SLAB ALLOCATOR Stage 1. This brings in a slab allocator written from scratch
[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 *
28 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.1 2003/08/27 01:43:07 dillon Exp $
29 *
30 * This module implements a slab allocator drop-in replacement for the
31 * kernel malloc().
32 *
33 * A slab allocator reserves a ZONE for each chunk size, then lays the
34 * chunks out in an array within the zone. Allocation and deallocation
35 * is nearly instantanious, and fragmentation/overhead losses are limited
36 * to a fixed worst-case amount.
37 *
38 * The downside of this slab implementation is in the chunk size
39 * multiplied by the number of zones. ~80 zones * 128K = 10MB of VM per cpu.
40 * In a kernel implementation all this memory will be physical so
41 * the zone size is adjusted downward on machines with less physical
42 * memory. The upside is that overhead is bounded... this is the *worst*
43 * case overhead.
44 *
45 * Slab management is done on a per-cpu basis and no locking or mutexes
46 * are required, only a critical section. When one cpu frees memory
47 * belonging to another cpu's slab manager an asynchronous IPI message
48 * will be queued to execute the operation. In addition, both the
49 * high level slab allocator and the low level zone allocator optimize
50 * M_ZERO requests, and the slab allocator does not have to pre initialize
51 * the linked list of chunks.
52 *
53 * XXX Balancing is needed between cpus. Balance will be handled through
54 * asynchronous IPIs primarily by reassigning the z_Cpu ownership of chunks.
55 *
56 * XXX If we have to allocate a new zone and M_USE_RESERVE is set, use of
57 * the new zone should be restricted to M_USE_RESERVE requests only.
58 *
59 * Alloc Size Chunking Number of zones
60 * 0-127 8 16
61 * 128-255 16 8
62 * 256-511 32 8
63 * 512-1023 64 8
64 * 1024-2047 128 8
65 * 2048-4095 256 8
66 * 4096-8191 512 8
67 * 8192-16383 1024 8
68 * 16384-32767 2048 8
69 * (if PAGE_SIZE is 4K the maximum zone allocation is 16383)
70 *
71 * Allocations >= ZALLOC_ZONE_LIMIT go directly to kmem.
72 *
73 * API REQUIREMENTS AND SIDE EFFECTS
74 *
75 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we
76 * have remained compatible with the following API requirements:
77 *
78 * + small power-of-2 sized allocations are power-of-2 aligned (kern_tty)
79 * + malloc(0) is allowed and returns non-NULL (ahc driver)
80 * + ability to allocate arbitrarily large chunks of memory
81 */
82
83#include "opt_vm.h"
84
85#if defined(USE_SLAB_ALLOCATOR)
86
87#if !defined(NO_KMEM_MAP)
88#error "NO_KMEM_MAP must be defined when USE_SLAB_ALLOCATOR is defined"
89#endif
90
91#include <sys/param.h>
92#include <sys/systm.h>
93#include <sys/kernel.h>
94#include <sys/slaballoc.h>
95#include <sys/mbuf.h>
96#include <sys/vmmeter.h>
97#include <sys/lock.h>
98#include <sys/thread.h>
99#include <sys/globaldata.h>
100
101#include <vm/vm.h>
102#include <vm/vm_param.h>
103#include <vm/vm_kern.h>
104#include <vm/vm_extern.h>
105#include <vm/vm_object.h>
106#include <vm/pmap.h>
107#include <vm/vm_map.h>
108#include <vm/vm_page.h>
109#include <vm/vm_pageout.h>
110
111#include <machine/cpu.h>
112
113#include <sys/thread2.h>
114
115#define arysize(ary) (sizeof(ary)/sizeof((ary)[0]))
116
117/*
118 * Fixed globals (not per-cpu)
119 */
120static int ZoneSize;
121static int ZonePageCount;
122static int ZonePageLimit;
123static int ZoneMask;
124static struct malloc_type *kmemstatistics;
125static struct kmemusage *kmemusage;
126static int32_t weirdary[16];
127
128static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
129static void kmem_slab_free(void *ptr, vm_size_t bytes);
130
131/*
132 * Misc constants. Note that allocations that are exact multiples of
133 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
134 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
135 */
136#define MIN_CHUNK_SIZE 8 /* in bytes */
137#define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
138#define ZONE_RELS_THRESH 2 /* threshold number of zones */
139#define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
140
141/*
142 * The WEIRD_ADDR is used as known text to copy into free objects to
143 * try to create deterministic failure cases if the data is accessed after
144 * free.
145 */
146#define WEIRD_ADDR 0xdeadc0de
147#define MAX_COPY sizeof(weirdary)
148#define ZERO_LENGTH_PTR ((void *)-8)
149
150/*
151 * Misc global malloc buckets
152 */
153
154MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
155MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
156MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
157
158MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
159MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
160
161/*
162 * Initialize the slab memory allocator. We have to choose a zone size based
163 * on available physical memory. We choose a zone side which is approximately
164 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
165 * 128K. The zone size is limited to the bounds set in slaballoc.h
166 * (typically 32K min, 128K max).
167 */
168static void kmeminit(void *dummy);
169
170SYSINIT(kmem, SI_SUB_KMEM, SI_ORDER_FIRST, kmeminit, NULL)
171
172static void
173kmeminit(void *dummy)
174{
175 vm_poff_t limsize;
176 int usesize;
177 int i;
178 vm_pindex_t npg;
179
180 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
181 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS)
182 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
183
184 usesize = (int)(limsize / 1024); /* convert to KB */
185
186 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
187 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
188 ZoneSize <<= 1;
189 ZoneMask = ZoneSize - 1;
190 ZonePageLimit = PAGE_SIZE * 4;
191 ZonePageCount = ZoneSize / PAGE_SIZE;
192
193 npg = (VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE;
194 kmemusage = kmem_slab_alloc(npg * sizeof(struct kmemusage), PAGE_SIZE, M_ZERO);
195
196 for (i = 0; i < arysize(weirdary); ++i)
197 weirdary[i] = WEIRD_ADDR;
198
199 if (bootverbose)
200 printf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
201}
202
203/*
204 * Initialize a malloc type tracking structure. NOTE! counters and such
205 * need to be made per-cpu (maybe with a MAXCPU array).
206 */
207void
208malloc_init(void *data)
209{
210 struct malloc_type *type = data;
211 vm_poff_t limsize;
212
213 if (type->ks_magic != M_MAGIC)
214 panic("malloc type lacks magic");
215
216 if (type->ks_limit != 0)
217 return;
218
219 if (vmstats.v_page_count == 0)
220 panic("malloc_init not allowed before vm init");
221
222 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
223 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS)
224 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
225 type->ks_limit = limsize / 10;
226
227 type->ks_next = kmemstatistics;
228 kmemstatistics = type;
229}
230
231void
232malloc_uninit(void *data)
233{
234 struct malloc_type *type = data;
235 struct malloc_type *t;
236
237 if (type->ks_magic != M_MAGIC)
238 panic("malloc type lacks magic");
239
240 if (vmstats.v_page_count == 0)
241 panic("malloc_uninit not allowed before vm init");
242
243 if (type->ks_limit == 0)
244 panic("malloc_uninit on uninitialized type");
245
246#ifdef INVARIANTS
247 if (type->ks_memuse != 0) {
248 printf("malloc_uninit: %ld bytes of '%s' still allocated\n",
249 type->ks_memuse, type->ks_shortdesc);
250 }
251#endif
252 if (type == kmemstatistics) {
253 kmemstatistics = type->ks_next;
254 } else {
255 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
256 if (t->ks_next == type) {
257 t->ks_next = type->ks_next;
258 break;
259 }
260 }
261 }
262 type->ks_next = NULL;
263 type->ks_limit = 0;
264}
265
266/*
267 * Calculate the zone index for the allocation request size and set the
268 * allocation request size to that particular zone's chunk size.
269 */
270static __inline int
271zoneindex(unsigned long *bytes)
272{
273 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
274 if (n < 128) {
275 *bytes = n = (n + 7) & ~7;
276 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
277 }
278 if (n < 256) {
279 *bytes = n = (n + 15) & ~15;
280 return(n / 16 + 7);
281 }
282 if (n < 8192) {
283 if (n < 512) {
284 *bytes = n = (n + 31) & ~31;
285 return(n / 32 + 15);
286 }
287 if (n < 1024) {
288 *bytes = n = (n + 63) & ~63;
289 return(n / 64 + 23);
290 }
291 if (n < 2048) {
292 *bytes = n = (n + 127) & ~127;
293 return(n / 128 + 31);
294 }
295 if (n < 4096) {
296 *bytes = n = (n + 255) & ~255;
297 return(n / 256 + 39);
298 }
299 *bytes = n = (n + 511) & ~511;
300 return(n / 512 + 47);
301 }
302#if ZALLOC_ZONE_LIMIT > 8192
303 if (n < 16384) {
304 *bytes = n = (n + 1023) & ~1023;
305 return(n / 1024 + 55);
306 }
307#endif
308#if ZALLOC_ZONE_LIMIT > 16384
309 if (n < 32768) {
310 *bytes = n = (n + 2047) & ~2047;
311 return(n / 2048 + 63);
312 }
313#endif
314 panic("Unexpected byte count %d", n);
315 return(0);
316}
317
318/*
319 * malloc() (SLAB ALLOCATOR)
320 *
321 * Allocate memory via the slab allocator. If the request is too large,
322 * or if it page-aligned beyond a certain size, we fall back to the
323 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
324 * &SlabMisc if you don't care.
325 *
326 * M_NOWAIT - return NULL instead of blocking.
327 * M_ZERO - zero the returned memory.
328 * M_USE_RESERVE - allocate out of the system reserve if necessary
329 */
330void *
331malloc(unsigned long size, struct malloc_type *type, int flags)
332{
333 SLZone *z;
334 SLChunk *chunk;
335 SLGlobalData *slgd;
336 int zi;
337
338 slgd = &mycpu->gd_slab;
339
340 /*
341 * XXX silly to have this in the critical path.
342 */
343 if (type->ks_limit == 0) {
344 crit_enter();
345 if (type->ks_limit == 0)
346 malloc_init(type);
347 crit_exit();
348 }
349 ++type->ks_calls;
350
351 /*
352 * Handle the case where the limit is reached. Panic if can't return
353 * NULL. XXX the original malloc code looped, but this tended to
354 * simply deadlock the computer.
355 */
356 while (type->ks_memuse >= type->ks_limit) {
357 if (flags & (M_NOWAIT|M_NULLOK))
358 return(NULL);
359 panic("%s: malloc limit exceeded", type->ks_shortdesc);
360 }
361
362 /*
363 * Handle the degenerate size == 0 case. Yes, this does happen.
364 * Return a special pointer. This is to maintain compatibility with
365 * the original malloc implementation. Certain devices, such as the
366 * adaptec driver, not only allocate 0 bytes, they check for NULL and
367 * also realloc() later on. Joy.
368 */
369 if (size == 0)
370 return(ZERO_LENGTH_PTR);
371
372 /*
373 * Handle large allocations directly. There should not be very many of
374 * these so performance is not a big issue.
375 *
376 * Guarentee page alignment for allocations in multiples of PAGE_SIZE
377 */
378 if (size >= ZALLOC_ZONE_LIMIT || (size & PAGE_MASK) == 0) {
379 struct kmemusage *kup;
380
381 size = round_page(size);
382 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
383 if (chunk == NULL)
384 return(NULL);
385 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
386 kup = btokup(chunk);
387 kup->ku_pagecnt = size / PAGE_SIZE;
388 crit_enter();
389 goto done;
390 }
391
392 /*
393 * Attempt to allocate out of an existing zone. First try the free list,
394 * then allocate out of unallocated space. If we find a good zone move
395 * it to the head of the list so later allocations find it quickly
396 * (we might have thousands of zones in the list).
397 *
398 * Note: zoneindex() will panic of size is too large.
399 */
400 zi = zoneindex(&size);
401 KKASSERT(zi < NZONES);
402 crit_enter();
403 if ((z = slgd->ZoneAry[zi]) != NULL) {
404 KKASSERT(z->z_NFree > 0);
405
406 /*
407 * Remove us from the ZoneAry[] when we become empty
408 */
409 if (--z->z_NFree == 0) {
410 slgd->ZoneAry[zi] = z->z_Next;
411 z->z_Next = NULL;
412 }
413
414 /*
415 * Locate a chunk in a free page. This attempts to localize
416 * reallocations into earlier pages without us having to sort
417 * the chunk list. A chunk may still overlap a page boundary.
418 */
419 while (z->z_FirstFreePg < ZonePageCount) {
420 if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) {
421#ifdef DIAGNOSTIC
422 /*
423 * Diagnostic: c_Next is not total garbage.
424 */
425 KKASSERT(chunk->c_Next == NULL ||
426 ((intptr_t)chunk->c_Next & IN_SAME_PAGE_MASK) ==
427 ((intptr_t)chunk & IN_SAME_PAGE_MASK));
428#endif
429 if ((uintptr_t)chunk < 0xC0000000U)
430 panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount);
431 if (chunk->c_Next && (uintptr_t)chunk->c_Next < 0xC0000000U)
432 panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount);
433 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next;
434 goto done;
435 }
436 ++z->z_FirstFreePg;
437 }
438
439 /*
440 * Never before used memory is available at the UAlloc. This
441 * memory has already been zero'd.
442 */
443 chunk = (SLChunk *)((char *)z + z->z_UAlloc);
444 z->z_UAlloc += size;
445 KKASSERT(z->z_UAlloc <= ZoneSize);
446 flags &= ~M_ZERO;
447 goto done;
448 }
449
450 /*
451 * If all zones are exhausted we need to allocate a new zone for this
452 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
453 * UAlloc use above in regards to M_ZERO.
454 *
455 * At least one subsystem, the tty code (see CROUND) expects power-of-2
456 * allocations to be power-of-2 aligned. We maintain compatibility by
457 * adjusting the base offset below.
458 */
459 {
460 int off;
461
462 if ((z = slgd->FreeZones) != NULL) {
463 slgd->FreeZones = z->z_Next;
464 --slgd->NFreeZones;
465 bzero(z, sizeof(SLZone));
466 } else {
467 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
468 if (z == NULL)
469 goto fail;
470 }
471
472 /*
473 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
474 * Otherwise just 8-byte align the data.
475 */
476 if ((size | (size - 1)) + 1 == (size << 1))
477 off = (sizeof(SLZone) + size - 1) & ~(size - 1);
478 else
479 off = (sizeof(SLZone) + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
480 z->z_Magic = ZALLOC_SLAB_MAGIC;
481 z->z_ZoneIndex = zi;
482 z->z_NMax = (ZoneSize - off) / size;
483 z->z_NFree = z->z_NMax - 1;
484 z->z_UAlloc = off + size;
485 z->z_ChunkSize = size;
486 z->z_FirstFreePg = ZonePageCount;
487 z->z_Cpu = mycpu->gd_cpuid;
488 chunk = (SLChunk *)((char *)z + off);
489 z->z_Next = slgd->ZoneAry[zi];
490 slgd->ZoneAry[zi] = z;
491 flags &= ~M_ZERO; /* already zero'd */
492 }
493done:
494 crit_exit();
495 if (flags & M_ZERO)
496 bzero(chunk, size);
497 ++type->ks_inuse;
498 type->ks_memuse += size;
499 return(chunk);
500fail:
501 crit_exit();
502 return(NULL);
503}
504
505void *
506realloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
507{
508 SLZone *z;
509 void *nptr;
510 unsigned long osize;
511
512 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
513 return(malloc(size, type, flags));
514 if (size == 0) {
515 free(ptr, type);
516 return(NULL);
517 }
518
519 /*
520 * Handle oversized allocations. XXX we really should require that a
521 * size be passed to free() instead of this nonsense.
522 */
523 {
524 struct kmemusage *kup;
525
526 kup = btokup(ptr);
527 if (kup->ku_pagecnt) {
528 osize = kup->ku_pagecnt << PAGE_SHIFT;
529 if (osize == round_page(size))
530 return(ptr);
531 if ((nptr = malloc(size, type, flags)) == NULL)
532 return(NULL);
533 bcopy(ptr, nptr, min(size, osize));
534 free(ptr, type);
535 return(nptr);
536 }
537 }
538
539 /*
540 * Get the original allocation's zone. If the new request winds up
541 * using the same chunk size we do not have to do anything.
542 */
543 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
544 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
545
546 zoneindex(&size);
547 if (z->z_ChunkSize == size)
548 return(ptr);
549
550 /*
551 * Allocate memory for the new request size. Note that zoneindex has
552 * already adjusted the request size to the appropriate chunk size, which
553 * should optimize our bcopy(). Then copy and return the new pointer.
554 */
555 if ((nptr = malloc(size, type, flags)) == NULL)
556 return(NULL);
557 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
558 free(ptr, type);
559 return(nptr);
560}
561
562/*
563 * free() (SLAB ALLOCATOR)
564 *
565 * Free the specified chunk of memory. The byte count is not strictly
566 * required but if DIAGNOSTIC is set we use it as a sanity check.
567 */
568static
569void
570free_remote(void *ptr)
571{
572 free(ptr, *(struct malloc_type **)ptr);
573}
574
575void
576free(void *ptr, struct malloc_type *type)
577{
578 SLZone *z;
579 SLChunk *chunk;
580 SLGlobalData *slgd;
581 int pgno;
582
583 slgd = &mycpu->gd_slab;
584
585 /*
586 * Handle special 0-byte allocations
587 */
588 if (ptr == ZERO_LENGTH_PTR)
589 return;
590
591 /*
592 * Handle oversized allocations. XXX we really should require that a
593 * size be passed to free() instead of this nonsense.
594 */
595 {
596 struct kmemusage *kup;
597 unsigned long size;
598
599 kup = btokup(ptr);
600 if (kup->ku_pagecnt) {
601 size = kup->ku_pagecnt << PAGE_SHIFT;
602 kup->ku_pagecnt = 0;
603 --type->ks_inuse;
604 type->ks_memuse -= size;
605#ifdef INVARIANTS
606 KKASSERT(sizeof(weirdary) <= size);
607 bcopy(weirdary, ptr, sizeof(weirdary));
608#endif
609 kmem_slab_free(ptr, size); /* may block */
610 return;
611 }
612 }
613
614 /*
615 * Zone case. Figure out the zone based on the fact that it is
616 * ZoneSize aligned.
617 */
618 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
619 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
620
621 /*
622 * If we do not own the zone then forward the request to the
623 * cpu that does. The freeing code does not need the byte count
624 * unless DIAGNOSTIC is set.
625 */
626 if (z->z_Cpu != mycpu->gd_cpuid) {
627 *(struct malloc_type **)ptr = type;
628 lwkt_send_ipiq(z->z_Cpu, free_remote, ptr);
629 return;
630 }
631
632 if (type->ks_magic != M_MAGIC)
633 panic("free: malloc type lacks magic");
634
635 crit_enter();
636 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT;
637 chunk = ptr;
638
639#ifdef DIAGNOSTIC
640 /*
641 * Diagnostic: attempt to detect a double-free (not perfect).
642 */
643 if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) {
644 SLChunk *scan;
645 for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) {
646 if (scan == chunk)
647 panic("Double free at %p", chunk);
648 }
649 }
650#endif
651
652 /*
653 * Put weird data into the memory to detect modifications after freeing,
654 * illegal pointer use after freeing (we should fault on the odd address),
655 * and so forth. XXX needs more work, see the old malloc code.
656 */
657#ifdef INVARIANTS
658 if (z->z_ChunkSize < sizeof(weirdary))
659 bcopy(weirdary, chunk, z->z_ChunkSize);
660 else
661 bcopy(weirdary, chunk, sizeof(weirdary));
662#endif
663
664 /*
665 * Add this free non-zero'd chunk to a linked list for reuse, adjust
666 * z_FirstFreePg.
667 */
668 if ((uintptr_t)chunk < 0xC0000000U)
669 panic("BADFREE %p\n", chunk);
670#if 0
671 if (type->ks_inuse == 34 && type->ks_memuse == 600 && (uint32_t)ptr == 0xc11600b8)
672 Debugger("Freeing");
673#endif
674 chunk->c_Next = z->z_PageAry[pgno];
675 z->z_PageAry[pgno] = chunk;
676 if (chunk->c_Next && (uintptr_t)chunk->c_Next < 0xC0000000U)
677 panic("BADFREE2");
678 if (z->z_FirstFreePg > pgno)
679 z->z_FirstFreePg = pgno;
680
681 /*
682 * Bump the number of free chunks. If it becomes non-zero the zone
683 * must be added back onto the appropriate list.
684 */
685 if (z->z_NFree++ == 0) {
686 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
687 slgd->ZoneAry[z->z_ZoneIndex] = z;
688 }
689
690 --type->ks_inuse;
691 type->ks_memuse -= z->z_ChunkSize;
692
693 /*
694 * If the zone becomes totally free, and there are other zones we
695 * can allocate from, move this zone to the FreeZones list. Implement
696 * hysteresis on the FreeZones list to improve performance.
697 *
698 * XXX try not to block on the kernel_map lock.
699 */
700 if (z->z_NFree == z->z_NMax &&
701 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z)
702 ) {
703 SLZone **pz;
704
705 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
706 ;
707 *pz = z->z_Next;
708 z->z_Magic = -1;
709 if (slgd->NFreeZones == ZONE_RELS_THRESH &&
710 lockstatus(&kernel_map->lock, NULL) == 0) {
711 SLZone *oz;
712
713 z->z_Next = slgd->FreeZones->z_Next;
714 oz = slgd->FreeZones;
715 slgd->FreeZones = z;
716 kmem_slab_free(oz, ZoneSize); /* may block */
717 } else {
718 z->z_Next = slgd->FreeZones;
719 slgd->FreeZones = z;
720 ++slgd->NFreeZones;
721 }
722 }
723 crit_exit();
724}
725
726/*
727 * kmem_slab_alloc()
728 *
729 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
730 * specified alignment. M_* flags are expected in the flags field.
731 *
732 * Alignment must be a multiple of PAGE_SIZE.
733 *
734 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
735 * but when we move zalloc() over to use this function as its backend
736 * we will have to switch to kreserve/krelease and call reserve(0)
737 * after the new space is made available.
738 */
739static void *
740kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
741{
742 vm_size_t i;
743 vm_offset_t addr;
744 vm_offset_t offset;
745 int count;
746 vm_map_t map = kernel_map;
747
748 size = round_page(size);
749 addr = vm_map_min(map);
750
751 /*
752 * Reserve properly aligned space from kernel_map
753 */
754 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
755 crit_enter();
756 vm_map_lock(map);
757 if (vm_map_findspace(map, vm_map_min(map), size, align, &addr)) {
758 vm_map_unlock(map);
759 if ((flags & (M_NOWAIT|M_NULLOK)) == 0)
760 panic("kmem_slab_alloc(): kernel_map ran out of space!");
761 crit_exit();
762 vm_map_entry_release(count);
763 return(NULL);
764 }
765 offset = addr - VM_MIN_KERNEL_ADDRESS;
766 vm_object_reference(kernel_object);
767 vm_map_insert(map, &count,
768 kernel_object, offset, addr, addr + size,
769 VM_PROT_ALL, VM_PROT_ALL, 0);
770
771 /*
772 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
773 */
774 for (i = 0; i < size; i += PAGE_SIZE) {
775 vm_page_t m;
776 vm_pindex_t idx = OFF_TO_IDX(offset + i);
777 int zero = (flags & M_ZERO) ? VM_ALLOC_ZERO : 0;
778
779 if ((flags & (M_NOWAIT|M_USE_RESERVE)) == M_NOWAIT)
780 m = vm_page_alloc(kernel_object, idx, VM_ALLOC_INTERRUPT|zero);
781 else
782 m = vm_page_alloc(kernel_object, idx, VM_ALLOC_SYSTEM|zero);
783 if (m == NULL) {
784 if ((flags & M_NOWAIT) == 0) {
785 vm_map_unlock(map);
786 vm_wait();
787 vm_map_lock(map);
788 i -= PAGE_SIZE; /* retry */
789 continue;
790 }
791 while (i != 0) {
792 i -= PAGE_SIZE;
793 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
794 vm_page_free(m);
795 }
796 vm_map_delete(map, addr, addr + size, &count);
797 vm_map_unlock(map);
798 crit_exit();
799 vm_map_entry_release(count);
800 return(NULL);
801 }
802 }
803
804 /*
805 * Mark the map entry as non-pageable using a routine that allows us to
806 * populate the underlying pages.
807 */
808 vm_map_set_wired_quick(map, addr, size, &count);
809 crit_exit();
810
811 /*
812 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
813 */
814 for (i = 0; i < size; i += PAGE_SIZE) {
815 vm_page_t m;
816
817 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
818 m->valid = VM_PAGE_BITS_ALL;
819 vm_page_wire(m);
820 vm_page_wakeup(m);
821 pmap_enter(kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
822 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
823 bzero((char *)addr + i, PAGE_SIZE);
824 vm_page_flag_clear(m, PG_ZERO);
825 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_REFERENCED);
826 }
827 vm_map_unlock(map);
828 vm_map_entry_release(count);
829 return((void *)addr);
830}
831
832static void
833kmem_slab_free(void *ptr, vm_size_t size)
834{
835 crit_enter();
836 vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
837 crit_exit();
838}
839
840#endif