kmalloc: Use 'fls' to round up the size to the nearest power of 2
[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
MD
80 *
81 * API REQUIREMENTS AND SIDE EFFECTS
82 *
83 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we
84 * have remained compatible with the following API requirements:
85 *
86 * + small power-of-2 sized allocations are power-of-2 aligned (kern_tty)
3d177b31 87 * + all power-of-2 sized allocations are power-of-2 aligned (twe)
a108bf71
MD
88 * + malloc(0) is allowed and returns non-NULL (ahc driver)
89 * + ability to allocate arbitrarily large chunks of memory
90 */
91
92#include "opt_vm.h"
93
a108bf71
MD
94#include <sys/param.h>
95#include <sys/systm.h>
96#include <sys/kernel.h>
97#include <sys/slaballoc.h>
98#include <sys/mbuf.h>
99#include <sys/vmmeter.h>
100#include <sys/lock.h>
101#include <sys/thread.h>
102#include <sys/globaldata.h>
d2182dc1 103#include <sys/sysctl.h>
f2b5daf9 104#include <sys/ktr.h>
a108bf71
MD
105
106#include <vm/vm.h>
107#include <vm/vm_param.h>
108#include <vm/vm_kern.h>
109#include <vm/vm_extern.h>
110#include <vm/vm_object.h>
111#include <vm/pmap.h>
112#include <vm/vm_map.h>
113#include <vm/vm_page.h>
114#include <vm/vm_pageout.h>
115
116#include <machine/cpu.h>
117
118#include <sys/thread2.h>
119
722871d3
MD
120#define btokup(z) (&pmap_kvtom((vm_offset_t)(z))->ku_pagecnt)
121
5bf48697
AE
122#define MEMORY_STRING "ptr=%p type=%p size=%lu flags=%04x"
123#define MEMORY_ARGS void *ptr, void *type, unsigned long size, int flags
f2b5daf9
MD
124
125#if !defined(KTR_MEMORY)
126#define KTR_MEMORY KTR_ALL
127#endif
128KTR_INFO_MASTER(memory);
5bf48697
AE
129KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin");
130KTR_INFO(KTR_MEMORY, memory, malloc_end, 1, MEMORY_STRING, MEMORY_ARGS);
131KTR_INFO(KTR_MEMORY, memory, free_zero, 2, MEMORY_STRING, MEMORY_ARGS);
132KTR_INFO(KTR_MEMORY, memory, free_ovsz, 3, MEMORY_STRING, MEMORY_ARGS);
133KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 4, MEMORY_STRING, MEMORY_ARGS);
134KTR_INFO(KTR_MEMORY, memory, free_chunk, 5, MEMORY_STRING, MEMORY_ARGS);
f2b5daf9 135#ifdef SMP
5bf48697
AE
136KTR_INFO(KTR_MEMORY, memory, free_request, 6, MEMORY_STRING, MEMORY_ARGS);
137KTR_INFO(KTR_MEMORY, memory, free_rem_beg, 7, MEMORY_STRING, MEMORY_ARGS);
138KTR_INFO(KTR_MEMORY, memory, free_rem_end, 8, MEMORY_STRING, MEMORY_ARGS);
f2b5daf9 139#endif
5bf48697
AE
140KTR_INFO(KTR_MEMORY, memory, free_beg, 9, "free begin");
141KTR_INFO(KTR_MEMORY, memory, free_end, 10, "free end");
f2b5daf9
MD
142
143#define logmemory(name, ptr, type, size, flags) \
144 KTR_LOG(memory_ ## name, ptr, type, size, flags)
b68ad50c
MD
145#define logmemory_quick(name) \
146 KTR_LOG(memory_ ## name)
f2b5daf9 147
a108bf71
MD
148/*
149 * Fixed globals (not per-cpu)
150 */
151static int ZoneSize;
46a3f46d 152static int ZoneLimit;
a108bf71 153static int ZonePageCount;
5fee07e6 154static uintptr_t ZoneMask;
665206ee
MD
155static int ZoneBigAlloc; /* in KB */
156static int ZoneGenAlloc; /* in KB */
460426e6 157struct malloc_type *kmemstatistics; /* exported to vmstat */
a108bf71
MD
158static int32_t weirdary[16];
159
160static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
161static void kmem_slab_free(void *ptr, vm_size_t bytes);
5fee07e6 162
10cc6608
MD
163#if defined(INVARIANTS)
164static void chunk_mark_allocated(SLZone *z, void *chunk);
165static void chunk_mark_free(SLZone *z, void *chunk);
5fee07e6
MD
166#else
167#define chunk_mark_allocated(z, chunk)
168#define chunk_mark_free(z, chunk)
10cc6608 169#endif
a108bf71
MD
170
171/*
172 * Misc constants. Note that allocations that are exact multiples of
173 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
174 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
175 */
176#define MIN_CHUNK_SIZE 8 /* in bytes */
177#define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
ad94a851 178#define ZONE_RELS_THRESH 32 /* threshold number of zones */
a108bf71
MD
179#define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
180
181/*
182 * The WEIRD_ADDR is used as known text to copy into free objects to
183 * try to create deterministic failure cases if the data is accessed after
184 * free.
185 */
186#define WEIRD_ADDR 0xdeadc0de
187#define MAX_COPY sizeof(weirdary)
188#define ZERO_LENGTH_PTR ((void *)-8)
189
190/*
191 * Misc global malloc buckets
192 */
193
194MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
195MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
196MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
197
198MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
199MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
200
201/*
202 * Initialize the slab memory allocator. We have to choose a zone size based
203 * on available physical memory. We choose a zone side which is approximately
204 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
205 * 128K. The zone size is limited to the bounds set in slaballoc.h
206 * (typically 32K min, 128K max).
207 */
208static void kmeminit(void *dummy);
209
c7841cbe
MD
210char *ZeroPage;
211
ba39e2e0 212SYSINIT(kmem, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL)
a108bf71 213
d2182dc1
MD
214#ifdef INVARIANTS
215/*
216 * If enabled any memory allocated without M_ZERO is initialized to -1.
217 */
218static int use_malloc_pattern;
219SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
0c52fa62
SG
220 &use_malloc_pattern, 0,
221 "Initialize memory to -1 if M_ZERO not specified");
d2182dc1
MD
222#endif
223
ad94a851 224static int ZoneRelsThresh = ZONE_RELS_THRESH;
665206ee
MD
225SYSCTL_INT(_kern, OID_AUTO, zone_big_alloc, CTLFLAG_RD, &ZoneBigAlloc, 0, "");
226SYSCTL_INT(_kern, OID_AUTO, zone_gen_alloc, CTLFLAG_RD, &ZoneGenAlloc, 0, "");
ad94a851 227SYSCTL_INT(_kern, OID_AUTO, zone_cache, CTLFLAG_RW, &ZoneRelsThresh, 0, "");
665206ee 228
b12defdc
MD
229/*
230 * Returns the kernel memory size limit for the purposes of initializing
231 * various subsystem caches. The smaller of available memory and the KVM
232 * memory space is returned.
233 *
234 * The size in megabytes is returned.
235 */
236size_t
237kmem_lim_size(void)
238{
239 size_t limsize;
240
241 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
242 if (limsize > KvaSize)
243 limsize = KvaSize;
244 return (limsize / (1024 * 1024));
245}
246
a108bf71
MD
247static void
248kmeminit(void *dummy)
249{
7c457ac8 250 size_t limsize;
a108bf71
MD
251 int usesize;
252 int i;
a108bf71 253
b12defdc
MD
254 limsize = kmem_lim_size();
255 usesize = (int)(limsize * 1024); /* convert to KB */
a108bf71 256
b12defdc
MD
257 /*
258 * If the machine has a large KVM space and more than 8G of ram,
259 * double the zone release threshold to reduce SMP invalidations.
260 * If more than 16G of ram, do it again.
261 *
262 * The BIOS eats a little ram so add some slop. We want 8G worth of
263 * memory sticks to trigger the first adjustment.
264 */
265 if (ZoneRelsThresh == ZONE_RELS_THRESH) {
266 if (limsize >= 7 * 1024)
267 ZoneRelsThresh *= 2;
268 if (limsize >= 15 * 1024)
269 ZoneRelsThresh *= 2;
270 }
a108bf71 271
b12defdc
MD
272 /*
273 * Calculate the zone size. This typically calculates to
274 * ZALLOC_MAX_ZONE_SIZE
275 */
a108bf71
MD
276 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
277 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
278 ZoneSize <<= 1;
46a3f46d
MD
279 ZoneLimit = ZoneSize / 4;
280 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
281 ZoneLimit = ZALLOC_ZONE_LIMIT;
5fee07e6 282 ZoneMask = ~(uintptr_t)(ZoneSize - 1);
a108bf71
MD
283 ZonePageCount = ZoneSize / PAGE_SIZE;
284
a3034532 285 for (i = 0; i < NELEM(weirdary); ++i)
a108bf71
MD
286 weirdary[i] = WEIRD_ADDR;
287
c7841cbe
MD
288 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
289
a108bf71 290 if (bootverbose)
6ea70f76 291 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
a108bf71
MD
292}
293
294/*
bba6a44d 295 * Initialize a malloc type tracking structure.
a108bf71
MD
296 */
297void
298malloc_init(void *data)
299{
300 struct malloc_type *type = data;
7c457ac8 301 size_t limsize;
a108bf71
MD
302
303 if (type->ks_magic != M_MAGIC)
304 panic("malloc type lacks magic");
305
306 if (type->ks_limit != 0)
307 return;
308
309 if (vmstats.v_page_count == 0)
310 panic("malloc_init not allowed before vm init");
311
b12defdc 312 limsize = kmem_lim_size() * (1024 * 1024);
a108bf71
MD
313 type->ks_limit = limsize / 10;
314
315 type->ks_next = kmemstatistics;
316 kmemstatistics = type;
317}
318
319void
320malloc_uninit(void *data)
321{
322 struct malloc_type *type = data;
323 struct malloc_type *t;
bba6a44d
MD
324#ifdef INVARIANTS
325 int i;
1d712609 326 long ttl;
bba6a44d 327#endif
a108bf71
MD
328
329 if (type->ks_magic != M_MAGIC)
330 panic("malloc type lacks magic");
331
332 if (vmstats.v_page_count == 0)
333 panic("malloc_uninit not allowed before vm init");
334
335 if (type->ks_limit == 0)
336 panic("malloc_uninit on uninitialized type");
337
6c92c1f2
SZ
338#ifdef SMP
339 /* Make sure that all pending kfree()s are finished. */
340 lwkt_synchronize_ipiqs("muninit");
341#endif
342
a108bf71 343#ifdef INVARIANTS
1d712609
MD
344 /*
345 * memuse is only correct in aggregation. Due to memory being allocated
346 * on one cpu and freed on another individual array entries may be
347 * negative or positive (canceling each other out).
348 */
349 for (i = ttl = 0; i < ncpus; ++i)
350 ttl += type->ks_memuse[i];
351 if (ttl) {
6ea70f76 352 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
1d712609 353 ttl, type->ks_shortdesc, i);
a108bf71
MD
354 }
355#endif
356 if (type == kmemstatistics) {
357 kmemstatistics = type->ks_next;
358 } else {
359 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
360 if (t->ks_next == type) {
361 t->ks_next = type->ks_next;
362 break;
363 }
364 }
365 }
366 type->ks_next = NULL;
367 type->ks_limit = 0;
368}
369
370/*
40153c65
MD
371 * Increase the kmalloc pool limit for the specified pool. No changes
372 * are the made if the pool would shrink.
373 */
374void
375kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
376{
377 if (type->ks_limit == 0)
378 malloc_init(type);
7c457ac8
MD
379 if (bytes == 0)
380 bytes = KvaSize;
40153c65
MD
381 if (type->ks_limit < bytes)
382 type->ks_limit = bytes;
383}
384
385/*
ebe36cfe
MD
386 * Dynamically create a malloc pool. This function is a NOP if *typep is
387 * already non-NULL.
388 */
389void
390kmalloc_create(struct malloc_type **typep, const char *descr)
391{
392 struct malloc_type *type;
393
394 if (*typep == NULL) {
395 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
396 type->ks_magic = M_MAGIC;
397 type->ks_shortdesc = descr;
398 malloc_init(type);
399 *typep = type;
400 }
401}
402
403/*
404 * Destroy a dynamically created malloc pool. This function is a NOP if
405 * the pool has already been destroyed.
406 */
407void
408kmalloc_destroy(struct malloc_type **typep)
409{
410 if (*typep != NULL) {
411 malloc_uninit(*typep);
412 kfree(*typep, M_TEMP);
413 *typep = NULL;
414 }
415}
416
417/*
a108bf71
MD
418 * Calculate the zone index for the allocation request size and set the
419 * allocation request size to that particular zone's chunk size.
420 */
421static __inline int
422zoneindex(unsigned long *bytes)
423{
424 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
425 if (n < 128) {
426 *bytes = n = (n + 7) & ~7;
427 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
428 }
429 if (n < 256) {
430 *bytes = n = (n + 15) & ~15;
431 return(n / 16 + 7);
432 }
433 if (n < 8192) {
434 if (n < 512) {
435 *bytes = n = (n + 31) & ~31;
436 return(n / 32 + 15);
437 }
438 if (n < 1024) {
439 *bytes = n = (n + 63) & ~63;
440 return(n / 64 + 23);
441 }
442 if (n < 2048) {
443 *bytes = n = (n + 127) & ~127;
444 return(n / 128 + 31);
445 }
446 if (n < 4096) {
447 *bytes = n = (n + 255) & ~255;
448 return(n / 256 + 39);
449 }
450 *bytes = n = (n + 511) & ~511;
451 return(n / 512 + 47);
452 }
453#if ZALLOC_ZONE_LIMIT > 8192
454 if (n < 16384) {
455 *bytes = n = (n + 1023) & ~1023;
456 return(n / 1024 + 55);
457 }
458#endif
459#if ZALLOC_ZONE_LIMIT > 16384
460 if (n < 32768) {
461 *bytes = n = (n + 2047) & ~2047;
462 return(n / 2048 + 63);
463 }
464#endif
465 panic("Unexpected byte count %d", n);
466 return(0);
467}
468
bbb201fd
MD
469#ifdef SLAB_DEBUG
470/*
471 * Used to debug memory corruption issues. Record up to (typically 32)
472 * allocation sources for this zone (for a particular chunk size).
473 */
474
475static void
476slab_record_source(SLZone *z, const char *file, int line)
477{
478 int i;
479 int b = line & (SLAB_DEBUG_ENTRIES - 1);
480
481 i = b;
482 do {
483 if (z->z_Sources[i].file == file && z->z_Sources[i].line == line)
484 return;
485 if (z->z_Sources[i].file == NULL)
486 break;
487 i = (i + 1) & (SLAB_DEBUG_ENTRIES - 1);
488 } while (i != b);
489 z->z_Sources[i].file = file;
490 z->z_Sources[i].line = line;
491}
492
493#endif
494
1e57f867
SZ
495static __inline unsigned long
496powerof2_size(unsigned long size)
497{
498 int i, wt;
499
500 if (size == 0)
501 return 0;
502
503 i = flsl(size);
504 wt = (size & ~(1 << (i - 1)));
505 if (!wt)
506 --i;
507
508 return (1UL << i);
509}
510
a108bf71 511/*
5fee07e6 512 * kmalloc() (SLAB ALLOCATOR)
a108bf71
MD
513 *
514 * Allocate memory via the slab allocator. If the request is too large,
515 * or if it page-aligned beyond a certain size, we fall back to the
516 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
517 * &SlabMisc if you don't care.
518 *
8cb2bf45
JS
519 * M_RNOWAIT - don't block.
520 * M_NULLOK - return NULL instead of blocking.
a108bf71 521 * M_ZERO - zero the returned memory.
dc1fd4b3
MD
522 * M_USE_RESERVE - allow greater drawdown of the free list
523 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
1e57f867 524 * M_POWEROF2 - roundup size to the nearest power of 2
5b287bba
MD
525 *
526 * MPSAFE
a108bf71 527 */
bbb201fd
MD
528
529#ifdef SLAB_DEBUG
530void *
531kmalloc_debug(unsigned long size, struct malloc_type *type, int flags,
532 const char *file, int line)
533#else
a108bf71 534void *
8aca2bd4 535kmalloc(unsigned long size, struct malloc_type *type, int flags)
bbb201fd 536#endif
a108bf71
MD
537{
538 SLZone *z;
539 SLChunk *chunk;
d8100bdc 540#ifdef SMP
5fee07e6 541 SLChunk *bchunk;
d8100bdc 542#endif
a108bf71 543 SLGlobalData *slgd;
bba6a44d 544 struct globaldata *gd;
a108bf71 545 int zi;
d2182dc1
MD
546#ifdef INVARIANTS
547 int i;
548#endif
a108bf71 549
b68ad50c 550 logmemory_quick(malloc_beg);
bba6a44d
MD
551 gd = mycpu;
552 slgd = &gd->gd_slab;
a108bf71
MD
553
554 /*
555 * XXX silly to have this in the critical path.
556 */
557 if (type->ks_limit == 0) {
558 crit_enter();
559 if (type->ks_limit == 0)
560 malloc_init(type);
561 crit_exit();
562 }
563 ++type->ks_calls;
564
1e57f867
SZ
565 if (flags & M_POWEROF2)
566 size = powerof2_size(size);
567
a108bf71 568 /*
38e34349
MD
569 * Handle the case where the limit is reached. Panic if we can't return
570 * NULL. The original malloc code looped, but this tended to
a108bf71 571 * simply deadlock the computer.
38e34349
MD
572 *
573 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
574 * to determine if a more complete limit check should be done. The
575 * actual memory use is tracked via ks_memuse[cpu].
a108bf71 576 */
bba6a44d
MD
577 while (type->ks_loosememuse >= type->ks_limit) {
578 int i;
579 long ttl;
580
581 for (i = ttl = 0; i < ncpus; ++i)
582 ttl += type->ks_memuse[i];
38e34349 583 type->ks_loosememuse = ttl; /* not MP synchronized */
28135cc2
MD
584 if ((ssize_t)ttl < 0) /* deal with occassional race */
585 ttl = 0;
bba6a44d 586 if (ttl >= type->ks_limit) {
f2b5daf9 587 if (flags & M_NULLOK) {
5fee07e6 588 logmemory(malloc_end, NULL, type, size, flags);
bba6a44d 589 return(NULL);
f2b5daf9 590 }
bba6a44d
MD
591 panic("%s: malloc limit exceeded", type->ks_shortdesc);
592 }
a108bf71
MD
593 }
594
595 /*
596 * Handle the degenerate size == 0 case. Yes, this does happen.
597 * Return a special pointer. This is to maintain compatibility with
598 * the original malloc implementation. Certain devices, such as the
599 * adaptec driver, not only allocate 0 bytes, they check for NULL and
600 * also realloc() later on. Joy.
601 */
f2b5daf9 602 if (size == 0) {
5fee07e6 603 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
a108bf71 604 return(ZERO_LENGTH_PTR);
f2b5daf9 605 }
a108bf71
MD
606
607 /*
a7cf0021
MD
608 * Handle hysteresis from prior frees here in malloc(). We cannot
609 * safely manipulate the kernel_map in free() due to free() possibly
610 * being called via an IPI message or from sensitive interrupt code.
5fee07e6
MD
611 *
612 * NOTE: ku_pagecnt must be cleared before we free the slab or we
613 * might race another cpu allocating the kva and setting
614 * ku_pagecnt.
a7cf0021 615 */
ad94a851 616 while (slgd->NFreeZones > ZoneRelsThresh && (flags & M_RNOWAIT) == 0) {
46a3f46d 617 crit_enter();
ad94a851 618 if (slgd->NFreeZones > ZoneRelsThresh) { /* crit sect race */
722871d3 619 int *kup;
5fee07e6 620
46a3f46d
MD
621 z = slgd->FreeZones;
622 slgd->FreeZones = z->z_Next;
623 --slgd->NFreeZones;
5fee07e6 624 kup = btokup(z);
722871d3 625 *kup = 0;
46a3f46d 626 kmem_slab_free(z, ZoneSize); /* may block */
51295aee 627 atomic_add_int(&ZoneGenAlloc, -ZoneSize / 1024);
46a3f46d
MD
628 }
629 crit_exit();
630 }
5fee07e6 631
46a3f46d 632 /*
5fee07e6 633 * XXX handle oversized frees that were queued from kfree().
46a3f46d 634 */
dc1fd4b3 635 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) {
46a3f46d
MD
636 crit_enter();
637 if ((z = slgd->FreeOvZones) != NULL) {
7e83df33
MD
638 vm_size_t tsize;
639
46a3f46d
MD
640 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
641 slgd->FreeOvZones = z->z_Next;
7e83df33
MD
642 tsize = z->z_ChunkSize;
643 kmem_slab_free(z, tsize); /* may block */
644 atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024);
46a3f46d
MD
645 }
646 crit_exit();
a7cf0021
MD
647 }
648
649 /*
a108bf71
MD
650 * Handle large allocations directly. There should not be very many of
651 * these so performance is not a big issue.
652 *
b543eeed
MD
653 * The backend allocator is pretty nasty on a SMP system. Use the
654 * slab allocator for one and two page-sized chunks even though we lose
655 * some efficiency. XXX maybe fix mmio and the elf loader instead.
a108bf71 656 */
b543eeed 657 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
722871d3 658 int *kup;
a108bf71
MD
659
660 size = round_page(size);
661 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
f2b5daf9 662 if (chunk == NULL) {
5fee07e6 663 logmemory(malloc_end, NULL, type, size, flags);
a108bf71 664 return(NULL);
f2b5daf9 665 }
665206ee 666 atomic_add_int(&ZoneBigAlloc, (int)size / 1024);
a108bf71 667 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
8f1d5415 668 flags |= M_PASSIVE_ZERO;
a108bf71 669 kup = btokup(chunk);
722871d3 670 *kup = size / PAGE_SIZE;
a108bf71
MD
671 crit_enter();
672 goto done;
673 }
674
675 /*
676 * Attempt to allocate out of an existing zone. First try the free list,
677 * then allocate out of unallocated space. If we find a good zone move
678 * it to the head of the list so later allocations find it quickly
679 * (we might have thousands of zones in the list).
680 *
681 * Note: zoneindex() will panic of size is too large.
682 */
683 zi = zoneindex(&size);
684 KKASSERT(zi < NZONES);
685 crit_enter();
a108bf71 686
5fee07e6 687 if ((z = slgd->ZoneAry[zi]) != NULL) {
a108bf71 688 /*
5fee07e6
MD
689 * Locate a chunk - we have to have at least one. If this is the
690 * last chunk go ahead and do the work to retrieve chunks freed
691 * from remote cpus, and if the zone is still empty move it off
692 * the ZoneAry.
a108bf71 693 */
5fee07e6
MD
694 if (--z->z_NFree <= 0) {
695 KKASSERT(z->z_NFree == 0);
696
697#ifdef SMP
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 }
724#endif
725 /*
726 * Remove from the zone list if no free chunks remain.
727 * Clear RSignal
728 */
729 if (z->z_NFree == 0) {
730 slgd->ZoneAry[zi] = z->z_Next;
731 z->z_Next = NULL;
732 } else {
733 z->z_RSignal = 0;
734 }
a108bf71
MD
735 }
736
737 /*
5fee07e6 738 * Fast path, we have chunks available in z_LChunks.
a108bf71 739 */
5fee07e6
MD
740 chunk = z->z_LChunks;
741 if (chunk) {
10cc6608 742 chunk_mark_allocated(z, chunk);
5fee07e6
MD
743 z->z_LChunks = chunk->c_Next;
744 if (z->z_LChunks == NULL)
745 z->z_LChunksp = &z->z_LChunks;
bbb201fd
MD
746#ifdef SLAB_DEBUG
747 slab_record_source(z, file, line);
748#endif
a108bf71 749 goto done;
a108bf71
MD
750 }
751
752 /*
5fee07e6
MD
753 * No chunks are available in LChunks, the free chunk MUST be
754 * in the never-before-used memory area, controlled by UIndex.
755 *
756 * The consequences are very serious if our zone got corrupted so
757 * we use an explicit panic rather than a KASSERT.
a108bf71 758 */
1c5ca4f3 759 if (z->z_UIndex + 1 != z->z_NMax)
5fee07e6 760 ++z->z_UIndex;
1c5ca4f3
MD
761 else
762 z->z_UIndex = 0;
5fee07e6 763
1c5ca4f3
MD
764 if (z->z_UIndex == z->z_UEndIndex)
765 panic("slaballoc: corrupted zone");
5fee07e6 766
1c5ca4f3 767 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
8f1d5415 768 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 769 flags &= ~M_ZERO;
8f1d5415
MD
770 flags |= M_PASSIVE_ZERO;
771 }
10cc6608 772 chunk_mark_allocated(z, chunk);
bbb201fd
MD
773#ifdef SLAB_DEBUG
774 slab_record_source(z, file, line);
775#endif
a108bf71
MD
776 goto done;
777 }
778
779 /*
780 * If all zones are exhausted we need to allocate a new zone for this
781 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
6ab8e1da
MD
782 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
783 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
784 * we do not pre-zero it because we do not want to mess up the L1 cache.
a108bf71
MD
785 *
786 * At least one subsystem, the tty code (see CROUND) expects power-of-2
787 * allocations to be power-of-2 aligned. We maintain compatibility by
788 * adjusting the base offset below.
789 */
790 {
791 int off;
722871d3 792 int *kup;
a108bf71
MD
793
794 if ((z = slgd->FreeZones) != NULL) {
795 slgd->FreeZones = z->z_Next;
796 --slgd->NFreeZones;
797 bzero(z, sizeof(SLZone));
6ab8e1da 798 z->z_Flags |= SLZF_UNOTZEROD;
a108bf71
MD
799 } else {
800 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
801 if (z == NULL)
802 goto fail;
51295aee 803 atomic_add_int(&ZoneGenAlloc, ZoneSize / 1024);
a108bf71
MD
804 }
805
806 /*
10cc6608
MD
807 * How big is the base structure?
808 */
809#if defined(INVARIANTS)
810 /*
811 * Make room for z_Bitmap. An exact calculation is somewhat more
812 * complicated so don't make an exact calculation.
813 */
814 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
815 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
816#else
817 off = sizeof(SLZone);
818#endif
819
820 /*
a108bf71
MD
821 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
822 * Otherwise just 8-byte align the data.
823 */
824 if ((size | (size - 1)) + 1 == (size << 1))
10cc6608 825 off = (off + size - 1) & ~(size - 1);
a108bf71 826 else
10cc6608 827 off = (off + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
a108bf71
MD
828 z->z_Magic = ZALLOC_SLAB_MAGIC;
829 z->z_ZoneIndex = zi;
830 z->z_NMax = (ZoneSize - off) / size;
831 z->z_NFree = z->z_NMax - 1;
1c5ca4f3
MD
832 z->z_BasePtr = (char *)z + off;
833 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
a108bf71 834 z->z_ChunkSize = size;
2db3b277 835 z->z_CpuGd = gd;
bba6a44d 836 z->z_Cpu = gd->gd_cpuid;
5fee07e6 837 z->z_LChunksp = &z->z_LChunks;
bbb201fd
MD
838#ifdef SLAB_DEBUG
839 bcopy(z->z_Sources, z->z_AltSources, sizeof(z->z_Sources));
840 bzero(z->z_Sources, sizeof(z->z_Sources));
841#endif
1c5ca4f3 842 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
a108bf71
MD
843 z->z_Next = slgd->ZoneAry[zi];
844 slgd->ZoneAry[zi] = z;
8f1d5415 845 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 846 flags &= ~M_ZERO; /* already zero'd */
8f1d5415
MD
847 flags |= M_PASSIVE_ZERO;
848 }
5fee07e6 849 kup = btokup(z);
722871d3 850 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */
10cc6608 851 chunk_mark_allocated(z, chunk);
bbb201fd
MD
852#ifdef SLAB_DEBUG
853 slab_record_source(z, file, line);
854#endif
1c5ca4f3
MD
855
856 /*
857 * Slide the base index for initial allocations out of the next
858 * zone we create so we do not over-weight the lower part of the
859 * cpu memory caches.
860 */
861 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
862 & (ZALLOC_MAX_ZONE_SIZE - 1);
a108bf71 863 }
5fee07e6 864
a108bf71 865done:
bba6a44d
MD
866 ++type->ks_inuse[gd->gd_cpuid];
867 type->ks_memuse[gd->gd_cpuid] += size;
38e34349 868 type->ks_loosememuse += size; /* not MP synchronized */
a108bf71 869 crit_exit();
5fee07e6 870
a108bf71
MD
871 if (flags & M_ZERO)
872 bzero(chunk, size);
bba6a44d 873#ifdef INVARIANTS
d2182dc1
MD
874 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
875 if (use_malloc_pattern) {
876 for (i = 0; i < size; i += sizeof(int)) {
877 *(int *)((char *)chunk + i) = -1;
878 }
879 }
bba6a44d 880 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
d2182dc1 881 }
bba6a44d 882#endif
5fee07e6 883 logmemory(malloc_end, chunk, type, size, flags);
a108bf71
MD
884 return(chunk);
885fail:
886 crit_exit();
5fee07e6 887 logmemory(malloc_end, NULL, type, size, flags);
a108bf71
MD
888 return(NULL);
889}
890
38e34349
MD
891/*
892 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
893 *
894 * Generally speaking this routine is not called very often and we do
895 * not attempt to optimize it beyond reusing the same pointer if the
896 * new size fits within the chunking of the old pointer's zone.
897 */
bbb201fd
MD
898#ifdef SLAB_DEBUG
899void *
900krealloc_debug(void *ptr, unsigned long size,
901 struct malloc_type *type, int flags,
902 const char *file, int line)
903#else
a108bf71 904void *
8aca2bd4 905krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
bbb201fd 906#endif
a108bf71 907{
722871d3 908 unsigned long osize;
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
MD
958 if (size < ZoneLimit) {
959 zoneindex(&size);
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
38e34349 970/*
45d2b1d8
MD
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
985/*
38e34349
MD
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
1d712609 1010#ifdef SMP
a108bf71 1011/*
5fee07e6 1012 * Notify our cpu that a remote cpu has freed some chunks in a zone that
df9daea8
MD
1013 * we own. RCount will be bumped so the memory should be good, but validate
1014 * that it really is.
a108bf71
MD
1015 */
1016static
1017void
5fee07e6 1018kfree_remote(void *ptr)
a108bf71 1019{
5fee07e6
MD
1020 SLGlobalData *slgd;
1021 SLChunk *bchunk;
1022 SLZone *z;
1023 int nfree;
722871d3 1024 int *kup;
5fee07e6 1025
5fee07e6
MD
1026 slgd = &mycpu->gd_slab;
1027 z = ptr;
1028 kup = btokup(z);
df9daea8
MD
1029 KKASSERT(*kup == -((int)mycpuid + 1));
1030 KKASSERT(z->z_RCount > 0);
1031 atomic_subtract_int(&z->z_RCount, 1);
5fee07e6 1032
5bf48697 1033 logmemory(free_rem_beg, z, NULL, 0L, 0);
df9daea8
MD
1034 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1035 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
1036 nfree = z->z_NFree;
5fee07e6 1037
df9daea8
MD
1038 /*
1039 * Indicate that we will no longer be off of the ZoneAry by
1040 * clearing RSignal.
1041 */
1042 if (z->z_RChunks)
1043 z->z_RSignal = 0;
5fee07e6 1044
df9daea8
MD
1045 /*
1046 * Atomically extract the bchunks list and then process it back
1047 * into the lchunks list. We want to append our bchunks to the
1048 * lchunks list and not prepend since we likely do not have
1049 * cache mastership of the related data (not that it helps since
1050 * we are using c_Next).
1051 */
1052 while ((bchunk = z->z_RChunks) != NULL) {
1053 cpu_ccfence();
1054 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
1055 *z->z_LChunksp = bchunk;
1056 while (bchunk) {
1057 chunk_mark_free(z, bchunk);
1058 z->z_LChunksp = &bchunk->c_Next;
1059 bchunk = bchunk->c_Next;
1060 ++z->z_NFree;
5fee07e6 1061 }
df9daea8 1062 break;
5fee07e6 1063 }
df9daea8
MD
1064 }
1065 if (z->z_NFree && nfree == 0) {
1066 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1067 slgd->ZoneAry[z->z_ZoneIndex] = z;
1068 }
5fee07e6 1069
df9daea8
MD
1070 /*
1071 * If the zone becomes totally free, and there are other zones we
1072 * can allocate from, move this zone to the FreeZones list. Since
1073 * this code can be called from an IPI callback, do *NOT* try to mess
1074 * with kernel_map here. Hysteresis will be performed at malloc() time.
1075 *
1076 * Do not move the zone if there is an IPI inflight, otherwise MP
1077 * races can result in our free_remote code accessing a destroyed
1078 * zone.
1079 */
1080 if (z->z_NFree == z->z_NMax &&
1081 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1082 z->z_RCount == 0
1083 ) {
1084 SLZone **pz;
1085 int *kup;
5fee07e6 1086
df9daea8
MD
1087 for (pz = &slgd->ZoneAry[z->z_ZoneIndex];
1088 z != *pz;
1089 pz = &(*pz)->z_Next) {
1090 ;
5fee07e6 1091 }
df9daea8
MD
1092 *pz = z->z_Next;
1093 z->z_Magic = -1;
1094 z->z_Next = slgd->FreeZones;
1095 slgd->FreeZones = z;
1096 ++slgd->NFreeZones;
1097 kup = btokup(z);
1098 *kup = 0;
5fee07e6 1099 }
5bf48697 1100 logmemory(free_rem_end, z, bchunk, 0L, 0);
a108bf71
MD
1101}
1102
1d712609
MD
1103#endif
1104
38e34349 1105/*
5b287bba 1106 * free (SLAB ALLOCATOR)
38e34349
MD
1107 *
1108 * Free a memory block previously allocated by malloc. Note that we do not
5fee07e6 1109 * attempt to update ks_loosememuse as MP races could prevent us from
38e34349 1110 * checking memory limits in malloc.
5b287bba
MD
1111 *
1112 * MPSAFE
38e34349 1113 */
a108bf71 1114void
8aca2bd4 1115kfree(void *ptr, struct malloc_type *type)
a108bf71
MD
1116{
1117 SLZone *z;
1118 SLChunk *chunk;
1119 SLGlobalData *slgd;
bba6a44d 1120 struct globaldata *gd;
722871d3 1121 int *kup;
5fee07e6 1122 unsigned long size;
d8100bdc
SW
1123#ifdef SMP
1124 SLChunk *bchunk;
5fee07e6 1125 int rsignal;
d8100bdc 1126#endif
a108bf71 1127
b68ad50c 1128 logmemory_quick(free_beg);
bba6a44d
MD
1129 gd = mycpu;
1130 slgd = &gd->gd_slab;
a108bf71 1131
d39911d9
JS
1132 if (ptr == NULL)
1133 panic("trying to free NULL pointer");
1134
a108bf71
MD
1135 /*
1136 * Handle special 0-byte allocations
1137 */
f2b5daf9 1138 if (ptr == ZERO_LENGTH_PTR) {
5bf48697 1139 logmemory(free_zero, ptr, type, -1UL, 0);
b68ad50c 1140 logmemory_quick(free_end);
a108bf71 1141 return;
f2b5daf9 1142 }
a108bf71
MD
1143
1144 /*
5fee07e6
MD
1145 * Panic on bad malloc type
1146 */
1147 if (type->ks_magic != M_MAGIC)
1148 panic("free: malloc type lacks magic");
1149
1150 /*
a108bf71
MD
1151 * Handle oversized allocations. XXX we really should require that a
1152 * size be passed to free() instead of this nonsense.
bba6a44d
MD
1153 *
1154 * This code is never called via an ipi.
a108bf71 1155 */
5fee07e6 1156 kup = btokup(ptr);
722871d3
MD
1157 if (*kup > 0) {
1158 size = *kup << PAGE_SHIFT;
1159 *kup = 0;
a108bf71 1160#ifdef INVARIANTS
5fee07e6
MD
1161 KKASSERT(sizeof(weirdary) <= size);
1162 bcopy(weirdary, ptr, sizeof(weirdary));
a108bf71 1163#endif
5fee07e6
MD
1164 /*
1165 * NOTE: For oversized allocations we do not record the
1166 * originating cpu. It gets freed on the cpu calling
1167 * kfree(). The statistics are in aggregate.
1168 *
1169 * note: XXX we have still inherited the interrupts-can't-block
1170 * assumption. An interrupt thread does not bump
1171 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1172 * primarily until we can fix softupdate's assumptions about free().
1173 */
1174 crit_enter();
1175 --type->ks_inuse[gd->gd_cpuid];
1176 type->ks_memuse[gd->gd_cpuid] -= size;
1177 if (mycpu->gd_intr_nesting_level ||
1178 (gd->gd_curthread->td_flags & TDF_INTTHREAD))
1179 {
1180 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1181 z = (SLZone *)ptr;
1182 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1183 z->z_Next = slgd->FreeOvZones;
1184 z->z_ChunkSize = size;
1185 slgd->FreeOvZones = z;
1186 crit_exit();
1187 } else {
1188 crit_exit();
1189 logmemory(free_ovsz, ptr, type, size, 0);
1190 kmem_slab_free(ptr, size); /* may block */
1191 atomic_add_int(&ZoneBigAlloc, -(int)size / 1024);
a108bf71 1192 }
5fee07e6
MD
1193 logmemory_quick(free_end);
1194 return;
a108bf71
MD
1195 }
1196
1197 /*
1198 * Zone case. Figure out the zone based on the fact that it is
1199 * ZoneSize aligned.
1200 */
5fee07e6
MD
1201 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1202 kup = btokup(z);
722871d3 1203 KKASSERT(*kup < 0);
a108bf71
MD
1204 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1205
1206 /*
5fee07e6
MD
1207 * If we do not own the zone then use atomic ops to free to the
1208 * remote cpu linked list and notify the target zone using a
1209 * passive message.
1210 *
1211 * The target zone cannot be deallocated while we own a chunk of it,
1212 * so the zone header's storage is stable until the very moment
1213 * we adjust z_RChunks. After that we cannot safely dereference (z).
1214 *
1215 * (no critical section needed)
a108bf71 1216 */
2db3b277 1217 if (z->z_CpuGd != gd) {
75c7ffea 1218#ifdef SMP
5fee07e6
MD
1219 /*
1220 * Making these adjustments now allow us to avoid passing (type)
1221 * to the remote cpu. Note that ks_inuse/ks_memuse is being
28135cc2
MD
1222 * adjusted on OUR cpu, not the zone cpu, but it should all still
1223 * sum up properly and cancel out.
5fee07e6 1224 */
28135cc2
MD
1225 crit_enter();
1226 --type->ks_inuse[gd->gd_cpuid];
1227 type->ks_memuse[gd->gd_cpuid] -= z->z_ChunkSize;
1228 crit_exit();
5fee07e6
MD
1229
1230 /*
1231 * WARNING! This code competes with other cpus. Once we
1232 * successfully link the chunk to RChunks the remote
1233 * cpu can rip z's storage out from under us.
df9daea8
MD
1234 *
1235 * Bumping RCount prevents z's storage from getting
1236 * ripped out.
5fee07e6
MD
1237 */
1238 rsignal = z->z_RSignal;
1239 cpu_lfence();
df9daea8
MD
1240 if (rsignal)
1241 atomic_add_int(&z->z_RCount, 1);
5fee07e6
MD
1242
1243 chunk = ptr;
1244 for (;;) {
1245 bchunk = z->z_RChunks;
1246 cpu_ccfence();
1247 chunk->c_Next = bchunk;
1248 cpu_sfence();
1249
1250 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1251 break;
1252 }
5fee07e6
MD
1253
1254 /*
1255 * We have to signal the remote cpu if our actions will cause
1256 * the remote zone to be placed back on ZoneAry so it can
1257 * move the zone back on.
1258 *
1259 * We only need to deal with NULL->non-NULL RChunk transitions
1260 * and only if z_RSignal is set. We interlock by reading rsignal
1261 * before adding our chunk to RChunks. This should result in
1262 * virtually no IPI traffic.
1263 *
1264 * We can use a passive IPI to reduce overhead even further.
1265 */
1266 if (bchunk == NULL && rsignal) {
5bf48697 1267 logmemory(free_request, ptr, type, (unsigned long)z->z_ChunkSize, 0);
5fee07e6 1268 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
df9daea8
MD
1269 /* z can get ripped out from under us from this point on */
1270 } else if (rsignal) {
1271 atomic_subtract_int(&z->z_RCount, 1);
1272 /* z can get ripped out from under us from this point on */
5fee07e6 1273 }
75c7ffea
MD
1274#else
1275 panic("Corrupt SLZone");
1276#endif
b68ad50c 1277 logmemory_quick(free_end);
a108bf71
MD
1278 return;
1279 }
1280
5fee07e6
MD
1281 /*
1282 * kfree locally
1283 */
5bf48697 1284 logmemory(free_chunk, ptr, type, (unsigned long)z->z_ChunkSize, 0);
f2b5daf9 1285
a108bf71 1286 crit_enter();
a108bf71 1287 chunk = ptr;
10cc6608 1288 chunk_mark_free(z, chunk);
a108bf71
MD
1289
1290 /*
1291 * Put weird data into the memory to detect modifications after freeing,
1292 * illegal pointer use after freeing (we should fault on the odd address),
1293 * and so forth. XXX needs more work, see the old malloc code.
1294 */
1295#ifdef INVARIANTS
1296 if (z->z_ChunkSize < sizeof(weirdary))
1297 bcopy(weirdary, chunk, z->z_ChunkSize);
1298 else
1299 bcopy(weirdary, chunk, sizeof(weirdary));
1300#endif
1301
1302 /*
5fee07e6
MD
1303 * Add this free non-zero'd chunk to a linked list for reuse. Add
1304 * to the front of the linked list so it is more likely to be
1305 * reallocated, since it is already in our L1 cache.
a108bf71 1306 */
6ab8e1da 1307#ifdef INVARIANTS
c439ad8f 1308 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
fc92d4aa 1309 panic("BADFREE %p", chunk);
a108bf71 1310#endif
5fee07e6
MD
1311 chunk->c_Next = z->z_LChunks;
1312 z->z_LChunks = chunk;
1313 if (chunk->c_Next == NULL)
1314 z->z_LChunksp = &chunk->c_Next;
1315
6ab8e1da 1316#ifdef INVARIANTS
c439ad8f 1317 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
a108bf71 1318 panic("BADFREE2");
6ab8e1da 1319#endif
a108bf71
MD
1320
1321 /*
1322 * Bump the number of free chunks. If it becomes non-zero the zone
1323 * must be added back onto the appropriate list.
1324 */
1325 if (z->z_NFree++ == 0) {
1326 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1327 slgd->ZoneAry[z->z_ZoneIndex] = z;
1328 }
1329
bba6a44d
MD
1330 --type->ks_inuse[z->z_Cpu];
1331 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
a108bf71
MD
1332
1333 /*
1334 * If the zone becomes totally free, and there are other zones we
a7cf0021
MD
1335 * can allocate from, move this zone to the FreeZones list. Since
1336 * this code can be called from an IPI callback, do *NOT* try to mess
1337 * with kernel_map here. Hysteresis will be performed at malloc() time.
a108bf71
MD
1338 */
1339 if (z->z_NFree == z->z_NMax &&
df9daea8
MD
1340 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1341 z->z_RCount == 0
a108bf71
MD
1342 ) {
1343 SLZone **pz;
722871d3 1344 int *kup;
a108bf71
MD
1345
1346 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
1347 ;
1348 *pz = z->z_Next;
1349 z->z_Magic = -1;
a7cf0021
MD
1350 z->z_Next = slgd->FreeZones;
1351 slgd->FreeZones = z;
1352 ++slgd->NFreeZones;
5fee07e6 1353 kup = btokup(z);
722871d3 1354 *kup = 0;
a108bf71 1355 }
b68ad50c 1356 logmemory_quick(free_end);
a108bf71
MD
1357 crit_exit();
1358}
1359
10cc6608 1360#if defined(INVARIANTS)
5fee07e6 1361
10cc6608
MD
1362/*
1363 * Helper routines for sanity checks
1364 */
1365static
1366void
1367chunk_mark_allocated(SLZone *z, void *chunk)
1368{
1369 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1370 __uint32_t *bitptr;
1371
5fee07e6
MD
1372 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1373 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1374 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
10cc6608
MD
1375 bitptr = &z->z_Bitmap[bitdex >> 5];
1376 bitdex &= 31;
5fee07e6
MD
1377 KASSERT((*bitptr & (1 << bitdex)) == 0,
1378 ("memory chunk %p is already allocated!", chunk));
10cc6608
MD
1379 *bitptr |= 1 << bitdex;
1380}
1381
1382static
1383void
1384chunk_mark_free(SLZone *z, void *chunk)
1385{
1386 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1387 __uint32_t *bitptr;
1388
5fee07e6
MD
1389 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1390 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1391 ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
10cc6608
MD
1392 bitptr = &z->z_Bitmap[bitdex >> 5];
1393 bitdex &= 31;
5fee07e6
MD
1394 KASSERT((*bitptr & (1 << bitdex)) != 0,
1395 ("memory chunk %p is already free!", chunk));
10cc6608
MD
1396 *bitptr &= ~(1 << bitdex);
1397}
1398
1399#endif
1400
a108bf71 1401/*
5b287bba 1402 * kmem_slab_alloc()
a108bf71
MD
1403 *
1404 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1405 * specified alignment. M_* flags are expected in the flags field.
1406 *
1407 * Alignment must be a multiple of PAGE_SIZE.
1408 *
1409 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1410 * but when we move zalloc() over to use this function as its backend
1411 * we will have to switch to kreserve/krelease and call reserve(0)
1412 * after the new space is made available.
dc1fd4b3
MD
1413 *
1414 * Interrupt code which has preempted other code is not allowed to
c397c465
MD
1415 * use PQ_CACHE pages. However, if an interrupt thread is run
1416 * non-preemptively or blocks and then runs non-preemptively, then
b12defdc 1417 * it is free to use PQ_CACHE pages. <--- may not apply any longer XXX
a108bf71
MD
1418 */
1419static void *
1420kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1421{
1422 vm_size_t i;
1423 vm_offset_t addr;
1de1e800 1424 int count, vmflags, base_vmflags;
b12defdc
MD
1425 vm_page_t mbase = NULL;
1426 vm_page_t m;
dc1fd4b3 1427 thread_t td;
a108bf71
MD
1428
1429 size = round_page(size);
e4846942 1430 addr = vm_map_min(&kernel_map);
a108bf71 1431
a108bf71
MD
1432 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1433 crit_enter();
e4846942 1434 vm_map_lock(&kernel_map);
c809941b 1435 if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) {
e4846942 1436 vm_map_unlock(&kernel_map);
8cb2bf45 1437 if ((flags & M_NULLOK) == 0)
a108bf71 1438 panic("kmem_slab_alloc(): kernel_map ran out of space!");
a108bf71 1439 vm_map_entry_release(count);
2de4f77e 1440 crit_exit();
a108bf71
MD
1441 return(NULL);
1442 }
e4846942
MD
1443
1444 /*
1445 * kernel_object maps 1:1 to kernel_map.
1446 */
b12defdc
MD
1447 vm_object_hold(&kernel_object);
1448 vm_object_reference_locked(&kernel_object);
e4846942
MD
1449 vm_map_insert(&kernel_map, &count,
1450 &kernel_object, addr, addr, addr + size,
1b874851
MD
1451 VM_MAPTYPE_NORMAL,
1452 VM_PROT_ALL, VM_PROT_ALL,
1453 0);
b12defdc
MD
1454 vm_object_drop(&kernel_object);
1455 vm_map_set_wired_quick(&kernel_map, addr, size, &count);
1456 vm_map_unlock(&kernel_map);
a108bf71 1457
dc1fd4b3 1458 td = curthread;
dc1fd4b3 1459
1de1e800
JS
1460 base_vmflags = 0;
1461 if (flags & M_ZERO)
1462 base_vmflags |= VM_ALLOC_ZERO;
1463 if (flags & M_USE_RESERVE)
1464 base_vmflags |= VM_ALLOC_SYSTEM;
1465 if (flags & M_USE_INTERRUPT_RESERVE)
1466 base_vmflags |= VM_ALLOC_INTERRUPT;
77912481
MD
1467 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1468 panic("kmem_slab_alloc: bad flags %08x (%p)",
1469 flags, ((int **)&size)[-1]);
1470 }
1de1e800 1471
a108bf71 1472 /*
b12defdc
MD
1473 * Allocate the pages. Do not mess with the PG_ZERO flag or map
1474 * them yet. VM_ALLOC_NORMAL can only be set if we are not preempting.
1475 *
1476 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1477 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1478 * implied in this case), though I'm not sure if we really need to
1479 * do that.
a108bf71 1480 */
b12defdc
MD
1481 vmflags = base_vmflags;
1482 if (flags & M_WAITOK) {
1483 if (td->td_preempted)
1484 vmflags |= VM_ALLOC_SYSTEM;
1485 else
1486 vmflags |= VM_ALLOC_NORMAL;
1487 }
a108bf71 1488
b12defdc
MD
1489 vm_object_hold(&kernel_object);
1490 for (i = 0; i < size; i += PAGE_SIZE) {
e4846942 1491 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
b12defdc
MD
1492 if (i == 0)
1493 mbase = m;
dc1fd4b3
MD
1494
1495 /*
1496 * If the allocation failed we either return NULL or we retry.
1497 *
c397c465
MD
1498 * If M_WAITOK is specified we wait for more memory and retry.
1499 * If M_WAITOK is specified from a preemption we yield instead of
1500 * wait. Livelock will not occur because the interrupt thread
1501 * will not be preempting anyone the second time around after the
1502 * yield.
dc1fd4b3 1503 */
a108bf71 1504 if (m == NULL) {
c397c465 1505 if (flags & M_WAITOK) {
fe1e98d0 1506 if (td->td_preempted) {
77912481 1507 lwkt_switch();
dc1fd4b3 1508 } else {
4ecf7cc9 1509 vm_wait(0);
dc1fd4b3 1510 }
a108bf71
MD
1511 i -= PAGE_SIZE; /* retry */
1512 continue;
1513 }
b12defdc
MD
1514 break;
1515 }
1516 }
dc1fd4b3 1517
b12defdc
MD
1518 /*
1519 * Check and deal with an allocation failure
1520 */
1521 if (i != size) {
1522 while (i != 0) {
1523 i -= PAGE_SIZE;
1524 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1525 /* page should already be busy */
1526 vm_page_free(m);
a108bf71 1527 }
b12defdc
MD
1528 vm_map_lock(&kernel_map);
1529 vm_map_delete(&kernel_map, addr, addr + size, &count);
1530 vm_map_unlock(&kernel_map);
1531 vm_object_drop(&kernel_object);
1532
1533 vm_map_entry_release(count);
1534 crit_exit();
1535 return(NULL);
a108bf71
MD
1536 }
1537
1538 /*
dc1fd4b3
MD
1539 * Success!
1540 *
b12defdc
MD
1541 * NOTE: The VM pages are still busied. mbase points to the first one
1542 * but we have to iterate via vm_page_next()
a108bf71 1543 */
b12defdc 1544 vm_object_drop(&kernel_object);
a108bf71
MD
1545 crit_exit();
1546
1547 /*
1548 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1549 */
b12defdc
MD
1550 m = mbase;
1551 i = 0;
a108bf71 1552
b12defdc
MD
1553 while (i < size) {
1554 /*
1555 * page should already be busy
1556 */
a108bf71
MD
1557 m->valid = VM_PAGE_BITS_ALL;
1558 vm_page_wire(m);
921c891e
MD
1559 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL | VM_PROT_NOSYNC,
1560 1, NULL);
a108bf71
MD
1561 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1562 bzero((char *)addr + i, PAGE_SIZE);
1563 vm_page_flag_clear(m, PG_ZERO);
17cde63e
MD
1564 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1565 vm_page_flag_set(m, PG_REFERENCED);
a491077e 1566 vm_page_wakeup(m);
b12defdc
MD
1567
1568 i += PAGE_SIZE;
1569 vm_object_hold(&kernel_object);
1570 m = vm_page_next(m);
1571 vm_object_drop(&kernel_object);
a108bf71 1572 }
b12defdc 1573 smp_invltlb();
a108bf71
MD
1574 vm_map_entry_release(count);
1575 return((void *)addr);
1576}
1577
38e34349 1578/*
5b287bba 1579 * kmem_slab_free()
38e34349 1580 */
a108bf71
MD
1581static void
1582kmem_slab_free(void *ptr, vm_size_t size)
1583{
1584 crit_enter();
e4846942 1585 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
a108bf71
MD
1586 crit_exit();
1587}
1588
55126ffe 1589void *
55126ffe
SZ
1590kmalloc_cachealign(unsigned long size_alloc, struct malloc_type *type,
1591 int flags)
1592{
1593 if (size_alloc < __VM_CACHELINE_SIZE)
1594 size_alloc = __VM_CACHELINE_SIZE;
1e57f867 1595 return kmalloc(size_alloc, type, flags | M_POWEROF2);
55126ffe 1596}