1/*-
2 * Copyright (c) 2002, 2003, 2004, 2005 Jeffrey Roberson <jeff@FreeBSD.org>
3 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
4 * Copyright (c) 2004-2006 Robert N. M. Watson
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice unmodified, this list of conditions, and the following
12 * disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 */
28
29/*
30 * uma_core.c Implementation of the Universal Memory allocator
31 *
32 * This allocator is intended to replace the multitude of similar object caches
33 * in the standard FreeBSD kernel. The intent is to be flexible as well as
34 * effecient. A primary design goal is to return unused memory to the rest of
35 * the system. This will make the system as a whole more flexible due to the
36 * ability to move memory to subsystems which most need it instead of leaving
37 * pools of reserved memory unused.
38 *
39 * The basic ideas stem from similar slab/zone based allocators whose algorithms
40 * are well known.
41 *
42 */
43
44/*
45 * TODO:
46 * - Improve memory usage for large allocations
47 * - Investigate cache size adjustments
48 */
49
50#include <sys/cdefs.h>
51__FBSDID("$FreeBSD: head/sys/vm/uma_core.c 166213 2007-01-25 01:05:23Z mohans $");
52
53/* I should really use ktr.. */
54/*
55#define UMA_DEBUG 1
56#define UMA_DEBUG_ALLOC 1
57#define UMA_DEBUG_ALLOC_1 1
58*/
59
60#include "opt_ddb.h"
61#include "opt_param.h"
62
63#include <sys/param.h>
64#include <sys/systm.h>
65#include <sys/kernel.h>
66#include <sys/types.h>
67#include <sys/queue.h>
68#include <sys/malloc.h>
69#include <sys/ktr.h>
70#include <sys/lock.h>
71#include <sys/sysctl.h>
72#include <sys/mutex.h>
73#include <sys/proc.h>
74#include <sys/sbuf.h>
75#include <sys/smp.h>
76#include <sys/vmmeter.h>
77
78#include <vm/vm.h>
79#include <vm/vm_object.h>
80#include <vm/vm_page.h>
81#include <vm/vm_param.h>
82#include <vm/vm_map.h>
83#include <vm/vm_kern.h>
84#include <vm/vm_extern.h>
85#include <vm/uma.h>
86#include <vm/uma_int.h>
87#include <vm/uma_dbg.h>
88
89#include <machine/vmparam.h>
90
91#include <ddb/ddb.h>
92
93/*
94 * This is the zone and keg from which all zones are spawned. The idea is that
95 * even the zone & keg heads are allocated from the allocator, so we use the
96 * bss section to bootstrap us.
97 */
98static struct uma_keg masterkeg;
99static struct uma_zone masterzone_k;
100static struct uma_zone masterzone_z;
101static uma_zone_t kegs = &masterzone_k;
102static uma_zone_t zones = &masterzone_z;
103
104/* This is the zone from which all of uma_slab_t's are allocated. */
105static uma_zone_t slabzone;
106static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
107
108/*
109 * The initial hash tables come out of this zone so they can be allocated
110 * prior to malloc coming up.
111 */
112static uma_zone_t hashzone;
113
114static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
115
116/*
117 * Are we allowed to allocate buckets?
118 */
119static int bucketdisable = 1;
120
121/* Linked list of all kegs in the system */
122static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(&uma_kegs);
123
124/* This mutex protects the keg list */
125static struct mtx uma_mtx;
126
127/* Linked list of boot time pages */
128static LIST_HEAD(,uma_slab) uma_boot_pages =
129 LIST_HEAD_INITIALIZER(&uma_boot_pages);
130
131/* This mutex protects the boot time pages list */
132static struct mtx uma_boot_pages_mtx;
133
134/* Is the VM done starting up? */
135static int booted = 0;
136
137/* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
138static u_int uma_max_ipers;
139static u_int uma_max_ipers_ref;
140
141/*
142 * This is the handle used to schedule events that need to happen
143 * outside of the allocation fast path.
144 */
145static struct callout uma_callout;
146#define UMA_TIMEOUT 20 /* Seconds for callout interval. */
147
148/*
149 * This structure is passed as the zone ctor arg so that I don't have to create
150 * a special allocation function just for zones.
151 */
152struct uma_zctor_args {
153 char *name;
154 size_t size;
155 uma_ctor ctor;
156 uma_dtor dtor;
157 uma_init uminit;
158 uma_fini fini;
159 uma_keg_t keg;
160 int align;
161 u_int32_t flags;
162};
163
164struct uma_kctor_args {
165 uma_zone_t zone;
166 size_t size;
167 uma_init uminit;
168 uma_fini fini;
169 int align;
170 u_int32_t flags;
171};
172
173struct uma_bucket_zone {
174 uma_zone_t ubz_zone;
175 char *ubz_name;
176 int ubz_entries;
177};
178
179#define BUCKET_MAX 128
180
181struct uma_bucket_zone bucket_zones[] = {
182 { NULL, "16 Bucket", 16 },
183 { NULL, "32 Bucket", 32 },
184 { NULL, "64 Bucket", 64 },
185 { NULL, "128 Bucket", 128 },
186 { NULL, NULL, 0}
187};
188
189#define BUCKET_SHIFT 4
190#define BUCKET_ZONES ((BUCKET_MAX >> BUCKET_SHIFT) + 1)
191
192/*
193 * bucket_size[] maps requested bucket sizes to zones that allocate a bucket
194 * of approximately the right size.
195 */
196static uint8_t bucket_size[BUCKET_ZONES];
197
198/*
199 * Flags and enumerations to be passed to internal functions.
200 */
201enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI };
202
203#define ZFREE_STATFAIL 0x00000001 /* Update zone failure statistic. */
204#define ZFREE_STATFREE 0x00000002 /* Update zone free statistic. */
205
206/* Prototypes.. */
207
208static void *obj_alloc(uma_zone_t, int, u_int8_t *, int);
209static void *page_alloc(uma_zone_t, int, u_int8_t *, int);
210static void *startup_alloc(uma_zone_t, int, u_int8_t *, int);
211static void page_free(void *, int, u_int8_t);
212static uma_slab_t slab_zalloc(uma_zone_t, int);
213static void cache_drain(uma_zone_t);
214static void bucket_drain(uma_zone_t, uma_bucket_t);
215static void bucket_cache_drain(uma_zone_t zone);
216static int keg_ctor(void *, int, void *, int);
217static void keg_dtor(void *, int, void *);
218static int zone_ctor(void *, int, void *, int);
219static void zone_dtor(void *, int, void *);
220static int zero_init(void *, int, int);
221static void zone_small_init(uma_zone_t zone);
222static void zone_large_init(uma_zone_t zone);
223static void zone_foreach(void (*zfunc)(uma_zone_t));
224static void zone_timeout(uma_zone_t zone);
225static int hash_alloc(struct uma_hash *);
226static int hash_expand(struct uma_hash *, struct uma_hash *);
227static void hash_free(struct uma_hash *hash);
228static void uma_timeout(void *);
229static void uma_startup3(void);
230static void *uma_zalloc_internal(uma_zone_t, void *, int);
231static void uma_zfree_internal(uma_zone_t, void *, void *, enum zfreeskip,
232 int);
233static void bucket_enable(void);
234static void bucket_init(void);
235static uma_bucket_t bucket_alloc(int, int);
236static void bucket_free(uma_bucket_t);
237static void bucket_zone_drain(void);
238static int uma_zalloc_bucket(uma_zone_t zone, int flags);
239static uma_slab_t uma_zone_slab(uma_zone_t zone, int flags);
240static void *uma_slab_alloc(uma_zone_t zone, uma_slab_t slab);
241static uma_zone_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
242 uma_fini fini, int align, u_int32_t flags);
243
244void uma_print_zone(uma_zone_t);
245void uma_print_stats(void);
246static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
247static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
248
249#ifdef WITNESS
250static int nosleepwithlocks = 1;
251#else
252static int nosleepwithlocks = 0;
253#endif
254SYSCTL_INT(_debug, OID_AUTO, nosleepwithlocks, CTLFLAG_RW, &nosleepwithlocks,
255 0, "Convert M_WAITOK to M_NOWAIT to avoid lock-held-across-sleep paths");
256SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
257
258SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
259 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
260
261SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
262 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
263
264/*
265 * This routine checks to see whether or not it's safe to enable buckets.
266 */
267
268static void
269bucket_enable(void)
270{
271 if (cnt.v_free_count < cnt.v_free_min)
272 bucketdisable = 1;
273 else
274 bucketdisable = 0;
275}
276
277/*
278 * Initialize bucket_zones, the array of zones of buckets of various sizes.
279 *
280 * For each zone, calculate the memory required for each bucket, consisting
281 * of the header and an array of pointers. Initialize bucket_size[] to point
282 * the range of appropriate bucket sizes at the zone.
283 */
284static void
285bucket_init(void)
286{
287 struct uma_bucket_zone *ubz;
288 int i;
289 int j;
290
291 for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) {
292 int size;
293
294 ubz = &bucket_zones[j];
295 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
296 size += sizeof(void *) * ubz->ubz_entries;
297 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
298 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
299 for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT))
300 bucket_size[i >> BUCKET_SHIFT] = j;
301 }
302}
303
304/*
305 * Given a desired number of entries for a bucket, return the zone from which
306 * to allocate the bucket.
307 */
308static struct uma_bucket_zone *
309bucket_zone_lookup(int entries)
310{
311 int idx;
312
313 idx = howmany(entries, 1 << BUCKET_SHIFT);
314 return (&bucket_zones[bucket_size[idx]]);
315}
316
317static uma_bucket_t
318bucket_alloc(int entries, int bflags)
319{
320 struct uma_bucket_zone *ubz;
321 uma_bucket_t bucket;
322
323 /*
324 * This is to stop us from allocating per cpu buckets while we're
325 * running out of vm.boot_pages. Otherwise, we would exhaust the
326 * boot pages. This also prevents us from allocating buckets in
327 * low memory situations.
328 */
329 if (bucketdisable)
330 return (NULL);
331
332 ubz = bucket_zone_lookup(entries);
333 bucket = uma_zalloc_internal(ubz->ubz_zone, NULL, bflags);
334 if (bucket) {
335#ifdef INVARIANTS
336 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
337#endif
338 bucket->ub_cnt = 0;
339 bucket->ub_entries = ubz->ubz_entries;
340 }
341
342 return (bucket);
343}
344
345static void
346bucket_free(uma_bucket_t bucket)
347{
348 struct uma_bucket_zone *ubz;
349
350 ubz = bucket_zone_lookup(bucket->ub_entries);
351 uma_zfree_internal(ubz->ubz_zone, bucket, NULL, SKIP_NONE,
352 ZFREE_STATFREE);
353}
354
355static void
356bucket_zone_drain(void)
357{
358 struct uma_bucket_zone *ubz;
359
360 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
361 zone_drain(ubz->ubz_zone);
362}
363
364
365/*
366 * Routine called by timeout which is used to fire off some time interval
367 * based calculations. (stats, hash size, etc.)
368 *
369 * Arguments:
370 * arg Unused
371 *
372 * Returns:
373 * Nothing
374 */
375static void
376uma_timeout(void *unused)
377{
378 bucket_enable();
379 zone_foreach(zone_timeout);
380
381 /* Reschedule this event */
382 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
383}
384
385/*
386 * Routine to perform timeout driven calculations. This expands the
387 * hashes and does per cpu statistics aggregation.
388 *
389 * Arguments:
390 * zone The zone to operate on
391 *
392 * Returns:
393 * Nothing
394 */
395static void
396zone_timeout(uma_zone_t zone)
397{
398 uma_keg_t keg;
399 u_int64_t alloc;
400
401 keg = zone->uz_keg;
402 alloc = 0;
403
404 /*
405 * Expand the zone hash table.
406 *
407 * This is done if the number of slabs is larger than the hash size.
408 * What I'm trying to do here is completely reduce collisions. This
409 * may be a little aggressive. Should I allow for two collisions max?
410 */
411 ZONE_LOCK(zone);
412 if (keg->uk_flags & UMA_ZONE_HASH &&
413 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
414 struct uma_hash newhash;
415 struct uma_hash oldhash;
416 int ret;
417
418 /*
419 * This is so involved because allocating and freeing
420 * while the zone lock is held will lead to deadlock.
421 * I have to do everything in stages and check for
422 * races.
423 */
424 newhash = keg->uk_hash;
425 ZONE_UNLOCK(zone);
426 ret = hash_alloc(&newhash);
427 ZONE_LOCK(zone);
428 if (ret) {
429 if (hash_expand(&keg->uk_hash, &newhash)) {
430 oldhash = keg->uk_hash;
431 keg->uk_hash = newhash;
432 } else
433 oldhash = newhash;
434
435 ZONE_UNLOCK(zone);
436 hash_free(&oldhash);
437 ZONE_LOCK(zone);
438 }
439 }
440 ZONE_UNLOCK(zone);
441}
442
443/*
444 * Allocate and zero fill the next sized hash table from the appropriate
445 * backing store.
446 *
447 * Arguments:
448 * hash A new hash structure with the old hash size in uh_hashsize
449 *
450 * Returns:
451 * 1 on sucess and 0 on failure.
452 */
453static int
454hash_alloc(struct uma_hash *hash)
455{
456 int oldsize;
457 int alloc;
458
459 oldsize = hash->uh_hashsize;
460
461 /* We're just going to go to a power of two greater */
462 if (oldsize) {
463 hash->uh_hashsize = oldsize * 2;
464 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
465 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
466 M_UMAHASH, M_NOWAIT);
467 } else {
468 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
469 hash->uh_slab_hash = uma_zalloc_internal(hashzone, NULL,
470 M_WAITOK);
471 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
472 }
473 if (hash->uh_slab_hash) {
474 bzero(hash->uh_slab_hash, alloc);
475 hash->uh_hashmask = hash->uh_hashsize - 1;
476 return (1);
477 }
478
479 return (0);
480}
481
482/*
483 * Expands the hash table for HASH zones. This is done from zone_timeout
484 * to reduce collisions. This must not be done in the regular allocation
485 * path, otherwise, we can recurse on the vm while allocating pages.
486 *
487 * Arguments:
488 * oldhash The hash you want to expand
489 * newhash The hash structure for the new table
490 *
491 * Returns:
492 * Nothing
493 *
494 * Discussion:
495 */
496static int
497hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
498{
499 uma_slab_t slab;
500 int hval;
501 int i;
502
503 if (!newhash->uh_slab_hash)
504 return (0);
505
506 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
507 return (0);
508
509 /*
510 * I need to investigate hash algorithms for resizing without a
511 * full rehash.
512 */
513
514 for (i = 0; i < oldhash->uh_hashsize; i++)
515 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
516 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
517 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
518 hval = UMA_HASH(newhash, slab->us_data);
519 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
520 slab, us_hlink);
521 }
522
523 return (1);
524}
525
526/*
527 * Free the hash bucket to the appropriate backing store.
528 *
529 * Arguments:
530 * slab_hash The hash bucket we're freeing
531 * hashsize The number of entries in that hash bucket
532 *
533 * Returns:
534 * Nothing
535 */
536static void
537hash_free(struct uma_hash *hash)
538{
539 if (hash->uh_slab_hash == NULL)
540 return;
541 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
542 uma_zfree_internal(hashzone,
543 hash->uh_slab_hash, NULL, SKIP_NONE, ZFREE_STATFREE);
544 else
545 free(hash->uh_slab_hash, M_UMAHASH);
546}
547
548/*
549 * Frees all outstanding items in a bucket
550 *
551 * Arguments:
552 * zone The zone to free to, must be unlocked.
553 * bucket The free/alloc bucket with items, cpu queue must be locked.
554 *
555 * Returns:
556 * Nothing
557 */
558
559static void
560bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
561{
562 uma_slab_t slab;
563 int mzone;
564 void *item;
565
566 if (bucket == NULL)
567 return;
568
569 slab = NULL;
570 mzone = 0;
571
572 /* We have to lookup the slab again for malloc.. */
573 if (zone->uz_keg->uk_flags & UMA_ZONE_MALLOC)
574 mzone = 1;
575
576 while (bucket->ub_cnt > 0) {
577 bucket->ub_cnt--;
578 item = bucket->ub_bucket[bucket->ub_cnt];
579#ifdef INVARIANTS
580 bucket->ub_bucket[bucket->ub_cnt] = NULL;
581 KASSERT(item != NULL,
582 ("bucket_drain: botched ptr, item is NULL"));
583#endif
584 /*
585 * This is extremely inefficient. The slab pointer was passed
586 * to uma_zfree_arg, but we lost it because the buckets don't
587 * hold them. This will go away when free() gets a size passed
588 * to it.
589 */
590 if (mzone)
591 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
592 uma_zfree_internal(zone, item, slab, SKIP_DTOR, 0);
593 }
594}
595
596/*
597 * Drains the per cpu caches for a zone.
598 *
599 * NOTE: This may only be called while the zone is being turn down, and not
600 * during normal operation. This is necessary in order that we do not have
601 * to migrate CPUs to drain the per-CPU caches.
602 *
603 * Arguments:
604 * zone The zone to drain, must be unlocked.
605 *
606 * Returns:
607 * Nothing
608 */
609static void
610cache_drain(uma_zone_t zone)
611{
612 uma_cache_t cache;
613 int cpu;
614
615 /*
616 * XXX: It is safe to not lock the per-CPU caches, because we're
617 * tearing down the zone anyway. I.e., there will be no further use
618 * of the caches at this point.
619 *
620 * XXX: It would good to be able to assert that the zone is being
621 * torn down to prevent improper use of cache_drain().
622 *
623 * XXX: We lock the zone before passing into bucket_cache_drain() as
624 * it is used elsewhere. Should the tear-down path be made special
625 * there in some form?
626 */
627 for (cpu = 0; cpu <= mp_maxid; cpu++) {
628 if (CPU_ABSENT(cpu))
629 continue;
630 cache = &zone->uz_cpu[cpu];
631 bucket_drain(zone, cache->uc_allocbucket);
632 bucket_drain(zone, cache->uc_freebucket);
633 if (cache->uc_allocbucket != NULL)
634 bucket_free(cache->uc_allocbucket);
635 if (cache->uc_freebucket != NULL)
636 bucket_free(cache->uc_freebucket);
637 cache->uc_allocbucket = cache->uc_freebucket = NULL;
638 }
639 ZONE_LOCK(zone);
640 bucket_cache_drain(zone);
641 ZONE_UNLOCK(zone);
642}
643
644/*
645 * Drain the cached buckets from a zone. Expects a locked zone on entry.
646 */
647static void
648bucket_cache_drain(uma_zone_t zone)
649{
650 uma_bucket_t bucket;
651
652 /*
653 * Drain the bucket queues and free the buckets, we just keep two per
654 * cpu (alloc/free).
655 */
656 while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
657 LIST_REMOVE(bucket, ub_link);
658 ZONE_UNLOCK(zone);
659 bucket_drain(zone, bucket);
660 bucket_free(bucket);
661 ZONE_LOCK(zone);
662 }
663
664 /* Now we do the free queue.. */
665 while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
666 LIST_REMOVE(bucket, ub_link);
667 bucket_free(bucket);
668 }
669}
670
671/*
672 * Frees pages from a zone back to the system. This is done on demand from
673 * the pageout daemon.
674 *
675 * Arguments:
676 * zone The zone to free pages from
677 * all Should we drain all items?
678 *
679 * Returns:
680 * Nothing.
681 */
682void
683zone_drain(uma_zone_t zone)
684{
685 struct slabhead freeslabs = { 0 };
686 uma_keg_t keg;
687 uma_slab_t slab;
688 uma_slab_t n;
689 u_int8_t flags;
690 u_int8_t *mem;
691 int i;
692
693 keg = zone->uz_keg;
694
695 /*
696 * We don't want to take pages from statically allocated zones at this
697 * time
698 */
699 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
700 return;
701
702 ZONE_LOCK(zone);
703
704#ifdef UMA_DEBUG
705 printf("%s free items: %u\n", zone->uz_name, keg->uk_free);
706#endif
707 bucket_cache_drain(zone);
708 if (keg->uk_free == 0)
709 goto finished;
710
711 slab = LIST_FIRST(&keg->uk_free_slab);
712 while (slab) {
713 n = LIST_NEXT(slab, us_link);
714
715 /* We have no where to free these to */
716 if (slab->us_flags & UMA_SLAB_BOOT) {
717 slab = n;
718 continue;
719 }
720
721 LIST_REMOVE(slab, us_link);
722 keg->uk_pages -= keg->uk_ppera;
723 keg->uk_free -= keg->uk_ipers;
724
725 if (keg->uk_flags & UMA_ZONE_HASH)
726 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
727
728 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
729
730 slab = n;
731 }
732finished:
733 ZONE_UNLOCK(zone);
734
735 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
736 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
737 if (keg->uk_fini)
738 for (i = 0; i < keg->uk_ipers; i++)
739 keg->uk_fini(
740 slab->us_data + (keg->uk_rsize * i),
741 keg->uk_size);
742 flags = slab->us_flags;
743 mem = slab->us_data;
744
745 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
746 (keg->uk_flags & UMA_ZONE_REFCNT)) {
747 vm_object_t obj;
748
749 if (flags & UMA_SLAB_KMEM)
750 obj = kmem_object;
751 else
752 obj = NULL;
753 for (i = 0; i < keg->uk_ppera; i++)
754 vsetobj((vm_offset_t)mem + (i * PAGE_SIZE),
755 obj);
756 }
757 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
758 uma_zfree_internal(keg->uk_slabzone, slab, NULL,
759 SKIP_NONE, ZFREE_STATFREE);
760#ifdef UMA_DEBUG
761 printf("%s: Returning %d bytes.\n",
762 zone->uz_name, UMA_SLAB_SIZE * keg->uk_ppera);
763#endif
764 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags);
765 }
766}
767
768/*
769 * Allocate a new slab for a zone. This does not insert the slab onto a list.
770 *
771 * Arguments:
772 * zone The zone to allocate slabs for
773 * wait Shall we wait?
774 *
775 * Returns:
776 * The slab that was allocated or NULL if there is no memory and the
777 * caller specified M_NOWAIT.
778 */
779static uma_slab_t
780slab_zalloc(uma_zone_t zone, int wait)
781{
782 uma_slabrefcnt_t slabref;
783 uma_slab_t slab;
784 uma_keg_t keg;
785 u_int8_t *mem;
786 u_int8_t flags;
787 int i;
788
789 slab = NULL;
790 keg = zone->uz_keg;
791
792#ifdef UMA_DEBUG
793 printf("slab_zalloc: Allocating a new slab for %s\n", zone->uz_name);
794#endif
795 ZONE_UNLOCK(zone);
796
797 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
798 slab = uma_zalloc_internal(keg->uk_slabzone, NULL, wait);
799 if (slab == NULL) {
800 ZONE_LOCK(zone);
801 return NULL;
802 }
803 }
804
805 /*
806 * This reproduces the old vm_zone behavior of zero filling pages the
807 * first time they are added to a zone.
808 *
809 * Malloced items are zeroed in uma_zalloc.
810 */
811
812 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
813 wait |= M_ZERO;
814 else
815 wait &= ~M_ZERO;
816
817 mem = keg->uk_allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE,
818 &flags, wait);
819 if (mem == NULL) {
820 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
821 uma_zfree_internal(keg->uk_slabzone, slab, NULL,
822 SKIP_NONE, ZFREE_STATFREE);
823 ZONE_LOCK(zone);
824 return (NULL);
825 }
826
827 /* Point the slab into the allocated memory */
828 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
829 slab = (uma_slab_t )(mem + keg->uk_pgoff);
830
831 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
832 (keg->uk_flags & UMA_ZONE_REFCNT))
833 for (i = 0; i < keg->uk_ppera; i++)
834 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
835
836 slab->us_keg = keg;
837 slab->us_data = mem;
838 slab->us_freecount = keg->uk_ipers;
839 slab->us_firstfree = 0;
840 slab->us_flags = flags;
841
842 if (keg->uk_flags & UMA_ZONE_REFCNT) {
843 slabref = (uma_slabrefcnt_t)slab;
844 for (i = 0; i < keg->uk_ipers; i++) {
845 slabref->us_freelist[i].us_refcnt = 0;
846 slabref->us_freelist[i].us_item = i+1;
847 }
848 } else {
849 for (i = 0; i < keg->uk_ipers; i++)
850 slab->us_freelist[i].us_item = i+1;
851 }
852
853 if (keg->uk_init != NULL) {
854 for (i = 0; i < keg->uk_ipers; i++)
855 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
856 keg->uk_size, wait) != 0)
857 break;
858 if (i != keg->uk_ipers) {
859 if (keg->uk_fini != NULL) {
860 for (i--; i > -1; i--)
861 keg->uk_fini(slab->us_data +
862 (keg->uk_rsize * i),
863 keg->uk_size);
864 }
865 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
866 (keg->uk_flags & UMA_ZONE_REFCNT)) {
867 vm_object_t obj;
868
869 if (flags & UMA_SLAB_KMEM)
870 obj = kmem_object;
871 else
872 obj = NULL;
873 for (i = 0; i < keg->uk_ppera; i++)
874 vsetobj((vm_offset_t)mem +
875 (i * PAGE_SIZE), obj);
876 }
877 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
878 uma_zfree_internal(keg->uk_slabzone, slab,
879 NULL, SKIP_NONE, ZFREE_STATFREE);
880 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera,
881 flags);
882 ZONE_LOCK(zone);
883 return (NULL);
884 }
885 }
886 ZONE_LOCK(zone);
887
888 if (keg->uk_flags & UMA_ZONE_HASH)
889 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
890
891 keg->uk_pages += keg->uk_ppera;
892 keg->uk_free += keg->uk_ipers;
893
894 return (slab);
895}
896
897/*
898 * This function is intended to be used early on in place of page_alloc() so
899 * that we may use the boot time page cache to satisfy allocations before
900 * the VM is ready.
901 */
902static void *
903startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
904{
905 uma_keg_t keg;
906 uma_slab_t tmps;
907
908 keg = zone->uz_keg;
909
910 /*
911 * Check our small startup cache to see if it has pages remaining.
912 */
913 mtx_lock(&uma_boot_pages_mtx);
914 if ((tmps = LIST_FIRST(&uma_boot_pages)) != NULL) {
915 LIST_REMOVE(tmps, us_link);
916 mtx_unlock(&uma_boot_pages_mtx);
917 *pflag = tmps->us_flags;
918 return (tmps->us_data);
919 }
920 mtx_unlock(&uma_boot_pages_mtx);
921 if (booted == 0)
922 panic("UMA: Increase vm.boot_pages");
923 /*
924 * Now that we've booted reset these users to their real allocator.
925 */
926#ifdef UMA_MD_SMALL_ALLOC
927 keg->uk_allocf = uma_small_alloc;
928#else
929 keg->uk_allocf = page_alloc;
930#endif
931 return keg->uk_allocf(zone, bytes, pflag, wait);
932}
933
934/*
935 * Allocates a number of pages from the system
936 *
937 * Arguments:
938 * zone Unused
939 * bytes The number of bytes requested
940 * wait Shall we wait?
941 *
942 * Returns:
943 * A pointer to the alloced memory or possibly
944 * NULL if M_NOWAIT is set.
945 */
946static void *
947page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
948{
949 void *p; /* Returned page */
950
951 *pflag = UMA_SLAB_KMEM;
952 p = (void *) kmem_malloc(kmem_map, bytes, wait);
953
954 return (p);
955}
956
957/*
958 * Allocates a number of pages from within an object
959 *
960 * Arguments:
961 * zone Unused
962 * bytes The number of bytes requested
963 * wait Shall we wait?
964 *
965 * Returns:
966 * A pointer to the alloced memory or possibly
967 * NULL if M_NOWAIT is set.
968 */
969static void *
970obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
971{
972 vm_object_t object;
973 vm_offset_t retkva, zkva;
974 vm_page_t p;
975 int pages, startpages;
976
977 object = zone->uz_keg->uk_obj;
978 retkva = 0;
979
980 /*
981 * This looks a little weird since we're getting one page at a time.
982 */
983 VM_OBJECT_LOCK(object);
984 p = TAILQ_LAST(&object->memq, pglist);
985 pages = p != NULL ? p->pindex + 1 : 0;
986 startpages = pages;
987 zkva = zone->uz_keg->uk_kva + pages * PAGE_SIZE;
988 for (; bytes > 0; bytes -= PAGE_SIZE) {
989 p = vm_page_alloc(object, pages,
990 VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED);
991 if (p == NULL) {
992 if (pages != startpages)
993 pmap_qremove(retkva, pages - startpages);
994 while (pages != startpages) {
995 pages--;
996 p = TAILQ_LAST(&object->memq, pglist);
997 vm_page_lock_queues();
998 vm_page_unwire(p, 0);
999 vm_page_free(p);
1000 vm_page_unlock_queues();
1001 }
1002 retkva = 0;
1003 goto done;
1004 }
1005 pmap_qenter(zkva, &p, 1);
1006 if (retkva == 0)
1007 retkva = zkva;
1008 zkva += PAGE_SIZE;
1009 pages += 1;
1010 }
1011done:
1012 VM_OBJECT_UNLOCK(object);
1013 *flags = UMA_SLAB_PRIV;
1014
1015 return ((void *)retkva);
1016}
1017
1018/*
1019 * Frees a number of pages to the system
1020 *
1021 * Arguments:
1022 * mem A pointer to the memory to be freed
1023 * size The size of the memory being freed
1024 * flags The original p->us_flags field
1025 *
1026 * Returns:
1027 * Nothing
1028 */
1029static void
1030page_free(void *mem, int size, u_int8_t flags)
1031{
1032 vm_map_t map;
1033
1034 if (flags & UMA_SLAB_KMEM)
1035 map = kmem_map;
1036 else
1037 panic("UMA: page_free used with invalid flags %d\n", flags);
1038
1039 kmem_free(map, (vm_offset_t)mem, size);
1040}
1041
1042/*
1043 * Zero fill initializer
1044 *
1045 * Arguments/Returns follow uma_init specifications
1046 */
1047static int
1048zero_init(void *mem, int size, int flags)
1049{
1050 bzero(mem, size);
1051 return (0);
1052}
1053
1054/*
1055 * Finish creating a small uma zone. This calculates ipers, and the zone size.
1056 *
1057 * Arguments
1058 * zone The zone we should initialize
1059 *
1060 * Returns
1061 * Nothing
1062 */
1063static void
1064zone_small_init(uma_zone_t zone)
1065{
1066 uma_keg_t keg;
1067 u_int rsize;
1068 u_int memused;
1069 u_int wastedspace;
1070 u_int shsize;
1071
1072 keg = zone->uz_keg;
1073 KASSERT(keg != NULL, ("Keg is null in zone_small_init"));
1074 rsize = keg->uk_size;
1075
1076 if (rsize < UMA_SMALLEST_UNIT)
1077 rsize = UMA_SMALLEST_UNIT;
1078 if (rsize & keg->uk_align)
1079 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1080
1081 keg->uk_rsize = rsize;
1082 keg->uk_ppera = 1;
1083
1084 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1085 rsize += UMA_FRITMREF_SZ; /* linkage & refcnt */
1086 shsize = sizeof(struct uma_slab_refcnt);
1087 } else {
1088 rsize += UMA_FRITM_SZ; /* Account for linkage */
1089 shsize = sizeof(struct uma_slab);
1090 }
1091
1092 keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize;
1093 KASSERT(keg->uk_ipers != 0, ("zone_small_init: ipers is 0"));
1094 memused = keg->uk_ipers * rsize + shsize;
1095 wastedspace = UMA_SLAB_SIZE - memused;
1096
1097 /*
1098 * We can't do OFFPAGE if we're internal or if we've been
1099 * asked to not go to the VM for buckets. If we do this we
1100 * may end up going to the VM (kmem_map) for slabs which we
1101 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1102 * result of UMA_ZONE_VM, which clearly forbids it.
1103 */
1104 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1105 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1106 return;
1107
1108 if ((wastedspace >= UMA_MAX_WASTE) &&
1109 (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) {
1110 keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize;
1111 KASSERT(keg->uk_ipers <= 255,
1112 ("zone_small_init: keg->uk_ipers too high!"));
1113#ifdef UMA_DEBUG
1114 printf("UMA decided we need offpage slab headers for "
1115 "zone: %s, calculated wastedspace = %d, "
1116 "maximum wasted space allowed = %d, "
1117 "calculated ipers = %d, "
1118 "new wasted space = %d\n", zone->uz_name, wastedspace,
1119 UMA_MAX_WASTE, keg->uk_ipers,
1120 UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize);
1121#endif
1122 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1123 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1124 keg->uk_flags |= UMA_ZONE_HASH;
1125 }
1126}
1127
1128/*
1129 * Finish creating a large (> UMA_SLAB_SIZE) uma zone. Just give in and do
1130 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1131 * more complicated.
1132 *
1133 * Arguments
1134 * zone The zone we should initialize
1135 *
1136 * Returns
1137 * Nothing
1138 */
1139static void
1140zone_large_init(uma_zone_t zone)
1141{
1142 uma_keg_t keg;
1143 int pages;
1144
1145 keg = zone->uz_keg;
1146
1147 KASSERT(keg != NULL, ("Keg is null in zone_large_init"));
1148 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1149 ("zone_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY zone"));
1150
1151 pages = keg->uk_size / UMA_SLAB_SIZE;
1152
1153 /* Account for remainder */
1154 if ((pages * UMA_SLAB_SIZE) < keg->uk_size)
1155 pages++;
1156
1157 keg->uk_ppera = pages;
1158 keg->uk_ipers = 1;
1159
1160 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1161 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1162 keg->uk_flags |= UMA_ZONE_HASH;
1163
1164 keg->uk_rsize = keg->uk_size;
1165}
1166
1167/*
1168 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1169 * the keg onto the global keg list.
1170 *
1171 * Arguments/Returns follow uma_ctor specifications
1172 * udata Actually uma_kctor_args
1173 */
1174static int
1175keg_ctor(void *mem, int size, void *udata, int flags)
1176{
1177 struct uma_kctor_args *arg = udata;
1178 uma_keg_t keg = mem;
1179 uma_zone_t zone;
1180
1181 bzero(keg, size);
1182 keg->uk_size = arg->size;
1183 keg->uk_init = arg->uminit;
1184 keg->uk_fini = arg->fini;
1185 keg->uk_align = arg->align;
1186 keg->uk_free = 0;
1187 keg->uk_pages = 0;
1188 keg->uk_flags = arg->flags;
1189 keg->uk_allocf = page_alloc;
1190 keg->uk_freef = page_free;
1191 keg->uk_recurse = 0;
1192 keg->uk_slabzone = NULL;
1193
1194 /*
1195 * The master zone is passed to us at keg-creation time.
1196 */
1197 zone = arg->zone;
1198 zone->uz_keg = keg;
1199
1200 if (arg->flags & UMA_ZONE_VM)
1201 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1202
1203 if (arg->flags & UMA_ZONE_ZINIT)
1204 keg->uk_init = zero_init;
1205
1206 /*
1207 * The +UMA_FRITM_SZ added to uk_size is to account for the
1208 * linkage that is added to the size in zone_small_init(). If
1209 * we don't account for this here then we may end up in
1210 * zone_small_init() with a calculated 'ipers' of 0.
1211 */
1212 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1213 if ((keg->uk_size+UMA_FRITMREF_SZ) >
1214 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)))
1215 zone_large_init(zone);
1216 else
1217 zone_small_init(zone);
1218 } else {
1219 if ((keg->uk_size+UMA_FRITM_SZ) >
1220 (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1221 zone_large_init(zone);
1222 else
1223 zone_small_init(zone);
1224 }
1225
1226 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1227 if (keg->uk_flags & UMA_ZONE_REFCNT)
1228 keg->uk_slabzone = slabrefzone;
1229 else
1230 keg->uk_slabzone = slabzone;
1231 }
1232
1233 /*
1234 * If we haven't booted yet we need allocations to go through the
1235 * startup cache until the vm is ready.
1236 */
1237 if (keg->uk_ppera == 1) {
1238#ifdef UMA_MD_SMALL_ALLOC
1239 keg->uk_allocf = uma_small_alloc;
1240 keg->uk_freef = uma_small_free;
1241#endif
1242 if (booted == 0)
1243 keg->uk_allocf = startup_alloc;
1244 }
1245
1246 /*
1247 * Initialize keg's lock (shared among zones) through
1248 * Master zone
1249 */
1250 zone->uz_lock = &keg->uk_lock;
1251 if (arg->flags & UMA_ZONE_MTXCLASS)
1252 ZONE_LOCK_INIT(zone, 1);
1253 else
1254 ZONE_LOCK_INIT(zone, 0);
1255
1256 /*
1257 * If we're putting the slab header in the actual page we need to
1258 * figure out where in each page it goes. This calculates a right
1259 * justified offset into the memory on an ALIGN_PTR boundary.
1260 */
1261 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1262 u_int totsize;
1263
1264 /* Size of the slab struct and free list */
1265 if (keg->uk_flags & UMA_ZONE_REFCNT)
1266 totsize = sizeof(struct uma_slab_refcnt) +
1267 keg->uk_ipers * UMA_FRITMREF_SZ;
1268 else
1269 totsize = sizeof(struct uma_slab) +
1270 keg->uk_ipers * UMA_FRITM_SZ;
1271
1272 if (totsize & UMA_ALIGN_PTR)
1273 totsize = (totsize & ~UMA_ALIGN_PTR) +
1274 (UMA_ALIGN_PTR + 1);
1275 keg->uk_pgoff = UMA_SLAB_SIZE - totsize;
1276
1277 if (keg->uk_flags & UMA_ZONE_REFCNT)
1278 totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt)
1279 + keg->uk_ipers * UMA_FRITMREF_SZ;
1280 else
1281 totsize = keg->uk_pgoff + sizeof(struct uma_slab)
1282 + keg->uk_ipers * UMA_FRITM_SZ;
1283
1284 /*
1285 * The only way the following is possible is if with our
1286 * UMA_ALIGN_PTR adjustments we are now bigger than
1287 * UMA_SLAB_SIZE. I haven't checked whether this is
1288 * mathematically possible for all cases, so we make
1289 * sure here anyway.
1290 */
1291 if (totsize > UMA_SLAB_SIZE) {
1292 printf("zone %s ipers %d rsize %d size %d\n",
1293 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1294 keg->uk_size);
1295 panic("UMA slab won't fit.\n");
1296 }
1297 }
1298
1299 if (keg->uk_flags & UMA_ZONE_HASH)
1300 hash_alloc(&keg->uk_hash);
1301
1302#ifdef UMA_DEBUG
1303 printf("%s(%p) size = %d ipers = %d ppera = %d pgoff = %d\n",
1304 zone->uz_name, zone,
1305 keg->uk_size, keg->uk_ipers,
1306 keg->uk_ppera, keg->uk_pgoff);
1307#endif
1308
1309 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1310
1311 mtx_lock(&uma_mtx);
1312 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1313 mtx_unlock(&uma_mtx);
1314 return (0);
1315}
1316
1317/*
1318 * Zone header ctor. This initializes all fields, locks, etc.
1319 *
1320 * Arguments/Returns follow uma_ctor specifications
1321 * udata Actually uma_zctor_args
1322 */
1323
1324static int
1325zone_ctor(void *mem, int size, void *udata, int flags)
1326{
1327 struct uma_zctor_args *arg = udata;
1328 uma_zone_t zone = mem;
1329 uma_zone_t z;
1330 uma_keg_t keg;
1331
1332 bzero(zone, size);
1333 zone->uz_name = arg->name;
1334 zone->uz_ctor = arg->ctor;
1335 zone->uz_dtor = arg->dtor;
1336 zone->uz_init = NULL;
1337 zone->uz_fini = NULL;
1338 zone->uz_allocs = 0;
1339 zone->uz_frees = 0;
1340 zone->uz_fails = 0;
1341 zone->uz_fills = zone->uz_count = 0;
1342
1343 if (arg->flags & UMA_ZONE_SECONDARY) {
1344 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1345 keg = arg->keg;
1346 zone->uz_keg = keg;
1347 zone->uz_init = arg->uminit;
1348 zone->uz_fini = arg->fini;
1349 zone->uz_lock = &keg->uk_lock;
1350 mtx_lock(&uma_mtx);
1351 ZONE_LOCK(zone);
1352 keg->uk_flags |= UMA_ZONE_SECONDARY;
1353 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1354 if (LIST_NEXT(z, uz_link) == NULL) {
1355 LIST_INSERT_AFTER(z, zone, uz_link);
1356 break;
1357 }
1358 }
1359 ZONE_UNLOCK(zone);
1360 mtx_unlock(&uma_mtx);
1361 } else if (arg->keg == NULL) {
1362 if (uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1363 arg->align, arg->flags) == NULL)
1364 return (ENOMEM);
1365 } else {
1366 struct uma_kctor_args karg;
1367 int error;
1368
1369 /* We should only be here from uma_startup() */
1370 karg.size = arg->size;
1371 karg.uminit = arg->uminit;
1372 karg.fini = arg->fini;
1373 karg.align = arg->align;
1374 karg.flags = arg->flags;
1375 karg.zone = zone;
1376 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1377 flags);
1378 if (error)
1379 return (error);
1380 }
1381 keg = zone->uz_keg;
1382 zone->uz_lock = &keg->uk_lock;
1383
1384 /*
1385 * Some internal zones don't have room allocated for the per cpu
1386 * caches. If we're internal, bail out here.
1387 */
1388 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1389 KASSERT((keg->uk_flags & UMA_ZONE_SECONDARY) == 0,
1390 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1391 return (0);
1392 }
1393
1394 if (keg->uk_flags & UMA_ZONE_MAXBUCKET)
1395 zone->uz_count = BUCKET_MAX;
1396 else if (keg->uk_ipers <= BUCKET_MAX)
1397 zone->uz_count = keg->uk_ipers;
1398 else
1399 zone->uz_count = BUCKET_MAX;
1400 return (0);
1401}
1402
1403/*
1404 * Keg header dtor. This frees all data, destroys locks, frees the hash
1405 * table and removes the keg from the global list.
1406 *
1407 * Arguments/Returns follow uma_dtor specifications
1408 * udata unused
1409 */
1410static void
1411keg_dtor(void *arg, int size, void *udata)
1412{
1413 uma_keg_t keg;
1414
1415 keg = (uma_keg_t)arg;
1416 mtx_lock(&keg->uk_lock);
1417 if (keg->uk_free != 0) {
1418 printf("Freed UMA keg was not empty (%d items). "
1419 " Lost %d pages of memory.\n",
1420 keg->uk_free, keg->uk_pages);
1421 }
1422 mtx_unlock(&keg->uk_lock);
1423
1424 if (keg->uk_flags & UMA_ZONE_HASH)
1425 hash_free(&keg->uk_hash);
1426
1427 mtx_destroy(&keg->uk_lock);
1428}
1429
1430/*
1431 * Zone header dtor.
1432 *
1433 * Arguments/Returns follow uma_dtor specifications
1434 * udata unused
1435 */
1436static void
1437zone_dtor(void *arg, int size, void *udata)
1438{
1439 uma_zone_t zone;
1440 uma_keg_t keg;
1441
1442 zone = (uma_zone_t)arg;
1443 keg = zone->uz_keg;
1444
1445 if (!(keg->uk_flags & UMA_ZFLAG_INTERNAL))
1446 cache_drain(zone);
1447
1448 mtx_lock(&uma_mtx);
1449 zone_drain(zone);
1450 if (keg->uk_flags & UMA_ZONE_SECONDARY) {
1451 LIST_REMOVE(zone, uz_link);
1452 /*
1453 * XXX there are some races here where
1454 * the zone can be drained but zone lock
1455 * released and then refilled before we
1456 * remove it... we dont care for now
1457 */
1458 ZONE_LOCK(zone);
1459 if (LIST_EMPTY(&keg->uk_zones))
1460 keg->uk_flags &= ~UMA_ZONE_SECONDARY;
1461 ZONE_UNLOCK(zone);
1462 mtx_unlock(&uma_mtx);
1463 } else {
1464 LIST_REMOVE(keg, uk_link);
1465 LIST_REMOVE(zone, uz_link);
1466 mtx_unlock(&uma_mtx);
1467 uma_zfree_internal(kegs, keg, NULL, SKIP_NONE,
1468 ZFREE_STATFREE);
1469 }
1470 zone->uz_keg = NULL;
1471}
1472
1473/*
1474 * Traverses every zone in the system and calls a callback
1475 *
1476 * Arguments:
1477 * zfunc A pointer to a function which accepts a zone
1478 * as an argument.
1479 *
1480 * Returns:
1481 * Nothing
1482 */
1483static void
1484zone_foreach(void (*zfunc)(uma_zone_t))
1485{
1486 uma_keg_t keg;
1487 uma_zone_t zone;
1488
1489 mtx_lock(&uma_mtx);
1490 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1491 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1492 zfunc(zone);
1493 }
1494 mtx_unlock(&uma_mtx);
1495}
1496
1497/* Public functions */
1498/* See uma.h */
1499void
1500uma_startup(void *bootmem, int boot_pages)
1501{
1502 struct uma_zctor_args args;
1503 uma_slab_t slab;
1504 u_int slabsize;
1505 u_int objsize, totsize, wsize;
1506 int i;
1507
1508#ifdef UMA_DEBUG
1509 printf("Creating uma keg headers zone and keg.\n");
1510#endif
1511 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1512
1513 /*
1514 * Figure out the maximum number of items-per-slab we'll have if
1515 * we're using the OFFPAGE slab header to track free items, given
1516 * all possible object sizes and the maximum desired wastage
1517 * (UMA_MAX_WASTE).
1518 *
1519 * We iterate until we find an object size for
1520 * which the calculated wastage in zone_small_init() will be
1521 * enough to warrant OFFPAGE. Since wastedspace versus objsize
1522 * is an overall increasing see-saw function, we find the smallest
1523 * objsize such that the wastage is always acceptable for objects
1524 * with that objsize or smaller. Since a smaller objsize always
1525 * generates a larger possible uma_max_ipers, we use this computed
1526 * objsize to calculate the largest ipers possible. Since the
1527 * ipers calculated for OFFPAGE slab headers is always larger than
1528 * the ipers initially calculated in zone_small_init(), we use
1529 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to
1530 * obtain the maximum ipers possible for offpage slab headers.
1531 *
1532 * It should be noted that ipers versus objsize is an inversly
1533 * proportional function which drops off rather quickly so as
1534 * long as our UMA_MAX_WASTE is such that the objsize we calculate
1535 * falls into the portion of the inverse relation AFTER the steep
1536 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386).
1537 *
1538 * Note that we have 8-bits (1 byte) to use as a freelist index
1539 * inside the actual slab header itself and this is enough to
1540 * accomodate us. In the worst case, a UMA_SMALLEST_UNIT sized
1541 * object with offpage slab header would have ipers =
1542 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is
1543 * 1 greater than what our byte-integer freelist index can
1544 * accomodate, but we know that this situation never occurs as
1545 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate
1546 * that we need to go to offpage slab headers. Or, if we do,
1547 * then we trap that condition below and panic in the INVARIANTS case.
1548 */
1549 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE;
1550 totsize = wsize;
1551 objsize = UMA_SMALLEST_UNIT;
1552 while (totsize >= wsize) {
1553 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) /
1554 (objsize + UMA_FRITM_SZ);
1555 totsize *= (UMA_FRITM_SZ + objsize);
1556 objsize++;
1557 }
1558 if (objsize > UMA_SMALLEST_UNIT)
1559 objsize--;
1560 uma_max_ipers = UMA_SLAB_SIZE / objsize;
1561
1562 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE;
1563 totsize = wsize;
1564 objsize = UMA_SMALLEST_UNIT;
1565 while (totsize >= wsize) {
1566 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) /
1567 (objsize + UMA_FRITMREF_SZ);
1568 totsize *= (UMA_FRITMREF_SZ + objsize);
1569 objsize++;
1570 }
1571 if (objsize > UMA_SMALLEST_UNIT)
1572 objsize--;
1573 uma_max_ipers_ref = UMA_SLAB_SIZE / objsize;
1574
1575 KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255),
1576 ("uma_startup: calculated uma_max_ipers values too large!"));
1577
1578#ifdef UMA_DEBUG
1579 printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers);
1580 printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n",
1581 uma_max_ipers_ref);
1582#endif
1583
1584 /* "manually" create the initial zone */
1585 args.name = "UMA Kegs";
1586 args.size = sizeof(struct uma_keg);
1587 args.ctor = keg_ctor;
1588 args.dtor = keg_dtor;
1589 args.uminit = zero_init;
1590 args.fini = NULL;
1591 args.keg = &masterkeg;
1592 args.align = 32 - 1;
1593 args.flags = UMA_ZFLAG_INTERNAL;
1594 /* The initial zone has no Per cpu queues so it's smaller */
1595 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1596
1597#ifdef UMA_DEBUG
1598 printf("Filling boot free list.\n");
1599#endif
1600 for (i = 0; i < boot_pages; i++) {
1601 slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE));
1602 slab->us_data = (u_int8_t *)slab;
1603 slab->us_flags = UMA_SLAB_BOOT;
1604 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1605 }
1606 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1607
1608#ifdef UMA_DEBUG
1609 printf("Creating uma zone headers zone and keg.\n");
1610#endif
1611 args.name = "UMA Zones";
1612 args.size = sizeof(struct uma_zone) +
1613 (sizeof(struct uma_cache) * (mp_maxid + 1));
1614 args.ctor = zone_ctor;
1615 args.dtor = zone_dtor;
1616 args.uminit = zero_init;
1617 args.fini = NULL;
1618 args.keg = NULL;
1619 args.align = 32 - 1;
1620 args.flags = UMA_ZFLAG_INTERNAL;
1621 /* The initial zone has no Per cpu queues so it's smaller */
1622 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1623
1624#ifdef UMA_DEBUG
1625 printf("Initializing pcpu cache locks.\n");
1626#endif
1627#ifdef UMA_DEBUG
1628 printf("Creating slab and hash zones.\n");
1629#endif
1630
1631 /*
1632 * This is the max number of free list items we'll have with
1633 * offpage slabs.
1634 */
1635 slabsize = uma_max_ipers * UMA_FRITM_SZ;
1636 slabsize += sizeof(struct uma_slab);
1637
1638 /* Now make a zone for slab headers */
1639 slabzone = uma_zcreate("UMA Slabs",
1640 slabsize,
1641 NULL, NULL, NULL, NULL,
1642 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1643
1644 /*
1645 * We also create a zone for the bigger slabs with reference
1646 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1647 */
1648 slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ;
1649 slabsize += sizeof(struct uma_slab_refcnt);
1650 slabrefzone = uma_zcreate("UMA RCntSlabs",
1651 slabsize,
1652 NULL, NULL, NULL, NULL,
1653 UMA_ALIGN_PTR,
1654 UMA_ZFLAG_INTERNAL);
1655
1656 hashzone = uma_zcreate("UMA Hash",
1657 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1658 NULL, NULL, NULL, NULL,
1659 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1660
1661 bucket_init();
1662
1663#ifdef UMA_MD_SMALL_ALLOC
1664 booted = 1;
1665#endif
1666
1667#ifdef UMA_DEBUG
1668 printf("UMA startup complete.\n");
1669#endif
1670}
1671
1672/* see uma.h */
1673void
1674uma_startup2(void)
1675{
1676 booted = 1;
1677 bucket_enable();
1678#ifdef UMA_DEBUG
1679 printf("UMA startup2 complete.\n");
1680#endif
1681}
1682
1683/*
1684 * Initialize our callout handle
1685 *
1686 */
1687
1688static void
1689uma_startup3(void)
1690{
1691#ifdef UMA_DEBUG
1692 printf("Starting callout.\n");
1693#endif
1694 callout_init(&uma_callout, CALLOUT_MPSAFE);
1695 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1696#ifdef UMA_DEBUG
1697 printf("UMA startup3 complete.\n");
1698#endif
1699}
1700
1701static uma_zone_t
1702uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1703 int align, u_int32_t flags)
1704{
1705 struct uma_kctor_args args;
1706
1707 args.size = size;
1708 args.uminit = uminit;
1709 args.fini = fini;
1710 args.align = align;
1711 args.flags = flags;
1712 args.zone = zone;
1713 return (uma_zalloc_internal(kegs, &args, M_WAITOK));
1714}
1715
1716/* See uma.h */
1717uma_zone_t
1718uma_zcreate(char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1719 uma_init uminit, uma_fini fini, int align, u_int32_t flags)
1720
1721{
1722 struct uma_zctor_args args;
1723
1724 /* This stuff is essential for the zone ctor */
1725 args.name = name;
1726 args.size = size;
1727 args.ctor = ctor;
1728 args.dtor = dtor;
1729 args.uminit = uminit;
1730 args.fini = fini;
1731 args.align = align;
1732 args.flags = flags;
1733 args.keg = NULL;
1734
1735 return (uma_zalloc_internal(zones, &args, M_WAITOK));
1736}
1737
1738/* See uma.h */
1739uma_zone_t
1740uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1741 uma_init zinit, uma_fini zfini, uma_zone_t master)
1742{
1743 struct uma_zctor_args args;
1744
1745 args.name = name;
1746 args.size = master->uz_keg->uk_size;
1747 args.ctor = ctor;
1748 args.dtor = dtor;
1749 args.uminit = zinit;
1750 args.fini = zfini;
1751 args.align = master->uz_keg->uk_align;
1752 args.flags = master->uz_keg->uk_flags | UMA_ZONE_SECONDARY;
1753 args.keg = master->uz_keg;
1754
1755 return (uma_zalloc_internal(zones, &args, M_WAITOK));
1756}
1757
1758/* See uma.h */
1759void
1760uma_zdestroy(uma_zone_t zone)
1761{
1762
1763 uma_zfree_internal(zones, zone, NULL, SKIP_NONE, ZFREE_STATFREE);
1764}
1765
1766/* See uma.h */
1767void *
1768uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1769{
1770 void *item;
1771 uma_cache_t cache;
1772 uma_bucket_t bucket;
1773 int cpu;
1774
1775 /* This is the fast path allocation */
1776#ifdef UMA_DEBUG_ALLOC_1
1777 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1778#endif
1779 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1780 zone->uz_name, flags);
1781
1782 if (flags & M_WAITOK) {
1783 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1784 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
1785 }
1786
1787 /*
1788 * If possible, allocate from the per-CPU cache. There are two
1789 * requirements for safe access to the per-CPU cache: (1) the thread
1790 * accessing the cache must not be preempted or yield during access,
1791 * and (2) the thread must not migrate CPUs without switching which
1792 * cache it accesses. We rely on a critical section to prevent
1793 * preemption and migration. We release the critical section in
1794 * order to acquire the zone mutex if we are unable to allocate from
1795 * the current cache; when we re-acquire the critical section, we
1796 * must detect and handle migration if it has occurred.
1797 */
1798zalloc_restart:
1799 critical_enter();
1800 cpu = curcpu;
1801 cache = &zone->uz_cpu[cpu];
1802
1803zalloc_start:
1804 bucket = cache->uc_allocbucket;
1805
1806 if (bucket) {
1807 if (bucket->ub_cnt > 0) {
1808 bucket->ub_cnt--;
1809 item = bucket->ub_bucket[bucket->ub_cnt];
1810#ifdef INVARIANTS
1811 bucket->ub_bucket[bucket->ub_cnt] = NULL;
1812#endif
1813 KASSERT(item != NULL,
1814 ("uma_zalloc: Bucket pointer mangled."));
1815 cache->uc_allocs++;
1816 critical_exit();
1817#ifdef INVARIANTS
1818 ZONE_LOCK(zone);
1819 uma_dbg_alloc(zone, NULL, item);
1820 ZONE_UNLOCK(zone);
1821#endif
1822 if (zone->uz_ctor != NULL) {
1823 if (zone->uz_ctor(item, zone->uz_keg->uk_size,
1824 udata, flags) != 0) {
1825 uma_zfree_internal(zone, item, udata,
1826 SKIP_DTOR, ZFREE_STATFAIL |
1827 ZFREE_STATFREE);
1828 return (NULL);
1829 }
1830 }
1831 if (flags & M_ZERO)
1832 bzero(item, zone->uz_keg->uk_size);
1833 return (item);
1834 } else if (cache->uc_freebucket) {
1835 /*
1836 * We have run out of items in our allocbucket.
1837 * See if we can switch with our free bucket.
1838 */
1839 if (cache->uc_freebucket->ub_cnt > 0) {
1840#ifdef UMA_DEBUG_ALLOC
1841 printf("uma_zalloc: Swapping empty with"
1842 " alloc.\n");
1843#endif
1844 bucket = cache->uc_freebucket;
1845 cache->uc_freebucket = cache->uc_allocbucket;
1846 cache->uc_allocbucket = bucket;
1847
1848 goto zalloc_start;
1849 }
1850 }
1851 }
1852 /*
1853 * Attempt to retrieve the item from the per-CPU cache has failed, so
1854 * we must go back to the zone. This requires the zone lock, so we
1855 * must drop the critical section, then re-acquire it when we go back
1856 * to the cache. Since the critical section is released, we may be
1857 * preempted or migrate. As such, make sure not to maintain any
1858 * thread-local state specific to the cache from prior to releasing
1859 * the critical section.
1860 */
1861 critical_exit();
1862 ZONE_LOCK(zone);
1863 critical_enter();
1864 cpu = curcpu;
1865 cache = &zone->uz_cpu[cpu];
1866 bucket = cache->uc_allocbucket;
1867 if (bucket != NULL) {
1868 if (bucket->ub_cnt > 0) {
1869 ZONE_UNLOCK(zone);
1870 goto zalloc_start;
1871 }
1872 bucket = cache->uc_freebucket;
1873 if (bucket != NULL && bucket->ub_cnt > 0) {
1874 ZONE_UNLOCK(zone);
1875 goto zalloc_start;
1876 }
1877 }
1878
1879 /* Since we have locked the zone we may as well send back our stats */
1880 zone->uz_allocs += cache->uc_allocs;
1881 cache->uc_allocs = 0;
1882 zone->uz_frees += cache->uc_frees;
1883 cache->uc_frees = 0;
1884
1885 /* Our old one is now a free bucket */
1886 if (cache->uc_allocbucket) {
1887 KASSERT(cache->uc_allocbucket->ub_cnt == 0,
1888 ("uma_zalloc_arg: Freeing a non free bucket."));
1889 LIST_INSERT_HEAD(&zone->uz_free_bucket,
1890 cache->uc_allocbucket, ub_link);
1891 cache->uc_allocbucket = NULL;
1892 }
1893
1894 /* Check the free list for a new alloc bucket */
1895 if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
1896 KASSERT(bucket->ub_cnt != 0,
1897 ("uma_zalloc_arg: Returning an empty bucket."));
1898
1899 LIST_REMOVE(bucket, ub_link);
1900 cache->uc_allocbucket = bucket;
1901 ZONE_UNLOCK(zone);
1902 goto zalloc_start;
1903 }
1904 /* We are no longer associated with this CPU. */
1905 critical_exit();
1906
1907 /* Bump up our uz_count so we get here less */
1908 if (zone->uz_count < BUCKET_MAX)
1909 zone->uz_count++;
1910
1911 /*
1912 * Now lets just fill a bucket and put it on the free list. If that
1913 * works we'll restart the allocation from the begining.
1914 */
1915 if (uma_zalloc_bucket(zone, flags)) {
1916 ZONE_UNLOCK(zone);
1917 goto zalloc_restart;
1918 }
1919 ZONE_UNLOCK(zone);
1920 /*
1921 * We may not be able to get a bucket so return an actual item.
1922 */
1923#ifdef UMA_DEBUG
1924 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
1925#endif
1926
1927 return (uma_zalloc_internal(zone, udata, flags));
1928}
1929
1930static uma_slab_t
1931uma_zone_slab(uma_zone_t zone, int flags)
1932{
1933 uma_slab_t slab;
1934 uma_keg_t keg;
1935
1936 keg = zone->uz_keg;
1937
1938 /*
1939 * This is to prevent us from recursively trying to allocate
1940 * buckets. The problem is that if an allocation forces us to
1941 * grab a new bucket we will call page_alloc, which will go off
1942 * and cause the vm to allocate vm_map_entries. If we need new
1943 * buckets there too we will recurse in kmem_alloc and bad
1944 * things happen. So instead we return a NULL bucket, and make
1945 * the code that allocates buckets smart enough to deal with it
1946 *
1947 * XXX: While we want this protection for the bucket zones so that
1948 * recursion from the VM is handled (and the calling code that
1949 * allocates buckets knows how to deal with it), we do not want
1950 * to prevent allocation from the slab header zones (slabzone
1951 * and slabrefzone) if uk_recurse is not zero for them. The
1952 * reason is that it could lead to NULL being returned for
1953 * slab header allocations even in the M_WAITOK case, and the
1954 * caller can't handle that.
1955 */
1956 if (keg->uk_flags & UMA_ZFLAG_INTERNAL && keg->uk_recurse != 0)
1957 if ((zone != slabzone) && (zone != slabrefzone))
1958 return (NULL);
1959
1960 slab = NULL;
1961
1962 for (;;) {
1963 /*
1964 * Find a slab with some space. Prefer slabs that are partially
1965 * used over those that are totally full. This helps to reduce
1966 * fragmentation.
1967 */
1968 if (keg->uk_free != 0) {
1969 if (!LIST_EMPTY(&keg->uk_part_slab)) {
1970 slab = LIST_FIRST(&keg->uk_part_slab);
1971 } else {
1972 slab = LIST_FIRST(&keg->uk_free_slab);
1973 LIST_REMOVE(slab, us_link);
1974 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
1975 us_link);
1976 }
1977 return (slab);
1978 }
1979
1980 /*
1981 * M_NOVM means don't ask at all!
1982 */
1983 if (flags & M_NOVM)
1984 break;
1985
1986 if (keg->uk_maxpages &&
1987 keg->uk_pages >= keg->uk_maxpages) {
1988 keg->uk_flags |= UMA_ZFLAG_FULL;
1989
1990 if (flags & M_NOWAIT)
1991 break;
1992 else
1993 msleep(keg, &keg->uk_lock, PVM,
1994 "zonelimit", 0);
1995 continue;
1996 }
1997 keg->uk_recurse++;
1998 slab = slab_zalloc(zone, flags);
1999 keg->uk_recurse--;
2000
2001 /*
2002 * If we got a slab here it's safe to mark it partially used
2003 * and return. We assume that the caller is going to remove
2004 * at least one item.
2005 */
2006 if (slab) {
2007 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2008 return (slab);
2009 }
2010 /*
2011 * We might not have been able to get a slab but another cpu
2012 * could have while we were unlocked. Check again before we
2013 * fail.
2014 */
2015 if (flags & M_NOWAIT)
2016 flags |= M_NOVM;
2017 }
2018 return (slab);
2019}
2020
2021static void *
2022uma_slab_alloc(uma_zone_t zone, uma_slab_t slab)
2023{
2024 uma_keg_t keg;
2025 uma_slabrefcnt_t slabref;
2026 void *item;
2027 u_int8_t freei;
2028
2029 keg = zone->uz_keg;
2030
2031 freei = slab->us_firstfree;
2032 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2033 slabref = (uma_slabrefcnt_t)slab;
2034 slab->us_firstfree = slabref->us_freelist[freei].us_item;
2035 } else {
2036 slab->us_firstfree = slab->us_freelist[freei].us_item;
2037 }
2038 item = slab->us_data + (keg->uk_rsize * freei);
2039
2040 slab->us_freecount--;
2041 keg->uk_free--;
2042#ifdef INVARIANTS
2043 uma_dbg_alloc(zone, slab, item);
2044#endif
2045 /* Move this slab to the full list */
2046 if (slab->us_freecount == 0) {
2047 LIST_REMOVE(slab, us_link);
2048 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2049 }
2050
2051 return (item);
2052}
2053
2054static int
2055uma_zalloc_bucket(uma_zone_t zone, int flags)
2056{
2057 uma_bucket_t bucket;
2058 uma_slab_t slab;
2059 int16_t saved;
2060 int max, origflags = flags;
2061
2062 /*
2063 * Try this zone's free list first so we don't allocate extra buckets.
2064 */
2065 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2066 KASSERT(bucket->ub_cnt == 0,
2067 ("uma_zalloc_bucket: Bucket on free list is not empty."));
2068 LIST_REMOVE(bucket, ub_link);
2069 } else {
2070 int bflags;
2071
2072 bflags = (flags & ~M_ZERO);
2073 if (zone->uz_keg->uk_flags & UMA_ZFLAG_CACHEONLY)
2074 bflags |= M_NOVM;
2075
2076 ZONE_UNLOCK(zone);
2077 bucket = bucket_alloc(zone->uz_count, bflags);
2078 ZONE_LOCK(zone);
2079 }
2080
2081 if (bucket == NULL)
2082 return (0);
2083
2084#ifdef SMP
2085 /*
2086 * This code is here to limit the number of simultaneous bucket fills
2087 * for any given zone to the number of per cpu caches in this zone. This
2088 * is done so that we don't allocate more memory than we really need.
2089 */
2090 if (zone->uz_fills >= mp_ncpus)
2091 goto done;
2092
2093#endif
2094 zone->uz_fills++;
2095
2096 max = MIN(bucket->ub_entries, zone->uz_count);
2097 /* Try to keep the buckets totally full */
2098 saved = bucket->ub_cnt;
2099 while (bucket->ub_cnt < max &&
2100 (slab = uma_zone_slab(zone, flags)) != NULL) {
2101 while (slab->us_freecount && bucket->ub_cnt < max) {
2102 bucket->ub_bucket[bucket->ub_cnt++] =
2103 uma_slab_alloc(zone, slab);
2104 }
2105
2106 /* Don't block on the next fill */
2107 flags |= M_NOWAIT;
2108 }
2109
2110 /*
2111 * We unlock here because we need to call the zone's init.
2112 * It should be safe to unlock because the slab dealt with
2113 * above is already on the appropriate list within the keg
2114 * and the bucket we filled is not yet on any list, so we
2115 * own it.
2116 */
2117 if (zone->uz_init != NULL) {
2118 int i;
2119
2120 ZONE_UNLOCK(zone);
2121 for (i = saved; i < bucket->ub_cnt; i++)
2122 if (zone->uz_init(bucket->ub_bucket[i],
2123 zone->uz_keg->uk_size, origflags) != 0)
2124 break;
2125 /*
2126 * If we couldn't initialize the whole bucket, put the
2127 * rest back onto the freelist.
2128 */
2129 if (i != bucket->ub_cnt) {
2130 int j;
2131
2132 for (j = i; j < bucket->ub_cnt; j++) {
2133 uma_zfree_internal(zone, bucket->ub_bucket[j],
2134 NULL, SKIP_FINI, 0);
2135#ifdef INVARIANTS
2136 bucket->ub_bucket[j] = NULL;
2137#endif
2138 }
2139 bucket->ub_cnt = i;
2140 }
2141 ZONE_LOCK(zone);
2142 }
2143
2144 zone->uz_fills--;
2145 if (bucket->ub_cnt != 0) {
2146 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2147 bucket, ub_link);
2148 return (1);
2149 }
2150#ifdef SMP
2151done:
2152#endif
2153 bucket_free(bucket);
2154
2155 return (0);
2156}
2157/*
2158 * Allocates an item for an internal zone
2159 *
2160 * Arguments
2161 * zone The zone to alloc for.
2162 * udata The data to be passed to the constructor.
2163 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2164 *
2165 * Returns
2166 * NULL if there is no memory and M_NOWAIT is set
2167 * An item if successful
2168 */
2169
2170static void *
2171uma_zalloc_internal(uma_zone_t zone, void *udata, int flags)
2172{
2173 uma_keg_t keg;
2174 uma_slab_t slab;
2175 void *item;
2176
2177 item = NULL;
2178 keg = zone->uz_keg;
2179
2180#ifdef UMA_DEBUG_ALLOC
2181 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2182#endif
2183 ZONE_LOCK(zone);
2184
2185 slab = uma_zone_slab(zone, flags);
2186 if (slab == NULL) {
2187 zone->uz_fails++;
2188 ZONE_UNLOCK(zone);
2189 return (NULL);
2190 }
2191
2192 item = uma_slab_alloc(zone, slab);
2193
2194 zone->uz_allocs++;
2195
2196 ZONE_UNLOCK(zone);
2197
2198 /*
2199 * We have to call both the zone's init (not the keg's init)
2200 * and the zone's ctor. This is because the item is going from
2201 * a keg slab directly to the user, and the user is expecting it
2202 * to be both zone-init'd as well as zone-ctor'd.
2203 */
2204 if (zone->uz_init != NULL) {
2205 if (zone->uz_init(item, keg->uk_size, flags) != 0) {
2206 uma_zfree_internal(zone, item, udata, SKIP_FINI,
2207 ZFREE_STATFAIL | ZFREE_STATFREE);
2208 return (NULL);
2209 }
2210 }
2211 if (zone->uz_ctor != NULL) {
2212 if (zone->uz_ctor(item, keg->uk_size, udata, flags) != 0) {
2213 uma_zfree_internal(zone, item, udata, SKIP_DTOR,
2214 ZFREE_STATFAIL | ZFREE_STATFREE);
2215 return (NULL);
2216 }
2217 }
2218 if (flags & M_ZERO)
2219 bzero(item, keg->uk_size);
2220
2221 return (item);
2222}
2223
2224/* See uma.h */
2225void
2226uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2227{
2228 uma_keg_t keg;
2229 uma_cache_t cache;
2230 uma_bucket_t bucket;
2231 int bflags;
2232 int cpu;
2233
2234 keg = zone->uz_keg;
2235
2236#ifdef UMA_DEBUG_ALLOC_1
2237 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2238#endif
2239 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2240 zone->uz_name);
2241
2242 if (zone->uz_dtor)
2243 zone->uz_dtor(item, keg->uk_size, udata);
2244#ifdef INVARIANTS
2245 ZONE_LOCK(zone);
2246 if (keg->uk_flags & UMA_ZONE_MALLOC)
2247 uma_dbg_free(zone, udata, item);
2248 else
2249 uma_dbg_free(zone, NULL, item);
2250 ZONE_UNLOCK(zone);
2251#endif
2252 /*
2253 * The race here is acceptable. If we miss it we'll just have to wait
2254 * a little longer for the limits to be reset.
2255 */
2256 if (keg->uk_flags & UMA_ZFLAG_FULL)
2257 goto zfree_internal;
2258
2259 /*
2260 * If possible, free to the per-CPU cache. There are two
2261 * requirements for safe access to the per-CPU cache: (1) the thread
2262 * accessing the cache must not be preempted or yield during access,
2263 * and (2) the thread must not migrate CPUs without switching which
2264 * cache it accesses. We rely on a critical section to prevent
2265 * preemption and migration. We release the critical section in
2266 * order to acquire the zone mutex if we are unable to free to the
2267 * current cache; when we re-acquire the critical section, we must
2268 * detect and handle migration if it has occurred.
2269 */
2270zfree_restart:
2271 critical_enter();
2272 cpu = curcpu;
2273 cache = &zone->uz_cpu[cpu];
2274
2275zfree_start:
2276 bucket = cache->uc_freebucket;
2277
2278 if (bucket) {
2279 /*
2280 * Do we have room in our bucket? It is OK for this uz count
2281 * check to be slightly out of sync.
2282 */
2283
2284 if (bucket->ub_cnt < bucket->ub_entries) {
2285 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2286 ("uma_zfree: Freeing to non free bucket index."));
2287 bucket->ub_bucket[bucket->ub_cnt] = item;
2288 bucket->ub_cnt++;
2289 cache->uc_frees++;
2290 critical_exit();
2291 return;
2292 } else if (cache->uc_allocbucket) {
2293#ifdef UMA_DEBUG_ALLOC
2294 printf("uma_zfree: Swapping buckets.\n");
2295#endif
2296 /*
2297 * We have run out of space in our freebucket.
2298 * See if we can switch with our alloc bucket.
2299 */
2300 if (cache->uc_allocbucket->ub_cnt <
2301 cache->uc_freebucket->ub_cnt) {
2302 bucket = cache->uc_freebucket;
2303 cache->uc_freebucket = cache->uc_allocbucket;
2304 cache->uc_allocbucket = bucket;
2305 goto zfree_start;
2306 }
2307 }
2308 }
2309 /*
2310 * We can get here for two reasons:
2311 *
2312 * 1) The buckets are NULL
2313 * 2) The alloc and free buckets are both somewhat full.
2314 *
2315 * We must go back the zone, which requires acquiring the zone lock,
2316 * which in turn means we must release and re-acquire the critical
2317 * section. Since the critical section is released, we may be
2318 * preempted or migrate. As such, make sure not to maintain any
2319 * thread-local state specific to the cache from prior to releasing
2320 * the critical section.
2321 */
2322 critical_exit();
2323 ZONE_LOCK(zone);
2324 critical_enter();
2325 cpu = curcpu;
2326 cache = &zone->uz_cpu[cpu];
2327 if (cache->uc_freebucket != NULL) {
2328 if (cache->uc_freebucket->ub_cnt <
2329 cache->uc_freebucket->ub_entries) {
2330 ZONE_UNLOCK(zone);
2331 goto zfree_start;
2332 }
2333 if (cache->uc_allocbucket != NULL &&
2334 (cache->uc_allocbucket->ub_cnt <
2335 cache->uc_freebucket->ub_cnt)) {
2336 ZONE_UNLOCK(zone);
2337 goto zfree_start;
2338 }
2339 }
2340
2341 /* Since we have locked the zone we may as well send back our stats */
2342 zone->uz_allocs += cache->uc_allocs;
2343 cache->uc_allocs = 0;
2344 zone->uz_frees += cache->uc_frees;
2345 cache->uc_frees = 0;
2346
2347 bucket = cache->uc_freebucket;
2348 cache->uc_freebucket = NULL;
2349
2350 /* Can we throw this on the zone full list? */
2351 if (bucket != NULL) {
2352#ifdef UMA_DEBUG_ALLOC
2353 printf("uma_zfree: Putting old bucket on the free list.\n");
2354#endif
2355 /* ub_cnt is pointing to the last free item */
2356 KASSERT(bucket->ub_cnt != 0,
2357 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2358 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2359 bucket, ub_link);
2360 }
2361 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2362 LIST_REMOVE(bucket, ub_link);
2363 ZONE_UNLOCK(zone);
2364 cache->uc_freebucket = bucket;
2365 goto zfree_start;
2366 }
2367 /* We are no longer associated with this CPU. */
2368 critical_exit();
2369
2370 /* And the zone.. */
2371 ZONE_UNLOCK(zone);
2372
2373#ifdef UMA_DEBUG_ALLOC
2374 printf("uma_zfree: Allocating new free bucket.\n");
2375#endif
2376 bflags = M_NOWAIT;
2377
2378 if (keg->uk_flags & UMA_ZFLAG_CACHEONLY)
2379 bflags |= M_NOVM;
2380 bucket = bucket_alloc(zone->uz_count, bflags);
2381 if (bucket) {
2382 ZONE_LOCK(zone);
2383 LIST_INSERT_HEAD(&zone->uz_free_bucket,
2384 bucket, ub_link);
2385 ZONE_UNLOCK(zone);
2386 goto zfree_restart;
2387 }
2388
2389 /*
2390 * If nothing else caught this, we'll just do an internal free.
2391 */
2392zfree_internal:
2393 uma_zfree_internal(zone, item, udata, SKIP_DTOR, ZFREE_STATFREE);
2394
2395 return;
2396}
2397
2398/*
2399 * Frees an item to an INTERNAL zone or allocates a free bucket
2400 *
2401 * Arguments:
2402 * zone The zone to free to
2403 * item The item we're freeing
2404 * udata User supplied data for the dtor
2405 * skip Skip dtors and finis
2406 */
2407static void
2408uma_zfree_internal(uma_zone_t zone, void *item, void *udata,
2409 enum zfreeskip skip, int flags)
2410{
2411 uma_slab_t slab;
2412 uma_slabrefcnt_t slabref;
2413 uma_keg_t keg;
2414 u_int8_t *mem;
2415 u_int8_t freei;
2416
2417 keg = zone->uz_keg;
2418
2419 if (skip < SKIP_DTOR && zone->uz_dtor)
2420 zone->uz_dtor(item, keg->uk_size, udata);
2421 if (skip < SKIP_FINI && zone->uz_fini)
2422 zone->uz_fini(item, keg->uk_size);
2423
2424 ZONE_LOCK(zone);
2425
2426 if (flags & ZFREE_STATFAIL)
2427 zone->uz_fails++;
2428 if (flags & ZFREE_STATFREE)
2429 zone->uz_frees++;
2430
2431 if (!(keg->uk_flags & UMA_ZONE_MALLOC)) {
2432 mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK));
2433 if (keg->uk_flags & UMA_ZONE_HASH)
2434 slab = hash_sfind(&keg->uk_hash, mem);
2435 else {
2436 mem += keg->uk_pgoff;
2437 slab = (uma_slab_t)mem;
2438 }
2439 } else {
2440 slab = (uma_slab_t)udata;
2441 }
2442
2443 /* Do we need to remove from any lists? */
2444 if (slab->us_freecount+1 == keg->uk_ipers) {
2445 LIST_REMOVE(slab, us_link);
2446 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2447 } else if (slab->us_freecount == 0) {
2448 LIST_REMOVE(slab, us_link);
2449 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2450 }
2451
2452 /* Slab management stuff */
2453 freei = ((unsigned long)item - (unsigned long)slab->us_data)
2454 / keg->uk_rsize;
2455
2456#ifdef INVARIANTS
2457 if (!skip)
2458 uma_dbg_free(zone, slab, item);
2459#endif
2460
2461 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2462 slabref = (uma_slabrefcnt_t)slab;
2463 slabref->us_freelist[freei].us_item = slab->us_firstfree;
2464 } else {
2465 slab->us_freelist[freei].us_item = slab->us_firstfree;
2466 }
2467 slab->us_firstfree = freei;
2468 slab->us_freecount++;
2469
2470 /* Zone statistics */
2471 keg->uk_free++;
2472
2473 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2474 if (keg->uk_pages < keg->uk_maxpages)
2475 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2476
2477 /*
2478 * We can handle one more allocation. Since we're clearing ZFLAG_FULL,
2479 * wake up all procs blocked on pages. This should be uncommon, so
2480 * keeping this simple for now (rather than adding count of blocked
2481 * threads etc).
2482 */
2483 wakeup(keg);
2484 }
2485
2486 ZONE_UNLOCK(zone);
2487}
2488
2489/* See uma.h */
2490void
2491uma_zone_set_max(uma_zone_t zone, int nitems)
2492{
2493 uma_keg_t keg;
2494
2495 keg = zone->uz_keg;
2496 ZONE_LOCK(zone);
2497 if (keg->uk_ppera > 1)
2498 keg->uk_maxpages = nitems * keg->uk_ppera;
2499 else
2500 keg->uk_maxpages = nitems / keg->uk_ipers;
2501
2502 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2503 keg->uk_maxpages++;
2504
2505 ZONE_UNLOCK(zone);
2506}
2507
2508/* See uma.h */
2509void
2510uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2511{
2512 ZONE_LOCK(zone);
2513 KASSERT(zone->uz_keg->uk_pages == 0,
2514 ("uma_zone_set_init on non-empty keg"));
2515 zone->uz_keg->uk_init = uminit;
2516 ZONE_UNLOCK(zone);
2517}
2518
2519/* See uma.h */
2520void
2521uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2522{
2523 ZONE_LOCK(zone);
2524 KASSERT(zone->uz_keg->uk_pages == 0,
2525 ("uma_zone_set_fini on non-empty keg"));
2526 zone->uz_keg->uk_fini = fini;
2527 ZONE_UNLOCK(zone);
2528}
2529
2530/* See uma.h */
2531void
2532uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2533{
2534 ZONE_LOCK(zone);
2535 KASSERT(zone->uz_keg->uk_pages == 0,
2536 ("uma_zone_set_zinit on non-empty keg"));
2537 zone->uz_init = zinit;
2538 ZONE_UNLOCK(zone);
2539}
2540
2541/* See uma.h */
2542void
2543uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2544{
2545 ZONE_LOCK(zone);
2546 KASSERT(zone->uz_keg->uk_pages == 0,
2547 ("uma_zone_set_zfini on non-empty keg"));
2548 zone->uz_fini = zfini;
2549 ZONE_UNLOCK(zone);
2550}
2551
2552/* See uma.h */
2553/* XXX uk_freef is not actually used with the zone locked */
2554void
2555uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2556{
2557 ZONE_LOCK(zone);
2558 zone->uz_keg->uk_freef = freef;
2559 ZONE_UNLOCK(zone);
2560}
2561
2562/* See uma.h */
2563/* XXX uk_allocf is not actually used with the zone locked */
2564void
2565uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2566{
2567 ZONE_LOCK(zone);
2568 zone->uz_keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2569 zone->uz_keg->uk_allocf = allocf;
2570 ZONE_UNLOCK(zone);
2571}
2572
2573/* See uma.h */
2574int
2575uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count)
2576{
2577 uma_keg_t keg;
2578 vm_offset_t kva;
2579 int pages;
2580
2581 keg = zone->uz_keg;
2582 pages = count / keg->uk_ipers;
2583
2584 if (pages * keg->uk_ipers < count)
2585 pages++;
2586
2587 kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2588
2589 if (kva == 0)
2590 return (0);
2591 if (obj == NULL) {
2592 obj = vm_object_allocate(OBJT_DEFAULT,
2593 pages);
2594 } else {
2595 VM_OBJECT_LOCK_INIT(obj, "uma object");
2596 _vm_object_allocate(OBJT_DEFAULT,
2597 pages, obj);
2598 }
2599 ZONE_LOCK(zone);
2600 keg->uk_kva = kva;
2601 keg->uk_obj = obj;
2602 keg->uk_maxpages = pages;
2603 keg->uk_allocf = obj_alloc;
2604 keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
2605 ZONE_UNLOCK(zone);
2606 return (1);
2607}
2608
2609/* See uma.h */
2610void
2611uma_prealloc(uma_zone_t zone, int items)
2612{
2613 int slabs;
2614 uma_slab_t slab;
2615 uma_keg_t keg;
2616
2617 keg = zone->uz_keg;
2618 ZONE_LOCK(zone);
2619 slabs = items / keg->uk_ipers;
2620 if (slabs * keg->uk_ipers < items)
2621 slabs++;
2622 while (slabs > 0) {
2623 slab = slab_zalloc(zone, M_WAITOK);
2624 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2625 slabs--;
2626 }
2627 ZONE_UNLOCK(zone);
2628}
2629
2630/* See uma.h */
2631u_int32_t *
2632uma_find_refcnt(uma_zone_t zone, void *item)
2633{
2634 uma_slabrefcnt_t slabref;
2635 uma_keg_t keg;
2636 u_int32_t *refcnt;
2637 int idx;
2638
2639 keg = zone->uz_keg;
2640 slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item &
2641 (~UMA_SLAB_MASK));
2642 KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT,
2643 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
2644 idx = ((unsigned long)item - (unsigned long)slabref->us_data)
2645 / keg->uk_rsize;
2646 refcnt = &slabref->us_freelist[idx].us_refcnt;
2647 return refcnt;
2648}
2649
2650/* See uma.h */
2651void
2652uma_reclaim(void)
2653{
2654#ifdef UMA_DEBUG
2655 printf("UMA: vm asked us to release pages!\n");
2656#endif
2657 bucket_enable();
2658 zone_foreach(zone_drain);
2659 /*
2660 * Some slabs may have been freed but this zone will be visited early
2661 * we visit again so that we can free pages that are empty once other
2662 * zones are drained. We have to do the same for buckets.
2663 */
2664 zone_drain(slabzone);
2665 zone_drain(slabrefzone);
2666 bucket_zone_drain();
2667}
2668
2669/* See uma.h */
2670int
2671uma_zone_exhausted(uma_zone_t zone)
2672{
2673 int full;
2674
2675 ZONE_LOCK(zone);
2676 full = (zone->uz_keg->uk_flags & UMA_ZFLAG_FULL);
2677 ZONE_UNLOCK(zone);
2678 return (full);
2679}
2680
2681int
2682uma_zone_exhausted_nolock(uma_zone_t zone)
2683{
2684 return (zone->uz_keg->uk_flags & UMA_ZFLAG_FULL);
2685}
2686
2687void *
2688uma_large_malloc(int size, int wait)
2689{
2690 void *mem;
2691 uma_slab_t slab;
2692 u_int8_t flags;
2693
2694 slab = uma_zalloc_internal(slabzone, NULL, wait);
2695 if (slab == NULL)
2696 return (NULL);
2697 mem = page_alloc(NULL, size, &flags, wait);
2698 if (mem) {
2699 vsetslab((vm_offset_t)mem, slab);
2700 slab->us_data = mem;
2701 slab->us_flags = flags | UMA_SLAB_MALLOC;
2702 slab->us_size = size;
2703 } else {
2704 uma_zfree_internal(slabzone, slab, NULL, SKIP_NONE,
2705 ZFREE_STATFAIL | ZFREE_STATFREE);
2706 }
2707
2708 return (mem);
2709}
2710
2711void
2712uma_large_free(uma_slab_t slab)
2713{
2714 vsetobj((vm_offset_t)slab->us_data, kmem_object);
2715 page_free(slab->us_data, slab->us_size, slab->us_flags);
2716 uma_zfree_internal(slabzone, slab, NULL, SKIP_NONE, ZFREE_STATFREE);
2717}
2718
2719void
2720uma_print_stats(void)
2721{
2722 zone_foreach(uma_print_zone);
2723}
2724
2725static void
2726slab_print(uma_slab_t slab)
2727{
2728 printf("slab: keg %p, data %p, freecount %d, firstfree %d\n",
2729 slab->us_keg, slab->us_data, slab->us_freecount,
2730 slab->us_firstfree);
2731}
2732
2733static void
2734cache_print(uma_cache_t cache)
2735{
2736 printf("alloc: %p(%d), free: %p(%d)\n",
2737 cache->uc_allocbucket,
2738 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
2739 cache->uc_freebucket,
2740 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
2741}
2742
2743void
2744uma_print_zone(uma_zone_t zone)
2745{
2746 uma_cache_t cache;
2747 uma_keg_t keg;
2748 uma_slab_t slab;
2749 int i;
2750
2751 keg = zone->uz_keg;
2752 printf("%s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n",
2753 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
2754 keg->uk_ipers, keg->uk_ppera,
2755 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
2756 printf("Part slabs:\n");
2757 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
2758 slab_print(slab);
2759 printf("Free slabs:\n");
2760 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
2761 slab_print(slab);
2762 printf("Full slabs:\n");
2763 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
2764 slab_print(slab);
2765 for (i = 0; i <= mp_maxid; i++) {
2766 if (CPU_ABSENT(i))
2767 continue;
2768 cache = &zone->uz_cpu[i];
2769 printf("CPU %d Cache:\n", i);
2770 cache_print(cache);
2771 }
2772}
2773
2774#ifdef DDB
2775/*
2776 * Generate statistics across both the zone and its per-cpu cache's. Return
2777 * desired statistics if the pointer is non-NULL for that statistic.
2778 *
2779 * Note: does not update the zone statistics, as it can't safely clear the
2780 * per-CPU cache statistic.
2781 *
2782 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
2783 * safe from off-CPU; we should modify the caches to track this information
2784 * directly so that we don't have to.
2785 */
2786static void
2787uma_zone_sumstat(uma_zone_t z, int *cachefreep, u_int64_t *allocsp,
2788 u_int64_t *freesp)
2789{
2790 uma_cache_t cache;
2791 u_int64_t allocs, frees;
2792 int cachefree, cpu;
2793
2794 allocs = frees = 0;
2795 cachefree = 0;
2796 for (cpu = 0; cpu <= mp_maxid; cpu++) {
2797 if (CPU_ABSENT(cpu))
2798 continue;
2799 cache = &z->uz_cpu[cpu];
2800 if (cache->uc_allocbucket != NULL)
2801 cachefree += cache->uc_allocbucket->ub_cnt;
2802 if (cache->uc_freebucket != NULL)
2803 cachefree += cache->uc_freebucket->ub_cnt;
2804 allocs += cache->uc_allocs;
2805 frees += cache->uc_frees;
2806 }
2807 allocs += z->uz_allocs;
2808 frees += z->uz_frees;
2809 if (cachefreep != NULL)
2810 *cachefreep = cachefree;
2811 if (allocsp != NULL)
2812 *allocsp = allocs;
2813 if (freesp != NULL)
2814 *freesp = frees;
2815}
2816#endif /* DDB */
2817
2818static int
2819sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
2820{
2821 uma_keg_t kz;
2822 uma_zone_t z;
2823 int count;
2824
2825 count = 0;
2826 mtx_lock(&uma_mtx);
2827 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2828 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2829 count++;
2830 }
2831 mtx_unlock(&uma_mtx);
2832 return (sysctl_handle_int(oidp, &count, 0, req));
2833}
2834
2835static int
2836sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
2837{
2838 struct uma_stream_header ush;
2839 struct uma_type_header uth;
2840 struct uma_percpu_stat ups;
2841 uma_bucket_t bucket;
2842 struct sbuf sbuf;
2843 uma_cache_t cache;
2844 uma_keg_t kz;
2845 uma_zone_t z;
2846 char *buffer;
2847 int buflen, count, error, i;
2848
2849 mtx_lock(&uma_mtx);
2850restart:
2851 mtx_assert(&uma_mtx, MA_OWNED);
2852 count = 0;
2853 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2854 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2855 count++;
2856 }
2857 mtx_unlock(&uma_mtx);
2858
2859 buflen = sizeof(ush) + count * (sizeof(uth) + sizeof(ups) *
2860 (mp_maxid + 1)) + 1;
2861 buffer = malloc(buflen, M_TEMP, M_WAITOK | M_ZERO);
2862
2863 mtx_lock(&uma_mtx);
2864 i = 0;
2865 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2866 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2867 i++;
2868 }
2869 if (i > count) {
2870 free(buffer, M_TEMP);
2871 goto restart;
2872 }
2873 count = i;
2874
2875 sbuf_new(&sbuf, buffer, buflen, SBUF_FIXEDLEN);
2876
2877 /*
2878 * Insert stream header.
2879 */
2880 bzero(&ush, sizeof(ush));
2881 ush.ush_version = UMA_STREAM_VERSION;
2882 ush.ush_maxcpus = (mp_maxid + 1);
2883 ush.ush_count = count;
2884 if (sbuf_bcat(&sbuf, &ush, sizeof(ush)) < 0) {
2885 mtx_unlock(&uma_mtx);
2886 error = ENOMEM;
2887 goto out;
2888 }
2889
2890 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2891 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
2892 bzero(&uth, sizeof(uth));
2893 ZONE_LOCK(z);
2894 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
2895 uth.uth_align = kz->uk_align;
2896 uth.uth_pages = kz->uk_pages;
2897 uth.uth_keg_free = kz->uk_free;
2898 uth.uth_size = kz->uk_size;
2899 uth.uth_rsize = kz->uk_rsize;
2900 uth.uth_maxpages = kz->uk_maxpages;
2901 if (kz->uk_ppera > 1)
2902 uth.uth_limit = kz->uk_maxpages /
2903 kz->uk_ppera;
2904 else
2905 uth.uth_limit = kz->uk_maxpages *
2906 kz->uk_ipers;
2907
2908 /*
2909 * A zone is secondary is it is not the first entry
2910 * on the keg's zone list.
2911 */
2912 if ((kz->uk_flags & UMA_ZONE_SECONDARY) &&
2913 (LIST_FIRST(&kz->uk_zones) != z))
2914 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
2915
2916 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
2917 uth.uth_zone_free += bucket->ub_cnt;
2918 uth.uth_allocs = z->uz_allocs;
2919 uth.uth_frees = z->uz_frees;
2920 uth.uth_fails = z->uz_fails;
2921 if (sbuf_bcat(&sbuf, &uth, sizeof(uth)) < 0) {
2922 ZONE_UNLOCK(z);
2923 mtx_unlock(&uma_mtx);
2924 error = ENOMEM;
2925 goto out;
2926 }
2927 /*
2928 * While it is not normally safe to access the cache
2929 * bucket pointers while not on the CPU that owns the
2930 * cache, we only allow the pointers to be exchanged
2931 * without the zone lock held, not invalidated, so
2932 * accept the possible race associated with bucket
2933 * exchange during monitoring.
2934 */
2935 for (i = 0; i < (mp_maxid + 1); i++) {
2936 bzero(&ups, sizeof(ups));
2937 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
2938 goto skip;
2939 if (CPU_ABSENT(i))
2940 goto skip;
2941 cache = &z->uz_cpu[i];
2942 if (cache->uc_allocbucket != NULL)
2943 ups.ups_cache_free +=
2944 cache->uc_allocbucket->ub_cnt;
2945 if (cache->uc_freebucket != NULL)
2946 ups.ups_cache_free +=
2947 cache->uc_freebucket->ub_cnt;
2948 ups.ups_allocs = cache->uc_allocs;
2949 ups.ups_frees = cache->uc_frees;
2950skip:
2951 if (sbuf_bcat(&sbuf, &ups, sizeof(ups)) < 0) {
2952 ZONE_UNLOCK(z);
2953 mtx_unlock(&uma_mtx);
2954 error = ENOMEM;
2955 goto out;
2956 }
2957 }
2958 ZONE_UNLOCK(z);
2959 }
2960 }
2961 mtx_unlock(&uma_mtx);
2962 sbuf_finish(&sbuf);
2963 error = SYSCTL_OUT(req, sbuf_data(&sbuf), sbuf_len(&sbuf));
2964out:
2965 free(buffer, M_TEMP);
2966 return (error);
2967}
2968
2969#ifdef DDB
2970DB_SHOW_COMMAND(uma, db_show_uma)
2971{
2972 u_int64_t allocs, frees;
2973 uma_bucket_t bucket;
2974 uma_keg_t kz;
2975 uma_zone_t z;
2976 int cachefree;
2977
2978 db_printf("%18s %8s %8s %8s %12s\n", "Zone", "Size", "Used", "Free",
2979 "Requests");
2980 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2981 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
2982 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
2983 allocs = z->uz_allocs;
2984 frees = z->uz_frees;
2985 cachefree = 0;
2986 } else
2987 uma_zone_sumstat(z, &cachefree, &allocs,
2988 &frees);
2989 if (!((kz->uk_flags & UMA_ZONE_SECONDARY) &&
2990 (LIST_FIRST(&kz->uk_zones) != z)))
2991 cachefree += kz->uk_free;
2992 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
2993 cachefree += bucket->ub_cnt;
2994 db_printf("%18s %8ju %8jd %8d %12ju\n", z->uz_name,
2995 (uintmax_t)kz->uk_size,
2996 (intmax_t)(allocs - frees), cachefree,
2997 (uintmax_t)allocs);
2998 }
2999 }
3000}
3001#endif
2 * Copyright (c) 2002, 2003, 2004, 2005 Jeffrey Roberson <jeff@FreeBSD.org>
3 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
4 * Copyright (c) 2004-2006 Robert N. M. Watson
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice unmodified, this list of conditions, and the following
12 * disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 */
28
29/*
30 * uma_core.c Implementation of the Universal Memory allocator
31 *
32 * This allocator is intended to replace the multitude of similar object caches
33 * in the standard FreeBSD kernel. The intent is to be flexible as well as
34 * effecient. A primary design goal is to return unused memory to the rest of
35 * the system. This will make the system as a whole more flexible due to the
36 * ability to move memory to subsystems which most need it instead of leaving
37 * pools of reserved memory unused.
38 *
39 * The basic ideas stem from similar slab/zone based allocators whose algorithms
40 * are well known.
41 *
42 */
43
44/*
45 * TODO:
46 * - Improve memory usage for large allocations
47 * - Investigate cache size adjustments
48 */
49
50#include <sys/cdefs.h>
51__FBSDID("$FreeBSD: head/sys/vm/uma_core.c 166213 2007-01-25 01:05:23Z mohans $");
52
53/* I should really use ktr.. */
54/*
55#define UMA_DEBUG 1
56#define UMA_DEBUG_ALLOC 1
57#define UMA_DEBUG_ALLOC_1 1
58*/
59
60#include "opt_ddb.h"
61#include "opt_param.h"
62
63#include <sys/param.h>
64#include <sys/systm.h>
65#include <sys/kernel.h>
66#include <sys/types.h>
67#include <sys/queue.h>
68#include <sys/malloc.h>
69#include <sys/ktr.h>
70#include <sys/lock.h>
71#include <sys/sysctl.h>
72#include <sys/mutex.h>
73#include <sys/proc.h>
74#include <sys/sbuf.h>
75#include <sys/smp.h>
76#include <sys/vmmeter.h>
77
78#include <vm/vm.h>
79#include <vm/vm_object.h>
80#include <vm/vm_page.h>
81#include <vm/vm_param.h>
82#include <vm/vm_map.h>
83#include <vm/vm_kern.h>
84#include <vm/vm_extern.h>
85#include <vm/uma.h>
86#include <vm/uma_int.h>
87#include <vm/uma_dbg.h>
88
89#include <machine/vmparam.h>
90
91#include <ddb/ddb.h>
92
93/*
94 * This is the zone and keg from which all zones are spawned. The idea is that
95 * even the zone & keg heads are allocated from the allocator, so we use the
96 * bss section to bootstrap us.
97 */
98static struct uma_keg masterkeg;
99static struct uma_zone masterzone_k;
100static struct uma_zone masterzone_z;
101static uma_zone_t kegs = &masterzone_k;
102static uma_zone_t zones = &masterzone_z;
103
104/* This is the zone from which all of uma_slab_t's are allocated. */
105static uma_zone_t slabzone;
106static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
107
108/*
109 * The initial hash tables come out of this zone so they can be allocated
110 * prior to malloc coming up.
111 */
112static uma_zone_t hashzone;
113
114static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
115
116/*
117 * Are we allowed to allocate buckets?
118 */
119static int bucketdisable = 1;
120
121/* Linked list of all kegs in the system */
122static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(&uma_kegs);
123
124/* This mutex protects the keg list */
125static struct mtx uma_mtx;
126
127/* Linked list of boot time pages */
128static LIST_HEAD(,uma_slab) uma_boot_pages =
129 LIST_HEAD_INITIALIZER(&uma_boot_pages);
130
131/* This mutex protects the boot time pages list */
132static struct mtx uma_boot_pages_mtx;
133
134/* Is the VM done starting up? */
135static int booted = 0;
136
137/* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
138static u_int uma_max_ipers;
139static u_int uma_max_ipers_ref;
140
141/*
142 * This is the handle used to schedule events that need to happen
143 * outside of the allocation fast path.
144 */
145static struct callout uma_callout;
146#define UMA_TIMEOUT 20 /* Seconds for callout interval. */
147
148/*
149 * This structure is passed as the zone ctor arg so that I don't have to create
150 * a special allocation function just for zones.
151 */
152struct uma_zctor_args {
153 char *name;
154 size_t size;
155 uma_ctor ctor;
156 uma_dtor dtor;
157 uma_init uminit;
158 uma_fini fini;
159 uma_keg_t keg;
160 int align;
161 u_int32_t flags;
162};
163
164struct uma_kctor_args {
165 uma_zone_t zone;
166 size_t size;
167 uma_init uminit;
168 uma_fini fini;
169 int align;
170 u_int32_t flags;
171};
172
173struct uma_bucket_zone {
174 uma_zone_t ubz_zone;
175 char *ubz_name;
176 int ubz_entries;
177};
178
179#define BUCKET_MAX 128
180
181struct uma_bucket_zone bucket_zones[] = {
182 { NULL, "16 Bucket", 16 },
183 { NULL, "32 Bucket", 32 },
184 { NULL, "64 Bucket", 64 },
185 { NULL, "128 Bucket", 128 },
186 { NULL, NULL, 0}
187};
188
189#define BUCKET_SHIFT 4
190#define BUCKET_ZONES ((BUCKET_MAX >> BUCKET_SHIFT) + 1)
191
192/*
193 * bucket_size[] maps requested bucket sizes to zones that allocate a bucket
194 * of approximately the right size.
195 */
196static uint8_t bucket_size[BUCKET_ZONES];
197
198/*
199 * Flags and enumerations to be passed to internal functions.
200 */
201enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI };
202
203#define ZFREE_STATFAIL 0x00000001 /* Update zone failure statistic. */
204#define ZFREE_STATFREE 0x00000002 /* Update zone free statistic. */
205
206/* Prototypes.. */
207
208static void *obj_alloc(uma_zone_t, int, u_int8_t *, int);
209static void *page_alloc(uma_zone_t, int, u_int8_t *, int);
210static void *startup_alloc(uma_zone_t, int, u_int8_t *, int);
211static void page_free(void *, int, u_int8_t);
212static uma_slab_t slab_zalloc(uma_zone_t, int);
213static void cache_drain(uma_zone_t);
214static void bucket_drain(uma_zone_t, uma_bucket_t);
215static void bucket_cache_drain(uma_zone_t zone);
216static int keg_ctor(void *, int, void *, int);
217static void keg_dtor(void *, int, void *);
218static int zone_ctor(void *, int, void *, int);
219static void zone_dtor(void *, int, void *);
220static int zero_init(void *, int, int);
221static void zone_small_init(uma_zone_t zone);
222static void zone_large_init(uma_zone_t zone);
223static void zone_foreach(void (*zfunc)(uma_zone_t));
224static void zone_timeout(uma_zone_t zone);
225static int hash_alloc(struct uma_hash *);
226static int hash_expand(struct uma_hash *, struct uma_hash *);
227static void hash_free(struct uma_hash *hash);
228static void uma_timeout(void *);
229static void uma_startup3(void);
230static void *uma_zalloc_internal(uma_zone_t, void *, int);
231static void uma_zfree_internal(uma_zone_t, void *, void *, enum zfreeskip,
232 int);
233static void bucket_enable(void);
234static void bucket_init(void);
235static uma_bucket_t bucket_alloc(int, int);
236static void bucket_free(uma_bucket_t);
237static void bucket_zone_drain(void);
238static int uma_zalloc_bucket(uma_zone_t zone, int flags);
239static uma_slab_t uma_zone_slab(uma_zone_t zone, int flags);
240static void *uma_slab_alloc(uma_zone_t zone, uma_slab_t slab);
241static uma_zone_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
242 uma_fini fini, int align, u_int32_t flags);
243
244void uma_print_zone(uma_zone_t);
245void uma_print_stats(void);
246static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
247static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
248
249#ifdef WITNESS
250static int nosleepwithlocks = 1;
251#else
252static int nosleepwithlocks = 0;
253#endif
254SYSCTL_INT(_debug, OID_AUTO, nosleepwithlocks, CTLFLAG_RW, &nosleepwithlocks,
255 0, "Convert M_WAITOK to M_NOWAIT to avoid lock-held-across-sleep paths");
256SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
257
258SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
259 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
260
261SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
262 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
263
264/*
265 * This routine checks to see whether or not it's safe to enable buckets.
266 */
267
268static void
269bucket_enable(void)
270{
271 if (cnt.v_free_count < cnt.v_free_min)
272 bucketdisable = 1;
273 else
274 bucketdisable = 0;
275}
276
277/*
278 * Initialize bucket_zones, the array of zones of buckets of various sizes.
279 *
280 * For each zone, calculate the memory required for each bucket, consisting
281 * of the header and an array of pointers. Initialize bucket_size[] to point
282 * the range of appropriate bucket sizes at the zone.
283 */
284static void
285bucket_init(void)
286{
287 struct uma_bucket_zone *ubz;
288 int i;
289 int j;
290
291 for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) {
292 int size;
293
294 ubz = &bucket_zones[j];
295 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
296 size += sizeof(void *) * ubz->ubz_entries;
297 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
298 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
299 for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT))
300 bucket_size[i >> BUCKET_SHIFT] = j;
301 }
302}
303
304/*
305 * Given a desired number of entries for a bucket, return the zone from which
306 * to allocate the bucket.
307 */
308static struct uma_bucket_zone *
309bucket_zone_lookup(int entries)
310{
311 int idx;
312
313 idx = howmany(entries, 1 << BUCKET_SHIFT);
314 return (&bucket_zones[bucket_size[idx]]);
315}
316
317static uma_bucket_t
318bucket_alloc(int entries, int bflags)
319{
320 struct uma_bucket_zone *ubz;
321 uma_bucket_t bucket;
322
323 /*
324 * This is to stop us from allocating per cpu buckets while we're
325 * running out of vm.boot_pages. Otherwise, we would exhaust the
326 * boot pages. This also prevents us from allocating buckets in
327 * low memory situations.
328 */
329 if (bucketdisable)
330 return (NULL);
331
332 ubz = bucket_zone_lookup(entries);
333 bucket = uma_zalloc_internal(ubz->ubz_zone, NULL, bflags);
334 if (bucket) {
335#ifdef INVARIANTS
336 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
337#endif
338 bucket->ub_cnt = 0;
339 bucket->ub_entries = ubz->ubz_entries;
340 }
341
342 return (bucket);
343}
344
345static void
346bucket_free(uma_bucket_t bucket)
347{
348 struct uma_bucket_zone *ubz;
349
350 ubz = bucket_zone_lookup(bucket->ub_entries);
351 uma_zfree_internal(ubz->ubz_zone, bucket, NULL, SKIP_NONE,
352 ZFREE_STATFREE);
353}
354
355static void
356bucket_zone_drain(void)
357{
358 struct uma_bucket_zone *ubz;
359
360 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
361 zone_drain(ubz->ubz_zone);
362}
363
364
365/*
366 * Routine called by timeout which is used to fire off some time interval
367 * based calculations. (stats, hash size, etc.)
368 *
369 * Arguments:
370 * arg Unused
371 *
372 * Returns:
373 * Nothing
374 */
375static void
376uma_timeout(void *unused)
377{
378 bucket_enable();
379 zone_foreach(zone_timeout);
380
381 /* Reschedule this event */
382 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
383}
384
385/*
386 * Routine to perform timeout driven calculations. This expands the
387 * hashes and does per cpu statistics aggregation.
388 *
389 * Arguments:
390 * zone The zone to operate on
391 *
392 * Returns:
393 * Nothing
394 */
395static void
396zone_timeout(uma_zone_t zone)
397{
398 uma_keg_t keg;
399 u_int64_t alloc;
400
401 keg = zone->uz_keg;
402 alloc = 0;
403
404 /*
405 * Expand the zone hash table.
406 *
407 * This is done if the number of slabs is larger than the hash size.
408 * What I'm trying to do here is completely reduce collisions. This
409 * may be a little aggressive. Should I allow for two collisions max?
410 */
411 ZONE_LOCK(zone);
412 if (keg->uk_flags & UMA_ZONE_HASH &&
413 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
414 struct uma_hash newhash;
415 struct uma_hash oldhash;
416 int ret;
417
418 /*
419 * This is so involved because allocating and freeing
420 * while the zone lock is held will lead to deadlock.
421 * I have to do everything in stages and check for
422 * races.
423 */
424 newhash = keg->uk_hash;
425 ZONE_UNLOCK(zone);
426 ret = hash_alloc(&newhash);
427 ZONE_LOCK(zone);
428 if (ret) {
429 if (hash_expand(&keg->uk_hash, &newhash)) {
430 oldhash = keg->uk_hash;
431 keg->uk_hash = newhash;
432 } else
433 oldhash = newhash;
434
435 ZONE_UNLOCK(zone);
436 hash_free(&oldhash);
437 ZONE_LOCK(zone);
438 }
439 }
440 ZONE_UNLOCK(zone);
441}
442
443/*
444 * Allocate and zero fill the next sized hash table from the appropriate
445 * backing store.
446 *
447 * Arguments:
448 * hash A new hash structure with the old hash size in uh_hashsize
449 *
450 * Returns:
451 * 1 on sucess and 0 on failure.
452 */
453static int
454hash_alloc(struct uma_hash *hash)
455{
456 int oldsize;
457 int alloc;
458
459 oldsize = hash->uh_hashsize;
460
461 /* We're just going to go to a power of two greater */
462 if (oldsize) {
463 hash->uh_hashsize = oldsize * 2;
464 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
465 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
466 M_UMAHASH, M_NOWAIT);
467 } else {
468 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
469 hash->uh_slab_hash = uma_zalloc_internal(hashzone, NULL,
470 M_WAITOK);
471 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
472 }
473 if (hash->uh_slab_hash) {
474 bzero(hash->uh_slab_hash, alloc);
475 hash->uh_hashmask = hash->uh_hashsize - 1;
476 return (1);
477 }
478
479 return (0);
480}
481
482/*
483 * Expands the hash table for HASH zones. This is done from zone_timeout
484 * to reduce collisions. This must not be done in the regular allocation
485 * path, otherwise, we can recurse on the vm while allocating pages.
486 *
487 * Arguments:
488 * oldhash The hash you want to expand
489 * newhash The hash structure for the new table
490 *
491 * Returns:
492 * Nothing
493 *
494 * Discussion:
495 */
496static int
497hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
498{
499 uma_slab_t slab;
500 int hval;
501 int i;
502
503 if (!newhash->uh_slab_hash)
504 return (0);
505
506 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
507 return (0);
508
509 /*
510 * I need to investigate hash algorithms for resizing without a
511 * full rehash.
512 */
513
514 for (i = 0; i < oldhash->uh_hashsize; i++)
515 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
516 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
517 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
518 hval = UMA_HASH(newhash, slab->us_data);
519 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
520 slab, us_hlink);
521 }
522
523 return (1);
524}
525
526/*
527 * Free the hash bucket to the appropriate backing store.
528 *
529 * Arguments:
530 * slab_hash The hash bucket we're freeing
531 * hashsize The number of entries in that hash bucket
532 *
533 * Returns:
534 * Nothing
535 */
536static void
537hash_free(struct uma_hash *hash)
538{
539 if (hash->uh_slab_hash == NULL)
540 return;
541 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
542 uma_zfree_internal(hashzone,
543 hash->uh_slab_hash, NULL, SKIP_NONE, ZFREE_STATFREE);
544 else
545 free(hash->uh_slab_hash, M_UMAHASH);
546}
547
548/*
549 * Frees all outstanding items in a bucket
550 *
551 * Arguments:
552 * zone The zone to free to, must be unlocked.
553 * bucket The free/alloc bucket with items, cpu queue must be locked.
554 *
555 * Returns:
556 * Nothing
557 */
558
559static void
560bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
561{
562 uma_slab_t slab;
563 int mzone;
564 void *item;
565
566 if (bucket == NULL)
567 return;
568
569 slab = NULL;
570 mzone = 0;
571
572 /* We have to lookup the slab again for malloc.. */
573 if (zone->uz_keg->uk_flags & UMA_ZONE_MALLOC)
574 mzone = 1;
575
576 while (bucket->ub_cnt > 0) {
577 bucket->ub_cnt--;
578 item = bucket->ub_bucket[bucket->ub_cnt];
579#ifdef INVARIANTS
580 bucket->ub_bucket[bucket->ub_cnt] = NULL;
581 KASSERT(item != NULL,
582 ("bucket_drain: botched ptr, item is NULL"));
583#endif
584 /*
585 * This is extremely inefficient. The slab pointer was passed
586 * to uma_zfree_arg, but we lost it because the buckets don't
587 * hold them. This will go away when free() gets a size passed
588 * to it.
589 */
590 if (mzone)
591 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
592 uma_zfree_internal(zone, item, slab, SKIP_DTOR, 0);
593 }
594}
595
596/*
597 * Drains the per cpu caches for a zone.
598 *
599 * NOTE: This may only be called while the zone is being turn down, and not
600 * during normal operation. This is necessary in order that we do not have
601 * to migrate CPUs to drain the per-CPU caches.
602 *
603 * Arguments:
604 * zone The zone to drain, must be unlocked.
605 *
606 * Returns:
607 * Nothing
608 */
609static void
610cache_drain(uma_zone_t zone)
611{
612 uma_cache_t cache;
613 int cpu;
614
615 /*
616 * XXX: It is safe to not lock the per-CPU caches, because we're
617 * tearing down the zone anyway. I.e., there will be no further use
618 * of the caches at this point.
619 *
620 * XXX: It would good to be able to assert that the zone is being
621 * torn down to prevent improper use of cache_drain().
622 *
623 * XXX: We lock the zone before passing into bucket_cache_drain() as
624 * it is used elsewhere. Should the tear-down path be made special
625 * there in some form?
626 */
627 for (cpu = 0; cpu <= mp_maxid; cpu++) {
628 if (CPU_ABSENT(cpu))
629 continue;
630 cache = &zone->uz_cpu[cpu];
631 bucket_drain(zone, cache->uc_allocbucket);
632 bucket_drain(zone, cache->uc_freebucket);
633 if (cache->uc_allocbucket != NULL)
634 bucket_free(cache->uc_allocbucket);
635 if (cache->uc_freebucket != NULL)
636 bucket_free(cache->uc_freebucket);
637 cache->uc_allocbucket = cache->uc_freebucket = NULL;
638 }
639 ZONE_LOCK(zone);
640 bucket_cache_drain(zone);
641 ZONE_UNLOCK(zone);
642}
643
644/*
645 * Drain the cached buckets from a zone. Expects a locked zone on entry.
646 */
647static void
648bucket_cache_drain(uma_zone_t zone)
649{
650 uma_bucket_t bucket;
651
652 /*
653 * Drain the bucket queues and free the buckets, we just keep two per
654 * cpu (alloc/free).
655 */
656 while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
657 LIST_REMOVE(bucket, ub_link);
658 ZONE_UNLOCK(zone);
659 bucket_drain(zone, bucket);
660 bucket_free(bucket);
661 ZONE_LOCK(zone);
662 }
663
664 /* Now we do the free queue.. */
665 while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
666 LIST_REMOVE(bucket, ub_link);
667 bucket_free(bucket);
668 }
669}
670
671/*
672 * Frees pages from a zone back to the system. This is done on demand from
673 * the pageout daemon.
674 *
675 * Arguments:
676 * zone The zone to free pages from
677 * all Should we drain all items?
678 *
679 * Returns:
680 * Nothing.
681 */
682void
683zone_drain(uma_zone_t zone)
684{
685 struct slabhead freeslabs = { 0 };
686 uma_keg_t keg;
687 uma_slab_t slab;
688 uma_slab_t n;
689 u_int8_t flags;
690 u_int8_t *mem;
691 int i;
692
693 keg = zone->uz_keg;
694
695 /*
696 * We don't want to take pages from statically allocated zones at this
697 * time
698 */
699 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
700 return;
701
702 ZONE_LOCK(zone);
703
704#ifdef UMA_DEBUG
705 printf("%s free items: %u\n", zone->uz_name, keg->uk_free);
706#endif
707 bucket_cache_drain(zone);
708 if (keg->uk_free == 0)
709 goto finished;
710
711 slab = LIST_FIRST(&keg->uk_free_slab);
712 while (slab) {
713 n = LIST_NEXT(slab, us_link);
714
715 /* We have no where to free these to */
716 if (slab->us_flags & UMA_SLAB_BOOT) {
717 slab = n;
718 continue;
719 }
720
721 LIST_REMOVE(slab, us_link);
722 keg->uk_pages -= keg->uk_ppera;
723 keg->uk_free -= keg->uk_ipers;
724
725 if (keg->uk_flags & UMA_ZONE_HASH)
726 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
727
728 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
729
730 slab = n;
731 }
732finished:
733 ZONE_UNLOCK(zone);
734
735 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
736 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
737 if (keg->uk_fini)
738 for (i = 0; i < keg->uk_ipers; i++)
739 keg->uk_fini(
740 slab->us_data + (keg->uk_rsize * i),
741 keg->uk_size);
742 flags = slab->us_flags;
743 mem = slab->us_data;
744
745 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
746 (keg->uk_flags & UMA_ZONE_REFCNT)) {
747 vm_object_t obj;
748
749 if (flags & UMA_SLAB_KMEM)
750 obj = kmem_object;
751 else
752 obj = NULL;
753 for (i = 0; i < keg->uk_ppera; i++)
754 vsetobj((vm_offset_t)mem + (i * PAGE_SIZE),
755 obj);
756 }
757 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
758 uma_zfree_internal(keg->uk_slabzone, slab, NULL,
759 SKIP_NONE, ZFREE_STATFREE);
760#ifdef UMA_DEBUG
761 printf("%s: Returning %d bytes.\n",
762 zone->uz_name, UMA_SLAB_SIZE * keg->uk_ppera);
763#endif
764 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags);
765 }
766}
767
768/*
769 * Allocate a new slab for a zone. This does not insert the slab onto a list.
770 *
771 * Arguments:
772 * zone The zone to allocate slabs for
773 * wait Shall we wait?
774 *
775 * Returns:
776 * The slab that was allocated or NULL if there is no memory and the
777 * caller specified M_NOWAIT.
778 */
779static uma_slab_t
780slab_zalloc(uma_zone_t zone, int wait)
781{
782 uma_slabrefcnt_t slabref;
783 uma_slab_t slab;
784 uma_keg_t keg;
785 u_int8_t *mem;
786 u_int8_t flags;
787 int i;
788
789 slab = NULL;
790 keg = zone->uz_keg;
791
792#ifdef UMA_DEBUG
793 printf("slab_zalloc: Allocating a new slab for %s\n", zone->uz_name);
794#endif
795 ZONE_UNLOCK(zone);
796
797 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
798 slab = uma_zalloc_internal(keg->uk_slabzone, NULL, wait);
799 if (slab == NULL) {
800 ZONE_LOCK(zone);
801 return NULL;
802 }
803 }
804
805 /*
806 * This reproduces the old vm_zone behavior of zero filling pages the
807 * first time they are added to a zone.
808 *
809 * Malloced items are zeroed in uma_zalloc.
810 */
811
812 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
813 wait |= M_ZERO;
814 else
815 wait &= ~M_ZERO;
816
817 mem = keg->uk_allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE,
818 &flags, wait);
819 if (mem == NULL) {
820 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
821 uma_zfree_internal(keg->uk_slabzone, slab, NULL,
822 SKIP_NONE, ZFREE_STATFREE);
823 ZONE_LOCK(zone);
824 return (NULL);
825 }
826
827 /* Point the slab into the allocated memory */
828 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
829 slab = (uma_slab_t )(mem + keg->uk_pgoff);
830
831 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
832 (keg->uk_flags & UMA_ZONE_REFCNT))
833 for (i = 0; i < keg->uk_ppera; i++)
834 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
835
836 slab->us_keg = keg;
837 slab->us_data = mem;
838 slab->us_freecount = keg->uk_ipers;
839 slab->us_firstfree = 0;
840 slab->us_flags = flags;
841
842 if (keg->uk_flags & UMA_ZONE_REFCNT) {
843 slabref = (uma_slabrefcnt_t)slab;
844 for (i = 0; i < keg->uk_ipers; i++) {
845 slabref->us_freelist[i].us_refcnt = 0;
846 slabref->us_freelist[i].us_item = i+1;
847 }
848 } else {
849 for (i = 0; i < keg->uk_ipers; i++)
850 slab->us_freelist[i].us_item = i+1;
851 }
852
853 if (keg->uk_init != NULL) {
854 for (i = 0; i < keg->uk_ipers; i++)
855 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
856 keg->uk_size, wait) != 0)
857 break;
858 if (i != keg->uk_ipers) {
859 if (keg->uk_fini != NULL) {
860 for (i--; i > -1; i--)
861 keg->uk_fini(slab->us_data +
862 (keg->uk_rsize * i),
863 keg->uk_size);
864 }
865 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
866 (keg->uk_flags & UMA_ZONE_REFCNT)) {
867 vm_object_t obj;
868
869 if (flags & UMA_SLAB_KMEM)
870 obj = kmem_object;
871 else
872 obj = NULL;
873 for (i = 0; i < keg->uk_ppera; i++)
874 vsetobj((vm_offset_t)mem +
875 (i * PAGE_SIZE), obj);
876 }
877 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
878 uma_zfree_internal(keg->uk_slabzone, slab,
879 NULL, SKIP_NONE, ZFREE_STATFREE);
880 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera,
881 flags);
882 ZONE_LOCK(zone);
883 return (NULL);
884 }
885 }
886 ZONE_LOCK(zone);
887
888 if (keg->uk_flags & UMA_ZONE_HASH)
889 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
890
891 keg->uk_pages += keg->uk_ppera;
892 keg->uk_free += keg->uk_ipers;
893
894 return (slab);
895}
896
897/*
898 * This function is intended to be used early on in place of page_alloc() so
899 * that we may use the boot time page cache to satisfy allocations before
900 * the VM is ready.
901 */
902static void *
903startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
904{
905 uma_keg_t keg;
906 uma_slab_t tmps;
907
908 keg = zone->uz_keg;
909
910 /*
911 * Check our small startup cache to see if it has pages remaining.
912 */
913 mtx_lock(&uma_boot_pages_mtx);
914 if ((tmps = LIST_FIRST(&uma_boot_pages)) != NULL) {
915 LIST_REMOVE(tmps, us_link);
916 mtx_unlock(&uma_boot_pages_mtx);
917 *pflag = tmps->us_flags;
918 return (tmps->us_data);
919 }
920 mtx_unlock(&uma_boot_pages_mtx);
921 if (booted == 0)
922 panic("UMA: Increase vm.boot_pages");
923 /*
924 * Now that we've booted reset these users to their real allocator.
925 */
926#ifdef UMA_MD_SMALL_ALLOC
927 keg->uk_allocf = uma_small_alloc;
928#else
929 keg->uk_allocf = page_alloc;
930#endif
931 return keg->uk_allocf(zone, bytes, pflag, wait);
932}
933
934/*
935 * Allocates a number of pages from the system
936 *
937 * Arguments:
938 * zone Unused
939 * bytes The number of bytes requested
940 * wait Shall we wait?
941 *
942 * Returns:
943 * A pointer to the alloced memory or possibly
944 * NULL if M_NOWAIT is set.
945 */
946static void *
947page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
948{
949 void *p; /* Returned page */
950
951 *pflag = UMA_SLAB_KMEM;
952 p = (void *) kmem_malloc(kmem_map, bytes, wait);
953
954 return (p);
955}
956
957/*
958 * Allocates a number of pages from within an object
959 *
960 * Arguments:
961 * zone Unused
962 * bytes The number of bytes requested
963 * wait Shall we wait?
964 *
965 * Returns:
966 * A pointer to the alloced memory or possibly
967 * NULL if M_NOWAIT is set.
968 */
969static void *
970obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
971{
972 vm_object_t object;
973 vm_offset_t retkva, zkva;
974 vm_page_t p;
975 int pages, startpages;
976
977 object = zone->uz_keg->uk_obj;
978 retkva = 0;
979
980 /*
981 * This looks a little weird since we're getting one page at a time.
982 */
983 VM_OBJECT_LOCK(object);
984 p = TAILQ_LAST(&object->memq, pglist);
985 pages = p != NULL ? p->pindex + 1 : 0;
986 startpages = pages;
987 zkva = zone->uz_keg->uk_kva + pages * PAGE_SIZE;
988 for (; bytes > 0; bytes -= PAGE_SIZE) {
989 p = vm_page_alloc(object, pages,
990 VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED);
991 if (p == NULL) {
992 if (pages != startpages)
993 pmap_qremove(retkva, pages - startpages);
994 while (pages != startpages) {
995 pages--;
996 p = TAILQ_LAST(&object->memq, pglist);
997 vm_page_lock_queues();
998 vm_page_unwire(p, 0);
999 vm_page_free(p);
1000 vm_page_unlock_queues();
1001 }
1002 retkva = 0;
1003 goto done;
1004 }
1005 pmap_qenter(zkva, &p, 1);
1006 if (retkva == 0)
1007 retkva = zkva;
1008 zkva += PAGE_SIZE;
1009 pages += 1;
1010 }
1011done:
1012 VM_OBJECT_UNLOCK(object);
1013 *flags = UMA_SLAB_PRIV;
1014
1015 return ((void *)retkva);
1016}
1017
1018/*
1019 * Frees a number of pages to the system
1020 *
1021 * Arguments:
1022 * mem A pointer to the memory to be freed
1023 * size The size of the memory being freed
1024 * flags The original p->us_flags field
1025 *
1026 * Returns:
1027 * Nothing
1028 */
1029static void
1030page_free(void *mem, int size, u_int8_t flags)
1031{
1032 vm_map_t map;
1033
1034 if (flags & UMA_SLAB_KMEM)
1035 map = kmem_map;
1036 else
1037 panic("UMA: page_free used with invalid flags %d\n", flags);
1038
1039 kmem_free(map, (vm_offset_t)mem, size);
1040}
1041
1042/*
1043 * Zero fill initializer
1044 *
1045 * Arguments/Returns follow uma_init specifications
1046 */
1047static int
1048zero_init(void *mem, int size, int flags)
1049{
1050 bzero(mem, size);
1051 return (0);
1052}
1053
1054/*
1055 * Finish creating a small uma zone. This calculates ipers, and the zone size.
1056 *
1057 * Arguments
1058 * zone The zone we should initialize
1059 *
1060 * Returns
1061 * Nothing
1062 */
1063static void
1064zone_small_init(uma_zone_t zone)
1065{
1066 uma_keg_t keg;
1067 u_int rsize;
1068 u_int memused;
1069 u_int wastedspace;
1070 u_int shsize;
1071
1072 keg = zone->uz_keg;
1073 KASSERT(keg != NULL, ("Keg is null in zone_small_init"));
1074 rsize = keg->uk_size;
1075
1076 if (rsize < UMA_SMALLEST_UNIT)
1077 rsize = UMA_SMALLEST_UNIT;
1078 if (rsize & keg->uk_align)
1079 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1080
1081 keg->uk_rsize = rsize;
1082 keg->uk_ppera = 1;
1083
1084 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1085 rsize += UMA_FRITMREF_SZ; /* linkage & refcnt */
1086 shsize = sizeof(struct uma_slab_refcnt);
1087 } else {
1088 rsize += UMA_FRITM_SZ; /* Account for linkage */
1089 shsize = sizeof(struct uma_slab);
1090 }
1091
1092 keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize;
1093 KASSERT(keg->uk_ipers != 0, ("zone_small_init: ipers is 0"));
1094 memused = keg->uk_ipers * rsize + shsize;
1095 wastedspace = UMA_SLAB_SIZE - memused;
1096
1097 /*
1098 * We can't do OFFPAGE if we're internal or if we've been
1099 * asked to not go to the VM for buckets. If we do this we
1100 * may end up going to the VM (kmem_map) for slabs which we
1101 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1102 * result of UMA_ZONE_VM, which clearly forbids it.
1103 */
1104 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1105 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1106 return;
1107
1108 if ((wastedspace >= UMA_MAX_WASTE) &&
1109 (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) {
1110 keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize;
1111 KASSERT(keg->uk_ipers <= 255,
1112 ("zone_small_init: keg->uk_ipers too high!"));
1113#ifdef UMA_DEBUG
1114 printf("UMA decided we need offpage slab headers for "
1115 "zone: %s, calculated wastedspace = %d, "
1116 "maximum wasted space allowed = %d, "
1117 "calculated ipers = %d, "
1118 "new wasted space = %d\n", zone->uz_name, wastedspace,
1119 UMA_MAX_WASTE, keg->uk_ipers,
1120 UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize);
1121#endif
1122 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1123 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1124 keg->uk_flags |= UMA_ZONE_HASH;
1125 }
1126}
1127
1128/*
1129 * Finish creating a large (> UMA_SLAB_SIZE) uma zone. Just give in and do
1130 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1131 * more complicated.
1132 *
1133 * Arguments
1134 * zone The zone we should initialize
1135 *
1136 * Returns
1137 * Nothing
1138 */
1139static void
1140zone_large_init(uma_zone_t zone)
1141{
1142 uma_keg_t keg;
1143 int pages;
1144
1145 keg = zone->uz_keg;
1146
1147 KASSERT(keg != NULL, ("Keg is null in zone_large_init"));
1148 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1149 ("zone_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY zone"));
1150
1151 pages = keg->uk_size / UMA_SLAB_SIZE;
1152
1153 /* Account for remainder */
1154 if ((pages * UMA_SLAB_SIZE) < keg->uk_size)
1155 pages++;
1156
1157 keg->uk_ppera = pages;
1158 keg->uk_ipers = 1;
1159
1160 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1161 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1162 keg->uk_flags |= UMA_ZONE_HASH;
1163
1164 keg->uk_rsize = keg->uk_size;
1165}
1166
1167/*
1168 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1169 * the keg onto the global keg list.
1170 *
1171 * Arguments/Returns follow uma_ctor specifications
1172 * udata Actually uma_kctor_args
1173 */
1174static int
1175keg_ctor(void *mem, int size, void *udata, int flags)
1176{
1177 struct uma_kctor_args *arg = udata;
1178 uma_keg_t keg = mem;
1179 uma_zone_t zone;
1180
1181 bzero(keg, size);
1182 keg->uk_size = arg->size;
1183 keg->uk_init = arg->uminit;
1184 keg->uk_fini = arg->fini;
1185 keg->uk_align = arg->align;
1186 keg->uk_free = 0;
1187 keg->uk_pages = 0;
1188 keg->uk_flags = arg->flags;
1189 keg->uk_allocf = page_alloc;
1190 keg->uk_freef = page_free;
1191 keg->uk_recurse = 0;
1192 keg->uk_slabzone = NULL;
1193
1194 /*
1195 * The master zone is passed to us at keg-creation time.
1196 */
1197 zone = arg->zone;
1198 zone->uz_keg = keg;
1199
1200 if (arg->flags & UMA_ZONE_VM)
1201 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1202
1203 if (arg->flags & UMA_ZONE_ZINIT)
1204 keg->uk_init = zero_init;
1205
1206 /*
1207 * The +UMA_FRITM_SZ added to uk_size is to account for the
1208 * linkage that is added to the size in zone_small_init(). If
1209 * we don't account for this here then we may end up in
1210 * zone_small_init() with a calculated 'ipers' of 0.
1211 */
1212 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1213 if ((keg->uk_size+UMA_FRITMREF_SZ) >
1214 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)))
1215 zone_large_init(zone);
1216 else
1217 zone_small_init(zone);
1218 } else {
1219 if ((keg->uk_size+UMA_FRITM_SZ) >
1220 (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1221 zone_large_init(zone);
1222 else
1223 zone_small_init(zone);
1224 }
1225
1226 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1227 if (keg->uk_flags & UMA_ZONE_REFCNT)
1228 keg->uk_slabzone = slabrefzone;
1229 else
1230 keg->uk_slabzone = slabzone;
1231 }
1232
1233 /*
1234 * If we haven't booted yet we need allocations to go through the
1235 * startup cache until the vm is ready.
1236 */
1237 if (keg->uk_ppera == 1) {
1238#ifdef UMA_MD_SMALL_ALLOC
1239 keg->uk_allocf = uma_small_alloc;
1240 keg->uk_freef = uma_small_free;
1241#endif
1242 if (booted == 0)
1243 keg->uk_allocf = startup_alloc;
1244 }
1245
1246 /*
1247 * Initialize keg's lock (shared among zones) through
1248 * Master zone
1249 */
1250 zone->uz_lock = &keg->uk_lock;
1251 if (arg->flags & UMA_ZONE_MTXCLASS)
1252 ZONE_LOCK_INIT(zone, 1);
1253 else
1254 ZONE_LOCK_INIT(zone, 0);
1255
1256 /*
1257 * If we're putting the slab header in the actual page we need to
1258 * figure out where in each page it goes. This calculates a right
1259 * justified offset into the memory on an ALIGN_PTR boundary.
1260 */
1261 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1262 u_int totsize;
1263
1264 /* Size of the slab struct and free list */
1265 if (keg->uk_flags & UMA_ZONE_REFCNT)
1266 totsize = sizeof(struct uma_slab_refcnt) +
1267 keg->uk_ipers * UMA_FRITMREF_SZ;
1268 else
1269 totsize = sizeof(struct uma_slab) +
1270 keg->uk_ipers * UMA_FRITM_SZ;
1271
1272 if (totsize & UMA_ALIGN_PTR)
1273 totsize = (totsize & ~UMA_ALIGN_PTR) +
1274 (UMA_ALIGN_PTR + 1);
1275 keg->uk_pgoff = UMA_SLAB_SIZE - totsize;
1276
1277 if (keg->uk_flags & UMA_ZONE_REFCNT)
1278 totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt)
1279 + keg->uk_ipers * UMA_FRITMREF_SZ;
1280 else
1281 totsize = keg->uk_pgoff + sizeof(struct uma_slab)
1282 + keg->uk_ipers * UMA_FRITM_SZ;
1283
1284 /*
1285 * The only way the following is possible is if with our
1286 * UMA_ALIGN_PTR adjustments we are now bigger than
1287 * UMA_SLAB_SIZE. I haven't checked whether this is
1288 * mathematically possible for all cases, so we make
1289 * sure here anyway.
1290 */
1291 if (totsize > UMA_SLAB_SIZE) {
1292 printf("zone %s ipers %d rsize %d size %d\n",
1293 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1294 keg->uk_size);
1295 panic("UMA slab won't fit.\n");
1296 }
1297 }
1298
1299 if (keg->uk_flags & UMA_ZONE_HASH)
1300 hash_alloc(&keg->uk_hash);
1301
1302#ifdef UMA_DEBUG
1303 printf("%s(%p) size = %d ipers = %d ppera = %d pgoff = %d\n",
1304 zone->uz_name, zone,
1305 keg->uk_size, keg->uk_ipers,
1306 keg->uk_ppera, keg->uk_pgoff);
1307#endif
1308
1309 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1310
1311 mtx_lock(&uma_mtx);
1312 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1313 mtx_unlock(&uma_mtx);
1314 return (0);
1315}
1316
1317/*
1318 * Zone header ctor. This initializes all fields, locks, etc.
1319 *
1320 * Arguments/Returns follow uma_ctor specifications
1321 * udata Actually uma_zctor_args
1322 */
1323
1324static int
1325zone_ctor(void *mem, int size, void *udata, int flags)
1326{
1327 struct uma_zctor_args *arg = udata;
1328 uma_zone_t zone = mem;
1329 uma_zone_t z;
1330 uma_keg_t keg;
1331
1332 bzero(zone, size);
1333 zone->uz_name = arg->name;
1334 zone->uz_ctor = arg->ctor;
1335 zone->uz_dtor = arg->dtor;
1336 zone->uz_init = NULL;
1337 zone->uz_fini = NULL;
1338 zone->uz_allocs = 0;
1339 zone->uz_frees = 0;
1340 zone->uz_fails = 0;
1341 zone->uz_fills = zone->uz_count = 0;
1342
1343 if (arg->flags & UMA_ZONE_SECONDARY) {
1344 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1345 keg = arg->keg;
1346 zone->uz_keg = keg;
1347 zone->uz_init = arg->uminit;
1348 zone->uz_fini = arg->fini;
1349 zone->uz_lock = &keg->uk_lock;
1350 mtx_lock(&uma_mtx);
1351 ZONE_LOCK(zone);
1352 keg->uk_flags |= UMA_ZONE_SECONDARY;
1353 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1354 if (LIST_NEXT(z, uz_link) == NULL) {
1355 LIST_INSERT_AFTER(z, zone, uz_link);
1356 break;
1357 }
1358 }
1359 ZONE_UNLOCK(zone);
1360 mtx_unlock(&uma_mtx);
1361 } else if (arg->keg == NULL) {
1362 if (uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1363 arg->align, arg->flags) == NULL)
1364 return (ENOMEM);
1365 } else {
1366 struct uma_kctor_args karg;
1367 int error;
1368
1369 /* We should only be here from uma_startup() */
1370 karg.size = arg->size;
1371 karg.uminit = arg->uminit;
1372 karg.fini = arg->fini;
1373 karg.align = arg->align;
1374 karg.flags = arg->flags;
1375 karg.zone = zone;
1376 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1377 flags);
1378 if (error)
1379 return (error);
1380 }
1381 keg = zone->uz_keg;
1382 zone->uz_lock = &keg->uk_lock;
1383
1384 /*
1385 * Some internal zones don't have room allocated for the per cpu
1386 * caches. If we're internal, bail out here.
1387 */
1388 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1389 KASSERT((keg->uk_flags & UMA_ZONE_SECONDARY) == 0,
1390 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1391 return (0);
1392 }
1393
1394 if (keg->uk_flags & UMA_ZONE_MAXBUCKET)
1395 zone->uz_count = BUCKET_MAX;
1396 else if (keg->uk_ipers <= BUCKET_MAX)
1397 zone->uz_count = keg->uk_ipers;
1398 else
1399 zone->uz_count = BUCKET_MAX;
1400 return (0);
1401}
1402
1403/*
1404 * Keg header dtor. This frees all data, destroys locks, frees the hash
1405 * table and removes the keg from the global list.
1406 *
1407 * Arguments/Returns follow uma_dtor specifications
1408 * udata unused
1409 */
1410static void
1411keg_dtor(void *arg, int size, void *udata)
1412{
1413 uma_keg_t keg;
1414
1415 keg = (uma_keg_t)arg;
1416 mtx_lock(&keg->uk_lock);
1417 if (keg->uk_free != 0) {
1418 printf("Freed UMA keg was not empty (%d items). "
1419 " Lost %d pages of memory.\n",
1420 keg->uk_free, keg->uk_pages);
1421 }
1422 mtx_unlock(&keg->uk_lock);
1423
1424 if (keg->uk_flags & UMA_ZONE_HASH)
1425 hash_free(&keg->uk_hash);
1426
1427 mtx_destroy(&keg->uk_lock);
1428}
1429
1430/*
1431 * Zone header dtor.
1432 *
1433 * Arguments/Returns follow uma_dtor specifications
1434 * udata unused
1435 */
1436static void
1437zone_dtor(void *arg, int size, void *udata)
1438{
1439 uma_zone_t zone;
1440 uma_keg_t keg;
1441
1442 zone = (uma_zone_t)arg;
1443 keg = zone->uz_keg;
1444
1445 if (!(keg->uk_flags & UMA_ZFLAG_INTERNAL))
1446 cache_drain(zone);
1447
1448 mtx_lock(&uma_mtx);
1449 zone_drain(zone);
1450 if (keg->uk_flags & UMA_ZONE_SECONDARY) {
1451 LIST_REMOVE(zone, uz_link);
1452 /*
1453 * XXX there are some races here where
1454 * the zone can be drained but zone lock
1455 * released and then refilled before we
1456 * remove it... we dont care for now
1457 */
1458 ZONE_LOCK(zone);
1459 if (LIST_EMPTY(&keg->uk_zones))
1460 keg->uk_flags &= ~UMA_ZONE_SECONDARY;
1461 ZONE_UNLOCK(zone);
1462 mtx_unlock(&uma_mtx);
1463 } else {
1464 LIST_REMOVE(keg, uk_link);
1465 LIST_REMOVE(zone, uz_link);
1466 mtx_unlock(&uma_mtx);
1467 uma_zfree_internal(kegs, keg, NULL, SKIP_NONE,
1468 ZFREE_STATFREE);
1469 }
1470 zone->uz_keg = NULL;
1471}
1472
1473/*
1474 * Traverses every zone in the system and calls a callback
1475 *
1476 * Arguments:
1477 * zfunc A pointer to a function which accepts a zone
1478 * as an argument.
1479 *
1480 * Returns:
1481 * Nothing
1482 */
1483static void
1484zone_foreach(void (*zfunc)(uma_zone_t))
1485{
1486 uma_keg_t keg;
1487 uma_zone_t zone;
1488
1489 mtx_lock(&uma_mtx);
1490 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1491 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1492 zfunc(zone);
1493 }
1494 mtx_unlock(&uma_mtx);
1495}
1496
1497/* Public functions */
1498/* See uma.h */
1499void
1500uma_startup(void *bootmem, int boot_pages)
1501{
1502 struct uma_zctor_args args;
1503 uma_slab_t slab;
1504 u_int slabsize;
1505 u_int objsize, totsize, wsize;
1506 int i;
1507
1508#ifdef UMA_DEBUG
1509 printf("Creating uma keg headers zone and keg.\n");
1510#endif
1511 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1512
1513 /*
1514 * Figure out the maximum number of items-per-slab we'll have if
1515 * we're using the OFFPAGE slab header to track free items, given
1516 * all possible object sizes and the maximum desired wastage
1517 * (UMA_MAX_WASTE).
1518 *
1519 * We iterate until we find an object size for
1520 * which the calculated wastage in zone_small_init() will be
1521 * enough to warrant OFFPAGE. Since wastedspace versus objsize
1522 * is an overall increasing see-saw function, we find the smallest
1523 * objsize such that the wastage is always acceptable for objects
1524 * with that objsize or smaller. Since a smaller objsize always
1525 * generates a larger possible uma_max_ipers, we use this computed
1526 * objsize to calculate the largest ipers possible. Since the
1527 * ipers calculated for OFFPAGE slab headers is always larger than
1528 * the ipers initially calculated in zone_small_init(), we use
1529 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to
1530 * obtain the maximum ipers possible for offpage slab headers.
1531 *
1532 * It should be noted that ipers versus objsize is an inversly
1533 * proportional function which drops off rather quickly so as
1534 * long as our UMA_MAX_WASTE is such that the objsize we calculate
1535 * falls into the portion of the inverse relation AFTER the steep
1536 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386).
1537 *
1538 * Note that we have 8-bits (1 byte) to use as a freelist index
1539 * inside the actual slab header itself and this is enough to
1540 * accomodate us. In the worst case, a UMA_SMALLEST_UNIT sized
1541 * object with offpage slab header would have ipers =
1542 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is
1543 * 1 greater than what our byte-integer freelist index can
1544 * accomodate, but we know that this situation never occurs as
1545 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate
1546 * that we need to go to offpage slab headers. Or, if we do,
1547 * then we trap that condition below and panic in the INVARIANTS case.
1548 */
1549 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE;
1550 totsize = wsize;
1551 objsize = UMA_SMALLEST_UNIT;
1552 while (totsize >= wsize) {
1553 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) /
1554 (objsize + UMA_FRITM_SZ);
1555 totsize *= (UMA_FRITM_SZ + objsize);
1556 objsize++;
1557 }
1558 if (objsize > UMA_SMALLEST_UNIT)
1559 objsize--;
1560 uma_max_ipers = UMA_SLAB_SIZE / objsize;
1561
1562 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE;
1563 totsize = wsize;
1564 objsize = UMA_SMALLEST_UNIT;
1565 while (totsize >= wsize) {
1566 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) /
1567 (objsize + UMA_FRITMREF_SZ);
1568 totsize *= (UMA_FRITMREF_SZ + objsize);
1569 objsize++;
1570 }
1571 if (objsize > UMA_SMALLEST_UNIT)
1572 objsize--;
1573 uma_max_ipers_ref = UMA_SLAB_SIZE / objsize;
1574
1575 KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255),
1576 ("uma_startup: calculated uma_max_ipers values too large!"));
1577
1578#ifdef UMA_DEBUG
1579 printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers);
1580 printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n",
1581 uma_max_ipers_ref);
1582#endif
1583
1584 /* "manually" create the initial zone */
1585 args.name = "UMA Kegs";
1586 args.size = sizeof(struct uma_keg);
1587 args.ctor = keg_ctor;
1588 args.dtor = keg_dtor;
1589 args.uminit = zero_init;
1590 args.fini = NULL;
1591 args.keg = &masterkeg;
1592 args.align = 32 - 1;
1593 args.flags = UMA_ZFLAG_INTERNAL;
1594 /* The initial zone has no Per cpu queues so it's smaller */
1595 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1596
1597#ifdef UMA_DEBUG
1598 printf("Filling boot free list.\n");
1599#endif
1600 for (i = 0; i < boot_pages; i++) {
1601 slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE));
1602 slab->us_data = (u_int8_t *)slab;
1603 slab->us_flags = UMA_SLAB_BOOT;
1604 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1605 }
1606 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1607
1608#ifdef UMA_DEBUG
1609 printf("Creating uma zone headers zone and keg.\n");
1610#endif
1611 args.name = "UMA Zones";
1612 args.size = sizeof(struct uma_zone) +
1613 (sizeof(struct uma_cache) * (mp_maxid + 1));
1614 args.ctor = zone_ctor;
1615 args.dtor = zone_dtor;
1616 args.uminit = zero_init;
1617 args.fini = NULL;
1618 args.keg = NULL;
1619 args.align = 32 - 1;
1620 args.flags = UMA_ZFLAG_INTERNAL;
1621 /* The initial zone has no Per cpu queues so it's smaller */
1622 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1623
1624#ifdef UMA_DEBUG
1625 printf("Initializing pcpu cache locks.\n");
1626#endif
1627#ifdef UMA_DEBUG
1628 printf("Creating slab and hash zones.\n");
1629#endif
1630
1631 /*
1632 * This is the max number of free list items we'll have with
1633 * offpage slabs.
1634 */
1635 slabsize = uma_max_ipers * UMA_FRITM_SZ;
1636 slabsize += sizeof(struct uma_slab);
1637
1638 /* Now make a zone for slab headers */
1639 slabzone = uma_zcreate("UMA Slabs",
1640 slabsize,
1641 NULL, NULL, NULL, NULL,
1642 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1643
1644 /*
1645 * We also create a zone for the bigger slabs with reference
1646 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1647 */
1648 slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ;
1649 slabsize += sizeof(struct uma_slab_refcnt);
1650 slabrefzone = uma_zcreate("UMA RCntSlabs",
1651 slabsize,
1652 NULL, NULL, NULL, NULL,
1653 UMA_ALIGN_PTR,
1654 UMA_ZFLAG_INTERNAL);
1655
1656 hashzone = uma_zcreate("UMA Hash",
1657 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1658 NULL, NULL, NULL, NULL,
1659 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1660
1661 bucket_init();
1662
1663#ifdef UMA_MD_SMALL_ALLOC
1664 booted = 1;
1665#endif
1666
1667#ifdef UMA_DEBUG
1668 printf("UMA startup complete.\n");
1669#endif
1670}
1671
1672/* see uma.h */
1673void
1674uma_startup2(void)
1675{
1676 booted = 1;
1677 bucket_enable();
1678#ifdef UMA_DEBUG
1679 printf("UMA startup2 complete.\n");
1680#endif
1681}
1682
1683/*
1684 * Initialize our callout handle
1685 *
1686 */
1687
1688static void
1689uma_startup3(void)
1690{
1691#ifdef UMA_DEBUG
1692 printf("Starting callout.\n");
1693#endif
1694 callout_init(&uma_callout, CALLOUT_MPSAFE);
1695 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1696#ifdef UMA_DEBUG
1697 printf("UMA startup3 complete.\n");
1698#endif
1699}
1700
1701static uma_zone_t
1702uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1703 int align, u_int32_t flags)
1704{
1705 struct uma_kctor_args args;
1706
1707 args.size = size;
1708 args.uminit = uminit;
1709 args.fini = fini;
1710 args.align = align;
1711 args.flags = flags;
1712 args.zone = zone;
1713 return (uma_zalloc_internal(kegs, &args, M_WAITOK));
1714}
1715
1716/* See uma.h */
1717uma_zone_t
1718uma_zcreate(char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1719 uma_init uminit, uma_fini fini, int align, u_int32_t flags)
1720
1721{
1722 struct uma_zctor_args args;
1723
1724 /* This stuff is essential for the zone ctor */
1725 args.name = name;
1726 args.size = size;
1727 args.ctor = ctor;
1728 args.dtor = dtor;
1729 args.uminit = uminit;
1730 args.fini = fini;
1731 args.align = align;
1732 args.flags = flags;
1733 args.keg = NULL;
1734
1735 return (uma_zalloc_internal(zones, &args, M_WAITOK));
1736}
1737
1738/* See uma.h */
1739uma_zone_t
1740uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1741 uma_init zinit, uma_fini zfini, uma_zone_t master)
1742{
1743 struct uma_zctor_args args;
1744
1745 args.name = name;
1746 args.size = master->uz_keg->uk_size;
1747 args.ctor = ctor;
1748 args.dtor = dtor;
1749 args.uminit = zinit;
1750 args.fini = zfini;
1751 args.align = master->uz_keg->uk_align;
1752 args.flags = master->uz_keg->uk_flags | UMA_ZONE_SECONDARY;
1753 args.keg = master->uz_keg;
1754
1755 return (uma_zalloc_internal(zones, &args, M_WAITOK));
1756}
1757
1758/* See uma.h */
1759void
1760uma_zdestroy(uma_zone_t zone)
1761{
1762
1763 uma_zfree_internal(zones, zone, NULL, SKIP_NONE, ZFREE_STATFREE);
1764}
1765
1766/* See uma.h */
1767void *
1768uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1769{
1770 void *item;
1771 uma_cache_t cache;
1772 uma_bucket_t bucket;
1773 int cpu;
1774
1775 /* This is the fast path allocation */
1776#ifdef UMA_DEBUG_ALLOC_1
1777 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1778#endif
1779 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1780 zone->uz_name, flags);
1781
1782 if (flags & M_WAITOK) {
1783 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1784 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
1785 }
1786
1787 /*
1788 * If possible, allocate from the per-CPU cache. There are two
1789 * requirements for safe access to the per-CPU cache: (1) the thread
1790 * accessing the cache must not be preempted or yield during access,
1791 * and (2) the thread must not migrate CPUs without switching which
1792 * cache it accesses. We rely on a critical section to prevent
1793 * preemption and migration. We release the critical section in
1794 * order to acquire the zone mutex if we are unable to allocate from
1795 * the current cache; when we re-acquire the critical section, we
1796 * must detect and handle migration if it has occurred.
1797 */
1798zalloc_restart:
1799 critical_enter();
1800 cpu = curcpu;
1801 cache = &zone->uz_cpu[cpu];
1802
1803zalloc_start:
1804 bucket = cache->uc_allocbucket;
1805
1806 if (bucket) {
1807 if (bucket->ub_cnt > 0) {
1808 bucket->ub_cnt--;
1809 item = bucket->ub_bucket[bucket->ub_cnt];
1810#ifdef INVARIANTS
1811 bucket->ub_bucket[bucket->ub_cnt] = NULL;
1812#endif
1813 KASSERT(item != NULL,
1814 ("uma_zalloc: Bucket pointer mangled."));
1815 cache->uc_allocs++;
1816 critical_exit();
1817#ifdef INVARIANTS
1818 ZONE_LOCK(zone);
1819 uma_dbg_alloc(zone, NULL, item);
1820 ZONE_UNLOCK(zone);
1821#endif
1822 if (zone->uz_ctor != NULL) {
1823 if (zone->uz_ctor(item, zone->uz_keg->uk_size,
1824 udata, flags) != 0) {
1825 uma_zfree_internal(zone, item, udata,
1826 SKIP_DTOR, ZFREE_STATFAIL |
1827 ZFREE_STATFREE);
1828 return (NULL);
1829 }
1830 }
1831 if (flags & M_ZERO)
1832 bzero(item, zone->uz_keg->uk_size);
1833 return (item);
1834 } else if (cache->uc_freebucket) {
1835 /*
1836 * We have run out of items in our allocbucket.
1837 * See if we can switch with our free bucket.
1838 */
1839 if (cache->uc_freebucket->ub_cnt > 0) {
1840#ifdef UMA_DEBUG_ALLOC
1841 printf("uma_zalloc: Swapping empty with"
1842 " alloc.\n");
1843#endif
1844 bucket = cache->uc_freebucket;
1845 cache->uc_freebucket = cache->uc_allocbucket;
1846 cache->uc_allocbucket = bucket;
1847
1848 goto zalloc_start;
1849 }
1850 }
1851 }
1852 /*
1853 * Attempt to retrieve the item from the per-CPU cache has failed, so
1854 * we must go back to the zone. This requires the zone lock, so we
1855 * must drop the critical section, then re-acquire it when we go back
1856 * to the cache. Since the critical section is released, we may be
1857 * preempted or migrate. As such, make sure not to maintain any
1858 * thread-local state specific to the cache from prior to releasing
1859 * the critical section.
1860 */
1861 critical_exit();
1862 ZONE_LOCK(zone);
1863 critical_enter();
1864 cpu = curcpu;
1865 cache = &zone->uz_cpu[cpu];
1866 bucket = cache->uc_allocbucket;
1867 if (bucket != NULL) {
1868 if (bucket->ub_cnt > 0) {
1869 ZONE_UNLOCK(zone);
1870 goto zalloc_start;
1871 }
1872 bucket = cache->uc_freebucket;
1873 if (bucket != NULL && bucket->ub_cnt > 0) {
1874 ZONE_UNLOCK(zone);
1875 goto zalloc_start;
1876 }
1877 }
1878
1879 /* Since we have locked the zone we may as well send back our stats */
1880 zone->uz_allocs += cache->uc_allocs;
1881 cache->uc_allocs = 0;
1882 zone->uz_frees += cache->uc_frees;
1883 cache->uc_frees = 0;
1884
1885 /* Our old one is now a free bucket */
1886 if (cache->uc_allocbucket) {
1887 KASSERT(cache->uc_allocbucket->ub_cnt == 0,
1888 ("uma_zalloc_arg: Freeing a non free bucket."));
1889 LIST_INSERT_HEAD(&zone->uz_free_bucket,
1890 cache->uc_allocbucket, ub_link);
1891 cache->uc_allocbucket = NULL;
1892 }
1893
1894 /* Check the free list for a new alloc bucket */
1895 if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
1896 KASSERT(bucket->ub_cnt != 0,
1897 ("uma_zalloc_arg: Returning an empty bucket."));
1898
1899 LIST_REMOVE(bucket, ub_link);
1900 cache->uc_allocbucket = bucket;
1901 ZONE_UNLOCK(zone);
1902 goto zalloc_start;
1903 }
1904 /* We are no longer associated with this CPU. */
1905 critical_exit();
1906
1907 /* Bump up our uz_count so we get here less */
1908 if (zone->uz_count < BUCKET_MAX)
1909 zone->uz_count++;
1910
1911 /*
1912 * Now lets just fill a bucket and put it on the free list. If that
1913 * works we'll restart the allocation from the begining.
1914 */
1915 if (uma_zalloc_bucket(zone, flags)) {
1916 ZONE_UNLOCK(zone);
1917 goto zalloc_restart;
1918 }
1919 ZONE_UNLOCK(zone);
1920 /*
1921 * We may not be able to get a bucket so return an actual item.
1922 */
1923#ifdef UMA_DEBUG
1924 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
1925#endif
1926
1927 return (uma_zalloc_internal(zone, udata, flags));
1928}
1929
1930static uma_slab_t
1931uma_zone_slab(uma_zone_t zone, int flags)
1932{
1933 uma_slab_t slab;
1934 uma_keg_t keg;
1935
1936 keg = zone->uz_keg;
1937
1938 /*
1939 * This is to prevent us from recursively trying to allocate
1940 * buckets. The problem is that if an allocation forces us to
1941 * grab a new bucket we will call page_alloc, which will go off
1942 * and cause the vm to allocate vm_map_entries. If we need new
1943 * buckets there too we will recurse in kmem_alloc and bad
1944 * things happen. So instead we return a NULL bucket, and make
1945 * the code that allocates buckets smart enough to deal with it
1946 *
1947 * XXX: While we want this protection for the bucket zones so that
1948 * recursion from the VM is handled (and the calling code that
1949 * allocates buckets knows how to deal with it), we do not want
1950 * to prevent allocation from the slab header zones (slabzone
1951 * and slabrefzone) if uk_recurse is not zero for them. The
1952 * reason is that it could lead to NULL being returned for
1953 * slab header allocations even in the M_WAITOK case, and the
1954 * caller can't handle that.
1955 */
1956 if (keg->uk_flags & UMA_ZFLAG_INTERNAL && keg->uk_recurse != 0)
1957 if ((zone != slabzone) && (zone != slabrefzone))
1958 return (NULL);
1959
1960 slab = NULL;
1961
1962 for (;;) {
1963 /*
1964 * Find a slab with some space. Prefer slabs that are partially
1965 * used over those that are totally full. This helps to reduce
1966 * fragmentation.
1967 */
1968 if (keg->uk_free != 0) {
1969 if (!LIST_EMPTY(&keg->uk_part_slab)) {
1970 slab = LIST_FIRST(&keg->uk_part_slab);
1971 } else {
1972 slab = LIST_FIRST(&keg->uk_free_slab);
1973 LIST_REMOVE(slab, us_link);
1974 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
1975 us_link);
1976 }
1977 return (slab);
1978 }
1979
1980 /*
1981 * M_NOVM means don't ask at all!
1982 */
1983 if (flags & M_NOVM)
1984 break;
1985
1986 if (keg->uk_maxpages &&
1987 keg->uk_pages >= keg->uk_maxpages) {
1988 keg->uk_flags |= UMA_ZFLAG_FULL;
1989
1990 if (flags & M_NOWAIT)
1991 break;
1992 else
1993 msleep(keg, &keg->uk_lock, PVM,
1994 "zonelimit", 0);
1995 continue;
1996 }
1997 keg->uk_recurse++;
1998 slab = slab_zalloc(zone, flags);
1999 keg->uk_recurse--;
2000
2001 /*
2002 * If we got a slab here it's safe to mark it partially used
2003 * and return. We assume that the caller is going to remove
2004 * at least one item.
2005 */
2006 if (slab) {
2007 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2008 return (slab);
2009 }
2010 /*
2011 * We might not have been able to get a slab but another cpu
2012 * could have while we were unlocked. Check again before we
2013 * fail.
2014 */
2015 if (flags & M_NOWAIT)
2016 flags |= M_NOVM;
2017 }
2018 return (slab);
2019}
2020
2021static void *
2022uma_slab_alloc(uma_zone_t zone, uma_slab_t slab)
2023{
2024 uma_keg_t keg;
2025 uma_slabrefcnt_t slabref;
2026 void *item;
2027 u_int8_t freei;
2028
2029 keg = zone->uz_keg;
2030
2031 freei = slab->us_firstfree;
2032 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2033 slabref = (uma_slabrefcnt_t)slab;
2034 slab->us_firstfree = slabref->us_freelist[freei].us_item;
2035 } else {
2036 slab->us_firstfree = slab->us_freelist[freei].us_item;
2037 }
2038 item = slab->us_data + (keg->uk_rsize * freei);
2039
2040 slab->us_freecount--;
2041 keg->uk_free--;
2042#ifdef INVARIANTS
2043 uma_dbg_alloc(zone, slab, item);
2044#endif
2045 /* Move this slab to the full list */
2046 if (slab->us_freecount == 0) {
2047 LIST_REMOVE(slab, us_link);
2048 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2049 }
2050
2051 return (item);
2052}
2053
2054static int
2055uma_zalloc_bucket(uma_zone_t zone, int flags)
2056{
2057 uma_bucket_t bucket;
2058 uma_slab_t slab;
2059 int16_t saved;
2060 int max, origflags = flags;
2061
2062 /*
2063 * Try this zone's free list first so we don't allocate extra buckets.
2064 */
2065 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2066 KASSERT(bucket->ub_cnt == 0,
2067 ("uma_zalloc_bucket: Bucket on free list is not empty."));
2068 LIST_REMOVE(bucket, ub_link);
2069 } else {
2070 int bflags;
2071
2072 bflags = (flags & ~M_ZERO);
2073 if (zone->uz_keg->uk_flags & UMA_ZFLAG_CACHEONLY)
2074 bflags |= M_NOVM;
2075
2076 ZONE_UNLOCK(zone);
2077 bucket = bucket_alloc(zone->uz_count, bflags);
2078 ZONE_LOCK(zone);
2079 }
2080
2081 if (bucket == NULL)
2082 return (0);
2083
2084#ifdef SMP
2085 /*
2086 * This code is here to limit the number of simultaneous bucket fills
2087 * for any given zone to the number of per cpu caches in this zone. This
2088 * is done so that we don't allocate more memory than we really need.
2089 */
2090 if (zone->uz_fills >= mp_ncpus)
2091 goto done;
2092
2093#endif
2094 zone->uz_fills++;
2095
2096 max = MIN(bucket->ub_entries, zone->uz_count);
2097 /* Try to keep the buckets totally full */
2098 saved = bucket->ub_cnt;
2099 while (bucket->ub_cnt < max &&
2100 (slab = uma_zone_slab(zone, flags)) != NULL) {
2101 while (slab->us_freecount && bucket->ub_cnt < max) {
2102 bucket->ub_bucket[bucket->ub_cnt++] =
2103 uma_slab_alloc(zone, slab);
2104 }
2105
2106 /* Don't block on the next fill */
2107 flags |= M_NOWAIT;
2108 }
2109
2110 /*
2111 * We unlock here because we need to call the zone's init.
2112 * It should be safe to unlock because the slab dealt with
2113 * above is already on the appropriate list within the keg
2114 * and the bucket we filled is not yet on any list, so we
2115 * own it.
2116 */
2117 if (zone->uz_init != NULL) {
2118 int i;
2119
2120 ZONE_UNLOCK(zone);
2121 for (i = saved; i < bucket->ub_cnt; i++)
2122 if (zone->uz_init(bucket->ub_bucket[i],
2123 zone->uz_keg->uk_size, origflags) != 0)
2124 break;
2125 /*
2126 * If we couldn't initialize the whole bucket, put the
2127 * rest back onto the freelist.
2128 */
2129 if (i != bucket->ub_cnt) {
2130 int j;
2131
2132 for (j = i; j < bucket->ub_cnt; j++) {
2133 uma_zfree_internal(zone, bucket->ub_bucket[j],
2134 NULL, SKIP_FINI, 0);
2135#ifdef INVARIANTS
2136 bucket->ub_bucket[j] = NULL;
2137#endif
2138 }
2139 bucket->ub_cnt = i;
2140 }
2141 ZONE_LOCK(zone);
2142 }
2143
2144 zone->uz_fills--;
2145 if (bucket->ub_cnt != 0) {
2146 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2147 bucket, ub_link);
2148 return (1);
2149 }
2150#ifdef SMP
2151done:
2152#endif
2153 bucket_free(bucket);
2154
2155 return (0);
2156}
2157/*
2158 * Allocates an item for an internal zone
2159 *
2160 * Arguments
2161 * zone The zone to alloc for.
2162 * udata The data to be passed to the constructor.
2163 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2164 *
2165 * Returns
2166 * NULL if there is no memory and M_NOWAIT is set
2167 * An item if successful
2168 */
2169
2170static void *
2171uma_zalloc_internal(uma_zone_t zone, void *udata, int flags)
2172{
2173 uma_keg_t keg;
2174 uma_slab_t slab;
2175 void *item;
2176
2177 item = NULL;
2178 keg = zone->uz_keg;
2179
2180#ifdef UMA_DEBUG_ALLOC
2181 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2182#endif
2183 ZONE_LOCK(zone);
2184
2185 slab = uma_zone_slab(zone, flags);
2186 if (slab == NULL) {
2187 zone->uz_fails++;
2188 ZONE_UNLOCK(zone);
2189 return (NULL);
2190 }
2191
2192 item = uma_slab_alloc(zone, slab);
2193
2194 zone->uz_allocs++;
2195
2196 ZONE_UNLOCK(zone);
2197
2198 /*
2199 * We have to call both the zone's init (not the keg's init)
2200 * and the zone's ctor. This is because the item is going from
2201 * a keg slab directly to the user, and the user is expecting it
2202 * to be both zone-init'd as well as zone-ctor'd.
2203 */
2204 if (zone->uz_init != NULL) {
2205 if (zone->uz_init(item, keg->uk_size, flags) != 0) {
2206 uma_zfree_internal(zone, item, udata, SKIP_FINI,
2207 ZFREE_STATFAIL | ZFREE_STATFREE);
2208 return (NULL);
2209 }
2210 }
2211 if (zone->uz_ctor != NULL) {
2212 if (zone->uz_ctor(item, keg->uk_size, udata, flags) != 0) {
2213 uma_zfree_internal(zone, item, udata, SKIP_DTOR,
2214 ZFREE_STATFAIL | ZFREE_STATFREE);
2215 return (NULL);
2216 }
2217 }
2218 if (flags & M_ZERO)
2219 bzero(item, keg->uk_size);
2220
2221 return (item);
2222}
2223
2224/* See uma.h */
2225void
2226uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2227{
2228 uma_keg_t keg;
2229 uma_cache_t cache;
2230 uma_bucket_t bucket;
2231 int bflags;
2232 int cpu;
2233
2234 keg = zone->uz_keg;
2235
2236#ifdef UMA_DEBUG_ALLOC_1
2237 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2238#endif
2239 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2240 zone->uz_name);
2241
2242 if (zone->uz_dtor)
2243 zone->uz_dtor(item, keg->uk_size, udata);
2244#ifdef INVARIANTS
2245 ZONE_LOCK(zone);
2246 if (keg->uk_flags & UMA_ZONE_MALLOC)
2247 uma_dbg_free(zone, udata, item);
2248 else
2249 uma_dbg_free(zone, NULL, item);
2250 ZONE_UNLOCK(zone);
2251#endif
2252 /*
2253 * The race here is acceptable. If we miss it we'll just have to wait
2254 * a little longer for the limits to be reset.
2255 */
2256 if (keg->uk_flags & UMA_ZFLAG_FULL)
2257 goto zfree_internal;
2258
2259 /*
2260 * If possible, free to the per-CPU cache. There are two
2261 * requirements for safe access to the per-CPU cache: (1) the thread
2262 * accessing the cache must not be preempted or yield during access,
2263 * and (2) the thread must not migrate CPUs without switching which
2264 * cache it accesses. We rely on a critical section to prevent
2265 * preemption and migration. We release the critical section in
2266 * order to acquire the zone mutex if we are unable to free to the
2267 * current cache; when we re-acquire the critical section, we must
2268 * detect and handle migration if it has occurred.
2269 */
2270zfree_restart:
2271 critical_enter();
2272 cpu = curcpu;
2273 cache = &zone->uz_cpu[cpu];
2274
2275zfree_start:
2276 bucket = cache->uc_freebucket;
2277
2278 if (bucket) {
2279 /*
2280 * Do we have room in our bucket? It is OK for this uz count
2281 * check to be slightly out of sync.
2282 */
2283
2284 if (bucket->ub_cnt < bucket->ub_entries) {
2285 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2286 ("uma_zfree: Freeing to non free bucket index."));
2287 bucket->ub_bucket[bucket->ub_cnt] = item;
2288 bucket->ub_cnt++;
2289 cache->uc_frees++;
2290 critical_exit();
2291 return;
2292 } else if (cache->uc_allocbucket) {
2293#ifdef UMA_DEBUG_ALLOC
2294 printf("uma_zfree: Swapping buckets.\n");
2295#endif
2296 /*
2297 * We have run out of space in our freebucket.
2298 * See if we can switch with our alloc bucket.
2299 */
2300 if (cache->uc_allocbucket->ub_cnt <
2301 cache->uc_freebucket->ub_cnt) {
2302 bucket = cache->uc_freebucket;
2303 cache->uc_freebucket = cache->uc_allocbucket;
2304 cache->uc_allocbucket = bucket;
2305 goto zfree_start;
2306 }
2307 }
2308 }
2309 /*
2310 * We can get here for two reasons:
2311 *
2312 * 1) The buckets are NULL
2313 * 2) The alloc and free buckets are both somewhat full.
2314 *
2315 * We must go back the zone, which requires acquiring the zone lock,
2316 * which in turn means we must release and re-acquire the critical
2317 * section. Since the critical section is released, we may be
2318 * preempted or migrate. As such, make sure not to maintain any
2319 * thread-local state specific to the cache from prior to releasing
2320 * the critical section.
2321 */
2322 critical_exit();
2323 ZONE_LOCK(zone);
2324 critical_enter();
2325 cpu = curcpu;
2326 cache = &zone->uz_cpu[cpu];
2327 if (cache->uc_freebucket != NULL) {
2328 if (cache->uc_freebucket->ub_cnt <
2329 cache->uc_freebucket->ub_entries) {
2330 ZONE_UNLOCK(zone);
2331 goto zfree_start;
2332 }
2333 if (cache->uc_allocbucket != NULL &&
2334 (cache->uc_allocbucket->ub_cnt <
2335 cache->uc_freebucket->ub_cnt)) {
2336 ZONE_UNLOCK(zone);
2337 goto zfree_start;
2338 }
2339 }
2340
2341 /* Since we have locked the zone we may as well send back our stats */
2342 zone->uz_allocs += cache->uc_allocs;
2343 cache->uc_allocs = 0;
2344 zone->uz_frees += cache->uc_frees;
2345 cache->uc_frees = 0;
2346
2347 bucket = cache->uc_freebucket;
2348 cache->uc_freebucket = NULL;
2349
2350 /* Can we throw this on the zone full list? */
2351 if (bucket != NULL) {
2352#ifdef UMA_DEBUG_ALLOC
2353 printf("uma_zfree: Putting old bucket on the free list.\n");
2354#endif
2355 /* ub_cnt is pointing to the last free item */
2356 KASSERT(bucket->ub_cnt != 0,
2357 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2358 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2359 bucket, ub_link);
2360 }
2361 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2362 LIST_REMOVE(bucket, ub_link);
2363 ZONE_UNLOCK(zone);
2364 cache->uc_freebucket = bucket;
2365 goto zfree_start;
2366 }
2367 /* We are no longer associated with this CPU. */
2368 critical_exit();
2369
2370 /* And the zone.. */
2371 ZONE_UNLOCK(zone);
2372
2373#ifdef UMA_DEBUG_ALLOC
2374 printf("uma_zfree: Allocating new free bucket.\n");
2375#endif
2376 bflags = M_NOWAIT;
2377
2378 if (keg->uk_flags & UMA_ZFLAG_CACHEONLY)
2379 bflags |= M_NOVM;
2380 bucket = bucket_alloc(zone->uz_count, bflags);
2381 if (bucket) {
2382 ZONE_LOCK(zone);
2383 LIST_INSERT_HEAD(&zone->uz_free_bucket,
2384 bucket, ub_link);
2385 ZONE_UNLOCK(zone);
2386 goto zfree_restart;
2387 }
2388
2389 /*
2390 * If nothing else caught this, we'll just do an internal free.
2391 */
2392zfree_internal:
2393 uma_zfree_internal(zone, item, udata, SKIP_DTOR, ZFREE_STATFREE);
2394
2395 return;
2396}
2397
2398/*
2399 * Frees an item to an INTERNAL zone or allocates a free bucket
2400 *
2401 * Arguments:
2402 * zone The zone to free to
2403 * item The item we're freeing
2404 * udata User supplied data for the dtor
2405 * skip Skip dtors and finis
2406 */
2407static void
2408uma_zfree_internal(uma_zone_t zone, void *item, void *udata,
2409 enum zfreeskip skip, int flags)
2410{
2411 uma_slab_t slab;
2412 uma_slabrefcnt_t slabref;
2413 uma_keg_t keg;
2414 u_int8_t *mem;
2415 u_int8_t freei;
2416
2417 keg = zone->uz_keg;
2418
2419 if (skip < SKIP_DTOR && zone->uz_dtor)
2420 zone->uz_dtor(item, keg->uk_size, udata);
2421 if (skip < SKIP_FINI && zone->uz_fini)
2422 zone->uz_fini(item, keg->uk_size);
2423
2424 ZONE_LOCK(zone);
2425
2426 if (flags & ZFREE_STATFAIL)
2427 zone->uz_fails++;
2428 if (flags & ZFREE_STATFREE)
2429 zone->uz_frees++;
2430
2431 if (!(keg->uk_flags & UMA_ZONE_MALLOC)) {
2432 mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK));
2433 if (keg->uk_flags & UMA_ZONE_HASH)
2434 slab = hash_sfind(&keg->uk_hash, mem);
2435 else {
2436 mem += keg->uk_pgoff;
2437 slab = (uma_slab_t)mem;
2438 }
2439 } else {
2440 slab = (uma_slab_t)udata;
2441 }
2442
2443 /* Do we need to remove from any lists? */
2444 if (slab->us_freecount+1 == keg->uk_ipers) {
2445 LIST_REMOVE(slab, us_link);
2446 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2447 } else if (slab->us_freecount == 0) {
2448 LIST_REMOVE(slab, us_link);
2449 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2450 }
2451
2452 /* Slab management stuff */
2453 freei = ((unsigned long)item - (unsigned long)slab->us_data)
2454 / keg->uk_rsize;
2455
2456#ifdef INVARIANTS
2457 if (!skip)
2458 uma_dbg_free(zone, slab, item);
2459#endif
2460
2461 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2462 slabref = (uma_slabrefcnt_t)slab;
2463 slabref->us_freelist[freei].us_item = slab->us_firstfree;
2464 } else {
2465 slab->us_freelist[freei].us_item = slab->us_firstfree;
2466 }
2467 slab->us_firstfree = freei;
2468 slab->us_freecount++;
2469
2470 /* Zone statistics */
2471 keg->uk_free++;
2472
2473 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2474 if (keg->uk_pages < keg->uk_maxpages)
2475 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2476
2477 /*
2478 * We can handle one more allocation. Since we're clearing ZFLAG_FULL,
2479 * wake up all procs blocked on pages. This should be uncommon, so
2480 * keeping this simple for now (rather than adding count of blocked
2481 * threads etc).
2482 */
2483 wakeup(keg);
2484 }
2485
2486 ZONE_UNLOCK(zone);
2487}
2488
2489/* See uma.h */
2490void
2491uma_zone_set_max(uma_zone_t zone, int nitems)
2492{
2493 uma_keg_t keg;
2494
2495 keg = zone->uz_keg;
2496 ZONE_LOCK(zone);
2497 if (keg->uk_ppera > 1)
2498 keg->uk_maxpages = nitems * keg->uk_ppera;
2499 else
2500 keg->uk_maxpages = nitems / keg->uk_ipers;
2501
2502 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2503 keg->uk_maxpages++;
2504
2505 ZONE_UNLOCK(zone);
2506}
2507
2508/* See uma.h */
2509void
2510uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2511{
2512 ZONE_LOCK(zone);
2513 KASSERT(zone->uz_keg->uk_pages == 0,
2514 ("uma_zone_set_init on non-empty keg"));
2515 zone->uz_keg->uk_init = uminit;
2516 ZONE_UNLOCK(zone);
2517}
2518
2519/* See uma.h */
2520void
2521uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2522{
2523 ZONE_LOCK(zone);
2524 KASSERT(zone->uz_keg->uk_pages == 0,
2525 ("uma_zone_set_fini on non-empty keg"));
2526 zone->uz_keg->uk_fini = fini;
2527 ZONE_UNLOCK(zone);
2528}
2529
2530/* See uma.h */
2531void
2532uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2533{
2534 ZONE_LOCK(zone);
2535 KASSERT(zone->uz_keg->uk_pages == 0,
2536 ("uma_zone_set_zinit on non-empty keg"));
2537 zone->uz_init = zinit;
2538 ZONE_UNLOCK(zone);
2539}
2540
2541/* See uma.h */
2542void
2543uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2544{
2545 ZONE_LOCK(zone);
2546 KASSERT(zone->uz_keg->uk_pages == 0,
2547 ("uma_zone_set_zfini on non-empty keg"));
2548 zone->uz_fini = zfini;
2549 ZONE_UNLOCK(zone);
2550}
2551
2552/* See uma.h */
2553/* XXX uk_freef is not actually used with the zone locked */
2554void
2555uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2556{
2557 ZONE_LOCK(zone);
2558 zone->uz_keg->uk_freef = freef;
2559 ZONE_UNLOCK(zone);
2560}
2561
2562/* See uma.h */
2563/* XXX uk_allocf is not actually used with the zone locked */
2564void
2565uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2566{
2567 ZONE_LOCK(zone);
2568 zone->uz_keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2569 zone->uz_keg->uk_allocf = allocf;
2570 ZONE_UNLOCK(zone);
2571}
2572
2573/* See uma.h */
2574int
2575uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count)
2576{
2577 uma_keg_t keg;
2578 vm_offset_t kva;
2579 int pages;
2580
2581 keg = zone->uz_keg;
2582 pages = count / keg->uk_ipers;
2583
2584 if (pages * keg->uk_ipers < count)
2585 pages++;
2586
2587 kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2588
2589 if (kva == 0)
2590 return (0);
2591 if (obj == NULL) {
2592 obj = vm_object_allocate(OBJT_DEFAULT,
2593 pages);
2594 } else {
2595 VM_OBJECT_LOCK_INIT(obj, "uma object");
2596 _vm_object_allocate(OBJT_DEFAULT,
2597 pages, obj);
2598 }
2599 ZONE_LOCK(zone);
2600 keg->uk_kva = kva;
2601 keg->uk_obj = obj;
2602 keg->uk_maxpages = pages;
2603 keg->uk_allocf = obj_alloc;
2604 keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
2605 ZONE_UNLOCK(zone);
2606 return (1);
2607}
2608
2609/* See uma.h */
2610void
2611uma_prealloc(uma_zone_t zone, int items)
2612{
2613 int slabs;
2614 uma_slab_t slab;
2615 uma_keg_t keg;
2616
2617 keg = zone->uz_keg;
2618 ZONE_LOCK(zone);
2619 slabs = items / keg->uk_ipers;
2620 if (slabs * keg->uk_ipers < items)
2621 slabs++;
2622 while (slabs > 0) {
2623 slab = slab_zalloc(zone, M_WAITOK);
2624 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2625 slabs--;
2626 }
2627 ZONE_UNLOCK(zone);
2628}
2629
2630/* See uma.h */
2631u_int32_t *
2632uma_find_refcnt(uma_zone_t zone, void *item)
2633{
2634 uma_slabrefcnt_t slabref;
2635 uma_keg_t keg;
2636 u_int32_t *refcnt;
2637 int idx;
2638
2639 keg = zone->uz_keg;
2640 slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item &
2641 (~UMA_SLAB_MASK));
2642 KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT,
2643 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
2644 idx = ((unsigned long)item - (unsigned long)slabref->us_data)
2645 / keg->uk_rsize;
2646 refcnt = &slabref->us_freelist[idx].us_refcnt;
2647 return refcnt;
2648}
2649
2650/* See uma.h */
2651void
2652uma_reclaim(void)
2653{
2654#ifdef UMA_DEBUG
2655 printf("UMA: vm asked us to release pages!\n");
2656#endif
2657 bucket_enable();
2658 zone_foreach(zone_drain);
2659 /*
2660 * Some slabs may have been freed but this zone will be visited early
2661 * we visit again so that we can free pages that are empty once other
2662 * zones are drained. We have to do the same for buckets.
2663 */
2664 zone_drain(slabzone);
2665 zone_drain(slabrefzone);
2666 bucket_zone_drain();
2667}
2668
2669/* See uma.h */
2670int
2671uma_zone_exhausted(uma_zone_t zone)
2672{
2673 int full;
2674
2675 ZONE_LOCK(zone);
2676 full = (zone->uz_keg->uk_flags & UMA_ZFLAG_FULL);
2677 ZONE_UNLOCK(zone);
2678 return (full);
2679}
2680
2681int
2682uma_zone_exhausted_nolock(uma_zone_t zone)
2683{
2684 return (zone->uz_keg->uk_flags & UMA_ZFLAG_FULL);
2685}
2686
2687void *
2688uma_large_malloc(int size, int wait)
2689{
2690 void *mem;
2691 uma_slab_t slab;
2692 u_int8_t flags;
2693
2694 slab = uma_zalloc_internal(slabzone, NULL, wait);
2695 if (slab == NULL)
2696 return (NULL);
2697 mem = page_alloc(NULL, size, &flags, wait);
2698 if (mem) {
2699 vsetslab((vm_offset_t)mem, slab);
2700 slab->us_data = mem;
2701 slab->us_flags = flags | UMA_SLAB_MALLOC;
2702 slab->us_size = size;
2703 } else {
2704 uma_zfree_internal(slabzone, slab, NULL, SKIP_NONE,
2705 ZFREE_STATFAIL | ZFREE_STATFREE);
2706 }
2707
2708 return (mem);
2709}
2710
2711void
2712uma_large_free(uma_slab_t slab)
2713{
2714 vsetobj((vm_offset_t)slab->us_data, kmem_object);
2715 page_free(slab->us_data, slab->us_size, slab->us_flags);
2716 uma_zfree_internal(slabzone, slab, NULL, SKIP_NONE, ZFREE_STATFREE);
2717}
2718
2719void
2720uma_print_stats(void)
2721{
2722 zone_foreach(uma_print_zone);
2723}
2724
2725static void
2726slab_print(uma_slab_t slab)
2727{
2728 printf("slab: keg %p, data %p, freecount %d, firstfree %d\n",
2729 slab->us_keg, slab->us_data, slab->us_freecount,
2730 slab->us_firstfree);
2731}
2732
2733static void
2734cache_print(uma_cache_t cache)
2735{
2736 printf("alloc: %p(%d), free: %p(%d)\n",
2737 cache->uc_allocbucket,
2738 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
2739 cache->uc_freebucket,
2740 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
2741}
2742
2743void
2744uma_print_zone(uma_zone_t zone)
2745{
2746 uma_cache_t cache;
2747 uma_keg_t keg;
2748 uma_slab_t slab;
2749 int i;
2750
2751 keg = zone->uz_keg;
2752 printf("%s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n",
2753 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
2754 keg->uk_ipers, keg->uk_ppera,
2755 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
2756 printf("Part slabs:\n");
2757 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
2758 slab_print(slab);
2759 printf("Free slabs:\n");
2760 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
2761 slab_print(slab);
2762 printf("Full slabs:\n");
2763 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
2764 slab_print(slab);
2765 for (i = 0; i <= mp_maxid; i++) {
2766 if (CPU_ABSENT(i))
2767 continue;
2768 cache = &zone->uz_cpu[i];
2769 printf("CPU %d Cache:\n", i);
2770 cache_print(cache);
2771 }
2772}
2773
2774#ifdef DDB
2775/*
2776 * Generate statistics across both the zone and its per-cpu cache's. Return
2777 * desired statistics if the pointer is non-NULL for that statistic.
2778 *
2779 * Note: does not update the zone statistics, as it can't safely clear the
2780 * per-CPU cache statistic.
2781 *
2782 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
2783 * safe from off-CPU; we should modify the caches to track this information
2784 * directly so that we don't have to.
2785 */
2786static void
2787uma_zone_sumstat(uma_zone_t z, int *cachefreep, u_int64_t *allocsp,
2788 u_int64_t *freesp)
2789{
2790 uma_cache_t cache;
2791 u_int64_t allocs, frees;
2792 int cachefree, cpu;
2793
2794 allocs = frees = 0;
2795 cachefree = 0;
2796 for (cpu = 0; cpu <= mp_maxid; cpu++) {
2797 if (CPU_ABSENT(cpu))
2798 continue;
2799 cache = &z->uz_cpu[cpu];
2800 if (cache->uc_allocbucket != NULL)
2801 cachefree += cache->uc_allocbucket->ub_cnt;
2802 if (cache->uc_freebucket != NULL)
2803 cachefree += cache->uc_freebucket->ub_cnt;
2804 allocs += cache->uc_allocs;
2805 frees += cache->uc_frees;
2806 }
2807 allocs += z->uz_allocs;
2808 frees += z->uz_frees;
2809 if (cachefreep != NULL)
2810 *cachefreep = cachefree;
2811 if (allocsp != NULL)
2812 *allocsp = allocs;
2813 if (freesp != NULL)
2814 *freesp = frees;
2815}
2816#endif /* DDB */
2817
2818static int
2819sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
2820{
2821 uma_keg_t kz;
2822 uma_zone_t z;
2823 int count;
2824
2825 count = 0;
2826 mtx_lock(&uma_mtx);
2827 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2828 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2829 count++;
2830 }
2831 mtx_unlock(&uma_mtx);
2832 return (sysctl_handle_int(oidp, &count, 0, req));
2833}
2834
2835static int
2836sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
2837{
2838 struct uma_stream_header ush;
2839 struct uma_type_header uth;
2840 struct uma_percpu_stat ups;
2841 uma_bucket_t bucket;
2842 struct sbuf sbuf;
2843 uma_cache_t cache;
2844 uma_keg_t kz;
2845 uma_zone_t z;
2846 char *buffer;
2847 int buflen, count, error, i;
2848
2849 mtx_lock(&uma_mtx);
2850restart:
2851 mtx_assert(&uma_mtx, MA_OWNED);
2852 count = 0;
2853 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2854 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2855 count++;
2856 }
2857 mtx_unlock(&uma_mtx);
2858
2859 buflen = sizeof(ush) + count * (sizeof(uth) + sizeof(ups) *
2860 (mp_maxid + 1)) + 1;
2861 buffer = malloc(buflen, M_TEMP, M_WAITOK | M_ZERO);
2862
2863 mtx_lock(&uma_mtx);
2864 i = 0;
2865 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2866 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2867 i++;
2868 }
2869 if (i > count) {
2870 free(buffer, M_TEMP);
2871 goto restart;
2872 }
2873 count = i;
2874
2875 sbuf_new(&sbuf, buffer, buflen, SBUF_FIXEDLEN);
2876
2877 /*
2878 * Insert stream header.
2879 */
2880 bzero(&ush, sizeof(ush));
2881 ush.ush_version = UMA_STREAM_VERSION;
2882 ush.ush_maxcpus = (mp_maxid + 1);
2883 ush.ush_count = count;
2884 if (sbuf_bcat(&sbuf, &ush, sizeof(ush)) < 0) {
2885 mtx_unlock(&uma_mtx);
2886 error = ENOMEM;
2887 goto out;
2888 }
2889
2890 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2891 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
2892 bzero(&uth, sizeof(uth));
2893 ZONE_LOCK(z);
2894 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
2895 uth.uth_align = kz->uk_align;
2896 uth.uth_pages = kz->uk_pages;
2897 uth.uth_keg_free = kz->uk_free;
2898 uth.uth_size = kz->uk_size;
2899 uth.uth_rsize = kz->uk_rsize;
2900 uth.uth_maxpages = kz->uk_maxpages;
2901 if (kz->uk_ppera > 1)
2902 uth.uth_limit = kz->uk_maxpages /
2903 kz->uk_ppera;
2904 else
2905 uth.uth_limit = kz->uk_maxpages *
2906 kz->uk_ipers;
2907
2908 /*
2909 * A zone is secondary is it is not the first entry
2910 * on the keg's zone list.
2911 */
2912 if ((kz->uk_flags & UMA_ZONE_SECONDARY) &&
2913 (LIST_FIRST(&kz->uk_zones) != z))
2914 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
2915
2916 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
2917 uth.uth_zone_free += bucket->ub_cnt;
2918 uth.uth_allocs = z->uz_allocs;
2919 uth.uth_frees = z->uz_frees;
2920 uth.uth_fails = z->uz_fails;
2921 if (sbuf_bcat(&sbuf, &uth, sizeof(uth)) < 0) {
2922 ZONE_UNLOCK(z);
2923 mtx_unlock(&uma_mtx);
2924 error = ENOMEM;
2925 goto out;
2926 }
2927 /*
2928 * While it is not normally safe to access the cache
2929 * bucket pointers while not on the CPU that owns the
2930 * cache, we only allow the pointers to be exchanged
2931 * without the zone lock held, not invalidated, so
2932 * accept the possible race associated with bucket
2933 * exchange during monitoring.
2934 */
2935 for (i = 0; i < (mp_maxid + 1); i++) {
2936 bzero(&ups, sizeof(ups));
2937 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
2938 goto skip;
2939 if (CPU_ABSENT(i))
2940 goto skip;
2941 cache = &z->uz_cpu[i];
2942 if (cache->uc_allocbucket != NULL)
2943 ups.ups_cache_free +=
2944 cache->uc_allocbucket->ub_cnt;
2945 if (cache->uc_freebucket != NULL)
2946 ups.ups_cache_free +=
2947 cache->uc_freebucket->ub_cnt;
2948 ups.ups_allocs = cache->uc_allocs;
2949 ups.ups_frees = cache->uc_frees;
2950skip:
2951 if (sbuf_bcat(&sbuf, &ups, sizeof(ups)) < 0) {
2952 ZONE_UNLOCK(z);
2953 mtx_unlock(&uma_mtx);
2954 error = ENOMEM;
2955 goto out;
2956 }
2957 }
2958 ZONE_UNLOCK(z);
2959 }
2960 }
2961 mtx_unlock(&uma_mtx);
2962 sbuf_finish(&sbuf);
2963 error = SYSCTL_OUT(req, sbuf_data(&sbuf), sbuf_len(&sbuf));
2964out:
2965 free(buffer, M_TEMP);
2966 return (error);
2967}
2968
2969#ifdef DDB
2970DB_SHOW_COMMAND(uma, db_show_uma)
2971{
2972 u_int64_t allocs, frees;
2973 uma_bucket_t bucket;
2974 uma_keg_t kz;
2975 uma_zone_t z;
2976 int cachefree;
2977
2978 db_printf("%18s %8s %8s %8s %12s\n", "Zone", "Size", "Used", "Free",
2979 "Requests");
2980 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2981 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
2982 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
2983 allocs = z->uz_allocs;
2984 frees = z->uz_frees;
2985 cachefree = 0;
2986 } else
2987 uma_zone_sumstat(z, &cachefree, &allocs,
2988 &frees);
2989 if (!((kz->uk_flags & UMA_ZONE_SECONDARY) &&
2990 (LIST_FIRST(&kz->uk_zones) != z)))
2991 cachefree += kz->uk_free;
2992 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
2993 cachefree += bucket->ub_cnt;
2994 db_printf("%18s %8ju %8jd %8d %12ju\n", z->uz_name,
2995 (uintmax_t)kz->uk_size,
2996 (intmax_t)(allocs - frees), cachefree,
2997 (uintmax_t)allocs);
2998 }
2999 }
3000}
3001#endif