kernel - Provide descriptions for lwkt.* and debug.* sysctl's
[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
120#define arysize(ary) (sizeof(ary)/sizeof((ary)[0]))
121
722871d3
MD
122#define btokup(z) (&pmap_kvtom((vm_offset_t)(z))->ku_pagecnt)
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 int32_t weirdary[16];
162
163static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
164static void kmem_slab_free(void *ptr, vm_size_t bytes);
5fee07e6 165
10cc6608
MD
166#if defined(INVARIANTS)
167static void chunk_mark_allocated(SLZone *z, void *chunk);
168static void chunk_mark_free(SLZone *z, void *chunk);
5fee07e6
MD
169#else
170#define chunk_mark_allocated(z, chunk)
171#define chunk_mark_free(z, chunk)
10cc6608 172#endif
a108bf71
MD
173
174/*
175 * Misc constants. Note that allocations that are exact multiples of
176 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
177 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
178 */
179#define MIN_CHUNK_SIZE 8 /* in bytes */
180#define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
181#define ZONE_RELS_THRESH 2 /* threshold number of zones */
182#define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
183
184/*
185 * The WEIRD_ADDR is used as known text to copy into free objects to
186 * try to create deterministic failure cases if the data is accessed after
187 * free.
188 */
189#define WEIRD_ADDR 0xdeadc0de
190#define MAX_COPY sizeof(weirdary)
191#define ZERO_LENGTH_PTR ((void *)-8)
192
193/*
194 * Misc global malloc buckets
195 */
196
197MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
198MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
199MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
200
201MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
202MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
203
204/*
205 * Initialize the slab memory allocator. We have to choose a zone size based
206 * on available physical memory. We choose a zone side which is approximately
207 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
208 * 128K. The zone size is limited to the bounds set in slaballoc.h
209 * (typically 32K min, 128K max).
210 */
211static void kmeminit(void *dummy);
212
c7841cbe
MD
213char *ZeroPage;
214
ba39e2e0 215SYSINIT(kmem, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL)
a108bf71 216
d2182dc1
MD
217#ifdef INVARIANTS
218/*
219 * If enabled any memory allocated without M_ZERO is initialized to -1.
220 */
221static int use_malloc_pattern;
222SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
0c52fa62
SG
223 &use_malloc_pattern, 0,
224 "Initialize memory to -1 if M_ZERO not specified");
d2182dc1
MD
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;
a108bf71 236
7c457ac8 237 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
c439ad8f
MD
238 if (limsize > KvaSize)
239 limsize = KvaSize;
a108bf71
MD
240
241 usesize = (int)(limsize / 1024); /* convert to KB */
242
243 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
244 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
245 ZoneSize <<= 1;
46a3f46d
MD
246 ZoneLimit = ZoneSize / 4;
247 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
248 ZoneLimit = ZALLOC_ZONE_LIMIT;
5fee07e6 249 ZoneMask = ~(uintptr_t)(ZoneSize - 1);
a108bf71
MD
250 ZonePageCount = ZoneSize / PAGE_SIZE;
251
a108bf71
MD
252 for (i = 0; i < arysize(weirdary); ++i)
253 weirdary[i] = WEIRD_ADDR;
254
c7841cbe
MD
255 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
256
a108bf71 257 if (bootverbose)
6ea70f76 258 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
a108bf71
MD
259}
260
261/*
bba6a44d 262 * Initialize a malloc type tracking structure.
a108bf71
MD
263 */
264void
265malloc_init(void *data)
266{
267 struct malloc_type *type = data;
7c457ac8 268 size_t limsize;
a108bf71
MD
269
270 if (type->ks_magic != M_MAGIC)
271 panic("malloc type lacks magic");
272
273 if (type->ks_limit != 0)
274 return;
275
276 if (vmstats.v_page_count == 0)
277 panic("malloc_init not allowed before vm init");
278
7c457ac8 279 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
c439ad8f
MD
280 if (limsize > KvaSize)
281 limsize = KvaSize;
a108bf71
MD
282 type->ks_limit = limsize / 10;
283
284 type->ks_next = kmemstatistics;
285 kmemstatistics = type;
286}
287
288void
289malloc_uninit(void *data)
290{
291 struct malloc_type *type = data;
292 struct malloc_type *t;
bba6a44d
MD
293#ifdef INVARIANTS
294 int i;
1d712609 295 long ttl;
bba6a44d 296#endif
a108bf71
MD
297
298 if (type->ks_magic != M_MAGIC)
299 panic("malloc type lacks magic");
300
301 if (vmstats.v_page_count == 0)
302 panic("malloc_uninit not allowed before vm init");
303
304 if (type->ks_limit == 0)
305 panic("malloc_uninit on uninitialized type");
306
6c92c1f2
SZ
307#ifdef SMP
308 /* Make sure that all pending kfree()s are finished. */
309 lwkt_synchronize_ipiqs("muninit");
310#endif
311
a108bf71 312#ifdef INVARIANTS
1d712609
MD
313 /*
314 * memuse is only correct in aggregation. Due to memory being allocated
315 * on one cpu and freed on another individual array entries may be
316 * negative or positive (canceling each other out).
317 */
318 for (i = ttl = 0; i < ncpus; ++i)
319 ttl += type->ks_memuse[i];
320 if (ttl) {
6ea70f76 321 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
1d712609 322 ttl, type->ks_shortdesc, i);
a108bf71
MD
323 }
324#endif
325 if (type == kmemstatistics) {
326 kmemstatistics = type->ks_next;
327 } else {
328 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
329 if (t->ks_next == type) {
330 t->ks_next = type->ks_next;
331 break;
332 }
333 }
334 }
335 type->ks_next = NULL;
336 type->ks_limit = 0;
337}
338
339/*
40153c65
MD
340 * Increase the kmalloc pool limit for the specified pool. No changes
341 * are the made if the pool would shrink.
342 */
343void
344kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
345{
346 if (type->ks_limit == 0)
347 malloc_init(type);
7c457ac8
MD
348 if (bytes == 0)
349 bytes = KvaSize;
40153c65
MD
350 if (type->ks_limit < bytes)
351 type->ks_limit = bytes;
352}
353
354/*
ebe36cfe
MD
355 * Dynamically create a malloc pool. This function is a NOP if *typep is
356 * already non-NULL.
357 */
358void
359kmalloc_create(struct malloc_type **typep, const char *descr)
360{
361 struct malloc_type *type;
362
363 if (*typep == NULL) {
364 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
365 type->ks_magic = M_MAGIC;
366 type->ks_shortdesc = descr;
367 malloc_init(type);
368 *typep = type;
369 }
370}
371
372/*
373 * Destroy a dynamically created malloc pool. This function is a NOP if
374 * the pool has already been destroyed.
375 */
376void
377kmalloc_destroy(struct malloc_type **typep)
378{
379 if (*typep != NULL) {
380 malloc_uninit(*typep);
381 kfree(*typep, M_TEMP);
382 *typep = NULL;
383 }
384}
385
386/*
a108bf71
MD
387 * Calculate the zone index for the allocation request size and set the
388 * allocation request size to that particular zone's chunk size.
389 */
390static __inline int
391zoneindex(unsigned long *bytes)
392{
393 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
394 if (n < 128) {
395 *bytes = n = (n + 7) & ~7;
396 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
397 }
398 if (n < 256) {
399 *bytes = n = (n + 15) & ~15;
400 return(n / 16 + 7);
401 }
402 if (n < 8192) {
403 if (n < 512) {
404 *bytes = n = (n + 31) & ~31;
405 return(n / 32 + 15);
406 }
407 if (n < 1024) {
408 *bytes = n = (n + 63) & ~63;
409 return(n / 64 + 23);
410 }
411 if (n < 2048) {
412 *bytes = n = (n + 127) & ~127;
413 return(n / 128 + 31);
414 }
415 if (n < 4096) {
416 *bytes = n = (n + 255) & ~255;
417 return(n / 256 + 39);
418 }
419 *bytes = n = (n + 511) & ~511;
420 return(n / 512 + 47);
421 }
422#if ZALLOC_ZONE_LIMIT > 8192
423 if (n < 16384) {
424 *bytes = n = (n + 1023) & ~1023;
425 return(n / 1024 + 55);
426 }
427#endif
428#if ZALLOC_ZONE_LIMIT > 16384
429 if (n < 32768) {
430 *bytes = n = (n + 2047) & ~2047;
431 return(n / 2048 + 63);
432 }
433#endif
434 panic("Unexpected byte count %d", n);
435 return(0);
436}
437
438/*
5fee07e6 439 * kmalloc() (SLAB ALLOCATOR)
a108bf71
MD
440 *
441 * Allocate memory via the slab allocator. If the request is too large,
442 * or if it page-aligned beyond a certain size, we fall back to the
443 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
444 * &SlabMisc if you don't care.
445 *
8cb2bf45
JS
446 * M_RNOWAIT - don't block.
447 * M_NULLOK - return NULL instead of blocking.
a108bf71 448 * M_ZERO - zero the returned memory.
dc1fd4b3
MD
449 * M_USE_RESERVE - allow greater drawdown of the free list
450 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
5b287bba
MD
451 *
452 * MPSAFE
a108bf71
MD
453 */
454void *
8aca2bd4 455kmalloc(unsigned long size, struct malloc_type *type, int flags)
a108bf71
MD
456{
457 SLZone *z;
458 SLChunk *chunk;
d8100bdc 459#ifdef SMP
5fee07e6 460 SLChunk *bchunk;
d8100bdc 461#endif
a108bf71 462 SLGlobalData *slgd;
bba6a44d 463 struct globaldata *gd;
a108bf71 464 int zi;
d2182dc1
MD
465#ifdef INVARIANTS
466 int i;
467#endif
a108bf71 468
b68ad50c 469 logmemory_quick(malloc_beg);
bba6a44d
MD
470 gd = mycpu;
471 slgd = &gd->gd_slab;
a108bf71
MD
472
473 /*
474 * XXX silly to have this in the critical path.
475 */
476 if (type->ks_limit == 0) {
477 crit_enter();
478 if (type->ks_limit == 0)
479 malloc_init(type);
480 crit_exit();
481 }
482 ++type->ks_calls;
483
484 /*
38e34349
MD
485 * Handle the case where the limit is reached. Panic if we can't return
486 * NULL. The original malloc code looped, but this tended to
a108bf71 487 * simply deadlock the computer.
38e34349
MD
488 *
489 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
490 * to determine if a more complete limit check should be done. The
491 * actual memory use is tracked via ks_memuse[cpu].
a108bf71 492 */
bba6a44d
MD
493 while (type->ks_loosememuse >= type->ks_limit) {
494 int i;
495 long ttl;
496
497 for (i = ttl = 0; i < ncpus; ++i)
498 ttl += type->ks_memuse[i];
38e34349 499 type->ks_loosememuse = ttl; /* not MP synchronized */
28135cc2
MD
500 if ((ssize_t)ttl < 0) /* deal with occassional race */
501 ttl = 0;
bba6a44d 502 if (ttl >= type->ks_limit) {
f2b5daf9 503 if (flags & M_NULLOK) {
5fee07e6 504 logmemory(malloc_end, NULL, type, size, flags);
bba6a44d 505 return(NULL);
f2b5daf9 506 }
bba6a44d
MD
507 panic("%s: malloc limit exceeded", type->ks_shortdesc);
508 }
a108bf71
MD
509 }
510
511 /*
512 * Handle the degenerate size == 0 case. Yes, this does happen.
513 * Return a special pointer. This is to maintain compatibility with
514 * the original malloc implementation. Certain devices, such as the
515 * adaptec driver, not only allocate 0 bytes, they check for NULL and
516 * also realloc() later on. Joy.
517 */
f2b5daf9 518 if (size == 0) {
5fee07e6 519 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
a108bf71 520 return(ZERO_LENGTH_PTR);
f2b5daf9 521 }
a108bf71
MD
522
523 /*
a7cf0021
MD
524 * Handle hysteresis from prior frees here in malloc(). We cannot
525 * safely manipulate the kernel_map in free() due to free() possibly
526 * being called via an IPI message or from sensitive interrupt code.
5fee07e6
MD
527 *
528 * NOTE: ku_pagecnt must be cleared before we free the slab or we
529 * might race another cpu allocating the kva and setting
530 * ku_pagecnt.
a7cf0021 531 */
dc1fd4b3 532 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_RNOWAIT) == 0) {
46a3f46d
MD
533 crit_enter();
534 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */
722871d3 535 int *kup;
5fee07e6 536
46a3f46d
MD
537 z = slgd->FreeZones;
538 slgd->FreeZones = z->z_Next;
539 --slgd->NFreeZones;
5fee07e6 540 kup = btokup(z);
722871d3 541 *kup = 0;
46a3f46d 542 kmem_slab_free(z, ZoneSize); /* may block */
665206ee 543 atomic_add_int(&ZoneGenAlloc, -(int)ZoneSize / 1024);
46a3f46d
MD
544 }
545 crit_exit();
546 }
5fee07e6 547
46a3f46d 548 /*
5fee07e6 549 * XXX handle oversized frees that were queued from kfree().
46a3f46d 550 */
dc1fd4b3 551 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) {
46a3f46d
MD
552 crit_enter();
553 if ((z = slgd->FreeOvZones) != NULL) {
7e83df33
MD
554 vm_size_t tsize;
555
46a3f46d
MD
556 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
557 slgd->FreeOvZones = z->z_Next;
7e83df33
MD
558 tsize = z->z_ChunkSize;
559 kmem_slab_free(z, tsize); /* may block */
560 atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024);
46a3f46d
MD
561 }
562 crit_exit();
a7cf0021
MD
563 }
564
565 /*
a108bf71
MD
566 * Handle large allocations directly. There should not be very many of
567 * these so performance is not a big issue.
568 *
b543eeed
MD
569 * The backend allocator is pretty nasty on a SMP system. Use the
570 * slab allocator for one and two page-sized chunks even though we lose
571 * some efficiency. XXX maybe fix mmio and the elf loader instead.
a108bf71 572 */
b543eeed 573 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
722871d3 574 int *kup;
a108bf71
MD
575
576 size = round_page(size);
577 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
f2b5daf9 578 if (chunk == NULL) {
5fee07e6 579 logmemory(malloc_end, NULL, type, size, flags);
a108bf71 580 return(NULL);
f2b5daf9 581 }
665206ee 582 atomic_add_int(&ZoneBigAlloc, (int)size / 1024);
a108bf71 583 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
8f1d5415 584 flags |= M_PASSIVE_ZERO;
a108bf71 585 kup = btokup(chunk);
722871d3 586 *kup = size / PAGE_SIZE;
a108bf71
MD
587 crit_enter();
588 goto done;
589 }
590
591 /*
592 * Attempt to allocate out of an existing zone. First try the free list,
593 * then allocate out of unallocated space. If we find a good zone move
594 * it to the head of the list so later allocations find it quickly
595 * (we might have thousands of zones in the list).
596 *
597 * Note: zoneindex() will panic of size is too large.
598 */
599 zi = zoneindex(&size);
600 KKASSERT(zi < NZONES);
601 crit_enter();
a108bf71 602
5fee07e6 603 if ((z = slgd->ZoneAry[zi]) != NULL) {
a108bf71 604 /*
5fee07e6
MD
605 * Locate a chunk - we have to have at least one. If this is the
606 * last chunk go ahead and do the work to retrieve chunks freed
607 * from remote cpus, and if the zone is still empty move it off
608 * the ZoneAry.
a108bf71 609 */
5fee07e6
MD
610 if (--z->z_NFree <= 0) {
611 KKASSERT(z->z_NFree == 0);
612
613#ifdef SMP
614 /*
615 * WARNING! This code competes with other cpus. It is ok
616 * for us to not drain RChunks here but we might as well, and
617 * it is ok if more accumulate after we're done.
618 *
619 * Set RSignal before pulling rchunks off, indicating that we
620 * will be moving ourselves off of the ZoneAry. Remote ends will
621 * read RSignal before putting rchunks on thus interlocking
622 * their IPI signaling.
623 */
624 if (z->z_RChunks == NULL)
625 atomic_swap_int(&z->z_RSignal, 1);
626
627 while ((bchunk = z->z_RChunks) != NULL) {
628 cpu_ccfence();
629 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
630 *z->z_LChunksp = bchunk;
631 while (bchunk) {
632 chunk_mark_free(z, bchunk);
633 z->z_LChunksp = &bchunk->c_Next;
634 bchunk = bchunk->c_Next;
635 ++z->z_NFree;
636 }
637 break;
638 }
639 }
640#endif
641 /*
642 * Remove from the zone list if no free chunks remain.
643 * Clear RSignal
644 */
645 if (z->z_NFree == 0) {
646 slgd->ZoneAry[zi] = z->z_Next;
647 z->z_Next = NULL;
648 } else {
649 z->z_RSignal = 0;
650 }
a108bf71
MD
651 }
652
653 /*
5fee07e6 654 * Fast path, we have chunks available in z_LChunks.
a108bf71 655 */
5fee07e6
MD
656 chunk = z->z_LChunks;
657 if (chunk) {
10cc6608 658 chunk_mark_allocated(z, chunk);
5fee07e6
MD
659 z->z_LChunks = chunk->c_Next;
660 if (z->z_LChunks == NULL)
661 z->z_LChunksp = &z->z_LChunks;
a108bf71 662 goto done;
a108bf71
MD
663 }
664
665 /*
5fee07e6
MD
666 * No chunks are available in LChunks, the free chunk MUST be
667 * in the never-before-used memory area, controlled by UIndex.
668 *
669 * The consequences are very serious if our zone got corrupted so
670 * we use an explicit panic rather than a KASSERT.
a108bf71 671 */
1c5ca4f3 672 if (z->z_UIndex + 1 != z->z_NMax)
5fee07e6 673 ++z->z_UIndex;
1c5ca4f3
MD
674 else
675 z->z_UIndex = 0;
5fee07e6 676
1c5ca4f3
MD
677 if (z->z_UIndex == z->z_UEndIndex)
678 panic("slaballoc: corrupted zone");
5fee07e6 679
1c5ca4f3 680 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
8f1d5415 681 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 682 flags &= ~M_ZERO;
8f1d5415
MD
683 flags |= M_PASSIVE_ZERO;
684 }
10cc6608 685 chunk_mark_allocated(z, chunk);
a108bf71
MD
686 goto done;
687 }
688
689 /*
690 * If all zones are exhausted we need to allocate a new zone for this
691 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
6ab8e1da
MD
692 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
693 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
694 * we do not pre-zero it because we do not want to mess up the L1 cache.
a108bf71
MD
695 *
696 * At least one subsystem, the tty code (see CROUND) expects power-of-2
697 * allocations to be power-of-2 aligned. We maintain compatibility by
698 * adjusting the base offset below.
699 */
700 {
701 int off;
722871d3 702 int *kup;
a108bf71
MD
703
704 if ((z = slgd->FreeZones) != NULL) {
705 slgd->FreeZones = z->z_Next;
706 --slgd->NFreeZones;
707 bzero(z, sizeof(SLZone));
6ab8e1da 708 z->z_Flags |= SLZF_UNOTZEROD;
a108bf71
MD
709 } else {
710 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
711 if (z == NULL)
712 goto fail;
665206ee 713 atomic_add_int(&ZoneGenAlloc, (int)ZoneSize / 1024);
a108bf71
MD
714 }
715
716 /*
10cc6608
MD
717 * How big is the base structure?
718 */
719#if defined(INVARIANTS)
720 /*
721 * Make room for z_Bitmap. An exact calculation is somewhat more
722 * complicated so don't make an exact calculation.
723 */
724 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
725 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
726#else
727 off = sizeof(SLZone);
728#endif
729
730 /*
a108bf71
MD
731 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
732 * Otherwise just 8-byte align the data.
733 */
734 if ((size | (size - 1)) + 1 == (size << 1))
10cc6608 735 off = (off + size - 1) & ~(size - 1);
a108bf71 736 else
10cc6608 737 off = (off + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
a108bf71
MD
738 z->z_Magic = ZALLOC_SLAB_MAGIC;
739 z->z_ZoneIndex = zi;
740 z->z_NMax = (ZoneSize - off) / size;
741 z->z_NFree = z->z_NMax - 1;
1c5ca4f3
MD
742 z->z_BasePtr = (char *)z + off;
743 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
a108bf71 744 z->z_ChunkSize = size;
2db3b277 745 z->z_CpuGd = gd;
bba6a44d 746 z->z_Cpu = gd->gd_cpuid;
5fee07e6 747 z->z_LChunksp = &z->z_LChunks;
1c5ca4f3 748 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
a108bf71
MD
749 z->z_Next = slgd->ZoneAry[zi];
750 slgd->ZoneAry[zi] = z;
8f1d5415 751 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 752 flags &= ~M_ZERO; /* already zero'd */
8f1d5415
MD
753 flags |= M_PASSIVE_ZERO;
754 }
5fee07e6 755 kup = btokup(z);
722871d3 756 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */
10cc6608 757 chunk_mark_allocated(z, chunk);
1c5ca4f3
MD
758
759 /*
760 * Slide the base index for initial allocations out of the next
761 * zone we create so we do not over-weight the lower part of the
762 * cpu memory caches.
763 */
764 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
765 & (ZALLOC_MAX_ZONE_SIZE - 1);
a108bf71 766 }
5fee07e6 767
a108bf71 768done:
bba6a44d
MD
769 ++type->ks_inuse[gd->gd_cpuid];
770 type->ks_memuse[gd->gd_cpuid] += size;
38e34349 771 type->ks_loosememuse += size; /* not MP synchronized */
a108bf71 772 crit_exit();
5fee07e6 773
a108bf71
MD
774 if (flags & M_ZERO)
775 bzero(chunk, size);
bba6a44d 776#ifdef INVARIANTS
d2182dc1
MD
777 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
778 if (use_malloc_pattern) {
779 for (i = 0; i < size; i += sizeof(int)) {
780 *(int *)((char *)chunk + i) = -1;
781 }
782 }
bba6a44d 783 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
d2182dc1 784 }
bba6a44d 785#endif
5fee07e6 786 logmemory(malloc_end, chunk, type, size, flags);
a108bf71
MD
787 return(chunk);
788fail:
789 crit_exit();
5fee07e6 790 logmemory(malloc_end, NULL, type, size, flags);
a108bf71
MD
791 return(NULL);
792}
793
38e34349
MD
794/*
795 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
796 *
797 * Generally speaking this routine is not called very often and we do
798 * not attempt to optimize it beyond reusing the same pointer if the
799 * new size fits within the chunking of the old pointer's zone.
800 */
a108bf71 801void *
8aca2bd4 802krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
a108bf71 803{
722871d3 804 unsigned long osize;
a108bf71
MD
805 SLZone *z;
806 void *nptr;
722871d3 807 int *kup;
a108bf71 808
eb7f3e3c
MD
809 KKASSERT((flags & M_ZERO) == 0); /* not supported */
810
a108bf71 811 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
efda3bd0 812 return(kmalloc(size, type, flags));
a108bf71 813 if (size == 0) {
efda3bd0 814 kfree(ptr, type);
a108bf71
MD
815 return(NULL);
816 }
817
818 /*
819 * Handle oversized allocations. XXX we really should require that a
820 * size be passed to free() instead of this nonsense.
821 */
5fee07e6 822 kup = btokup(ptr);
722871d3
MD
823 if (*kup > 0) {
824 osize = *kup << PAGE_SHIFT;
5fee07e6
MD
825 if (osize == round_page(size))
826 return(ptr);
827 if ((nptr = kmalloc(size, type, flags)) == NULL)
828 return(NULL);
829 bcopy(ptr, nptr, min(size, osize));
830 kfree(ptr, type);
831 return(nptr);
a108bf71
MD
832 }
833
834 /*
835 * Get the original allocation's zone. If the new request winds up
836 * using the same chunk size we do not have to do anything.
837 */
5fee07e6
MD
838 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
839 kup = btokup(z);
722871d3 840 KKASSERT(*kup < 0);
a108bf71
MD
841 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
842
a108bf71
MD
843 /*
844 * Allocate memory for the new request size. Note that zoneindex has
845 * already adjusted the request size to the appropriate chunk size, which
846 * should optimize our bcopy(). Then copy and return the new pointer.
1ea6580d
MD
847 *
848 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
849 * necessary align the result.
850 *
851 * We can only zoneindex (to align size to the chunk size) if the new
852 * size is not too large.
a108bf71 853 */
1ea6580d
MD
854 if (size < ZoneLimit) {
855 zoneindex(&size);
856 if (z->z_ChunkSize == size)
857 return(ptr);
858 }
efda3bd0 859 if ((nptr = kmalloc(size, type, flags)) == NULL)
a108bf71
MD
860 return(NULL);
861 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
efda3bd0 862 kfree(ptr, type);
a108bf71
MD
863 return(nptr);
864}
865
38e34349 866/*
45d2b1d8
MD
867 * Return the kmalloc limit for this type, in bytes.
868 */
869long
870kmalloc_limit(struct malloc_type *type)
871{
872 if (type->ks_limit == 0) {
873 crit_enter();
874 if (type->ks_limit == 0)
875 malloc_init(type);
876 crit_exit();
877 }
878 return(type->ks_limit);
879}
880
881/*
38e34349
MD
882 * Allocate a copy of the specified string.
883 *
884 * (MP SAFE) (MAY BLOCK)
885 */
1ac06773 886char *
59302080 887kstrdup(const char *str, struct malloc_type *type)
1ac06773
MD
888{
889 int zlen; /* length inclusive of terminating NUL */
890 char *nstr;
891
892 if (str == NULL)
893 return(NULL);
894 zlen = strlen(str) + 1;
efda3bd0 895 nstr = kmalloc(zlen, type, M_WAITOK);
1ac06773
MD
896 bcopy(str, nstr, zlen);
897 return(nstr);
898}
899
1d712609 900#ifdef SMP
a108bf71 901/*
5fee07e6 902 * Notify our cpu that a remote cpu has freed some chunks in a zone that
df9daea8
MD
903 * we own. RCount will be bumped so the memory should be good, but validate
904 * that it really is.
a108bf71
MD
905 */
906static
907void
5fee07e6 908kfree_remote(void *ptr)
a108bf71 909{
5fee07e6
MD
910 SLGlobalData *slgd;
911 SLChunk *bchunk;
912 SLZone *z;
913 int nfree;
722871d3 914 int *kup;
5fee07e6 915
5fee07e6
MD
916 slgd = &mycpu->gd_slab;
917 z = ptr;
918 kup = btokup(z);
df9daea8
MD
919 KKASSERT(*kup == -((int)mycpuid + 1));
920 KKASSERT(z->z_RCount > 0);
921 atomic_subtract_int(&z->z_RCount, 1);
5fee07e6 922
df9daea8
MD
923 logmemory(free_rem_beg, z, NULL, 0, 0);
924 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
925 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
926 nfree = z->z_NFree;
5fee07e6 927
df9daea8
MD
928 /*
929 * Indicate that we will no longer be off of the ZoneAry by
930 * clearing RSignal.
931 */
932 if (z->z_RChunks)
933 z->z_RSignal = 0;
5fee07e6 934
df9daea8
MD
935 /*
936 * Atomically extract the bchunks list and then process it back
937 * into the lchunks list. We want to append our bchunks to the
938 * lchunks list and not prepend since we likely do not have
939 * cache mastership of the related data (not that it helps since
940 * we are using c_Next).
941 */
942 while ((bchunk = z->z_RChunks) != NULL) {
943 cpu_ccfence();
944 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
945 *z->z_LChunksp = bchunk;
946 while (bchunk) {
947 chunk_mark_free(z, bchunk);
948 z->z_LChunksp = &bchunk->c_Next;
949 bchunk = bchunk->c_Next;
950 ++z->z_NFree;
5fee07e6 951 }
df9daea8 952 break;
5fee07e6 953 }
df9daea8
MD
954 }
955 if (z->z_NFree && nfree == 0) {
956 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
957 slgd->ZoneAry[z->z_ZoneIndex] = z;
958 }
5fee07e6 959
df9daea8
MD
960 /*
961 * If the zone becomes totally free, and there are other zones we
962 * can allocate from, move this zone to the FreeZones list. Since
963 * this code can be called from an IPI callback, do *NOT* try to mess
964 * with kernel_map here. Hysteresis will be performed at malloc() time.
965 *
966 * Do not move the zone if there is an IPI inflight, otherwise MP
967 * races can result in our free_remote code accessing a destroyed
968 * zone.
969 */
970 if (z->z_NFree == z->z_NMax &&
971 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
972 z->z_RCount == 0
973 ) {
974 SLZone **pz;
975 int *kup;
5fee07e6 976
df9daea8
MD
977 for (pz = &slgd->ZoneAry[z->z_ZoneIndex];
978 z != *pz;
979 pz = &(*pz)->z_Next) {
980 ;
5fee07e6 981 }
df9daea8
MD
982 *pz = z->z_Next;
983 z->z_Magic = -1;
984 z->z_Next = slgd->FreeZones;
985 slgd->FreeZones = z;
986 ++slgd->NFreeZones;
987 kup = btokup(z);
988 *kup = 0;
5fee07e6 989 }
df9daea8 990 logmemory(free_rem_end, z, bchunk, 0, 0);
a108bf71
MD
991}
992
1d712609
MD
993#endif
994
38e34349 995/*
5b287bba 996 * free (SLAB ALLOCATOR)
38e34349
MD
997 *
998 * Free a memory block previously allocated by malloc. Note that we do not
5fee07e6 999 * attempt to update ks_loosememuse as MP races could prevent us from
38e34349 1000 * checking memory limits in malloc.
5b287bba
MD
1001 *
1002 * MPSAFE
38e34349 1003 */
a108bf71 1004void
8aca2bd4 1005kfree(void *ptr, struct malloc_type *type)
a108bf71
MD
1006{
1007 SLZone *z;
1008 SLChunk *chunk;
1009 SLGlobalData *slgd;
bba6a44d 1010 struct globaldata *gd;
722871d3 1011 int *kup;
5fee07e6 1012 unsigned long size;
d8100bdc
SW
1013#ifdef SMP
1014 SLChunk *bchunk;
5fee07e6 1015 int rsignal;
d8100bdc 1016#endif
a108bf71 1017
b68ad50c 1018 logmemory_quick(free_beg);
bba6a44d
MD
1019 gd = mycpu;
1020 slgd = &gd->gd_slab;
a108bf71 1021
d39911d9
JS
1022 if (ptr == NULL)
1023 panic("trying to free NULL pointer");
1024
a108bf71
MD
1025 /*
1026 * Handle special 0-byte allocations
1027 */
f2b5daf9
MD
1028 if (ptr == ZERO_LENGTH_PTR) {
1029 logmemory(free_zero, ptr, type, -1, 0);
b68ad50c 1030 logmemory_quick(free_end);
a108bf71 1031 return;
f2b5daf9 1032 }
a108bf71
MD
1033
1034 /*
5fee07e6
MD
1035 * Panic on bad malloc type
1036 */
1037 if (type->ks_magic != M_MAGIC)
1038 panic("free: malloc type lacks magic");
1039
1040 /*
a108bf71
MD
1041 * Handle oversized allocations. XXX we really should require that a
1042 * size be passed to free() instead of this nonsense.
bba6a44d
MD
1043 *
1044 * This code is never called via an ipi.
a108bf71 1045 */
5fee07e6 1046 kup = btokup(ptr);
722871d3
MD
1047 if (*kup > 0) {
1048 size = *kup << PAGE_SHIFT;
1049 *kup = 0;
a108bf71 1050#ifdef INVARIANTS
5fee07e6
MD
1051 KKASSERT(sizeof(weirdary) <= size);
1052 bcopy(weirdary, ptr, sizeof(weirdary));
a108bf71 1053#endif
5fee07e6
MD
1054 /*
1055 * NOTE: For oversized allocations we do not record the
1056 * originating cpu. It gets freed on the cpu calling
1057 * kfree(). The statistics are in aggregate.
1058 *
1059 * note: XXX we have still inherited the interrupts-can't-block
1060 * assumption. An interrupt thread does not bump
1061 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1062 * primarily until we can fix softupdate's assumptions about free().
1063 */
1064 crit_enter();
1065 --type->ks_inuse[gd->gd_cpuid];
1066 type->ks_memuse[gd->gd_cpuid] -= size;
1067 if (mycpu->gd_intr_nesting_level ||
1068 (gd->gd_curthread->td_flags & TDF_INTTHREAD))
1069 {
1070 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1071 z = (SLZone *)ptr;
1072 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1073 z->z_Next = slgd->FreeOvZones;
1074 z->z_ChunkSize = size;
1075 slgd->FreeOvZones = z;
1076 crit_exit();
1077 } else {
1078 crit_exit();
1079 logmemory(free_ovsz, ptr, type, size, 0);
1080 kmem_slab_free(ptr, size); /* may block */
1081 atomic_add_int(&ZoneBigAlloc, -(int)size / 1024);
a108bf71 1082 }
5fee07e6
MD
1083 logmemory_quick(free_end);
1084 return;
a108bf71
MD
1085 }
1086
1087 /*
1088 * Zone case. Figure out the zone based on the fact that it is
1089 * ZoneSize aligned.
1090 */
5fee07e6
MD
1091 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1092 kup = btokup(z);
722871d3 1093 KKASSERT(*kup < 0);
a108bf71
MD
1094 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1095
1096 /*
5fee07e6
MD
1097 * If we do not own the zone then use atomic ops to free to the
1098 * remote cpu linked list and notify the target zone using a
1099 * passive message.
1100 *
1101 * The target zone cannot be deallocated while we own a chunk of it,
1102 * so the zone header's storage is stable until the very moment
1103 * we adjust z_RChunks. After that we cannot safely dereference (z).
1104 *
1105 * (no critical section needed)
a108bf71 1106 */
2db3b277 1107 if (z->z_CpuGd != gd) {
75c7ffea 1108#ifdef SMP
5fee07e6
MD
1109 /*
1110 * Making these adjustments now allow us to avoid passing (type)
1111 * to the remote cpu. Note that ks_inuse/ks_memuse is being
28135cc2
MD
1112 * adjusted on OUR cpu, not the zone cpu, but it should all still
1113 * sum up properly and cancel out.
5fee07e6 1114 */
28135cc2
MD
1115 crit_enter();
1116 --type->ks_inuse[gd->gd_cpuid];
1117 type->ks_memuse[gd->gd_cpuid] -= z->z_ChunkSize;
1118 crit_exit();
5fee07e6
MD
1119
1120 /*
1121 * WARNING! This code competes with other cpus. Once we
1122 * successfully link the chunk to RChunks the remote
1123 * cpu can rip z's storage out from under us.
df9daea8
MD
1124 *
1125 * Bumping RCount prevents z's storage from getting
1126 * ripped out.
5fee07e6
MD
1127 */
1128 rsignal = z->z_RSignal;
1129 cpu_lfence();
df9daea8
MD
1130 if (rsignal)
1131 atomic_add_int(&z->z_RCount, 1);
5fee07e6
MD
1132
1133 chunk = ptr;
1134 for (;;) {
1135 bchunk = z->z_RChunks;
1136 cpu_ccfence();
1137 chunk->c_Next = bchunk;
1138 cpu_sfence();
1139
1140 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1141 break;
1142 }
5fee07e6
MD
1143
1144 /*
1145 * We have to signal the remote cpu if our actions will cause
1146 * the remote zone to be placed back on ZoneAry so it can
1147 * move the zone back on.
1148 *
1149 * We only need to deal with NULL->non-NULL RChunk transitions
1150 * and only if z_RSignal is set. We interlock by reading rsignal
1151 * before adding our chunk to RChunks. This should result in
1152 * virtually no IPI traffic.
1153 *
1154 * We can use a passive IPI to reduce overhead even further.
1155 */
1156 if (bchunk == NULL && rsignal) {
1157 logmemory(free_request, ptr, type, z->z_ChunkSize, 0);
1158 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
df9daea8
MD
1159 /* z can get ripped out from under us from this point on */
1160 } else if (rsignal) {
1161 atomic_subtract_int(&z->z_RCount, 1);
1162 /* z can get ripped out from under us from this point on */
5fee07e6 1163 }
75c7ffea
MD
1164#else
1165 panic("Corrupt SLZone");
1166#endif
b68ad50c 1167 logmemory_quick(free_end);
a108bf71
MD
1168 return;
1169 }
1170
5fee07e6
MD
1171 /*
1172 * kfree locally
1173 */
f2b5daf9
MD
1174 logmemory(free_chunk, ptr, type, z->z_ChunkSize, 0);
1175
a108bf71 1176 crit_enter();
a108bf71 1177 chunk = ptr;
10cc6608 1178 chunk_mark_free(z, chunk);
a108bf71
MD
1179
1180 /*
1181 * Put weird data into the memory to detect modifications after freeing,
1182 * illegal pointer use after freeing (we should fault on the odd address),
1183 * and so forth. XXX needs more work, see the old malloc code.
1184 */
1185#ifdef INVARIANTS
1186 if (z->z_ChunkSize < sizeof(weirdary))
1187 bcopy(weirdary, chunk, z->z_ChunkSize);
1188 else
1189 bcopy(weirdary, chunk, sizeof(weirdary));
1190#endif
1191
1192 /*
5fee07e6
MD
1193 * Add this free non-zero'd chunk to a linked list for reuse. Add
1194 * to the front of the linked list so it is more likely to be
1195 * reallocated, since it is already in our L1 cache.
a108bf71 1196 */
6ab8e1da 1197#ifdef INVARIANTS
c439ad8f 1198 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
fc92d4aa 1199 panic("BADFREE %p", chunk);
a108bf71 1200#endif
5fee07e6
MD
1201 chunk->c_Next = z->z_LChunks;
1202 z->z_LChunks = chunk;
1203 if (chunk->c_Next == NULL)
1204 z->z_LChunksp = &chunk->c_Next;
1205
6ab8e1da 1206#ifdef INVARIANTS
c439ad8f 1207 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
a108bf71 1208 panic("BADFREE2");
6ab8e1da 1209#endif
a108bf71
MD
1210
1211 /*
1212 * Bump the number of free chunks. If it becomes non-zero the zone
1213 * must be added back onto the appropriate list.
1214 */
1215 if (z->z_NFree++ == 0) {
1216 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
1217 slgd->ZoneAry[z->z_ZoneIndex] = z;
1218 }
1219
bba6a44d
MD
1220 --type->ks_inuse[z->z_Cpu];
1221 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
a108bf71
MD
1222
1223 /*
1224 * If the zone becomes totally free, and there are other zones we
a7cf0021
MD
1225 * can allocate from, move this zone to the FreeZones list. Since
1226 * this code can be called from an IPI callback, do *NOT* try to mess
1227 * with kernel_map here. Hysteresis will be performed at malloc() time.
a108bf71
MD
1228 */
1229 if (z->z_NFree == z->z_NMax &&
df9daea8
MD
1230 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) &&
1231 z->z_RCount == 0
a108bf71
MD
1232 ) {
1233 SLZone **pz;
722871d3 1234 int *kup;
a108bf71
MD
1235
1236 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
1237 ;
1238 *pz = z->z_Next;
1239 z->z_Magic = -1;
a7cf0021
MD
1240 z->z_Next = slgd->FreeZones;
1241 slgd->FreeZones = z;
1242 ++slgd->NFreeZones;
5fee07e6 1243 kup = btokup(z);
722871d3 1244 *kup = 0;
a108bf71 1245 }
b68ad50c 1246 logmemory_quick(free_end);
a108bf71
MD
1247 crit_exit();
1248}
1249
10cc6608 1250#if defined(INVARIANTS)
5fee07e6 1251
10cc6608
MD
1252/*
1253 * Helper routines for sanity checks
1254 */
1255static
1256void
1257chunk_mark_allocated(SLZone *z, void *chunk)
1258{
1259 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1260 __uint32_t *bitptr;
1261
5fee07e6
MD
1262 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1263 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1264 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
10cc6608
MD
1265 bitptr = &z->z_Bitmap[bitdex >> 5];
1266 bitdex &= 31;
5fee07e6
MD
1267 KASSERT((*bitptr & (1 << bitdex)) == 0,
1268 ("memory chunk %p is already allocated!", chunk));
10cc6608
MD
1269 *bitptr |= 1 << bitdex;
1270}
1271
1272static
1273void
1274chunk_mark_free(SLZone *z, void *chunk)
1275{
1276 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1277 __uint32_t *bitptr;
1278
5fee07e6
MD
1279 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1280 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1281 ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
10cc6608
MD
1282 bitptr = &z->z_Bitmap[bitdex >> 5];
1283 bitdex &= 31;
5fee07e6
MD
1284 KASSERT((*bitptr & (1 << bitdex)) != 0,
1285 ("memory chunk %p is already free!", chunk));
10cc6608
MD
1286 *bitptr &= ~(1 << bitdex);
1287}
1288
1289#endif
1290
a108bf71 1291/*
5b287bba 1292 * kmem_slab_alloc()
a108bf71
MD
1293 *
1294 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1295 * specified alignment. M_* flags are expected in the flags field.
1296 *
1297 * Alignment must be a multiple of PAGE_SIZE.
1298 *
1299 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1300 * but when we move zalloc() over to use this function as its backend
1301 * we will have to switch to kreserve/krelease and call reserve(0)
1302 * after the new space is made available.
dc1fd4b3
MD
1303 *
1304 * Interrupt code which has preempted other code is not allowed to
c397c465
MD
1305 * use PQ_CACHE pages. However, if an interrupt thread is run
1306 * non-preemptively or blocks and then runs non-preemptively, then
1307 * it is free to use PQ_CACHE pages.
a108bf71
MD
1308 */
1309static void *
1310kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1311{
1312 vm_size_t i;
1313 vm_offset_t addr;
1de1e800 1314 int count, vmflags, base_vmflags;
bf1a4675 1315 vm_page_t mp[ZALLOC_MAX_ZONE_SIZE / PAGE_SIZE];
dc1fd4b3 1316 thread_t td;
a108bf71
MD
1317
1318 size = round_page(size);
e4846942 1319 addr = vm_map_min(&kernel_map);
a108bf71
MD
1320
1321 /*
5c39c498
MD
1322 * Reserve properly aligned space from kernel_map. RNOWAIT allocations
1323 * cannot block.
a108bf71 1324 */
5c39c498 1325 if (flags & M_RNOWAIT) {
ed2013d8 1326 if (lwkt_trytoken(&vm_token) == 0)
5c39c498
MD
1327 return(NULL);
1328 } else {
ed2013d8 1329 lwkt_gettoken(&vm_token);
5c39c498 1330 }
a108bf71
MD
1331 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1332 crit_enter();
e4846942 1333 vm_map_lock(&kernel_map);
c809941b 1334 if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) {
e4846942 1335 vm_map_unlock(&kernel_map);
8cb2bf45 1336 if ((flags & M_NULLOK) == 0)
a108bf71 1337 panic("kmem_slab_alloc(): kernel_map ran out of space!");
a108bf71 1338 vm_map_entry_release(count);
2de4f77e 1339 crit_exit();
ed2013d8 1340 lwkt_reltoken(&vm_token);
a108bf71
MD
1341 return(NULL);
1342 }
e4846942
MD
1343
1344 /*
1345 * kernel_object maps 1:1 to kernel_map.
1346 */
c439ad8f 1347 vm_object_reference(&kernel_object);
e4846942
MD
1348 vm_map_insert(&kernel_map, &count,
1349 &kernel_object, addr, addr, addr + size,
1b874851
MD
1350 VM_MAPTYPE_NORMAL,
1351 VM_PROT_ALL, VM_PROT_ALL,
1352 0);
a108bf71 1353
dc1fd4b3 1354 td = curthread;
dc1fd4b3 1355
1de1e800
JS
1356 base_vmflags = 0;
1357 if (flags & M_ZERO)
1358 base_vmflags |= VM_ALLOC_ZERO;
1359 if (flags & M_USE_RESERVE)
1360 base_vmflags |= VM_ALLOC_SYSTEM;
1361 if (flags & M_USE_INTERRUPT_RESERVE)
1362 base_vmflags |= VM_ALLOC_INTERRUPT;
77912481
MD
1363 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1364 panic("kmem_slab_alloc: bad flags %08x (%p)",
1365 flags, ((int **)&size)[-1]);
1366 }
1de1e800
JS
1367
1368
a108bf71
MD
1369 /*
1370 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
1371 */
1372 for (i = 0; i < size; i += PAGE_SIZE) {
1373 vm_page_t m;
fe1e98d0
MD
1374
1375 /*
c397c465
MD
1376 * VM_ALLOC_NORMAL can only be set if we are not preempting.
1377 *
1378 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1379 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
4ecf7cc9
MD
1380 * implied in this case), though I'm not sure if we really need to
1381 * do that.
fe1e98d0 1382 */
1de1e800 1383 vmflags = base_vmflags;
c397c465 1384 if (flags & M_WAITOK) {
1de1e800 1385 if (td->td_preempted)
fe1e98d0 1386 vmflags |= VM_ALLOC_SYSTEM;
1de1e800 1387 else
dc1fd4b3 1388 vmflags |= VM_ALLOC_NORMAL;
dc1fd4b3 1389 }
a108bf71 1390
e4846942 1391 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
bf1a4675
VS
1392 if ((i / PAGE_SIZE) < (sizeof(mp) / sizeof(mp[0])))
1393 mp[i / PAGE_SIZE] = m;
dc1fd4b3
MD
1394
1395 /*
1396 * If the allocation failed we either return NULL or we retry.
1397 *
c397c465
MD
1398 * If M_WAITOK is specified we wait for more memory and retry.
1399 * If M_WAITOK is specified from a preemption we yield instead of
1400 * wait. Livelock will not occur because the interrupt thread
1401 * will not be preempting anyone the second time around after the
1402 * yield.
dc1fd4b3 1403 */
a108bf71 1404 if (m == NULL) {
c397c465 1405 if (flags & M_WAITOK) {
fe1e98d0 1406 if (td->td_preempted) {
e4846942 1407 vm_map_unlock(&kernel_map);
77912481 1408 lwkt_switch();
e4846942 1409 vm_map_lock(&kernel_map);
dc1fd4b3 1410 } else {
e4846942 1411 vm_map_unlock(&kernel_map);
4ecf7cc9 1412 vm_wait(0);
e4846942 1413 vm_map_lock(&kernel_map);
dc1fd4b3 1414 }
a108bf71
MD
1415 i -= PAGE_SIZE; /* retry */
1416 continue;
1417 }
dc1fd4b3
MD
1418
1419 /*
1420 * We were unable to recover, cleanup and return NULL
2de4f77e
MD
1421 *
1422 * (vm_token already held)
dc1fd4b3 1423 */
a108bf71
MD
1424 while (i != 0) {
1425 i -= PAGE_SIZE;
e4846942 1426 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
17cde63e 1427 /* page should already be busy */
a108bf71
MD
1428 vm_page_free(m);
1429 }
e4846942
MD
1430 vm_map_delete(&kernel_map, addr, addr + size, &count);
1431 vm_map_unlock(&kernel_map);
a108bf71 1432 vm_map_entry_release(count);
2de4f77e 1433 crit_exit();
ed2013d8 1434 lwkt_reltoken(&vm_token);
a108bf71
MD
1435 return(NULL);
1436 }
1437 }
1438
1439 /*
dc1fd4b3
MD
1440 * Success!
1441 *
a108bf71
MD
1442 * Mark the map entry as non-pageable using a routine that allows us to
1443 * populate the underlying pages.
17cde63e
MD
1444 *
1445 * The pages were busied by the allocations above.
a108bf71 1446 */
e4846942 1447 vm_map_set_wired_quick(&kernel_map, addr, size, &count);
a108bf71
MD
1448 crit_exit();
1449
1450 /*
1451 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1452 */
1453 for (i = 0; i < size; i += PAGE_SIZE) {
1454 vm_page_t m;
1455
bf1a4675
VS
1456 if ((i / PAGE_SIZE) < (sizeof(mp) / sizeof(mp[0])))
1457 m = mp[i / PAGE_SIZE];
1458 else
1459 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
a108bf71 1460 m->valid = VM_PAGE_BITS_ALL;
17cde63e 1461 /* page should already be busy */
a108bf71
MD
1462 vm_page_wire(m);
1463 vm_page_wakeup(m);
fbbaeba3 1464 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
a108bf71
MD
1465 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1466 bzero((char *)addr + i, PAGE_SIZE);
1467 vm_page_flag_clear(m, PG_ZERO);
17cde63e
MD
1468 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1469 vm_page_flag_set(m, PG_REFERENCED);
a108bf71 1470 }
e4846942 1471 vm_map_unlock(&kernel_map);
a108bf71 1472 vm_map_entry_release(count);
ed2013d8 1473 lwkt_reltoken(&vm_token);
a108bf71
MD
1474 return((void *)addr);
1475}
1476
38e34349 1477/*
5b287bba 1478 * kmem_slab_free()
38e34349 1479 */
a108bf71
MD
1480static void
1481kmem_slab_free(void *ptr, vm_size_t size)
1482{
1483 crit_enter();
ed2013d8 1484 lwkt_gettoken(&vm_token);
e4846942 1485 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
ed2013d8 1486 lwkt_reltoken(&vm_token);
a108bf71
MD
1487 crit_exit();
1488}
1489