Add the pim(4) and multicast(4) manual pages, following
[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 *
bba6a44d 28 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.9 2003/10/18 05:48:42 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/*
bba6a44d 209 * Initialize a malloc type tracking structure.
a108bf71
MD
210 */
211void
212malloc_init(void *data)
213{
214 struct malloc_type *type = data;
215 vm_poff_t limsize;
216
217 if (type->ks_magic != M_MAGIC)
218 panic("malloc type lacks magic");
219
220 if (type->ks_limit != 0)
221 return;
222
223 if (vmstats.v_page_count == 0)
224 panic("malloc_init not allowed before vm init");
225
226 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
227 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS)
228 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
229 type->ks_limit = limsize / 10;
230
231 type->ks_next = kmemstatistics;
232 kmemstatistics = type;
233}
234
235void
236malloc_uninit(void *data)
237{
238 struct malloc_type *type = data;
239 struct malloc_type *t;
bba6a44d
MD
240#ifdef INVARIANTS
241 int i;
242#endif
a108bf71
MD
243
244 if (type->ks_magic != M_MAGIC)
245 panic("malloc type lacks magic");
246
247 if (vmstats.v_page_count == 0)
248 panic("malloc_uninit not allowed before vm init");
249
250 if (type->ks_limit == 0)
251 panic("malloc_uninit on uninitialized type");
252
253#ifdef INVARIANTS
bba6a44d
MD
254 for (i = 0; i < ncpus; ++i) {
255 if (type->ks_memuse[i] != 0) {
256 printf(
257 "malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
258 type->ks_memuse[i], type->ks_shortdesc, i);
259 }
a108bf71
MD
260 }
261#endif
262 if (type == kmemstatistics) {
263 kmemstatistics = type->ks_next;
264 } else {
265 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
266 if (t->ks_next == type) {
267 t->ks_next = type->ks_next;
268 break;
269 }
270 }
271 }
272 type->ks_next = NULL;
273 type->ks_limit = 0;
274}
275
276/*
277 * Calculate the zone index for the allocation request size and set the
278 * allocation request size to that particular zone's chunk size.
279 */
280static __inline int
281zoneindex(unsigned long *bytes)
282{
283 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
284 if (n < 128) {
285 *bytes = n = (n + 7) & ~7;
286 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
287 }
288 if (n < 256) {
289 *bytes = n = (n + 15) & ~15;
290 return(n / 16 + 7);
291 }
292 if (n < 8192) {
293 if (n < 512) {
294 *bytes = n = (n + 31) & ~31;
295 return(n / 32 + 15);
296 }
297 if (n < 1024) {
298 *bytes = n = (n + 63) & ~63;
299 return(n / 64 + 23);
300 }
301 if (n < 2048) {
302 *bytes = n = (n + 127) & ~127;
303 return(n / 128 + 31);
304 }
305 if (n < 4096) {
306 *bytes = n = (n + 255) & ~255;
307 return(n / 256 + 39);
308 }
309 *bytes = n = (n + 511) & ~511;
310 return(n / 512 + 47);
311 }
312#if ZALLOC_ZONE_LIMIT > 8192
313 if (n < 16384) {
314 *bytes = n = (n + 1023) & ~1023;
315 return(n / 1024 + 55);
316 }
317#endif
318#if ZALLOC_ZONE_LIMIT > 16384
319 if (n < 32768) {
320 *bytes = n = (n + 2047) & ~2047;
321 return(n / 2048 + 63);
322 }
323#endif
324 panic("Unexpected byte count %d", n);
325 return(0);
326}
327
328/*
329 * malloc() (SLAB ALLOCATOR)
330 *
331 * Allocate memory via the slab allocator. If the request is too large,
332 * or if it page-aligned beyond a certain size, we fall back to the
333 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
334 * &SlabMisc if you don't care.
335 *
336 * M_NOWAIT - return NULL instead of blocking.
337 * M_ZERO - zero the returned memory.
338 * M_USE_RESERVE - allocate out of the system reserve if necessary
339 */
340void *
341malloc(unsigned long size, struct malloc_type *type, int flags)
342{
343 SLZone *z;
344 SLChunk *chunk;
345 SLGlobalData *slgd;
bba6a44d 346 struct globaldata *gd;
a108bf71
MD
347 int zi;
348
bba6a44d
MD
349 gd = mycpu;
350 slgd = &gd->gd_slab;
a108bf71
MD
351
352 /*
353 * XXX silly to have this in the critical path.
354 */
355 if (type->ks_limit == 0) {
356 crit_enter();
357 if (type->ks_limit == 0)
358 malloc_init(type);
359 crit_exit();
360 }
361 ++type->ks_calls;
362
363 /*
364 * Handle the case where the limit is reached. Panic if can't return
365 * NULL. XXX the original malloc code looped, but this tended to
366 * simply deadlock the computer.
367 */
bba6a44d
MD
368 while (type->ks_loosememuse >= type->ks_limit) {
369 int i;
370 long ttl;
371
372 for (i = ttl = 0; i < ncpus; ++i)
373 ttl += type->ks_memuse[i];
374 type->ks_loosememuse = ttl;
375 if (ttl >= type->ks_limit) {
376 if (flags & (M_NOWAIT|M_NULLOK))
377 return(NULL);
378 panic("%s: malloc limit exceeded", type->ks_shortdesc);
379 }
a108bf71
MD
380 }
381
382 /*
383 * Handle the degenerate size == 0 case. Yes, this does happen.
384 * Return a special pointer. This is to maintain compatibility with
385 * the original malloc implementation. Certain devices, such as the
386 * adaptec driver, not only allocate 0 bytes, they check for NULL and
387 * also realloc() later on. Joy.
388 */
389 if (size == 0)
390 return(ZERO_LENGTH_PTR);
391
a7cf0021
MD
392 /*
393 * Handle hysteresis from prior frees here in malloc(). We cannot
394 * safely manipulate the kernel_map in free() due to free() possibly
395 * being called via an IPI message or from sensitive interrupt code.
396 */
397 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_NOWAIT) == 0) {
46a3f46d
MD
398 crit_enter();
399 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */
400 z = slgd->FreeZones;
401 slgd->FreeZones = z->z_Next;
402 --slgd->NFreeZones;
403 kmem_slab_free(z, ZoneSize); /* may block */
404 }
405 crit_exit();
406 }
407 /*
408 * XXX handle oversized frees that were queued from free().
409 */
410 while (slgd->FreeOvZones && (flags & M_NOWAIT) == 0) {
411 crit_enter();
412 if ((z = slgd->FreeOvZones) != NULL) {
413 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
414 slgd->FreeOvZones = z->z_Next;
415 kmem_slab_free(z, z->z_ChunkSize); /* may block */
416 }
417 crit_exit();
a7cf0021
MD
418 }
419
a108bf71
MD
420 /*
421 * Handle large allocations directly. There should not be very many of
422 * these so performance is not a big issue.
423 *
424 * Guarentee page alignment for allocations in multiples of PAGE_SIZE
425 */
46a3f46d 426 if (size >= ZoneLimit || (size & PAGE_MASK) == 0) {
a108bf71
MD
427 struct kmemusage *kup;
428
429 size = round_page(size);
430 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
431 if (chunk == NULL)
432 return(NULL);
433 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
434 kup = btokup(chunk);
435 kup->ku_pagecnt = size / PAGE_SIZE;
bba6a44d 436 kup->ku_cpu = gd->gd_cpuid;
a108bf71
MD
437 crit_enter();
438 goto done;
439 }
440
441 /*
442 * Attempt to allocate out of an existing zone. First try the free list,
443 * then allocate out of unallocated space. If we find a good zone move
444 * it to the head of the list so later allocations find it quickly
445 * (we might have thousands of zones in the list).
446 *
447 * Note: zoneindex() will panic of size is too large.
448 */
449 zi = zoneindex(&size);
450 KKASSERT(zi < NZONES);
451 crit_enter();
452 if ((z = slgd->ZoneAry[zi]) != NULL) {
453 KKASSERT(z->z_NFree > 0);
454
455 /*
456 * Remove us from the ZoneAry[] when we become empty
457 */
458 if (--z->z_NFree == 0) {
459 slgd->ZoneAry[zi] = z->z_Next;
460 z->z_Next = NULL;
461 }
462
463 /*
464 * Locate a chunk in a free page. This attempts to localize
465 * reallocations into earlier pages without us having to sort
466 * the chunk list. A chunk may still overlap a page boundary.
467 */
468 while (z->z_FirstFreePg < ZonePageCount) {
469 if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) {
470#ifdef DIAGNOSTIC
471 /*
472 * Diagnostic: c_Next is not total garbage.
473 */
474 KKASSERT(chunk->c_Next == NULL ||
475 ((intptr_t)chunk->c_Next & IN_SAME_PAGE_MASK) ==
476 ((intptr_t)chunk & IN_SAME_PAGE_MASK));
477#endif
6ab8e1da
MD
478#ifdef INVARIANTS
479 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
a108bf71 480 panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount);
6ab8e1da 481 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
a108bf71 482 panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount);
6ab8e1da 483#endif
a108bf71
MD
484 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next;
485 goto done;
486 }
487 ++z->z_FirstFreePg;
488 }
489
490 /*
1c5ca4f3
MD
491 * No chunks are available but NFree said we had some memory, so
492 * it must be available in the never-before-used-memory area
493 * governed by UIndex. The consequences are very serious if our zone
494 * got corrupted so we use an explicit panic rather then a KASSERT.
a108bf71 495 */
1c5ca4f3
MD
496 if (z->z_UIndex + 1 != z->z_NMax)
497 z->z_UIndex = z->z_UIndex + 1;
498 else
499 z->z_UIndex = 0;
500 if (z->z_UIndex == z->z_UEndIndex)
501 panic("slaballoc: corrupted zone");
502 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
6ab8e1da
MD
503 if ((z->z_Flags & SLZF_UNOTZEROD) == 0)
504 flags &= ~M_ZERO;
a108bf71
MD
505 goto done;
506 }
507
508 /*
509 * If all zones are exhausted we need to allocate a new zone for this
510 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
6ab8e1da
MD
511 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
512 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
513 * we do not pre-zero it because we do not want to mess up the L1 cache.
a108bf71
MD
514 *
515 * At least one subsystem, the tty code (see CROUND) expects power-of-2
516 * allocations to be power-of-2 aligned. We maintain compatibility by
517 * adjusting the base offset below.
518 */
519 {
520 int off;
521
522 if ((z = slgd->FreeZones) != NULL) {
523 slgd->FreeZones = z->z_Next;
524 --slgd->NFreeZones;
525 bzero(z, sizeof(SLZone));
6ab8e1da 526 z->z_Flags |= SLZF_UNOTZEROD;
a108bf71
MD
527 } else {
528 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
529 if (z == NULL)
530 goto fail;
531 }
532
533 /*
534 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
535 * Otherwise just 8-byte align the data.
536 */
537 if ((size | (size - 1)) + 1 == (size << 1))
538 off = (sizeof(SLZone) + size - 1) & ~(size - 1);
539 else
540 off = (sizeof(SLZone) + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
541 z->z_Magic = ZALLOC_SLAB_MAGIC;
542 z->z_ZoneIndex = zi;
543 z->z_NMax = (ZoneSize - off) / size;
544 z->z_NFree = z->z_NMax - 1;
1c5ca4f3
MD
545 z->z_BasePtr = (char *)z + off;
546 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
a108bf71
MD
547 z->z_ChunkSize = size;
548 z->z_FirstFreePg = ZonePageCount;
bba6a44d 549 z->z_Cpu = gd->gd_cpuid;
1c5ca4f3 550 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
a108bf71
MD
551 z->z_Next = slgd->ZoneAry[zi];
552 slgd->ZoneAry[zi] = z;
6ab8e1da
MD
553 if ((z->z_Flags & SLZF_UNOTZEROD) == 0)
554 flags &= ~M_ZERO; /* already zero'd */
1c5ca4f3
MD
555
556 /*
557 * Slide the base index for initial allocations out of the next
558 * zone we create so we do not over-weight the lower part of the
559 * cpu memory caches.
560 */
561 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
562 & (ZALLOC_MAX_ZONE_SIZE - 1);
a108bf71
MD
563 }
564done:
bba6a44d
MD
565 ++type->ks_inuse[gd->gd_cpuid];
566 type->ks_memuse[gd->gd_cpuid] += size;
567 type->ks_loosememuse += size;
a108bf71
MD
568 crit_exit();
569 if (flags & M_ZERO)
570 bzero(chunk, size);
bba6a44d
MD
571#ifdef INVARIANTS
572 else
573 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
574#endif
a108bf71
MD
575 return(chunk);
576fail:
577 crit_exit();
578 return(NULL);
579}
580
581void *
582realloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
583{
584 SLZone *z;
585 void *nptr;
586 unsigned long osize;
587
588 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
589 return(malloc(size, type, flags));
590 if (size == 0) {
591 free(ptr, type);
592 return(NULL);
593 }
594
595 /*
596 * Handle oversized allocations. XXX we really should require that a
597 * size be passed to free() instead of this nonsense.
598 */
599 {
600 struct kmemusage *kup;
601
602 kup = btokup(ptr);
603 if (kup->ku_pagecnt) {
604 osize = kup->ku_pagecnt << PAGE_SHIFT;
605 if (osize == round_page(size))
606 return(ptr);
607 if ((nptr = malloc(size, type, flags)) == NULL)
608 return(NULL);
609 bcopy(ptr, nptr, min(size, osize));
610 free(ptr, type);
611 return(nptr);
612 }
613 }
614
615 /*
616 * Get the original allocation's zone. If the new request winds up
617 * using the same chunk size we do not have to do anything.
618 */
619 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
620 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
621
622 zoneindex(&size);
623 if (z->z_ChunkSize == size)
624 return(ptr);
625
626 /*
627 * Allocate memory for the new request size. Note that zoneindex has
628 * already adjusted the request size to the appropriate chunk size, which
629 * should optimize our bcopy(). Then copy and return the new pointer.
630 */
631 if ((nptr = malloc(size, type, flags)) == NULL)
632 return(NULL);
633 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
634 free(ptr, type);
635 return(nptr);
636}
637
638/*
639 * free() (SLAB ALLOCATOR)
640 *
bba6a44d 641 * Free the specified chunk of memory.
a108bf71
MD
642 */
643static
644void
645free_remote(void *ptr)
646{
647 free(ptr, *(struct malloc_type **)ptr);
648}
649
650void
651free(void *ptr, struct malloc_type *type)
652{
653 SLZone *z;
654 SLChunk *chunk;
655 SLGlobalData *slgd;
bba6a44d 656 struct globaldata *gd;
a108bf71
MD
657 int pgno;
658
bba6a44d
MD
659 gd = mycpu;
660 slgd = &gd->gd_slab;
a108bf71
MD
661
662 /*
663 * Handle special 0-byte allocations
664 */
665 if (ptr == ZERO_LENGTH_PTR)
666 return;
667
668 /*
669 * Handle oversized allocations. XXX we really should require that a
670 * size be passed to free() instead of this nonsense.
bba6a44d
MD
671 *
672 * This code is never called via an ipi.
a108bf71
MD
673 */
674 {
675 struct kmemusage *kup;
676 unsigned long size;
677
678 kup = btokup(ptr);
679 if (kup->ku_pagecnt) {
680 size = kup->ku_pagecnt << PAGE_SHIFT;
681 kup->ku_pagecnt = 0;
a108bf71
MD
682#ifdef INVARIANTS
683 KKASSERT(sizeof(weirdary) <= size);
684 bcopy(weirdary, ptr, sizeof(weirdary));
685#endif
bba6a44d
MD
686 /*
687 * note: we always adjust our cpu's slot, not the originating
688 * cpu (kup->ku_cpuid). The statistics are in aggregate.
689 */
690 crit_enter();
691 --type->ks_inuse[gd->gd_cpuid];
692 type->ks_memuse[gd->gd_cpuid] -= size;
46a3f46d 693 if (mycpu->gd_intr_nesting_level) {
46a3f46d
MD
694 z = (SLZone *)ptr;
695 z->z_Magic = ZALLOC_OVSZ_MAGIC;
696 z->z_Next = slgd->FreeOvZones;
697 z->z_ChunkSize = size;
698 slgd->FreeOvZones = z;
699 crit_exit();
700 } else {
bba6a44d 701 crit_exit();
46a3f46d
MD
702 kmem_slab_free(ptr, size); /* may block */
703 }
a108bf71
MD
704 return;
705 }
706 }
707
708 /*
709 * Zone case. Figure out the zone based on the fact that it is
710 * ZoneSize aligned.
711 */
712 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
713 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
714
715 /*
716 * If we do not own the zone then forward the request to the
717 * cpu that does. The freeing code does not need the byte count
718 * unless DIAGNOSTIC is set.
719 */
bba6a44d 720 if (z->z_Cpu != gd->gd_cpuid) {
a108bf71
MD
721 *(struct malloc_type **)ptr = type;
722 lwkt_send_ipiq(z->z_Cpu, free_remote, ptr);
723 return;
724 }
725
726 if (type->ks_magic != M_MAGIC)
727 panic("free: malloc type lacks magic");
728
729 crit_enter();
730 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT;
731 chunk = ptr;
732
bba6a44d 733#ifdef INVARIANTS
a108bf71 734 /*
bba6a44d
MD
735 * Attempt to detect a double-free. To reduce overhead we only check
736 * if there appears to be link pointer at the base of the data.
a108bf71
MD
737 */
738 if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) {
739 SLChunk *scan;
740 for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) {
741 if (scan == chunk)
742 panic("Double free at %p", chunk);
743 }
744 }
745#endif
746
747 /*
748 * Put weird data into the memory to detect modifications after freeing,
749 * illegal pointer use after freeing (we should fault on the odd address),
750 * and so forth. XXX needs more work, see the old malloc code.
751 */
752#ifdef INVARIANTS
753 if (z->z_ChunkSize < sizeof(weirdary))
754 bcopy(weirdary, chunk, z->z_ChunkSize);
755 else
756 bcopy(weirdary, chunk, sizeof(weirdary));
757#endif
758
759 /*
760 * Add this free non-zero'd chunk to a linked list for reuse, adjust
761 * z_FirstFreePg.
762 */
6ab8e1da
MD
763#ifdef INVARIANTS
764 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
a108bf71 765 panic("BADFREE %p\n", chunk);
a108bf71
MD
766#endif
767 chunk->c_Next = z->z_PageAry[pgno];
768 z->z_PageAry[pgno] = chunk;
6ab8e1da
MD
769#ifdef INVARIANTS
770 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
a108bf71 771 panic("BADFREE2");
6ab8e1da 772#endif
a108bf71
MD
773 if (z->z_FirstFreePg > pgno)
774 z->z_FirstFreePg = pgno;
775
776 /*
777 * Bump the number of free chunks. If it becomes non-zero the zone
778 * must be added back onto the appropriate list.
779 */
780 if (z->z_NFree++ == 0) {
781 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
782 slgd->ZoneAry[z->z_ZoneIndex] = z;
783 }
784
bba6a44d
MD
785 --type->ks_inuse[z->z_Cpu];
786 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
a108bf71
MD
787
788 /*
789 * If the zone becomes totally free, and there are other zones we
a7cf0021
MD
790 * can allocate from, move this zone to the FreeZones list. Since
791 * this code can be called from an IPI callback, do *NOT* try to mess
792 * with kernel_map here. Hysteresis will be performed at malloc() time.
a108bf71
MD
793 */
794 if (z->z_NFree == z->z_NMax &&
795 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z)
796 ) {
797 SLZone **pz;
798
799 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
800 ;
801 *pz = z->z_Next;
802 z->z_Magic = -1;
a7cf0021
MD
803 z->z_Next = slgd->FreeZones;
804 slgd->FreeZones = z;
805 ++slgd->NFreeZones;
a108bf71
MD
806 }
807 crit_exit();
808}
809
810/*
811 * kmem_slab_alloc()
812 *
813 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
814 * specified alignment. M_* flags are expected in the flags field.
815 *
816 * Alignment must be a multiple of PAGE_SIZE.
817 *
818 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
819 * but when we move zalloc() over to use this function as its backend
820 * we will have to switch to kreserve/krelease and call reserve(0)
821 * after the new space is made available.
822 */
823static void *
824kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
825{
826 vm_size_t i;
827 vm_offset_t addr;
828 vm_offset_t offset;
829 int count;
830 vm_map_t map = kernel_map;
831
832 size = round_page(size);
833 addr = vm_map_min(map);
834
835 /*
836 * Reserve properly aligned space from kernel_map
837 */
838 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
839 crit_enter();
840 vm_map_lock(map);
841 if (vm_map_findspace(map, vm_map_min(map), size, align, &addr)) {
842 vm_map_unlock(map);
843 if ((flags & (M_NOWAIT|M_NULLOK)) == 0)
844 panic("kmem_slab_alloc(): kernel_map ran out of space!");
845 crit_exit();
846 vm_map_entry_release(count);
847 return(NULL);
848 }
849 offset = addr - VM_MIN_KERNEL_ADDRESS;
850 vm_object_reference(kernel_object);
851 vm_map_insert(map, &count,
852 kernel_object, offset, addr, addr + size,
853 VM_PROT_ALL, VM_PROT_ALL, 0);
854
855 /*
856 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
857 */
858 for (i = 0; i < size; i += PAGE_SIZE) {
859 vm_page_t m;
860 vm_pindex_t idx = OFF_TO_IDX(offset + i);
861 int zero = (flags & M_ZERO) ? VM_ALLOC_ZERO : 0;
862
863 if ((flags & (M_NOWAIT|M_USE_RESERVE)) == M_NOWAIT)
864 m = vm_page_alloc(kernel_object, idx, VM_ALLOC_INTERRUPT|zero);
865 else
866 m = vm_page_alloc(kernel_object, idx, VM_ALLOC_SYSTEM|zero);
867 if (m == NULL) {
868 if ((flags & M_NOWAIT) == 0) {
869 vm_map_unlock(map);
870 vm_wait();
871 vm_map_lock(map);
872 i -= PAGE_SIZE; /* retry */
873 continue;
874 }
875 while (i != 0) {
876 i -= PAGE_SIZE;
877 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
878 vm_page_free(m);
879 }
880 vm_map_delete(map, addr, addr + size, &count);
881 vm_map_unlock(map);
882 crit_exit();
883 vm_map_entry_release(count);
884 return(NULL);
885 }
886 }
887
888 /*
889 * Mark the map entry as non-pageable using a routine that allows us to
890 * populate the underlying pages.
891 */
892 vm_map_set_wired_quick(map, addr, size, &count);
893 crit_exit();
894
895 /*
896 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
897 */
898 for (i = 0; i < size; i += PAGE_SIZE) {
899 vm_page_t m;
900
901 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
902 m->valid = VM_PAGE_BITS_ALL;
903 vm_page_wire(m);
904 vm_page_wakeup(m);
905 pmap_enter(kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
906 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
907 bzero((char *)addr + i, PAGE_SIZE);
908 vm_page_flag_clear(m, PG_ZERO);
909 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_REFERENCED);
910 }
911 vm_map_unlock(map);
912 vm_map_entry_release(count);
913 return((void *)addr);
914}
915
916static void
917kmem_slab_free(void *ptr, vm_size_t size)
918{
919 crit_enter();
920 vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
921 crit_exit();
922}
923
924#endif