Add polling support to BGE.
[dragonfly.git] / sys / kern / kern_slaballoc.c
CommitLineData
a108bf71 1/*
5b287bba 2 * KERN_SLABALLOC.C - Kernel SLAB memory allocator
8c10bfcf
MD
3 *
4 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
5 *
6 * This code is derived from software contributed to The DragonFly Project
7 * by Matthew Dillon <dillon@backplane.com>
8 *
a108bf71
MD
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
8c10bfcf 12 *
a108bf71
MD
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
18 * distribution.
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
22 *
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
a108bf71 34 * SUCH DAMAGE.
8c10bfcf 35 *
ba39e2e0 36 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.49 2007/04/30 07:18:53 dillon Exp $
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
f2b5daf9
MD
122#define MEMORY_STRING "ptr=%p type=%p size=%d flags=%04x"
123#define MEMORY_ARG_SIZE (sizeof(void *) * 2 + sizeof(unsigned long) + \
124 sizeof(int))
125
126#if !defined(KTR_MEMORY)
127#define KTR_MEMORY KTR_ALL
128#endif
129KTR_INFO_MASTER(memory);
130KTR_INFO(KTR_MEMORY, memory, malloc, 0, MEMORY_STRING, MEMORY_ARG_SIZE);
131KTR_INFO(KTR_MEMORY, memory, free_zero, 1, MEMORY_STRING, MEMORY_ARG_SIZE);
017ba73b
MD
132KTR_INFO(KTR_MEMORY, memory, free_ovsz, 2, MEMORY_STRING, MEMORY_ARG_SIZE);
133KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 3, MEMORY_STRING, MEMORY_ARG_SIZE);
134KTR_INFO(KTR_MEMORY, memory, free_chunk, 4, MEMORY_STRING, MEMORY_ARG_SIZE);
f2b5daf9 135#ifdef SMP
017ba73b
MD
136KTR_INFO(KTR_MEMORY, memory, free_request, 5, MEMORY_STRING, MEMORY_ARG_SIZE);
137KTR_INFO(KTR_MEMORY, memory, free_remote, 6, MEMORY_STRING, MEMORY_ARG_SIZE);
f2b5daf9 138#endif
b68ad50c
MD
139KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin", 0);
140KTR_INFO(KTR_MEMORY, memory, free_beg, 0, "free begin", 0);
141KTR_INFO(KTR_MEMORY, memory, free_end, 0, "free end", 0);
f2b5daf9
MD
142
143#define logmemory(name, ptr, type, size, flags) \
144 KTR_LOG(memory_ ## name, ptr, type, size, flags)
b68ad50c
MD
145#define logmemory_quick(name) \
146 KTR_LOG(memory_ ## name)
f2b5daf9 147
a108bf71
MD
148/*
149 * Fixed globals (not per-cpu)
150 */
151static int ZoneSize;
46a3f46d 152static int ZoneLimit;
a108bf71 153static int ZonePageCount;
a108bf71 154static int ZoneMask;
460426e6 155struct malloc_type *kmemstatistics; /* exported to vmstat */
a108bf71
MD
156static struct kmemusage *kmemusage;
157static int32_t weirdary[16];
158
159static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
160static void kmem_slab_free(void *ptr, vm_size_t bytes);
10cc6608
MD
161#if defined(INVARIANTS)
162static void chunk_mark_allocated(SLZone *z, void *chunk);
163static void chunk_mark_free(SLZone *z, void *chunk);
164#endif
a108bf71
MD
165
166/*
167 * Misc constants. Note that allocations that are exact multiples of
168 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
169 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists.
170 */
171#define MIN_CHUNK_SIZE 8 /* in bytes */
172#define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1)
173#define ZONE_RELS_THRESH 2 /* threshold number of zones */
174#define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK)
175
176/*
177 * The WEIRD_ADDR is used as known text to copy into free objects to
178 * try to create deterministic failure cases if the data is accessed after
179 * free.
180 */
181#define WEIRD_ADDR 0xdeadc0de
182#define MAX_COPY sizeof(weirdary)
183#define ZERO_LENGTH_PTR ((void *)-8)
184
185/*
186 * Misc global malloc buckets
187 */
188
189MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
190MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
191MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
192
193MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
194MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
195
196/*
197 * Initialize the slab memory allocator. We have to choose a zone size based
198 * on available physical memory. We choose a zone side which is approximately
199 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
200 * 128K. The zone size is limited to the bounds set in slaballoc.h
201 * (typically 32K min, 128K max).
202 */
203static void kmeminit(void *dummy);
204
ba39e2e0 205SYSINIT(kmem, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL)
a108bf71 206
d2182dc1
MD
207#ifdef INVARIANTS
208/*
209 * If enabled any memory allocated without M_ZERO is initialized to -1.
210 */
211static int use_malloc_pattern;
212SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
213 &use_malloc_pattern, 0, "");
214#endif
215
a108bf71
MD
216static void
217kmeminit(void *dummy)
218{
219 vm_poff_t limsize;
220 int usesize;
221 int i;
222 vm_pindex_t npg;
223
224 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
c439ad8f
MD
225 if (limsize > KvaSize)
226 limsize = KvaSize;
a108bf71
MD
227
228 usesize = (int)(limsize / 1024); /* convert to KB */
229
230 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
231 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
232 ZoneSize <<= 1;
46a3f46d
MD
233 ZoneLimit = ZoneSize / 4;
234 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
235 ZoneLimit = ZALLOC_ZONE_LIMIT;
a108bf71 236 ZoneMask = ZoneSize - 1;
a108bf71
MD
237 ZonePageCount = ZoneSize / PAGE_SIZE;
238
c439ad8f
MD
239 npg = KvaSize / PAGE_SIZE;
240 kmemusage = kmem_slab_alloc(npg * sizeof(struct kmemusage),
241 PAGE_SIZE, M_WAITOK|M_ZERO);
a108bf71
MD
242
243 for (i = 0; i < arysize(weirdary); ++i)
244 weirdary[i] = WEIRD_ADDR;
245
246 if (bootverbose)
6ea70f76 247 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
a108bf71
MD
248}
249
250/*
bba6a44d 251 * Initialize a malloc type tracking structure.
a108bf71
MD
252 */
253void
254malloc_init(void *data)
255{
256 struct malloc_type *type = data;
257 vm_poff_t limsize;
258
259 if (type->ks_magic != M_MAGIC)
260 panic("malloc type lacks magic");
261
262 if (type->ks_limit != 0)
263 return;
264
265 if (vmstats.v_page_count == 0)
266 panic("malloc_init not allowed before vm init");
267
268 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE;
c439ad8f
MD
269 if (limsize > KvaSize)
270 limsize = KvaSize;
a108bf71
MD
271 type->ks_limit = limsize / 10;
272
273 type->ks_next = kmemstatistics;
274 kmemstatistics = type;
275}
276
277void
278malloc_uninit(void *data)
279{
280 struct malloc_type *type = data;
281 struct malloc_type *t;
bba6a44d
MD
282#ifdef INVARIANTS
283 int i;
1d712609 284 long ttl;
bba6a44d 285#endif
a108bf71
MD
286
287 if (type->ks_magic != M_MAGIC)
288 panic("malloc type lacks magic");
289
290 if (vmstats.v_page_count == 0)
291 panic("malloc_uninit not allowed before vm init");
292
293 if (type->ks_limit == 0)
294 panic("malloc_uninit on uninitialized type");
295
296#ifdef INVARIANTS
1d712609
MD
297 /*
298 * memuse is only correct in aggregation. Due to memory being allocated
299 * on one cpu and freed on another individual array entries may be
300 * negative or positive (canceling each other out).
301 */
302 for (i = ttl = 0; i < ncpus; ++i)
303 ttl += type->ks_memuse[i];
304 if (ttl) {
6ea70f76 305 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
1d712609 306 ttl, type->ks_shortdesc, i);
a108bf71
MD
307 }
308#endif
309 if (type == kmemstatistics) {
310 kmemstatistics = type->ks_next;
311 } else {
312 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
313 if (t->ks_next == type) {
314 t->ks_next = type->ks_next;
315 break;
316 }
317 }
318 }
319 type->ks_next = NULL;
320 type->ks_limit = 0;
321}
322
323/*
324 * Calculate the zone index for the allocation request size and set the
325 * allocation request size to that particular zone's chunk size.
326 */
327static __inline int
328zoneindex(unsigned long *bytes)
329{
330 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
331 if (n < 128) {
332 *bytes = n = (n + 7) & ~7;
333 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
334 }
335 if (n < 256) {
336 *bytes = n = (n + 15) & ~15;
337 return(n / 16 + 7);
338 }
339 if (n < 8192) {
340 if (n < 512) {
341 *bytes = n = (n + 31) & ~31;
342 return(n / 32 + 15);
343 }
344 if (n < 1024) {
345 *bytes = n = (n + 63) & ~63;
346 return(n / 64 + 23);
347 }
348 if (n < 2048) {
349 *bytes = n = (n + 127) & ~127;
350 return(n / 128 + 31);
351 }
352 if (n < 4096) {
353 *bytes = n = (n + 255) & ~255;
354 return(n / 256 + 39);
355 }
356 *bytes = n = (n + 511) & ~511;
357 return(n / 512 + 47);
358 }
359#if ZALLOC_ZONE_LIMIT > 8192
360 if (n < 16384) {
361 *bytes = n = (n + 1023) & ~1023;
362 return(n / 1024 + 55);
363 }
364#endif
365#if ZALLOC_ZONE_LIMIT > 16384
366 if (n < 32768) {
367 *bytes = n = (n + 2047) & ~2047;
368 return(n / 2048 + 63);
369 }
370#endif
371 panic("Unexpected byte count %d", n);
372 return(0);
373}
374
375/*
5b287bba 376 * malloc() (SLAB ALLOCATOR)
a108bf71
MD
377 *
378 * Allocate memory via the slab allocator. If the request is too large,
379 * or if it page-aligned beyond a certain size, we fall back to the
380 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
381 * &SlabMisc if you don't care.
382 *
8cb2bf45
JS
383 * M_RNOWAIT - don't block.
384 * M_NULLOK - return NULL instead of blocking.
a108bf71 385 * M_ZERO - zero the returned memory.
dc1fd4b3
MD
386 * M_USE_RESERVE - allow greater drawdown of the free list
387 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
5b287bba
MD
388 *
389 * MPSAFE
a108bf71 390 */
8aca2bd4 391
a108bf71 392void *
8aca2bd4 393kmalloc(unsigned long size, struct malloc_type *type, int flags)
a108bf71
MD
394{
395 SLZone *z;
396 SLChunk *chunk;
397 SLGlobalData *slgd;
bba6a44d 398 struct globaldata *gd;
a108bf71 399 int zi;
d2182dc1
MD
400#ifdef INVARIANTS
401 int i;
402#endif
a108bf71 403
b68ad50c 404 logmemory_quick(malloc_beg);
bba6a44d
MD
405 gd = mycpu;
406 slgd = &gd->gd_slab;
a108bf71
MD
407
408 /*
409 * XXX silly to have this in the critical path.
410 */
411 if (type->ks_limit == 0) {
412 crit_enter();
413 if (type->ks_limit == 0)
414 malloc_init(type);
415 crit_exit();
416 }
417 ++type->ks_calls;
418
419 /*
38e34349
MD
420 * Handle the case where the limit is reached. Panic if we can't return
421 * NULL. The original malloc code looped, but this tended to
a108bf71 422 * simply deadlock the computer.
38e34349
MD
423 *
424 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
425 * to determine if a more complete limit check should be done. The
426 * actual memory use is tracked via ks_memuse[cpu].
a108bf71 427 */
bba6a44d
MD
428 while (type->ks_loosememuse >= type->ks_limit) {
429 int i;
430 long ttl;
431
432 for (i = ttl = 0; i < ncpus; ++i)
433 ttl += type->ks_memuse[i];
38e34349 434 type->ks_loosememuse = ttl; /* not MP synchronized */
bba6a44d 435 if (ttl >= type->ks_limit) {
f2b5daf9
MD
436 if (flags & M_NULLOK) {
437 logmemory(malloc, NULL, type, size, flags);
bba6a44d 438 return(NULL);
f2b5daf9 439 }
bba6a44d
MD
440 panic("%s: malloc limit exceeded", type->ks_shortdesc);
441 }
a108bf71
MD
442 }
443
444 /*
445 * Handle the degenerate size == 0 case. Yes, this does happen.
446 * Return a special pointer. This is to maintain compatibility with
447 * the original malloc implementation. Certain devices, such as the
448 * adaptec driver, not only allocate 0 bytes, they check for NULL and
449 * also realloc() later on. Joy.
450 */
f2b5daf9
MD
451 if (size == 0) {
452 logmemory(malloc, ZERO_LENGTH_PTR, type, size, flags);
a108bf71 453 return(ZERO_LENGTH_PTR);
f2b5daf9 454 }
a108bf71 455
a7cf0021
MD
456 /*
457 * Handle hysteresis from prior frees here in malloc(). We cannot
458 * safely manipulate the kernel_map in free() due to free() possibly
459 * being called via an IPI message or from sensitive interrupt code.
460 */
dc1fd4b3 461 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_RNOWAIT) == 0) {
46a3f46d
MD
462 crit_enter();
463 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */
464 z = slgd->FreeZones;
465 slgd->FreeZones = z->z_Next;
466 --slgd->NFreeZones;
467 kmem_slab_free(z, ZoneSize); /* may block */
468 }
469 crit_exit();
470 }
471 /*
472 * XXX handle oversized frees that were queued from free().
473 */
dc1fd4b3 474 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) {
46a3f46d
MD
475 crit_enter();
476 if ((z = slgd->FreeOvZones) != NULL) {
477 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
478 slgd->FreeOvZones = z->z_Next;
479 kmem_slab_free(z, z->z_ChunkSize); /* may block */
480 }
481 crit_exit();
a7cf0021
MD
482 }
483
a108bf71
MD
484 /*
485 * Handle large allocations directly. There should not be very many of
486 * these so performance is not a big issue.
487 *
488 * Guarentee page alignment for allocations in multiples of PAGE_SIZE
489 */
46a3f46d 490 if (size >= ZoneLimit || (size & PAGE_MASK) == 0) {
a108bf71
MD
491 struct kmemusage *kup;
492
493 size = round_page(size);
494 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
f2b5daf9
MD
495 if (chunk == NULL) {
496 logmemory(malloc, NULL, type, size, flags);
a108bf71 497 return(NULL);
f2b5daf9 498 }
a108bf71 499 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
8f1d5415 500 flags |= M_PASSIVE_ZERO;
a108bf71
MD
501 kup = btokup(chunk);
502 kup->ku_pagecnt = size / PAGE_SIZE;
bba6a44d 503 kup->ku_cpu = gd->gd_cpuid;
a108bf71
MD
504 crit_enter();
505 goto done;
506 }
507
508 /*
509 * Attempt to allocate out of an existing zone. First try the free list,
510 * then allocate out of unallocated space. If we find a good zone move
511 * it to the head of the list so later allocations find it quickly
512 * (we might have thousands of zones in the list).
513 *
514 * Note: zoneindex() will panic of size is too large.
515 */
516 zi = zoneindex(&size);
517 KKASSERT(zi < NZONES);
518 crit_enter();
519 if ((z = slgd->ZoneAry[zi]) != NULL) {
520 KKASSERT(z->z_NFree > 0);
521
522 /*
523 * Remove us from the ZoneAry[] when we become empty
524 */
525 if (--z->z_NFree == 0) {
526 slgd->ZoneAry[zi] = z->z_Next;
527 z->z_Next = NULL;
528 }
529
530 /*
531 * Locate a chunk in a free page. This attempts to localize
532 * reallocations into earlier pages without us having to sort
533 * the chunk list. A chunk may still overlap a page boundary.
534 */
535 while (z->z_FirstFreePg < ZonePageCount) {
536 if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) {
537#ifdef DIAGNOSTIC
538 /*
539 * Diagnostic: c_Next is not total garbage.
540 */
541 KKASSERT(chunk->c_Next == NULL ||
542 ((intptr_t)chunk->c_Next & IN_SAME_PAGE_MASK) ==
543 ((intptr_t)chunk & IN_SAME_PAGE_MASK));
544#endif
6ab8e1da 545#ifdef INVARIANTS
c439ad8f 546 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
a108bf71 547 panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount);
c439ad8f 548 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
a108bf71 549 panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount);
10cc6608 550 chunk_mark_allocated(z, chunk);
6ab8e1da 551#endif
a108bf71
MD
552 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next;
553 goto done;
554 }
555 ++z->z_FirstFreePg;
556 }
557
558 /*
1c5ca4f3
MD
559 * No chunks are available but NFree said we had some memory, so
560 * it must be available in the never-before-used-memory area
561 * governed by UIndex. The consequences are very serious if our zone
562 * got corrupted so we use an explicit panic rather then a KASSERT.
a108bf71 563 */
1c5ca4f3
MD
564 if (z->z_UIndex + 1 != z->z_NMax)
565 z->z_UIndex = z->z_UIndex + 1;
566 else
567 z->z_UIndex = 0;
568 if (z->z_UIndex == z->z_UEndIndex)
569 panic("slaballoc: corrupted zone");
570 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
8f1d5415 571 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 572 flags &= ~M_ZERO;
8f1d5415
MD
573 flags |= M_PASSIVE_ZERO;
574 }
10cc6608
MD
575#if defined(INVARIANTS)
576 chunk_mark_allocated(z, chunk);
577#endif
a108bf71
MD
578 goto done;
579 }
580
581 /*
582 * If all zones are exhausted we need to allocate a new zone for this
583 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
6ab8e1da
MD
584 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
585 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
586 * we do not pre-zero it because we do not want to mess up the L1 cache.
a108bf71
MD
587 *
588 * At least one subsystem, the tty code (see CROUND) expects power-of-2
589 * allocations to be power-of-2 aligned. We maintain compatibility by
590 * adjusting the base offset below.
591 */
592 {
593 int off;
594
595 if ((z = slgd->FreeZones) != NULL) {
596 slgd->FreeZones = z->z_Next;
597 --slgd->NFreeZones;
598 bzero(z, sizeof(SLZone));
6ab8e1da 599 z->z_Flags |= SLZF_UNOTZEROD;
a108bf71
MD
600 } else {
601 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
602 if (z == NULL)
603 goto fail;
604 }
605
10cc6608
MD
606 /*
607 * How big is the base structure?
608 */
609#if defined(INVARIANTS)
610 /*
611 * Make room for z_Bitmap. An exact calculation is somewhat more
612 * complicated so don't make an exact calculation.
613 */
614 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
615 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
616#else
617 off = sizeof(SLZone);
618#endif
619
a108bf71
MD
620 /*
621 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
622 * Otherwise just 8-byte align the data.
623 */
624 if ((size | (size - 1)) + 1 == (size << 1))
10cc6608 625 off = (off + size - 1) & ~(size - 1);
a108bf71 626 else
10cc6608 627 off = (off + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK;
a108bf71
MD
628 z->z_Magic = ZALLOC_SLAB_MAGIC;
629 z->z_ZoneIndex = zi;
630 z->z_NMax = (ZoneSize - off) / size;
631 z->z_NFree = z->z_NMax - 1;
1c5ca4f3
MD
632 z->z_BasePtr = (char *)z + off;
633 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
a108bf71
MD
634 z->z_ChunkSize = size;
635 z->z_FirstFreePg = ZonePageCount;
2db3b277 636 z->z_CpuGd = gd;
bba6a44d 637 z->z_Cpu = gd->gd_cpuid;
1c5ca4f3 638 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
a108bf71
MD
639 z->z_Next = slgd->ZoneAry[zi];
640 slgd->ZoneAry[zi] = z;
8f1d5415 641 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
6ab8e1da 642 flags &= ~M_ZERO; /* already zero'd */
8f1d5415
MD
643 flags |= M_PASSIVE_ZERO;
644 }
10cc6608
MD
645#if defined(INVARIANTS)
646 chunk_mark_allocated(z, chunk);
647#endif
1c5ca4f3
MD
648
649 /*
650 * Slide the base index for initial allocations out of the next
651 * zone we create so we do not over-weight the lower part of the
652 * cpu memory caches.
653 */
654 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
655 & (ZALLOC_MAX_ZONE_SIZE - 1);
a108bf71
MD
656 }
657done:
bba6a44d
MD
658 ++type->ks_inuse[gd->gd_cpuid];
659 type->ks_memuse[gd->gd_cpuid] += size;
38e34349 660 type->ks_loosememuse += size; /* not MP synchronized */
a108bf71
MD
661 crit_exit();
662 if (flags & M_ZERO)
663 bzero(chunk, size);
bba6a44d 664#ifdef INVARIANTS
d2182dc1
MD
665 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
666 if (use_malloc_pattern) {
667 for (i = 0; i < size; i += sizeof(int)) {
668 *(int *)((char *)chunk + i) = -1;
669 }
670 }
bba6a44d 671 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
d2182dc1 672 }
bba6a44d 673#endif
f2b5daf9 674 logmemory(malloc, chunk, type, size, flags);
a108bf71
MD
675 return(chunk);
676fail:
677 crit_exit();
f2b5daf9 678 logmemory(malloc, NULL, type, size, flags);
a108bf71
MD
679 return(NULL);
680}
681
38e34349
MD
682/*
683 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
684 *
685 * Generally speaking this routine is not called very often and we do
686 * not attempt to optimize it beyond reusing the same pointer if the
687 * new size fits within the chunking of the old pointer's zone.
688 */
a108bf71 689void *
8aca2bd4 690krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
a108bf71
MD
691{
692 SLZone *z;
693 void *nptr;
694 unsigned long osize;
695
eb7f3e3c
MD
696 KKASSERT((flags & M_ZERO) == 0); /* not supported */
697
a108bf71 698 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
efda3bd0 699 return(kmalloc(size, type, flags));
a108bf71 700 if (size == 0) {
efda3bd0 701 kfree(ptr, type);
a108bf71
MD
702 return(NULL);
703 }
704
705 /*
706 * Handle oversized allocations. XXX we really should require that a
707 * size be passed to free() instead of this nonsense.
708 */
709 {
710 struct kmemusage *kup;
711
712 kup = btokup(ptr);
713 if (kup->ku_pagecnt) {
714 osize = kup->ku_pagecnt << PAGE_SHIFT;
715 if (osize == round_page(size))
716 return(ptr);
efda3bd0 717 if ((nptr = kmalloc(size, type, flags)) == NULL)
a108bf71
MD
718 return(NULL);
719 bcopy(ptr, nptr, min(size, osize));
efda3bd0 720 kfree(ptr, type);
a108bf71
MD
721 return(nptr);
722 }
723 }
724
725 /*
726 * Get the original allocation's zone. If the new request winds up
727 * using the same chunk size we do not have to do anything.
728 */
729 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
730 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
731
732 zoneindex(&size);
733 if (z->z_ChunkSize == size)
734 return(ptr);
735
736 /*
737 * Allocate memory for the new request size. Note that zoneindex has
738 * already adjusted the request size to the appropriate chunk size, which
739 * should optimize our bcopy(). Then copy and return the new pointer.
740 */
efda3bd0 741 if ((nptr = kmalloc(size, type, flags)) == NULL)
a108bf71
MD
742 return(NULL);
743 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
efda3bd0 744 kfree(ptr, type);
a108bf71
MD
745 return(nptr);
746}
747
38e34349
MD
748/*
749 * Allocate a copy of the specified string.
750 *
751 * (MP SAFE) (MAY BLOCK)
752 */
1ac06773 753char *
59302080 754kstrdup(const char *str, struct malloc_type *type)
1ac06773
MD
755{
756 int zlen; /* length inclusive of terminating NUL */
757 char *nstr;
758
759 if (str == NULL)
760 return(NULL);
761 zlen = strlen(str) + 1;
efda3bd0 762 nstr = kmalloc(zlen, type, M_WAITOK);
1ac06773
MD
763 bcopy(str, nstr, zlen);
764 return(nstr);
765}
766
1d712609 767#ifdef SMP
a108bf71
MD
768/*
769 * free() (SLAB ALLOCATOR)
770 *
bba6a44d 771 * Free the specified chunk of memory.
a108bf71
MD
772 */
773static
774void
775free_remote(void *ptr)
776{
f2b5daf9 777 logmemory(free_remote, ptr, *(struct malloc_type **)ptr, -1, 0);
efda3bd0 778 kfree(ptr, *(struct malloc_type **)ptr);
a108bf71
MD
779}
780
1d712609
MD
781#endif
782
38e34349 783/*
5b287bba 784 * free (SLAB ALLOCATOR)
38e34349
MD
785 *
786 * Free a memory block previously allocated by malloc. Note that we do not
787 * attempt to uplodate ks_loosememuse as MP races could prevent us from
788 * checking memory limits in malloc.
5b287bba
MD
789 *
790 * MPSAFE
38e34349 791 */
a108bf71 792void
8aca2bd4 793kfree(void *ptr, struct malloc_type *type)
a108bf71
MD
794{
795 SLZone *z;
796 SLChunk *chunk;
797 SLGlobalData *slgd;
bba6a44d 798 struct globaldata *gd;
a108bf71
MD
799 int pgno;
800
b68ad50c 801 logmemory_quick(free_beg);
bba6a44d
MD
802 gd = mycpu;
803 slgd = &gd->gd_slab;
a108bf71 804
d39911d9
JS
805 if (ptr == NULL)
806 panic("trying to free NULL pointer");
807
a108bf71
MD
808 /*
809 * Handle special 0-byte allocations
810 */
f2b5daf9
MD
811 if (ptr == ZERO_LENGTH_PTR) {
812 logmemory(free_zero, ptr, type, -1, 0);
b68ad50c 813 logmemory_quick(free_end);
a108bf71 814 return;
f2b5daf9 815 }
a108bf71
MD
816
817 /*
818 * Handle oversized allocations. XXX we really should require that a
819 * size be passed to free() instead of this nonsense.
bba6a44d
MD
820 *
821 * This code is never called via an ipi.
a108bf71
MD
822 */
823 {
824 struct kmemusage *kup;
825 unsigned long size;
826
827 kup = btokup(ptr);
828 if (kup->ku_pagecnt) {
829 size = kup->ku_pagecnt << PAGE_SHIFT;
830 kup->ku_pagecnt = 0;
a108bf71
MD
831#ifdef INVARIANTS
832 KKASSERT(sizeof(weirdary) <= size);
833 bcopy(weirdary, ptr, sizeof(weirdary));
834#endif
bba6a44d
MD
835 /*
836 * note: we always adjust our cpu's slot, not the originating
837 * cpu (kup->ku_cpuid). The statistics are in aggregate.
81f5fc99
MD
838 *
839 * note: XXX we have still inherited the interrupts-can't-block
840 * assumption. An interrupt thread does not bump
841 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
842 * primarily until we can fix softupdate's assumptions about free().
bba6a44d
MD
843 */
844 crit_enter();
845 --type->ks_inuse[gd->gd_cpuid];
846 type->ks_memuse[gd->gd_cpuid] -= size;
81f5fc99 847 if (mycpu->gd_intr_nesting_level || (gd->gd_curthread->td_flags & TDF_INTTHREAD)) {
f2b5daf9 848 logmemory(free_ovsz_delayed, ptr, type, size, 0);
46a3f46d
MD
849 z = (SLZone *)ptr;
850 z->z_Magic = ZALLOC_OVSZ_MAGIC;
851 z->z_Next = slgd->FreeOvZones;
852 z->z_ChunkSize = size;
853 slgd->FreeOvZones = z;
854 crit_exit();
855 } else {
bba6a44d 856 crit_exit();
f2b5daf9 857 logmemory(free_ovsz, ptr, type, size, 0);
46a3f46d
MD
858 kmem_slab_free(ptr, size); /* may block */
859 }
b68ad50c 860 logmemory_quick(free_end);
a108bf71
MD
861 return;
862 }
863 }
864
865 /*
866 * Zone case. Figure out the zone based on the fact that it is
867 * ZoneSize aligned.
868 */
869 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask);
870 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
871
872 /*
873 * If we do not own the zone then forward the request to the
4c9f5a7f
MD
874 * cpu that does. Since the timing is non-critical, a passive
875 * message is sent.
a108bf71 876 */
2db3b277 877 if (z->z_CpuGd != gd) {
a108bf71 878 *(struct malloc_type **)ptr = type;
75c7ffea 879#ifdef SMP
f2b5daf9 880 logmemory(free_request, ptr, type, z->z_ChunkSize, 0);
4c9f5a7f 881 lwkt_send_ipiq_passive(z->z_CpuGd, free_remote, ptr);
75c7ffea
MD
882#else
883 panic("Corrupt SLZone");
884#endif
b68ad50c 885 logmemory_quick(free_end);
a108bf71
MD
886 return;
887 }
888
f2b5daf9
MD
889 logmemory(free_chunk, ptr, type, z->z_ChunkSize, 0);
890
a108bf71
MD
891 if (type->ks_magic != M_MAGIC)
892 panic("free: malloc type lacks magic");
893
894 crit_enter();
895 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT;
896 chunk = ptr;
897
bba6a44d 898#ifdef INVARIANTS
a108bf71 899 /*
bba6a44d
MD
900 * Attempt to detect a double-free. To reduce overhead we only check
901 * if there appears to be link pointer at the base of the data.
a108bf71
MD
902 */
903 if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) {
904 SLChunk *scan;
905 for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) {
906 if (scan == chunk)
907 panic("Double free at %p", chunk);
908 }
909 }
10cc6608 910 chunk_mark_free(z, chunk);
a108bf71
MD
911#endif
912
913 /*
914 * Put weird data into the memory to detect modifications after freeing,
915 * illegal pointer use after freeing (we should fault on the odd address),
916 * and so forth. XXX needs more work, see the old malloc code.
917 */
918#ifdef INVARIANTS
919 if (z->z_ChunkSize < sizeof(weirdary))
920 bcopy(weirdary, chunk, z->z_ChunkSize);
921 else
922 bcopy(weirdary, chunk, sizeof(weirdary));
923#endif
924
925 /*
926 * Add this free non-zero'd chunk to a linked list for reuse, adjust
927 * z_FirstFreePg.
928 */
6ab8e1da 929#ifdef INVARIANTS
c439ad8f 930 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
fc92d4aa 931 panic("BADFREE %p", chunk);
a108bf71
MD
932#endif
933 chunk->c_Next = z->z_PageAry[pgno];
934 z->z_PageAry[pgno] = chunk;
6ab8e1da 935#ifdef INVARIANTS
c439ad8f 936 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
a108bf71 937 panic("BADFREE2");
6ab8e1da 938#endif
a108bf71
MD
939 if (z->z_FirstFreePg > pgno)
940 z->z_FirstFreePg = pgno;
941
942 /*
943 * Bump the number of free chunks. If it becomes non-zero the zone
944 * must be added back onto the appropriate list.
945 */
946 if (z->z_NFree++ == 0) {
947 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex];
948 slgd->ZoneAry[z->z_ZoneIndex] = z;
949 }
950
bba6a44d
MD
951 --type->ks_inuse[z->z_Cpu];
952 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize;
a108bf71
MD
953
954 /*
955 * If the zone becomes totally free, and there are other zones we
a7cf0021
MD
956 * can allocate from, move this zone to the FreeZones list. Since
957 * this code can be called from an IPI callback, do *NOT* try to mess
958 * with kernel_map here. Hysteresis will be performed at malloc() time.
a108bf71
MD
959 */
960 if (z->z_NFree == z->z_NMax &&
961 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z)
962 ) {
963 SLZone **pz;
964
965 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next)
966 ;
967 *pz = z->z_Next;
968 z->z_Magic = -1;
a7cf0021
MD
969 z->z_Next = slgd->FreeZones;
970 slgd->FreeZones = z;
971 ++slgd->NFreeZones;
a108bf71 972 }
b68ad50c 973 logmemory_quick(free_end);
a108bf71
MD
974 crit_exit();
975}
976
10cc6608
MD
977#if defined(INVARIANTS)
978/*
979 * Helper routines for sanity checks
980 */
981static
982void
983chunk_mark_allocated(SLZone *z, void *chunk)
984{
985 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
986 __uint32_t *bitptr;
987
988 KASSERT(bitdex >= 0 && bitdex < z->z_NMax, ("memory chunk %p bit index %d is illegal", chunk, bitdex));
989 bitptr = &z->z_Bitmap[bitdex >> 5];
990 bitdex &= 31;
991 KASSERT((*bitptr & (1 << bitdex)) == 0, ("memory chunk %p is already allocated!", chunk));
992 *bitptr |= 1 << bitdex;
993}
994
995static
996void
997chunk_mark_free(SLZone *z, void *chunk)
998{
999 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1000 __uint32_t *bitptr;
1001
1002 KASSERT(bitdex >= 0 && bitdex < z->z_NMax, ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
1003 bitptr = &z->z_Bitmap[bitdex >> 5];
1004 bitdex &= 31;
1005 KASSERT((*bitptr & (1 << bitdex)) != 0, ("memory chunk %p is already free!", chunk));
1006 *bitptr &= ~(1 << bitdex);
1007}
1008
1009#endif
1010
a108bf71 1011/*
5b287bba 1012 * kmem_slab_alloc()
a108bf71
MD
1013 *
1014 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1015 * specified alignment. M_* flags are expected in the flags field.
1016 *
1017 * Alignment must be a multiple of PAGE_SIZE.
1018 *
1019 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1020 * but when we move zalloc() over to use this function as its backend
1021 * we will have to switch to kreserve/krelease and call reserve(0)
1022 * after the new space is made available.
dc1fd4b3
MD
1023 *
1024 * Interrupt code which has preempted other code is not allowed to
c397c465
MD
1025 * use PQ_CACHE pages. However, if an interrupt thread is run
1026 * non-preemptively or blocks and then runs non-preemptively, then
1027 * it is free to use PQ_CACHE pages.
38e34349
MD
1028 *
1029 * This routine will currently obtain the BGL.
5b287bba
MD
1030 *
1031 * MPALMOSTSAFE - acquires mplock
a108bf71
MD
1032 */
1033static void *
1034kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1035{
1036 vm_size_t i;
1037 vm_offset_t addr;
1de1e800 1038 int count, vmflags, base_vmflags;
dc1fd4b3 1039 thread_t td;
a108bf71
MD
1040
1041 size = round_page(size);
e4846942 1042 addr = vm_map_min(&kernel_map);
a108bf71
MD
1043
1044 /*
5c39c498
MD
1045 * Reserve properly aligned space from kernel_map. RNOWAIT allocations
1046 * cannot block.
a108bf71 1047 */
5c39c498
MD
1048 if (flags & M_RNOWAIT) {
1049 if (try_mplock() == 0)
1050 return(NULL);
1051 } else {
1052 get_mplock();
1053 }
a108bf71
MD
1054 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1055 crit_enter();
e4846942
MD
1056 vm_map_lock(&kernel_map);
1057 if (vm_map_findspace(&kernel_map, addr, size, align, &addr)) {
1058 vm_map_unlock(&kernel_map);
8cb2bf45 1059 if ((flags & M_NULLOK) == 0)
a108bf71
MD
1060 panic("kmem_slab_alloc(): kernel_map ran out of space!");
1061 crit_exit();
1062 vm_map_entry_release(count);
38e34349 1063 rel_mplock();
a108bf71
MD
1064 return(NULL);
1065 }
e4846942
MD
1066
1067 /*
1068 * kernel_object maps 1:1 to kernel_map.
1069 */
c439ad8f 1070 vm_object_reference(&kernel_object);
e4846942
MD
1071 vm_map_insert(&kernel_map, &count,
1072 &kernel_object, addr, addr, addr + size,
1b874851
MD
1073 VM_MAPTYPE_NORMAL,
1074 VM_PROT_ALL, VM_PROT_ALL,
1075 0);
a108bf71 1076
dc1fd4b3 1077 td = curthread;
dc1fd4b3 1078
1de1e800
JS
1079 base_vmflags = 0;
1080 if (flags & M_ZERO)
1081 base_vmflags |= VM_ALLOC_ZERO;
1082 if (flags & M_USE_RESERVE)
1083 base_vmflags |= VM_ALLOC_SYSTEM;
1084 if (flags & M_USE_INTERRUPT_RESERVE)
1085 base_vmflags |= VM_ALLOC_INTERRUPT;
1086 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0)
1087 panic("kmem_slab_alloc: bad flags %08x (%p)", flags, ((int **)&size)[-1]);
1088
1089
a108bf71
MD
1090 /*
1091 * Allocate the pages. Do not mess with the PG_ZERO flag yet.
1092 */
1093 for (i = 0; i < size; i += PAGE_SIZE) {
1094 vm_page_t m;
fe1e98d0
MD
1095
1096 /*
c397c465
MD
1097 * VM_ALLOC_NORMAL can only be set if we are not preempting.
1098 *
1099 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1100 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1101 * implied in this case), though I'm sure if we really need to do
1102 * that.
fe1e98d0 1103 */
1de1e800 1104 vmflags = base_vmflags;
c397c465 1105 if (flags & M_WAITOK) {
1de1e800 1106 if (td->td_preempted)
fe1e98d0 1107 vmflags |= VM_ALLOC_SYSTEM;
1de1e800 1108 else
dc1fd4b3 1109 vmflags |= VM_ALLOC_NORMAL;
dc1fd4b3 1110 }
a108bf71 1111
e4846942 1112 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
dc1fd4b3
MD
1113
1114 /*
1115 * If the allocation failed we either return NULL or we retry.
1116 *
c397c465
MD
1117 * If M_WAITOK is specified we wait for more memory and retry.
1118 * If M_WAITOK is specified from a preemption we yield instead of
1119 * wait. Livelock will not occur because the interrupt thread
1120 * will not be preempting anyone the second time around after the
1121 * yield.
dc1fd4b3 1122 */
a108bf71 1123 if (m == NULL) {
c397c465 1124 if (flags & M_WAITOK) {
fe1e98d0 1125 if (td->td_preempted) {
e4846942 1126 vm_map_unlock(&kernel_map);
dc1fd4b3 1127 lwkt_yield();
e4846942 1128 vm_map_lock(&kernel_map);
dc1fd4b3 1129 } else {
e4846942 1130 vm_map_unlock(&kernel_map);
dc1fd4b3 1131 vm_wait();
e4846942 1132 vm_map_lock(&kernel_map);
dc1fd4b3 1133 }
a108bf71
MD
1134 i -= PAGE_SIZE; /* retry */
1135 continue;
1136 }
dc1fd4b3
MD
1137
1138 /*
1139 * We were unable to recover, cleanup and return NULL
1140 */
a108bf71
MD
1141 while (i != 0) {
1142 i -= PAGE_SIZE;
e4846942 1143 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
a108bf71
MD
1144 vm_page_free(m);
1145 }
e4846942
MD
1146 vm_map_delete(&kernel_map, addr, addr + size, &count);
1147 vm_map_unlock(&kernel_map);
a108bf71
MD
1148 crit_exit();
1149 vm_map_entry_release(count);
38e34349 1150 rel_mplock();
a108bf71
MD
1151 return(NULL);
1152 }
1153 }
1154
1155 /*
dc1fd4b3
MD
1156 * Success!
1157 *
a108bf71
MD
1158 * Mark the map entry as non-pageable using a routine that allows us to
1159 * populate the underlying pages.
1160 */
e4846942 1161 vm_map_set_wired_quick(&kernel_map, addr, size, &count);
a108bf71
MD
1162 crit_exit();
1163
1164 /*
1165 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO.
1166 */
1167 for (i = 0; i < size; i += PAGE_SIZE) {
1168 vm_page_t m;
1169
e4846942 1170 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
a108bf71
MD
1171 m->valid = VM_PAGE_BITS_ALL;
1172 vm_page_wire(m);
1173 vm_page_wakeup(m);
fbbaeba3 1174 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL, 1);
a108bf71
MD
1175 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO))
1176 bzero((char *)addr + i, PAGE_SIZE);
1177 vm_page_flag_clear(m, PG_ZERO);
1178 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_REFERENCED);
1179 }
e4846942 1180 vm_map_unlock(&kernel_map);
a108bf71 1181 vm_map_entry_release(count);
38e34349 1182 rel_mplock();
a108bf71
MD
1183 return((void *)addr);
1184}
1185
38e34349 1186/*
5b287bba
MD
1187 * kmem_slab_free()
1188 *
1189 * MPALMOSTSAFE - acquires mplock
38e34349 1190 */
a108bf71
MD
1191static void
1192kmem_slab_free(void *ptr, vm_size_t size)
1193{
38e34349 1194 get_mplock();
a108bf71 1195 crit_enter();
e4846942 1196 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
a108bf71 1197 crit_exit();
38e34349 1198 rel_mplock();
a108bf71
MD
1199}
1200