kernel - Correct bug in malloc_type ks_memuse/ks_inuse management
[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.
8c10bfcf 37 *
40153c65 38 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.55 2008/10/22 01:42:17 dillon Exp $
a108bf71
MD
39 *
40 * This module implements a slab allocator drop-in replacement for the
41 * kernel malloc().
42 *
43 * A slab allocator reserves a ZONE for each chunk size, then lays the
44 * chunks out in an array within the zone. Allocation and deallocation
45 * is nearly instantanious, and fragmentation/overhead losses are limited
46 * to a fixed worst-case amount.
47 *
48 * The downside of this slab implementation is in the chunk size
49 * multiplied by the number of zones. ~80 zones * 128K = 10MB of VM per cpu.
50 * In a kernel implementation all this memory will be physical so
51 * the zone size is adjusted downward on machines with less physical
52 * memory. The upside is that overhead is bounded... this is the *worst*
53 * case overhead.
54 *
55 * Slab management is done on a per-cpu basis and no locking or mutexes
56 * are required, only a critical section. When one cpu frees memory
57 * belonging to another cpu's slab manager an asynchronous IPI message
58 * will be queued to execute the operation. In addition, both the
59 * high level slab allocator and the low level zone allocator optimize
60 * M_ZERO requests, and the slab allocator does not have to pre initialize
61 * the linked list of chunks.
62 *
63 * XXX Balancing is needed between cpus. Balance will be handled through
64 * asynchronous IPIs primarily by reassigning the z_Cpu ownership of chunks.
65 *
66 * XXX If we have to allocate a new zone and M_USE_RESERVE is set, use of
67 * the new zone should be restricted to M_USE_RESERVE requests only.
68 *
69 * Alloc Size Chunking Number of zones
70 * 0-127 8 16
71 * 128-255 16 8
72 * 256-511 32 8
73 * 512-1023 64 8
74 * 1024-2047 128 8
75 * 2048-4095 256 8
76 * 4096-8191 512 8
77 * 8192-16383 1024 8
78 * 16384-32767 2048 8
79 * (if PAGE_SIZE is 4K the maximum zone allocation is 16383)
80 *
46a3f46d 81 * Allocations >= ZoneLimit go directly to kmem.
a108bf71
MD
82 *
83 * API REQUIREMENTS AND SIDE EFFECTS
84 *
85 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we
86 * have remained compatible with the following API requirements:
87 *
88 * + small power-of-2 sized allocations are power-of-2 aligned (kern_tty)
3d177b31 89 * + all power-of-2 sized allocations are power-of-2 aligned (twe)
a108bf71
MD
90 * + malloc(0) is allowed and returns non-NULL (ahc driver)
91 * + ability to allocate arbitrarily large chunks of memory
92 */
93
94#include "opt_vm.h"
95
a108bf71
MD
96#include <sys/param.h>
97#include <sys/systm.h>
98#include <sys/kernel.h>
99#include <sys/slaballoc.h>
100#include <sys/mbuf.h>
101#include <sys/vmmeter.h>
102#include <sys/lock.h>
103#include <sys/thread.h>
104#include <sys/globaldata.h>
d2182dc1 105#include <sys/sysctl.h>
f2b5daf9 106#include <sys/ktr.h>
a108bf71
MD
107
108#include <vm/vm.h>
109#include <vm/vm_param.h>
110#include <vm/vm_kern.h>
111#include <vm/vm_extern.h>
112#include <vm/vm_object.h>
113#include <vm/pmap.h>
114#include <vm/vm_map.h>
115#include <vm/vm_page.h>
116#include <vm/vm_pageout.h>
117
118#include <machine/cpu.h>
119
120#include <sys/thread2.h>
121
122#define arysize(ary) (sizeof(ary)/sizeof((ary)[0]))
123
f2b5daf9
MD
124#define MEMORY_STRING "ptr=%p type=%p size=%d flags=%04x"
125#define MEMORY_ARG_SIZE (sizeof(void *) * 2 + sizeof(unsigned long) + \
126 sizeof(int))
127
128#if !defined(KTR_MEMORY)
129#define KTR_MEMORY KTR_ALL
130#endif
131KTR_INFO_MASTER(memory);
5fee07e6
MD
132KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin", 0);
133KTR_INFO(KTR_MEMORY, memory, malloc_end, 1, MEMORY_STRING, MEMORY_ARG_SIZE);
134KTR_INFO(KTR_MEMORY, memory, free_zero, 2, MEMORY_STRING, MEMORY_ARG_SIZE);
135KTR_INFO(KTR_MEMORY, memory, free_ovsz, 3, MEMORY_STRING, MEMORY_ARG_SIZE);
136KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 4, MEMORY_STRING, MEMORY_ARG_SIZE);
137KTR_INFO(KTR_MEMORY, memory, free_chunk, 5, MEMORY_STRING, MEMORY_ARG_SIZE);
f2b5daf9 138#ifdef SMP
5fee07e6
MD
139KTR_INFO(KTR_MEMORY, memory, free_request, 6, MEMORY_STRING, MEMORY_ARG_SIZE);
140KTR_INFO(KTR_MEMORY, memory, free_rem_beg, 7, MEMORY_STRING, MEMORY_ARG_SIZE);
141KTR_INFO(KTR_MEMORY, memory, free_rem_end, 8, MEMORY_STRING, MEMORY_ARG_SIZE);
f2b5daf9 142#endif
5fee07e6
MD
143KTR_INFO(KTR_MEMORY, memory, free_beg, 9, "free begin", 0);
144KTR_INFO(KTR_MEMORY, memory, free_end, 10, "free end", 0);
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 struct kmemusage *kmemusage;
162static int32_t weirdary[16];
163
164static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
165static void kmem_slab_free(void *ptr, vm_size_t bytes);
5fee07e6 166
10cc6608
MD
167#if defined(INVARIANTS)
168static void chunk_mark_allocated(SLZone *z, void *chunk);
169static void chunk_mark_free(SLZone *z, void *chunk);
5fee07e6
MD
170#else
171#define chunk_mark_allocated(z, chunk)
172#define chunk_mark_free(z, chunk)
10cc6608 173#endif
a108bf71
MD
174
175/*
176 * Misc constants. Note that allocations that are exact multiples of
177 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
178 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
179 */
180#define MIN_CHUNK_SIZE 8 /* in bytes */
181#define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
182#define ZONE_RELS_THRESH 2 /* threshold number of zones */
183#define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
184
185/*
186 * The WEIRD_ADDR is used as known text to copy into free objects to
187 * try to create deterministic failure cases if the data is accessed after
188 * free.
189 */
190#define WEIRD_ADDR 0xdeadc0de
191#define MAX_COPY sizeof(weirdary)
192#define ZERO_LENGTH_PTR ((void *)-8)
193
194/*
195 * Misc global malloc buckets
196 */
197
198MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
199MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
200MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
201
202MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
203MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
204
205/*
206 * Initialize the slab memory allocator. We have to choose a zone size based
207 * on available physical memory. We choose a zone side which is approximately
208 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
209 * 128K. The zone size is limited to the bounds set in slaballoc.h
210 * (typically 32K min, 128K max).
211 */
212static void kmeminit(void *dummy);
213
c7841cbe
MD
214char *ZeroPage;
215
ba39e2e0 216SYSINIT(kmem, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL)
a108bf71 217
d2182dc1
MD
218#ifdef INVARIANTS
219/*
220 * If enabled any memory allocated without M_ZERO is initialized to -1.
221 */
222static int use_malloc_pattern;
223SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
224 &use_malloc_pattern, 0, "");
225#endif
226
665206ee
MD
227SYSCTL_INT(_kern, OID_AUTO, zone_big_alloc, CTLFLAG_RD, &ZoneBigAlloc, 0, "");
228SYSCTL_INT(_kern, OID_AUTO, zone_gen_alloc, CTLFLAG_RD, &ZoneGenAlloc, 0, "");
229
a108bf71
MD
230static void
231kmeminit(void *dummy)
232{
7c457ac8 233 size_t limsize;
a108bf71
MD
234 int usesize;
235 int i;
f9ab53b8 236 vm_offset_t npg;
a108bf71 237
7c457ac8 238 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
c439ad8f
MD
239 if (limsize > KvaSize)
240 limsize = KvaSize;
a108bf71
MD
241
242 usesize = (int)(limsize / 1024); /* convert to KB */
243
244 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
245 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
246 ZoneSize <<= 1;
46a3f46d
MD
247 ZoneLimit = ZoneSize / 4;
248 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
249 ZoneLimit = ZALLOC_ZONE_LIMIT;
5fee07e6 250 ZoneMask = ~(uintptr_t)(ZoneSize - 1);
a108bf71
MD
251 ZonePageCount = ZoneSize / PAGE_SIZE;
252
c439ad8f
MD
253 npg = KvaSize / PAGE_SIZE;
254 kmemusage = kmem_slab_alloc(npg * sizeof(struct kmemusage),
255 PAGE_SIZE, M_WAITOK|M_ZERO);
a108bf71
MD
256
257 for (i = 0; i < arysize(weirdary); ++i)
258 weirdary[i] = WEIRD_ADDR;
259
c7841cbe
MD
260 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
261
a108bf71 262 if (bootverbose)
6ea70f76 263 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
a108bf71
MD
264}
265
266/*
bba6a44d 267 * Initialize a malloc type tracking structure.
a108bf71
MD
268 */
269void
270malloc_init(void *data)
271{
272 struct malloc_type *type = data;
7c457ac8 273 size_t limsize;
a108bf71
MD
274
275 if (type->ks_magic != M_MAGIC)
276 panic("malloc type lacks magic");
277
278 if (type->ks_limit != 0)
279 return;
280
281 if (vmstats.v_page_count == 0)
282 panic("malloc_init not allowed before vm init");
283
7c457ac8 284 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
c439ad8f
MD
285 if (limsize > KvaSize)
286 limsize = KvaSize;
a108bf71
MD
287 type->ks_limit = limsize / 10;
288
289 type->ks_next = kmemstatistics;
290 kmemstatistics = type;
291}
292
293void
294malloc_uninit(void *data)
295{
296 struct malloc_type *type = data;
297 struct malloc_type *t;
bba6a44d
MD
298#ifdef INVARIANTS
299 int i;
1d712609 300 long ttl;
bba6a44d 301#endif
a108bf71
MD
302
303 if (type->ks_magic != M_MAGIC)
304 panic("malloc type lacks magic");
305
306 if (vmstats.v_page_count == 0)
307 panic("malloc_uninit not allowed before vm init");
308
309 if (type->ks_limit == 0)
310 panic("malloc_uninit on uninitialized type");
311
6c92c1f2
SZ
312#ifdef SMP
313 /* Make sure that all pending kfree()s are finished. */
314 lwkt_synchronize_ipiqs("muninit");
315#endif
316
a108bf71 317#ifdef INVARIANTS
1d712609
MD
318 /*
319 * memuse is only correct in aggregation. Due to memory being allocated
320 * on one cpu and freed on another individual array entries may be
321 * negative or positive (canceling each other out).
322 */
323 for (i = ttl = 0; i < ncpus; ++i)
324 ttl += type->ks_memuse[i];
325 if (ttl) {
6ea70f76 326 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
1d712609 327 ttl, type->ks_shortdesc, i);
a108bf71
MD
328 }
329#endif
330 if (type == kmemstatistics) {
331 kmemstatistics = type->ks_next;
332 } else {
333 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
334 if (t->ks_next == type) {
335 t->ks_next = type->ks_next;
336 break;
337 }
338 }
339 }
340 type->ks_next = NULL;
341 type->ks_limit = 0;
342}
343
344/*
40153c65
MD
345 * Increase the kmalloc pool limit for the specified pool. No changes
346 * are the made if the pool would shrink.
347 */
348void
349kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
350{
351 if (type->ks_limit == 0)
352 malloc_init(type);
7c457ac8
MD
353 if (bytes == 0)
354 bytes = KvaSize;
40153c65
MD
355 if (type->ks_limit < bytes)
356 type->ks_limit = bytes;
357}
358
359/*
ebe36cfe
MD
360 * Dynamically create a malloc pool. This function is a NOP if *typep is
361 * already non-NULL.
362 */
363void
364kmalloc_create(struct malloc_type **typep, const char *descr)
365{
366 struct malloc_type *type;
367
368 if (*typep == NULL) {
369 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
370 type->ks_magic = M_MAGIC;
371 type->ks_shortdesc = descr;
372 malloc_init(type);
373 *typep = type;
374 }
375}
376
377/*
378 * Destroy a dynamically created malloc pool. This function is a NOP if
379 * the pool has already been destroyed.
380 */
381void
382kmalloc_destroy(struct malloc_type **typep)
383{
384 if (*typep != NULL) {
385 malloc_uninit(*typep);
386 kfree(*typep, M_TEMP);
387 *typep = NULL;
388 }
389}
390
391/*
a108bf71
MD
392 * Calculate the zone index for the allocation request size and set the
393 * allocation request size to that particular zone's chunk size.
394 */
395static __inline int
396zoneindex(unsigned long *bytes)
397{
398 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
399 if (n < 128) {
400 *bytes = n = (n + 7) & ~7;
401 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
402 }
403 if (n < 256) {
404 *bytes = n = (n + 15) & ~15;
405 return(n / 16 + 7);
406 }
407 if (n < 8192) {
408 if (n < 512) {
409 *bytes = n = (n + 31) & ~31;
410 return(n / 32 + 15);
411 }
412 if (n < 1024) {
413 *bytes = n = (n + 63) & ~63;
414 return(n / 64 + 23);
415 }
416 if (n < 2048) {
417 *bytes = n = (n + 127) & ~127;
418 return(n / 128 + 31);
419 }
420 if (n < 4096) {
421 *bytes = n = (n + 255) & ~255;
422 return(n / 256 + 39);
423 }
424 *bytes = n = (n + 511) & ~511;
425 return(n / 512 + 47);
426 }
427#if ZALLOC_ZONE_LIMIT > 8192
428 if (n < 16384) {
429 *bytes = n = (n + 1023) & ~1023;
430 return(n / 1024 + 55);
431 }
432#endif
433#if ZALLOC_ZONE_LIMIT > 16384
434 if (n < 32768) {
435 *bytes = n = (n + 2047) & ~2047;
436 return(n / 2048 + 63);
437 }
438#endif
439 panic("Unexpected byte count %d", n);
440 return(0);
441}
442
443/*
5fee07e6 444 * kmalloc() (SLAB ALLOCATOR)
a108bf71
MD
445 *
446 * Allocate memory via the slab allocator. If the request is too large,
447 * or if it page-aligned beyond a certain size, we fall back to the
448 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
449 * &SlabMisc if you don't care.
450 *
8cb2bf45
JS
451 * M_RNOWAIT - don't block.
452 * M_NULLOK - return NULL instead of blocking.
a108bf71 453 * M_ZERO - zero the returned memory.
dc1fd4b3
MD
454 * M_USE_RESERVE - allow greater drawdown of the free list
455 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
5b287bba
MD
456 *
457 * MPSAFE
a108bf71
MD
458 */
459void *
8aca2bd4 460kmalloc(unsigned long size, struct malloc_type *type, int flags)
a108bf71
MD
461{
462 SLZone *z;
463 SLChunk *chunk;
d8100bdc 464#ifdef SMP
5fee07e6 465 SLChunk *bchunk;
d8100bdc 466#endif
a108bf71 467 SLGlobalData *slgd;
bba6a44d 468 struct globaldata *gd;
a108bf71 469 int zi;
d2182dc1
MD
470#ifdef INVARIANTS
471 int i;
472#endif
a108bf71 473
b68ad50c 474 logmemory_quick(malloc_beg);
bba6a44d
MD
475 gd = mycpu;
476 slgd = &gd->gd_slab;
a108bf71
MD
477
478 /*
479 * XXX silly to have this in the critical path.
480 */
481 if (type->ks_limit == 0) {
482 crit_enter();
483 if (type->ks_limit == 0)
484 malloc_init(type);
485 crit_exit();
486 }
487 ++type->ks_calls;
488
489 /*
38e34349
MD
490 * Handle the case where the limit is reached. Panic if we can't return
491 * NULL. The original malloc code looped, but this tended to
a108bf71 492 * simply deadlock the computer.
38e34349
MD
493 *
494 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
495 * to determine if a more complete limit check should be done. The
496 * actual memory use is tracked via ks_memuse[cpu].
a108bf71 497 */
bba6a44d
MD
498 while (type->ks_loosememuse >= type->ks_limit) {
499 int i;
500 long ttl;
501
502 for (i = ttl = 0; i < ncpus; ++i)
503 ttl += type->ks_memuse[i];
38e34349 504 type->ks_loosememuse = ttl; /* not MP synchronized */
28135cc2
MD
505 if ((ssize_t)ttl < 0) /* deal with occassional race */
506 ttl = 0;
bba6a44d 507 if (ttl >= type->ks_limit) {
f2b5daf9 508 if (flags & M_NULLOK) {
5fee07e6 509 logmemory(malloc_end, NULL, type, size, flags);
bba6a44d 510 return(NULL);
f2b5daf9 511 }
bba6a44d
MD
512 panic("%s: malloc limit exceeded", type->ks_shortdesc);
513 }
a108bf71
MD
514 }
515
516 /*
517 * Handle the degenerate size == 0 case. Yes, this does happen.
518 * Return a special pointer. This is to maintain compatibility with
519 * the original malloc implementation. Certain devices, such as the
520 * adaptec driver, not only allocate 0 bytes, they check for NULL and
521 * also realloc() later on. Joy.
522 */
f2b5daf9 523 if (size == 0) {
5fee07e6 524 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
a108bf71 525 return(ZERO_LENGTH_PTR);
f2b5daf9 526 }
a108bf71
MD
527
528 /*
a7cf0021
MD
529 * Handle hysteresis from prior frees here in malloc(). We cannot
530 * safely manipulate the kernel_map in free() due to free() possibly
531 * being called via an IPI message or from sensitive interrupt code.
5fee07e6
MD
532 *
533 * NOTE: ku_pagecnt must be cleared before we free the slab or we
534 * might race another cpu allocating the kva and setting
535 * ku_pagecnt.
a7cf0021 536 */
dc1fd4b3 537 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_RNOWAIT) == 0) {
46a3f46d
MD
538 crit_enter();
539 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */
5fee07e6
MD
540 struct kmemusage *kup;
541
46a3f46d
MD
542 z = slgd->FreeZones;
543 slgd->FreeZones = z->z_Next;
544 --slgd->NFreeZones;
5fee07e6
MD
545 kup = btokup(z);
546 kup->ku_pagecnt = 0;
46a3f46d 547 kmem_slab_free(z, ZoneSize); /* may block */
665206ee 548 atomic_add_int(&ZoneGenAlloc, -(int)ZoneSize / 1024);
46a3f46d
MD
549 }
550 crit_exit();
551 }
5fee07e6 552
46a3f46d 553 /*
5fee07e6 554 * XXX handle oversized frees that were queued from kfree().
46a3f46d 555 */
dc1fd4b3 556 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) {
46a3f46d
MD
557 crit_enter();
558 if ((z = slgd->FreeOvZones) != NULL) {
7e83df33
MD
559 vm_size_t tsize;
560
46a3f46d
MD
561 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
562 slgd->FreeOvZones = z->z_Next;
7e83df33
MD
563 tsize = z->z_ChunkSize;
564 kmem_slab_free(z, tsize); /* may block */
565 atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024);
46a3f46d
MD
566 }
567 crit_exit();
a7cf0021
MD
568 }
569
570 /*
a108bf71
MD
571 * Handle large allocations directly. There should not be very many of
572 * these so performance is not a big issue.
573 *
b543eeed
MD
574 * The backend allocator is pretty nasty on a SMP system. Use the
575 * slab allocator for one and two page-sized chunks even though we lose
576 * some efficiency. XXX maybe fix mmio and the elf loader instead.
a108bf71 577 */
b543eeed 578 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
a108bf71
MD
579 struct kmemusage *kup;
580
581 size = round_page(size);
582 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
f2b5daf9 583 if (chunk == NULL) {
5fee07e6 584 logmemory(malloc_end, NULL, type, size, flags);
a108bf71 585 return(NULL);
f2b5daf9 586 }
665206ee 587 atomic_add_int(&ZoneBigAlloc, (int)size / 1024);
a108bf71 588 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
8f1d5415 589 flags |= M_PASSIVE_ZERO;
a108bf71
MD
590 kup = btokup(chunk);
591 kup->ku_pagecnt = size / PAGE_SIZE;
592 crit_enter();
593 goto done;
594 }
595
596 /*
597 * Attempt to allocate out of an existing zone. First try the free list,
598 * then allocate out of unallocated space. If we find a good zone move
599 * it to the head of the list so later allocations find it quickly
600 * (we might have thousands of zones in the list).
601 *
602 * Note: zoneindex() will panic of size is too large.
603 */
604 zi = zoneindex(&size);
605 KKASSERT(zi < NZONES);
606 crit_enter();
a108bf71 607
5fee07e6 608 if ((z = slgd->ZoneAry[zi]) != NULL) {
a108bf71 609 /*
5fee07e6
MD
610 * Locate a chunk - we have to have at least one. If this is the
611 * last chunk go ahead and do the work to retrieve chunks freed
612 * from remote cpus, and if the zone is still empty move it off
613 * the ZoneAry.
a108bf71 614 */
5fee07e6
MD
615 if (--z->z_NFree <= 0) {
616 KKASSERT(z->z_NFree == 0);
617
618#ifdef SMP
619 /*
620 * WARNING! This code competes with other cpus. It is ok
621 * for us to not drain RChunks here but we might as well, and
622 * it is ok if more accumulate after we're done.
623 *
624 * Set RSignal before pulling rchunks off, indicating that we
625 * will be moving ourselves off of the ZoneAry. Remote ends will
626 * read RSignal before putting rchunks on thus interlocking
627 * their IPI signaling.
628 */
629 if (z->z_RChunks == NULL)
630 atomic_swap_int(&z->z_RSignal, 1);
631
632 while ((bchunk = z->z_RChunks) != NULL) {
633 cpu_ccfence();
634 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
635 *z->z_LChunksp = bchunk;
636 while (bchunk) {
637 chunk_mark_free(z, bchunk);
638 z->z_LChunksp = &bchunk->c_Next;
639 bchunk = bchunk->c_Next;
640 ++z->z_NFree;
641 }
642 break;
643 }
644 }
645#endif
646 /*
647 * Remove from the zone list if no free chunks remain.
648 * Clear RSignal
649 */
650 if (z->z_NFree == 0) {
651 slgd->ZoneAry[zi] = z->z_Next;
652 z->z_Next = NULL;
653 } else {
654 z->z_RSignal = 0;
655 }
a108bf71
MD
656 }
657
658 /*
5fee07e6 659 * Fast path, we have chunks available in z_LChunks.
a108bf71 660 */
5fee07e6
MD
661 chunk = z->z_LChunks;
662 if (chunk) {
10cc6608 663 chunk_mark_allocated(z, chunk);
5fee07e6
MD
664 z->z_LChunks = chunk->c_Next;
665 if (z->z_LChunks == NULL)
666 z->z_LChunksp = &z->z_LChunks;
a108bf71 667 goto done;
a108bf71
MD
668 }
669
670 /*
5fee07e6
MD
671 * No chunks are available in LChunks, the free chunk MUST be
672 * in the never-before-used memory area, controlled by UIndex.
673 *
674 * The consequences are very serious if our zone got corrupted so
675 * we use an explicit panic rather than a KASSERT.
a108bf71 676 */
1c5ca4f3 677 if (z->z_UIndex + 1 != z->z_NMax)
5fee07e6 678 ++z->z_UIndex;
1c5ca4f3
MD
679 else
680 z->z_UIndex = 0;
5fee07e6 681
1c5ca4f3
MD
682 if (z->z_UIndex == z->z_UEndIndex)
683 panic("slaballoc: corrupted zone");
5fee07e6 684
1c5ca4f3 685 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
8f1d5415 686 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 687 flags &= ~M_ZERO;
8f1d5415
MD
688 flags |= M_PASSIVE_ZERO;
689 }
10cc6608 690 chunk_mark_allocated(z, chunk);
a108bf71
MD
691 goto done;
692 }
693
694 /*
695 * If all zones are exhausted we need to allocate a new zone for this
696 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
6ab8e1da
MD
697 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
698 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
699 * we do not pre-zero it because we do not want to mess up the L1 cache.
a108bf71
MD
700 *
701 * At least one subsystem, the tty code (see CROUND) expects power-of-2
702 * allocations to be power-of-2 aligned. We maintain compatibility by
703 * adjusting the base offset below.
704 */
705 {
706 int off;
5fee07e6 707 struct kmemusage *kup;
a108bf71
MD
708
709 if ((z = slgd->FreeZones) != NULL) {
710 slgd->FreeZones = z->z_Next;
711 --slgd->NFreeZones;
712 bzero(z, sizeof(SLZone));
6ab8e1da 713 z->z_Flags |= SLZF_UNOTZEROD;
a108bf71
MD
714 } else {
715 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
716 if (z == NULL)
717 goto fail;
665206ee 718 atomic_add_int(&ZoneGenAlloc, (int)ZoneSize / 1024);
a108bf71
MD
719 }
720
721 /*
10cc6608
MD
722 * How big is the base structure?
723 */
724#if defined(INVARIANTS)
725 /*
726 * Make room for z_Bitmap. An exact calculation is somewhat more
727 * complicated so don't make an exact calculation.
728 */
729 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
730 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
731#else
732 off = sizeof(SLZone);
733#endif
734
735 /*
a108bf71
MD
736 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
737 * Otherwise just 8-byte align the data.
738 */
739 if ((size | (size - 1)) + 1 == (size << 1))
10cc6608 740 off = (off + size - 1) & ~(size - 1);
a108bf71 741 else
10cc6608 742 off = (off + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
a108bf71
MD
743 z->z_Magic = ZALLOC_SLAB_MAGIC;
744 z->z_ZoneIndex = zi;
745 z->z_NMax = (ZoneSize - off) / size;
746 z->z_NFree = z->z_NMax - 1;
1c5ca4f3
MD
747 z->z_BasePtr = (char *)z + off;
748 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
a108bf71 749 z->z_ChunkSize = size;
2db3b277 750 z->z_CpuGd = gd;
bba6a44d 751 z->z_Cpu = gd->gd_cpuid;
5fee07e6 752 z->z_LChunksp = &z->z_LChunks;
1c5ca4f3 753 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
a108bf71
MD
754 z->z_Next = slgd->ZoneAry[zi];
755 slgd->ZoneAry[zi] = z;
8f1d5415 756 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 757 flags &= ~M_ZERO; /* already zero'd */
8f1d5415
MD
758 flags |= M_PASSIVE_ZERO;
759 }
5fee07e6
MD
760 kup = btokup(z);
761 kup->ku_pagecnt = -(z->z_Cpu + 1); /* -1 to -(N+1) */
10cc6608 762 chunk_mark_allocated(z, chunk);
1c5ca4f3
MD
763
764 /*
765 * Slide the base index for initial allocations out of the next
766 * zone we create so we do not over-weight the lower part of the
767 * cpu memory caches.
768 */
769 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
770 & (ZALLOC_MAX_ZONE_SIZE - 1);
a108bf71 771 }
5fee07e6 772
a108bf71 773done:
bba6a44d
MD
774 ++type->ks_inuse[gd->gd_cpuid];
775 type->ks_memuse[gd->gd_cpuid] += size;
38e34349 776 type->ks_loosememuse += size; /* not MP synchronized */
a108bf71 777 crit_exit();
5fee07e6 778
a108bf71
MD
779 if (flags & M_ZERO)
780 bzero(chunk, size);
bba6a44d 781#ifdef INVARIANTS
d2182dc1
MD
782 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
783 if (use_malloc_pattern) {
784 for (i = 0; i < size; i += sizeof(int)) {
785 *(int *)((char *)chunk + i) = -1;
786 }
787 }
bba6a44d 788 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
d2182dc1 789 }
bba6a44d 790#endif
5fee07e6 791 logmemory(malloc_end, chunk, type, size, flags);
a108bf71
MD
792 return(chunk);
793fail:
794 crit_exit();
5fee07e6 795 logmemory(malloc_end, NULL, type, size, flags);
a108bf71
MD
796 return(NULL);
797}
798
38e34349
MD
799/*
800 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
801 *
802 * Generally speaking this routine is not called very often and we do
803 * not attempt to optimize it beyond reusing the same pointer if the
804 * new size fits within the chunking of the old pointer's zone.
805 */
a108bf71 806void *
8aca2bd4 807krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
a108bf71 808{
5fee07e6 809 struct kmemusage *kup;
a108bf71
MD
810 SLZone *z;
811 void *nptr;
812 unsigned long osize;
813
eb7f3e3c
MD
814 KKASSERT((flags & M_ZERO) == 0); /* not supported */
815
a108bf71 816 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
efda3bd0 817 return(kmalloc(size, type, flags));
a108bf71 818 if (size == 0) {
efda3bd0 819 kfree(ptr, type);
a108bf71
MD
820 return(NULL);
821 }
822
823 /*
824 * Handle oversized allocations. XXX we really should require that a
825 * size be passed to free() instead of this nonsense.
826 */
5fee07e6
MD
827 kup = btokup(ptr);
828 if (kup->ku_pagecnt > 0) {
829 osize = kup->ku_pagecnt << PAGE_SHIFT;
830 if (osize == round_page(size))
831 return(ptr);
832 if ((nptr = kmalloc(size, type, flags)) == NULL)
833 return(NULL);
834 bcopy(ptr, nptr, min(size, osize));
835 kfree(ptr, type);
836 return(nptr);
a108bf71
MD
837 }
838
839 /*
840 * Get the original allocation's zone. If the new request winds up
841 * using the same chunk size we do not have to do anything.
842 */
5fee07e6
MD
843 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
844 kup = btokup(z);
845 KKASSERT(kup->ku_pagecnt < 0);
a108bf71
MD
846 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
847
a108bf71
MD
848 /*
849 * Allocate memory for the new request size. Note that zoneindex has
850 * already adjusted the request size to the appropriate chunk size, which
851 * should optimize our bcopy(). Then copy and return the new pointer.
1ea6580d
MD
852 *
853 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
854 * necessary align the result.
855 *
856 * We can only zoneindex (to align size to the chunk size) if the new
857 * size is not too large.
a108bf71 858 */
1ea6580d
MD
859 if (size < ZoneLimit) {
860 zoneindex(&size);
861 if (z->z_ChunkSize == size)
862 return(ptr);
863 }
efda3bd0 864 if ((nptr = kmalloc(size, type, flags)) == NULL)
a108bf71
MD
865 return(NULL);
866 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
efda3bd0 867 kfree(ptr, type);
a108bf71
MD
868 return(nptr);
869}
870
38e34349 871/*
45d2b1d8
MD
872 * Return the kmalloc limit for this type, in bytes.
873 */
874long
875kmalloc_limit(struct malloc_type *type)
876{
877 if (type->ks_limit == 0) {
878 crit_enter();
879 if (type->ks_limit == 0)
880 malloc_init(type);
881 crit_exit();
882 }
883 return(type->ks_limit);
884}
885
886/*
38e34349
MD
887 * Allocate a copy of the specified string.
888 *
889 * (MP SAFE) (MAY BLOCK)
890 */
1ac06773 891char *
59302080 892kstrdup(const char *str, struct malloc_type *type)
1ac06773
MD
893{
894 int zlen; /* length inclusive of terminating NUL */
895 char *nstr;
896
897 if (str == NULL)
898 return(NULL);
899 zlen = strlen(str) + 1;
efda3bd0 900 nstr = kmalloc(zlen, type, M_WAITOK);
1ac06773
MD
901 bcopy(str, nstr, zlen);
902 return(nstr);
903}
904
1d712609 905#ifdef SMP
a108bf71 906/*
5fee07e6
MD
907 * Notify our cpu that a remote cpu has freed some chunks in a zone that
908 * we own. Due to MP races we might no longer own the zone, use the
909 * kmemusage array to check.
a108bf71
MD
910 */
911static
912void
5fee07e6 913kfree_remote(void *ptr)
a108bf71 914{
5fee07e6
MD
915 struct kmemusage *kup;
916 SLGlobalData *slgd;
917 SLChunk *bchunk;
918 SLZone *z;
919 int nfree;
920
921 /*
922 * Do not dereference (z) until we validate that its storage is
923 * still around.
924 */
925 slgd = &mycpu->gd_slab;
926 z = ptr;
927 kup = btokup(z);
928
929 if (kup->ku_pagecnt == -((int)mycpuid + 1)) { /* -1 to -(N+1) */
930 logmemory(free_rem_beg, z, NULL, 0, 0);
931 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
932 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
933 nfree = z->z_NFree;
934
935 /*
936 * Indicate that we will no longer be off of the ZoneAry by
937 * clearing RSignal.
938 */
939 if (z->z_RChunks)
940 z->z_RSignal = 0;
941
942 /*
943 * Atomically extract the bchunks list and then process it back
944 * into the lchunks list. We want to append our bchunks to the
945 * lchunks list and not prepend since we likely do not have
946 * cache mastership of the related data (not that it helps since
947 * we are using c_Next).
948 */
949 while ((bchunk = z->z_RChunks) != NULL) {
950 cpu_ccfence();
951 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
952 *z->z_LChunksp = bchunk;
953 while (bchunk) {
954 chunk_mark_free(z, bchunk);
955 z->z_LChunksp = &bchunk->c_Next;
956 bchunk = bchunk->c_Next;
957 ++z->z_NFree;
958 }
959 break;
960 }
961 }
962 if (z->z_NFree && nfree == 0) {
963 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
964 slgd->ZoneAry[z->z_ZoneIndex] = z;
965 }
966
967 /*
968 * If the zone becomes totally free, and there are other zones we
969 * can allocate from, move this zone to the FreeZones list. Since
970 * this code can be called from an IPI callback, do *NOT* try to mess
971 * with kernel_map here. Hysteresis will be performed at malloc() time.
972 */
973 if (z->z_NFree == z->z_NMax &&
974 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z)
975 ) {
976 struct kmemusage *kup;
977 SLZone **pz;
978
979 for (pz = &slgd->ZoneAry[z->z_ZoneIndex];
980 z != *pz;
981 pz = &(*pz)->z_Next) {
982 ;
983 }
984 *pz = z->z_Next;
985 z->z_Magic = -1;
986 z->z_Next = slgd->FreeZones;
987 slgd->FreeZones = z;
988 ++slgd->NFreeZones;
989 kup = btokup(z);
990 kup->ku_pagecnt = 0;
991 }
992 logmemory(free_rem_end, z, bchunk, 0, 0);
993 }
a108bf71
MD
994}
995
1d712609
MD
996#endif
997
38e34349 998/*
5b287bba 999 * free (SLAB ALLOCATOR)
38e34349
MD
1000 *
1001 * Free a memory block previously allocated by malloc. Note that we do not
5fee07e6 1002 * attempt to update ks_loosememuse as MP races could prevent us from
38e34349 1003 * checking memory limits in malloc.
5b287bba
MD
1004 *
1005 * MPSAFE
38e34349 1006 */
a108bf71 1007void
8aca2bd4 1008kfree(void *ptr, struct malloc_type *type)
a108bf71
MD
1009{
1010 SLZone *z;
1011 SLChunk *chunk;
1012 SLGlobalData *slgd;
bba6a44d 1013 struct globaldata *gd;
5fee07e6
MD
1014 struct kmemusage *kup;
1015 unsigned long size;
d8100bdc
SW
1016#ifdef SMP
1017 SLChunk *bchunk;
5fee07e6 1018 int rsignal;
d8100bdc 1019#endif
a108bf71 1020
b68ad50c 1021 logmemory_quick(free_beg);
bba6a44d
MD
1022 gd = mycpu;
1023 slgd = &gd->gd_slab;
a108bf71 1024
d39911d9
JS
1025 if (ptr == NULL)
1026 panic("trying to free NULL pointer");
1027
a108bf71
MD
1028 /*
1029 * Handle special 0-byte allocations
1030 */
f2b5daf9
MD
1031 if (ptr == ZERO_LENGTH_PTR) {
1032 logmemory(free_zero, ptr, type, -1, 0);
b68ad50c 1033 logmemory_quick(free_end);
a108bf71 1034 return;
f2b5daf9 1035 }
a108bf71
MD
1036
1037 /*
5fee07e6
MD
1038 * Panic on bad malloc type
1039 */
1040 if (type->ks_magic != M_MAGIC)
1041 panic("free: malloc type lacks magic");
1042
1043 /*
a108bf71
MD
1044 * Handle oversized allocations. XXX we really should require that a
1045 * size be passed to free() instead of this nonsense.
bba6a44d
MD
1046 *
1047 * This code is never called via an ipi.
a108bf71 1048 */
5fee07e6
MD
1049 kup = btokup(ptr);
1050 if (kup->ku_pagecnt > 0) {
1051 size = kup->ku_pagecnt << PAGE_SHIFT;
1052 kup->ku_pagecnt = 0;
a108bf71 1053#ifdef INVARIANTS
5fee07e6
MD
1054 KKASSERT(sizeof(weirdary) <= size);
1055 bcopy(weirdary, ptr, sizeof(weirdary));
a108bf71 1056#endif
5fee07e6
MD
1057 /*
1058 * NOTE: For oversized allocations we do not record the
1059 * originating cpu. It gets freed on the cpu calling
1060 * kfree(). The statistics are in aggregate.
1061 *
1062 * note: XXX we have still inherited the interrupts-can't-block
1063 * assumption. An interrupt thread does not bump
1064 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1065 * primarily until we can fix softupdate's assumptions about free().
1066 */
1067 crit_enter();
1068 --type->ks_inuse[gd->gd_cpuid];
1069 type->ks_memuse[gd->gd_cpuid] -= size;
1070 if (mycpu->gd_intr_nesting_level ||
1071 (gd->gd_curthread->td_flags & TDF_INTTHREAD))
1072 {
1073 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1074 z = (SLZone *)ptr;
1075 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1076 z->z_Next = slgd->FreeOvZones;
1077 z->z_ChunkSize = size;
1078 slgd->FreeOvZones = z;
1079 crit_exit();
1080 } else {
1081 crit_exit();
1082 logmemory(free_ovsz, ptr, type, size, 0);
1083 kmem_slab_free(ptr, size); /* may block */
1084 atomic_add_int(&ZoneBigAlloc, -(int)size / 1024);
a108bf71 1085 }
5fee07e6
MD
1086 logmemory_quick(free_end);
1087 return;
a108bf71
MD
1088 }
1089
1090 /*
1091 * Zone case. Figure out the zone based on the fact that it is
1092 * ZoneSize aligned.
1093 */
5fee07e6
MD
1094 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1095 kup = btokup(z);
1096 KKASSERT(kup->ku_pagecnt < 0);
a108bf71
MD
1097 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1098
1099 /*
5fee07e6
MD
1100 * If we do not own the zone then use atomic ops to free to the
1101 * remote cpu linked list and notify the target zone using a
1102 * passive message.
1103 *
1104 * The target zone cannot be deallocated while we own a chunk of it,
1105 * so the zone header's storage is stable until the very moment
1106 * we adjust z_RChunks. After that we cannot safely dereference (z).
1107 *
1108 * (no critical section needed)
a108bf71 1109 */
2db3b277 1110 if (z->z_CpuGd != gd) {
75c7ffea 1111#ifdef SMP
5fee07e6
MD
1112 /*
1113 * Making these adjustments now allow us to avoid passing (type)
1114 * to the remote cpu. Note that ks_inuse/ks_memuse is being
28135cc2
MD
1115 * adjusted on OUR cpu, not the zone cpu, but it should all still
1116 * sum up properly and cancel out.
5fee07e6 1117 */
28135cc2
MD
1118 crit_enter();
1119 --type->ks_inuse[gd->gd_cpuid];
1120 type->ks_memuse[gd->gd_cpuid] -= z->z_ChunkSize;
1121 crit_exit();
5fee07e6
MD
1122
1123 /*
1124 * WARNING! This code competes with other cpus. Once we
1125 * successfully link the chunk to RChunks the remote
1126 * cpu can rip z's storage out from under us.
1127 */
1128 rsignal = z->z_RSignal;
1129 cpu_lfence();
1130
1131 chunk = ptr;
1132 for (;;) {
1133 bchunk = z->z_RChunks;
1134 cpu_ccfence();
1135 chunk->c_Next = bchunk;
1136 cpu_sfence();
1137
1138 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1139 break;
1140 }
1141 /* z cannot be dereferenced now */
1142
1143 /*
1144 * We have to signal the remote cpu if our actions will cause
1145 * the remote zone to be placed back on ZoneAry so it can
1146 * move the zone back on.
1147 *
1148 * We only need to deal with NULL->non-NULL RChunk transitions
1149 * and only if z_RSignal is set. We interlock by reading rsignal
1150 * before adding our chunk to RChunks. This should result in
1151 * virtually no IPI traffic.
1152 *
1153 * We can use a passive IPI to reduce overhead even further.
1154 */
1155 if (bchunk == NULL && rsignal) {
1156 logmemory(free_request, ptr, type, z->z_ChunkSize, 0);
1157 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
1158 }
75c7ffea
MD
1159#else
1160 panic("Corrupt SLZone");
1161#endif
b68ad50c 1162 logmemory_quick(free_end);
a108bf71
MD
1163 return;
1164 }
1165
5fee07e6
MD
1166 /*
1167 * kfree locally
1168 */
f2b5daf9
MD
1169 logmemory(free_chunk, ptr, type, z->z_ChunkSize, 0);
1170
a108bf71 1171 crit_enter();
a108bf71 1172 chunk = ptr;
10cc6608 1173 chunk_mark_free(z, chunk);
a108bf71
MD
1174
1175 /*
1176 * Put weird data into the memory to detect modifications after freeing,
1177 * illegal pointer use after freeing (we should fault on the odd address),
1178 * and so forth. XXX needs more work, see the old malloc code.
1179 */
1180#ifdef INVARIANTS
1181 if (z->z_ChunkSize < sizeof(weirdary))
1182 bcopy(weirdary, chunk, z->z_ChunkSize);
1183 else
1184 bcopy(weirdary, chunk, sizeof(weirdary));
1185#endif
1186
1187 /*
5fee07e6
MD
1188 * Add this free non-zero'd chunk to a linked list for reuse. Add
1189 * to the front of the linked list so it is more likely to be
1190 * reallocated, since it is already in our L1 cache.
a108bf71 1191 */
6ab8e1da 1192#ifdef INVARIANTS
c439ad8f 1193 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
fc92d4aa 1194 panic("BADFREE %p", chunk);
a108bf71 1195#endif
5fee07e6
MD
1196 chunk->c_Next = z->z_LChunks;
1197 z->z_LChunks = chunk;
1198 if (chunk->c_Next == NULL)
1199 z->z_LChunksp = &chunk->c_Next;
1200
6ab8e1da 1201#ifdef INVARIANTS
c439ad8f 1202 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
a108bf71 1203 panic("BADFREE2");
6ab8e1da 1204#endif
a108bf71
MD
1205
1206 /*
1207 * Bump the number of free chunks. If it becomes non-zero the zone
1208 * must be added back onto the appropriate list.
1209 */
1210 if (z->z_NFree++ == 0) {
1211 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1212 slgd->ZoneAry[z->z_ZoneIndex] = z;
1213 }
1214
bba6a44d
MD
1215 --type->ks_inuse[z->z_Cpu];
1216 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
a108bf71
MD
1217
1218 /*
1219 * If the zone becomes totally free, and there are other zones we
a7cf0021
MD
1220 * can allocate from, move this zone to the FreeZones list. Since
1221 * this code can be called from an IPI callback, do *NOT* try to mess
1222 * with kernel_map here. Hysteresis will be performed at malloc() time.
a108bf71
MD
1223 */
1224 if (z->z_NFree == z->z_NMax &&
1225 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z)
1226 ) {
1227 SLZone **pz;
5fee07e6 1228 struct kmemusage *kup;
a108bf71
MD
1229
1230 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
1231 ;
1232 *pz = z->z_Next;
1233 z->z_Magic = -1;
a7cf0021
MD
1234 z->z_Next = slgd->FreeZones;
1235 slgd->FreeZones = z;
1236 ++slgd->NFreeZones;
5fee07e6
MD
1237 kup = btokup(z);
1238 kup->ku_pagecnt = 0;
a108bf71 1239 }
b68ad50c 1240 logmemory_quick(free_end);
a108bf71
MD
1241 crit_exit();
1242}
1243
10cc6608 1244#if defined(INVARIANTS)
5fee07e6 1245
10cc6608
MD
1246/*
1247 * Helper routines for sanity checks
1248 */
1249static
1250void
1251chunk_mark_allocated(SLZone *z, void *chunk)
1252{
1253 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1254 __uint32_t *bitptr;
1255
5fee07e6
MD
1256 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1257 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1258 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
10cc6608
MD
1259 bitptr = &z->z_Bitmap[bitdex >> 5];
1260 bitdex &= 31;
5fee07e6
MD
1261 KASSERT((*bitptr & (1 << bitdex)) == 0,
1262 ("memory chunk %p is already allocated!", chunk));
10cc6608
MD
1263 *bitptr |= 1 << bitdex;
1264}
1265
1266static
1267void
1268chunk_mark_free(SLZone *z, void *chunk)
1269{
1270 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1271 __uint32_t *bitptr;
1272
5fee07e6
MD
1273 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1274 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1275 ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
10cc6608
MD
1276 bitptr = &z->z_Bitmap[bitdex >> 5];
1277 bitdex &= 31;
5fee07e6
MD
1278 KASSERT((*bitptr & (1 << bitdex)) != 0,
1279 ("memory chunk %p is already free!", chunk));
10cc6608
MD
1280 *bitptr &= ~(1 << bitdex);
1281}
1282
1283#endif
1284
a108bf71 1285/*
5b287bba 1286 * kmem_slab_alloc()
a108bf71
MD
1287 *
1288 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1289 * specified alignment. M_* flags are expected in the flags field.
1290 *
1291 * Alignment must be a multiple of PAGE_SIZE.
1292 *
1293 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1294 * but when we move zalloc() over to use this function as its backend
1295 * we will have to switch to kreserve/krelease and call reserve(0)
1296 * after the new space is made available.
dc1fd4b3
MD
1297 *
1298 * Interrupt code which has preempted other code is not allowed to
c397c465
MD
1299 * use PQ_CACHE pages. However, if an interrupt thread is run
1300 * non-preemptively or blocks and then runs non-preemptively, then
1301 * it is free to use PQ_CACHE pages.
a108bf71
MD
1302 */
1303static void *
1304kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1305{
1306 vm_size_t i;
1307 vm_offset_t addr;
1de1e800 1308 int count, vmflags, base_vmflags;
dc1fd4b3 1309 thread_t td;
a108bf71
MD
1310
1311 size = round_page(size);
e4846942 1312 addr = vm_map_min(&kernel_map);
a108bf71
MD
1313
1314 /*
5c39c498
MD
1315 * Reserve properly aligned space from kernel_map. RNOWAIT allocations
1316 * cannot block.
a108bf71 1317 */
5c39c498 1318 if (flags & M_RNOWAIT) {
ed2013d8 1319 if (lwkt_trytoken(&vm_token) == 0)
5c39c498
MD
1320 return(NULL);
1321 } else {
ed2013d8 1322 lwkt_gettoken(&vm_token);
5c39c498 1323 }
a108bf71
MD
1324 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1325 crit_enter();
e4846942 1326 vm_map_lock(&kernel_map);
c809941b 1327 if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) {
e4846942 1328 vm_map_unlock(&kernel_map);
8cb2bf45 1329 if ((flags & M_NULLOK) == 0)
a108bf71 1330 panic("kmem_slab_alloc(): kernel_map ran out of space!");
a108bf71 1331 vm_map_entry_release(count);
2de4f77e 1332 crit_exit();
ed2013d8 1333 lwkt_reltoken(&vm_token);
a108bf71
MD
1334 return(NULL);
1335 }
e4846942
MD
1336
1337 /*
1338 * kernel_object maps 1:1 to kernel_map.
1339 */
c439ad8f 1340 vm_object_reference(&kernel_object);
e4846942
MD
1341 vm_map_insert(&kernel_map, &count,
1342 &kernel_object, addr, addr, addr + size,
1b874851
MD
1343 VM_MAPTYPE_NORMAL,
1344 VM_PROT_ALL, VM_PROT_ALL,
1345 0);
a108bf71 1346
dc1fd4b3 1347 td = curthread;
dc1fd4b3 1348
1de1e800
JS
1349 base_vmflags = 0;
1350 if (flags & M_ZERO)
1351 base_vmflags |= VM_ALLOC_ZERO;
1352 if (flags & M_USE_RESERVE)
1353 base_vmflags |= VM_ALLOC_SYSTEM;
1354 if (flags & M_USE_INTERRUPT_RESERVE)
1355 base_vmflags |= VM_ALLOC_INTERRUPT;
77912481
MD
1356 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1357 panic("kmem_slab_alloc: bad flags %08x (%p)",
1358 flags, ((int **)&size)[-1]);
1359 }
1de1e800
JS
1360
1361
a108bf71
MD
1362 /*
1363 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
1364 */
1365 for (i = 0; i < size; i += PAGE_SIZE) {
1366 vm_page_t m;
fe1e98d0
MD
1367
1368 /*
c397c465
MD
1369 * VM_ALLOC_NORMAL can only be set if we are not preempting.
1370 *
1371 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1372 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
4ecf7cc9
MD
1373 * implied in this case), though I'm not sure if we really need to
1374 * do that.
fe1e98d0 1375 */
1de1e800 1376 vmflags = base_vmflags;
c397c465 1377 if (flags & M_WAITOK) {
1de1e800 1378 if (td->td_preempted)
fe1e98d0 1379 vmflags |= VM_ALLOC_SYSTEM;
1de1e800 1380 else
dc1fd4b3 1381 vmflags |= VM_ALLOC_NORMAL;
dc1fd4b3 1382 }
a108bf71 1383
e4846942 1384 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
dc1fd4b3
MD
1385
1386 /*
1387 * If the allocation failed we either return NULL or we retry.
1388 *
c397c465
MD
1389 * If M_WAITOK is specified we wait for more memory and retry.
1390 * If M_WAITOK is specified from a preemption we yield instead of
1391 * wait. Livelock will not occur because the interrupt thread
1392 * will not be preempting anyone the second time around after the
1393 * yield.
dc1fd4b3 1394 */
a108bf71 1395 if (m == NULL) {
c397c465 1396 if (flags & M_WAITOK) {
fe1e98d0 1397 if (td->td_preempted) {
e4846942 1398 vm_map_unlock(&kernel_map);
77912481 1399 lwkt_switch();
e4846942 1400 vm_map_lock(&kernel_map);
dc1fd4b3 1401 } else {
e4846942 1402 vm_map_unlock(&kernel_map);
4ecf7cc9 1403 vm_wait(0);
e4846942 1404 vm_map_lock(&kernel_map);
dc1fd4b3 1405 }
a108bf71
MD
1406 i -= PAGE_SIZE; /* retry */
1407 continue;
1408 }
dc1fd4b3
MD
1409
1410 /*
1411 * We were unable to recover, cleanup and return NULL
2de4f77e
MD
1412 *
1413 * (vm_token already held)
dc1fd4b3 1414 */
a108bf71
MD
1415 while (i != 0) {
1416 i -= PAGE_SIZE;
e4846942 1417 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
17cde63e 1418 /* page should already be busy */
a108bf71
MD
1419 vm_page_free(m);
1420 }
e4846942
MD
1421 vm_map_delete(&kernel_map, addr, addr + size, &count);
1422 vm_map_unlock(&kernel_map);
a108bf71 1423 vm_map_entry_release(count);
2de4f77e 1424 crit_exit();
ed2013d8 1425 lwkt_reltoken(&vm_token);
a108bf71
MD
1426 return(NULL);
1427 }
1428 }
1429
1430 /*
dc1fd4b3
MD
1431 * Success!
1432 *
a108bf71
MD
1433 * Mark the map entry as non-pageable using a routine that allows us to
1434 * populate the underlying pages.
17cde63e
MD
1435 *
1436 * The pages were busied by the allocations above.
a108bf71 1437 */
e4846942 1438 vm_map_set_wired_quick(&kernel_map, addr, size, &count);
a108bf71
MD
1439 crit_exit();
1440
1441 /*
1442 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1443 */
77912481 1444 lwkt_gettoken(&vm_token);
a108bf71
MD
1445 for (i = 0; i < size; i += PAGE_SIZE) {
1446 vm_page_t m;
1447
e4846942 1448 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
a108bf71 1449 m->valid = VM_PAGE_BITS_ALL;
17cde63e 1450 /* page should already be busy */
a108bf71
MD
1451 vm_page_wire(m);
1452 vm_page_wakeup(m);
fbbaeba3 1453 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
a108bf71
MD
1454 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1455 bzero((char *)addr + i, PAGE_SIZE);
1456 vm_page_flag_clear(m, PG_ZERO);
17cde63e
MD
1457 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1458 vm_page_flag_set(m, PG_REFERENCED);
a108bf71 1459 }
77912481 1460 lwkt_reltoken(&vm_token);
e4846942 1461 vm_map_unlock(&kernel_map);
a108bf71 1462 vm_map_entry_release(count);
ed2013d8 1463 lwkt_reltoken(&vm_token);
a108bf71
MD
1464 return((void *)addr);
1465}
1466
38e34349 1467/*
5b287bba 1468 * kmem_slab_free()
38e34349 1469 */
a108bf71
MD
1470static void
1471kmem_slab_free(void *ptr, vm_size_t size)
1472{
1473 crit_enter();
ed2013d8 1474 lwkt_gettoken(&vm_token);
e4846942 1475 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
ed2013d8 1476 lwkt_reltoken(&vm_token);
a108bf71
MD
1477 crit_exit();
1478}
1479