEL_ADDRESS
- VM_MIN_KERNEL_ADDRESS
;
270 type
->ks_limit
= limsize
/ 10;
272 type
->ks_next
= kmemstatistics
;
273 kmemstatistics
= type
;
277 malloc_uninit(void *data
)
279 struct malloc_type
*type
= data
;
280 struct malloc_type
*t
;
286 if (type
->ks_magic
!= M_MAGIC
)
287 panic("malloc type lacks magic");
289 if (vmstats
.v_page_count
== 0)
290 panic("malloc_uninit not allowed before vm init");
292 if (type
->ks_limit
== 0)
293 panic("malloc_uninit on uninitialized type");
297 * memuse is only correct in aggregation. Due to memory being allocated
298 * on one cpu and freed on another individual array entries may be
299 * negative or positive (canceling each other out).
301 for (i
= ttl
= 0; i
< ncpus
; ++i
)
302 ttl
+= type
->ks_memuse
[i
];
304 printf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
305 ttl
, type
->ks_shortdesc
, i
);
308 if (type
== kmemstatistics
) {
309 kmemstatistics
= type
->ks_next
;
311 for (t
= kmemstatistics
; t
->ks_next
!= NULL
; t
= t
->ks_next
) {
312 if (t
->ks_next
== type
) {
313 t
->ks_next
= type
->ks_next
;
318 type
->ks_next
= NULL
;
323 * Calculate the zone index for the allocation request size and set the
324 * allocation request size to that particular zone's chunk size.
327 zoneindex(unsigned long *bytes
)
329 unsigned int n
= (unsigned int)*bytes
; /* unsigned for shift opt */
331 *bytes
= n
= (n
+ 7) & ~7;
332 return(n
/ 8 - 1); /* 8 byte chunks, 16 zones */
335 *bytes
= n
= (n
+ 15) & ~15;
340 *bytes
= n
= (n
+ 31) & ~31;
344 *bytes
= n
= (n
+ 63) & ~63;
348 *bytes
= n
= (n
+ 127) & ~127;
349 return(n
/ 128 + 31);
352 *bytes
= n
= (n
+ 255) & ~255;
353 return(n
/ 256 + 39);
355 *bytes
= n
= (n
+ 511) & ~511;
356 return(n
/ 512 + 47);
358 #if ZALLOC_ZONE_LIMIT > 8192
360 *bytes
= n
= (n
+ 1023) & ~1023;
361 return(n
/ 1024 + 55);
364 #if ZALLOC_ZONE_LIMIT > 16384
366 *bytes
= n
= (n
+ 2047) & ~2047;
367 return(n
/ 2048 + 63);
370 panic("Unexpected byte count %d", n
);
375 * malloc() (SLAB ALLOCATOR) (MPSAFE)
377 * Allocate memory via the slab allocator. If the request is too large,
378 * or if it page-aligned beyond a certain size, we fall back to the
379 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
380 * &SlabMisc if you don't care.
382 * M_RNOWAIT - don't block.
383 * M_NULLOK - return NULL instead of blocking.
384 * M_ZERO - zero the returned memory.
385 * M_USE_RESERVE - allow greater drawdown of the free list
386 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
389 malloc(unsigned long size
, struct malloc_type
*type
, int flags
)
394 struct globaldata
*gd
;
400 logmemory_quick(malloc_beg
);
405 * XXX silly to have this in the critical path.
407 if (type
->ks_limit
== 0) {
409 if (type
->ks_limit
== 0)
416 * Handle the case where the limit is reached. Panic if we can't return
417 * NULL. The original malloc code looped, but this tended to
418 * simply deadlock the computer.
420 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
421 * to determine if a more complete limit check should be done. The
422 * actual memory use is tracked via ks_memuse[cpu].
424 while (type
->ks_loosememuse
>= type
->ks_limit
) {
428 for (i
= ttl
= 0; i
< ncpus
; ++i
)
429 ttl
+= type
->ks_memuse
[i
];
430 type
->ks_loosememuse
= ttl
; /* not MP synchronized */
431 if (ttl
>= type
->ks_limit
) {
432 if (flags
& M_NULLOK
) {
433 logmemory(malloc
, NULL
, type
, size
, flags
);
436 panic("%s: malloc limit exceeded", type
->ks_shortdesc
);
441 * Handle the degenerate size == 0 case. Yes, this does happen.
442 * Return a special pointer. This is to maintain compatibility with
443 * the original malloc implementation. Certain devices, such as the
444 * adaptec driver, not only allocate 0 bytes, they check for NULL and
445 * also realloc() later on. Joy.
448 logmemory(malloc
, ZERO_LENGTH_PTR
, type
, size
, flags
);
449 return(ZERO_LENGTH_PTR
);
453 * Handle hysteresis from prior frees here in malloc(). We cannot
454 * safely manipulate the kernel_map in free() due to free() possibly
455 * being called via an IPI message or from sensitive interrupt code.
457 while (slgd
->NFreeZones
> ZONE_RELS_THRESH
&& (flags
& M_RNOWAIT
) == 0) {
459 if (slgd
->NFreeZones
> ZONE_RELS_THRESH
) { /* crit sect race */
461 slgd
->FreeZones
= z
->z_Next
;
463 kmem_slab_free(z
, ZoneSize
); /* may block */
468 * XXX handle oversized frees that were queued from free().
470 while (slgd
->FreeOvZones
&& (flags
& M_RNOWAIT
) == 0) {
472 if ((z
= slgd
->FreeOvZones
) != NULL
) {
473 KKASSERT(z
->z_Magic
== ZALLOC_OVSZ_MAGIC
);
474 slgd
->FreeOvZones
= z
->z_Next
;
475 kmem_slab_free(z
, z
->z_ChunkSize
); /* may block */
481 * Handle large allocations directly. There should not be very many of
482 * these so performance is not a big issue.
484 * Guarentee page alignment for allocations in multiples of PAGE_SIZE
486 if (size
>= ZoneLimit
|| (size
& PAGE_MASK
) == 0) {
487 struct kmemusage
*kup
;
489 size
= round_page(size
);
490 chunk
= kmem_slab_alloc(size
, PAGE_SIZE
, flags
);
492 logmemory(malloc
, NULL
, type
, size
, flags
);
495 flags
&= ~M_ZERO
; /* result already zero'd if M_ZERO was set */
496 flags
|= M_PASSIVE_ZERO
;
498 kup
->ku_pagecnt
= size
/ PAGE_SIZE
;
499 kup
->ku_cpu
= gd
->gd_cpuid
;
505 * Attempt to allocate out of an existing zone. First try the free list,
506 * then allocate out of unallocated space. If we find a good zone move
507 * it to the head of the list so later allocations find it quickly
508 * (we might have thousands of zones in the list).
510 * Note: zoneindex() will panic of size is too large.
512 zi
= zoneindex(&size
);
513 KKASSERT(zi
< NZONES
);
515 if ((z
= slgd
->ZoneAry
[zi
]) != NULL
) {
516 KKASSERT(z
->z_NFree
> 0);
519 * Remove us from the ZoneAry[] when we become empty
521 if (--z
->z_NFree
== 0) {
522 slgd
->ZoneAry
[zi
] = z
->z_Next
;
527 * Locate a chunk in a free page. This attempts to localize
528 * reallocations into earlier pages without us having to sort
529 * the chunk list. A chunk may still overlap a page boundary.
531 while (z
->z_FirstFreePg
< ZonePageCount
) {
532 if ((chunk
= z
->z_PageAry
[z
->z_FirstFreePg
]) != NULL
) {
535 * Diagnostic: c_Next is not total garbage.
537 KKASSERT(chunk
->c_Next
== NULL
||
538 ((intptr_t)chunk
->c_Next
& IN_SAME_PAGE_MASK
) ==
539 ((intptr_t)chunk
& IN_SAME_PAGE_MASK
));
542 if ((uintptr_t)chunk
< VM_MIN_KERNEL_ADDRESS
)
543 panic("chunk %p FFPG %d/%d", chunk
, z
->z_FirstFreePg
, ZonePageCount
);
544 if (chunk
->c_Next
&& (uintptr_t)chunk
->c_Next
< VM_MIN_KERNEL_ADDRESS
)
545 panic("chunkNEXT %p %p FFPG %d/%d", chunk
, chunk
->c_Next
, z
->z_FirstFreePg
, ZonePageCount
);
546 chunk_mark_allocated(z
, chunk
);
548 z
->z_PageAry
[z
->z_FirstFreePg
] = chunk
->c_Next
;
555 * No chunks are available but NFree said we had some memory, so
556 * it must be available in the never-before-used-memory area
557 * governed by UIndex. The consequences are very serious if our zone
558 * got corrupted so we use an explicit panic rather then a KASSERT.
560 if (z
->z_UIndex
+ 1 != z
->z_NMax
)
561 z
->z_UIndex
= z
->z_UIndex
+ 1;
564 if (z
->z_UIndex
== z
->z_UEndIndex
)
565 panic("slaballoc: corrupted zone");
566 chunk
= (SLChunk
*)(z
->z_BasePtr
+ z
->z_UIndex
* size
);
567 if ((z
->z_Flags
& SLZF_UNOTZEROD
) == 0) {
569 flags
|= M_PASSIVE_ZERO
;
571 #if defined(INVARIANTS)
572 chunk_mark_allocated(z
, chunk
);
578 * If all zones are exhausted we need to allocate a new zone for this
579 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
580 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
581 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
582 * we do not pre-zero it because we do not want to mess up the L1 cache.
584 * At least one subsystem, the tty code (see CROUND) expects power-of-2
585 * allocations to be power-of-2 aligned. We maintain compatibility by
586 * adjusting the base offset below.
591 if ((z
= slgd
->FreeZones
) != NULL
) {
592 slgd
->FreeZones
= z
->z_Next
;
594 bzero(z
, sizeof(SLZone
));
595 z
->z_Flags
|= SLZF_UNOTZEROD
;
597 z
= kmem_slab_alloc(ZoneSize
, ZoneSize
, flags
|M_ZERO
);
603 * How big is the base structure?
605 #if defined(INVARIANTS)
607 * Make room for z_Bitmap. An exact calculation is somewhat more
608 * complicated so don't make an exact calculation.
610 off
= offsetof(SLZone
, z_Bitmap
[(ZoneSize
/ size
+ 31) / 32]);
611 bzero(z
->z_Bitmap
, (ZoneSize
/ size
+ 31) / 8);
613 off
= sizeof(SLZone
);
617 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
618 * Otherwise just 8-byte align the data.
620 if ((size
| (size
- 1)) + 1 == (size
<< 1))
621 off
= (off
+ size
- 1) & ~(size
- 1);
623 off
= (off
+ MIN_CHUNK_MASK
) & ~MIN_CHUNK_MASK
;
624 z
->z_Magic
= ZALLOC_SLAB_MAGIC
;
626 z
->z_NMax
= (ZoneSize
- off
) / size
;
627 z
->z_NFree
= z
->z_NMax
- 1;
628 z
->z_BasePtr
= (char *)z
+ off
;
629 z
->z_UIndex
= z
->z_UEndIndex
= slgd
->JunkIndex
% z
->z_NMax
;
630 z
->z_ChunkSize
= size
;
631 z
->z_FirstFreePg
= ZonePageCount
;
633 z
->z_Cpu
= gd
->gd_cpuid
;
634 chunk
= (SLChunk
*)(z
->z_BasePtr
+ z
->z_UIndex
* size
);
635 z
->z_Next
= slgd
->ZoneAry
[zi
];
636 slgd
->ZoneAry
[zi
] = z
;
637 if ((z
->z_Flags
& SLZF_UNOTZEROD
) == 0) {
638 flags
&= ~M_ZERO
; /* already zero'd */
639 flags
|= M_PASSIVE_ZERO
;
641 #if defined(INVARIANTS)
642 chunk_mark_allocated(z
, chunk
);
646 * Slide the base index for initial allocations out of the next
647 * zone we create so we do not over-weight the lower part of the
650 slgd
->JunkIndex
= (slgd
->JunkIndex
+ ZALLOC_SLAB_SLIDE
)
651 & (ZALLOC_MAX_ZONE_SIZE
- 1);
654 ++type
->ks_inuse
[gd
->gd_cpuid
];
655 type
->ks_memuse
[gd
->gd_cpuid
] += size
;
656 type
->ks_loosememuse
+= size
; /* not MP synchronized */
661 else if ((flags
& (M_ZERO
|M_PASSIVE_ZERO
)) == 0) {
662 if (use_malloc_pattern
) {
663 for (i
= 0; i
< size
; i
+= sizeof(int)) {
664 *(int *)((char *)chunk
+ i
) = -1;
667 chunk
->c_Next
= (void *)-1; /* avoid accidental double-free check */
670 logmemory(malloc
, chunk
, type
, size
, flags
);
674 logmemory(malloc
, NULL
, type
, size
, flags
);
679 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
681 * Generally speaking this routine is not called very often and we do
682 * not attempt to optimize it beyond reusing the same pointer if the
683 * new size fits within the chunking of the old pointer's zone.
686 realloc(void *ptr
, unsigned long size
, struct malloc_type
*type
, int flags
)
692 KKASSERT((flags
& M_ZERO
) == 0); /* not supported */
694 if (ptr
== NULL
|| ptr
== ZERO_LENGTH_PTR
)
695 return(malloc(size
, type
, flags
));
702 * Handle oversized allocations. XXX we really should require that a
703 * size be passed to free() instead of this nonsense.
706 struct kmemusage
*kup
;
709 if (kup
->ku_pagecnt
) {
710 osize
= kup
->ku_pagecnt
<< PAGE_SHIFT
;
711 if (osize
== round_page(size
))
713 if ((nptr
= malloc(size
, type
, flags
)) == NULL
)
715 bcopy(ptr
, nptr
, min(size
, osize
));
722 * Get the original allocation's zone. If the new request winds up
723 * using the same chunk size we do not have to do anything.
725 z
= (SLZone
*)((uintptr_t)ptr
& ~(uintptr_t)ZoneMask
);
726 KKASSERT(z
->z_Magic
== ZALLOC_SLAB_MAGIC
);
729 if (z
->z_ChunkSize
== size
)
733 * Allocate memory for the new request size. Note that zoneindex has
734 * already adjusted the request size to the appropriate chunk size, which
735 * should optimize our bcopy(). Then copy and return the new pointer.
737 if ((nptr
= malloc(size
, type
, flags
)) == NULL
)
739 bcopy(ptr
, nptr
, min(size
, z
->z_ChunkSize
));
745 * Allocate a copy of the specified string.
747 * (MP SAFE) (MAY BLOCK)
750 strdup(const char *str
, struct malloc_type
*type
)
752 int zlen
; /* length inclusive of terminating NUL */
757 zlen
= strlen(str
) + 1;
758 nstr
= malloc(zlen
, type
, M_WAITOK
);
759 bcopy(str
, nstr
, zlen
);
765 * free() (SLAB ALLOCATOR)
767 * Free the specified chunk of memory.
771 free_remote(void *ptr
)
773 logmemory(free_remote
, ptr
, *(struct malloc_type
**)ptr
, -1, 0);
774 free(ptr
, *(struct malloc_type
**)ptr
);
780 * free (SLAB ALLOCATOR) (MP SAFE)
782 * Free a memory block previously allocated by malloc. Note that we do not
783 * attempt to uplodate ks_loosememuse as MP races could prevent us from
784 * checking memory limits in malloc.
787 free(void *ptr
, struct malloc_type
*type
)
792 struct globaldata
*gd
;
795 logmemory_quick(free_beg
);
800 panic("trying to free NULL pointer");
803 * Handle special 0-byte allocations
805 if (ptr
== ZERO_LENGTH_PTR
) {
806 logmemory(free_zero
, ptr
, type
, -1, 0);
807 logmemory_quick(free_end
);
812 * Handle oversized allocations. XXX we really should require that a
813 * size be passed to free() instead of this nonsense.
815 * This code is never called via an ipi.
818 struct kmemusage
*kup
;
822 if (kup
->ku_pagecnt
) {
823 size
= kup
->ku_pagecnt
<< PAGE_SHIFT
;
826 KKASSERT(sizeof(weirdary
) <= size
);
827 bcopy(weirdary
, ptr
, sizeof(weirdary
));
830 * note: we always adjust our cpu's slot, not the originating
831 * cpu (kup->ku_cpuid). The statistics are in aggregate.
833 * note: XXX we have still inherited the interrupts-can't-block
834 * assumption. An interrupt thread does not bump
835 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
836 * primarily until we can fix softupdate's assumptions about free().
839 --type
->ks_inuse
[gd
->gd_cpuid
];
840 type
->ks_memuse
[gd
->gd_cpuid
] -= size
;
841 if (mycpu
->gd_intr_nesting_level
|| (gd
->gd_curthread
->td_flags
& TDF_INTTHREAD
)) {
842 logmemory(free_ovsz_delayed
, ptr
, type
, size
, 0);
844 z
->z_Magic
= ZALLOC_OVSZ_MAGIC
;
845 z
->z_Next
= slgd
->FreeOvZones
;
846 z
->z_ChunkSize
= size
;
847 slgd
->FreeOvZones
= z
;
851 logmemory(free_ovsz
, ptr
, type
, size
, 0);
852 kmem_slab_free(ptr
, size
); /* may block */
854 logmemory_quick(free_end
);
860 * Zone case. Figure out the zone based on the fact that it is
863 z
= (SLZone
*)((uintptr_t)ptr
& ~(uintptr_t)ZoneMask
);
864 KKASSERT(z
->z_Magic
== ZALLOC_SLAB_MAGIC
);
867 * If we do not own the zone then forward the request to the
868 * cpu that does. Since the timing is non-critical, a passive
871 if (z
->z_CpuGd
!= gd
) {
872 *(struct malloc_type
**)ptr
= type
;
874 logmemory(free_request
, ptr
, type
, z
->z_ChunkSize
, 0);
875 lwkt_send_ipiq_passive(z
->z_CpuGd
, free_remote
, ptr
);
877 panic("Corrupt SLZone");
879 logmemory_quick(free_end
);
883 logmemory(free_chunk
, ptr
, type
, z
->z_ChunkSize
, 0);
885 if (type
->ks_magic
!= M_MAGIC
)
886 panic("free: malloc type lacks magic");
889 pgno
= ((char *)ptr
- (char *)z
) >> PAGE_SHIFT
;
894 * Attempt to detect a double-free. To reduce overhead we only check
895 * if there appears to be link pointer at the base of the data.
897 if (((intptr_t)chunk
->c_Next
- (intptr_t)z
) >> PAGE_SHIFT
== pgno
) {
899 for (scan
= z
->z_PageAry
[pgno
]; scan
; scan
= scan
->c_Next
) {
901 panic("Double free at %p", chunk
);
904 chunk_mark_free(z
, chunk
);
908 * Put weird data into the memory to detect modifications after freeing,
909 * illegal pointer use after freeing (we should fault on the odd address),
910 * and so forth. XXX needs more work, see the old malloc code.
913 if (z
->z_ChunkSize
< sizeof(weirdary
))
914 bcopy(weirdary
, chunk
, z
->z_ChunkSize
);
916 bcopy(weirdary
, chunk
, sizeof(weirdary
));
920 * Add this free non-zero'd chunk to a linked list for reuse, adjust
924 if ((uintptr_t)chunk
< VM_MIN_KERNEL_ADDRESS
)
925 panic("BADFREE %p", chunk
);
927 chunk
->c_Next
= z
->z_PageAry
[pgno
];
928 z
->z_PageAry
[pgno
] = chunk
;
930 if (chunk
->c_Next
&& (uintptr_t)chunk
->c_Next
< VM_MIN_KERNEL_ADDRESS
)
933 if (z
->z_FirstFreePg
> pgno
)
934 z
->z_FirstFreePg
= pgno
;
937 * Bump the number of free chunks. If it becomes non-zero the zone
938 * must be added back onto the appropriate list.
940 if (z
->z_NFree
++ == 0) {
941 z
->z_Next
= slgd
->ZoneAry
[z
->z_ZoneIndex
];
942 slgd
->ZoneAry
[z
->z_ZoneIndex
] = z
;
945 --type
->ks_inuse
[z
->z_Cpu
];
946 type
->ks_memuse
[z
->z_Cpu
] -= z
->z_ChunkSize
;
949 * If the zone becomes totally free, and there are other zones we
950 * can allocate from, move this zone to the FreeZones list. Since
951 * this code can be called from an IPI callback, do *NOT* try to mess
952 * with kernel_map here. Hysteresis will be performed at malloc() time.
954 if (z
->z_NFree
== z
->z_NMax
&&
955 (z
->z_Next
|| slgd
->ZoneAry
[z
->z_ZoneIndex
] != z
)
959 for (pz
= &slgd
->ZoneAry
[z
->z_ZoneIndex
]; z
!= *pz
; pz
= &(*pz
)->z_Next
)
963 z
->z_Next
= slgd
->FreeZones
;
967 logmemory_quick(free_end
);
971 #if defined(INVARIANTS)
973 * Helper routines for sanity checks
977 chunk_mark_allocated(SLZone
*z
, void *chunk
)
979 int bitdex
= ((char *)chunk
- (char *)z
->z_BasePtr
) / z
->z_ChunkSize
;
982 KASSERT(bitdex
>= 0 && bitdex
< z
->z_NMax
, ("memory chunk %p bit index %d is illegal", chunk
, bitdex
));
983 bitptr
= &z
->z_Bitmap
[bitdex
>> 5];
985 KASSERT((*bitptr
& (1 << bitdex
)) == 0, ("memory chunk %p is already allocated!", chunk
));
986 *bitptr
|= 1 << bitdex
;
991 chunk_mark_free(SLZone
*z
, void *chunk
)
993 int bitdex
= ((char *)chunk
- (char *)z
->z_BasePtr
) / z
->z_ChunkSize
;
996 KASSERT(bitdex
>= 0 && bitdex
< z
->z_NMax
, ("memory chunk %p bit index %d is illegal!", chunk
, bitdex
));
997 bitptr
= &z
->z_Bitmap
[bitdex
>> 5];
999 KASSERT((*bitptr
& (1 << bitdex
)) != 0, ("memory chunk %p is already free!", chunk
));
1000 *bitptr
&= ~(1 << bitdex
);
1006 * kmem_slab_alloc() (MP SAFE) (GETS BGL)
1008 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1009 * specified alignment. M_* flags are expected in the flags field.
1011 * Alignment must be a multiple of PAGE_SIZE.
1013 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1014 * but when we move zalloc() over to use this function as its backend
1015 * we will have to switch to kreserve/krelease and call reserve(0)
1016 * after the new space is made available.
1018 * Interrupt code which has preempted other code is not allowed to
1019 * use PQ_CACHE pages. However, if an interrupt thread is run
1020 * non-preemptively or blocks and then runs non-preemptively, then
1021 * it is free to use PQ_CACHE pages.
1023 * This routine will currently obtain the BGL.
1026 kmem_slab_alloc(vm_size_t size
, vm_offset_t align
, int flags
)
1031 int count
, vmflags
, base_vmflags
;
1033 vm_map_t map
= kernel_map
;
1035 size
= round_page(size
);
1036 addr
= vm_map_min(map
);
1039 * Reserve properly aligned space from kernel_map. RNOWAIT allocations
1042 if (flags
& M_RNOWAIT
) {
1043 if (try_mplock() == 0)
1048 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1051 if (vm_map_findspace(map
, vm_map_min(map
), size
, align
, &addr
)) {
1053 if ((flags
& M_NULLOK
) == 0)
1054 panic("kmem_slab_alloc(): kernel_map ran out of space!");
1056 vm_map_entry_release(count
);
1060 offset
= addr
- VM_MIN_KERNEL_ADDRESS
;
1061 vm_object_reference(kernel_object
);
1062 vm_map_insert(map
, &count
,
1063 kernel_object
, offset
, addr
, addr
+ size
,
1064 VM_PROT_ALL
, VM_PROT_ALL
, 0);
1070 base_vmflags
|= VM_ALLOC_ZERO
;
1071 if (flags
& M_USE_RESERVE
)
1072 base_vmflags
|= VM_ALLOC_SYSTEM
;
1073 if (flags
& M_USE_INTERRUPT_RESERVE
)
1074 base_vmflags
|= VM_ALLOC_INTERRUPT
;
1075 if ((flags
& (M_RNOWAIT
|M_WAITOK
)) == 0)
1076 panic("kmem_slab_alloc: bad flags %08x (%p)", flags
, ((int **)&size
)[-1]);
1080 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
1082 for (i
= 0; i
< size
; i
+= PAGE_SIZE
) {
1084 vm_pindex_t idx
= OFF_TO_IDX(offset
+ i
);
1087 * VM_ALLOC_NORMAL can only be set if we are not preempting.
1089 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1090 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1091 * implied in this case), though I'm sure if we really need to do
1094 vmflags
= base_vmflags
;
1095 if (flags
& M_WAITOK
) {
1096 if (td
->td_preempted
)
1097 vmflags
|= VM_ALLOC_SYSTEM
;
1099 vmflags
|= VM_ALLOC_NORMAL
;
1102 m
= vm_page_alloc(kernel_object
, idx
, vmflags
);
1105 * If the allocation failed we either return NULL or we retry.
1107 * If M_WAITOK is specified we wait for more memory and retry.
1108 * If M_WAITOK is specified from a preemption we yield instead of
1109 * wait. Livelock will not occur because the interrupt thread
1110 * will not be preempting anyone the second time around after the
1114 if (flags
& M_WAITOK
) {
1115 if (td
->td_preempted
) {
1124 i
-= PAGE_SIZE
; /* retry */
1129 * We were unable to recover, cleanup and return NULL
1133 m
= vm_page_lookup(kernel_object
, OFF_TO_IDX(offset
+ i
));
1136 vm_map_delete(map
, addr
, addr
+ size
, &count
);
1139 vm_map_entry_release(count
);
1148 * Mark the map entry as non-pageable using a routine that allows us to
1149 * populate the underlying pages.
1151 vm_map_set_wired_quick(map
, addr
, size
, &count
);
1155 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1157 for (i
= 0; i
< size
; i
+= PAGE_SIZE
) {
1160 m
= vm_page_lookup(kernel_object
, OFF_TO_IDX(offset
+ i
));
1161 m
->valid
= VM_PAGE_BITS_ALL
;
1164 pmap_enter(kernel_pmap
, addr
+ i
, m
, VM_PROT_ALL
, 1);
1165 if ((m
->flags
& PG_ZERO
) == 0 && (flags
& M_ZERO
))
1166 bzero((char *)addr
+ i
, PAGE_SIZE
);
1167 vm_page_flag_clear(m
, PG_ZERO
);
1168 vm_page_flag_set(m
, PG_MAPPED
| PG_WRITEABLE
| PG_REFERENCED
);
1171 vm_map_entry_release(count
);
1173 return((void *)addr
);
1177 * kmem_slab_free() (MP SAFE) (GETS BGL)
1180 kmem_slab_free(void *ptr
, vm_size_t size
)
1184 vm_map_remove(kernel_map
, (vm_offset_t
)ptr
, (vm_offset_t
)ptr
+ size
);