kmalloc: Add comment about alignment property
[dragonfly.git] / sys / kern / kern_slaballoc.c
CommitLineData
a108bf71 1/*
ed2013d8
VS
2 * (MPSAFE)
3 *
5b287bba 4 * KERN_SLABALLOC.C - Kernel SLAB memory allocator
8c10bfcf 5 *
ed2013d8 6 * Copyright (c) 2003,2004,2010 The DragonFly Project. All rights reserved.
8c10bfcf
MD
7 *
8 * This code is derived from software contributed to The DragonFly Project
9 * by Matthew Dillon <dillon@backplane.com>
10 *
a108bf71
MD
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
13 * are met:
8c10bfcf 14 *
a108bf71
MD
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
18 * notice, this list of conditions and the following disclaimer in
19 * the documentation and/or other materials provided with the
20 * distribution.
21 * 3. Neither the name of The DragonFly Project nor the names of its
22 * contributors may be used to endorse or promote products derived
23 * from this software without specific, prior written permission.
24 *
25 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
26 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
27 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
28 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
29 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
30 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
31 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
32 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
33 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
34 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
35 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
a108bf71 36 * SUCH DAMAGE.
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 80 *
1f3dc3c2
SZ
81 * Alignment properties:
82 * - All power-of-2 sized allocations are power-of-2 aligned.
83 * - Allocations with M_POWEROF2 are power-of-2 aligned on the nearest
84 * power-of-2 round up of 'size'.
85 * - Non-power-of-2 sized allocations are zone chunk size aligned (see the
86 * above table 'Chunking' column).
87 *
a108bf71
MD
88 * API REQUIREMENTS AND SIDE EFFECTS
89 *
90 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we
91 * have remained compatible with the following API requirements:
92 *
a108bf71
MD
93 * + malloc(0) is allowed and returns non-NULL (ahc driver)
94 * + ability to allocate arbitrarily large chunks of memory
95 */
96
97#include "opt_vm.h"
98
a108bf71
MD
99#include <sys/param.h>
100#include <sys/systm.h>
101#include <sys/kernel.h>
102#include <sys/slaballoc.h>
103#include <sys/mbuf.h>
104#include <sys/vmmeter.h>
105#include <sys/lock.h>
106#include <sys/thread.h>
107#include <sys/globaldata.h>
d2182dc1 108#include <sys/sysctl.h>
f2b5daf9 109#include <sys/ktr.h>
a108bf71
MD
110
111#include <vm/vm.h>
112#include <vm/vm_param.h>
113#include <vm/vm_kern.h>
114#include <vm/vm_extern.h>
115#include <vm/vm_object.h>
116#include <vm/pmap.h>
117#include <vm/vm_map.h>
118#include <vm/vm_page.h>
119#include <vm/vm_pageout.h>
120
121#include <machine/cpu.h>
122
123#include <sys/thread2.h>
124
722871d3
MD
125#define btokup(z) (&pmap_kvtom((vm_offset_t)(z))->ku_pagecnt)
126
5bf48697
AE
127#define MEMORY_STRING "ptr=%p type=%p size=%lu flags=%04x"
128#define MEMORY_ARGS void *ptr, void *type, unsigned long size, int flags
f2b5daf9
MD
129
130#if !defined(KTR_MEMORY)
131#define KTR_MEMORY KTR_ALL
132#endif
133KTR_INFO_MASTER(memory);
5bf48697
AE
134KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin");
135KTR_INFO(KTR_MEMORY, memory, malloc_end, 1, MEMORY_STRING, MEMORY_ARGS);
136KTR_INFO(KTR_MEMORY, memory, free_zero, 2, MEMORY_STRING, MEMORY_ARGS);
137KTR_INFO(KTR_MEMORY, memory, free_ovsz, 3, MEMORY_STRING, MEMORY_ARGS);
138KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 4, MEMORY_STRING, MEMORY_ARGS);
139KTR_INFO(KTR_MEMORY, memory, free_chunk, 5, MEMORY_STRING, MEMORY_ARGS);
f2b5daf9 140#ifdef SMP
5bf48697
AE
141KTR_INFO(KTR_MEMORY, memory, free_request, 6, MEMORY_STRING, MEMORY_ARGS);
142KTR_INFO(KTR_MEMORY, memory, free_rem_beg, 7, MEMORY_STRING, MEMORY_ARGS);
143KTR_INFO(KTR_MEMORY, memory, free_rem_end, 8, MEMORY_STRING, MEMORY_ARGS);
f2b5daf9 144#endif
5bf48697
AE
145KTR_INFO(KTR_MEMORY, memory, free_beg, 9, "free begin");
146KTR_INFO(KTR_MEMORY, memory, free_end, 10, "free end");
f2b5daf9
MD
147
148#define logmemory(name, ptr, type, size, flags) \
149 KTR_LOG(memory_ ## name, ptr, type, size, flags)
b68ad50c
MD
150#define logmemory_quick(name) \
151 KTR_LOG(memory_ ## name)
f2b5daf9 152
a108bf71
MD
153/*
154 * Fixed globals (not per-cpu)
155 */
156static int ZoneSize;
46a3f46d 157static int ZoneLimit;
a108bf71 158static int ZonePageCount;
5fee07e6 159static uintptr_t ZoneMask;
665206ee
MD
160static int ZoneBigAlloc; /* in KB */
161static int ZoneGenAlloc; /* in KB */
460426e6 162struct malloc_type *kmemstatistics; /* exported to vmstat */
a108bf71
MD
163static int32_t weirdary[16];
164
165static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
166static void kmem_slab_free(void *ptr, vm_size_t bytes);
5fee07e6 167
10cc6608
MD
168#if defined(INVARIANTS)
169static void chunk_mark_allocated(SLZone *z, void *chunk);
170static void chunk_mark_free(SLZone *z, void *chunk);
5fee07e6
MD
171#else
172#define chunk_mark_allocated(z, chunk)
173#define chunk_mark_free(z, chunk)
10cc6608 174#endif
a108bf71
MD
175
176/*
177 * Misc constants. Note that allocations that are exact multiples of
178 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
a108bf71 179 */
ad94a851 180#define ZONE_RELS_THRESH 32 /* threshold number of zones */
a108bf71
MD
181
182/*
183 * The WEIRD_ADDR is used as known text to copy into free objects to
184 * try to create deterministic failure cases if the data is accessed after
185 * free.
186 */
187#define WEIRD_ADDR 0xdeadc0de
188#define MAX_COPY sizeof(weirdary)
189#define ZERO_LENGTH_PTR ((void *)-8)
190
191/*
192 * Misc global malloc buckets
193 */
194
195MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
196MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
197MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
198
199MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
200MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
201
202/*
203 * Initialize the slab memory allocator. We have to choose a zone size based
204 * on available physical memory. We choose a zone side which is approximately
205 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
206 * 128K. The zone size is limited to the bounds set in slaballoc.h
207 * (typically 32K min, 128K max).
208 */
209static void kmeminit(void *dummy);
210
c7841cbe
MD
211char *ZeroPage;
212
ba39e2e0 213SYSINIT(kmem, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL)
a108bf71 214
d2182dc1
MD
215#ifdef INVARIANTS
216/*
217 * If enabled any memory allocated without M_ZERO is initialized to -1.
218 */
219static int use_malloc_pattern;
220SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
0c52fa62
SG
221 &use_malloc_pattern, 0,
222 "Initialize memory to -1 if M_ZERO not specified");
d2182dc1
MD
223#endif
224
ad94a851 225static int ZoneRelsThresh = ZONE_RELS_THRESH;
665206ee
MD
226SYSCTL_INT(_kern, OID_AUTO, zone_big_alloc, CTLFLAG_RD, &ZoneBigAlloc, 0, "");
227SYSCTL_INT(_kern, OID_AUTO, zone_gen_alloc, CTLFLAG_RD, &ZoneGenAlloc, 0, "");
ad94a851 228SYSCTL_INT(_kern, OID_AUTO, zone_cache, CTLFLAG_RW, &ZoneRelsThresh, 0, "");
665206ee 229
b12defdc
MD
230/*
231 * Returns the kernel memory size limit for the purposes of initializing
232 * various subsystem caches. The smaller of available memory and the KVM
233 * memory space is returned.
234 *
235 * The size in megabytes is returned.
236 */
237size_t
238kmem_lim_size(void)
239{
240 size_t limsize;
241
242 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
243 if (limsize > KvaSize)
244 limsize = KvaSize;
245 return (limsize / (1024 * 1024));
246}
247
a108bf71
MD
248static void
249kmeminit(void *dummy)
250{
7c457ac8 251 size_t limsize;
a108bf71
MD
252 int usesize;
253 int i;
a108bf71 254
b12defdc
MD
255 limsize = kmem_lim_size();
256 usesize = (int)(limsize * 1024); /* convert to KB */
a108bf71 257
b12defdc
MD
258 /*
259 * If the machine has a large KVM space and more than 8G of ram,
260 * double the zone release threshold to reduce SMP invalidations.
261 * If more than 16G of ram, do it again.
262 *
263 * The BIOS eats a little ram so add some slop. We want 8G worth of
264 * memory sticks to trigger the first adjustment.
265 */
266 if (ZoneRelsThresh == ZONE_RELS_THRESH) {
267 if (limsize >= 7 * 1024)
268 ZoneRelsThresh *= 2;
269 if (limsize >= 15 * 1024)
270 ZoneRelsThresh *= 2;
271 }
a108bf71 272
b12defdc
MD
273 /*
274 * Calculate the zone size. This typically calculates to
275 * ZALLOC_MAX_ZONE_SIZE
276 */
a108bf71
MD
277 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
278 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
279 ZoneSize <<= 1;
46a3f46d
MD
280 ZoneLimit = ZoneSize / 4;
281 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
282 ZoneLimit = ZALLOC_ZONE_LIMIT;
5fee07e6 283 ZoneMask = ~(uintptr_t)(ZoneSize - 1);
a108bf71
MD
284 ZonePageCount = ZoneSize / PAGE_SIZE;
285
a3034532 286 for (i = 0; i < NELEM(weirdary); ++i)
a108bf71
MD
287 weirdary[i] = WEIRD_ADDR;
288
c7841cbe
MD
289 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
290
a108bf71 291 if (bootverbose)
6ea70f76 292 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
a108bf71
MD
293}
294
295/*
bba6a44d 296 * Initialize a malloc type tracking structure.
a108bf71
MD
297 */
298void
299malloc_init(void *data)
300{
301 struct malloc_type *type = data;
7c457ac8 302 size_t limsize;
a108bf71
MD
303
304 if (type->ks_magic != M_MAGIC)
305 panic("malloc type lacks magic");
306
307 if (type->ks_limit != 0)
308 return;
309
310 if (vmstats.v_page_count == 0)
311 panic("malloc_init not allowed before vm init");
312
b12defdc 313 limsize = kmem_lim_size() * (1024 * 1024);
a108bf71
MD
314 type->ks_limit = limsize / 10;
315
316 type->ks_next = kmemstatistics;
317 kmemstatistics = type;
318}
319
320void
321malloc_uninit(void *data)
322{
323 struct malloc_type *type = data;
324 struct malloc_type *t;
bba6a44d
MD
325#ifdef INVARIANTS
326 int i;
1d712609 327 long ttl;
bba6a44d 328#endif
a108bf71
MD
329
330 if (type->ks_magic != M_MAGIC)
331 panic("malloc type lacks magic");
332
333 if (vmstats.v_page_count == 0)
334 panic("malloc_uninit not allowed before vm init");
335
336 if (type->ks_limit == 0)
337 panic("malloc_uninit on uninitialized type");
338
6c92c1f2
SZ
339#ifdef SMP
340 /* Make sure that all pending kfree()s are finished. */
341 lwkt_synchronize_ipiqs("muninit");
342#endif
343
a108bf71 344#ifdef INVARIANTS
1d712609
MD
345 /*
346 * memuse is only correct in aggregation. Due to memory being allocated
347 * on one cpu and freed on another individual array entries may be
348 * negative or positive (canceling each other out).
349 */
350 for (i = ttl = 0; i < ncpus; ++i)
351 ttl += type->ks_memuse[i];
352 if (ttl) {
6ea70f76 353 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
1d712609 354 ttl, type->ks_shortdesc, i);
a108bf71
MD
355 }
356#endif
357 if (type == kmemstatistics) {
358 kmemstatistics = type->ks_next;
359 } else {
360 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
361 if (t->ks_next == type) {
362 t->ks_next = type->ks_next;
363 break;
364 }
365 }
366 }
367 type->ks_next = NULL;
368 type->ks_limit = 0;
369}
370
371/*
40153c65
MD
372 * Increase the kmalloc pool limit for the specified pool. No changes
373 * are the made if the pool would shrink.
374 */
375void
376kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
377{
378 if (type->ks_limit == 0)
379 malloc_init(type);
7c457ac8
MD
380 if (bytes == 0)
381 bytes = KvaSize;
40153c65
MD
382 if (type->ks_limit < bytes)
383 type->ks_limit = bytes;
384}
385
386/*
ebe36cfe
MD
387 * Dynamically create a malloc pool. This function is a NOP if *typep is
388 * already non-NULL.
389 */
390void
391kmalloc_create(struct malloc_type **typep, const char *descr)
392{
393 struct malloc_type *type;
394
395 if (*typep == NULL) {
396 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
397 type->ks_magic = M_MAGIC;
398 type->ks_shortdesc = descr;
399 malloc_init(type);
400 *typep = type;
401 }
402}
403
404/*
405 * Destroy a dynamically created malloc pool. This function is a NOP if
406 * the pool has already been destroyed.
407 */
408void
409kmalloc_destroy(struct malloc_type **typep)
410{
411 if (*typep != NULL) {
412 malloc_uninit(*typep);
413 kfree(*typep, M_TEMP);
414 *typep = NULL;
415 }
416}
417
418/*
a108bf71
MD
419 * Calculate the zone index for the allocation request size and set the
420 * allocation request size to that particular zone's chunk size.
421 */
422static __inline int
81dedbc2 423zoneindex(unsigned long *bytes, unsigned long *align)
a108bf71
MD
424{
425 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
426 if (n < 128) {
427 *bytes = n = (n + 7) & ~7;
81dedbc2 428 *align = 8;
a108bf71
MD
429 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
430 }
431 if (n < 256) {
432 *bytes = n = (n + 15) & ~15;
81dedbc2 433 *align = 16;
a108bf71
MD
434 return(n / 16 + 7);
435 }
436 if (n < 8192) {
437 if (n < 512) {
438 *bytes = n = (n + 31) & ~31;
81dedbc2 439 *align = 32;
a108bf71
MD
440 return(n / 32 + 15);
441 }
442 if (n < 1024) {
443 *bytes = n = (n + 63) & ~63;
81dedbc2 444 *align = 64;
a108bf71
MD
445 return(n / 64 + 23);
446 }
447 if (n < 2048) {
448 *bytes = n = (n + 127) & ~127;
81dedbc2 449 *align = 128;
a108bf71
MD
450 return(n / 128 + 31);
451 }
452 if (n < 4096) {
453 *bytes = n = (n + 255) & ~255;
81dedbc2 454 *align = 256;
a108bf71
MD
455 return(n / 256 + 39);
456 }
457 *bytes = n = (n + 511) & ~511;
81dedbc2 458 *align = 512;
a108bf71
MD
459 return(n / 512 + 47);
460 }
461#if ZALLOC_ZONE_LIMIT > 8192
462 if (n < 16384) {
463 *bytes = n = (n + 1023) & ~1023;
81dedbc2 464 *align = 1024;
a108bf71
MD
465 return(n / 1024 + 55);
466 }
467#endif
468#if ZALLOC_ZONE_LIMIT > 16384
469 if (n < 32768) {
470 *bytes = n = (n + 2047) & ~2047;
81dedbc2 471 *align = 2048;
a108bf71
MD
472 return(n / 2048 + 63);
473 }
474#endif
475 panic("Unexpected byte count %d", n);
476 return(0);
477}
478
bbb201fd
MD
479#ifdef SLAB_DEBUG
480/*
481 * Used to debug memory corruption issues. Record up to (typically 32)
482 * allocation sources for this zone (for a particular chunk size).
483 */
484
485static void
486slab_record_source(SLZone *z, const char *file, int line)
487{
488 int i;
489 int b = line & (SLAB_DEBUG_ENTRIES - 1);
490
491 i = b;
492 do {
493 if (z->z_Sources[i].file == file && z->z_Sources[i].line == line)
494 return;
495 if (z->z_Sources[i].file == NULL)
496 break;
497 i = (i + 1) & (SLAB_DEBUG_ENTRIES - 1);
498 } while (i != b);
499 z->z_Sources[i].file = file;
500 z->z_Sources[i].line = line;
501}
502
503#endif
504
1e57f867
SZ
505static __inline unsigned long
506powerof2_size(unsigned long size)
507{
508 int i, wt;
509
510 if (size == 0)
511 return 0;
512
513 i = flsl(size);
514 wt = (size & ~(1 << (i - 1)));
515 if (!wt)
516 --i;
517
518 return (1UL << i);
519}
520
a108bf71 521/*
5fee07e6 522 * kmalloc() (SLAB ALLOCATOR)
a108bf71
MD
523 *
524 * Allocate memory via the slab allocator. If the request is too large,
525 * or if it page-aligned beyond a certain size, we fall back to the
526 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
527 * &SlabMisc if you don't care.
528 *
8cb2bf45
JS
529 * M_RNOWAIT - don't block.
530 * M_NULLOK - return NULL instead of blocking.
a108bf71 531 * M_ZERO - zero the returned memory.
dc1fd4b3
MD
532 * M_USE_RESERVE - allow greater drawdown of the free list
533 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
1e57f867 534 * M_POWEROF2 - roundup size to the nearest power of 2
5b287bba
MD
535 *
536 * MPSAFE
a108bf71 537 */
bbb201fd
MD
538
539#ifdef SLAB_DEBUG
540void *
541kmalloc_debug(unsigned long size, struct malloc_type *type, int flags,
542 const char *file, int line)
543#else
a108bf71 544void *
8aca2bd4 545kmalloc(unsigned long size, struct malloc_type *type, int flags)
bbb201fd 546#endif
a108bf71
MD
547{
548 SLZone *z;
549 SLChunk *chunk;
d8100bdc 550#ifdef SMP
5fee07e6 551 SLChunk *bchunk;
d8100bdc 552#endif
a108bf71 553 SLGlobalData *slgd;
bba6a44d 554 struct globaldata *gd;
81dedbc2 555 unsigned long align;
a108bf71 556 int zi;
d2182dc1
MD
557#ifdef INVARIANTS
558 int i;
559#endif
a108bf71 560
b68ad50c 561 logmemory_quick(malloc_beg);
bba6a44d
MD
562 gd = mycpu;
563 slgd = &gd->gd_slab;
a108bf71
MD
564
565 /*
566 * XXX silly to have this in the critical path.
567 */
568 if (type->ks_limit == 0) {
569 crit_enter();
570 if (type->ks_limit == 0)
571 malloc_init(type);
572 crit_exit();
573 }
574 ++type->ks_calls;
575
1e57f867
SZ
576 if (flags & M_POWEROF2)
577 size = powerof2_size(size);
578
a108bf71 579 /*
38e34349
MD
580 * Handle the case where the limit is reached. Panic if we can't return
581 * NULL. The original malloc code looped, but this tended to
a108bf71 582 * simply deadlock the computer.
38e34349
MD
583 *
584 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
585 * to determine if a more complete limit check should be done. The
586 * actual memory use is tracked via ks_memuse[cpu].
a108bf71 587 */
bba6a44d
MD
588 while (type->ks_loosememuse >= type->ks_limit) {
589 int i;
590 long ttl;
591
592 for (i = ttl = 0; i < ncpus; ++i)
593 ttl += type->ks_memuse[i];
38e34349 594 type->ks_loosememuse = ttl; /* not MP synchronized */
28135cc2
MD
595 if ((ssize_t)ttl < 0) /* deal with occassional race */
596 ttl = 0;
bba6a44d 597 if (ttl >= type->ks_limit) {
f2b5daf9 598 if (flags & M_NULLOK) {
5fee07e6 599 logmemory(malloc_end, NULL, type, size, flags);
bba6a44d 600 return(NULL);
f2b5daf9 601 }
bba6a44d
MD
602 panic("%s: malloc limit exceeded", type->ks_shortdesc);
603 }
a108bf71
MD
604 }
605
606 /*
607 * Handle the degenerate size == 0 case. Yes, this does happen.
608 * Return a special pointer. This is to maintain compatibility with
609 * the original malloc implementation. Certain devices, such as the
610 * adaptec driver, not only allocate 0 bytes, they check for NULL and
611 * also realloc() later on. Joy.
612 */
f2b5daf9 613 if (size == 0) {
5fee07e6 614 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
a108bf71 615 return(ZERO_LENGTH_PTR);
f2b5daf9 616 }
a108bf71
MD
617
618 /*
a7cf0021
MD
619 * Handle hysteresis from prior frees here in malloc(). We cannot
620 * safely manipulate the kernel_map in free() due to free() possibly
621 * being called via an IPI message or from sensitive interrupt code.
5fee07e6
MD
622 *
623 * NOTE: ku_pagecnt must be cleared before we free the slab or we
624 * might race another cpu allocating the kva and setting
625 * ku_pagecnt.
a7cf0021 626 */
ad94a851 627 while (slgd->NFreeZones > ZoneRelsThresh && (flags & M_RNOWAIT) == 0) {
46a3f46d 628 crit_enter();
ad94a851 629 if (slgd->NFreeZones > ZoneRelsThresh) { /* crit sect race */
722871d3 630 int *kup;
5fee07e6 631
46a3f46d
MD
632 z = slgd->FreeZones;
633 slgd->FreeZones = z->z_Next;
634 --slgd->NFreeZones;
5fee07e6 635 kup = btokup(z);
722871d3 636 *kup = 0;
46a3f46d 637 kmem_slab_free(z, ZoneSize); /* may block */
51295aee 638 atomic_add_int(&ZoneGenAlloc, -ZoneSize / 1024);
46a3f46d
MD
639 }
640 crit_exit();
641 }
5fee07e6 642
46a3f46d 643 /*
5fee07e6 644 * XXX handle oversized frees that were queued from kfree().
46a3f46d 645 */
dc1fd4b3 646 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) {
46a3f46d
MD
647 crit_enter();
648 if ((z = slgd->FreeOvZones) != NULL) {
7e83df33
MD
649 vm_size_t tsize;
650
46a3f46d
MD
651 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
652 slgd->FreeOvZones = z->z_Next;
7e83df33
MD
653 tsize = z->z_ChunkSize;
654 kmem_slab_free(z, tsize); /* may block */
655 atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024);
46a3f46d
MD
656 }
657 crit_exit();
a7cf0021
MD
658 }
659
660 /*
a108bf71
MD
661 * Handle large allocations directly. There should not be very many of
662 * these so performance is not a big issue.
663 *
b543eeed
MD
664 * The backend allocator is pretty nasty on a SMP system. Use the
665 * slab allocator for one and two page-sized chunks even though we lose
666 * some efficiency. XXX maybe fix mmio and the elf loader instead.
a108bf71 667 */
b543eeed 668 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
722871d3 669 int *kup;
a108bf71
MD
670
671 size = round_page(size);
672 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
f2b5daf9 673 if (chunk == NULL) {
5fee07e6 674 logmemory(malloc_end, NULL, type, size, flags);
a108bf71 675 return(NULL);
f2b5daf9 676 }
665206ee 677 atomic_add_int(&ZoneBigAlloc, (int)size / 1024);
a108bf71 678 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
8f1d5415 679 flags |= M_PASSIVE_ZERO;
a108bf71 680 kup = btokup(chunk);
722871d3 681 *kup = size / PAGE_SIZE;
a108bf71
MD
682 crit_enter();
683 goto done;
684 }
685
686 /*
687 * Attempt to allocate out of an existing zone. First try the free list,
688 * then allocate out of unallocated space. If we find a good zone move
689 * it to the head of the list so later allocations find it quickly
690 * (we might have thousands of zones in the list).
691 *
692 * Note: zoneindex() will panic of size is too large.
693 */
81dedbc2 694 zi = zoneindex(&size, &align);
a108bf71
MD
695 KKASSERT(zi < NZONES);
696 crit_enter();
a108bf71 697
5fee07e6 698 if ((z = slgd->ZoneAry[zi]) != NULL) {
a108bf71 699 /*
5fee07e6
MD
700 * Locate a chunk - we have to have at least one. If this is the
701 * last chunk go ahead and do the work to retrieve chunks freed
702 * from remote cpus, and if the zone is still empty move it off
703 * the ZoneAry.
a108bf71 704 */
5fee07e6
MD
705 if (--z->z_NFree <= 0) {
706 KKASSERT(z->z_NFree == 0);
707
708#ifdef SMP
709 /*
710 * WARNING! This code competes with other cpus. It is ok
711 * for us to not drain RChunks here but we might as well, and
712 * it is ok if more accumulate after we're done.
713 *
714 * Set RSignal before pulling rchunks off, indicating that we
715 * will be moving ourselves off of the ZoneAry. Remote ends will
716 * read RSignal before putting rchunks on thus interlocking
717 * their IPI signaling.
718 */
719 if (z->z_RChunks == NULL)
720 atomic_swap_int(&z->z_RSignal, 1);
721
722 while ((bchunk = z->z_RChunks) != NULL) {
723 cpu_ccfence();
724 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
725 *z->z_LChunksp = bchunk;
726 while (bchunk) {
727 chunk_mark_free(z, bchunk);
728 z->z_LChunksp = &bchunk->c_Next;
729 bchunk = bchunk->c_Next;
730 ++z->z_NFree;
731 }
732 break;
733 }
734 }
735#endif
736 /*
737 * Remove from the zone list if no free chunks remain.
738 * Clear RSignal
739 */
740 if (z->z_NFree == 0) {
741 slgd->ZoneAry[zi] = z->z_Next;
742 z->z_Next = NULL;
743 } else {
744 z->z_RSignal = 0;
745 }
a108bf71
MD
746 }
747
748 /*
5fee07e6 749 * Fast path, we have chunks available in z_LChunks.
a108bf71 750 */
5fee07e6
MD
751 chunk = z->z_LChunks;
752 if (chunk) {
10cc6608 753 chunk_mark_allocated(z, chunk);
5fee07e6
MD
754 z->z_LChunks = chunk->c_Next;
755 if (z->z_LChunks == NULL)
756 z->z_LChunksp = &z->z_LChunks;
bbb201fd
MD
757#ifdef SLAB_DEBUG
758 slab_record_source(z, file, line);
759#endif
a108bf71 760 goto done;
a108bf71
MD
761 }
762
763 /*
5fee07e6
MD
764 * No chunks are available in LChunks, the free chunk MUST be
765 * in the never-before-used memory area, controlled by UIndex.
766 *
767 * The consequences are very serious if our zone got corrupted so
768 * we use an explicit panic rather than a KASSERT.
a108bf71 769 */
1c5ca4f3 770 if (z->z_UIndex + 1 != z->z_NMax)
5fee07e6 771 ++z->z_UIndex;
1c5ca4f3
MD
772 else
773 z->z_UIndex = 0;
5fee07e6 774
1c5ca4f3
MD
775 if (z->z_UIndex == z->z_UEndIndex)
776 panic("slaballoc: corrupted zone");
5fee07e6 777
1c5ca4f3 778 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
8f1d5415 779 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 780 flags &= ~M_ZERO;
8f1d5415
MD
781 flags |= M_PASSIVE_ZERO;
782 }
10cc6608 783 chunk_mark_allocated(z, chunk);
bbb201fd
MD
784#ifdef SLAB_DEBUG
785 slab_record_source(z, file, line);
786#endif
a108bf71
MD
787 goto done;
788 }
789
790 /*
791 * If all zones are exhausted we need to allocate a new zone for this
792 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
6ab8e1da
MD
793 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
794 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
795 * we do not pre-zero it because we do not want to mess up the L1 cache.
a108bf71
MD
796 *
797 * At least one subsystem, the tty code (see CROUND) expects power-of-2
798 * allocations to be power-of-2 aligned. We maintain compatibility by
799 * adjusting the base offset below.
800 */
801 {
802 int off;
722871d3 803 int *kup;
a108bf71
MD
804
805 if ((z = slgd->FreeZones) != NULL) {
806 slgd->FreeZones = z->z_Next;
807 --slgd->NFreeZones;
808 bzero(z, sizeof(SLZone));
6ab8e1da 809 z->z_Flags |= SLZF_UNOTZEROD;
a108bf71
MD
810 } else {
811 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
812 if (z == NULL)
813 goto fail;
51295aee 814 atomic_add_int(&ZoneGenAlloc, ZoneSize / 1024);
a108bf71
MD
815 }
816
817 /*
10cc6608
MD
818 * How big is the base structure?
819 */
820#if defined(INVARIANTS)
821 /*
822 * Make room for z_Bitmap. An exact calculation is somewhat more
823 * complicated so don't make an exact calculation.
824 */
825 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
826 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
827#else
828 off = sizeof(SLZone);
829#endif
830
831 /*
a108bf71
MD
832 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
833 * Otherwise just 8-byte align the data.
834 */
835 if ((size | (size - 1)) + 1 == (size << 1))
10cc6608 836 off = (off + size - 1) & ~(size - 1);
a108bf71 837 else
81dedbc2 838 off = (off + align - 1) & ~(align - 1);
a108bf71
MD
839 z->z_Magic = ZALLOC_SLAB_MAGIC;
840 z->z_ZoneIndex = zi;
841 z->z_NMax = (ZoneSize - off) / size;
842 z->z_NFree = z->z_NMax - 1;
1c5ca4f3
MD
843 z->z_BasePtr = (char *)z + off;
844 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
a108bf71 845 z->z_ChunkSize = size;
2db3b277 846 z->z_CpuGd = gd;
bba6a44d 847 z->z_Cpu = gd->gd_cpuid;
5fee07e6 848 z->z_LChunksp = &z->z_LChunks;
bbb201fd
MD
849#ifdef SLAB_DEBUG
850 bcopy(z->z_Sources, z->z_AltSources, sizeof(z->z_Sources));
851 bzero(z->z_Sources, sizeof(z->z_Sources));
852#endif
1c5ca4f3 853 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
a108bf71
MD
854 z->z_Next = slgd->ZoneAry[zi];
855 slgd->ZoneAry[zi] = z;
8f1d5415 856 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 857 flags &= ~M_ZERO; /* already zero'd */
8f1d5415
MD
858 flags |= M_PASSIVE_ZERO;
859 }
5fee07e6 860 kup = btokup(z);
722871d3 861 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */
10cc6608 862 chunk_mark_allocated(z, chunk);
bbb201fd
MD
863#ifdef SLAB_DEBUG
864 slab_record_source(z, file, line);
865#endif
1c5ca4f3
MD
866
867 /*
868 * Slide the base index for initial allocations out of the next
869 * zone we create so we do not over-weight the lower part of the
870 * cpu memory caches.
871 */
872 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
873 & (ZALLOC_MAX_ZONE_SIZE - 1);
a108bf71 874 }
5fee07e6 875
a108bf71 876done:
bba6a44d
MD
877 ++type->ks_inuse[gd->gd_cpuid];
878 type->ks_memuse[gd->gd_cpuid] += size;
38e34349 879 type->ks_loosememuse += size; /* not MP synchronized */
a108bf71 880 crit_exit();
5fee07e6 881
a108bf71
MD
882 if (flags & M_ZERO)
883 bzero(chunk, size);
bba6a44d 884#ifdef INVARIANTS
d2182dc1
MD
885 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
886 if (use_malloc_pattern) {
887 for (i = 0; i < size; i += sizeof(int)) {
888 *(int *)((char *)chunk + i) = -1;
889 }
890 }
bba6a44d 891 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
d2182dc1 892 }
bba6a44d 893#endif
5fee07e6 894 logmemory(malloc_end, chunk, type, size, flags);
a108bf71
MD
895 return(chunk);
896fail:
897 crit_exit();
5fee07e6 898 logmemory(malloc_end, NULL, type, size, flags);
a108bf71
MD
899 return(NULL);
900}
901
38e34349
MD
902/*
903 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
904 *
905 * Generally speaking this routine is not called very often and we do
906 * not attempt to optimize it beyond reusing the same pointer if the
907 * new size fits within the chunking of the old pointer's zone.
908 */
bbb201fd
MD
909#ifdef SLAB_DEBUG
910void *
911krealloc_debug(void *ptr, unsigned long size,
912 struct malloc_type *type, int flags,
913 const char *file, int line)
914#else
a108bf71 915void *
8aca2bd4 916krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
bbb201fd 917#endif
a108bf71 918{
722871d3 919 unsigned long osize;
81dedbc2 920 unsigned long align;
a108bf71
MD
921 SLZone *z;
922 void *nptr;
722871d3 923 int *kup;
a108bf71 924
eb7f3e3c
MD
925 KKASSERT((flags & M_ZERO) == 0); /* not supported */
926
a108bf71 927 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
bbb201fd 928 return(kmalloc_debug(size, type, flags, file, line));
a108bf71 929 if (size == 0) {
efda3bd0 930 kfree(ptr, type);
a108bf71
MD
931 return(NULL);
932 }
933
934 /*
935 * Handle oversized allocations. XXX we really should require that a
936 * size be passed to free() instead of this nonsense.
937 */
5fee07e6 938 kup = btokup(ptr);
722871d3
MD
939 if (*kup > 0) {
940 osize = *kup << PAGE_SHIFT;
5fee07e6
MD
941 if (osize == round_page(size))
942 return(ptr);
bbb201fd 943 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
5fee07e6
MD
944 return(NULL);
945 bcopy(ptr, nptr, min(size, osize));
946 kfree(ptr, type);
947 return(nptr);
a108bf71
MD
948 }
949
950 /*
951 * Get the original allocation's zone. If the new request winds up
952 * using the same chunk size we do not have to do anything.
953 */
5fee07e6
MD
954 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
955 kup = btokup(z);
722871d3 956 KKASSERT(*kup < 0);
a108bf71
MD
957 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
958
a108bf71
MD
959 /*
960 * Allocate memory for the new request size. Note that zoneindex has
961 * already adjusted the request size to the appropriate chunk size, which
962 * should optimize our bcopy(). Then copy and return the new pointer.
1ea6580d
MD
963 *
964 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
965 * necessary align the result.
966 *
967 * We can only zoneindex (to align size to the chunk size) if the new
968 * size is not too large.
a108bf71 969 */
1ea6580d 970 if (size < ZoneLimit) {
81dedbc2 971 zoneindex(&size, &align);
1ea6580d
MD
972 if (z->z_ChunkSize == size)
973 return(ptr);
974 }
bbb201fd 975 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
a108bf71
MD
976 return(NULL);
977 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
efda3bd0 978 kfree(ptr, type);
a108bf71
MD
979 return(nptr);
980}
981
38e34349 982/*
45d2b1d8
MD
983 * Return the kmalloc limit for this type, in bytes.
984 */
985long
986kmalloc_limit(struct malloc_type *type)
987{
988 if (type->ks_limit == 0) {
989 crit_enter();
990 if (type->ks_limit == 0)
991 malloc_init(type);
992 crit_exit();
993 }
994 return(type->ks_limit);
995}
996
997/*
38e34349
MD
998 * Allocate a copy of the specified string.
999 *
1000 * (MP SAFE) (MAY BLOCK)
1001 */
bbb201fd
MD
1002#ifdef SLAB_DEBUG
1003char *
1004kstrdup_debug(const char *str, struct malloc_type *type,
1005 const char *file, int line)
1006#else
1ac06773 1007char *
59302080 1008kstrdup(const char *str, struct malloc_type *type)
bbb201fd 1009#endif
1ac06773
MD
1010{
1011 int zlen; /* length inclusive of terminating NUL */
1012 char *nstr;
1013
1014 if (str == NULL)
1015 return(NULL);
1016 zlen = strlen(str) + 1;
bbb201fd 1017 nstr = kmalloc_debug(zlen, type, M_WAITOK, file, line);
1ac06773
MD
1018 bcopy(str, nstr, zlen);
1019 return(nstr);
1020}
1021
1d712609 1022#ifdef SMP
a108bf71 1023/*
5fee07e6 1024 * Notify our cpu that a remote cpu has freed some chunks in a zone that
df9daea8
MD
1025 * we own. RCount will be bumped so the memory should be good, but validate
1026 * that it really is.
a108bf71
MD
1027 */
1028static
1029void
5fee07e6 1030kfree_remote(void *ptr)
a108bf71 1031{
5fee07e6
MD
1032 SLGlobalData *slgd;
1033 SLChunk *bchunk;
1034 SLZone *z;
1035 int nfree;
722871d3 1036 int *kup;
5fee07e6 1037
5fee07e6
MD
1038 slgd = &mycpu->gd_slab;
1039 z = ptr;
1040 kup = btokup(z);
df9daea8
MD
1041 KKASSERT(*kup == -((int)mycpuid + 1));
1042 KKASSERT(z->z_RCount > 0);
1043 atomic_subtract_int(&z->z_RCount, 1);
5fee07e6 1044
5bf48697 1045 logmemory(free_rem_beg, z, NULL, 0L, 0);
df9daea8
MD
1046 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1047 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
1048 nfree = z->z_NFree;
5fee07e6 1049
df9daea8
MD
1050 /*
1051 * Indicate that we will no longer be off of the ZoneAry by
1052 * clearing RSignal.
1053 */
1054 if (z->z_RChunks)
1055 z->z_RSignal = 0;
5fee07e6 1056
df9daea8
MD
1057 /*
1058 * Atomically extract the bchunks list and then process it back
1059 * into the lchunks list. We want to append our bchunks to the
1060 * lchunks list and not prepend since we likely do not have
1061 * cache mastership of the related data (not that it helps since
1062 * we are using c_Next).
1063 */
1064 while ((bchunk = z->z_RChunks) != NULL) {
1065 cpu_ccfence();
1066 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
1067 *z->z_LChunksp = bchunk;
1068 while (bchunk) {
1069 chunk_mark_free(z, bchunk);
1070 z->z_LChunksp = &bchunk->c_Next;
1071 bchunk = bchunk->c_Next;
1072 ++z->z_NFree;
5fee07e6 1073 }
df9daea8 1074 break;
5fee07e6 1075 }
df9daea8
MD
1076 }
1077 if (z->z_NFree && nfree == 0) {
1078 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1079 slgd->ZoneAry[z->z_ZoneIndex] = z;
1080 }
5fee07e6 1081
df9daea8
MD
1082 /*
1083 * If the zone becomes totally free, and there are other zones we
1084 * can allocate from, move this zone to the FreeZones list. Since
1085 * this code can be called from an IPI callback, do *NOT* try to mess
1086 * with kernel_map here. Hysteresis will be performed at malloc() time.
1087 *
1088 * Do not move the zone if there is an IPI inflight, otherwise MP
1089 * races can result in our free_remote code accessing a destroyed
1090 * zone.
1091 */
1092 if (z->z_NFree == z->z_NMax &&
1093 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1094 z->z_RCount == 0
1095 ) {
1096 SLZone **pz;
1097 int *kup;
5fee07e6 1098
df9daea8
MD
1099 for (pz = &slgd->ZoneAry[z->z_ZoneIndex];
1100 z != *pz;
1101 pz = &(*pz)->z_Next) {
1102 ;
5fee07e6 1103 }
df9daea8
MD
1104 *pz = z->z_Next;
1105 z->z_Magic = -1;
1106 z->z_Next = slgd->FreeZones;
1107 slgd->FreeZones = z;
1108 ++slgd->NFreeZones;
1109 kup = btokup(z);
1110 *kup = 0;
5fee07e6 1111 }
5bf48697 1112 logmemory(free_rem_end, z, bchunk, 0L, 0);
a108bf71
MD
1113}
1114
1d712609
MD
1115#endif
1116
38e34349 1117/*
5b287bba 1118 * free (SLAB ALLOCATOR)
38e34349
MD
1119 *
1120 * Free a memory block previously allocated by malloc. Note that we do not
5fee07e6 1121 * attempt to update ks_loosememuse as MP races could prevent us from
38e34349 1122 * checking memory limits in malloc.
5b287bba
MD
1123 *
1124 * MPSAFE
38e34349 1125 */
a108bf71 1126void
8aca2bd4 1127kfree(void *ptr, struct malloc_type *type)
a108bf71
MD
1128{
1129 SLZone *z;
1130 SLChunk *chunk;
1131 SLGlobalData *slgd;
bba6a44d 1132 struct globaldata *gd;
722871d3 1133 int *kup;
5fee07e6 1134 unsigned long size;
d8100bdc
SW
1135#ifdef SMP
1136 SLChunk *bchunk;
5fee07e6 1137 int rsignal;
d8100bdc 1138#endif
a108bf71 1139
b68ad50c 1140 logmemory_quick(free_beg);
bba6a44d
MD
1141 gd = mycpu;
1142 slgd = &gd->gd_slab;
a108bf71 1143
d39911d9
JS
1144 if (ptr == NULL)
1145 panic("trying to free NULL pointer");
1146
a108bf71
MD
1147 /*
1148 * Handle special 0-byte allocations
1149 */
f2b5daf9 1150 if (ptr == ZERO_LENGTH_PTR) {
5bf48697 1151 logmemory(free_zero, ptr, type, -1UL, 0);
b68ad50c 1152 logmemory_quick(free_end);
a108bf71 1153 return;
f2b5daf9 1154 }
a108bf71
MD
1155
1156 /*
5fee07e6
MD
1157 * Panic on bad malloc type
1158 */
1159 if (type->ks_magic != M_MAGIC)
1160 panic("free: malloc type lacks magic");
1161
1162 /*
a108bf71
MD
1163 * Handle oversized allocations. XXX we really should require that a
1164 * size be passed to free() instead of this nonsense.
bba6a44d
MD
1165 *
1166 * This code is never called via an ipi.
a108bf71 1167 */
5fee07e6 1168 kup = btokup(ptr);
722871d3
MD
1169 if (*kup > 0) {
1170 size = *kup << PAGE_SHIFT;
1171 *kup = 0;
a108bf71 1172#ifdef INVARIANTS
5fee07e6
MD
1173 KKASSERT(sizeof(weirdary) <= size);
1174 bcopy(weirdary, ptr, sizeof(weirdary));
a108bf71 1175#endif
5fee07e6
MD
1176 /*
1177 * NOTE: For oversized allocations we do not record the
1178 * originating cpu. It gets freed on the cpu calling
1179 * kfree(). The statistics are in aggregate.
1180 *
1181 * note: XXX we have still inherited the interrupts-can't-block
1182 * assumption. An interrupt thread does not bump
1183 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1184 * primarily until we can fix softupdate's assumptions about free().
1185 */
1186 crit_enter();
1187 --type->ks_inuse[gd->gd_cpuid];
1188 type->ks_memuse[gd->gd_cpuid] -= size;
1189 if (mycpu->gd_intr_nesting_level ||
1190 (gd->gd_curthread->td_flags & TDF_INTTHREAD))
1191 {
1192 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1193 z = (SLZone *)ptr;
1194 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1195 z->z_Next = slgd->FreeOvZones;
1196 z->z_ChunkSize = size;
1197 slgd->FreeOvZones = z;
1198 crit_exit();
1199 } else {
1200 crit_exit();
1201 logmemory(free_ovsz, ptr, type, size, 0);
1202 kmem_slab_free(ptr, size); /* may block */
1203 atomic_add_int(&ZoneBigAlloc, -(int)size / 1024);
a108bf71 1204 }
5fee07e6
MD
1205 logmemory_quick(free_end);
1206 return;
a108bf71
MD
1207 }
1208
1209 /*
1210 * Zone case. Figure out the zone based on the fact that it is
1211 * ZoneSize aligned.
1212 */
5fee07e6
MD
1213 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1214 kup = btokup(z);
722871d3 1215 KKASSERT(*kup < 0);
a108bf71
MD
1216 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1217
1218 /*
5fee07e6
MD
1219 * If we do not own the zone then use atomic ops to free to the
1220 * remote cpu linked list and notify the target zone using a
1221 * passive message.
1222 *
1223 * The target zone cannot be deallocated while we own a chunk of it,
1224 * so the zone header's storage is stable until the very moment
1225 * we adjust z_RChunks. After that we cannot safely dereference (z).
1226 *
1227 * (no critical section needed)
a108bf71 1228 */
2db3b277 1229 if (z->z_CpuGd != gd) {
75c7ffea 1230#ifdef SMP
5fee07e6
MD
1231 /*
1232 * Making these adjustments now allow us to avoid passing (type)
1233 * to the remote cpu. Note that ks_inuse/ks_memuse is being
28135cc2
MD
1234 * adjusted on OUR cpu, not the zone cpu, but it should all still
1235 * sum up properly and cancel out.
5fee07e6 1236 */
28135cc2
MD
1237 crit_enter();
1238 --type->ks_inuse[gd->gd_cpuid];
1239 type->ks_memuse[gd->gd_cpuid] -= z->z_ChunkSize;
1240 crit_exit();
5fee07e6
MD
1241
1242 /*
1243 * WARNING! This code competes with other cpus. Once we
1244 * successfully link the chunk to RChunks the remote
1245 * cpu can rip z's storage out from under us.
df9daea8
MD
1246 *
1247 * Bumping RCount prevents z's storage from getting
1248 * ripped out.
5fee07e6
MD
1249 */
1250 rsignal = z->z_RSignal;
1251 cpu_lfence();
df9daea8
MD
1252 if (rsignal)
1253 atomic_add_int(&z->z_RCount, 1);
5fee07e6
MD
1254
1255 chunk = ptr;
1256 for (;;) {
1257 bchunk = z->z_RChunks;
1258 cpu_ccfence();
1259 chunk->c_Next = bchunk;
1260 cpu_sfence();
1261
1262 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1263 break;
1264 }
5fee07e6
MD
1265
1266 /*
1267 * We have to signal the remote cpu if our actions will cause
1268 * the remote zone to be placed back on ZoneAry so it can
1269 * move the zone back on.
1270 *
1271 * We only need to deal with NULL->non-NULL RChunk transitions
1272 * and only if z_RSignal is set. We interlock by reading rsignal
1273 * before adding our chunk to RChunks. This should result in
1274 * virtually no IPI traffic.
1275 *
1276 * We can use a passive IPI to reduce overhead even further.
1277 */
1278 if (bchunk == NULL && rsignal) {
5bf48697 1279 logmemory(free_request, ptr, type, (unsigned long)z->z_ChunkSize, 0);
5fee07e6 1280 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
df9daea8
MD
1281 /* z can get ripped out from under us from this point on */
1282 } else if (rsignal) {
1283 atomic_subtract_int(&z->z_RCount, 1);
1284 /* z can get ripped out from under us from this point on */
5fee07e6 1285 }
75c7ffea
MD
1286#else
1287 panic("Corrupt SLZone");
1288#endif
b68ad50c 1289 logmemory_quick(free_end);
a108bf71
MD
1290 return;
1291 }
1292
5fee07e6
MD
1293 /*
1294 * kfree locally
1295 */
5bf48697 1296 logmemory(free_chunk, ptr, type, (unsigned long)z->z_ChunkSize, 0);
f2b5daf9 1297
a108bf71 1298 crit_enter();
a108bf71 1299 chunk = ptr;
10cc6608 1300 chunk_mark_free(z, chunk);
a108bf71
MD
1301
1302 /*
1303 * Put weird data into the memory to detect modifications after freeing,
1304 * illegal pointer use after freeing (we should fault on the odd address),
1305 * and so forth. XXX needs more work, see the old malloc code.
1306 */
1307#ifdef INVARIANTS
1308 if (z->z_ChunkSize < sizeof(weirdary))
1309 bcopy(weirdary, chunk, z->z_ChunkSize);
1310 else
1311 bcopy(weirdary, chunk, sizeof(weirdary));
1312#endif
1313
1314 /*
5fee07e6
MD
1315 * Add this free non-zero'd chunk to a linked list for reuse. Add
1316 * to the front of the linked list so it is more likely to be
1317 * reallocated, since it is already in our L1 cache.
a108bf71 1318 */
6ab8e1da 1319#ifdef INVARIANTS
c439ad8f 1320 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
fc92d4aa 1321 panic("BADFREE %p", chunk);
a108bf71 1322#endif
5fee07e6
MD
1323 chunk->c_Next = z->z_LChunks;
1324 z->z_LChunks = chunk;
1325 if (chunk->c_Next == NULL)
1326 z->z_LChunksp = &chunk->c_Next;
1327
6ab8e1da 1328#ifdef INVARIANTS
c439ad8f 1329 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
a108bf71 1330 panic("BADFREE2");
6ab8e1da 1331#endif
a108bf71
MD
1332
1333 /*
1334 * Bump the number of free chunks. If it becomes non-zero the zone
1335 * must be added back onto the appropriate list.
1336 */
1337 if (z->z_NFree++ == 0) {
1338 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1339 slgd->ZoneAry[z->z_ZoneIndex] = z;
1340 }
1341
bba6a44d
MD
1342 --type->ks_inuse[z->z_Cpu];
1343 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
a108bf71
MD
1344
1345 /*
1346 * If the zone becomes totally free, and there are other zones we
a7cf0021
MD
1347 * can allocate from, move this zone to the FreeZones list. Since
1348 * this code can be called from an IPI callback, do *NOT* try to mess
1349 * with kernel_map here. Hysteresis will be performed at malloc() time.
a108bf71
MD
1350 */
1351 if (z->z_NFree == z->z_NMax &&
df9daea8
MD
1352 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1353 z->z_RCount == 0
a108bf71
MD
1354 ) {
1355 SLZone **pz;
722871d3 1356 int *kup;
a108bf71
MD
1357
1358 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
1359 ;
1360 *pz = z->z_Next;
1361 z->z_Magic = -1;
a7cf0021
MD
1362 z->z_Next = slgd->FreeZones;
1363 slgd->FreeZones = z;
1364 ++slgd->NFreeZones;
5fee07e6 1365 kup = btokup(z);
722871d3 1366 *kup = 0;
a108bf71 1367 }
b68ad50c 1368 logmemory_quick(free_end);
a108bf71
MD
1369 crit_exit();
1370}
1371
10cc6608 1372#if defined(INVARIANTS)
5fee07e6 1373
10cc6608
MD
1374/*
1375 * Helper routines for sanity checks
1376 */
1377static
1378void
1379chunk_mark_allocated(SLZone *z, void *chunk)
1380{
1381 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1382 __uint32_t *bitptr;
1383
5fee07e6
MD
1384 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1385 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1386 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
10cc6608
MD
1387 bitptr = &z->z_Bitmap[bitdex >> 5];
1388 bitdex &= 31;
5fee07e6
MD
1389 KASSERT((*bitptr & (1 << bitdex)) == 0,
1390 ("memory chunk %p is already allocated!", chunk));
10cc6608
MD
1391 *bitptr |= 1 << bitdex;
1392}
1393
1394static
1395void
1396chunk_mark_free(SLZone *z, void *chunk)
1397{
1398 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1399 __uint32_t *bitptr;
1400
5fee07e6
MD
1401 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1402 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1403 ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
10cc6608
MD
1404 bitptr = &z->z_Bitmap[bitdex >> 5];
1405 bitdex &= 31;
5fee07e6
MD
1406 KASSERT((*bitptr & (1 << bitdex)) != 0,
1407 ("memory chunk %p is already free!", chunk));
10cc6608
MD
1408 *bitptr &= ~(1 << bitdex);
1409}
1410
1411#endif
1412
a108bf71 1413/*
5b287bba 1414 * kmem_slab_alloc()
a108bf71
MD
1415 *
1416 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1417 * specified alignment. M_* flags are expected in the flags field.
1418 *
1419 * Alignment must be a multiple of PAGE_SIZE.
1420 *
1421 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1422 * but when we move zalloc() over to use this function as its backend
1423 * we will have to switch to kreserve/krelease and call reserve(0)
1424 * after the new space is made available.
dc1fd4b3
MD
1425 *
1426 * Interrupt code which has preempted other code is not allowed to
c397c465
MD
1427 * use PQ_CACHE pages. However, if an interrupt thread is run
1428 * non-preemptively or blocks and then runs non-preemptively, then
b12defdc 1429 * it is free to use PQ_CACHE pages. <--- may not apply any longer XXX
a108bf71
MD
1430 */
1431static void *
1432kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1433{
1434 vm_size_t i;
1435 vm_offset_t addr;
1de1e800 1436 int count, vmflags, base_vmflags;
b12defdc
MD
1437 vm_page_t mbase = NULL;
1438 vm_page_t m;
dc1fd4b3 1439 thread_t td;
a108bf71
MD
1440
1441 size = round_page(size);
e4846942 1442 addr = vm_map_min(&kernel_map);
a108bf71 1443
a108bf71
MD
1444 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1445 crit_enter();
e4846942 1446 vm_map_lock(&kernel_map);
c809941b 1447 if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) {
e4846942 1448 vm_map_unlock(&kernel_map);
8cb2bf45 1449 if ((flags & M_NULLOK) == 0)
a108bf71 1450 panic("kmem_slab_alloc(): kernel_map ran out of space!");
a108bf71 1451 vm_map_entry_release(count);
2de4f77e 1452 crit_exit();
a108bf71
MD
1453 return(NULL);
1454 }
e4846942
MD
1455
1456 /*
1457 * kernel_object maps 1:1 to kernel_map.
1458 */
b12defdc
MD
1459 vm_object_hold(&kernel_object);
1460 vm_object_reference_locked(&kernel_object);
e4846942
MD
1461 vm_map_insert(&kernel_map, &count,
1462 &kernel_object, addr, addr, addr + size,
1b874851
MD
1463 VM_MAPTYPE_NORMAL,
1464 VM_PROT_ALL, VM_PROT_ALL,
1465 0);
b12defdc
MD
1466 vm_object_drop(&kernel_object);
1467 vm_map_set_wired_quick(&kernel_map, addr, size, &count);
1468 vm_map_unlock(&kernel_map);
a108bf71 1469
dc1fd4b3 1470 td = curthread;
dc1fd4b3 1471
1de1e800
JS
1472 base_vmflags = 0;
1473 if (flags & M_ZERO)
1474 base_vmflags |= VM_ALLOC_ZERO;
1475 if (flags & M_USE_RESERVE)
1476 base_vmflags |= VM_ALLOC_SYSTEM;
1477 if (flags & M_USE_INTERRUPT_RESERVE)
1478 base_vmflags |= VM_ALLOC_INTERRUPT;
77912481
MD
1479 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1480 panic("kmem_slab_alloc: bad flags %08x (%p)",
1481 flags, ((int **)&size)[-1]);
1482 }
1de1e800 1483
a108bf71 1484 /*
b12defdc
MD
1485 * Allocate the pages. Do not mess with the PG_ZERO flag or map
1486 * them yet. VM_ALLOC_NORMAL can only be set if we are not preempting.
1487 *
1488 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1489 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1490 * implied in this case), though I'm not sure if we really need to
1491 * do that.
a108bf71 1492 */
b12defdc
MD
1493 vmflags = base_vmflags;
1494 if (flags & M_WAITOK) {
1495 if (td->td_preempted)
1496 vmflags |= VM_ALLOC_SYSTEM;
1497 else
1498 vmflags |= VM_ALLOC_NORMAL;
1499 }
a108bf71 1500
b12defdc
MD
1501 vm_object_hold(&kernel_object);
1502 for (i = 0; i < size; i += PAGE_SIZE) {
e4846942 1503 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
b12defdc
MD
1504 if (i == 0)
1505 mbase = m;
dc1fd4b3
MD
1506
1507 /*
1508 * If the allocation failed we either return NULL or we retry.
1509 *
c397c465
MD
1510 * If M_WAITOK is specified we wait for more memory and retry.
1511 * If M_WAITOK is specified from a preemption we yield instead of
1512 * wait. Livelock will not occur because the interrupt thread
1513 * will not be preempting anyone the second time around after the
1514 * yield.
dc1fd4b3 1515 */
a108bf71 1516 if (m == NULL) {
c397c465 1517 if (flags & M_WAITOK) {
fe1e98d0 1518 if (td->td_preempted) {
77912481 1519 lwkt_switch();
dc1fd4b3 1520 } else {
4ecf7cc9 1521 vm_wait(0);
dc1fd4b3 1522 }
a108bf71
MD
1523 i -= PAGE_SIZE; /* retry */
1524 continue;
1525 }
b12defdc
MD
1526 break;
1527 }
1528 }
dc1fd4b3 1529
b12defdc
MD
1530 /*
1531 * Check and deal with an allocation failure
1532 */
1533 if (i != size) {
1534 while (i != 0) {
1535 i -= PAGE_SIZE;
1536 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1537 /* page should already be busy */
1538 vm_page_free(m);
a108bf71 1539 }
b12defdc
MD
1540 vm_map_lock(&kernel_map);
1541 vm_map_delete(&kernel_map, addr, addr + size, &count);
1542 vm_map_unlock(&kernel_map);
1543 vm_object_drop(&kernel_object);
1544
1545 vm_map_entry_release(count);
1546 crit_exit();
1547 return(NULL);
a108bf71
MD
1548 }
1549
1550 /*
dc1fd4b3
MD
1551 * Success!
1552 *
b12defdc
MD
1553 * NOTE: The VM pages are still busied. mbase points to the first one
1554 * but we have to iterate via vm_page_next()
a108bf71 1555 */
b12defdc 1556 vm_object_drop(&kernel_object);
a108bf71
MD
1557 crit_exit();
1558
1559 /*
1560 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1561 */
b12defdc
MD
1562 m = mbase;
1563 i = 0;
a108bf71 1564
b12defdc
MD
1565 while (i < size) {
1566 /*
1567 * page should already be busy
1568 */
a108bf71
MD
1569 m->valid = VM_PAGE_BITS_ALL;
1570 vm_page_wire(m);
921c891e
MD
1571 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL | VM_PROT_NOSYNC,
1572 1, NULL);
a108bf71
MD
1573 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1574 bzero((char *)addr + i, PAGE_SIZE);
1575 vm_page_flag_clear(m, PG_ZERO);
17cde63e
MD
1576 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1577 vm_page_flag_set(m, PG_REFERENCED);
a491077e 1578 vm_page_wakeup(m);
b12defdc
MD
1579
1580 i += PAGE_SIZE;
1581 vm_object_hold(&kernel_object);
1582 m = vm_page_next(m);
1583 vm_object_drop(&kernel_object);
a108bf71 1584 }
b12defdc 1585 smp_invltlb();
a108bf71
MD
1586 vm_map_entry_release(count);
1587 return((void *)addr);
1588}
1589
38e34349 1590/*
5b287bba 1591 * kmem_slab_free()
38e34349 1592 */
a108bf71
MD
1593static void
1594kmem_slab_free(void *ptr, vm_size_t size)
1595{
1596 crit_enter();
e4846942 1597 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
a108bf71
MD
1598 crit_exit();
1599}
1600
55126ffe 1601void *
55126ffe
SZ
1602kmalloc_cachealign(unsigned long size_alloc, struct malloc_type *type,
1603 int flags)
1604{
1605 if (size_alloc < __VM_CACHELINE_SIZE)
1606 size_alloc = __VM_CACHELINE_SIZE;
1e57f867 1607 return kmalloc(size_alloc, type, flags | M_POWEROF2);
55126ffe 1608}