Fix the case when `-p' option of truss can be passed the pid of
[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 *
6ab8e1da 28 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.2 2003/08/27 07:00:27 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 *
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
6ab8e1da
MD
429#ifdef INVARIANTS
430 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
a108bf71 431 panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount);
6ab8e1da 432 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
a108bf71 433 panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount);
6ab8e1da 434#endif
a108bf71
MD
435 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next;
436 goto done;
437 }
438 ++z->z_FirstFreePg;
439 }
440
441 /*
442 * Never before used memory is available at the UAlloc. This
6ab8e1da 443 * memory may already have been zero'd.
a108bf71
MD
444 */
445 chunk = (SLChunk *)((char *)z + z->z_UAlloc);
446 z->z_UAlloc += size;
447 KKASSERT(z->z_UAlloc <= ZoneSize);
6ab8e1da
MD
448 if ((z->z_Flags & SLZF_UNOTZEROD) == 0)
449 flags &= ~M_ZERO;
a108bf71
MD
450 goto done;
451 }
452
453 /*
454 * If all zones are exhausted we need to allocate a new zone for this
455 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
6ab8e1da
MD
456 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
457 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
458 * we do not pre-zero it because we do not want to mess up the L1 cache.
a108bf71
MD
459 *
460 * At least one subsystem, the tty code (see CROUND) expects power-of-2
461 * allocations to be power-of-2 aligned. We maintain compatibility by
462 * adjusting the base offset below.
463 */
464 {
465 int off;
466
467 if ((z = slgd->FreeZones) != NULL) {
468 slgd->FreeZones = z->z_Next;
469 --slgd->NFreeZones;
470 bzero(z, sizeof(SLZone));
6ab8e1da 471 z->z_Flags |= SLZF_UNOTZEROD;
a108bf71
MD
472 } else {
473 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
474 if (z == NULL)
475 goto fail;
476 }
477
478 /*
479 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
480 * Otherwise just 8-byte align the data.
481 */
482 if ((size | (size - 1)) + 1 == (size << 1))
483 off = (sizeof(SLZone) + size - 1) & ~(size - 1);
484 else
485 off = (sizeof(SLZone) + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
486 z->z_Magic = ZALLOC_SLAB_MAGIC;
487 z->z_ZoneIndex = zi;
488 z->z_NMax = (ZoneSize - off) / size;
489 z->z_NFree = z->z_NMax - 1;
490 z->z_UAlloc = off + size;
491 z->z_ChunkSize = size;
492 z->z_FirstFreePg = ZonePageCount;
493 z->z_Cpu = mycpu->gd_cpuid;
494 chunk = (SLChunk *)((char *)z + off);
495 z->z_Next = slgd->ZoneAry[zi];
496 slgd->ZoneAry[zi] = z;
6ab8e1da
MD
497 if ((z->z_Flags & SLZF_UNOTZEROD) == 0)
498 flags &= ~M_ZERO; /* already zero'd */
a108bf71
MD
499 }
500done:
501 crit_exit();
502 if (flags & M_ZERO)
503 bzero(chunk, size);
504 ++type->ks_inuse;
505 type->ks_memuse += size;
506 return(chunk);
507fail:
508 crit_exit();
509 return(NULL);
510}
511
512void *
513realloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
514{
515 SLZone *z;
516 void *nptr;
517 unsigned long osize;
518
519 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
520 return(malloc(size, type, flags));
521 if (size == 0) {
522 free(ptr, type);
523 return(NULL);
524 }
525
526 /*
527 * Handle oversized allocations. XXX we really should require that a
528 * size be passed to free() instead of this nonsense.
529 */
530 {
531 struct kmemusage *kup;
532
533 kup = btokup(ptr);
534 if (kup->ku_pagecnt) {
535 osize = kup->ku_pagecnt << PAGE_SHIFT;
536 if (osize == round_page(size))
537 return(ptr);
538 if ((nptr = malloc(size, type, flags)) == NULL)
539 return(NULL);
540 bcopy(ptr, nptr, min(size, osize));
541 free(ptr, type);
542 return(nptr);
543 }
544 }
545
546 /*
547 * Get the original allocation's zone. If the new request winds up
548 * using the same chunk size we do not have to do anything.
549 */
550 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
551 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
552
553 zoneindex(&size);
554 if (z->z_ChunkSize == size)
555 return(ptr);
556
557 /*
558 * Allocate memory for the new request size. Note that zoneindex has
559 * already adjusted the request size to the appropriate chunk size, which
560 * should optimize our bcopy(). Then copy and return the new pointer.
561 */
562 if ((nptr = malloc(size, type, flags)) == NULL)
563 return(NULL);
564 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
565 free(ptr, type);
566 return(nptr);
567}
568
569/*
570 * free() (SLAB ALLOCATOR)
571 *
572 * Free the specified chunk of memory. The byte count is not strictly
573 * required but if DIAGNOSTIC is set we use it as a sanity check.
574 */
575static
576void
577free_remote(void *ptr)
578{
579 free(ptr, *(struct malloc_type **)ptr);
580}
581
582void
583free(void *ptr, struct malloc_type *type)
584{
585 SLZone *z;
586 SLChunk *chunk;
587 SLGlobalData *slgd;
588 int pgno;
589
590 slgd = &mycpu->gd_slab;
591
592 /*
593 * Handle special 0-byte allocations
594 */
595 if (ptr == ZERO_LENGTH_PTR)
596 return;
597
598 /*
599 * Handle oversized allocations. XXX we really should require that a
600 * size be passed to free() instead of this nonsense.
601 */
602 {
603 struct kmemusage *kup;
604 unsigned long size;
605
606 kup = btokup(ptr);
607 if (kup->ku_pagecnt) {
608 size = kup->ku_pagecnt << PAGE_SHIFT;
609 kup->ku_pagecnt = 0;
610 --type->ks_inuse;
611 type->ks_memuse -= size;
612#ifdef INVARIANTS
613 KKASSERT(sizeof(weirdary) <= size);
614 bcopy(weirdary, ptr, sizeof(weirdary));
615#endif
616 kmem_slab_free(ptr, size); /* may block */
617 return;
618 }
619 }
620
621 /*
622 * Zone case. Figure out the zone based on the fact that it is
623 * ZoneSize aligned.
624 */
625 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
626 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
627
628 /*
629 * If we do not own the zone then forward the request to the
630 * cpu that does. The freeing code does not need the byte count
631 * unless DIAGNOSTIC is set.
632 */
633 if (z->z_Cpu != mycpu->gd_cpuid) {
634 *(struct malloc_type **)ptr = type;
635 lwkt_send_ipiq(z->z_Cpu, free_remote, ptr);
636 return;
637 }
638
639 if (type->ks_magic != M_MAGIC)
640 panic("free: malloc type lacks magic");
641
642 crit_enter();
643 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT;
644 chunk = ptr;
645
646#ifdef DIAGNOSTIC
647 /*
648 * Diagnostic: attempt to detect a double-free (not perfect).
649 */
650 if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) {
651 SLChunk *scan;
652 for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) {
653 if (scan == chunk)
654 panic("Double free at %p", chunk);
655 }
656 }
657#endif
658
659 /*
660 * Put weird data into the memory to detect modifications after freeing,
661 * illegal pointer use after freeing (we should fault on the odd address),
662 * and so forth. XXX needs more work, see the old malloc code.
663 */
664#ifdef INVARIANTS
665 if (z->z_ChunkSize < sizeof(weirdary))
666 bcopy(weirdary, chunk, z->z_ChunkSize);
667 else
668 bcopy(weirdary, chunk, sizeof(weirdary));
669#endif
670
671 /*
672 * Add this free non-zero'd chunk to a linked list for reuse, adjust
673 * z_FirstFreePg.
674 */
6ab8e1da
MD
675#ifdef INVARIANTS
676 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
a108bf71 677 panic("BADFREE %p\n", chunk);
a108bf71
MD
678#endif
679 chunk->c_Next = z->z_PageAry[pgno];
680 z->z_PageAry[pgno] = chunk;
6ab8e1da
MD
681#ifdef INVARIANTS
682 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
a108bf71 683 panic("BADFREE2");
6ab8e1da 684#endif
a108bf71
MD
685 if (z->z_FirstFreePg > pgno)
686 z->z_FirstFreePg = pgno;
687
688 /*
689 * Bump the number of free chunks. If it becomes non-zero the zone
690 * must be added back onto the appropriate list.
691 */
692 if (z->z_NFree++ == 0) {
693 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
694 slgd->ZoneAry[z->z_ZoneIndex] = z;
695 }
696
697 --type->ks_inuse;
698 type->ks_memuse -= z->z_ChunkSize;
699
700 /*
701 * If the zone becomes totally free, and there are other zones we
702 * can allocate from, move this zone to the FreeZones list. Implement
703 * hysteresis on the FreeZones list to improve performance.
704 *
705 * XXX try not to block on the kernel_map lock.
706 */
707 if (z->z_NFree == z->z_NMax &&
708 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z)
709 ) {
710 SLZone **pz;
711
712 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
713 ;
714 *pz = z->z_Next;
715 z->z_Magic = -1;
716 if (slgd->NFreeZones == ZONE_RELS_THRESH &&
717 lockstatus(&kernel_map->lock, NULL) == 0) {
718 SLZone *oz;
719
720 z->z_Next = slgd->FreeZones->z_Next;
721 oz = slgd->FreeZones;
722 slgd->FreeZones = z;
723 kmem_slab_free(oz, ZoneSize); /* may block */
724 } else {
725 z->z_Next = slgd->FreeZones;
726 slgd->FreeZones = z;
727 ++slgd->NFreeZones;
728 }
729 }
730 crit_exit();
731}
732
733/*
734 * kmem_slab_alloc()
735 *
736 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
737 * specified alignment. M_* flags are expected in the flags field.
738 *
739 * Alignment must be a multiple of PAGE_SIZE.
740 *
741 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
742 * but when we move zalloc() over to use this function as its backend
743 * we will have to switch to kreserve/krelease and call reserve(0)
744 * after the new space is made available.
745 */
746static void *
747kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
748{
749 vm_size_t i;
750 vm_offset_t addr;
751 vm_offset_t offset;
752 int count;
753 vm_map_t map = kernel_map;
754
755 size = round_page(size);
756 addr = vm_map_min(map);
757
758 /*
759 * Reserve properly aligned space from kernel_map
760 */
761 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
762 crit_enter();
763 vm_map_lock(map);
764 if (vm_map_findspace(map, vm_map_min(map), size, align, &addr)) {
765 vm_map_unlock(map);
766 if ((flags & (M_NOWAIT|M_NULLOK)) == 0)
767 panic("kmem_slab_alloc(): kernel_map ran out of space!");
768 crit_exit();
769 vm_map_entry_release(count);
770 return(NULL);
771 }
772 offset = addr - VM_MIN_KERNEL_ADDRESS;
773 vm_object_reference(kernel_object);
774 vm_map_insert(map, &count,
775 kernel_object, offset, addr, addr + size,
776 VM_PROT_ALL, VM_PROT_ALL, 0);
777
778 /*
779 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
780 */
781 for (i = 0; i < size; i += PAGE_SIZE) {
782 vm_page_t m;
783 vm_pindex_t idx = OFF_TO_IDX(offset + i);
784 int zero = (flags & M_ZERO) ? VM_ALLOC_ZERO : 0;
785
786 if ((flags & (M_NOWAIT|M_USE_RESERVE)) == M_NOWAIT)
787 m = vm_page_alloc(kernel_object, idx, VM_ALLOC_INTERRUPT|zero);
788 else
789 m = vm_page_alloc(kernel_object, idx, VM_ALLOC_SYSTEM|zero);
790 if (m == NULL) {
791 if ((flags & M_NOWAIT) == 0) {
792 vm_map_unlock(map);
793 vm_wait();
794 vm_map_lock(map);
795 i -= PAGE_SIZE; /* retry */
796 continue;
797 }
798 while (i != 0) {
799 i -= PAGE_SIZE;
800 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
801 vm_page_free(m);
802 }
803 vm_map_delete(map, addr, addr + size, &count);
804 vm_map_unlock(map);
805 crit_exit();
806 vm_map_entry_release(count);
807 return(NULL);
808 }
809 }
810
811 /*
812 * Mark the map entry as non-pageable using a routine that allows us to
813 * populate the underlying pages.
814 */
815 vm_map_set_wired_quick(map, addr, size, &count);
816 crit_exit();
817
818 /*
819 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
820 */
821 for (i = 0; i < size; i += PAGE_SIZE) {
822 vm_page_t m;
823
824 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
825 m->valid = VM_PAGE_BITS_ALL;
826 vm_page_wire(m);
827 vm_page_wakeup(m);
828 pmap_enter(kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
829 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
830 bzero((char *)addr + i, PAGE_SIZE);
831 vm_page_flag_clear(m, PG_ZERO);
832 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_REFERENCED);
833 }
834 vm_map_unlock(map);
835 vm_map_entry_release(count);
836 return((void *)addr);
837}
838
839static void
840kmem_slab_free(void *ptr, vm_size_t size)
841{
842 crit_enter();
843 vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
844 crit_exit();
845}
846
847#endif