Use KTR's built-in call chain recording rather then hacking it up ourselves.
[dragonfly.git] / sys / kern / kern_slaballoc.c
CommitLineData
a108bf71 1/*
38e34349 2 * KERN_SLABALLOC.C - Kernel SLAB memory allocator (MP SAFE)
8c10bfcf
MD
3 *
4 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
5 *
6 * This code is derived from software contributed to The DragonFly Project
7 * by Matthew Dillon <dillon@backplane.com>
8 *
a108bf71
MD
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
8c10bfcf 12 *
a108bf71
MD
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
18 * distribution.
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
22 *
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
a108bf71 34 * SUCH DAMAGE.
8c10bfcf 35 *
d2182dc1 36 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.32 2005/06/03 22:55:58 dillon Exp $
a108bf71
MD
37 *
38 * This module implements a slab allocator drop-in replacement for the
39 * kernel malloc().
40 *
41 * A slab allocator reserves a ZONE for each chunk size, then lays the
42 * chunks out in an array within the zone. Allocation and deallocation
43 * is nearly instantanious, and fragmentation/overhead losses are limited
44 * to a fixed worst-case amount.
45 *
46 * The downside of this slab implementation is in the chunk size
47 * multiplied by the number of zones. ~80 zones * 128K = 10MB of VM per cpu.
48 * In a kernel implementation all this memory will be physical so
49 * the zone size is adjusted downward on machines with less physical
50 * memory. The upside is that overhead is bounded... this is the *worst*
51 * case overhead.
52 *
53 * Slab management is done on a per-cpu basis and no locking or mutexes
54 * are required, only a critical section. When one cpu frees memory
55 * belonging to another cpu's slab manager an asynchronous IPI message
56 * will be queued to execute the operation. In addition, both the
57 * high level slab allocator and the low level zone allocator optimize
58 * M_ZERO requests, and the slab allocator does not have to pre initialize
59 * the linked list of chunks.
60 *
61 * XXX Balancing is needed between cpus. Balance will be handled through
62 * asynchronous IPIs primarily by reassigning the z_Cpu ownership of chunks.
63 *
64 * XXX If we have to allocate a new zone and M_USE_RESERVE is set, use of
65 * the new zone should be restricted to M_USE_RESERVE requests only.
66 *
67 * Alloc Size Chunking Number of zones
68 * 0-127 8 16
69 * 128-255 16 8
70 * 256-511 32 8
71 * 512-1023 64 8
72 * 1024-2047 128 8
73 * 2048-4095 256 8
74 * 4096-8191 512 8
75 * 8192-16383 1024 8
76 * 16384-32767 2048 8
77 * (if PAGE_SIZE is 4K the maximum zone allocation is 16383)
78 *
46a3f46d 79 * Allocations >= ZoneLimit go directly to kmem.
a108bf71
MD
80 *
81 * API REQUIREMENTS AND SIDE EFFECTS
82 *
83 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we
84 * have remained compatible with the following API requirements:
85 *
86 * + small power-of-2 sized allocations are power-of-2 aligned (kern_tty)
3d177b31 87 * + all power-of-2 sized allocations are power-of-2 aligned (twe)
a108bf71
MD
88 * + malloc(0) is allowed and returns non-NULL (ahc driver)
89 * + ability to allocate arbitrarily large chunks of memory
90 */
91
92#include "opt_vm.h"
93
a108bf71
MD
94#include <sys/param.h>
95#include <sys/systm.h>
96#include <sys/kernel.h>
97#include <sys/slaballoc.h>
98#include <sys/mbuf.h>
99#include <sys/vmmeter.h>
100#include <sys/lock.h>
101#include <sys/thread.h>
102#include <sys/globaldata.h>
d2182dc1 103#include <sys/sysctl.h>
a108bf71
MD
104
105#include <vm/vm.h>
106#include <vm/vm_param.h>
107#include <vm/vm_kern.h>
108#include <vm/vm_extern.h>
109#include <vm/vm_object.h>
110#include <vm/pmap.h>
111#include <vm/vm_map.h>
112#include <vm/vm_page.h>
113#include <vm/vm_pageout.h>
114
115#include <machine/cpu.h>
116
117#include <sys/thread2.h>
118
119#define arysize(ary) (sizeof(ary)/sizeof((ary)[0]))
120
121/*
122 * Fixed globals (not per-cpu)
123 */
124static int ZoneSize;
46a3f46d 125static int ZoneLimit;
a108bf71 126static int ZonePageCount;
a108bf71
MD
127static int ZoneMask;
128static struct malloc_type *kmemstatistics;
129static struct kmemusage *kmemusage;
130static int32_t weirdary[16];
131
132static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
133static void kmem_slab_free(void *ptr, vm_size_t bytes);
134
135/*
136 * Misc constants. Note that allocations that are exact multiples of
137 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
138 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
139 */
140#define MIN_CHUNK_SIZE 8 /* in bytes */
141#define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
142#define ZONE_RELS_THRESH 2 /* threshold number of zones */
143#define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
144
145/*
146 * The WEIRD_ADDR is used as known text to copy into free objects to
147 * try to create deterministic failure cases if the data is accessed after
148 * free.
149 */
150#define WEIRD_ADDR 0xdeadc0de
151#define MAX_COPY sizeof(weirdary)
152#define ZERO_LENGTH_PTR ((void *)-8)
153
154/*
155 * Misc global malloc buckets
156 */
157
158MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
159MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
160MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
161
162MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
163MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
164
165/*
166 * Initialize the slab memory allocator. We have to choose a zone size based
167 * on available physical memory. We choose a zone side which is approximately
168 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
169 * 128K. The zone size is limited to the bounds set in slaballoc.h
170 * (typically 32K min, 128K max).
171 */
172static void kmeminit(void *dummy);
173
174SYSINIT(kmem, SI_SUB_KMEM, SI_ORDER_FIRST, kmeminit, NULL)
175
d2182dc1
MD
176#ifdef INVARIANTS
177/*
178 * If enabled any memory allocated without M_ZERO is initialized to -1.
179 */
180static int use_malloc_pattern;
181SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
182 &use_malloc_pattern, 0, "");
183#endif
184
a108bf71
MD
185static void
186kmeminit(void *dummy)
187{
188 vm_poff_t limsize;
189 int usesize;
190 int i;
191 vm_pindex_t npg;
192
193 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
194 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS)
195 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
196
197 usesize = (int)(limsize / 1024); /* convert to KB */
198
199 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
200 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
201 ZoneSize <<= 1;
46a3f46d
MD
202 ZoneLimit = ZoneSize / 4;
203 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
204 ZoneLimit = ZALLOC_ZONE_LIMIT;
a108bf71 205 ZoneMask = ZoneSize - 1;
a108bf71
MD
206 ZonePageCount = ZoneSize / PAGE_SIZE;
207
208 npg = (VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE;
dc1fd4b3 209 kmemusage = kmem_slab_alloc(npg * sizeof(struct kmemusage), PAGE_SIZE, M_WAITOK|M_ZERO);
a108bf71
MD
210
211 for (i = 0; i < arysize(weirdary); ++i)
212 weirdary[i] = WEIRD_ADDR;
213
214 if (bootverbose)
215 printf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
216}
217
218/*
bba6a44d 219 * Initialize a malloc type tracking structure.
a108bf71
MD
220 */
221void
222malloc_init(void *data)
223{
224 struct malloc_type *type = data;
225 vm_poff_t limsize;
226
227 if (type->ks_magic != M_MAGIC)
228 panic("malloc type lacks magic");
229
230 if (type->ks_limit != 0)
231 return;
232
233 if (vmstats.v_page_count == 0)
234 panic("malloc_init not allowed before vm init");
235
236 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
237 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS)
238 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
239 type->ks_limit = limsize / 10;
240
241 type->ks_next = kmemstatistics;
242 kmemstatistics = type;
243}
244
245void
246malloc_uninit(void *data)
247{
248 struct malloc_type *type = data;
249 struct malloc_type *t;
bba6a44d
MD
250#ifdef INVARIANTS
251 int i;
1d712609 252 long ttl;
bba6a44d 253#endif
a108bf71
MD
254
255 if (type->ks_magic != M_MAGIC)
256 panic("malloc type lacks magic");
257
258 if (vmstats.v_page_count == 0)
259 panic("malloc_uninit not allowed before vm init");
260
261 if (type->ks_limit == 0)
262 panic("malloc_uninit on uninitialized type");
263
264#ifdef INVARIANTS
1d712609
MD
265 /*
266 * memuse is only correct in aggregation. Due to memory being allocated
267 * on one cpu and freed on another individual array entries may be
268 * negative or positive (canceling each other out).
269 */
270 for (i = ttl = 0; i < ncpus; ++i)
271 ttl += type->ks_memuse[i];
272 if (ttl) {
273 printf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
274 ttl, type->ks_shortdesc, i);
a108bf71
MD
275 }
276#endif
277 if (type == kmemstatistics) {
278 kmemstatistics = type->ks_next;
279 } else {
280 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
281 if (t->ks_next == type) {
282 t->ks_next = type->ks_next;
283 break;
284 }
285 }
286 }
287 type->ks_next = NULL;
288 type->ks_limit = 0;
289}
290
291/*
292 * Calculate the zone index for the allocation request size and set the
293 * allocation request size to that particular zone's chunk size.
294 */
295static __inline int
296zoneindex(unsigned long *bytes)
297{
298 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
299 if (n < 128) {
300 *bytes = n = (n + 7) & ~7;
301 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
302 }
303 if (n < 256) {
304 *bytes = n = (n + 15) & ~15;
305 return(n / 16 + 7);
306 }
307 if (n < 8192) {
308 if (n < 512) {
309 *bytes = n = (n + 31) & ~31;
310 return(n / 32 + 15);
311 }
312 if (n < 1024) {
313 *bytes = n = (n + 63) & ~63;
314 return(n / 64 + 23);
315 }
316 if (n < 2048) {
317 *bytes = n = (n + 127) & ~127;
318 return(n / 128 + 31);
319 }
320 if (n < 4096) {
321 *bytes = n = (n + 255) & ~255;
322 return(n / 256 + 39);
323 }
324 *bytes = n = (n + 511) & ~511;
325 return(n / 512 + 47);
326 }
327#if ZALLOC_ZONE_LIMIT > 8192
328 if (n < 16384) {
329 *bytes = n = (n + 1023) & ~1023;
330 return(n / 1024 + 55);
331 }
332#endif
333#if ZALLOC_ZONE_LIMIT > 16384
334 if (n < 32768) {
335 *bytes = n = (n + 2047) & ~2047;
336 return(n / 2048 + 63);
337 }
338#endif
339 panic("Unexpected byte count %d", n);
340 return(0);
341}
342
343/*
38e34349 344 * malloc() (SLAB ALLOCATOR) (MP SAFE)
a108bf71
MD
345 *
346 * Allocate memory via the slab allocator. If the request is too large,
347 * or if it page-aligned beyond a certain size, we fall back to the
348 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
349 * &SlabMisc if you don't care.
350 *
8cb2bf45
JS
351 * M_RNOWAIT - don't block.
352 * M_NULLOK - return NULL instead of blocking.
a108bf71 353 * M_ZERO - zero the returned memory.
dc1fd4b3
MD
354 * M_USE_RESERVE - allow greater drawdown of the free list
355 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
a108bf71
MD
356 */
357void *
358malloc(unsigned long size, struct malloc_type *type, int flags)
359{
360 SLZone *z;
361 SLChunk *chunk;
362 SLGlobalData *slgd;
bba6a44d 363 struct globaldata *gd;
a108bf71 364 int zi;
d2182dc1
MD
365#ifdef INVARIANTS
366 int i;
367#endif
a108bf71 368
bba6a44d
MD
369 gd = mycpu;
370 slgd = &gd->gd_slab;
a108bf71
MD
371
372 /*
373 * XXX silly to have this in the critical path.
374 */
375 if (type->ks_limit == 0) {
376 crit_enter();
377 if (type->ks_limit == 0)
378 malloc_init(type);
379 crit_exit();
380 }
381 ++type->ks_calls;
382
383 /*
38e34349
MD
384 * Handle the case where the limit is reached. Panic if we can't return
385 * NULL. The original malloc code looped, but this tended to
a108bf71 386 * simply deadlock the computer.
38e34349
MD
387 *
388 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
389 * to determine if a more complete limit check should be done. The
390 * actual memory use is tracked via ks_memuse[cpu].
a108bf71 391 */
bba6a44d
MD
392 while (type->ks_loosememuse >= type->ks_limit) {
393 int i;
394 long ttl;
395
396 for (i = ttl = 0; i < ncpus; ++i)
397 ttl += type->ks_memuse[i];
38e34349 398 type->ks_loosememuse = ttl; /* not MP synchronized */
bba6a44d 399 if (ttl >= type->ks_limit) {
8cb2bf45 400 if (flags & M_NULLOK)
bba6a44d
MD
401 return(NULL);
402 panic("%s: malloc limit exceeded", type->ks_shortdesc);
403 }
a108bf71
MD
404 }
405
406 /*
407 * Handle the degenerate size == 0 case. Yes, this does happen.
408 * Return a special pointer. This is to maintain compatibility with
409 * the original malloc implementation. Certain devices, such as the
410 * adaptec driver, not only allocate 0 bytes, they check for NULL and
411 * also realloc() later on. Joy.
412 */
413 if (size == 0)
414 return(ZERO_LENGTH_PTR);
415
a7cf0021
MD
416 /*
417 * Handle hysteresis from prior frees here in malloc(). We cannot
418 * safely manipulate the kernel_map in free() due to free() possibly
419 * being called via an IPI message or from sensitive interrupt code.
420 */
dc1fd4b3 421 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_RNOWAIT) == 0) {
46a3f46d
MD
422 crit_enter();
423 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */
424 z = slgd->FreeZones;
425 slgd->FreeZones = z->z_Next;
426 --slgd->NFreeZones;
427 kmem_slab_free(z, ZoneSize); /* may block */
428 }
429 crit_exit();
430 }
431 /*
432 * XXX handle oversized frees that were queued from free().
433 */
dc1fd4b3 434 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) {
46a3f46d
MD
435 crit_enter();
436 if ((z = slgd->FreeOvZones) != NULL) {
437 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
438 slgd->FreeOvZones = z->z_Next;
439 kmem_slab_free(z, z->z_ChunkSize); /* may block */
440 }
441 crit_exit();
a7cf0021
MD
442 }
443
a108bf71
MD
444 /*
445 * Handle large allocations directly. There should not be very many of
446 * these so performance is not a big issue.
447 *
448 * Guarentee page alignment for allocations in multiples of PAGE_SIZE
449 */
46a3f46d 450 if (size >= ZoneLimit || (size & PAGE_MASK) == 0) {
a108bf71
MD
451 struct kmemusage *kup;
452
453 size = round_page(size);
454 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
455 if (chunk == NULL)
456 return(NULL);
457 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
8f1d5415 458 flags |= M_PASSIVE_ZERO;
a108bf71
MD
459 kup = btokup(chunk);
460 kup->ku_pagecnt = size / PAGE_SIZE;
bba6a44d 461 kup->ku_cpu = gd->gd_cpuid;
a108bf71
MD
462 crit_enter();
463 goto done;
464 }
465
466 /*
467 * Attempt to allocate out of an existing zone. First try the free list,
468 * then allocate out of unallocated space. If we find a good zone move
469 * it to the head of the list so later allocations find it quickly
470 * (we might have thousands of zones in the list).
471 *
472 * Note: zoneindex() will panic of size is too large.
473 */
474 zi = zoneindex(&size);
475 KKASSERT(zi < NZONES);
476 crit_enter();
477 if ((z = slgd->ZoneAry[zi]) != NULL) {
478 KKASSERT(z->z_NFree > 0);
479
480 /*
481 * Remove us from the ZoneAry[] when we become empty
482 */
483 if (--z->z_NFree == 0) {
484 slgd->ZoneAry[zi] = z->z_Next;
485 z->z_Next = NULL;
486 }
487
488 /*
489 * Locate a chunk in a free page. This attempts to localize
490 * reallocations into earlier pages without us having to sort
491 * the chunk list. A chunk may still overlap a page boundary.
492 */
493 while (z->z_FirstFreePg < ZonePageCount) {
494 if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) {
495#ifdef DIAGNOSTIC
496 /*
497 * Diagnostic: c_Next is not total garbage.
498 */
499 KKASSERT(chunk->c_Next == NULL ||
500 ((intptr_t)chunk->c_Next & IN_SAME_PAGE_MASK) ==
501 ((intptr_t)chunk & IN_SAME_PAGE_MASK));
502#endif
6ab8e1da
MD
503#ifdef INVARIANTS
504 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
a108bf71 505 panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount);
6ab8e1da 506 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
a108bf71 507 panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount);
6ab8e1da 508#endif
a108bf71
MD
509 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next;
510 goto done;
511 }
512 ++z->z_FirstFreePg;
513 }
514
515 /*
1c5ca4f3
MD
516 * No chunks are available but NFree said we had some memory, so
517 * it must be available in the never-before-used-memory area
518 * governed by UIndex. The consequences are very serious if our zone
519 * got corrupted so we use an explicit panic rather then a KASSERT.
a108bf71 520 */
1c5ca4f3
MD
521 if (z->z_UIndex + 1 != z->z_NMax)
522 z->z_UIndex = z->z_UIndex + 1;
523 else
524 z->z_UIndex = 0;
525 if (z->z_UIndex == z->z_UEndIndex)
526 panic("slaballoc: corrupted zone");
527 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
8f1d5415 528 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 529 flags &= ~M_ZERO;
8f1d5415
MD
530 flags |= M_PASSIVE_ZERO;
531 }
a108bf71
MD
532 goto done;
533 }
534
535 /*
536 * If all zones are exhausted we need to allocate a new zone for this
537 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
6ab8e1da
MD
538 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
539 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
540 * we do not pre-zero it because we do not want to mess up the L1 cache.
a108bf71
MD
541 *
542 * At least one subsystem, the tty code (see CROUND) expects power-of-2
543 * allocations to be power-of-2 aligned. We maintain compatibility by
544 * adjusting the base offset below.
545 */
546 {
547 int off;
548
549 if ((z = slgd->FreeZones) != NULL) {
550 slgd->FreeZones = z->z_Next;
551 --slgd->NFreeZones;
552 bzero(z, sizeof(SLZone));
6ab8e1da 553 z->z_Flags |= SLZF_UNOTZEROD;
a108bf71
MD
554 } else {
555 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
556 if (z == NULL)
557 goto fail;
558 }
559
560 /*
561 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
562 * Otherwise just 8-byte align the data.
563 */
564 if ((size | (size - 1)) + 1 == (size << 1))
565 off = (sizeof(SLZone) + size - 1) & ~(size - 1);
566 else
567 off = (sizeof(SLZone) + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
568 z->z_Magic = ZALLOC_SLAB_MAGIC;
569 z->z_ZoneIndex = zi;
570 z->z_NMax = (ZoneSize - off) / size;
571 z->z_NFree = z->z_NMax - 1;
1c5ca4f3
MD
572 z->z_BasePtr = (char *)z + off;
573 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
a108bf71
MD
574 z->z_ChunkSize = size;
575 z->z_FirstFreePg = ZonePageCount;
2db3b277 576 z->z_CpuGd = gd;
bba6a44d 577 z->z_Cpu = gd->gd_cpuid;
1c5ca4f3 578 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
a108bf71
MD
579 z->z_Next = slgd->ZoneAry[zi];
580 slgd->ZoneAry[zi] = z;
8f1d5415 581 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 582 flags &= ~M_ZERO; /* already zero'd */
8f1d5415
MD
583 flags |= M_PASSIVE_ZERO;
584 }
1c5ca4f3
MD
585
586 /*
587 * Slide the base index for initial allocations out of the next
588 * zone we create so we do not over-weight the lower part of the
589 * cpu memory caches.
590 */
591 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
592 & (ZALLOC_MAX_ZONE_SIZE - 1);
a108bf71
MD
593 }
594done:
bba6a44d
MD
595 ++type->ks_inuse[gd->gd_cpuid];
596 type->ks_memuse[gd->gd_cpuid] += size;
38e34349 597 type->ks_loosememuse += size; /* not MP synchronized */
a108bf71
MD
598 crit_exit();
599 if (flags & M_ZERO)
600 bzero(chunk, size);
bba6a44d 601#ifdef INVARIANTS
d2182dc1
MD
602 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
603 if (use_malloc_pattern) {
604 for (i = 0; i < size; i += sizeof(int)) {
605 *(int *)((char *)chunk + i) = -1;
606 }
607 }
bba6a44d 608 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
d2182dc1 609 }
bba6a44d 610#endif
a108bf71
MD
611 return(chunk);
612fail:
613 crit_exit();
614 return(NULL);
615}
616
38e34349
MD
617/*
618 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
619 *
620 * Generally speaking this routine is not called very often and we do
621 * not attempt to optimize it beyond reusing the same pointer if the
622 * new size fits within the chunking of the old pointer's zone.
623 */
a108bf71
MD
624void *
625realloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
626{
627 SLZone *z;
628 void *nptr;
629 unsigned long osize;
630
eb7f3e3c
MD
631 KKASSERT((flags & M_ZERO) == 0); /* not supported */
632
a108bf71
MD
633 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
634 return(malloc(size, type, flags));
635 if (size == 0) {
636 free(ptr, type);
637 return(NULL);
638 }
639
640 /*
641 * Handle oversized allocations. XXX we really should require that a
642 * size be passed to free() instead of this nonsense.
643 */
644 {
645 struct kmemusage *kup;
646
647 kup = btokup(ptr);
648 if (kup->ku_pagecnt) {
649 osize = kup->ku_pagecnt << PAGE_SHIFT;
650 if (osize == round_page(size))
651 return(ptr);
652 if ((nptr = malloc(size, type, flags)) == NULL)
653 return(NULL);
654 bcopy(ptr, nptr, min(size, osize));
655 free(ptr, type);
656 return(nptr);
657 }
658 }
659
660 /*
661 * Get the original allocation's zone. If the new request winds up
662 * using the same chunk size we do not have to do anything.
663 */
664 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
665 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
666
667 zoneindex(&size);
668 if (z->z_ChunkSize == size)
669 return(ptr);
670
671 /*
672 * Allocate memory for the new request size. Note that zoneindex has
673 * already adjusted the request size to the appropriate chunk size, which
674 * should optimize our bcopy(). Then copy and return the new pointer.
675 */
676 if ((nptr = malloc(size, type, flags)) == NULL)
677 return(NULL);
678 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
679 free(ptr, type);
680 return(nptr);
681}
682
38e34349
MD
683/*
684 * Allocate a copy of the specified string.
685 *
686 * (MP SAFE) (MAY BLOCK)
687 */
1ac06773
MD
688char *
689strdup(const char *str, struct malloc_type *type)
690{
691 int zlen; /* length inclusive of terminating NUL */
692 char *nstr;
693
694 if (str == NULL)
695 return(NULL);
696 zlen = strlen(str) + 1;
697 nstr = malloc(zlen, type, M_WAITOK);
698 bcopy(str, nstr, zlen);
699 return(nstr);
700}
701
1d712609 702#ifdef SMP
a108bf71
MD
703/*
704 * free() (SLAB ALLOCATOR)
705 *
bba6a44d 706 * Free the specified chunk of memory.
a108bf71
MD
707 */
708static
709void
710free_remote(void *ptr)
711{
712 free(ptr, *(struct malloc_type **)ptr);
713}
714
1d712609
MD
715#endif
716
38e34349
MD
717/*
718 * free (SLAB ALLOCATOR) (MP SAFE)
719 *
720 * Free a memory block previously allocated by malloc. Note that we do not
721 * attempt to uplodate ks_loosememuse as MP races could prevent us from
722 * checking memory limits in malloc.
723 */
a108bf71
MD
724void
725free(void *ptr, struct malloc_type *type)
726{
727 SLZone *z;
728 SLChunk *chunk;
729 SLGlobalData *slgd;
bba6a44d 730 struct globaldata *gd;
a108bf71
MD
731 int pgno;
732
bba6a44d
MD
733 gd = mycpu;
734 slgd = &gd->gd_slab;
a108bf71 735
d39911d9
JS
736 if (ptr == NULL)
737 panic("trying to free NULL pointer");
738
a108bf71
MD
739 /*
740 * Handle special 0-byte allocations
741 */
742 if (ptr == ZERO_LENGTH_PTR)
743 return;
744
745 /*
746 * Handle oversized allocations. XXX we really should require that a
747 * size be passed to free() instead of this nonsense.
bba6a44d
MD
748 *
749 * This code is never called via an ipi.
a108bf71
MD
750 */
751 {
752 struct kmemusage *kup;
753 unsigned long size;
754
755 kup = btokup(ptr);
756 if (kup->ku_pagecnt) {
757 size = kup->ku_pagecnt << PAGE_SHIFT;
758 kup->ku_pagecnt = 0;
a108bf71
MD
759#ifdef INVARIANTS
760 KKASSERT(sizeof(weirdary) <= size);
761 bcopy(weirdary, ptr, sizeof(weirdary));
762#endif
bba6a44d
MD
763 /*
764 * note: we always adjust our cpu's slot, not the originating
765 * cpu (kup->ku_cpuid). The statistics are in aggregate.
81f5fc99
MD
766 *
767 * note: XXX we have still inherited the interrupts-can't-block
768 * assumption. An interrupt thread does not bump
769 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
770 * primarily until we can fix softupdate's assumptions about free().
bba6a44d
MD
771 */
772 crit_enter();
773 --type->ks_inuse[gd->gd_cpuid];
774 type->ks_memuse[gd->gd_cpuid] -= size;
81f5fc99 775 if (mycpu->gd_intr_nesting_level || (gd->gd_curthread->td_flags & TDF_INTTHREAD)) {
46a3f46d
MD
776 z = (SLZone *)ptr;
777 z->z_Magic = ZALLOC_OVSZ_MAGIC;
778 z->z_Next = slgd->FreeOvZones;
779 z->z_ChunkSize = size;
780 slgd->FreeOvZones = z;
781 crit_exit();
782 } else {
bba6a44d 783 crit_exit();
46a3f46d
MD
784 kmem_slab_free(ptr, size); /* may block */
785 }
a108bf71
MD
786 return;
787 }
788 }
789
790 /*
791 * Zone case. Figure out the zone based on the fact that it is
792 * ZoneSize aligned.
793 */
794 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
795 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
796
797 /*
798 * If we do not own the zone then forward the request to the
4c9f5a7f
MD
799 * cpu that does. Since the timing is non-critical, a passive
800 * message is sent.
a108bf71 801 */
2db3b277 802 if (z->z_CpuGd != gd) {
a108bf71 803 *(struct malloc_type **)ptr = type;
75c7ffea 804#ifdef SMP
4c9f5a7f 805 lwkt_send_ipiq_passive(z->z_CpuGd, free_remote, ptr);
75c7ffea
MD
806#else
807 panic("Corrupt SLZone");
808#endif
a108bf71
MD
809 return;
810 }
811
812 if (type->ks_magic != M_MAGIC)
813 panic("free: malloc type lacks magic");
814
815 crit_enter();
816 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT;
817 chunk = ptr;
818
bba6a44d 819#ifdef INVARIANTS
a108bf71 820 /*
bba6a44d
MD
821 * Attempt to detect a double-free. To reduce overhead we only check
822 * if there appears to be link pointer at the base of the data.
a108bf71
MD
823 */
824 if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) {
825 SLChunk *scan;
826 for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) {
827 if (scan == chunk)
828 panic("Double free at %p", chunk);
829 }
830 }
831#endif
832
833 /*
834 * Put weird data into the memory to detect modifications after freeing,
835 * illegal pointer use after freeing (we should fault on the odd address),
836 * and so forth. XXX needs more work, see the old malloc code.
837 */
838#ifdef INVARIANTS
839 if (z->z_ChunkSize < sizeof(weirdary))
840 bcopy(weirdary, chunk, z->z_ChunkSize);
841 else
842 bcopy(weirdary, chunk, sizeof(weirdary));
843#endif
844
845 /*
846 * Add this free non-zero'd chunk to a linked list for reuse, adjust
847 * z_FirstFreePg.
848 */
6ab8e1da
MD
849#ifdef INVARIANTS
850 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS)
fc92d4aa 851 panic("BADFREE %p", chunk);
a108bf71
MD
852#endif
853 chunk->c_Next = z->z_PageAry[pgno];
854 z->z_PageAry[pgno] = chunk;
6ab8e1da
MD
855#ifdef INVARIANTS
856 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS)
a108bf71 857 panic("BADFREE2");
6ab8e1da 858#endif
a108bf71
MD
859 if (z->z_FirstFreePg > pgno)
860 z->z_FirstFreePg = pgno;
861
862 /*
863 * Bump the number of free chunks. If it becomes non-zero the zone
864 * must be added back onto the appropriate list.
865 */
866 if (z->z_NFree++ == 0) {
867 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
868 slgd->ZoneAry[z->z_ZoneIndex] = z;
869 }
870
bba6a44d
MD
871 --type->ks_inuse[z->z_Cpu];
872 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
a108bf71
MD
873
874 /*
875 * If the zone becomes totally free, and there are other zones we
a7cf0021
MD
876 * can allocate from, move this zone to the FreeZones list. Since
877 * this code can be called from an IPI callback, do *NOT* try to mess
878 * with kernel_map here. Hysteresis will be performed at malloc() time.
a108bf71
MD
879 */
880 if (z->z_NFree == z->z_NMax &&
881 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z)
882 ) {
883 SLZone **pz;
884
885 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
886 ;
887 *pz = z->z_Next;
888 z->z_Magic = -1;
a7cf0021
MD
889 z->z_Next = slgd->FreeZones;
890 slgd->FreeZones = z;
891 ++slgd->NFreeZones;
a108bf71
MD
892 }
893 crit_exit();
894}
895
896/*
38e34349 897 * kmem_slab_alloc() (MP SAFE) (GETS BGL)
a108bf71
MD
898 *
899 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
900 * specified alignment. M_* flags are expected in the flags field.
901 *
902 * Alignment must be a multiple of PAGE_SIZE.
903 *
904 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
905 * but when we move zalloc() over to use this function as its backend
906 * we will have to switch to kreserve/krelease and call reserve(0)
907 * after the new space is made available.
dc1fd4b3
MD
908 *
909 * Interrupt code which has preempted other code is not allowed to
c397c465
MD
910 * use PQ_CACHE pages. However, if an interrupt thread is run
911 * non-preemptively or blocks and then runs non-preemptively, then
912 * it is free to use PQ_CACHE pages.
38e34349
MD
913 *
914 * This routine will currently obtain the BGL.
a108bf71
MD
915 */
916static void *
917kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
918{
919 vm_size_t i;
920 vm_offset_t addr;
921 vm_offset_t offset;
1de1e800 922 int count, vmflags, base_vmflags;
dc1fd4b3 923 thread_t td;
a108bf71
MD
924 vm_map_t map = kernel_map;
925
926 size = round_page(size);
927 addr = vm_map_min(map);
928
929 /*
5c39c498
MD
930 * Reserve properly aligned space from kernel_map. RNOWAIT allocations
931 * cannot block.
a108bf71 932 */
5c39c498
MD
933 if (flags & M_RNOWAIT) {
934 if (try_mplock() == 0)
935 return(NULL);
936 } else {
937 get_mplock();
938 }
a108bf71
MD
939 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
940 crit_enter();
941 vm_map_lock(map);
942 if (vm_map_findspace(map, vm_map_min(map), size, align, &addr)) {
943 vm_map_unlock(map);
8cb2bf45 944 if ((flags & M_NULLOK) == 0)
a108bf71
MD
945 panic("kmem_slab_alloc(): kernel_map ran out of space!");
946 crit_exit();
947 vm_map_entry_release(count);
38e34349 948 rel_mplock();
a108bf71
MD
949 return(NULL);
950 }
951 offset = addr - VM_MIN_KERNEL_ADDRESS;
952 vm_object_reference(kernel_object);
953 vm_map_insert(map, &count,
954 kernel_object, offset, addr, addr + size,
955 VM_PROT_ALL, VM_PROT_ALL, 0);
956
dc1fd4b3 957 td = curthread;
dc1fd4b3 958
1de1e800
JS
959 base_vmflags = 0;
960 if (flags & M_ZERO)
961 base_vmflags |= VM_ALLOC_ZERO;
962 if (flags & M_USE_RESERVE)
963 base_vmflags |= VM_ALLOC_SYSTEM;
964 if (flags & M_USE_INTERRUPT_RESERVE)
965 base_vmflags |= VM_ALLOC_INTERRUPT;
966 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0)
967 panic("kmem_slab_alloc: bad flags %08x (%p)", flags, ((int **)&size)[-1]);
968
969
a108bf71
MD
970 /*
971 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
972 */
973 for (i = 0; i < size; i += PAGE_SIZE) {
974 vm_page_t m;
975 vm_pindex_t idx = OFF_TO_IDX(offset + i);
fe1e98d0
MD
976
977 /*
c397c465
MD
978 * VM_ALLOC_NORMAL can only be set if we are not preempting.
979 *
980 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
981 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
982 * implied in this case), though I'm sure if we really need to do
983 * that.
fe1e98d0 984 */
1de1e800 985 vmflags = base_vmflags;
c397c465 986 if (flags & M_WAITOK) {
1de1e800 987 if (td->td_preempted)
fe1e98d0 988 vmflags |= VM_ALLOC_SYSTEM;
1de1e800 989 else
dc1fd4b3 990 vmflags |= VM_ALLOC_NORMAL;
dc1fd4b3 991 }
a108bf71 992
dc1fd4b3
MD
993 m = vm_page_alloc(kernel_object, idx, vmflags);
994
995 /*
996 * If the allocation failed we either return NULL or we retry.
997 *
c397c465
MD
998 * If M_WAITOK is specified we wait for more memory and retry.
999 * If M_WAITOK is specified from a preemption we yield instead of
1000 * wait. Livelock will not occur because the interrupt thread
1001 * will not be preempting anyone the second time around after the
1002 * yield.
dc1fd4b3 1003 */
a108bf71 1004 if (m == NULL) {
c397c465 1005 if (flags & M_WAITOK) {
fe1e98d0 1006 if (td->td_preempted) {
dc1fd4b3
MD
1007 vm_map_unlock(map);
1008 lwkt_yield();
1009 vm_map_lock(map);
1010 } else {
dc1fd4b3
MD
1011 vm_map_unlock(map);
1012 vm_wait();
1013 vm_map_lock(map);
1014 }
a108bf71
MD
1015 i -= PAGE_SIZE; /* retry */
1016 continue;
1017 }
dc1fd4b3
MD
1018
1019 /*
1020 * We were unable to recover, cleanup and return NULL
1021 */
a108bf71
MD
1022 while (i != 0) {
1023 i -= PAGE_SIZE;
1024 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
1025 vm_page_free(m);
1026 }
1027 vm_map_delete(map, addr, addr + size, &count);
1028 vm_map_unlock(map);
1029 crit_exit();
1030 vm_map_entry_release(count);
38e34349 1031 rel_mplock();
a108bf71
MD
1032 return(NULL);
1033 }
1034 }
1035
1036 /*
dc1fd4b3
MD
1037 * Success!
1038 *
a108bf71
MD
1039 * Mark the map entry as non-pageable using a routine that allows us to
1040 * populate the underlying pages.
1041 */
1042 vm_map_set_wired_quick(map, addr, size, &count);
1043 crit_exit();
1044
1045 /*
1046 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1047 */
1048 for (i = 0; i < size; i += PAGE_SIZE) {
1049 vm_page_t m;
1050
1051 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i));
1052 m->valid = VM_PAGE_BITS_ALL;
1053 vm_page_wire(m);
1054 vm_page_wakeup(m);
1055 pmap_enter(kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
1056 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1057 bzero((char *)addr + i, PAGE_SIZE);
1058 vm_page_flag_clear(m, PG_ZERO);
1059 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_REFERENCED);
1060 }
1061 vm_map_unlock(map);
1062 vm_map_entry_release(count);
38e34349 1063 rel_mplock();
a108bf71
MD
1064 return((void *)addr);
1065}
1066
38e34349
MD
1067/*
1068 * kmem_slab_free() (MP SAFE) (GETS BGL)
1069 */
a108bf71
MD
1070static void
1071kmem_slab_free(void *ptr, vm_size_t size)
1072{
38e34349 1073 get_mplock();
a108bf71
MD
1074 crit_enter();
1075 vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
1076 crit_exit();
38e34349 1077 rel_mplock();
a108bf71
MD
1078}
1079