Fix an error message.
[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 *
1d712609 28 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.12 2003/10/19 18:18:50 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
a108bf71
MD
86#include <sys/param.h>
87#include <sys/systm.h>
88#include <sys/kernel.h>
89#include <sys/slaballoc.h>
90#include <sys/mbuf.h>
91#include <sys/vmmeter.h>
92#include <sys/lock.h>
93#include <sys/thread.h>
94#include <sys/globaldata.h>
95
96#include <vm/vm.h>
97#include <vm/vm_param.h>
98#include <vm/vm_kern.h>
99#include <vm/vm_extern.h>
100#include <vm/vm_object.h>
101#include <vm/pmap.h>
102#include <vm/vm_map.h>
103#include <vm/vm_page.h>
104#include <vm/vm_pageout.h>
105
106#include <machine/cpu.h>
107
108#include <sys/thread2.h>
109
110#define arysize(ary) (sizeof(ary)/sizeof((ary)[0]))
111
112/*
113 * Fixed globals (not per-cpu)
114 */
115static int ZoneSize;
46a3f46d 116static int ZoneLimit;
a108bf71
MD
117static int ZonePageCount;
118static int ZonePageLimit;
119static int ZoneMask;
120static struct malloc_type *kmemstatistics;
121static struct kmemusage *kmemusage;
122static int32_t weirdary[16];
123
124static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
125static void kmem_slab_free(void *ptr, vm_size_t bytes);
126
127/*
128 * Misc constants. Note that allocations that are exact multiples of
129 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
130 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
131 */
132#define MIN_CHUNK_SIZE 8 /* in bytes */
133#define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
134#define ZONE_RELS_THRESH 2 /* threshold number of zones */
135#define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
136
137/*
138 * The WEIRD_ADDR is used as known text to copy into free objects to
139 * try to create deterministic failure cases if the data is accessed after
140 * free.
141 */
142#define WEIRD_ADDR 0xdeadc0de
143#define MAX_COPY sizeof(weirdary)
144#define ZERO_LENGTH_PTR ((void *)-8)
145
146/*
147 * Misc global malloc buckets
148 */
149
150MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
151MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
152MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
153
154MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
155MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
156
157/*
158 * Initialize the slab memory allocator. We have to choose a zone size based
159 * on available physical memory. We choose a zone side which is approximately
160 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
161 * 128K. The zone size is limited to the bounds set in slaballoc.h
162 * (typically 32K min, 128K max).
163 */
164static void kmeminit(void *dummy);
165
166SYSINIT(kmem, SI_SUB_KMEM, SI_ORDER_FIRST, kmeminit, NULL)
167
168static void
169kmeminit(void *dummy)
170{
171 vm_poff_t limsize;
172 int usesize;
173 int i;
174 vm_pindex_t npg;
175
176 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
177 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS)
178 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
179
180 usesize = (int)(limsize / 1024); /* convert to KB */
181
182 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
183 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
184 ZoneSize <<= 1;
46a3f46d
MD
185 ZoneLimit = ZoneSize / 4;
186 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
187 ZoneLimit = ZALLOC_ZONE_LIMIT;
a108bf71
MD
188 ZoneMask = ZoneSize - 1;
189 ZonePageLimit = PAGE_SIZE * 4;
190 ZonePageCount = ZoneSize / PAGE_SIZE;
191
192 npg = (VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE;
193 kmemusage = kmem_slab_alloc(npg * sizeof(struct kmemusage), PAGE_SIZE, M_ZERO);
194
195 for (i = 0; i < arysize(weirdary); ++i)
196 weirdary[i] = WEIRD_ADDR;
197
198 if (bootverbose)
199 printf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
200}
201
202/*
bba6a44d 203 * Initialize a malloc type tracking structure.
a108bf71
MD
204 */
205void
206malloc_init(void *data)
207{
208 struct malloc_type *type = data;
209 vm_poff_t limsize;
210
211 if (type->ks_magic != M_MAGIC)
212 panic("malloc type lacks magic");
213
214 if (type->ks_limit != 0)
215 return;
216
217 if (vmstats.v_page_count == 0)
218 panic("malloc_init not allowed before vm init");
219
220 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
221 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS)
222 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
223 type->ks_limit = limsize / 10;
224
225 type->ks_next = kmemstatistics;
226 kmemstatistics = type;
227}
228
229void
230malloc_uninit(void *data)
231{
232 struct malloc_type *type = data;
233 struct malloc_type *t;
bba6a44d
MD
234#ifdef INVARIANTS
235 int i;
1d712609 236 long ttl;
bba6a44d 237#endif
a108bf71
MD
238
239 if (type->ks_magic != M_MAGIC)
240 panic("malloc type lacks magic");
241
242 if (vmstats.v_page_count == 0)
243 panic("malloc_uninit not allowed before vm init");
244
245 if (type->ks_limit == 0)
246 panic("malloc_uninit on uninitialized type");
247
248#ifdef INVARIANTS
1d712609
MD
249 /*
250 * memuse is only correct in aggregation. Due to memory being allocated
251 * on one cpu and freed on another individual array entries may be
252 * negative or positive (canceling each other out).
253 */
254 for (i = ttl = 0; i < ncpus; ++i)
255 ttl += type->ks_memuse[i];
256 if (ttl) {
257 printf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
258 ttl, type->ks_shortdesc, i);
a108bf71
MD
259 }
260#endif
261 if (type == kmemstatistics) {
262 kmemstatistics = type->ks_next;
263 } else {
264 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
265 if (t->ks_next == type) {
266 t->ks_next = type->ks_next;
267 break;
268 }
269 }
270 }
271 type->ks_next = NULL;
272 type->ks_limit = 0;
273}
274
275/*
276 * Calculate the zone index for the allocation request size and set the
277 * allocation request size to that particular zone's chunk size.
278 */
279static __inline int
280zoneindex(unsigned long *bytes)
281{
282 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
283 if (n < 128) {
284 *bytes = n = (n + 7) & ~7;
285 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
286 }
287 if (n < 256) {
288 *bytes = n = (n + 15) & ~15;
289 return(n / 16 + 7);
290 }
291 if (n < 8192) {
292 if (n < 512) {
293 *bytes = n = (n + 31) & ~31;
294 return(n / 32 + 15);
295 }
296 if (n < 1024) {
297 *bytes = n = (n + 63) & ~63;
298 return(n / 64 + 23);
299 }
300 if (n < 2048) {
301 *bytes = n = (n + 127) & ~127;
302 return(n / 128 + 31);
303 }
304 if (n < 4096) {
305 *bytes = n = (n + 255) & ~255;
306 return(n / 256 + 39);
307 }
308 *bytes = n = (n + 511) & ~511;
309 return(n / 512 + 47);
310 }
311#if ZALLOC_ZONE_LIMIT > 8192
312 if (n < 16384) {
313 *bytes = n = (n + 1023) & ~1023;
314 return(n / 1024 + 55);
315 }
316#endif
317#if ZALLOC_ZONE_LIMIT > 16384
318 if (n < 32768) {
319 *bytes = n = (n + 2047) & ~2047;
320 return(n / 2048 + 63);
321 }
322#endif
323 panic("Unexpected byte count %d", n);
324 return(0);
325}
326
327/*
328 * malloc() (SLAB ALLOCATOR)
329 *
330 * Allocate memory via the slab allocator. If the request is too large,
331 * or if it page-aligned beyond a certain size, we fall back to the
332 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
333 * &SlabMisc if you don't care.
334 *
335 * M_NOWAIT - return NULL instead of blocking.
336 * M_ZERO - zero the returned memory.
337 * M_USE_RESERVE - allocate out of the system reserve if necessary
338 */
339void *
340malloc(unsigned long size, struct malloc_type *type, int flags)
341{
342 SLZone *z;
343 SLChunk *chunk;
344 SLGlobalData *slgd;
bba6a44d 345 struct globaldata *gd;
a108bf71
MD
346 int zi;
347
bba6a44d
MD
348 gd = mycpu;
349 slgd = &gd->gd_slab;
a108bf71
MD
350
351 /*
352 * XXX silly to have this in the critical path.
353 */
354 if (type->ks_limit == 0) {
355 crit_enter();
356 if (type->ks_limit == 0)
357 malloc_init(type);
358 crit_exit();
359 }
360 ++type->ks_calls;
361
362 /*
363 * Handle the case where the limit is reached. Panic if can't return
364 * NULL. XXX the original malloc code looped, but this tended to
365 * simply deadlock the computer.
366 */
bba6a44d
MD
367 while (type->ks_loosememuse >= type->ks_limit) {
368 int i;
369 long ttl;
370
371 for (i = ttl = 0; i < ncpus; ++i)
372 ttl += type->ks_memuse[i];
373 type->ks_loosememuse = ttl;
374 if (ttl >= type->ks_limit) {
375 if (flags & (M_NOWAIT|M_NULLOK))
376 return(NULL);
377 panic("%s: malloc limit exceeded", type->ks_shortdesc);
378 }
a108bf71
MD
379 }
380
381 /*
382 * Handle the degenerate size == 0 case. Yes, this does happen.
383 * Return a special pointer. This is to maintain compatibility with
384 * the original malloc implementation. Certain devices, such as the
385 * adaptec driver, not only allocate 0 bytes, they check for NULL and
386 * also realloc() later on. Joy.
387 */
388 if (size == 0)
389 return(ZERO_LENGTH_PTR);
390
a7cf0021
MD
391 /*
392 * Handle hysteresis from prior frees here in malloc(). We cannot
393 * safely manipulate the kernel_map in free() due to free() possibly
394 * being called via an IPI message or from sensitive interrupt code.
395 */
396 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_NOWAIT) == 0) {
46a3f46d
MD
397 crit_enter();
398 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */
399 z = slgd->FreeZones;
400 slgd->FreeZones = z->z_Next;
401 --slgd->NFreeZones;
402 kmem_slab_free(z, ZoneSize); /* may block */
403 }
404 crit_exit();
405 }
406 /*
407 * XXX handle oversized frees that were queued from free().
408 */
409 while (slgd->FreeOvZones && (flags & M_NOWAIT) == 0) {
410 crit_enter();
411 if ((z = slgd->FreeOvZones) != NULL) {
412 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
413 slgd->FreeOvZones = z->z_Next;
414 kmem_slab_free(z, z->z_ChunkSize); /* may block */
415 }
416 crit_exit();
a7cf0021
MD
417 }
418
a108bf71
MD
419 /*
420 * Handle large allocations directly. There should not be very many of
421 * these so performance is not a big issue.
422 *
423 * Guarentee page alignment for allocations in multiples of PAGE_SIZE
424 */
46a3f46d 425 if (size >= ZoneLimit || (size & PAGE_MASK) == 0) {
a108bf71
MD
426 struct kmemusage *kup;
427
428 size = round_page(size);
429 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
430 if (chunk == NULL)
431 return(NULL);
432 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
433 kup = btokup(chunk);
434 kup->ku_pagecnt = size / PAGE_SIZE;
bba6a44d 435 kup->ku_cpu = gd->gd_cpuid;
a108bf71
MD
436 crit_enter();
437 goto done;
438 }
439
440 /*
441 * Attempt to allocate out of an existing zone. First try the free list,
442 * then allocate out of unallocated space. If we find a good zone move
443 * it to the head of the list so later allocations find it quickly
444 * (we might have thousands of zones in the list).
445 *
446 * Note: zoneindex() will panic of size is too large.
447 */
448 zi = zoneindex(&size);
449 KKASSERT(zi < NZONES);
450 crit_enter();
451 if ((z = slgd->ZoneAry[zi]) != NULL) {
452 KKASSERT(z->z_NFree > 0);
453
454 /*
455 * Remove us from the ZoneAry[] when we become empty
456 */
457 if (--z->z_NFree == 0) {
458 slgd->ZoneAry[zi] = z->z_Next;
459 z->z_Next = NULL;
460 }
461
462 /*
463 * Locate a chunk in a free page. This attempts to localize
464 * reallocations into earlier pages without us having to sort
465 * the chunk list. A chunk may still overlap a page boundary.
466 */
467 while (z->z_FirstFreePg < ZonePageCount) {
468 if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) {
469#ifdef DIAGNOSTIC
470 /*
471 * Diagnostic: c_Next is not total garbage.
472 */
473 KKASSERT(chunk->c_Next == NULL ||
474 ((intptr_t)chunk->c_Next & IN_SAME_PAGE_MASK) ==
475 ((intptr_t)chunk & IN_SAME_PAGE_MASK));
476#endif
6ab8e1da
MD
477#ifdef INVARIANTS
478 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
a108bf71 479 panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount);
6ab8e1da 480 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
a108bf71 481 panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount);
6ab8e1da 482#endif
a108bf71
MD
483 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next;
484 goto done;
485 }
486 ++z->z_FirstFreePg;
487 }
488
489 /*
1c5ca4f3
MD
490 * No chunks are available but NFree said we had some memory, so
491 * it must be available in the never-before-used-memory area
492 * governed by UIndex. The consequences are very serious if our zone
493 * got corrupted so we use an explicit panic rather then a KASSERT.
a108bf71 494 */
1c5ca4f3
MD
495 if (z->z_UIndex + 1 != z->z_NMax)
496 z->z_UIndex = z->z_UIndex + 1;
497 else
498 z->z_UIndex = 0;
499 if (z->z_UIndex == z->z_UEndIndex)
500 panic("slaballoc: corrupted zone");
501 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
6ab8e1da
MD
502 if ((z->z_Flags & SLZF_UNOTZEROD) == 0)
503 flags &= ~M_ZERO;
a108bf71
MD
504 goto done;
505 }
506
507 /*
508 * If all zones are exhausted we need to allocate a new zone for this
509 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
6ab8e1da
MD
510 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
511 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
512 * we do not pre-zero it because we do not want to mess up the L1 cache.
a108bf71
MD
513 *
514 * At least one subsystem, the tty code (see CROUND) expects power-of-2
515 * allocations to be power-of-2 aligned. We maintain compatibility by
516 * adjusting the base offset below.
517 */
518 {
519 int off;
520
521 if ((z = slgd->FreeZones) != NULL) {
522 slgd->FreeZones = z->z_Next;
523 --slgd->NFreeZones;
524 bzero(z, sizeof(SLZone));
6ab8e1da 525 z->z_Flags |= SLZF_UNOTZEROD;
a108bf71
MD
526 } else {
527 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
528 if (z == NULL)
529 goto fail;
530 }
531
532 /*
533 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
534 * Otherwise just 8-byte align the data.
535 */
536 if ((size | (size - 1)) + 1 == (size << 1))
537 off = (sizeof(SLZone) + size - 1) & ~(size - 1);
538 else
539 off = (sizeof(SLZone) + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
540 z->z_Magic = ZALLOC_SLAB_MAGIC;
541 z->z_ZoneIndex = zi;
542 z->z_NMax = (ZoneSize - off) / size;
543 z->z_NFree = z->z_NMax - 1;
1c5ca4f3
MD
544 z->z_BasePtr = (char *)z + off;
545 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
a108bf71
MD
546 z->z_ChunkSize = size;
547 z->z_FirstFreePg = ZonePageCount;
bba6a44d 548 z->z_Cpu = gd->gd_cpuid;
1c5ca4f3 549 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
a108bf71
MD
550 z->z_Next = slgd->ZoneAry[zi];
551 slgd->ZoneAry[zi] = z;
6ab8e1da
MD
552 if ((z->z_Flags & SLZF_UNOTZEROD) == 0)
553 flags &= ~M_ZERO; /* already zero'd */
1c5ca4f3
MD
554
555 /*
556 * Slide the base index for initial allocations out of the next
557 * zone we create so we do not over-weight the lower part of the
558 * cpu memory caches.
559 */
560 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
561 & (ZALLOC_MAX_ZONE_SIZE - 1);
a108bf71
MD
562 }
563done:
bba6a44d
MD
564 ++type->ks_inuse[gd->gd_cpuid];
565 type->ks_memuse[gd->gd_cpuid] += size;
566 type->ks_loosememuse += size;
a108bf71
MD
567 crit_exit();
568 if (flags & M_ZERO)
569 bzero(chunk, size);
bba6a44d
MD
570#ifdef INVARIANTS
571 else
572 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
573#endif
a108bf71
MD
574 return(chunk);
575fail:
576 crit_exit();
577 return(NULL);
578}
579
580void *
581realloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
582{
583 SLZone *z;
584 void *nptr;
585 unsigned long osize;
586
587 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
588 return(malloc(size, type, flags));
589 if (size == 0) {
590 free(ptr, type);
591 return(NULL);
592 }
593
594 /*
595 * Handle oversized allocations. XXX we really should require that a
596 * size be passed to free() instead of this nonsense.
597 */
598 {
599 struct kmemusage *kup;
600
601 kup = btokup(ptr);
602 if (kup->ku_pagecnt) {
603 osize = kup->ku_pagecnt << PAGE_SHIFT;
604 if (osize == round_page(size))
605 return(ptr);
606 if ((nptr = malloc(size, type, flags)) == NULL)
607 return(NULL);
608 bcopy(ptr, nptr, min(size, osize));
609 free(ptr, type);
610 return(nptr);
611 }
612 }
613
614 /*
615 * Get the original allocation's zone. If the new request winds up
616 * using the same chunk size we do not have to do anything.
617 */
618 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
619 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
620
621 zoneindex(&size);
622 if (z->z_ChunkSize == size)
623 return(ptr);
624
625 /*
626 * Allocate memory for the new request size. Note that zoneindex has
627 * already adjusted the request size to the appropriate chunk size, which
628 * should optimize our bcopy(). Then copy and return the new pointer.
629 */
630 if ((nptr = malloc(size, type, flags)) == NULL)
631 return(NULL);
632 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
633 free(ptr, type);
634 return(nptr);
635}
636
1d712609 637#ifdef SMP
a108bf71
MD
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
1d712609
MD
650#endif
651
a108bf71
MD
652void
653free(void *ptr, struct malloc_type *type)
654{
655 SLZone *z;
656 SLChunk *chunk;
657 SLGlobalData *slgd;
bba6a44d 658 struct globaldata *gd;
a108bf71
MD
659 int pgno;
660
bba6a44d
MD
661 gd = mycpu;
662 slgd = &gd->gd_slab;
a108bf71
MD
663
664 /*
665 * Handle special 0-byte allocations
666 */
667 if (ptr == ZERO_LENGTH_PTR)
668 return;
669
670 /*
671 * Handle oversized allocations. XXX we really should require that a
672 * size be passed to free() instead of this nonsense.
bba6a44d
MD
673 *
674 * This code is never called via an ipi.
a108bf71
MD
675 */
676 {
677 struct kmemusage *kup;
678 unsigned long size;
679
680 kup = btokup(ptr);
681 if (kup->ku_pagecnt) {
682 size = kup->ku_pagecnt << PAGE_SHIFT;
683 kup->ku_pagecnt = 0;
a108bf71
MD
684#ifdef INVARIANTS
685 KKASSERT(sizeof(weirdary) <= size);
686 bcopy(weirdary, ptr, sizeof(weirdary));
687#endif
bba6a44d
MD
688 /*
689 * note: we always adjust our cpu's slot, not the originating
690 * cpu (kup->ku_cpuid). The statistics are in aggregate.
691 */
692 crit_enter();
693 --type->ks_inuse[gd->gd_cpuid];
694 type->ks_memuse[gd->gd_cpuid] -= size;
46a3f46d 695 if (mycpu->gd_intr_nesting_level) {
46a3f46d
MD
696 z = (SLZone *)ptr;
697 z->z_Magic = ZALLOC_OVSZ_MAGIC;
698 z->z_Next = slgd->FreeOvZones;
699 z->z_ChunkSize = size;
700 slgd->FreeOvZones = z;
701 crit_exit();
702 } else {
bba6a44d 703 crit_exit();
46a3f46d
MD
704 kmem_slab_free(ptr, size); /* may block */
705 }
a108bf71
MD
706 return;
707 }
708 }
709
710 /*
711 * Zone case. Figure out the zone based on the fact that it is
712 * ZoneSize aligned.
713 */
714 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
715 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
716
717 /*
718 * If we do not own the zone then forward the request to the
719 * cpu that does. The freeing code does not need the byte count
720 * unless DIAGNOSTIC is set.
721 */
bba6a44d 722 if (z->z_Cpu != gd->gd_cpuid) {
a108bf71 723 *(struct malloc_type **)ptr = type;
75c7ffea 724#ifdef SMP
a108bf71 725 lwkt_send_ipiq(z->z_Cpu, free_remote, ptr);
75c7ffea
MD
726#else
727 panic("Corrupt SLZone");
728#endif
a108bf71
MD
729 return;
730 }
731
732 if (type->ks_magic != M_MAGIC)
733 panic("free: malloc type lacks magic");
734
735 crit_enter();
736 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT;
737 chunk = ptr;
738
bba6a44d 739#ifdef INVARIANTS
a108bf71 740 /*
bba6a44d
MD
741 * Attempt to detect a double-free. To reduce overhead we only check
742 * if there appears to be link pointer at the base of the data.
a108bf71
MD
743 */
744 if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) {
745 SLChunk *scan;
746 for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) {
747 if (scan == chunk)
748 panic("Double free at %p", chunk);
749 }
750 }
751#endif
752
753 /*
754 * Put weird data into the memory to detect modifications after freeing,
755 * illegal pointer use after freeing (we should fault on the odd address),
756 * and so forth. XXX needs more work, see the old malloc code.
757 */
758#ifdef INVARIANTS
759 if (z->z_ChunkSize < sizeof(weirdary))
760 bcopy(weirdary, chunk, z->z_ChunkSize);
761 else
762 bcopy(weirdary, chunk, sizeof(weirdary));
763#endif
764
765 /*
766 * Add this free non-zero'd chunk to a linked list for reuse, adjust
767 * z_FirstFreePg.
768 */
6ab8e1da
MD
769#ifdef INVARIANTS
770 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
a108bf71 771 panic("BADFREE %p\n", chunk);
a108bf71
MD
772#endif
773 chunk->c_Next = z->z_PageAry[pgno];
774 z->z_PageAry[pgno] = chunk;
6ab8e1da
MD
775#ifdef INVARIANTS
776 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
a108bf71 777 panic("BADFREE2");
6ab8e1da 778#endif
a108bf71
MD
779 if (z->z_FirstFreePg > pgno)
780 z->z_FirstFreePg = pgno;
781
782 /*
783 * Bump the number of free chunks. If it becomes non-zero the zone
784 * must be added back onto the appropriate list.
785 */
786 if (z->z_NFree++ == 0) {
787 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
788 slgd->ZoneAry[z->z_ZoneIndex] = z;
789 }
790
bba6a44d
MD
791 --type->ks_inuse[z->z_Cpu];
792 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
a108bf71
MD
793
794 /*
795 * If the zone becomes totally free, and there are other zones we
a7cf0021
MD
796 * can allocate from, move this zone to the FreeZones list. Since
797 * this code can be called from an IPI callback, do *NOT* try to mess
798 * with kernel_map here. Hysteresis will be performed at malloc() time.
a108bf71
MD
799 */
800 if (z->z_NFree == z->z_NMax &&
801 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z)
802 ) {
803 SLZone **pz;
804
805 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
806 ;
807 *pz = z->z_Next;
808 z->z_Magic = -1;
a7cf0021
MD
809 z->z_Next = slgd->FreeZones;
810 slgd->FreeZones = z;
811 ++slgd->NFreeZones;
a108bf71
MD
812 }
813 crit_exit();
814}
815
816/*
817 * kmem_slab_alloc()
818 *
819 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
820 * specified alignment. M_* flags are expected in the flags field.
821 *
822 * Alignment must be a multiple of PAGE_SIZE.
823 *
824 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
825 * but when we move zalloc() over to use this function as its backend
826 * we will have to switch to kreserve/krelease and call reserve(0)
827 * after the new space is made available.
828 */
829static void *
830kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
831{
832 vm_size_t i;
833 vm_offset_t addr;
834 vm_offset_t offset;
835 int count;
836 vm_map_t map = kernel_map;
837
838 size = round_page(size);
839 addr = vm_map_min(map);
840
841 /*
842 * Reserve properly aligned space from kernel_map
843 */
844 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
845 crit_enter();
846 vm_map_lock(map);
847 if (vm_map_findspace(map, vm_map_min(map), size, align, &addr)) {
848 vm_map_unlock(map);
849 if ((flags & (M_NOWAIT|M_NULLOK)) == 0)
850 panic("kmem_slab_alloc(): kernel_map ran out of space!");
851 crit_exit();
852 vm_map_entry_release(count);
853 return(NULL);
854 }
855 offset = addr - VM_MIN_KERNEL_ADDRESS;
856 vm_object_reference(kernel_object);
857 vm_map_insert(map, &count,
858 kernel_object, offset, addr, addr + size,
859 VM_PROT_ALL, VM_PROT_ALL, 0);
860
861 /*
862 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
863 */
864 for (i = 0; i < size; i += PAGE_SIZE) {
865 vm_page_t m;
866 vm_pindex_t idx = OFF_TO_IDX(offset + i);
867 int zero = (flags & M_ZERO) ? VM_ALLOC_ZERO : 0;
868
869 if ((flags & (M_NOWAIT|M_USE_RESERVE)) == M_NOWAIT)
870 m = vm_page_alloc(kernel_object, idx, VM_ALLOC_INTERRUPT|zero);
871 else
872 m = vm_page_alloc(kernel_object, idx, VM_ALLOC_SYSTEM|zero);
873 if (m == NULL) {
874 if ((flags & M_NOWAIT) == 0) {
875 vm_map_unlock(map);
876 vm_wait();
877 vm_map_lock(map);
878 i -= PAGE_SIZE; /* retry */
879 continue;
880 }
881 while (i != 0) {
882 i -= PAGE_SIZE;
883 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
884 vm_page_free(m);
885 }
886 vm_map_delete(map, addr, addr + size, &count);
887 vm_map_unlock(map);
888 crit_exit();
889 vm_map_entry_release(count);
890 return(NULL);
891 }
892 }
893
894 /*
895 * Mark the map entry as non-pageable using a routine that allows us to
896 * populate the underlying pages.
897 */
898 vm_map_set_wired_quick(map, addr, size, &count);
899 crit_exit();
900
901 /*
902 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
903 */
904 for (i = 0; i < size; i += PAGE_SIZE) {
905 vm_page_t m;
906
907 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
908 m->valid = VM_PAGE_BITS_ALL;
909 vm_page_wire(m);
910 vm_page_wakeup(m);
911 pmap_enter(kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
912 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
913 bzero((char *)addr + i, PAGE_SIZE);
914 vm_page_flag_clear(m, PG_ZERO);
915 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_REFERENCED);
916 }
917 vm_map_unlock(map);
918 vm_map_entry_release(count);
919 return((void *)addr);
920}
921
922static void
923kmem_slab_free(void *ptr, vm_size_t size)
924{
925 crit_enter();
926 vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
927 crit_exit();
928}
929