Apply FreeBSD Security Advisory FreeBSD-SA-03:14.arp. Fix DOS crash due
[dragonfly.git] / sys / kern / kern_slaballoc.c
CommitLineData
a108bf71
MD
1/*
2 * KERN_SLABALLOC.C - Kernel SLAB memory allocator
3 *
4 * Copyright (c) 2003 Matthew Dillon <dillon@backplane.com>
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * SUCH DAMAGE.
27 *
3d177b31 28 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.4 2003/09/22 21:45:44 dillon Exp $
a108bf71
MD
29 *
30 * This module implements a slab allocator drop-in replacement for the
31 * kernel malloc().
32 *
33 * A slab allocator reserves a ZONE for each chunk size, then lays the
34 * chunks out in an array within the zone. Allocation and deallocation
35 * is nearly instantanious, and fragmentation/overhead losses are limited
36 * to a fixed worst-case amount.
37 *
38 * The downside of this slab implementation is in the chunk size
39 * multiplied by the number of zones. ~80 zones * 128K = 10MB of VM per cpu.
40 * In a kernel implementation all this memory will be physical so
41 * the zone size is adjusted downward on machines with less physical
42 * memory. The upside is that overhead is bounded... this is the *worst*
43 * case overhead.
44 *
45 * Slab management is done on a per-cpu basis and no locking or mutexes
46 * are required, only a critical section. When one cpu frees memory
47 * belonging to another cpu's slab manager an asynchronous IPI message
48 * will be queued to execute the operation. In addition, both the
49 * high level slab allocator and the low level zone allocator optimize
50 * M_ZERO requests, and the slab allocator does not have to pre initialize
51 * the linked list of chunks.
52 *
53 * XXX Balancing is needed between cpus. Balance will be handled through
54 * asynchronous IPIs primarily by reassigning the z_Cpu ownership of chunks.
55 *
56 * XXX If we have to allocate a new zone and M_USE_RESERVE is set, use of
57 * the new zone should be restricted to M_USE_RESERVE requests only.
58 *
59 * Alloc Size Chunking Number of zones
60 * 0-127 8 16
61 * 128-255 16 8
62 * 256-511 32 8
63 * 512-1023 64 8
64 * 1024-2047 128 8
65 * 2048-4095 256 8
66 * 4096-8191 512 8
67 * 8192-16383 1024 8
68 * 16384-32767 2048 8
69 * (if PAGE_SIZE is 4K the maximum zone allocation is 16383)
70 *
71 * Allocations >= ZALLOC_ZONE_LIMIT go directly to kmem.
72 *
73 * API REQUIREMENTS AND SIDE EFFECTS
74 *
75 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we
76 * have remained compatible with the following API requirements:
77 *
78 * + small power-of-2 sized allocations are power-of-2 aligned (kern_tty)
3d177b31 79 * + all power-of-2 sized allocations are power-of-2 aligned (twe)
a108bf71
MD
80 * + malloc(0) is allowed and returns non-NULL (ahc driver)
81 * + ability to allocate arbitrarily large chunks of memory
82 */
83
84#include "opt_vm.h"
85
86#if defined(USE_SLAB_ALLOCATOR)
87
88#if !defined(NO_KMEM_MAP)
89#error "NO_KMEM_MAP must be defined when USE_SLAB_ALLOCATOR is defined"
90#endif
91
92#include <sys/param.h>
93#include <sys/systm.h>
94#include <sys/kernel.h>
95#include <sys/slaballoc.h>
96#include <sys/mbuf.h>
97#include <sys/vmmeter.h>
98#include <sys/lock.h>
99#include <sys/thread.h>
100#include <sys/globaldata.h>
101
102#include <vm/vm.h>
103#include <vm/vm_param.h>
104#include <vm/vm_kern.h>
105#include <vm/vm_extern.h>
106#include <vm/vm_object.h>
107#include <vm/pmap.h>
108#include <vm/vm_map.h>
109#include <vm/vm_page.h>
110#include <vm/vm_pageout.h>
111
112#include <machine/cpu.h>
113
114#include <sys/thread2.h>
115
116#define arysize(ary) (sizeof(ary)/sizeof((ary)[0]))
117
118/*
119 * Fixed globals (not per-cpu)
120 */
121static int ZoneSize;
122static int ZonePageCount;
123static int ZonePageLimit;
124static int ZoneMask;
125static struct malloc_type *kmemstatistics;
126static struct kmemusage *kmemusage;
127static int32_t weirdary[16];
128
129static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
130static void kmem_slab_free(void *ptr, vm_size_t bytes);
131
132/*
133 * Misc constants. Note that allocations that are exact multiples of
134 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
135 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
136 */
137#define MIN_CHUNK_SIZE 8 /* in bytes */
138#define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
139#define ZONE_RELS_THRESH 2 /* threshold number of zones */
140#define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
141
142/*
143 * The WEIRD_ADDR is used as known text to copy into free objects to
144 * try to create deterministic failure cases if the data is accessed after
145 * free.
146 */
147#define WEIRD_ADDR 0xdeadc0de
148#define MAX_COPY sizeof(weirdary)
149#define ZERO_LENGTH_PTR ((void *)-8)
150
151/*
152 * Misc global malloc buckets
153 */
154
155MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
156MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
157MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
158
159MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
160MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
161
162/*
163 * Initialize the slab memory allocator. We have to choose a zone size based
164 * on available physical memory. We choose a zone side which is approximately
165 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
166 * 128K. The zone size is limited to the bounds set in slaballoc.h
167 * (typically 32K min, 128K max).
168 */
169static void kmeminit(void *dummy);
170
171SYSINIT(kmem, SI_SUB_KMEM, SI_ORDER_FIRST, kmeminit, NULL)
172
173static void
174kmeminit(void *dummy)
175{
176 vm_poff_t limsize;
177 int usesize;
178 int i;
179 vm_pindex_t npg;
180
181 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
182 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS)
183 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
184
185 usesize = (int)(limsize / 1024); /* convert to KB */
186
187 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
188 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
189 ZoneSize <<= 1;
190 ZoneMask = ZoneSize - 1;
191 ZonePageLimit = PAGE_SIZE * 4;
192 ZonePageCount = ZoneSize / PAGE_SIZE;
193
194 npg = (VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE;
195 kmemusage = kmem_slab_alloc(npg * sizeof(struct kmemusage), PAGE_SIZE, M_ZERO);
196
197 for (i = 0; i < arysize(weirdary); ++i)
198 weirdary[i] = WEIRD_ADDR;
199
200 if (bootverbose)
201 printf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
202}
203
204/*
205 * Initialize a malloc type tracking structure. NOTE! counters and such
206 * need to be made per-cpu (maybe with a MAXCPU array).
207 */
208void
209malloc_init(void *data)
210{
211 struct malloc_type *type = data;
212 vm_poff_t limsize;
213
214 if (type->ks_magic != M_MAGIC)
215 panic("malloc type lacks magic");
216
217 if (type->ks_limit != 0)
218 return;
219
220 if (vmstats.v_page_count == 0)
221 panic("malloc_init not allowed before vm init");
222
223 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
224 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS)
225 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
226 type->ks_limit = limsize / 10;
227
228 type->ks_next = kmemstatistics;
229 kmemstatistics = type;
230}
231
232void
233malloc_uninit(void *data)
234{
235 struct malloc_type *type = data;
236 struct malloc_type *t;
237
238 if (type->ks_magic != M_MAGIC)
239 panic("malloc type lacks magic");
240
241 if (vmstats.v_page_count == 0)
242 panic("malloc_uninit not allowed before vm init");
243
244 if (type->ks_limit == 0)
245 panic("malloc_uninit on uninitialized type");
246
247#ifdef INVARIANTS
248 if (type->ks_memuse != 0) {
249 printf("malloc_uninit: %ld bytes of '%s' still allocated\n",
250 type->ks_memuse, type->ks_shortdesc);
251 }
252#endif
253 if (type == kmemstatistics) {
254 kmemstatistics = type->ks_next;
255 } else {
256 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
257 if (t->ks_next == type) {
258 t->ks_next = type->ks_next;
259 break;
260 }
261 }
262 }
263 type->ks_next = NULL;
264 type->ks_limit = 0;
265}
266
267/*
268 * Calculate the zone index for the allocation request size and set the
269 * allocation request size to that particular zone's chunk size.
270 */
271static __inline int
272zoneindex(unsigned long *bytes)
273{
274 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
275 if (n < 128) {
276 *bytes = n = (n + 7) & ~7;
277 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
278 }
279 if (n < 256) {
280 *bytes = n = (n + 15) & ~15;
281 return(n / 16 + 7);
282 }
283 if (n < 8192) {
284 if (n < 512) {
285 *bytes = n = (n + 31) & ~31;
286 return(n / 32 + 15);
287 }
288 if (n < 1024) {
289 *bytes = n = (n + 63) & ~63;
290 return(n / 64 + 23);
291 }
292 if (n < 2048) {
293 *bytes = n = (n + 127) & ~127;
294 return(n / 128 + 31);
295 }
296 if (n < 4096) {
297 *bytes = n = (n + 255) & ~255;
298 return(n / 256 + 39);
299 }
300 *bytes = n = (n + 511) & ~511;
301 return(n / 512 + 47);
302 }
303#if ZALLOC_ZONE_LIMIT > 8192
304 if (n < 16384) {
305 *bytes = n = (n + 1023) & ~1023;
306 return(n / 1024 + 55);
307 }
308#endif
309#if ZALLOC_ZONE_LIMIT > 16384
310 if (n < 32768) {
311 *bytes = n = (n + 2047) & ~2047;
312 return(n / 2048 + 63);
313 }
314#endif
315 panic("Unexpected byte count %d", n);
316 return(0);
317}
318
319/*
320 * malloc() (SLAB ALLOCATOR)
321 *
322 * Allocate memory via the slab allocator. If the request is too large,
323 * or if it page-aligned beyond a certain size, we fall back to the
324 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
325 * &SlabMisc if you don't care.
326 *
327 * M_NOWAIT - return NULL instead of blocking.
328 * M_ZERO - zero the returned memory.
329 * M_USE_RESERVE - allocate out of the system reserve if necessary
330 */
331void *
332malloc(unsigned long size, struct malloc_type *type, int flags)
333{
334 SLZone *z;
335 SLChunk *chunk;
336 SLGlobalData *slgd;
337 int zi;
338
339 slgd = &mycpu->gd_slab;
340
341 /*
342 * XXX silly to have this in the critical path.
343 */
344 if (type->ks_limit == 0) {
345 crit_enter();
346 if (type->ks_limit == 0)
347 malloc_init(type);
348 crit_exit();
349 }
350 ++type->ks_calls;
351
352 /*
353 * Handle the case where the limit is reached. Panic if can't return
354 * NULL. XXX the original malloc code looped, but this tended to
355 * simply deadlock the computer.
356 */
357 while (type->ks_memuse >= type->ks_limit) {
358 if (flags & (M_NOWAIT|M_NULLOK))
359 return(NULL);
360 panic("%s: malloc limit exceeded", type->ks_shortdesc);
361 }
362
363 /*
364 * Handle the degenerate size == 0 case. Yes, this does happen.
365 * Return a special pointer. This is to maintain compatibility with
366 * the original malloc implementation. Certain devices, such as the
367 * adaptec driver, not only allocate 0 bytes, they check for NULL and
368 * also realloc() later on. Joy.
369 */
370 if (size == 0)
371 return(ZERO_LENGTH_PTR);
372
373 /*
374 * Handle large allocations directly. There should not be very many of
375 * these so performance is not a big issue.
376 *
377 * Guarentee page alignment for allocations in multiples of PAGE_SIZE
378 */
379 if (size >= ZALLOC_ZONE_LIMIT || (size & PAGE_MASK) == 0) {
380 struct kmemusage *kup;
381
382 size = round_page(size);
383 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
384 if (chunk == NULL)
385 return(NULL);
386 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
387 kup = btokup(chunk);
388 kup->ku_pagecnt = size / PAGE_SIZE;
389 crit_enter();
390 goto done;
391 }
392
393 /*
394 * Attempt to allocate out of an existing zone. First try the free list,
395 * then allocate out of unallocated space. If we find a good zone move
396 * it to the head of the list so later allocations find it quickly
397 * (we might have thousands of zones in the list).
398 *
399 * Note: zoneindex() will panic of size is too large.
400 */
401 zi = zoneindex(&size);
402 KKASSERT(zi < NZONES);
403 crit_enter();
404 if ((z = slgd->ZoneAry[zi]) != NULL) {
405 KKASSERT(z->z_NFree > 0);
406
407 /*
408 * Remove us from the ZoneAry[] when we become empty
409 */
410 if (--z->z_NFree == 0) {
411 slgd->ZoneAry[zi] = z->z_Next;
412 z->z_Next = NULL;
413 }
414
415 /*
416 * Locate a chunk in a free page. This attempts to localize
417 * reallocations into earlier pages without us having to sort
418 * the chunk list. A chunk may still overlap a page boundary.
419 */
420 while (z->z_FirstFreePg < ZonePageCount) {
421 if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) {
422#ifdef DIAGNOSTIC
423 /*
424 * Diagnostic: c_Next is not total garbage.
425 */
426 KKASSERT(chunk->c_Next == NULL ||
427 ((intptr_t)chunk->c_Next & IN_SAME_PAGE_MASK) ==
428 ((intptr_t)chunk & IN_SAME_PAGE_MASK));
429#endif
6ab8e1da
MD
430#ifdef INVARIANTS
431 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
a108bf71 432 panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount);
6ab8e1da 433 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
a108bf71 434 panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount);
6ab8e1da 435#endif
a108bf71
MD
436 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next;
437 goto done;
438 }
439 ++z->z_FirstFreePg;
440 }
441
442 /*
1c5ca4f3
MD
443 * No chunks are available but NFree said we had some memory, so
444 * it must be available in the never-before-used-memory area
445 * governed by UIndex. The consequences are very serious if our zone
446 * got corrupted so we use an explicit panic rather then a KASSERT.
a108bf71 447 */
1c5ca4f3
MD
448 if (z->z_UIndex + 1 != z->z_NMax)
449 z->z_UIndex = z->z_UIndex + 1;
450 else
451 z->z_UIndex = 0;
452 if (z->z_UIndex == z->z_UEndIndex)
453 panic("slaballoc: corrupted zone");
454 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
6ab8e1da
MD
455 if ((z->z_Flags & SLZF_UNOTZEROD) == 0)
456 flags &= ~M_ZERO;
a108bf71
MD
457 goto done;
458 }
459
460 /*
461 * If all zones are exhausted we need to allocate a new zone for this
462 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
6ab8e1da
MD
463 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
464 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
465 * we do not pre-zero it because we do not want to mess up the L1 cache.
a108bf71
MD
466 *
467 * At least one subsystem, the tty code (see CROUND) expects power-of-2
468 * allocations to be power-of-2 aligned. We maintain compatibility by
469 * adjusting the base offset below.
470 */
471 {
472 int off;
473
474 if ((z = slgd->FreeZones) != NULL) {
475 slgd->FreeZones = z->z_Next;
476 --slgd->NFreeZones;
477 bzero(z, sizeof(SLZone));
6ab8e1da 478 z->z_Flags |= SLZF_UNOTZEROD;
a108bf71
MD
479 } else {
480 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
481 if (z == NULL)
482 goto fail;
483 }
484
485 /*
486 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
487 * Otherwise just 8-byte align the data.
488 */
489 if ((size | (size - 1)) + 1 == (size << 1))
490 off = (sizeof(SLZone) + size - 1) & ~(size - 1);
491 else
492 off = (sizeof(SLZone) + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
493 z->z_Magic = ZALLOC_SLAB_MAGIC;
494 z->z_ZoneIndex = zi;
495 z->z_NMax = (ZoneSize - off) / size;
496 z->z_NFree = z->z_NMax - 1;
1c5ca4f3
MD
497 z->z_BasePtr = (char *)z + off;
498 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
a108bf71
MD
499 z->z_ChunkSize = size;
500 z->z_FirstFreePg = ZonePageCount;
501 z->z_Cpu = mycpu->gd_cpuid;
1c5ca4f3 502 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
a108bf71
MD
503 z->z_Next = slgd->ZoneAry[zi];
504 slgd->ZoneAry[zi] = z;
6ab8e1da
MD
505 if ((z->z_Flags & SLZF_UNOTZEROD) == 0)
506 flags &= ~M_ZERO; /* already zero'd */
1c5ca4f3
MD
507
508 /*
509 * Slide the base index for initial allocations out of the next
510 * zone we create so we do not over-weight the lower part of the
511 * cpu memory caches.
512 */
513 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
514 & (ZALLOC_MAX_ZONE_SIZE - 1);
a108bf71
MD
515 }
516done:
517 crit_exit();
518 if (flags & M_ZERO)
519 bzero(chunk, size);
520 ++type->ks_inuse;
521 type->ks_memuse += size;
522 return(chunk);
523fail:
524 crit_exit();
525 return(NULL);
526}
527
528void *
529realloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
530{
531 SLZone *z;
532 void *nptr;
533 unsigned long osize;
534
535 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
536 return(malloc(size, type, flags));
537 if (size == 0) {
538 free(ptr, type);
539 return(NULL);
540 }
541
542 /*
543 * Handle oversized allocations. XXX we really should require that a
544 * size be passed to free() instead of this nonsense.
545 */
546 {
547 struct kmemusage *kup;
548
549 kup = btokup(ptr);
550 if (kup->ku_pagecnt) {
551 osize = kup->ku_pagecnt << PAGE_SHIFT;
552 if (osize == round_page(size))
553 return(ptr);
554 if ((nptr = malloc(size, type, flags)) == NULL)
555 return(NULL);
556 bcopy(ptr, nptr, min(size, osize));
557 free(ptr, type);
558 return(nptr);
559 }
560 }
561
562 /*
563 * Get the original allocation's zone. If the new request winds up
564 * using the same chunk size we do not have to do anything.
565 */
566 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
567 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
568
569 zoneindex(&size);
570 if (z->z_ChunkSize == size)
571 return(ptr);
572
573 /*
574 * Allocate memory for the new request size. Note that zoneindex has
575 * already adjusted the request size to the appropriate chunk size, which
576 * should optimize our bcopy(). Then copy and return the new pointer.
577 */
578 if ((nptr = malloc(size, type, flags)) == NULL)
579 return(NULL);
580 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
581 free(ptr, type);
582 return(nptr);
583}
584
585/*
586 * free() (SLAB ALLOCATOR)
587 *
588 * Free the specified chunk of memory. The byte count is not strictly
589 * required but if DIAGNOSTIC is set we use it as a sanity check.
590 */
591static
592void
593free_remote(void *ptr)
594{
595 free(ptr, *(struct malloc_type **)ptr);
596}
597
598void
599free(void *ptr, struct malloc_type *type)
600{
601 SLZone *z;
602 SLChunk *chunk;
603 SLGlobalData *slgd;
604 int pgno;
605
606 slgd = &mycpu->gd_slab;
607
608 /*
609 * Handle special 0-byte allocations
610 */
611 if (ptr == ZERO_LENGTH_PTR)
612 return;
613
614 /*
615 * Handle oversized allocations. XXX we really should require that a
616 * size be passed to free() instead of this nonsense.
617 */
618 {
619 struct kmemusage *kup;
620 unsigned long size;
621
622 kup = btokup(ptr);
623 if (kup->ku_pagecnt) {
624 size = kup->ku_pagecnt << PAGE_SHIFT;
625 kup->ku_pagecnt = 0;
626 --type->ks_inuse;
627 type->ks_memuse -= size;
628#ifdef INVARIANTS
629 KKASSERT(sizeof(weirdary) <= size);
630 bcopy(weirdary, ptr, sizeof(weirdary));
631#endif
632 kmem_slab_free(ptr, size); /* may block */
633 return;
634 }
635 }
636
637 /*
638 * Zone case. Figure out the zone based on the fact that it is
639 * ZoneSize aligned.
640 */
641 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
642 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
643
644 /*
645 * If we do not own the zone then forward the request to the
646 * cpu that does. The freeing code does not need the byte count
647 * unless DIAGNOSTIC is set.
648 */
649 if (z->z_Cpu != mycpu->gd_cpuid) {
650 *(struct malloc_type **)ptr = type;
651 lwkt_send_ipiq(z->z_Cpu, free_remote, ptr);
652 return;
653 }
654
655 if (type->ks_magic != M_MAGIC)
656 panic("free: malloc type lacks magic");
657
658 crit_enter();
659 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT;
660 chunk = ptr;
661
662#ifdef DIAGNOSTIC
663 /*
664 * Diagnostic: attempt to detect a double-free (not perfect).
665 */
666 if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) {
667 SLChunk *scan;
668 for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) {
669 if (scan == chunk)
670 panic("Double free at %p", chunk);
671 }
672 }
673#endif
674
675 /*
676 * Put weird data into the memory to detect modifications after freeing,
677 * illegal pointer use after freeing (we should fault on the odd address),
678 * and so forth. XXX needs more work, see the old malloc code.
679 */
680#ifdef INVARIANTS
681 if (z->z_ChunkSize < sizeof(weirdary))
682 bcopy(weirdary, chunk, z->z_ChunkSize);
683 else
684 bcopy(weirdary, chunk, sizeof(weirdary));
685#endif
686
687 /*
688 * Add this free non-zero'd chunk to a linked list for reuse, adjust
689 * z_FirstFreePg.
690 */
6ab8e1da
MD
691#ifdef INVARIANTS
692 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
a108bf71 693 panic("BADFREE %p\n", chunk);
a108bf71
MD
694#endif
695 chunk->c_Next = z->z_PageAry[pgno];
696 z->z_PageAry[pgno] = chunk;
6ab8e1da
MD
697#ifdef INVARIANTS
698 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
a108bf71 699 panic("BADFREE2");
6ab8e1da 700#endif
a108bf71
MD
701 if (z->z_FirstFreePg > pgno)
702 z->z_FirstFreePg = pgno;
703
704 /*
705 * Bump the number of free chunks. If it becomes non-zero the zone
706 * must be added back onto the appropriate list.
707 */
708 if (z->z_NFree++ == 0) {
709 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
710 slgd->ZoneAry[z->z_ZoneIndex] = z;
711 }
712
713 --type->ks_inuse;
714 type->ks_memuse -= z->z_ChunkSize;
715
716 /*
717 * If the zone becomes totally free, and there are other zones we
718 * can allocate from, move this zone to the FreeZones list. Implement
719 * hysteresis on the FreeZones list to improve performance.
720 *
721 * XXX try not to block on the kernel_map lock.
722 */
723 if (z->z_NFree == z->z_NMax &&
724 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z)
725 ) {
726 SLZone **pz;
727
728 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
729 ;
730 *pz = z->z_Next;
731 z->z_Magic = -1;
732 if (slgd->NFreeZones == ZONE_RELS_THRESH &&
733 lockstatus(&kernel_map->lock, NULL) == 0) {
734 SLZone *oz;
735
736 z->z_Next = slgd->FreeZones->z_Next;
737 oz = slgd->FreeZones;
738 slgd->FreeZones = z;
739 kmem_slab_free(oz, ZoneSize); /* may block */
740 } else {
741 z->z_Next = slgd->FreeZones;
742 slgd->FreeZones = z;
743 ++slgd->NFreeZones;
744 }
745 }
746 crit_exit();
747}
748
749/*
750 * kmem_slab_alloc()
751 *
752 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
753 * specified alignment. M_* flags are expected in the flags field.
754 *
755 * Alignment must be a multiple of PAGE_SIZE.
756 *
757 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
758 * but when we move zalloc() over to use this function as its backend
759 * we will have to switch to kreserve/krelease and call reserve(0)
760 * after the new space is made available.
761 */
762static void *
763kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
764{
765 vm_size_t i;
766 vm_offset_t addr;
767 vm_offset_t offset;
768 int count;
769 vm_map_t map = kernel_map;
770
771 size = round_page(size);
772 addr = vm_map_min(map);
773
774 /*
775 * Reserve properly aligned space from kernel_map
776 */
777 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
778 crit_enter();
779 vm_map_lock(map);
780 if (vm_map_findspace(map, vm_map_min(map), size, align, &addr)) {
781 vm_map_unlock(map);
782 if ((flags & (M_NOWAIT|M_NULLOK)) == 0)
783 panic("kmem_slab_alloc(): kernel_map ran out of space!");
784 crit_exit();
785 vm_map_entry_release(count);
786 return(NULL);
787 }
788 offset = addr - VM_MIN_KERNEL_ADDRESS;
789 vm_object_reference(kernel_object);
790 vm_map_insert(map, &count,
791 kernel_object, offset, addr, addr + size,
792 VM_PROT_ALL, VM_PROT_ALL, 0);
793
794 /*
795 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
796 */
797 for (i = 0; i < size; i += PAGE_SIZE) {
798 vm_page_t m;
799 vm_pindex_t idx = OFF_TO_IDX(offset + i);
800 int zero = (flags & M_ZERO) ? VM_ALLOC_ZERO : 0;
801
802 if ((flags & (M_NOWAIT|M_USE_RESERVE)) == M_NOWAIT)
803 m = vm_page_alloc(kernel_object, idx, VM_ALLOC_INTERRUPT|zero);
804 else
805 m = vm_page_alloc(kernel_object, idx, VM_ALLOC_SYSTEM|zero);
806 if (m == NULL) {
807 if ((flags & M_NOWAIT) == 0) {
808 vm_map_unlock(map);
809 vm_wait();
810 vm_map_lock(map);
811 i -= PAGE_SIZE; /* retry */
812 continue;
813 }
814 while (i != 0) {
815 i -= PAGE_SIZE;
816 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
817 vm_page_free(m);
818 }
819 vm_map_delete(map, addr, addr + size, &count);
820 vm_map_unlock(map);
821 crit_exit();
822 vm_map_entry_release(count);
823 return(NULL);
824 }
825 }
826
827 /*
828 * Mark the map entry as non-pageable using a routine that allows us to
829 * populate the underlying pages.
830 */
831 vm_map_set_wired_quick(map, addr, size, &count);
832 crit_exit();
833
834 /*
835 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
836 */
837 for (i = 0; i < size; i += PAGE_SIZE) {
838 vm_page_t m;
839
840 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
841 m->valid = VM_PAGE_BITS_ALL;
842 vm_page_wire(m);
843 vm_page_wakeup(m);
844 pmap_enter(kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
845 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
846 bzero((char *)addr + i, PAGE_SIZE);
847 vm_page_flag_clear(m, PG_ZERO);
848 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_REFERENCED);
849 }
850 vm_map_unlock(map);
851 vm_map_entry_release(count);
852 return((void *)addr);
853}
854
855static void
856kmem_slab_free(void *ptr, vm_size_t size)
857{
858 crit_enter();
859 vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
860 crit_exit();
861}
862
863#endif