Fix a uni-processor bug with the last commit... we weren't rescheduling on
[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 *
3ffb0453 28 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.8 2003/10/02 22:29:15 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 *
46a3f46d 71 * Allocations >= ZoneLimit go directly to kmem.
a108bf71
MD
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
074dcfe8 86#if !defined(NO_SLAB_ALLOCATOR)
a108bf71 87
074dcfe8
MD
88#if defined(USE_KMEM_MAP)
89#error "If you define USE_KMEM_MAP you must also define NO_SLAB_ALLOCATOR"
a108bf71
MD
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;
46a3f46d 122static int ZoneLimit;
a108bf71
MD
123static int ZonePageCount;
124static int ZonePageLimit;
125static int ZoneMask;
126static struct malloc_type *kmemstatistics;
127static struct kmemusage *kmemusage;
128static int32_t weirdary[16];
129
130static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
131static void kmem_slab_free(void *ptr, vm_size_t bytes);
132
133/*
134 * Misc constants. Note that allocations that are exact multiples of
135 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
136 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
137 */
138#define MIN_CHUNK_SIZE 8 /* in bytes */
139#define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
140#define ZONE_RELS_THRESH 2 /* threshold number of zones */
141#define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
142
143/*
144 * The WEIRD_ADDR is used as known text to copy into free objects to
145 * try to create deterministic failure cases if the data is accessed after
146 * free.
147 */
148#define WEIRD_ADDR 0xdeadc0de
149#define MAX_COPY sizeof(weirdary)
150#define ZERO_LENGTH_PTR ((void *)-8)
151
152/*
153 * Misc global malloc buckets
154 */
155
156MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
157MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
158MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
159
160MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
161MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
162
163/*
164 * Initialize the slab memory allocator. We have to choose a zone size based
165 * on available physical memory. We choose a zone side which is approximately
166 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
167 * 128K. The zone size is limited to the bounds set in slaballoc.h
168 * (typically 32K min, 128K max).
169 */
170static void kmeminit(void *dummy);
171
172SYSINIT(kmem, SI_SUB_KMEM, SI_ORDER_FIRST, kmeminit, NULL)
173
174static void
175kmeminit(void *dummy)
176{
177 vm_poff_t limsize;
178 int usesize;
179 int i;
180 vm_pindex_t npg;
181
182 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
183 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS)
184 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
185
186 usesize = (int)(limsize / 1024); /* convert to KB */
187
188 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
189 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
190 ZoneSize <<= 1;
46a3f46d
MD
191 ZoneLimit = ZoneSize / 4;
192 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
193 ZoneLimit = ZALLOC_ZONE_LIMIT;
a108bf71
MD
194 ZoneMask = ZoneSize - 1;
195 ZonePageLimit = PAGE_SIZE * 4;
196 ZonePageCount = ZoneSize / PAGE_SIZE;
197
198 npg = (VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE;
199 kmemusage = kmem_slab_alloc(npg * sizeof(struct kmemusage), PAGE_SIZE, M_ZERO);
200
201 for (i = 0; i < arysize(weirdary); ++i)
202 weirdary[i] = WEIRD_ADDR;
203
204 if (bootverbose)
205 printf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
206}
207
208/*
209 * Initialize a malloc type tracking structure. NOTE! counters and such
210 * need to be made per-cpu (maybe with a MAXCPU array).
211 */
212void
213malloc_init(void *data)
214{
215 struct malloc_type *type = data;
216 vm_poff_t limsize;
217
218 if (type->ks_magic != M_MAGIC)
219 panic("malloc type lacks magic");
220
221 if (type->ks_limit != 0)
222 return;
223
224 if (vmstats.v_page_count == 0)
225 panic("malloc_init not allowed before vm init");
226
227 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
228 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS)
229 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
230 type->ks_limit = limsize / 10;
231
232 type->ks_next = kmemstatistics;
233 kmemstatistics = type;
234}
235
236void
237malloc_uninit(void *data)
238{
239 struct malloc_type *type = data;
240 struct malloc_type *t;
241
242 if (type->ks_magic != M_MAGIC)
243 panic("malloc type lacks magic");
244
245 if (vmstats.v_page_count == 0)
246 panic("malloc_uninit not allowed before vm init");
247
248 if (type->ks_limit == 0)
249 panic("malloc_uninit on uninitialized type");
250
251#ifdef INVARIANTS
252 if (type->ks_memuse != 0) {
253 printf("malloc_uninit: %ld bytes of '%s' still allocated\n",
254 type->ks_memuse, type->ks_shortdesc);
255 }
256#endif
257 if (type == kmemstatistics) {
258 kmemstatistics = type->ks_next;
259 } else {
260 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
261 if (t->ks_next == type) {
262 t->ks_next = type->ks_next;
263 break;
264 }
265 }
266 }
267 type->ks_next = NULL;
268 type->ks_limit = 0;
269}
270
271/*
272 * Calculate the zone index for the allocation request size and set the
273 * allocation request size to that particular zone's chunk size.
274 */
275static __inline int
276zoneindex(unsigned long *bytes)
277{
278 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
279 if (n < 128) {
280 *bytes = n = (n + 7) & ~7;
281 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
282 }
283 if (n < 256) {
284 *bytes = n = (n + 15) & ~15;
285 return(n / 16 + 7);
286 }
287 if (n < 8192) {
288 if (n < 512) {
289 *bytes = n = (n + 31) & ~31;
290 return(n / 32 + 15);
291 }
292 if (n < 1024) {
293 *bytes = n = (n + 63) & ~63;
294 return(n / 64 + 23);
295 }
296 if (n < 2048) {
297 *bytes = n = (n + 127) & ~127;
298 return(n / 128 + 31);
299 }
300 if (n < 4096) {
301 *bytes = n = (n + 255) & ~255;
302 return(n / 256 + 39);
303 }
304 *bytes = n = (n + 511) & ~511;
305 return(n / 512 + 47);
306 }
307#if ZALLOC_ZONE_LIMIT > 8192
308 if (n < 16384) {
309 *bytes = n = (n + 1023) & ~1023;
310 return(n / 1024 + 55);
311 }
312#endif
313#if ZALLOC_ZONE_LIMIT > 16384
314 if (n < 32768) {
315 *bytes = n = (n + 2047) & ~2047;
316 return(n / 2048 + 63);
317 }
318#endif
319 panic("Unexpected byte count %d", n);
320 return(0);
321}
322
323/*
324 * malloc() (SLAB ALLOCATOR)
325 *
326 * Allocate memory via the slab allocator. If the request is too large,
327 * or if it page-aligned beyond a certain size, we fall back to the
328 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
329 * &SlabMisc if you don't care.
330 *
331 * M_NOWAIT - return NULL instead of blocking.
332 * M_ZERO - zero the returned memory.
333 * M_USE_RESERVE - allocate out of the system reserve if necessary
334 */
335void *
336malloc(unsigned long size, struct malloc_type *type, int flags)
337{
338 SLZone *z;
339 SLChunk *chunk;
340 SLGlobalData *slgd;
341 int zi;
342
343 slgd = &mycpu->gd_slab;
344
345 /*
346 * XXX silly to have this in the critical path.
347 */
348 if (type->ks_limit == 0) {
349 crit_enter();
350 if (type->ks_limit == 0)
351 malloc_init(type);
352 crit_exit();
353 }
354 ++type->ks_calls;
355
356 /*
357 * Handle the case where the limit is reached. Panic if can't return
358 * NULL. XXX the original malloc code looped, but this tended to
359 * simply deadlock the computer.
360 */
361 while (type->ks_memuse >= type->ks_limit) {
362 if (flags & (M_NOWAIT|M_NULLOK))
363 return(NULL);
364 panic("%s: malloc limit exceeded", type->ks_shortdesc);
365 }
366
367 /*
368 * Handle the degenerate size == 0 case. Yes, this does happen.
369 * Return a special pointer. This is to maintain compatibility with
370 * the original malloc implementation. Certain devices, such as the
371 * adaptec driver, not only allocate 0 bytes, they check for NULL and
372 * also realloc() later on. Joy.
373 */
374 if (size == 0)
375 return(ZERO_LENGTH_PTR);
376
a7cf0021
MD
377 /*
378 * Handle hysteresis from prior frees here in malloc(). We cannot
379 * safely manipulate the kernel_map in free() due to free() possibly
380 * being called via an IPI message or from sensitive interrupt code.
381 */
382 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_NOWAIT) == 0) {
46a3f46d
MD
383 crit_enter();
384 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */
385 z = slgd->FreeZones;
386 slgd->FreeZones = z->z_Next;
387 --slgd->NFreeZones;
388 kmem_slab_free(z, ZoneSize); /* may block */
389 }
390 crit_exit();
391 }
392 /*
393 * XXX handle oversized frees that were queued from free().
394 */
395 while (slgd->FreeOvZones && (flags & M_NOWAIT) == 0) {
396 crit_enter();
397 if ((z = slgd->FreeOvZones) != NULL) {
398 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
399 slgd->FreeOvZones = z->z_Next;
400 kmem_slab_free(z, z->z_ChunkSize); /* may block */
401 }
402 crit_exit();
a7cf0021
MD
403 }
404
a108bf71
MD
405 /*
406 * Handle large allocations directly. There should not be very many of
407 * these so performance is not a big issue.
408 *
409 * Guarentee page alignment for allocations in multiples of PAGE_SIZE
410 */
46a3f46d 411 if (size >= ZoneLimit || (size & PAGE_MASK) == 0) {
a108bf71
MD
412 struct kmemusage *kup;
413
414 size = round_page(size);
415 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
416 if (chunk == NULL)
417 return(NULL);
418 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
419 kup = btokup(chunk);
420 kup->ku_pagecnt = size / PAGE_SIZE;
421 crit_enter();
422 goto done;
423 }
424
425 /*
426 * Attempt to allocate out of an existing zone. First try the free list,
427 * then allocate out of unallocated space. If we find a good zone move
428 * it to the head of the list so later allocations find it quickly
429 * (we might have thousands of zones in the list).
430 *
431 * Note: zoneindex() will panic of size is too large.
432 */
433 zi = zoneindex(&size);
434 KKASSERT(zi < NZONES);
435 crit_enter();
436 if ((z = slgd->ZoneAry[zi]) != NULL) {
437 KKASSERT(z->z_NFree > 0);
438
439 /*
440 * Remove us from the ZoneAry[] when we become empty
441 */
442 if (--z->z_NFree == 0) {
443 slgd->ZoneAry[zi] = z->z_Next;
444 z->z_Next = NULL;
445 }
446
447 /*
448 * Locate a chunk in a free page. This attempts to localize
449 * reallocations into earlier pages without us having to sort
450 * the chunk list. A chunk may still overlap a page boundary.
451 */
452 while (z->z_FirstFreePg < ZonePageCount) {
453 if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) {
454#ifdef DIAGNOSTIC
455 /*
456 * Diagnostic: c_Next is not total garbage.
457 */
458 KKASSERT(chunk->c_Next == NULL ||
459 ((intptr_t)chunk->c_Next & IN_SAME_PAGE_MASK) ==
460 ((intptr_t)chunk & IN_SAME_PAGE_MASK));
461#endif
6ab8e1da
MD
462#ifdef INVARIANTS
463 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
a108bf71 464 panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount);
6ab8e1da 465 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
a108bf71 466 panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount);
6ab8e1da 467#endif
a108bf71
MD
468 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next;
469 goto done;
470 }
471 ++z->z_FirstFreePg;
472 }
473
474 /*
1c5ca4f3
MD
475 * No chunks are available but NFree said we had some memory, so
476 * it must be available in the never-before-used-memory area
477 * governed by UIndex. The consequences are very serious if our zone
478 * got corrupted so we use an explicit panic rather then a KASSERT.
a108bf71 479 */
1c5ca4f3
MD
480 if (z->z_UIndex + 1 != z->z_NMax)
481 z->z_UIndex = z->z_UIndex + 1;
482 else
483 z->z_UIndex = 0;
484 if (z->z_UIndex == z->z_UEndIndex)
485 panic("slaballoc: corrupted zone");
486 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
6ab8e1da
MD
487 if ((z->z_Flags & SLZF_UNOTZEROD) == 0)
488 flags &= ~M_ZERO;
a108bf71
MD
489 goto done;
490 }
491
492 /*
493 * If all zones are exhausted we need to allocate a new zone for this
494 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
6ab8e1da
MD
495 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
496 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
497 * we do not pre-zero it because we do not want to mess up the L1 cache.
a108bf71
MD
498 *
499 * At least one subsystem, the tty code (see CROUND) expects power-of-2
500 * allocations to be power-of-2 aligned. We maintain compatibility by
501 * adjusting the base offset below.
502 */
503 {
504 int off;
505
506 if ((z = slgd->FreeZones) != NULL) {
507 slgd->FreeZones = z->z_Next;
508 --slgd->NFreeZones;
509 bzero(z, sizeof(SLZone));
6ab8e1da 510 z->z_Flags |= SLZF_UNOTZEROD;
a108bf71
MD
511 } else {
512 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
513 if (z == NULL)
514 goto fail;
515 }
516
517 /*
518 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
519 * Otherwise just 8-byte align the data.
520 */
521 if ((size | (size - 1)) + 1 == (size << 1))
522 off = (sizeof(SLZone) + size - 1) & ~(size - 1);
523 else
524 off = (sizeof(SLZone) + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
525 z->z_Magic = ZALLOC_SLAB_MAGIC;
526 z->z_ZoneIndex = zi;
527 z->z_NMax = (ZoneSize - off) / size;
528 z->z_NFree = z->z_NMax - 1;
1c5ca4f3
MD
529 z->z_BasePtr = (char *)z + off;
530 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
a108bf71
MD
531 z->z_ChunkSize = size;
532 z->z_FirstFreePg = ZonePageCount;
533 z->z_Cpu = mycpu->gd_cpuid;
1c5ca4f3 534 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
a108bf71
MD
535 z->z_Next = slgd->ZoneAry[zi];
536 slgd->ZoneAry[zi] = z;
6ab8e1da
MD
537 if ((z->z_Flags & SLZF_UNOTZEROD) == 0)
538 flags &= ~M_ZERO; /* already zero'd */
1c5ca4f3
MD
539
540 /*
541 * Slide the base index for initial allocations out of the next
542 * zone we create so we do not over-weight the lower part of the
543 * cpu memory caches.
544 */
545 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
546 & (ZALLOC_MAX_ZONE_SIZE - 1);
a108bf71
MD
547 }
548done:
549 crit_exit();
550 if (flags & M_ZERO)
551 bzero(chunk, size);
552 ++type->ks_inuse;
553 type->ks_memuse += size;
554 return(chunk);
555fail:
556 crit_exit();
557 return(NULL);
558}
559
560void *
561realloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
562{
563 SLZone *z;
564 void *nptr;
565 unsigned long osize;
566
567 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
568 return(malloc(size, type, flags));
569 if (size == 0) {
570 free(ptr, type);
571 return(NULL);
572 }
573
574 /*
575 * Handle oversized allocations. XXX we really should require that a
576 * size be passed to free() instead of this nonsense.
577 */
578 {
579 struct kmemusage *kup;
580
581 kup = btokup(ptr);
582 if (kup->ku_pagecnt) {
583 osize = kup->ku_pagecnt << PAGE_SHIFT;
584 if (osize == round_page(size))
585 return(ptr);
586 if ((nptr = malloc(size, type, flags)) == NULL)
587 return(NULL);
588 bcopy(ptr, nptr, min(size, osize));
589 free(ptr, type);
590 return(nptr);
591 }
592 }
593
594 /*
595 * Get the original allocation's zone. If the new request winds up
596 * using the same chunk size we do not have to do anything.
597 */
598 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
599 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
600
601 zoneindex(&size);
602 if (z->z_ChunkSize == size)
603 return(ptr);
604
605 /*
606 * Allocate memory for the new request size. Note that zoneindex has
607 * already adjusted the request size to the appropriate chunk size, which
608 * should optimize our bcopy(). Then copy and return the new pointer.
609 */
610 if ((nptr = malloc(size, type, flags)) == NULL)
611 return(NULL);
612 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
613 free(ptr, type);
614 return(nptr);
615}
616
617/*
618 * free() (SLAB ALLOCATOR)
619 *
620 * Free the specified chunk of memory. The byte count is not strictly
621 * required but if DIAGNOSTIC is set we use it as a sanity check.
622 */
623static
624void
625free_remote(void *ptr)
626{
627 free(ptr, *(struct malloc_type **)ptr);
628}
629
630void
631free(void *ptr, struct malloc_type *type)
632{
633 SLZone *z;
634 SLChunk *chunk;
635 SLGlobalData *slgd;
636 int pgno;
637
638 slgd = &mycpu->gd_slab;
639
640 /*
641 * Handle special 0-byte allocations
642 */
643 if (ptr == ZERO_LENGTH_PTR)
644 return;
645
646 /*
647 * Handle oversized allocations. XXX we really should require that a
648 * size be passed to free() instead of this nonsense.
649 */
650 {
651 struct kmemusage *kup;
652 unsigned long size;
653
654 kup = btokup(ptr);
655 if (kup->ku_pagecnt) {
656 size = kup->ku_pagecnt << PAGE_SHIFT;
657 kup->ku_pagecnt = 0;
658 --type->ks_inuse;
659 type->ks_memuse -= size;
660#ifdef INVARIANTS
661 KKASSERT(sizeof(weirdary) <= size);
662 bcopy(weirdary, ptr, sizeof(weirdary));
663#endif
46a3f46d
MD
664 if (mycpu->gd_intr_nesting_level) {
665 crit_enter();
666 z = (SLZone *)ptr;
667 z->z_Magic = ZALLOC_OVSZ_MAGIC;
668 z->z_Next = slgd->FreeOvZones;
669 z->z_ChunkSize = size;
670 slgd->FreeOvZones = z;
671 crit_exit();
672 } else {
673 kmem_slab_free(ptr, size); /* may block */
674 }
a108bf71
MD
675 return;
676 }
677 }
678
679 /*
680 * Zone case. Figure out the zone based on the fact that it is
681 * ZoneSize aligned.
682 */
683 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
684 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
685
686 /*
687 * If we do not own the zone then forward the request to the
688 * cpu that does. The freeing code does not need the byte count
689 * unless DIAGNOSTIC is set.
690 */
691 if (z->z_Cpu != mycpu->gd_cpuid) {
692 *(struct malloc_type **)ptr = type;
693 lwkt_send_ipiq(z->z_Cpu, free_remote, ptr);
694 return;
695 }
696
697 if (type->ks_magic != M_MAGIC)
698 panic("free: malloc type lacks magic");
699
700 crit_enter();
701 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT;
702 chunk = ptr;
703
704#ifdef DIAGNOSTIC
705 /*
706 * Diagnostic: attempt to detect a double-free (not perfect).
707 */
708 if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) {
709 SLChunk *scan;
710 for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) {
711 if (scan == chunk)
712 panic("Double free at %p", chunk);
713 }
714 }
715#endif
716
717 /*
718 * Put weird data into the memory to detect modifications after freeing,
719 * illegal pointer use after freeing (we should fault on the odd address),
720 * and so forth. XXX needs more work, see the old malloc code.
721 */
722#ifdef INVARIANTS
723 if (z->z_ChunkSize < sizeof(weirdary))
724 bcopy(weirdary, chunk, z->z_ChunkSize);
725 else
726 bcopy(weirdary, chunk, sizeof(weirdary));
727#endif
728
729 /*
730 * Add this free non-zero'd chunk to a linked list for reuse, adjust
731 * z_FirstFreePg.
732 */
6ab8e1da
MD
733#ifdef INVARIANTS
734 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
a108bf71 735 panic("BADFREE %p\n", chunk);
a108bf71
MD
736#endif
737 chunk->c_Next = z->z_PageAry[pgno];
738 z->z_PageAry[pgno] = chunk;
6ab8e1da
MD
739#ifdef INVARIANTS
740 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
a108bf71 741 panic("BADFREE2");
6ab8e1da 742#endif
a108bf71
MD
743 if (z->z_FirstFreePg > pgno)
744 z->z_FirstFreePg = pgno;
745
746 /*
747 * Bump the number of free chunks. If it becomes non-zero the zone
748 * must be added back onto the appropriate list.
749 */
750 if (z->z_NFree++ == 0) {
751 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
752 slgd->ZoneAry[z->z_ZoneIndex] = z;
753 }
754
755 --type->ks_inuse;
756 type->ks_memuse -= z->z_ChunkSize;
757
758 /*
759 * If the zone becomes totally free, and there are other zones we
a7cf0021
MD
760 * can allocate from, move this zone to the FreeZones list. Since
761 * this code can be called from an IPI callback, do *NOT* try to mess
762 * with kernel_map here. Hysteresis will be performed at malloc() time.
a108bf71
MD
763 */
764 if (z->z_NFree == z->z_NMax &&
765 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z)
766 ) {
767 SLZone **pz;
768
769 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
770 ;
771 *pz = z->z_Next;
772 z->z_Magic = -1;
a7cf0021
MD
773 z->z_Next = slgd->FreeZones;
774 slgd->FreeZones = z;
775 ++slgd->NFreeZones;
a108bf71
MD
776 }
777 crit_exit();
778}
779
780/*
781 * kmem_slab_alloc()
782 *
783 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
784 * specified alignment. M_* flags are expected in the flags field.
785 *
786 * Alignment must be a multiple of PAGE_SIZE.
787 *
788 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
789 * but when we move zalloc() over to use this function as its backend
790 * we will have to switch to kreserve/krelease and call reserve(0)
791 * after the new space is made available.
792 */
793static void *
794kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
795{
796 vm_size_t i;
797 vm_offset_t addr;
798 vm_offset_t offset;
799 int count;
800 vm_map_t map = kernel_map;
801
802 size = round_page(size);
803 addr = vm_map_min(map);
804
805 /*
806 * Reserve properly aligned space from kernel_map
807 */
808 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
809 crit_enter();
810 vm_map_lock(map);
811 if (vm_map_findspace(map, vm_map_min(map), size, align, &addr)) {
812 vm_map_unlock(map);
813 if ((flags & (M_NOWAIT|M_NULLOK)) == 0)
814 panic("kmem_slab_alloc(): kernel_map ran out of space!");
815 crit_exit();
816 vm_map_entry_release(count);
817 return(NULL);
818 }
819 offset = addr - VM_MIN_KERNEL_ADDRESS;
820 vm_object_reference(kernel_object);
821 vm_map_insert(map, &count,
822 kernel_object, offset, addr, addr + size,
823 VM_PROT_ALL, VM_PROT_ALL, 0);
824
825 /*
826 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
827 */
828 for (i = 0; i < size; i += PAGE_SIZE) {
829 vm_page_t m;
830 vm_pindex_t idx = OFF_TO_IDX(offset + i);
831 int zero = (flags & M_ZERO) ? VM_ALLOC_ZERO : 0;
832
833 if ((flags & (M_NOWAIT|M_USE_RESERVE)) == M_NOWAIT)
834 m = vm_page_alloc(kernel_object, idx, VM_ALLOC_INTERRUPT|zero);
835 else
836 m = vm_page_alloc(kernel_object, idx, VM_ALLOC_SYSTEM|zero);
837 if (m == NULL) {
838 if ((flags & M_NOWAIT) == 0) {
839 vm_map_unlock(map);
840 vm_wait();
841 vm_map_lock(map);
842 i -= PAGE_SIZE; /* retry */
843 continue;
844 }
845 while (i != 0) {
846 i -= PAGE_SIZE;
847 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
848 vm_page_free(m);
849 }
850 vm_map_delete(map, addr, addr + size, &count);
851 vm_map_unlock(map);
852 crit_exit();
853 vm_map_entry_release(count);
854 return(NULL);
855 }
856 }
857
858 /*
859 * Mark the map entry as non-pageable using a routine that allows us to
860 * populate the underlying pages.
861 */
862 vm_map_set_wired_quick(map, addr, size, &count);
863 crit_exit();
864
865 /*
866 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
867 */
868 for (i = 0; i < size; i += PAGE_SIZE) {
869 vm_page_t m;
870
871 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
872 m->valid = VM_PAGE_BITS_ALL;
873 vm_page_wire(m);
874 vm_page_wakeup(m);
875 pmap_enter(kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
876 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
877 bzero((char *)addr + i, PAGE_SIZE);
878 vm_page_flag_clear(m, PG_ZERO);
879 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_REFERENCED);
880 }
881 vm_map_unlock(map);
882 vm_map_entry_release(count);
883 return((void *)addr);
884}
885
886static void
887kmem_slab_free(void *ptr, vm_size_t size)
888{
889 crit_enter();
890 vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
891 crit_exit();
892}
893
894#endif