kernel: Make SMP support default (and non-optional).
[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);
5bf48697
AE
140KTR_INFO(KTR_MEMORY, memory, free_request, 6, MEMORY_STRING, MEMORY_ARGS);
141KTR_INFO(KTR_MEMORY, memory, free_rem_beg, 7, MEMORY_STRING, MEMORY_ARGS);
142KTR_INFO(KTR_MEMORY, memory, free_rem_end, 8, MEMORY_STRING, MEMORY_ARGS);
5bf48697
AE
143KTR_INFO(KTR_MEMORY, memory, free_beg, 9, "free begin");
144KTR_INFO(KTR_MEMORY, memory, free_end, 10, "free end");
f2b5daf9
MD
145
146#define logmemory(name, ptr, type, size, flags) \
147 KTR_LOG(memory_ ## name, ptr, type, size, flags)
b68ad50c
MD
148#define logmemory_quick(name) \
149 KTR_LOG(memory_ ## name)
f2b5daf9 150
a108bf71
MD
151/*
152 * Fixed globals (not per-cpu)
153 */
154static int ZoneSize;
46a3f46d 155static int ZoneLimit;
a108bf71 156static int ZonePageCount;
5fee07e6 157static uintptr_t ZoneMask;
665206ee
MD
158static int ZoneBigAlloc; /* in KB */
159static int ZoneGenAlloc; /* in KB */
460426e6 160struct malloc_type *kmemstatistics; /* exported to vmstat */
a108bf71
MD
161static int32_t weirdary[16];
162
163static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
164static void kmem_slab_free(void *ptr, vm_size_t bytes);
5fee07e6 165
10cc6608
MD
166#if defined(INVARIANTS)
167static void chunk_mark_allocated(SLZone *z, void *chunk);
168static void chunk_mark_free(SLZone *z, void *chunk);
5fee07e6
MD
169#else
170#define chunk_mark_allocated(z, chunk)
171#define chunk_mark_free(z, chunk)
10cc6608 172#endif
a108bf71
MD
173
174/*
175 * Misc constants. Note that allocations that are exact multiples of
176 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
a108bf71 177 */
ad94a851 178#define ZONE_RELS_THRESH 32 /* threshold number of zones */
a108bf71
MD
179
180/*
181 * The WEIRD_ADDR is used as known text to copy into free objects to
182 * try to create deterministic failure cases if the data is accessed after
183 * free.
184 */
185#define WEIRD_ADDR 0xdeadc0de
186#define MAX_COPY sizeof(weirdary)
187#define ZERO_LENGTH_PTR ((void *)-8)
188
189/*
190 * Misc global malloc buckets
191 */
192
193MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
194MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
195MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
196
197MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
198MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
199
200/*
201 * Initialize the slab memory allocator. We have to choose a zone size based
202 * on available physical memory. We choose a zone side which is approximately
203 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
204 * 128K. The zone size is limited to the bounds set in slaballoc.h
205 * (typically 32K min, 128K max).
206 */
207static void kmeminit(void *dummy);
208
c7841cbe
MD
209char *ZeroPage;
210
ba39e2e0 211SYSINIT(kmem, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL)
a108bf71 212
d2182dc1
MD
213#ifdef INVARIANTS
214/*
215 * If enabled any memory allocated without M_ZERO is initialized to -1.
216 */
217static int use_malloc_pattern;
218SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
0c52fa62
SG
219 &use_malloc_pattern, 0,
220 "Initialize memory to -1 if M_ZERO not specified");
d2182dc1
MD
221#endif
222
ad94a851 223static int ZoneRelsThresh = ZONE_RELS_THRESH;
665206ee
MD
224SYSCTL_INT(_kern, OID_AUTO, zone_big_alloc, CTLFLAG_RD, &ZoneBigAlloc, 0, "");
225SYSCTL_INT(_kern, OID_AUTO, zone_gen_alloc, CTLFLAG_RD, &ZoneGenAlloc, 0, "");
ad94a851 226SYSCTL_INT(_kern, OID_AUTO, zone_cache, CTLFLAG_RW, &ZoneRelsThresh, 0, "");
665206ee 227
b12defdc
MD
228/*
229 * Returns the kernel memory size limit for the purposes of initializing
230 * various subsystem caches. The smaller of available memory and the KVM
231 * memory space is returned.
232 *
233 * The size in megabytes is returned.
234 */
235size_t
236kmem_lim_size(void)
237{
238 size_t limsize;
239
240 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
241 if (limsize > KvaSize)
242 limsize = KvaSize;
243 return (limsize / (1024 * 1024));
244}
245
a108bf71
MD
246static void
247kmeminit(void *dummy)
248{
7c457ac8 249 size_t limsize;
a108bf71
MD
250 int usesize;
251 int i;
a108bf71 252
b12defdc
MD
253 limsize = kmem_lim_size();
254 usesize = (int)(limsize * 1024); /* convert to KB */
a108bf71 255
b12defdc
MD
256 /*
257 * If the machine has a large KVM space and more than 8G of ram,
258 * double the zone release threshold to reduce SMP invalidations.
259 * If more than 16G of ram, do it again.
260 *
261 * The BIOS eats a little ram so add some slop. We want 8G worth of
262 * memory sticks to trigger the first adjustment.
263 */
264 if (ZoneRelsThresh == ZONE_RELS_THRESH) {
265 if (limsize >= 7 * 1024)
266 ZoneRelsThresh *= 2;
267 if (limsize >= 15 * 1024)
268 ZoneRelsThresh *= 2;
269 }
a108bf71 270
b12defdc
MD
271 /*
272 * Calculate the zone size. This typically calculates to
273 * ZALLOC_MAX_ZONE_SIZE
274 */
a108bf71
MD
275 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
276 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
277 ZoneSize <<= 1;
46a3f46d
MD
278 ZoneLimit = ZoneSize / 4;
279 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
280 ZoneLimit = ZALLOC_ZONE_LIMIT;
5fee07e6 281 ZoneMask = ~(uintptr_t)(ZoneSize - 1);
a108bf71
MD
282 ZonePageCount = ZoneSize / PAGE_SIZE;
283
a3034532 284 for (i = 0; i < NELEM(weirdary); ++i)
a108bf71
MD
285 weirdary[i] = WEIRD_ADDR;
286
c7841cbe
MD
287 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
288
a108bf71 289 if (bootverbose)
6ea70f76 290 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
a108bf71
MD
291}
292
293/*
bba6a44d 294 * Initialize a malloc type tracking structure.
a108bf71
MD
295 */
296void
297malloc_init(void *data)
298{
299 struct malloc_type *type = data;
7c457ac8 300 size_t limsize;
a108bf71
MD
301
302 if (type->ks_magic != M_MAGIC)
303 panic("malloc type lacks magic");
304
305 if (type->ks_limit != 0)
306 return;
307
308 if (vmstats.v_page_count == 0)
309 panic("malloc_init not allowed before vm init");
310
b12defdc 311 limsize = kmem_lim_size() * (1024 * 1024);
a108bf71
MD
312 type->ks_limit = limsize / 10;
313
314 type->ks_next = kmemstatistics;
315 kmemstatistics = type;
316}
317
318void
319malloc_uninit(void *data)
320{
321 struct malloc_type *type = data;
322 struct malloc_type *t;
bba6a44d
MD
323#ifdef INVARIANTS
324 int i;
1d712609 325 long ttl;
bba6a44d 326#endif
a108bf71
MD
327
328 if (type->ks_magic != M_MAGIC)
329 panic("malloc type lacks magic");
330
331 if (vmstats.v_page_count == 0)
332 panic("malloc_uninit not allowed before vm init");
333
334 if (type->ks_limit == 0)
335 panic("malloc_uninit on uninitialized type");
336
6c92c1f2
SZ
337 /* Make sure that all pending kfree()s are finished. */
338 lwkt_synchronize_ipiqs("muninit");
6c92c1f2 339
a108bf71 340#ifdef INVARIANTS
1d712609
MD
341 /*
342 * memuse is only correct in aggregation. Due to memory being allocated
343 * on one cpu and freed on another individual array entries may be
344 * negative or positive (canceling each other out).
345 */
346 for (i = ttl = 0; i < ncpus; ++i)
347 ttl += type->ks_memuse[i];
348 if (ttl) {
6ea70f76 349 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
1d712609 350 ttl, type->ks_shortdesc, i);
a108bf71
MD
351 }
352#endif
353 if (type == kmemstatistics) {
354 kmemstatistics = type->ks_next;
355 } else {
356 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
357 if (t->ks_next == type) {
358 t->ks_next = type->ks_next;
359 break;
360 }
361 }
362 }
363 type->ks_next = NULL;
364 type->ks_limit = 0;
365}
366
40153c65
MD
367/*
368 * Increase the kmalloc pool limit for the specified pool. No changes
369 * are the made if the pool would shrink.
370 */
371void
372kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
373{
374 if (type->ks_limit == 0)
375 malloc_init(type);
7c457ac8
MD
376 if (bytes == 0)
377 bytes = KvaSize;
40153c65
MD
378 if (type->ks_limit < bytes)
379 type->ks_limit = bytes;
380}
381
ebe36cfe
MD
382/*
383 * Dynamically create a malloc pool. This function is a NOP if *typep is
384 * already non-NULL.
385 */
386void
387kmalloc_create(struct malloc_type **typep, const char *descr)
388{
389 struct malloc_type *type;
390
391 if (*typep == NULL) {
392 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
393 type->ks_magic = M_MAGIC;
394 type->ks_shortdesc = descr;
395 malloc_init(type);
396 *typep = type;
397 }
398}
399
400/*
401 * Destroy a dynamically created malloc pool. This function is a NOP if
402 * the pool has already been destroyed.
403 */
404void
405kmalloc_destroy(struct malloc_type **typep)
406{
407 if (*typep != NULL) {
408 malloc_uninit(*typep);
409 kfree(*typep, M_TEMP);
410 *typep = NULL;
411 }
412}
413
a108bf71
MD
414/*
415 * Calculate the zone index for the allocation request size and set the
416 * allocation request size to that particular zone's chunk size.
417 */
418static __inline int
81dedbc2 419zoneindex(unsigned long *bytes, unsigned long *align)
a108bf71
MD
420{
421 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
422 if (n < 128) {
423 *bytes = n = (n + 7) & ~7;
81dedbc2 424 *align = 8;
a108bf71
MD
425 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
426 }
427 if (n < 256) {
428 *bytes = n = (n + 15) & ~15;
81dedbc2 429 *align = 16;
a108bf71
MD
430 return(n / 16 + 7);
431 }
432 if (n < 8192) {
433 if (n < 512) {
434 *bytes = n = (n + 31) & ~31;
81dedbc2 435 *align = 32;
a108bf71
MD
436 return(n / 32 + 15);
437 }
438 if (n < 1024) {
439 *bytes = n = (n + 63) & ~63;
81dedbc2 440 *align = 64;
a108bf71
MD
441 return(n / 64 + 23);
442 }
443 if (n < 2048) {
444 *bytes = n = (n + 127) & ~127;
81dedbc2 445 *align = 128;
a108bf71
MD
446 return(n / 128 + 31);
447 }
448 if (n < 4096) {
449 *bytes = n = (n + 255) & ~255;
81dedbc2 450 *align = 256;
a108bf71
MD
451 return(n / 256 + 39);
452 }
453 *bytes = n = (n + 511) & ~511;
81dedbc2 454 *align = 512;
a108bf71
MD
455 return(n / 512 + 47);
456 }
457#if ZALLOC_ZONE_LIMIT > 8192
458 if (n < 16384) {
459 *bytes = n = (n + 1023) & ~1023;
81dedbc2 460 *align = 1024;
a108bf71
MD
461 return(n / 1024 + 55);
462 }
463#endif
464#if ZALLOC_ZONE_LIMIT > 16384
465 if (n < 32768) {
466 *bytes = n = (n + 2047) & ~2047;
81dedbc2 467 *align = 2048;
a108bf71
MD
468 return(n / 2048 + 63);
469 }
470#endif
471 panic("Unexpected byte count %d", n);
472 return(0);
473}
474
bbb201fd
MD
475#ifdef SLAB_DEBUG
476/*
477 * Used to debug memory corruption issues. Record up to (typically 32)
478 * allocation sources for this zone (for a particular chunk size).
479 */
480
481static void
482slab_record_source(SLZone *z, const char *file, int line)
483{
484 int i;
485 int b = line & (SLAB_DEBUG_ENTRIES - 1);
486
487 i = b;
488 do {
489 if (z->z_Sources[i].file == file && z->z_Sources[i].line == line)
490 return;
491 if (z->z_Sources[i].file == NULL)
492 break;
493 i = (i + 1) & (SLAB_DEBUG_ENTRIES - 1);
494 } while (i != b);
495 z->z_Sources[i].file = file;
496 z->z_Sources[i].line = line;
497}
498
499#endif
500
1e57f867
SZ
501static __inline unsigned long
502powerof2_size(unsigned long size)
503{
18042a75 504 int i;
1e57f867 505
18042a75
SZ
506 if (size == 0 || powerof2(size))
507 return size;
1e57f867
SZ
508
509 i = flsl(size);
1e57f867
SZ
510 return (1UL << i);
511}
512
a108bf71 513/*
5fee07e6 514 * kmalloc() (SLAB ALLOCATOR)
a108bf71
MD
515 *
516 * Allocate memory via the slab allocator. If the request is too large,
517 * or if it page-aligned beyond a certain size, we fall back to the
518 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
519 * &SlabMisc if you don't care.
520 *
8cb2bf45
JS
521 * M_RNOWAIT - don't block.
522 * M_NULLOK - return NULL instead of blocking.
a108bf71 523 * M_ZERO - zero the returned memory.
dc1fd4b3
MD
524 * M_USE_RESERVE - allow greater drawdown of the free list
525 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
1e57f867 526 * M_POWEROF2 - roundup size to the nearest power of 2
5b287bba
MD
527 *
528 * MPSAFE
a108bf71 529 */
bbb201fd
MD
530
531#ifdef SLAB_DEBUG
532void *
533kmalloc_debug(unsigned long size, struct malloc_type *type, int flags,
534 const char *file, int line)
535#else
a108bf71 536void *
8aca2bd4 537kmalloc(unsigned long size, struct malloc_type *type, int flags)
bbb201fd 538#endif
a108bf71
MD
539{
540 SLZone *z;
541 SLChunk *chunk;
5fee07e6 542 SLChunk *bchunk;
a108bf71 543 SLGlobalData *slgd;
bba6a44d 544 struct globaldata *gd;
81dedbc2 545 unsigned long align;
a108bf71 546 int zi;
d2182dc1
MD
547#ifdef INVARIANTS
548 int i;
549#endif
a108bf71 550
b68ad50c 551 logmemory_quick(malloc_beg);
bba6a44d
MD
552 gd = mycpu;
553 slgd = &gd->gd_slab;
a108bf71
MD
554
555 /*
556 * XXX silly to have this in the critical path.
557 */
558 if (type->ks_limit == 0) {
559 crit_enter();
560 if (type->ks_limit == 0)
561 malloc_init(type);
562 crit_exit();
563 }
564 ++type->ks_calls;
565
1e57f867
SZ
566 if (flags & M_POWEROF2)
567 size = powerof2_size(size);
568
a108bf71 569 /*
38e34349
MD
570 * Handle the case where the limit is reached. Panic if we can't return
571 * NULL. The original malloc code looped, but this tended to
a108bf71 572 * simply deadlock the computer.
38e34349
MD
573 *
574 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
575 * to determine if a more complete limit check should be done. The
576 * actual memory use is tracked via ks_memuse[cpu].
a108bf71 577 */
bba6a44d
MD
578 while (type->ks_loosememuse >= type->ks_limit) {
579 int i;
580 long ttl;
581
582 for (i = ttl = 0; i < ncpus; ++i)
583 ttl += type->ks_memuse[i];
38e34349 584 type->ks_loosememuse = ttl; /* not MP synchronized */
28135cc2
MD
585 if ((ssize_t)ttl < 0) /* deal with occassional race */
586 ttl = 0;
bba6a44d 587 if (ttl >= type->ks_limit) {
f2b5daf9 588 if (flags & M_NULLOK) {
5fee07e6 589 logmemory(malloc_end, NULL, type, size, flags);
bba6a44d 590 return(NULL);
f2b5daf9 591 }
bba6a44d
MD
592 panic("%s: malloc limit exceeded", type->ks_shortdesc);
593 }
a108bf71
MD
594 }
595
596 /*
597 * Handle the degenerate size == 0 case. Yes, this does happen.
598 * Return a special pointer. This is to maintain compatibility with
599 * the original malloc implementation. Certain devices, such as the
600 * adaptec driver, not only allocate 0 bytes, they check for NULL and
601 * also realloc() later on. Joy.
602 */
f2b5daf9 603 if (size == 0) {
5fee07e6 604 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
a108bf71 605 return(ZERO_LENGTH_PTR);
f2b5daf9 606 }
a108bf71 607
a7cf0021
MD
608 /*
609 * Handle hysteresis from prior frees here in malloc(). We cannot
610 * safely manipulate the kernel_map in free() due to free() possibly
611 * being called via an IPI message or from sensitive interrupt code.
5fee07e6
MD
612 *
613 * NOTE: ku_pagecnt must be cleared before we free the slab or we
614 * might race another cpu allocating the kva and setting
615 * ku_pagecnt.
a7cf0021 616 */
ad94a851 617 while (slgd->NFreeZones > ZoneRelsThresh && (flags & M_RNOWAIT) == 0) {
46a3f46d 618 crit_enter();
ad94a851 619 if (slgd->NFreeZones > ZoneRelsThresh) { /* crit sect race */
722871d3 620 int *kup;
5fee07e6 621
46a3f46d
MD
622 z = slgd->FreeZones;
623 slgd->FreeZones = z->z_Next;
624 --slgd->NFreeZones;
5fee07e6 625 kup = btokup(z);
722871d3 626 *kup = 0;
46a3f46d 627 kmem_slab_free(z, ZoneSize); /* may block */
51295aee 628 atomic_add_int(&ZoneGenAlloc, -ZoneSize / 1024);
46a3f46d
MD
629 }
630 crit_exit();
631 }
5fee07e6 632
46a3f46d 633 /*
5fee07e6 634 * XXX handle oversized frees that were queued from kfree().
46a3f46d 635 */
dc1fd4b3 636 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) {
46a3f46d
MD
637 crit_enter();
638 if ((z = slgd->FreeOvZones) != NULL) {
7e83df33
MD
639 vm_size_t tsize;
640
46a3f46d
MD
641 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
642 slgd->FreeOvZones = z->z_Next;
7e83df33
MD
643 tsize = z->z_ChunkSize;
644 kmem_slab_free(z, tsize); /* may block */
645 atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024);
46a3f46d
MD
646 }
647 crit_exit();
a7cf0021
MD
648 }
649
a108bf71
MD
650 /*
651 * Handle large allocations directly. There should not be very many of
652 * these so performance is not a big issue.
653 *
b543eeed
MD
654 * The backend allocator is pretty nasty on a SMP system. Use the
655 * slab allocator for one and two page-sized chunks even though we lose
656 * some efficiency. XXX maybe fix mmio and the elf loader instead.
a108bf71 657 */
b543eeed 658 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
722871d3 659 int *kup;
a108bf71
MD
660
661 size = round_page(size);
662 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
f2b5daf9 663 if (chunk == NULL) {
5fee07e6 664 logmemory(malloc_end, NULL, type, size, flags);
a108bf71 665 return(NULL);
f2b5daf9 666 }
665206ee 667 atomic_add_int(&ZoneBigAlloc, (int)size / 1024);
a108bf71 668 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
8f1d5415 669 flags |= M_PASSIVE_ZERO;
a108bf71 670 kup = btokup(chunk);
722871d3 671 *kup = size / PAGE_SIZE;
a108bf71
MD
672 crit_enter();
673 goto done;
674 }
675
676 /*
677 * Attempt to allocate out of an existing zone. First try the free list,
678 * then allocate out of unallocated space. If we find a good zone move
679 * it to the head of the list so later allocations find it quickly
680 * (we might have thousands of zones in the list).
681 *
682 * Note: zoneindex() will panic of size is too large.
683 */
81dedbc2 684 zi = zoneindex(&size, &align);
a108bf71
MD
685 KKASSERT(zi < NZONES);
686 crit_enter();
a108bf71 687
5fee07e6 688 if ((z = slgd->ZoneAry[zi]) != NULL) {
a108bf71 689 /*
5fee07e6
MD
690 * Locate a chunk - we have to have at least one. If this is the
691 * last chunk go ahead and do the work to retrieve chunks freed
692 * from remote cpus, and if the zone is still empty move it off
693 * the ZoneAry.
a108bf71 694 */
5fee07e6
MD
695 if (--z->z_NFree <= 0) {
696 KKASSERT(z->z_NFree == 0);
697
5fee07e6
MD
698 /*
699 * WARNING! This code competes with other cpus. It is ok
700 * for us to not drain RChunks here but we might as well, and
701 * it is ok if more accumulate after we're done.
702 *
703 * Set RSignal before pulling rchunks off, indicating that we
704 * will be moving ourselves off of the ZoneAry. Remote ends will
705 * read RSignal before putting rchunks on thus interlocking
706 * their IPI signaling.
707 */
708 if (z->z_RChunks == NULL)
709 atomic_swap_int(&z->z_RSignal, 1);
710
711 while ((bchunk = z->z_RChunks) != NULL) {
712 cpu_ccfence();
713 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
714 *z->z_LChunksp = bchunk;
715 while (bchunk) {
716 chunk_mark_free(z, bchunk);
717 z->z_LChunksp = &bchunk->c_Next;
718 bchunk = bchunk->c_Next;
719 ++z->z_NFree;
720 }
721 break;
722 }
723 }
5fee07e6
MD
724 /*
725 * Remove from the zone list if no free chunks remain.
726 * Clear RSignal
727 */
728 if (z->z_NFree == 0) {
729 slgd->ZoneAry[zi] = z->z_Next;
730 z->z_Next = NULL;
731 } else {
732 z->z_RSignal = 0;
733 }
a108bf71
MD
734 }
735
736 /*
5fee07e6 737 * Fast path, we have chunks available in z_LChunks.
a108bf71 738 */
5fee07e6
MD
739 chunk = z->z_LChunks;
740 if (chunk) {
10cc6608 741 chunk_mark_allocated(z, chunk);
5fee07e6
MD
742 z->z_LChunks = chunk->c_Next;
743 if (z->z_LChunks == NULL)
744 z->z_LChunksp = &z->z_LChunks;
bbb201fd
MD
745#ifdef SLAB_DEBUG
746 slab_record_source(z, file, line);
747#endif
a108bf71 748 goto done;
a108bf71
MD
749 }
750
751 /*
5fee07e6
MD
752 * No chunks are available in LChunks, the free chunk MUST be
753 * in the never-before-used memory area, controlled by UIndex.
754 *
755 * The consequences are very serious if our zone got corrupted so
756 * we use an explicit panic rather than a KASSERT.
a108bf71 757 */
1c5ca4f3 758 if (z->z_UIndex + 1 != z->z_NMax)
5fee07e6 759 ++z->z_UIndex;
1c5ca4f3
MD
760 else
761 z->z_UIndex = 0;
5fee07e6 762
1c5ca4f3
MD
763 if (z->z_UIndex == z->z_UEndIndex)
764 panic("slaballoc: corrupted zone");
5fee07e6 765
1c5ca4f3 766 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
8f1d5415 767 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 768 flags &= ~M_ZERO;
8f1d5415
MD
769 flags |= M_PASSIVE_ZERO;
770 }
10cc6608 771 chunk_mark_allocated(z, chunk);
bbb201fd
MD
772#ifdef SLAB_DEBUG
773 slab_record_source(z, file, line);
774#endif
a108bf71
MD
775 goto done;
776 }
777
778 /*
779 * If all zones are exhausted we need to allocate a new zone for this
780 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
6ab8e1da
MD
781 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
782 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
783 * we do not pre-zero it because we do not want to mess up the L1 cache.
a108bf71
MD
784 *
785 * At least one subsystem, the tty code (see CROUND) expects power-of-2
786 * allocations to be power-of-2 aligned. We maintain compatibility by
787 * adjusting the base offset below.
788 */
789 {
790 int off;
722871d3 791 int *kup;
a108bf71
MD
792
793 if ((z = slgd->FreeZones) != NULL) {
794 slgd->FreeZones = z->z_Next;
795 --slgd->NFreeZones;
796 bzero(z, sizeof(SLZone));
6ab8e1da 797 z->z_Flags |= SLZF_UNOTZEROD;
a108bf71
MD
798 } else {
799 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
800 if (z == NULL)
801 goto fail;
51295aee 802 atomic_add_int(&ZoneGenAlloc, ZoneSize / 1024);
a108bf71
MD
803 }
804
10cc6608
MD
805 /*
806 * How big is the base structure?
807 */
808#if defined(INVARIANTS)
809 /*
810 * Make room for z_Bitmap. An exact calculation is somewhat more
811 * complicated so don't make an exact calculation.
812 */
813 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
814 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
815#else
816 off = sizeof(SLZone);
817#endif
818
a108bf71
MD
819 /*
820 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
e0fe4b4e 821 * Otherwise properly align the data according to the chunk size.
a108bf71 822 */
298ff6e5 823 if (powerof2(size))
4aae2d75
SZ
824 align = size;
825 off = (off + align - 1) & ~(align - 1);
826
a108bf71
MD
827 z->z_Magic = ZALLOC_SLAB_MAGIC;
828 z->z_ZoneIndex = zi;
829 z->z_NMax = (ZoneSize - off) / size;
830 z->z_NFree = z->z_NMax - 1;
1c5ca4f3
MD
831 z->z_BasePtr = (char *)z + off;
832 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
a108bf71 833 z->z_ChunkSize = size;
2db3b277 834 z->z_CpuGd = gd;
bba6a44d 835 z->z_Cpu = gd->gd_cpuid;
5fee07e6 836 z->z_LChunksp = &z->z_LChunks;
bbb201fd
MD
837#ifdef SLAB_DEBUG
838 bcopy(z->z_Sources, z->z_AltSources, sizeof(z->z_Sources));
839 bzero(z->z_Sources, sizeof(z->z_Sources));
840#endif
1c5ca4f3 841 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
a108bf71
MD
842 z->z_Next = slgd->ZoneAry[zi];
843 slgd->ZoneAry[zi] = z;
8f1d5415 844 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 845 flags &= ~M_ZERO; /* already zero'd */
8f1d5415
MD
846 flags |= M_PASSIVE_ZERO;
847 }
5fee07e6 848 kup = btokup(z);
722871d3 849 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */
10cc6608 850 chunk_mark_allocated(z, chunk);
bbb201fd
MD
851#ifdef SLAB_DEBUG
852 slab_record_source(z, file, line);
853#endif
1c5ca4f3
MD
854
855 /*
856 * Slide the base index for initial allocations out of the next
857 * zone we create so we do not over-weight the lower part of the
858 * cpu memory caches.
859 */
860 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
861 & (ZALLOC_MAX_ZONE_SIZE - 1);
a108bf71 862 }
5fee07e6 863
a108bf71 864done:
bba6a44d
MD
865 ++type->ks_inuse[gd->gd_cpuid];
866 type->ks_memuse[gd->gd_cpuid] += size;
38e34349 867 type->ks_loosememuse += size; /* not MP synchronized */
a108bf71 868 crit_exit();
5fee07e6 869
a108bf71
MD
870 if (flags & M_ZERO)
871 bzero(chunk, size);
bba6a44d 872#ifdef INVARIANTS
d2182dc1
MD
873 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
874 if (use_malloc_pattern) {
875 for (i = 0; i < size; i += sizeof(int)) {
876 *(int *)((char *)chunk + i) = -1;
877 }
878 }
bba6a44d 879 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
d2182dc1 880 }
bba6a44d 881#endif
5fee07e6 882 logmemory(malloc_end, chunk, type, size, flags);
a108bf71
MD
883 return(chunk);
884fail:
885 crit_exit();
5fee07e6 886 logmemory(malloc_end, NULL, type, size, flags);
a108bf71
MD
887 return(NULL);
888}
889
38e34349
MD
890/*
891 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
892 *
893 * Generally speaking this routine is not called very often and we do
894 * not attempt to optimize it beyond reusing the same pointer if the
895 * new size fits within the chunking of the old pointer's zone.
896 */
bbb201fd
MD
897#ifdef SLAB_DEBUG
898void *
899krealloc_debug(void *ptr, unsigned long size,
900 struct malloc_type *type, int flags,
901 const char *file, int line)
902#else
a108bf71 903void *
8aca2bd4 904krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
bbb201fd 905#endif
a108bf71 906{
722871d3 907 unsigned long osize;
81dedbc2 908 unsigned long align;
a108bf71
MD
909 SLZone *z;
910 void *nptr;
722871d3 911 int *kup;
a108bf71 912
eb7f3e3c
MD
913 KKASSERT((flags & M_ZERO) == 0); /* not supported */
914
a108bf71 915 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
bbb201fd 916 return(kmalloc_debug(size, type, flags, file, line));
a108bf71 917 if (size == 0) {
efda3bd0 918 kfree(ptr, type);
a108bf71
MD
919 return(NULL);
920 }
921
922 /*
923 * Handle oversized allocations. XXX we really should require that a
924 * size be passed to free() instead of this nonsense.
925 */
5fee07e6 926 kup = btokup(ptr);
722871d3
MD
927 if (*kup > 0) {
928 osize = *kup << PAGE_SHIFT;
5fee07e6
MD
929 if (osize == round_page(size))
930 return(ptr);
bbb201fd 931 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
5fee07e6
MD
932 return(NULL);
933 bcopy(ptr, nptr, min(size, osize));
934 kfree(ptr, type);
935 return(nptr);
a108bf71
MD
936 }
937
938 /*
939 * Get the original allocation's zone. If the new request winds up
940 * using the same chunk size we do not have to do anything.
941 */
5fee07e6
MD
942 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
943 kup = btokup(z);
722871d3 944 KKASSERT(*kup < 0);
a108bf71
MD
945 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
946
a108bf71
MD
947 /*
948 * Allocate memory for the new request size. Note that zoneindex has
949 * already adjusted the request size to the appropriate chunk size, which
950 * should optimize our bcopy(). Then copy and return the new pointer.
1ea6580d
MD
951 *
952 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
953 * necessary align the result.
954 *
955 * We can only zoneindex (to align size to the chunk size) if the new
956 * size is not too large.
a108bf71 957 */
1ea6580d 958 if (size < ZoneLimit) {
81dedbc2 959 zoneindex(&size, &align);
1ea6580d
MD
960 if (z->z_ChunkSize == size)
961 return(ptr);
962 }
bbb201fd 963 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
a108bf71
MD
964 return(NULL);
965 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
efda3bd0 966 kfree(ptr, type);
a108bf71
MD
967 return(nptr);
968}
969
45d2b1d8
MD
970/*
971 * Return the kmalloc limit for this type, in bytes.
972 */
973long
974kmalloc_limit(struct malloc_type *type)
975{
976 if (type->ks_limit == 0) {
977 crit_enter();
978 if (type->ks_limit == 0)
979 malloc_init(type);
980 crit_exit();
981 }
982 return(type->ks_limit);
983}
984
38e34349
MD
985/*
986 * Allocate a copy of the specified string.
987 *
988 * (MP SAFE) (MAY BLOCK)
989 */
bbb201fd
MD
990#ifdef SLAB_DEBUG
991char *
992kstrdup_debug(const char *str, struct malloc_type *type,
993 const char *file, int line)
994#else
1ac06773 995char *
59302080 996kstrdup(const char *str, struct malloc_type *type)
bbb201fd 997#endif
1ac06773
MD
998{
999 int zlen; /* length inclusive of terminating NUL */
1000 char *nstr;
1001
1002 if (str == NULL)
1003 return(NULL);
1004 zlen = strlen(str) + 1;
bbb201fd 1005 nstr = kmalloc_debug(zlen, type, M_WAITOK, file, line);
1ac06773
MD
1006 bcopy(str, nstr, zlen);
1007 return(nstr);
1008}
1009
a108bf71 1010/*
5fee07e6 1011 * Notify our cpu that a remote cpu has freed some chunks in a zone that
df9daea8
MD
1012 * we own. RCount will be bumped so the memory should be good, but validate
1013 * that it really is.
a108bf71
MD
1014 */
1015static
1016void
5fee07e6 1017kfree_remote(void *ptr)
a108bf71 1018{
5fee07e6
MD
1019 SLGlobalData *slgd;
1020 SLChunk *bchunk;
1021 SLZone *z;
1022 int nfree;
722871d3 1023 int *kup;
5fee07e6 1024
5fee07e6
MD
1025 slgd = &mycpu->gd_slab;
1026 z = ptr;
1027 kup = btokup(z);
df9daea8
MD
1028 KKASSERT(*kup == -((int)mycpuid + 1));
1029 KKASSERT(z->z_RCount > 0);
1030 atomic_subtract_int(&z->z_RCount, 1);
5fee07e6 1031
5bf48697 1032 logmemory(free_rem_beg, z, NULL, 0L, 0);
df9daea8
MD
1033 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1034 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
1035 nfree = z->z_NFree;
5fee07e6 1036
df9daea8
MD
1037 /*
1038 * Indicate that we will no longer be off of the ZoneAry by
1039 * clearing RSignal.
1040 */
1041 if (z->z_RChunks)
1042 z->z_RSignal = 0;
5fee07e6 1043
df9daea8
MD
1044 /*
1045 * Atomically extract the bchunks list and then process it back
1046 * into the lchunks list. We want to append our bchunks to the
1047 * lchunks list and not prepend since we likely do not have
1048 * cache mastership of the related data (not that it helps since
1049 * we are using c_Next).
1050 */
1051 while ((bchunk = z->z_RChunks) != NULL) {
1052 cpu_ccfence();
1053 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
1054 *z->z_LChunksp = bchunk;
1055 while (bchunk) {
1056 chunk_mark_free(z, bchunk);
1057 z->z_LChunksp = &bchunk->c_Next;
1058 bchunk = bchunk->c_Next;
1059 ++z->z_NFree;
5fee07e6 1060 }
df9daea8 1061 break;
5fee07e6 1062 }
df9daea8
MD
1063 }
1064 if (z->z_NFree && nfree == 0) {
1065 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1066 slgd->ZoneAry[z->z_ZoneIndex] = z;
1067 }
5fee07e6 1068
df9daea8
MD
1069 /*
1070 * If the zone becomes totally free, and there are other zones we
1071 * can allocate from, move this zone to the FreeZones list. Since
1072 * this code can be called from an IPI callback, do *NOT* try to mess
1073 * with kernel_map here. Hysteresis will be performed at malloc() time.
1074 *
1075 * Do not move the zone if there is an IPI inflight, otherwise MP
1076 * races can result in our free_remote code accessing a destroyed
1077 * zone.
1078 */
1079 if (z->z_NFree == z->z_NMax &&
1080 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1081 z->z_RCount == 0
1082 ) {
1083 SLZone **pz;
1084 int *kup;
5fee07e6 1085
df9daea8
MD
1086 for (pz = &slgd->ZoneAry[z->z_ZoneIndex];
1087 z != *pz;
1088 pz = &(*pz)->z_Next) {
1089 ;
5fee07e6 1090 }
df9daea8
MD
1091 *pz = z->z_Next;
1092 z->z_Magic = -1;
1093 z->z_Next = slgd->FreeZones;
1094 slgd->FreeZones = z;
1095 ++slgd->NFreeZones;
1096 kup = btokup(z);
1097 *kup = 0;
5fee07e6 1098 }
5bf48697 1099 logmemory(free_rem_end, z, bchunk, 0L, 0);
a108bf71
MD
1100}
1101
38e34349 1102/*
5b287bba 1103 * free (SLAB ALLOCATOR)
38e34349
MD
1104 *
1105 * Free a memory block previously allocated by malloc. Note that we do not
5fee07e6 1106 * attempt to update ks_loosememuse as MP races could prevent us from
38e34349 1107 * checking memory limits in malloc.
5b287bba
MD
1108 *
1109 * MPSAFE
38e34349 1110 */
a108bf71 1111void
8aca2bd4 1112kfree(void *ptr, struct malloc_type *type)
a108bf71
MD
1113{
1114 SLZone *z;
1115 SLChunk *chunk;
1116 SLGlobalData *slgd;
bba6a44d 1117 struct globaldata *gd;
722871d3 1118 int *kup;
5fee07e6 1119 unsigned long size;
d8100bdc 1120 SLChunk *bchunk;
5fee07e6 1121 int rsignal;
a108bf71 1122
b68ad50c 1123 logmemory_quick(free_beg);
bba6a44d
MD
1124 gd = mycpu;
1125 slgd = &gd->gd_slab;
a108bf71 1126
d39911d9
JS
1127 if (ptr == NULL)
1128 panic("trying to free NULL pointer");
1129
a108bf71
MD
1130 /*
1131 * Handle special 0-byte allocations
1132 */
f2b5daf9 1133 if (ptr == ZERO_LENGTH_PTR) {
5bf48697 1134 logmemory(free_zero, ptr, type, -1UL, 0);
b68ad50c 1135 logmemory_quick(free_end);
a108bf71 1136 return;
f2b5daf9 1137 }
a108bf71 1138
5fee07e6
MD
1139 /*
1140 * Panic on bad malloc type
1141 */
1142 if (type->ks_magic != M_MAGIC)
1143 panic("free: malloc type lacks magic");
1144
a108bf71
MD
1145 /*
1146 * Handle oversized allocations. XXX we really should require that a
1147 * size be passed to free() instead of this nonsense.
bba6a44d
MD
1148 *
1149 * This code is never called via an ipi.
a108bf71 1150 */
5fee07e6 1151 kup = btokup(ptr);
722871d3
MD
1152 if (*kup > 0) {
1153 size = *kup << PAGE_SHIFT;
1154 *kup = 0;
a108bf71 1155#ifdef INVARIANTS
5fee07e6
MD
1156 KKASSERT(sizeof(weirdary) <= size);
1157 bcopy(weirdary, ptr, sizeof(weirdary));
a108bf71 1158#endif
5fee07e6
MD
1159 /*
1160 * NOTE: For oversized allocations we do not record the
1161 * originating cpu. It gets freed on the cpu calling
1162 * kfree(). The statistics are in aggregate.
1163 *
1164 * note: XXX we have still inherited the interrupts-can't-block
1165 * assumption. An interrupt thread does not bump
1166 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1167 * primarily until we can fix softupdate's assumptions about free().
1168 */
1169 crit_enter();
1170 --type->ks_inuse[gd->gd_cpuid];
1171 type->ks_memuse[gd->gd_cpuid] -= size;
1172 if (mycpu->gd_intr_nesting_level ||
1173 (gd->gd_curthread->td_flags & TDF_INTTHREAD))
1174 {
1175 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1176 z = (SLZone *)ptr;
1177 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1178 z->z_Next = slgd->FreeOvZones;
1179 z->z_ChunkSize = size;
1180 slgd->FreeOvZones = z;
1181 crit_exit();
1182 } else {
1183 crit_exit();
1184 logmemory(free_ovsz, ptr, type, size, 0);
1185 kmem_slab_free(ptr, size); /* may block */
1186 atomic_add_int(&ZoneBigAlloc, -(int)size / 1024);
a108bf71 1187 }
5fee07e6
MD
1188 logmemory_quick(free_end);
1189 return;
a108bf71
MD
1190 }
1191
1192 /*
1193 * Zone case. Figure out the zone based on the fact that it is
1194 * ZoneSize aligned.
1195 */
5fee07e6
MD
1196 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1197 kup = btokup(z);
722871d3 1198 KKASSERT(*kup < 0);
a108bf71
MD
1199 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1200
1201 /*
5fee07e6
MD
1202 * If we do not own the zone then use atomic ops to free to the
1203 * remote cpu linked list and notify the target zone using a
1204 * passive message.
1205 *
1206 * The target zone cannot be deallocated while we own a chunk of it,
1207 * so the zone header's storage is stable until the very moment
1208 * we adjust z_RChunks. After that we cannot safely dereference (z).
1209 *
1210 * (no critical section needed)
a108bf71 1211 */
2db3b277 1212 if (z->z_CpuGd != gd) {
5fee07e6
MD
1213 /*
1214 * Making these adjustments now allow us to avoid passing (type)
1215 * to the remote cpu. Note that ks_inuse/ks_memuse is being
28135cc2
MD
1216 * adjusted on OUR cpu, not the zone cpu, but it should all still
1217 * sum up properly and cancel out.
5fee07e6 1218 */
28135cc2
MD
1219 crit_enter();
1220 --type->ks_inuse[gd->gd_cpuid];
1221 type->ks_memuse[gd->gd_cpuid] -= z->z_ChunkSize;
1222 crit_exit();
5fee07e6
MD
1223
1224 /*
1225 * WARNING! This code competes with other cpus. Once we
1226 * successfully link the chunk to RChunks the remote
1227 * cpu can rip z's storage out from under us.
df9daea8
MD
1228 *
1229 * Bumping RCount prevents z's storage from getting
1230 * ripped out.
5fee07e6
MD
1231 */
1232 rsignal = z->z_RSignal;
1233 cpu_lfence();
df9daea8
MD
1234 if (rsignal)
1235 atomic_add_int(&z->z_RCount, 1);
5fee07e6
MD
1236
1237 chunk = ptr;
1238 for (;;) {
1239 bchunk = z->z_RChunks;
1240 cpu_ccfence();
1241 chunk->c_Next = bchunk;
1242 cpu_sfence();
1243
1244 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1245 break;
1246 }
5fee07e6
MD
1247
1248 /*
1249 * We have to signal the remote cpu if our actions will cause
1250 * the remote zone to be placed back on ZoneAry so it can
1251 * move the zone back on.
1252 *
1253 * We only need to deal with NULL->non-NULL RChunk transitions
1254 * and only if z_RSignal is set. We interlock by reading rsignal
1255 * before adding our chunk to RChunks. This should result in
1256 * virtually no IPI traffic.
1257 *
1258 * We can use a passive IPI to reduce overhead even further.
1259 */
1260 if (bchunk == NULL && rsignal) {
5bf48697 1261 logmemory(free_request, ptr, type, (unsigned long)z->z_ChunkSize, 0);
5fee07e6 1262 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
df9daea8
MD
1263 /* z can get ripped out from under us from this point on */
1264 } else if (rsignal) {
1265 atomic_subtract_int(&z->z_RCount, 1);
1266 /* z can get ripped out from under us from this point on */
5fee07e6 1267 }
b68ad50c 1268 logmemory_quick(free_end);
a108bf71
MD
1269 return;
1270 }
1271
5fee07e6
MD
1272 /*
1273 * kfree locally
1274 */
5bf48697 1275 logmemory(free_chunk, ptr, type, (unsigned long)z->z_ChunkSize, 0);
f2b5daf9 1276
a108bf71 1277 crit_enter();
a108bf71 1278 chunk = ptr;
10cc6608 1279 chunk_mark_free(z, chunk);
a108bf71
MD
1280
1281 /*
1282 * Put weird data into the memory to detect modifications after freeing,
1283 * illegal pointer use after freeing (we should fault on the odd address),
1284 * and so forth. XXX needs more work, see the old malloc code.
1285 */
1286#ifdef INVARIANTS
1287 if (z->z_ChunkSize < sizeof(weirdary))
1288 bcopy(weirdary, chunk, z->z_ChunkSize);
1289 else
1290 bcopy(weirdary, chunk, sizeof(weirdary));
1291#endif
1292
1293 /*
5fee07e6
MD
1294 * Add this free non-zero'd chunk to a linked list for reuse. Add
1295 * to the front of the linked list so it is more likely to be
1296 * reallocated, since it is already in our L1 cache.
a108bf71 1297 */
6ab8e1da 1298#ifdef INVARIANTS
c439ad8f 1299 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
fc92d4aa 1300 panic("BADFREE %p", chunk);
a108bf71 1301#endif
5fee07e6
MD
1302 chunk->c_Next = z->z_LChunks;
1303 z->z_LChunks = chunk;
1304 if (chunk->c_Next == NULL)
1305 z->z_LChunksp = &chunk->c_Next;
1306
6ab8e1da 1307#ifdef INVARIANTS
c439ad8f 1308 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
a108bf71 1309 panic("BADFREE2");
6ab8e1da 1310#endif
a108bf71
MD
1311
1312 /*
1313 * Bump the number of free chunks. If it becomes non-zero the zone
1314 * must be added back onto the appropriate list.
1315 */
1316 if (z->z_NFree++ == 0) {
1317 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1318 slgd->ZoneAry[z->z_ZoneIndex] = z;
1319 }
1320
bba6a44d
MD
1321 --type->ks_inuse[z->z_Cpu];
1322 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
a108bf71
MD
1323
1324 /*
1325 * If the zone becomes totally free, and there are other zones we
a7cf0021
MD
1326 * can allocate from, move this zone to the FreeZones list. Since
1327 * this code can be called from an IPI callback, do *NOT* try to mess
1328 * with kernel_map here. Hysteresis will be performed at malloc() time.
a108bf71
MD
1329 */
1330 if (z->z_NFree == z->z_NMax &&
df9daea8
MD
1331 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1332 z->z_RCount == 0
a108bf71
MD
1333 ) {
1334 SLZone **pz;
722871d3 1335 int *kup;
a108bf71
MD
1336
1337 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
1338 ;
1339 *pz = z->z_Next;
1340 z->z_Magic = -1;
a7cf0021
MD
1341 z->z_Next = slgd->FreeZones;
1342 slgd->FreeZones = z;
1343 ++slgd->NFreeZones;
5fee07e6 1344 kup = btokup(z);
722871d3 1345 *kup = 0;
a108bf71 1346 }
b68ad50c 1347 logmemory_quick(free_end);
a108bf71
MD
1348 crit_exit();
1349}
1350
10cc6608 1351#if defined(INVARIANTS)
5fee07e6 1352
10cc6608
MD
1353/*
1354 * Helper routines for sanity checks
1355 */
1356static
1357void
1358chunk_mark_allocated(SLZone *z, void *chunk)
1359{
1360 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1361 __uint32_t *bitptr;
1362
5fee07e6
MD
1363 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1364 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1365 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
10cc6608
MD
1366 bitptr = &z->z_Bitmap[bitdex >> 5];
1367 bitdex &= 31;
5fee07e6
MD
1368 KASSERT((*bitptr & (1 << bitdex)) == 0,
1369 ("memory chunk %p is already allocated!", chunk));
10cc6608
MD
1370 *bitptr |= 1 << bitdex;
1371}
1372
1373static
1374void
1375chunk_mark_free(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 free!", chunk));
10cc6608
MD
1387 *bitptr &= ~(1 << bitdex);
1388}
1389
1390#endif
1391
a108bf71 1392/*
5b287bba 1393 * kmem_slab_alloc()
a108bf71
MD
1394 *
1395 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1396 * specified alignment. M_* flags are expected in the flags field.
1397 *
1398 * Alignment must be a multiple of PAGE_SIZE.
1399 *
1400 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1401 * but when we move zalloc() over to use this function as its backend
1402 * we will have to switch to kreserve/krelease and call reserve(0)
1403 * after the new space is made available.
dc1fd4b3
MD
1404 *
1405 * Interrupt code which has preempted other code is not allowed to
c397c465
MD
1406 * use PQ_CACHE pages. However, if an interrupt thread is run
1407 * non-preemptively or blocks and then runs non-preemptively, then
b12defdc 1408 * it is free to use PQ_CACHE pages. <--- may not apply any longer XXX
a108bf71
MD
1409 */
1410static void *
1411kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1412{
1413 vm_size_t i;
1414 vm_offset_t addr;
1de1e800 1415 int count, vmflags, base_vmflags;
b12defdc
MD
1416 vm_page_t mbase = NULL;
1417 vm_page_t m;
dc1fd4b3 1418 thread_t td;
a108bf71
MD
1419
1420 size = round_page(size);
e4846942 1421 addr = vm_map_min(&kernel_map);
a108bf71 1422
a108bf71
MD
1423 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1424 crit_enter();
e4846942 1425 vm_map_lock(&kernel_map);
c809941b 1426 if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) {
e4846942 1427 vm_map_unlock(&kernel_map);
8cb2bf45 1428 if ((flags & M_NULLOK) == 0)
a108bf71 1429 panic("kmem_slab_alloc(): kernel_map ran out of space!");
a108bf71 1430 vm_map_entry_release(count);
2de4f77e 1431 crit_exit();
a108bf71
MD
1432 return(NULL);
1433 }
e4846942
MD
1434
1435 /*
1436 * kernel_object maps 1:1 to kernel_map.
1437 */
b12defdc
MD
1438 vm_object_hold(&kernel_object);
1439 vm_object_reference_locked(&kernel_object);
e4846942
MD
1440 vm_map_insert(&kernel_map, &count,
1441 &kernel_object, addr, addr, addr + size,
1b874851
MD
1442 VM_MAPTYPE_NORMAL,
1443 VM_PROT_ALL, VM_PROT_ALL,
1444 0);
b12defdc
MD
1445 vm_object_drop(&kernel_object);
1446 vm_map_set_wired_quick(&kernel_map, addr, size, &count);
1447 vm_map_unlock(&kernel_map);
a108bf71 1448
dc1fd4b3 1449 td = curthread;
dc1fd4b3 1450
1de1e800
JS
1451 base_vmflags = 0;
1452 if (flags & M_ZERO)
1453 base_vmflags |= VM_ALLOC_ZERO;
1454 if (flags & M_USE_RESERVE)
1455 base_vmflags |= VM_ALLOC_SYSTEM;
1456 if (flags & M_USE_INTERRUPT_RESERVE)
1457 base_vmflags |= VM_ALLOC_INTERRUPT;
77912481
MD
1458 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1459 panic("kmem_slab_alloc: bad flags %08x (%p)",
1460 flags, ((int **)&size)[-1]);
1461 }
1de1e800 1462
a108bf71 1463 /*
b12defdc
MD
1464 * Allocate the pages. Do not mess with the PG_ZERO flag or map
1465 * them yet. VM_ALLOC_NORMAL can only be set if we are not preempting.
1466 *
1467 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1468 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1469 * implied in this case), though I'm not sure if we really need to
1470 * do that.
a108bf71 1471 */
b12defdc
MD
1472 vmflags = base_vmflags;
1473 if (flags & M_WAITOK) {
1474 if (td->td_preempted)
1475 vmflags |= VM_ALLOC_SYSTEM;
1476 else
1477 vmflags |= VM_ALLOC_NORMAL;
1478 }
a108bf71 1479
b12defdc
MD
1480 vm_object_hold(&kernel_object);
1481 for (i = 0; i < size; i += PAGE_SIZE) {
e4846942 1482 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
b12defdc
MD
1483 if (i == 0)
1484 mbase = m;
dc1fd4b3
MD
1485
1486 /*
1487 * If the allocation failed we either return NULL or we retry.
1488 *
c397c465
MD
1489 * If M_WAITOK is specified we wait for more memory and retry.
1490 * If M_WAITOK is specified from a preemption we yield instead of
1491 * wait. Livelock will not occur because the interrupt thread
1492 * will not be preempting anyone the second time around after the
1493 * yield.
dc1fd4b3 1494 */
a108bf71 1495 if (m == NULL) {
c397c465 1496 if (flags & M_WAITOK) {
fe1e98d0 1497 if (td->td_preempted) {
77912481 1498 lwkt_switch();
dc1fd4b3 1499 } else {
4ecf7cc9 1500 vm_wait(0);
dc1fd4b3 1501 }
a108bf71
MD
1502 i -= PAGE_SIZE; /* retry */
1503 continue;
1504 }
b12defdc
MD
1505 break;
1506 }
1507 }
dc1fd4b3 1508
b12defdc
MD
1509 /*
1510 * Check and deal with an allocation failure
1511 */
1512 if (i != size) {
1513 while (i != 0) {
1514 i -= PAGE_SIZE;
1515 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1516 /* page should already be busy */
1517 vm_page_free(m);
a108bf71 1518 }
b12defdc
MD
1519 vm_map_lock(&kernel_map);
1520 vm_map_delete(&kernel_map, addr, addr + size, &count);
1521 vm_map_unlock(&kernel_map);
1522 vm_object_drop(&kernel_object);
1523
1524 vm_map_entry_release(count);
1525 crit_exit();
1526 return(NULL);
a108bf71
MD
1527 }
1528
1529 /*
dc1fd4b3
MD
1530 * Success!
1531 *
b12defdc
MD
1532 * NOTE: The VM pages are still busied. mbase points to the first one
1533 * but we have to iterate via vm_page_next()
a108bf71 1534 */
b12defdc 1535 vm_object_drop(&kernel_object);
a108bf71
MD
1536 crit_exit();
1537
1538 /*
1539 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1540 */
b12defdc
MD
1541 m = mbase;
1542 i = 0;
a108bf71 1543
b12defdc
MD
1544 while (i < size) {
1545 /*
1546 * page should already be busy
1547 */
a108bf71
MD
1548 m->valid = VM_PAGE_BITS_ALL;
1549 vm_page_wire(m);
921c891e
MD
1550 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL | VM_PROT_NOSYNC,
1551 1, NULL);
a108bf71
MD
1552 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1553 bzero((char *)addr + i, PAGE_SIZE);
1554 vm_page_flag_clear(m, PG_ZERO);
17cde63e
MD
1555 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1556 vm_page_flag_set(m, PG_REFERENCED);
a491077e 1557 vm_page_wakeup(m);
b12defdc
MD
1558
1559 i += PAGE_SIZE;
1560 vm_object_hold(&kernel_object);
1561 m = vm_page_next(m);
1562 vm_object_drop(&kernel_object);
a108bf71 1563 }
b12defdc 1564 smp_invltlb();
a108bf71
MD
1565 vm_map_entry_release(count);
1566 return((void *)addr);
1567}
1568
38e34349 1569/*
5b287bba 1570 * kmem_slab_free()
38e34349 1571 */
a108bf71
MD
1572static void
1573kmem_slab_free(void *ptr, vm_size_t size)
1574{
1575 crit_enter();
e4846942 1576 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
a108bf71
MD
1577 crit_exit();
1578}
1579
55126ffe
SZ
1580void *
1581kmalloc_cachealign(unsigned long size_alloc, struct malloc_type *type,
1582 int flags)
1583{
3b556fa0
SZ
1584#if (__VM_CACHELINE_SIZE == 32)
1585#define CAN_CACHEALIGN(sz) ((sz) >= 256)
1586#elif (__VM_CACHELINE_SIZE == 64)
1587#define CAN_CACHEALIGN(sz) ((sz) >= 512)
1588#elif (__VM_CACHELINE_SIZE == 128)
1589#define CAN_CACHEALIGN(sz) ((sz) >= 1024)
1590#else
1591#error "unsupported cacheline size"
1592#endif
1593
1594 void *ret;
1595
55126ffe
SZ
1596 if (size_alloc < __VM_CACHELINE_SIZE)
1597 size_alloc = __VM_CACHELINE_SIZE;
3b556fa0
SZ
1598 else if (!CAN_CACHEALIGN(size_alloc))
1599 flags |= M_POWEROF2;
1600
1601 ret = kmalloc(size_alloc, type, flags);
1602 KASSERT(((uintptr_t)ret & (__VM_CACHELINE_SIZE - 1)) == 0,
1603 ("%p(%lu) not cacheline %d aligned",
1604 ret, size_alloc, __VM_CACHELINE_SIZE));
1605 return ret;
1606
1607#undef CAN_CACHEALIGN
55126ffe 1608}