kmalloc: No need to call fls if the size is already power-of-2 aligned
[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{
18042a75 508 int i;
1e57f867 509
18042a75
SZ
510 if (size == 0 || powerof2(size))
511 return size;
1e57f867
SZ
512
513 i = flsl(size);
1e57f867
SZ
514 return (1UL << i);
515}
516
a108bf71 517/*
5fee07e6 518 * kmalloc() (SLAB ALLOCATOR)
a108bf71
MD
519 *
520 * Allocate memory via the slab allocator. If the request is too large,
521 * or if it page-aligned beyond a certain size, we fall back to the
522 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
523 * &SlabMisc if you don't care.
524 *
8cb2bf45
JS
525 * M_RNOWAIT - don't block.
526 * M_NULLOK - return NULL instead of blocking.
a108bf71 527 * M_ZERO - zero the returned memory.
dc1fd4b3
MD
528 * M_USE_RESERVE - allow greater drawdown of the free list
529 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
1e57f867 530 * M_POWEROF2 - roundup size to the nearest power of 2
5b287bba
MD
531 *
532 * MPSAFE
a108bf71 533 */
bbb201fd
MD
534
535#ifdef SLAB_DEBUG
536void *
537kmalloc_debug(unsigned long size, struct malloc_type *type, int flags,
538 const char *file, int line)
539#else
a108bf71 540void *
8aca2bd4 541kmalloc(unsigned long size, struct malloc_type *type, int flags)
bbb201fd 542#endif
a108bf71
MD
543{
544 SLZone *z;
545 SLChunk *chunk;
d8100bdc 546#ifdef SMP
5fee07e6 547 SLChunk *bchunk;
d8100bdc 548#endif
a108bf71 549 SLGlobalData *slgd;
bba6a44d 550 struct globaldata *gd;
81dedbc2 551 unsigned long align;
a108bf71 552 int zi;
d2182dc1
MD
553#ifdef INVARIANTS
554 int i;
555#endif
a108bf71 556
b68ad50c 557 logmemory_quick(malloc_beg);
bba6a44d
MD
558 gd = mycpu;
559 slgd = &gd->gd_slab;
a108bf71
MD
560
561 /*
562 * XXX silly to have this in the critical path.
563 */
564 if (type->ks_limit == 0) {
565 crit_enter();
566 if (type->ks_limit == 0)
567 malloc_init(type);
568 crit_exit();
569 }
570 ++type->ks_calls;
571
1e57f867
SZ
572 if (flags & M_POWEROF2)
573 size = powerof2_size(size);
574
a108bf71 575 /*
38e34349
MD
576 * Handle the case where the limit is reached. Panic if we can't return
577 * NULL. The original malloc code looped, but this tended to
a108bf71 578 * simply deadlock the computer.
38e34349
MD
579 *
580 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
581 * to determine if a more complete limit check should be done. The
582 * actual memory use is tracked via ks_memuse[cpu].
a108bf71 583 */
bba6a44d
MD
584 while (type->ks_loosememuse >= type->ks_limit) {
585 int i;
586 long ttl;
587
588 for (i = ttl = 0; i < ncpus; ++i)
589 ttl += type->ks_memuse[i];
38e34349 590 type->ks_loosememuse = ttl; /* not MP synchronized */
28135cc2
MD
591 if ((ssize_t)ttl < 0) /* deal with occassional race */
592 ttl = 0;
bba6a44d 593 if (ttl >= type->ks_limit) {
f2b5daf9 594 if (flags & M_NULLOK) {
5fee07e6 595 logmemory(malloc_end, NULL, type, size, flags);
bba6a44d 596 return(NULL);
f2b5daf9 597 }
bba6a44d
MD
598 panic("%s: malloc limit exceeded", type->ks_shortdesc);
599 }
a108bf71
MD
600 }
601
602 /*
603 * Handle the degenerate size == 0 case. Yes, this does happen.
604 * Return a special pointer. This is to maintain compatibility with
605 * the original malloc implementation. Certain devices, such as the
606 * adaptec driver, not only allocate 0 bytes, they check for NULL and
607 * also realloc() later on. Joy.
608 */
f2b5daf9 609 if (size == 0) {
5fee07e6 610 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
a108bf71 611 return(ZERO_LENGTH_PTR);
f2b5daf9 612 }
a108bf71
MD
613
614 /*
a7cf0021
MD
615 * Handle hysteresis from prior frees here in malloc(). We cannot
616 * safely manipulate the kernel_map in free() due to free() possibly
617 * being called via an IPI message or from sensitive interrupt code.
5fee07e6
MD
618 *
619 * NOTE: ku_pagecnt must be cleared before we free the slab or we
620 * might race another cpu allocating the kva and setting
621 * ku_pagecnt.
a7cf0021 622 */
ad94a851 623 while (slgd->NFreeZones > ZoneRelsThresh && (flags & M_RNOWAIT) == 0) {
46a3f46d 624 crit_enter();
ad94a851 625 if (slgd->NFreeZones > ZoneRelsThresh) { /* crit sect race */
722871d3 626 int *kup;
5fee07e6 627
46a3f46d
MD
628 z = slgd->FreeZones;
629 slgd->FreeZones = z->z_Next;
630 --slgd->NFreeZones;
5fee07e6 631 kup = btokup(z);
722871d3 632 *kup = 0;
46a3f46d 633 kmem_slab_free(z, ZoneSize); /* may block */
51295aee 634 atomic_add_int(&ZoneGenAlloc, -ZoneSize / 1024);
46a3f46d
MD
635 }
636 crit_exit();
637 }
5fee07e6 638
46a3f46d 639 /*
5fee07e6 640 * XXX handle oversized frees that were queued from kfree().
46a3f46d 641 */
dc1fd4b3 642 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) {
46a3f46d
MD
643 crit_enter();
644 if ((z = slgd->FreeOvZones) != NULL) {
7e83df33
MD
645 vm_size_t tsize;
646
46a3f46d
MD
647 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
648 slgd->FreeOvZones = z->z_Next;
7e83df33
MD
649 tsize = z->z_ChunkSize;
650 kmem_slab_free(z, tsize); /* may block */
651 atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024);
46a3f46d
MD
652 }
653 crit_exit();
a7cf0021
MD
654 }
655
656 /*
a108bf71
MD
657 * Handle large allocations directly. There should not be very many of
658 * these so performance is not a big issue.
659 *
b543eeed
MD
660 * The backend allocator is pretty nasty on a SMP system. Use the
661 * slab allocator for one and two page-sized chunks even though we lose
662 * some efficiency. XXX maybe fix mmio and the elf loader instead.
a108bf71 663 */
b543eeed 664 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
722871d3 665 int *kup;
a108bf71
MD
666
667 size = round_page(size);
668 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
f2b5daf9 669 if (chunk == NULL) {
5fee07e6 670 logmemory(malloc_end, NULL, type, size, flags);
a108bf71 671 return(NULL);
f2b5daf9 672 }
665206ee 673 atomic_add_int(&ZoneBigAlloc, (int)size / 1024);
a108bf71 674 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
8f1d5415 675 flags |= M_PASSIVE_ZERO;
a108bf71 676 kup = btokup(chunk);
722871d3 677 *kup = size / PAGE_SIZE;
a108bf71
MD
678 crit_enter();
679 goto done;
680 }
681
682 /*
683 * Attempt to allocate out of an existing zone. First try the free list,
684 * then allocate out of unallocated space. If we find a good zone move
685 * it to the head of the list so later allocations find it quickly
686 * (we might have thousands of zones in the list).
687 *
688 * Note: zoneindex() will panic of size is too large.
689 */
81dedbc2 690 zi = zoneindex(&size, &align);
a108bf71
MD
691 KKASSERT(zi < NZONES);
692 crit_enter();
a108bf71 693
5fee07e6 694 if ((z = slgd->ZoneAry[zi]) != NULL) {
a108bf71 695 /*
5fee07e6
MD
696 * Locate a chunk - we have to have at least one. If this is the
697 * last chunk go ahead and do the work to retrieve chunks freed
698 * from remote cpus, and if the zone is still empty move it off
699 * the ZoneAry.
a108bf71 700 */
5fee07e6
MD
701 if (--z->z_NFree <= 0) {
702 KKASSERT(z->z_NFree == 0);
703
704#ifdef SMP
705 /*
706 * WARNING! This code competes with other cpus. It is ok
707 * for us to not drain RChunks here but we might as well, and
708 * it is ok if more accumulate after we're done.
709 *
710 * Set RSignal before pulling rchunks off, indicating that we
711 * will be moving ourselves off of the ZoneAry. Remote ends will
712 * read RSignal before putting rchunks on thus interlocking
713 * their IPI signaling.
714 */
715 if (z->z_RChunks == NULL)
716 atomic_swap_int(&z->z_RSignal, 1);
717
718 while ((bchunk = z->z_RChunks) != NULL) {
719 cpu_ccfence();
720 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
721 *z->z_LChunksp = bchunk;
722 while (bchunk) {
723 chunk_mark_free(z, bchunk);
724 z->z_LChunksp = &bchunk->c_Next;
725 bchunk = bchunk->c_Next;
726 ++z->z_NFree;
727 }
728 break;
729 }
730 }
731#endif
732 /*
733 * Remove from the zone list if no free chunks remain.
734 * Clear RSignal
735 */
736 if (z->z_NFree == 0) {
737 slgd->ZoneAry[zi] = z->z_Next;
738 z->z_Next = NULL;
739 } else {
740 z->z_RSignal = 0;
741 }
a108bf71
MD
742 }
743
744 /*
5fee07e6 745 * Fast path, we have chunks available in z_LChunks.
a108bf71 746 */
5fee07e6
MD
747 chunk = z->z_LChunks;
748 if (chunk) {
10cc6608 749 chunk_mark_allocated(z, chunk);
5fee07e6
MD
750 z->z_LChunks = chunk->c_Next;
751 if (z->z_LChunks == NULL)
752 z->z_LChunksp = &z->z_LChunks;
bbb201fd
MD
753#ifdef SLAB_DEBUG
754 slab_record_source(z, file, line);
755#endif
a108bf71 756 goto done;
a108bf71
MD
757 }
758
759 /*
5fee07e6
MD
760 * No chunks are available in LChunks, the free chunk MUST be
761 * in the never-before-used memory area, controlled by UIndex.
762 *
763 * The consequences are very serious if our zone got corrupted so
764 * we use an explicit panic rather than a KASSERT.
a108bf71 765 */
1c5ca4f3 766 if (z->z_UIndex + 1 != z->z_NMax)
5fee07e6 767 ++z->z_UIndex;
1c5ca4f3
MD
768 else
769 z->z_UIndex = 0;
5fee07e6 770
1c5ca4f3
MD
771 if (z->z_UIndex == z->z_UEndIndex)
772 panic("slaballoc: corrupted zone");
5fee07e6 773
1c5ca4f3 774 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
8f1d5415 775 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 776 flags &= ~M_ZERO;
8f1d5415
MD
777 flags |= M_PASSIVE_ZERO;
778 }
10cc6608 779 chunk_mark_allocated(z, chunk);
bbb201fd
MD
780#ifdef SLAB_DEBUG
781 slab_record_source(z, file, line);
782#endif
a108bf71
MD
783 goto done;
784 }
785
786 /*
787 * If all zones are exhausted we need to allocate a new zone for this
788 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
6ab8e1da
MD
789 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
790 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
791 * we do not pre-zero it because we do not want to mess up the L1 cache.
a108bf71
MD
792 *
793 * At least one subsystem, the tty code (see CROUND) expects power-of-2
794 * allocations to be power-of-2 aligned. We maintain compatibility by
795 * adjusting the base offset below.
796 */
797 {
798 int off;
722871d3 799 int *kup;
a108bf71
MD
800
801 if ((z = slgd->FreeZones) != NULL) {
802 slgd->FreeZones = z->z_Next;
803 --slgd->NFreeZones;
804 bzero(z, sizeof(SLZone));
6ab8e1da 805 z->z_Flags |= SLZF_UNOTZEROD;
a108bf71
MD
806 } else {
807 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
808 if (z == NULL)
809 goto fail;
51295aee 810 atomic_add_int(&ZoneGenAlloc, ZoneSize / 1024);
a108bf71
MD
811 }
812
813 /*
10cc6608
MD
814 * How big is the base structure?
815 */
816#if defined(INVARIANTS)
817 /*
818 * Make room for z_Bitmap. An exact calculation is somewhat more
819 * complicated so don't make an exact calculation.
820 */
821 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
822 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
823#else
824 off = sizeof(SLZone);
825#endif
826
827 /*
a108bf71
MD
828 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
829 * Otherwise just 8-byte align the data.
830 */
831 if ((size | (size - 1)) + 1 == (size << 1))
10cc6608 832 off = (off + size - 1) & ~(size - 1);
a108bf71 833 else
81dedbc2 834 off = (off + align - 1) & ~(align - 1);
a108bf71
MD
835 z->z_Magic = ZALLOC_SLAB_MAGIC;
836 z->z_ZoneIndex = zi;
837 z->z_NMax = (ZoneSize - off) / size;
838 z->z_NFree = z->z_NMax - 1;
1c5ca4f3
MD
839 z->z_BasePtr = (char *)z + off;
840 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
a108bf71 841 z->z_ChunkSize = size;
2db3b277 842 z->z_CpuGd = gd;
bba6a44d 843 z->z_Cpu = gd->gd_cpuid;
5fee07e6 844 z->z_LChunksp = &z->z_LChunks;
bbb201fd
MD
845#ifdef SLAB_DEBUG
846 bcopy(z->z_Sources, z->z_AltSources, sizeof(z->z_Sources));
847 bzero(z->z_Sources, sizeof(z->z_Sources));
848#endif
1c5ca4f3 849 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
a108bf71
MD
850 z->z_Next = slgd->ZoneAry[zi];
851 slgd->ZoneAry[zi] = z;
8f1d5415 852 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 853 flags &= ~M_ZERO; /* already zero'd */
8f1d5415
MD
854 flags |= M_PASSIVE_ZERO;
855 }
5fee07e6 856 kup = btokup(z);
722871d3 857 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */
10cc6608 858 chunk_mark_allocated(z, chunk);
bbb201fd
MD
859#ifdef SLAB_DEBUG
860 slab_record_source(z, file, line);
861#endif
1c5ca4f3
MD
862
863 /*
864 * Slide the base index for initial allocations out of the next
865 * zone we create so we do not over-weight the lower part of the
866 * cpu memory caches.
867 */
868 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
869 & (ZALLOC_MAX_ZONE_SIZE - 1);
a108bf71 870 }
5fee07e6 871
a108bf71 872done:
bba6a44d
MD
873 ++type->ks_inuse[gd->gd_cpuid];
874 type->ks_memuse[gd->gd_cpuid] += size;
38e34349 875 type->ks_loosememuse += size; /* not MP synchronized */
a108bf71 876 crit_exit();
5fee07e6 877
a108bf71
MD
878 if (flags & M_ZERO)
879 bzero(chunk, size);
bba6a44d 880#ifdef INVARIANTS
d2182dc1
MD
881 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
882 if (use_malloc_pattern) {
883 for (i = 0; i < size; i += sizeof(int)) {
884 *(int *)((char *)chunk + i) = -1;
885 }
886 }
bba6a44d 887 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
d2182dc1 888 }
bba6a44d 889#endif
5fee07e6 890 logmemory(malloc_end, chunk, type, size, flags);
a108bf71
MD
891 return(chunk);
892fail:
893 crit_exit();
5fee07e6 894 logmemory(malloc_end, NULL, type, size, flags);
a108bf71
MD
895 return(NULL);
896}
897
38e34349
MD
898/*
899 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
900 *
901 * Generally speaking this routine is not called very often and we do
902 * not attempt to optimize it beyond reusing the same pointer if the
903 * new size fits within the chunking of the old pointer's zone.
904 */
bbb201fd
MD
905#ifdef SLAB_DEBUG
906void *
907krealloc_debug(void *ptr, unsigned long size,
908 struct malloc_type *type, int flags,
909 const char *file, int line)
910#else
a108bf71 911void *
8aca2bd4 912krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
bbb201fd 913#endif
a108bf71 914{
722871d3 915 unsigned long osize;
81dedbc2 916 unsigned long align;
a108bf71
MD
917 SLZone *z;
918 void *nptr;
722871d3 919 int *kup;
a108bf71 920
eb7f3e3c
MD
921 KKASSERT((flags & M_ZERO) == 0); /* not supported */
922
a108bf71 923 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
bbb201fd 924 return(kmalloc_debug(size, type, flags, file, line));
a108bf71 925 if (size == 0) {
efda3bd0 926 kfree(ptr, type);
a108bf71
MD
927 return(NULL);
928 }
929
930 /*
931 * Handle oversized allocations. XXX we really should require that a
932 * size be passed to free() instead of this nonsense.
933 */
5fee07e6 934 kup = btokup(ptr);
722871d3
MD
935 if (*kup > 0) {
936 osize = *kup << PAGE_SHIFT;
5fee07e6
MD
937 if (osize == round_page(size))
938 return(ptr);
bbb201fd 939 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
5fee07e6
MD
940 return(NULL);
941 bcopy(ptr, nptr, min(size, osize));
942 kfree(ptr, type);
943 return(nptr);
a108bf71
MD
944 }
945
946 /*
947 * Get the original allocation's zone. If the new request winds up
948 * using the same chunk size we do not have to do anything.
949 */
5fee07e6
MD
950 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
951 kup = btokup(z);
722871d3 952 KKASSERT(*kup < 0);
a108bf71
MD
953 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
954
a108bf71
MD
955 /*
956 * Allocate memory for the new request size. Note that zoneindex has
957 * already adjusted the request size to the appropriate chunk size, which
958 * should optimize our bcopy(). Then copy and return the new pointer.
1ea6580d
MD
959 *
960 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
961 * necessary align the result.
962 *
963 * We can only zoneindex (to align size to the chunk size) if the new
964 * size is not too large.
a108bf71 965 */
1ea6580d 966 if (size < ZoneLimit) {
81dedbc2 967 zoneindex(&size, &align);
1ea6580d
MD
968 if (z->z_ChunkSize == size)
969 return(ptr);
970 }
bbb201fd 971 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
a108bf71
MD
972 return(NULL);
973 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
efda3bd0 974 kfree(ptr, type);
a108bf71
MD
975 return(nptr);
976}
977
38e34349 978/*
45d2b1d8
MD
979 * Return the kmalloc limit for this type, in bytes.
980 */
981long
982kmalloc_limit(struct malloc_type *type)
983{
984 if (type->ks_limit == 0) {
985 crit_enter();
986 if (type->ks_limit == 0)
987 malloc_init(type);
988 crit_exit();
989 }
990 return(type->ks_limit);
991}
992
993/*
38e34349
MD
994 * Allocate a copy of the specified string.
995 *
996 * (MP SAFE) (MAY BLOCK)
997 */
bbb201fd
MD
998#ifdef SLAB_DEBUG
999char *
1000kstrdup_debug(const char *str, struct malloc_type *type,
1001 const char *file, int line)
1002#else
1ac06773 1003char *
59302080 1004kstrdup(const char *str, struct malloc_type *type)
bbb201fd 1005#endif
1ac06773
MD
1006{
1007 int zlen; /* length inclusive of terminating NUL */
1008 char *nstr;
1009
1010 if (str == NULL)
1011 return(NULL);
1012 zlen = strlen(str) + 1;
bbb201fd 1013 nstr = kmalloc_debug(zlen, type, M_WAITOK, file, line);
1ac06773
MD
1014 bcopy(str, nstr, zlen);
1015 return(nstr);
1016}
1017
1d712609 1018#ifdef SMP
a108bf71 1019/*
5fee07e6 1020 * Notify our cpu that a remote cpu has freed some chunks in a zone that
df9daea8
MD
1021 * we own. RCount will be bumped so the memory should be good, but validate
1022 * that it really is.
a108bf71
MD
1023 */
1024static
1025void
5fee07e6 1026kfree_remote(void *ptr)
a108bf71 1027{
5fee07e6
MD
1028 SLGlobalData *slgd;
1029 SLChunk *bchunk;
1030 SLZone *z;
1031 int nfree;
722871d3 1032 int *kup;
5fee07e6 1033
5fee07e6
MD
1034 slgd = &mycpu->gd_slab;
1035 z = ptr;
1036 kup = btokup(z);
df9daea8
MD
1037 KKASSERT(*kup == -((int)mycpuid + 1));
1038 KKASSERT(z->z_RCount > 0);
1039 atomic_subtract_int(&z->z_RCount, 1);
5fee07e6 1040
5bf48697 1041 logmemory(free_rem_beg, z, NULL, 0L, 0);
df9daea8
MD
1042 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1043 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
1044 nfree = z->z_NFree;
5fee07e6 1045
df9daea8
MD
1046 /*
1047 * Indicate that we will no longer be off of the ZoneAry by
1048 * clearing RSignal.
1049 */
1050 if (z->z_RChunks)
1051 z->z_RSignal = 0;
5fee07e6 1052
df9daea8
MD
1053 /*
1054 * Atomically extract the bchunks list and then process it back
1055 * into the lchunks list. We want to append our bchunks to the
1056 * lchunks list and not prepend since we likely do not have
1057 * cache mastership of the related data (not that it helps since
1058 * we are using c_Next).
1059 */
1060 while ((bchunk = z->z_RChunks) != NULL) {
1061 cpu_ccfence();
1062 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
1063 *z->z_LChunksp = bchunk;
1064 while (bchunk) {
1065 chunk_mark_free(z, bchunk);
1066 z->z_LChunksp = &bchunk->c_Next;
1067 bchunk = bchunk->c_Next;
1068 ++z->z_NFree;
5fee07e6 1069 }
df9daea8 1070 break;
5fee07e6 1071 }
df9daea8
MD
1072 }
1073 if (z->z_NFree && nfree == 0) {
1074 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1075 slgd->ZoneAry[z->z_ZoneIndex] = z;
1076 }
5fee07e6 1077
df9daea8
MD
1078 /*
1079 * If the zone becomes totally free, and there are other zones we
1080 * can allocate from, move this zone to the FreeZones list. Since
1081 * this code can be called from an IPI callback, do *NOT* try to mess
1082 * with kernel_map here. Hysteresis will be performed at malloc() time.
1083 *
1084 * Do not move the zone if there is an IPI inflight, otherwise MP
1085 * races can result in our free_remote code accessing a destroyed
1086 * zone.
1087 */
1088 if (z->z_NFree == z->z_NMax &&
1089 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1090 z->z_RCount == 0
1091 ) {
1092 SLZone **pz;
1093 int *kup;
5fee07e6 1094
df9daea8
MD
1095 for (pz = &slgd->ZoneAry[z->z_ZoneIndex];
1096 z != *pz;
1097 pz = &(*pz)->z_Next) {
1098 ;
5fee07e6 1099 }
df9daea8
MD
1100 *pz = z->z_Next;
1101 z->z_Magic = -1;
1102 z->z_Next = slgd->FreeZones;
1103 slgd->FreeZones = z;
1104 ++slgd->NFreeZones;
1105 kup = btokup(z);
1106 *kup = 0;
5fee07e6 1107 }
5bf48697 1108 logmemory(free_rem_end, z, bchunk, 0L, 0);
a108bf71
MD
1109}
1110
1d712609
MD
1111#endif
1112
38e34349 1113/*
5b287bba 1114 * free (SLAB ALLOCATOR)
38e34349
MD
1115 *
1116 * Free a memory block previously allocated by malloc. Note that we do not
5fee07e6 1117 * attempt to update ks_loosememuse as MP races could prevent us from
38e34349 1118 * checking memory limits in malloc.
5b287bba
MD
1119 *
1120 * MPSAFE
38e34349 1121 */
a108bf71 1122void
8aca2bd4 1123kfree(void *ptr, struct malloc_type *type)
a108bf71
MD
1124{
1125 SLZone *z;
1126 SLChunk *chunk;
1127 SLGlobalData *slgd;
bba6a44d 1128 struct globaldata *gd;
722871d3 1129 int *kup;
5fee07e6 1130 unsigned long size;
d8100bdc
SW
1131#ifdef SMP
1132 SLChunk *bchunk;
5fee07e6 1133 int rsignal;
d8100bdc 1134#endif
a108bf71 1135
b68ad50c 1136 logmemory_quick(free_beg);
bba6a44d
MD
1137 gd = mycpu;
1138 slgd = &gd->gd_slab;
a108bf71 1139
d39911d9
JS
1140 if (ptr == NULL)
1141 panic("trying to free NULL pointer");
1142
a108bf71
MD
1143 /*
1144 * Handle special 0-byte allocations
1145 */
f2b5daf9 1146 if (ptr == ZERO_LENGTH_PTR) {
5bf48697 1147 logmemory(free_zero, ptr, type, -1UL, 0);
b68ad50c 1148 logmemory_quick(free_end);
a108bf71 1149 return;
f2b5daf9 1150 }
a108bf71
MD
1151
1152 /*
5fee07e6
MD
1153 * Panic on bad malloc type
1154 */
1155 if (type->ks_magic != M_MAGIC)
1156 panic("free: malloc type lacks magic");
1157
1158 /*
a108bf71
MD
1159 * Handle oversized allocations. XXX we really should require that a
1160 * size be passed to free() instead of this nonsense.
bba6a44d
MD
1161 *
1162 * This code is never called via an ipi.
a108bf71 1163 */
5fee07e6 1164 kup = btokup(ptr);
722871d3
MD
1165 if (*kup > 0) {
1166 size = *kup << PAGE_SHIFT;
1167 *kup = 0;
a108bf71 1168#ifdef INVARIANTS
5fee07e6
MD
1169 KKASSERT(sizeof(weirdary) <= size);
1170 bcopy(weirdary, ptr, sizeof(weirdary));
a108bf71 1171#endif
5fee07e6
MD
1172 /*
1173 * NOTE: For oversized allocations we do not record the
1174 * originating cpu. It gets freed on the cpu calling
1175 * kfree(). The statistics are in aggregate.
1176 *
1177 * note: XXX we have still inherited the interrupts-can't-block
1178 * assumption. An interrupt thread does not bump
1179 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1180 * primarily until we can fix softupdate's assumptions about free().
1181 */
1182 crit_enter();
1183 --type->ks_inuse[gd->gd_cpuid];
1184 type->ks_memuse[gd->gd_cpuid] -= size;
1185 if (mycpu->gd_intr_nesting_level ||
1186 (gd->gd_curthread->td_flags & TDF_INTTHREAD))
1187 {
1188 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1189 z = (SLZone *)ptr;
1190 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1191 z->z_Next = slgd->FreeOvZones;
1192 z->z_ChunkSize = size;
1193 slgd->FreeOvZones = z;
1194 crit_exit();
1195 } else {
1196 crit_exit();
1197 logmemory(free_ovsz, ptr, type, size, 0);
1198 kmem_slab_free(ptr, size); /* may block */
1199 atomic_add_int(&ZoneBigAlloc, -(int)size / 1024);
a108bf71 1200 }
5fee07e6
MD
1201 logmemory_quick(free_end);
1202 return;
a108bf71
MD
1203 }
1204
1205 /*
1206 * Zone case. Figure out the zone based on the fact that it is
1207 * ZoneSize aligned.
1208 */
5fee07e6
MD
1209 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1210 kup = btokup(z);
722871d3 1211 KKASSERT(*kup < 0);
a108bf71
MD
1212 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1213
1214 /*
5fee07e6
MD
1215 * If we do not own the zone then use atomic ops to free to the
1216 * remote cpu linked list and notify the target zone using a
1217 * passive message.
1218 *
1219 * The target zone cannot be deallocated while we own a chunk of it,
1220 * so the zone header's storage is stable until the very moment
1221 * we adjust z_RChunks. After that we cannot safely dereference (z).
1222 *
1223 * (no critical section needed)
a108bf71 1224 */
2db3b277 1225 if (z->z_CpuGd != gd) {
75c7ffea 1226#ifdef SMP
5fee07e6
MD
1227 /*
1228 * Making these adjustments now allow us to avoid passing (type)
1229 * to the remote cpu. Note that ks_inuse/ks_memuse is being
28135cc2
MD
1230 * adjusted on OUR cpu, not the zone cpu, but it should all still
1231 * sum up properly and cancel out.
5fee07e6 1232 */
28135cc2
MD
1233 crit_enter();
1234 --type->ks_inuse[gd->gd_cpuid];
1235 type->ks_memuse[gd->gd_cpuid] -= z->z_ChunkSize;
1236 crit_exit();
5fee07e6
MD
1237
1238 /*
1239 * WARNING! This code competes with other cpus. Once we
1240 * successfully link the chunk to RChunks the remote
1241 * cpu can rip z's storage out from under us.
df9daea8
MD
1242 *
1243 * Bumping RCount prevents z's storage from getting
1244 * ripped out.
5fee07e6
MD
1245 */
1246 rsignal = z->z_RSignal;
1247 cpu_lfence();
df9daea8
MD
1248 if (rsignal)
1249 atomic_add_int(&z->z_RCount, 1);
5fee07e6
MD
1250
1251 chunk = ptr;
1252 for (;;) {
1253 bchunk = z->z_RChunks;
1254 cpu_ccfence();
1255 chunk->c_Next = bchunk;
1256 cpu_sfence();
1257
1258 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1259 break;
1260 }
5fee07e6
MD
1261
1262 /*
1263 * We have to signal the remote cpu if our actions will cause
1264 * the remote zone to be placed back on ZoneAry so it can
1265 * move the zone back on.
1266 *
1267 * We only need to deal with NULL->non-NULL RChunk transitions
1268 * and only if z_RSignal is set. We interlock by reading rsignal
1269 * before adding our chunk to RChunks. This should result in
1270 * virtually no IPI traffic.
1271 *
1272 * We can use a passive IPI to reduce overhead even further.
1273 */
1274 if (bchunk == NULL && rsignal) {
5bf48697 1275 logmemory(free_request, ptr, type, (unsigned long)z->z_ChunkSize, 0);
5fee07e6 1276 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
df9daea8
MD
1277 /* z can get ripped out from under us from this point on */
1278 } else if (rsignal) {
1279 atomic_subtract_int(&z->z_RCount, 1);
1280 /* z can get ripped out from under us from this point on */
5fee07e6 1281 }
75c7ffea
MD
1282#else
1283 panic("Corrupt SLZone");
1284#endif
b68ad50c 1285 logmemory_quick(free_end);
a108bf71
MD
1286 return;
1287 }
1288
5fee07e6
MD
1289 /*
1290 * kfree locally
1291 */
5bf48697 1292 logmemory(free_chunk, ptr, type, (unsigned long)z->z_ChunkSize, 0);
f2b5daf9 1293
a108bf71 1294 crit_enter();
a108bf71 1295 chunk = ptr;
10cc6608 1296 chunk_mark_free(z, chunk);
a108bf71
MD
1297
1298 /*
1299 * Put weird data into the memory to detect modifications after freeing,
1300 * illegal pointer use after freeing (we should fault on the odd address),
1301 * and so forth. XXX needs more work, see the old malloc code.
1302 */
1303#ifdef INVARIANTS
1304 if (z->z_ChunkSize < sizeof(weirdary))
1305 bcopy(weirdary, chunk, z->z_ChunkSize);
1306 else
1307 bcopy(weirdary, chunk, sizeof(weirdary));
1308#endif
1309
1310 /*
5fee07e6
MD
1311 * Add this free non-zero'd chunk to a linked list for reuse. Add
1312 * to the front of the linked list so it is more likely to be
1313 * reallocated, since it is already in our L1 cache.
a108bf71 1314 */
6ab8e1da 1315#ifdef INVARIANTS
c439ad8f 1316 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
fc92d4aa 1317 panic("BADFREE %p", chunk);
a108bf71 1318#endif
5fee07e6
MD
1319 chunk->c_Next = z->z_LChunks;
1320 z->z_LChunks = chunk;
1321 if (chunk->c_Next == NULL)
1322 z->z_LChunksp = &chunk->c_Next;
1323
6ab8e1da 1324#ifdef INVARIANTS
c439ad8f 1325 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
a108bf71 1326 panic("BADFREE2");
6ab8e1da 1327#endif
a108bf71
MD
1328
1329 /*
1330 * Bump the number of free chunks. If it becomes non-zero the zone
1331 * must be added back onto the appropriate list.
1332 */
1333 if (z->z_NFree++ == 0) {
1334 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1335 slgd->ZoneAry[z->z_ZoneIndex] = z;
1336 }
1337
bba6a44d
MD
1338 --type->ks_inuse[z->z_Cpu];
1339 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
a108bf71
MD
1340
1341 /*
1342 * If the zone becomes totally free, and there are other zones we
a7cf0021
MD
1343 * can allocate from, move this zone to the FreeZones list. Since
1344 * this code can be called from an IPI callback, do *NOT* try to mess
1345 * with kernel_map here. Hysteresis will be performed at malloc() time.
a108bf71
MD
1346 */
1347 if (z->z_NFree == z->z_NMax &&
df9daea8
MD
1348 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1349 z->z_RCount == 0
a108bf71
MD
1350 ) {
1351 SLZone **pz;
722871d3 1352 int *kup;
a108bf71
MD
1353
1354 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
1355 ;
1356 *pz = z->z_Next;
1357 z->z_Magic = -1;
a7cf0021
MD
1358 z->z_Next = slgd->FreeZones;
1359 slgd->FreeZones = z;
1360 ++slgd->NFreeZones;
5fee07e6 1361 kup = btokup(z);
722871d3 1362 *kup = 0;
a108bf71 1363 }
b68ad50c 1364 logmemory_quick(free_end);
a108bf71
MD
1365 crit_exit();
1366}
1367
10cc6608 1368#if defined(INVARIANTS)
5fee07e6 1369
10cc6608
MD
1370/*
1371 * Helper routines for sanity checks
1372 */
1373static
1374void
1375chunk_mark_allocated(SLZone *z, void *chunk)
1376{
1377 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1378 __uint32_t *bitptr;
1379
5fee07e6
MD
1380 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1381 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1382 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
10cc6608
MD
1383 bitptr = &z->z_Bitmap[bitdex >> 5];
1384 bitdex &= 31;
5fee07e6
MD
1385 KASSERT((*bitptr & (1 << bitdex)) == 0,
1386 ("memory chunk %p is already allocated!", chunk));
10cc6608
MD
1387 *bitptr |= 1 << bitdex;
1388}
1389
1390static
1391void
1392chunk_mark_free(SLZone *z, void *chunk)
1393{
1394 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1395 __uint32_t *bitptr;
1396
5fee07e6
MD
1397 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1398 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1399 ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
10cc6608
MD
1400 bitptr = &z->z_Bitmap[bitdex >> 5];
1401 bitdex &= 31;
5fee07e6
MD
1402 KASSERT((*bitptr & (1 << bitdex)) != 0,
1403 ("memory chunk %p is already free!", chunk));
10cc6608
MD
1404 *bitptr &= ~(1 << bitdex);
1405}
1406
1407#endif
1408
a108bf71 1409/*
5b287bba 1410 * kmem_slab_alloc()
a108bf71
MD
1411 *
1412 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1413 * specified alignment. M_* flags are expected in the flags field.
1414 *
1415 * Alignment must be a multiple of PAGE_SIZE.
1416 *
1417 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1418 * but when we move zalloc() over to use this function as its backend
1419 * we will have to switch to kreserve/krelease and call reserve(0)
1420 * after the new space is made available.
dc1fd4b3
MD
1421 *
1422 * Interrupt code which has preempted other code is not allowed to
c397c465
MD
1423 * use PQ_CACHE pages. However, if an interrupt thread is run
1424 * non-preemptively or blocks and then runs non-preemptively, then
b12defdc 1425 * it is free to use PQ_CACHE pages. <--- may not apply any longer XXX
a108bf71
MD
1426 */
1427static void *
1428kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1429{
1430 vm_size_t i;
1431 vm_offset_t addr;
1de1e800 1432 int count, vmflags, base_vmflags;
b12defdc
MD
1433 vm_page_t mbase = NULL;
1434 vm_page_t m;
dc1fd4b3 1435 thread_t td;
a108bf71
MD
1436
1437 size = round_page(size);
e4846942 1438 addr = vm_map_min(&kernel_map);
a108bf71 1439
a108bf71
MD
1440 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1441 crit_enter();
e4846942 1442 vm_map_lock(&kernel_map);
c809941b 1443 if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) {
e4846942 1444 vm_map_unlock(&kernel_map);
8cb2bf45 1445 if ((flags & M_NULLOK) == 0)
a108bf71 1446 panic("kmem_slab_alloc(): kernel_map ran out of space!");
a108bf71 1447 vm_map_entry_release(count);
2de4f77e 1448 crit_exit();
a108bf71
MD
1449 return(NULL);
1450 }
e4846942
MD
1451
1452 /*
1453 * kernel_object maps 1:1 to kernel_map.
1454 */
b12defdc
MD
1455 vm_object_hold(&kernel_object);
1456 vm_object_reference_locked(&kernel_object);
e4846942
MD
1457 vm_map_insert(&kernel_map, &count,
1458 &kernel_object, addr, addr, addr + size,
1b874851
MD
1459 VM_MAPTYPE_NORMAL,
1460 VM_PROT_ALL, VM_PROT_ALL,
1461 0);
b12defdc
MD
1462 vm_object_drop(&kernel_object);
1463 vm_map_set_wired_quick(&kernel_map, addr, size, &count);
1464 vm_map_unlock(&kernel_map);
a108bf71 1465
dc1fd4b3 1466 td = curthread;
dc1fd4b3 1467
1de1e800
JS
1468 base_vmflags = 0;
1469 if (flags & M_ZERO)
1470 base_vmflags |= VM_ALLOC_ZERO;
1471 if (flags & M_USE_RESERVE)
1472 base_vmflags |= VM_ALLOC_SYSTEM;
1473 if (flags & M_USE_INTERRUPT_RESERVE)
1474 base_vmflags |= VM_ALLOC_INTERRUPT;
77912481
MD
1475 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1476 panic("kmem_slab_alloc: bad flags %08x (%p)",
1477 flags, ((int **)&size)[-1]);
1478 }
1de1e800 1479
a108bf71 1480 /*
b12defdc
MD
1481 * Allocate the pages. Do not mess with the PG_ZERO flag or map
1482 * them yet. VM_ALLOC_NORMAL can only be set if we are not preempting.
1483 *
1484 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1485 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1486 * implied in this case), though I'm not sure if we really need to
1487 * do that.
a108bf71 1488 */
b12defdc
MD
1489 vmflags = base_vmflags;
1490 if (flags & M_WAITOK) {
1491 if (td->td_preempted)
1492 vmflags |= VM_ALLOC_SYSTEM;
1493 else
1494 vmflags |= VM_ALLOC_NORMAL;
1495 }
a108bf71 1496
b12defdc
MD
1497 vm_object_hold(&kernel_object);
1498 for (i = 0; i < size; i += PAGE_SIZE) {
e4846942 1499 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
b12defdc
MD
1500 if (i == 0)
1501 mbase = m;
dc1fd4b3
MD
1502
1503 /*
1504 * If the allocation failed we either return NULL or we retry.
1505 *
c397c465
MD
1506 * If M_WAITOK is specified we wait for more memory and retry.
1507 * If M_WAITOK is specified from a preemption we yield instead of
1508 * wait. Livelock will not occur because the interrupt thread
1509 * will not be preempting anyone the second time around after the
1510 * yield.
dc1fd4b3 1511 */
a108bf71 1512 if (m == NULL) {
c397c465 1513 if (flags & M_WAITOK) {
fe1e98d0 1514 if (td->td_preempted) {
77912481 1515 lwkt_switch();
dc1fd4b3 1516 } else {
4ecf7cc9 1517 vm_wait(0);
dc1fd4b3 1518 }
a108bf71
MD
1519 i -= PAGE_SIZE; /* retry */
1520 continue;
1521 }
b12defdc
MD
1522 break;
1523 }
1524 }
dc1fd4b3 1525
b12defdc
MD
1526 /*
1527 * Check and deal with an allocation failure
1528 */
1529 if (i != size) {
1530 while (i != 0) {
1531 i -= PAGE_SIZE;
1532 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1533 /* page should already be busy */
1534 vm_page_free(m);
a108bf71 1535 }
b12defdc
MD
1536 vm_map_lock(&kernel_map);
1537 vm_map_delete(&kernel_map, addr, addr + size, &count);
1538 vm_map_unlock(&kernel_map);
1539 vm_object_drop(&kernel_object);
1540
1541 vm_map_entry_release(count);
1542 crit_exit();
1543 return(NULL);
a108bf71
MD
1544 }
1545
1546 /*
dc1fd4b3
MD
1547 * Success!
1548 *
b12defdc
MD
1549 * NOTE: The VM pages are still busied. mbase points to the first one
1550 * but we have to iterate via vm_page_next()
a108bf71 1551 */
b12defdc 1552 vm_object_drop(&kernel_object);
a108bf71
MD
1553 crit_exit();
1554
1555 /*
1556 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1557 */
b12defdc
MD
1558 m = mbase;
1559 i = 0;
a108bf71 1560
b12defdc
MD
1561 while (i < size) {
1562 /*
1563 * page should already be busy
1564 */
a108bf71
MD
1565 m->valid = VM_PAGE_BITS_ALL;
1566 vm_page_wire(m);
921c891e
MD
1567 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL | VM_PROT_NOSYNC,
1568 1, NULL);
a108bf71
MD
1569 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1570 bzero((char *)addr + i, PAGE_SIZE);
1571 vm_page_flag_clear(m, PG_ZERO);
17cde63e
MD
1572 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1573 vm_page_flag_set(m, PG_REFERENCED);
a491077e 1574 vm_page_wakeup(m);
b12defdc
MD
1575
1576 i += PAGE_SIZE;
1577 vm_object_hold(&kernel_object);
1578 m = vm_page_next(m);
1579 vm_object_drop(&kernel_object);
a108bf71 1580 }
b12defdc 1581 smp_invltlb();
a108bf71
MD
1582 vm_map_entry_release(count);
1583 return((void *)addr);
1584}
1585
38e34349 1586/*
5b287bba 1587 * kmem_slab_free()
38e34349 1588 */
a108bf71
MD
1589static void
1590kmem_slab_free(void *ptr, vm_size_t size)
1591{
1592 crit_enter();
e4846942 1593 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
a108bf71
MD
1594 crit_exit();
1595}
1596
55126ffe 1597void *
55126ffe
SZ
1598kmalloc_cachealign(unsigned long size_alloc, struct malloc_type *type,
1599 int flags)
1600{
3b556fa0
SZ
1601#if (__VM_CACHELINE_SIZE == 32)
1602#define CAN_CACHEALIGN(sz) ((sz) >= 256)
1603#elif (__VM_CACHELINE_SIZE == 64)
1604#define CAN_CACHEALIGN(sz) ((sz) >= 512)
1605#elif (__VM_CACHELINE_SIZE == 128)
1606#define CAN_CACHEALIGN(sz) ((sz) >= 1024)
1607#else
1608#error "unsupported cacheline size"
1609#endif
1610
1611 void *ret;
1612
55126ffe
SZ
1613 if (size_alloc < __VM_CACHELINE_SIZE)
1614 size_alloc = __VM_CACHELINE_SIZE;
3b556fa0
SZ
1615 else if (!CAN_CACHEALIGN(size_alloc))
1616 flags |= M_POWEROF2;
1617
1618 ret = kmalloc(size_alloc, type, flags);
1619 KASSERT(((uintptr_t)ret & (__VM_CACHELINE_SIZE - 1)) == 0,
1620 ("%p(%lu) not cacheline %d aligned",
1621 ret, size_alloc, __VM_CACHELINE_SIZE));
1622 return ret;
1623
1624#undef CAN_CACHEALIGN
55126ffe 1625}