Fix a ``missing free''.
[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 *
b0d4001a 28 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.11 2003/10/19 00:23:24 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;
236#endif
a108bf71
MD
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
bba6a44d
MD
248 for (i = 0; i < ncpus; ++i) {
249 if (type->ks_memuse[i] != 0) {
250 printf(
251 "malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
252 type->ks_memuse[i], type->ks_shortdesc, i);
253 }
a108bf71
MD
254 }
255#endif
256 if (type == kmemstatistics) {
257 kmemstatistics = type->ks_next;
258 } else {
259 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
260 if (t->ks_next == type) {
261 t->ks_next = type->ks_next;
262 break;
263 }
264 }
265 }
266 type->ks_next = NULL;
267 type->ks_limit = 0;
268}
269
270/*
271 * Calculate the zone index for the allocation request size and set the
272 * allocation request size to that particular zone's chunk size.
273 */
274static __inline int
275zoneindex(unsigned long *bytes)
276{
277 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
278 if (n < 128) {
279 *bytes = n = (n + 7) & ~7;
280 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
281 }
282 if (n < 256) {
283 *bytes = n = (n + 15) & ~15;
284 return(n / 16 + 7);
285 }
286 if (n < 8192) {
287 if (n < 512) {
288 *bytes = n = (n + 31) & ~31;
289 return(n / 32 + 15);
290 }
291 if (n < 1024) {
292 *bytes = n = (n + 63) & ~63;
293 return(n / 64 + 23);
294 }
295 if (n < 2048) {
296 *bytes = n = (n + 127) & ~127;
297 return(n / 128 + 31);
298 }
299 if (n < 4096) {
300 *bytes = n = (n + 255) & ~255;
301 return(n / 256 + 39);
302 }
303 *bytes = n = (n + 511) & ~511;
304 return(n / 512 + 47);
305 }
306#if ZALLOC_ZONE_LIMIT > 8192
307 if (n < 16384) {
308 *bytes = n = (n + 1023) & ~1023;
309 return(n / 1024 + 55);
310 }
311#endif
312#if ZALLOC_ZONE_LIMIT > 16384
313 if (n < 32768) {
314 *bytes = n = (n + 2047) & ~2047;
315 return(n / 2048 + 63);
316 }
317#endif
318 panic("Unexpected byte count %d", n);
319 return(0);
320}
321
322/*
323 * malloc() (SLAB ALLOCATOR)
324 *
325 * Allocate memory via the slab allocator. If the request is too large,
326 * or if it page-aligned beyond a certain size, we fall back to the
327 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
328 * &SlabMisc if you don't care.
329 *
330 * M_NOWAIT - return NULL instead of blocking.
331 * M_ZERO - zero the returned memory.
332 * M_USE_RESERVE - allocate out of the system reserve if necessary
333 */
334void *
335malloc(unsigned long size, struct malloc_type *type, int flags)
336{
337 SLZone *z;
338 SLChunk *chunk;
339 SLGlobalData *slgd;
bba6a44d 340 struct globaldata *gd;
a108bf71
MD
341 int zi;
342
bba6a44d
MD
343 gd = mycpu;
344 slgd = &gd->gd_slab;
a108bf71
MD
345
346 /*
347 * XXX silly to have this in the critical path.
348 */
349 if (type->ks_limit == 0) {
350 crit_enter();
351 if (type->ks_limit == 0)
352 malloc_init(type);
353 crit_exit();
354 }
355 ++type->ks_calls;
356
357 /*
358 * Handle the case where the limit is reached. Panic if can't return
359 * NULL. XXX the original malloc code looped, but this tended to
360 * simply deadlock the computer.
361 */
bba6a44d
MD
362 while (type->ks_loosememuse >= type->ks_limit) {
363 int i;
364 long ttl;
365
366 for (i = ttl = 0; i < ncpus; ++i)
367 ttl += type->ks_memuse[i];
368 type->ks_loosememuse = ttl;
369 if (ttl >= type->ks_limit) {
370 if (flags & (M_NOWAIT|M_NULLOK))
371 return(NULL);
372 panic("%s: malloc limit exceeded", type->ks_shortdesc);
373 }
a108bf71
MD
374 }
375
376 /*
377 * Handle the degenerate size == 0 case. Yes, this does happen.
378 * Return a special pointer. This is to maintain compatibility with
379 * the original malloc implementation. Certain devices, such as the
380 * adaptec driver, not only allocate 0 bytes, they check for NULL and
381 * also realloc() later on. Joy.
382 */
383 if (size == 0)
384 return(ZERO_LENGTH_PTR);
385
a7cf0021
MD
386 /*
387 * Handle hysteresis from prior frees here in malloc(). We cannot
388 * safely manipulate the kernel_map in free() due to free() possibly
389 * being called via an IPI message or from sensitive interrupt code.
390 */
391 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_NOWAIT) == 0) {
46a3f46d
MD
392 crit_enter();
393 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */
394 z = slgd->FreeZones;
395 slgd->FreeZones = z->z_Next;
396 --slgd->NFreeZones;
397 kmem_slab_free(z, ZoneSize); /* may block */
398 }
399 crit_exit();
400 }
401 /*
402 * XXX handle oversized frees that were queued from free().
403 */
404 while (slgd->FreeOvZones && (flags & M_NOWAIT) == 0) {
405 crit_enter();
406 if ((z = slgd->FreeOvZones) != NULL) {
407 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
408 slgd->FreeOvZones = z->z_Next;
409 kmem_slab_free(z, z->z_ChunkSize); /* may block */
410 }
411 crit_exit();
a7cf0021
MD
412 }
413
a108bf71
MD
414 /*
415 * Handle large allocations directly. There should not be very many of
416 * these so performance is not a big issue.
417 *
418 * Guarentee page alignment for allocations in multiples of PAGE_SIZE
419 */
46a3f46d 420 if (size >= ZoneLimit || (size & PAGE_MASK) == 0) {
a108bf71
MD
421 struct kmemusage *kup;
422
423 size = round_page(size);
424 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
425 if (chunk == NULL)
426 return(NULL);
427 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
428 kup = btokup(chunk);
429 kup->ku_pagecnt = size / PAGE_SIZE;
bba6a44d 430 kup->ku_cpu = gd->gd_cpuid;
a108bf71
MD
431 crit_enter();
432 goto done;
433 }
434
435 /*
436 * Attempt to allocate out of an existing zone. First try the free list,
437 * then allocate out of unallocated space. If we find a good zone move
438 * it to the head of the list so later allocations find it quickly
439 * (we might have thousands of zones in the list).
440 *
441 * Note: zoneindex() will panic of size is too large.
442 */
443 zi = zoneindex(&size);
444 KKASSERT(zi < NZONES);
445 crit_enter();
446 if ((z = slgd->ZoneAry[zi]) != NULL) {
447 KKASSERT(z->z_NFree > 0);
448
449 /*
450 * Remove us from the ZoneAry[] when we become empty
451 */
452 if (--z->z_NFree == 0) {
453 slgd->ZoneAry[zi] = z->z_Next;
454 z->z_Next = NULL;
455 }
456
457 /*
458 * Locate a chunk in a free page. This attempts to localize
459 * reallocations into earlier pages without us having to sort
460 * the chunk list. A chunk may still overlap a page boundary.
461 */
462 while (z->z_FirstFreePg < ZonePageCount) {
463 if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) {
464#ifdef DIAGNOSTIC
465 /*
466 * Diagnostic: c_Next is not total garbage.
467 */
468 KKASSERT(chunk->c_Next == NULL ||
469 ((intptr_t)chunk->c_Next & IN_SAME_PAGE_MASK) ==
470 ((intptr_t)chunk & IN_SAME_PAGE_MASK));
471#endif
6ab8e1da
MD
472#ifdef INVARIANTS
473 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
a108bf71 474 panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount);
6ab8e1da 475 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
a108bf71 476 panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount);
6ab8e1da 477#endif
a108bf71
MD
478 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next;
479 goto done;
480 }
481 ++z->z_FirstFreePg;
482 }
483
484 /*
1c5ca4f3
MD
485 * No chunks are available but NFree said we had some memory, so
486 * it must be available in the never-before-used-memory area
487 * governed by UIndex. The consequences are very serious if our zone
488 * got corrupted so we use an explicit panic rather then a KASSERT.
a108bf71 489 */
1c5ca4f3
MD
490 if (z->z_UIndex + 1 != z->z_NMax)
491 z->z_UIndex = z->z_UIndex + 1;
492 else
493 z->z_UIndex = 0;
494 if (z->z_UIndex == z->z_UEndIndex)
495 panic("slaballoc: corrupted zone");
496 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
6ab8e1da
MD
497 if ((z->z_Flags & SLZF_UNOTZEROD) == 0)
498 flags &= ~M_ZERO;
a108bf71
MD
499 goto done;
500 }
501
502 /*
503 * If all zones are exhausted we need to allocate a new zone for this
504 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
6ab8e1da
MD
505 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
506 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
507 * we do not pre-zero it because we do not want to mess up the L1 cache.
a108bf71
MD
508 *
509 * At least one subsystem, the tty code (see CROUND) expects power-of-2
510 * allocations to be power-of-2 aligned. We maintain compatibility by
511 * adjusting the base offset below.
512 */
513 {
514 int off;
515
516 if ((z = slgd->FreeZones) != NULL) {
517 slgd->FreeZones = z->z_Next;
518 --slgd->NFreeZones;
519 bzero(z, sizeof(SLZone));
6ab8e1da 520 z->z_Flags |= SLZF_UNOTZEROD;
a108bf71
MD
521 } else {
522 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
523 if (z == NULL)
524 goto fail;
525 }
526
527 /*
528 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
529 * Otherwise just 8-byte align the data.
530 */
531 if ((size | (size - 1)) + 1 == (size << 1))
532 off = (sizeof(SLZone) + size - 1) & ~(size - 1);
533 else
534 off = (sizeof(SLZone) + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
535 z->z_Magic = ZALLOC_SLAB_MAGIC;
536 z->z_ZoneIndex = zi;
537 z->z_NMax = (ZoneSize - off) / size;
538 z->z_NFree = z->z_NMax - 1;
1c5ca4f3
MD
539 z->z_BasePtr = (char *)z + off;
540 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
a108bf71
MD
541 z->z_ChunkSize = size;
542 z->z_FirstFreePg = ZonePageCount;
bba6a44d 543 z->z_Cpu = gd->gd_cpuid;
1c5ca4f3 544 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
a108bf71
MD
545 z->z_Next = slgd->ZoneAry[zi];
546 slgd->ZoneAry[zi] = z;
6ab8e1da
MD
547 if ((z->z_Flags & SLZF_UNOTZEROD) == 0)
548 flags &= ~M_ZERO; /* already zero'd */
1c5ca4f3
MD
549
550 /*
551 * Slide the base index for initial allocations out of the next
552 * zone we create so we do not over-weight the lower part of the
553 * cpu memory caches.
554 */
555 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
556 & (ZALLOC_MAX_ZONE_SIZE - 1);
a108bf71
MD
557 }
558done:
bba6a44d
MD
559 ++type->ks_inuse[gd->gd_cpuid];
560 type->ks_memuse[gd->gd_cpuid] += size;
561 type->ks_loosememuse += size;
a108bf71
MD
562 crit_exit();
563 if (flags & M_ZERO)
564 bzero(chunk, size);
bba6a44d
MD
565#ifdef INVARIANTS
566 else
567 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
568#endif
a108bf71
MD
569 return(chunk);
570fail:
571 crit_exit();
572 return(NULL);
573}
574
575void *
576realloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
577{
578 SLZone *z;
579 void *nptr;
580 unsigned long osize;
581
582 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
583 return(malloc(size, type, flags));
584 if (size == 0) {
585 free(ptr, type);
586 return(NULL);
587 }
588
589 /*
590 * Handle oversized allocations. XXX we really should require that a
591 * size be passed to free() instead of this nonsense.
592 */
593 {
594 struct kmemusage *kup;
595
596 kup = btokup(ptr);
597 if (kup->ku_pagecnt) {
598 osize = kup->ku_pagecnt << PAGE_SHIFT;
599 if (osize == round_page(size))
600 return(ptr);
601 if ((nptr = malloc(size, type, flags)) == NULL)
602 return(NULL);
603 bcopy(ptr, nptr, min(size, osize));
604 free(ptr, type);
605 return(nptr);
606 }
607 }
608
609 /*
610 * Get the original allocation's zone. If the new request winds up
611 * using the same chunk size we do not have to do anything.
612 */
613 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
614 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
615
616 zoneindex(&size);
617 if (z->z_ChunkSize == size)
618 return(ptr);
619
620 /*
621 * Allocate memory for the new request size. Note that zoneindex has
622 * already adjusted the request size to the appropriate chunk size, which
623 * should optimize our bcopy(). Then copy and return the new pointer.
624 */
625 if ((nptr = malloc(size, type, flags)) == NULL)
626 return(NULL);
627 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
628 free(ptr, type);
629 return(nptr);
630}
631
632/*
633 * free() (SLAB ALLOCATOR)
634 *
bba6a44d 635 * Free the specified chunk of memory.
a108bf71
MD
636 */
637static
638void
639free_remote(void *ptr)
640{
641 free(ptr, *(struct malloc_type **)ptr);
642}
643
644void
645free(void *ptr, struct malloc_type *type)
646{
647 SLZone *z;
648 SLChunk *chunk;
649 SLGlobalData *slgd;
bba6a44d 650 struct globaldata *gd;
a108bf71
MD
651 int pgno;
652
bba6a44d
MD
653 gd = mycpu;
654 slgd = &gd->gd_slab;
a108bf71
MD
655
656 /*
657 * Handle special 0-byte allocations
658 */
659 if (ptr == ZERO_LENGTH_PTR)
660 return;
661
662 /*
663 * Handle oversized allocations. XXX we really should require that a
664 * size be passed to free() instead of this nonsense.
bba6a44d
MD
665 *
666 * This code is never called via an ipi.
a108bf71
MD
667 */
668 {
669 struct kmemusage *kup;
670 unsigned long size;
671
672 kup = btokup(ptr);
673 if (kup->ku_pagecnt) {
674 size = kup->ku_pagecnt << PAGE_SHIFT;
675 kup->ku_pagecnt = 0;
a108bf71
MD
676#ifdef INVARIANTS
677 KKASSERT(sizeof(weirdary) <= size);
678 bcopy(weirdary, ptr, sizeof(weirdary));
679#endif
bba6a44d
MD
680 /*
681 * note: we always adjust our cpu's slot, not the originating
682 * cpu (kup->ku_cpuid). The statistics are in aggregate.
683 */
684 crit_enter();
685 --type->ks_inuse[gd->gd_cpuid];
686 type->ks_memuse[gd->gd_cpuid] -= size;
46a3f46d 687 if (mycpu->gd_intr_nesting_level) {
46a3f46d
MD
688 z = (SLZone *)ptr;
689 z->z_Magic = ZALLOC_OVSZ_MAGIC;
690 z->z_Next = slgd->FreeOvZones;
691 z->z_ChunkSize = size;
692 slgd->FreeOvZones = z;
693 crit_exit();
694 } else {
bba6a44d 695 crit_exit();
46a3f46d
MD
696 kmem_slab_free(ptr, size); /* may block */
697 }
a108bf71
MD
698 return;
699 }
700 }
701
702 /*
703 * Zone case. Figure out the zone based on the fact that it is
704 * ZoneSize aligned.
705 */
706 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
707 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
708
709 /*
710 * If we do not own the zone then forward the request to the
711 * cpu that does. The freeing code does not need the byte count
712 * unless DIAGNOSTIC is set.
713 */
bba6a44d 714 if (z->z_Cpu != gd->gd_cpuid) {
a108bf71 715 *(struct malloc_type **)ptr = type;
75c7ffea 716#ifdef SMP
a108bf71 717 lwkt_send_ipiq(z->z_Cpu, free_remote, ptr);
75c7ffea
MD
718#else
719 panic("Corrupt SLZone");
720#endif
a108bf71
MD
721 return;
722 }
723
724 if (type->ks_magic != M_MAGIC)
725 panic("free: malloc type lacks magic");
726
727 crit_enter();
728 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT;
729 chunk = ptr;
730
bba6a44d 731#ifdef INVARIANTS
a108bf71 732 /*
bba6a44d
MD
733 * Attempt to detect a double-free. To reduce overhead we only check
734 * if there appears to be link pointer at the base of the data.
a108bf71
MD
735 */
736 if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) {
737 SLChunk *scan;
738 for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) {
739 if (scan == chunk)
740 panic("Double free at %p", chunk);
741 }
742 }
743#endif
744
745 /*
746 * Put weird data into the memory to detect modifications after freeing,
747 * illegal pointer use after freeing (we should fault on the odd address),
748 * and so forth. XXX needs more work, see the old malloc code.
749 */
750#ifdef INVARIANTS
751 if (z->z_ChunkSize < sizeof(weirdary))
752 bcopy(weirdary, chunk, z->z_ChunkSize);
753 else
754 bcopy(weirdary, chunk, sizeof(weirdary));
755#endif
756
757 /*
758 * Add this free non-zero'd chunk to a linked list for reuse, adjust
759 * z_FirstFreePg.
760 */
6ab8e1da
MD
761#ifdef INVARIANTS
762 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
a108bf71 763 panic("BADFREE %p\n", chunk);
a108bf71
MD
764#endif
765 chunk->c_Next = z->z_PageAry[pgno];
766 z->z_PageAry[pgno] = chunk;
6ab8e1da
MD
767#ifdef INVARIANTS
768 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
a108bf71 769 panic("BADFREE2");
6ab8e1da 770#endif
a108bf71
MD
771 if (z->z_FirstFreePg > pgno)
772 z->z_FirstFreePg = pgno;
773
774 /*
775 * Bump the number of free chunks. If it becomes non-zero the zone
776 * must be added back onto the appropriate list.
777 */
778 if (z->z_NFree++ == 0) {
779 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
780 slgd->ZoneAry[z->z_ZoneIndex] = z;
781 }
782
bba6a44d
MD
783 --type->ks_inuse[z->z_Cpu];
784 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
a108bf71
MD
785
786 /*
787 * If the zone becomes totally free, and there are other zones we
a7cf0021
MD
788 * can allocate from, move this zone to the FreeZones list. Since
789 * this code can be called from an IPI callback, do *NOT* try to mess
790 * with kernel_map here. Hysteresis will be performed at malloc() time.
a108bf71
MD
791 */
792 if (z->z_NFree == z->z_NMax &&
793 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z)
794 ) {
795 SLZone **pz;
796
797 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
798 ;
799 *pz = z->z_Next;
800 z->z_Magic = -1;
a7cf0021
MD
801 z->z_Next = slgd->FreeZones;
802 slgd->FreeZones = z;
803 ++slgd->NFreeZones;
a108bf71
MD
804 }
805 crit_exit();
806}
807
808/*
809 * kmem_slab_alloc()
810 *
811 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
812 * specified alignment. M_* flags are expected in the flags field.
813 *
814 * Alignment must be a multiple of PAGE_SIZE.
815 *
816 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
817 * but when we move zalloc() over to use this function as its backend
818 * we will have to switch to kreserve/krelease and call reserve(0)
819 * after the new space is made available.
820 */
821static void *
822kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
823{
824 vm_size_t i;
825 vm_offset_t addr;
826 vm_offset_t offset;
827 int count;
828 vm_map_t map = kernel_map;
829
830 size = round_page(size);
831 addr = vm_map_min(map);
832
833 /*
834 * Reserve properly aligned space from kernel_map
835 */
836 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
837 crit_enter();
838 vm_map_lock(map);
839 if (vm_map_findspace(map, vm_map_min(map), size, align, &addr)) {
840 vm_map_unlock(map);
841 if ((flags & (M_NOWAIT|M_NULLOK)) == 0)
842 panic("kmem_slab_alloc(): kernel_map ran out of space!");
843 crit_exit();
844 vm_map_entry_release(count);
845 return(NULL);
846 }
847 offset = addr - VM_MIN_KERNEL_ADDRESS;
848 vm_object_reference(kernel_object);
849 vm_map_insert(map, &count,
850 kernel_object, offset, addr, addr + size,
851 VM_PROT_ALL, VM_PROT_ALL, 0);
852
853 /*
854 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
855 */
856 for (i = 0; i < size; i += PAGE_SIZE) {
857 vm_page_t m;
858 vm_pindex_t idx = OFF_TO_IDX(offset + i);
859 int zero = (flags & M_ZERO) ? VM_ALLOC_ZERO : 0;
860
861 if ((flags & (M_NOWAIT|M_USE_RESERVE)) == M_NOWAIT)
862 m = vm_page_alloc(kernel_object, idx, VM_ALLOC_INTERRUPT|zero);
863 else
864 m = vm_page_alloc(kernel_object, idx, VM_ALLOC_SYSTEM|zero);
865 if (m == NULL) {
866 if ((flags & M_NOWAIT) == 0) {
867 vm_map_unlock(map);
868 vm_wait();
869 vm_map_lock(map);
870 i -= PAGE_SIZE; /* retry */
871 continue;
872 }
873 while (i != 0) {
874 i -= PAGE_SIZE;
875 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
876 vm_page_free(m);
877 }
878 vm_map_delete(map, addr, addr + size, &count);
879 vm_map_unlock(map);
880 crit_exit();
881 vm_map_entry_release(count);
882 return(NULL);
883 }
884 }
885
886 /*
887 * Mark the map entry as non-pageable using a routine that allows us to
888 * populate the underlying pages.
889 */
890 vm_map_set_wired_quick(map, addr, size, &count);
891 crit_exit();
892
893 /*
894 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
895 */
896 for (i = 0; i < size; i += PAGE_SIZE) {
897 vm_page_t m;
898
899 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
900 m->valid = VM_PAGE_BITS_ALL;
901 vm_page_wire(m);
902 vm_page_wakeup(m);
903 pmap_enter(kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
904 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
905 bzero((char *)addr + i, PAGE_SIZE);
906 vm_page_flag_clear(m, PG_ZERO);
907 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_REFERENCED);
908 }
909 vm_map_unlock(map);
910 vm_map_entry_release(count);
911 return((void *)addr);
912}
913
914static void
915kmem_slab_free(void *ptr, vm_size_t size)
916{
917 crit_enter();
918 vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
919 crit_exit();
920}
921