1/*-
2 * Copyright (c) 2002-2005, 2009 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 242152 2012-10-26 17:51:05Z mdf $");
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#include "opt_vm.h"
63
64#include <sys/param.h>
65#include <sys/systm.h>
66#include <sys/kernel.h>
67#include <sys/types.h>
68#include <sys/queue.h>
69#include <sys/malloc.h>
70#include <sys/ktr.h>
71#include <sys/lock.h>
72#include <sys/sysctl.h>
73#include <sys/mutex.h>
74#include <sys/proc.h>
75#include <sys/sbuf.h>
76#include <sys/smp.h>
77#include <sys/vmmeter.h>
78
79#include <vm/vm.h>
80#include <vm/vm_object.h>
81#include <vm/vm_page.h>
82#include <vm/vm_param.h>
83#include <vm/vm_map.h>
84#include <vm/vm_kern.h>
85#include <vm/vm_extern.h>
86#include <vm/uma.h>
87#include <vm/uma_int.h>
88#include <vm/uma_dbg.h>
89
90#include <ddb/ddb.h>
91
92#ifdef DEBUG_MEMGUARD
93#include <vm/memguard.h>
94#endif
95
96/*
97 * This is the zone and keg from which all zones are spawned. The idea is that
98 * even the zone & keg heads are allocated from the allocator, so we use the
99 * bss section to bootstrap us.
100 */
101static struct uma_keg masterkeg;
102static struct uma_zone masterzone_k;
103static struct uma_zone masterzone_z;
104static uma_zone_t kegs = &masterzone_k;
105static uma_zone_t zones = &masterzone_z;
106
107/* This is the zone from which all of uma_slab_t's are allocated. */
108static uma_zone_t slabzone;
109static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
110
111/*
112 * The initial hash tables come out of this zone so they can be allocated
113 * prior to malloc coming up.
114 */
115static uma_zone_t hashzone;
116
117/* The boot-time adjusted value for cache line alignment. */
118int uma_align_cache = 64 - 1;
119
120static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
121
122/*
123 * Are we allowed to allocate buckets?
124 */
125static int bucketdisable = 1;
126
127/* Linked list of all kegs in the system */
128static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
129
130/* This mutex protects the keg list */
131static struct mtx uma_mtx;
132
133/* Linked list of boot time pages */
134static LIST_HEAD(,uma_slab) uma_boot_pages =
135 LIST_HEAD_INITIALIZER(uma_boot_pages);
136
137/* This mutex protects the boot time pages list */
138static struct mtx uma_boot_pages_mtx;
139
140/* Is the VM done starting up? */
141static int booted = 0;
142#define UMA_STARTUP 1
143#define UMA_STARTUP2 2
144
145/* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
146static u_int uma_max_ipers;
147static u_int uma_max_ipers_ref;
148
149/*
150 * This is the handle used to schedule events that need to happen
151 * outside of the allocation fast path.
152 */
153static struct callout uma_callout;
154#define UMA_TIMEOUT 20 /* Seconds for callout interval. */
155
156/*
157 * This structure is passed as the zone ctor arg so that I don't have to create
158 * a special allocation function just for zones.
159 */
160struct uma_zctor_args {
161 const char *name;
162 size_t size;
163 uma_ctor ctor;
164 uma_dtor dtor;
165 uma_init uminit;
166 uma_fini fini;
167 uma_keg_t keg;
168 int align;
169 u_int32_t flags;
170};
171
172struct uma_kctor_args {
173 uma_zone_t zone;
174 size_t size;
175 uma_init uminit;
176 uma_fini fini;
177 int align;
178 u_int32_t flags;
179};
180
181struct uma_bucket_zone {
182 uma_zone_t ubz_zone;
183 char *ubz_name;
184 int ubz_entries;
185};
186
187#define BUCKET_MAX 128
188
189struct uma_bucket_zone bucket_zones[] = {
190 { NULL, "16 Bucket", 16 },
191 { NULL, "32 Bucket", 32 },
192 { NULL, "64 Bucket", 64 },
193 { NULL, "128 Bucket", 128 },
194 { NULL, NULL, 0}
195};
196
197#define BUCKET_SHIFT 4
198#define BUCKET_ZONES ((BUCKET_MAX >> BUCKET_SHIFT) + 1)
199
200/*
201 * bucket_size[] maps requested bucket sizes to zones that allocate a bucket
202 * of approximately the right size.
203 */
204static uint8_t bucket_size[BUCKET_ZONES];
205
206/*
207 * Flags and enumerations to be passed to internal functions.
208 */
209enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI };
210
211#define ZFREE_STATFAIL 0x00000001 /* Update zone failure statistic. */
212#define ZFREE_STATFREE 0x00000002 /* Update zone free statistic. */
213
214/* Prototypes.. */
215
216static void *obj_alloc(uma_zone_t, int, u_int8_t *, int);
217static void *page_alloc(uma_zone_t, int, u_int8_t *, int);
218static void *startup_alloc(uma_zone_t, int, u_int8_t *, int);
219static void page_free(void *, int, u_int8_t);
220static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
221static void cache_drain(uma_zone_t);
222static void bucket_drain(uma_zone_t, uma_bucket_t);
223static void bucket_cache_drain(uma_zone_t zone);
224static int keg_ctor(void *, int, void *, int);
225static void keg_dtor(void *, int, void *);
226static int zone_ctor(void *, int, void *, int);
227static void zone_dtor(void *, int, void *);
228static int zero_init(void *, int, int);
229static void keg_small_init(uma_keg_t keg);
230static void keg_large_init(uma_keg_t keg);
231static void zone_foreach(void (*zfunc)(uma_zone_t));
232static void zone_timeout(uma_zone_t zone);
233static int hash_alloc(struct uma_hash *);
234static int hash_expand(struct uma_hash *, struct uma_hash *);
235static void hash_free(struct uma_hash *hash);
236static void uma_timeout(void *);
237static void uma_startup3(void);
238static void *zone_alloc_item(uma_zone_t, void *, int);
239static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip,
240 int);
241static void bucket_enable(void);
242static void bucket_init(void);
243static uma_bucket_t bucket_alloc(int, int);
244static void bucket_free(uma_bucket_t);
245static void bucket_zone_drain(void);
246static int zone_alloc_bucket(uma_zone_t zone, int flags);
247static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
248static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
249static void *slab_alloc_item(uma_zone_t zone, uma_slab_t slab);
250static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
251 uma_fini fini, int align, u_int32_t flags);
252static inline void zone_relock(uma_zone_t zone, uma_keg_t keg);
253static inline void keg_relock(uma_keg_t keg, uma_zone_t zone);
254
255void uma_print_zone(uma_zone_t);
256void uma_print_stats(void);
257static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
258static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
259
260SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
261
262SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
263 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
264
265SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
266 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
267
268/*
269 * This routine checks to see whether or not it's safe to enable buckets.
270 */
271
272static void
273bucket_enable(void)
274{
275 bucketdisable = vm_page_count_min();
276}
277
278/*
279 * Initialize bucket_zones, the array of zones of buckets of various sizes.
280 *
281 * For each zone, calculate the memory required for each bucket, consisting
282 * of the header and an array of pointers. Initialize bucket_size[] to point
283 * the range of appropriate bucket sizes at the zone.
284 */
285static void
286bucket_init(void)
287{
288 struct uma_bucket_zone *ubz;
289 int i;
290 int j;
291
292 for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) {
293 int size;
294
295 ubz = &bucket_zones[j];
296 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
297 size += sizeof(void *) * ubz->ubz_entries;
298 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
299 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
300 UMA_ZFLAG_INTERNAL | UMA_ZFLAG_BUCKET);
301 for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT))
302 bucket_size[i >> BUCKET_SHIFT] = j;
303 }
304}
305
306/*
307 * Given a desired number of entries for a bucket, return the zone from which
308 * to allocate the bucket.
309 */
310static struct uma_bucket_zone *
311bucket_zone_lookup(int entries)
312{
313 int idx;
314
315 idx = howmany(entries, 1 << BUCKET_SHIFT);
316 return (&bucket_zones[bucket_size[idx]]);
317}
318
319static uma_bucket_t
320bucket_alloc(int entries, int bflags)
321{
322 struct uma_bucket_zone *ubz;
323 uma_bucket_t bucket;
324
325 /*
326 * This is to stop us from allocating per cpu buckets while we're
327 * running out of vm.boot_pages. Otherwise, we would exhaust the
328 * boot pages. This also prevents us from allocating buckets in
329 * low memory situations.
330 */
331 if (bucketdisable)
332 return (NULL);
333
334 ubz = bucket_zone_lookup(entries);
335 bucket = zone_alloc_item(ubz->ubz_zone, NULL, bflags);
336 if (bucket) {
337#ifdef INVARIANTS
338 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
339#endif
340 bucket->ub_cnt = 0;
341 bucket->ub_entries = ubz->ubz_entries;
342 }
343
344 return (bucket);
345}
346
347static void
348bucket_free(uma_bucket_t bucket)
349{
350 struct uma_bucket_zone *ubz;
351
352 ubz = bucket_zone_lookup(bucket->ub_entries);
353 zone_free_item(ubz->ubz_zone, bucket, NULL, SKIP_NONE,
354 ZFREE_STATFREE);
355}
356
357static void
358bucket_zone_drain(void)
359{
360 struct uma_bucket_zone *ubz;
361
362 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
363 zone_drain(ubz->ubz_zone);
364}
365
366static inline uma_keg_t
367zone_first_keg(uma_zone_t zone)
368{
369
370 return (LIST_FIRST(&zone->uz_kegs)->kl_keg);
371}
372
373static void
374zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
375{
376 uma_klink_t klink;
377
378 LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
379 kegfn(klink->kl_keg);
380}
381
382/*
383 * Routine called by timeout which is used to fire off some time interval
384 * based calculations. (stats, hash size, etc.)
385 *
386 * Arguments:
387 * arg Unused
388 *
389 * Returns:
390 * Nothing
391 */
392static void
393uma_timeout(void *unused)
394{
395 bucket_enable();
396 zone_foreach(zone_timeout);
397
398 /* Reschedule this event */
399 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
400}
401
402/*
403 * Routine to perform timeout driven calculations. This expands the
404 * hashes and does per cpu statistics aggregation.
405 *
406 * Returns nothing.
407 */
408static void
409keg_timeout(uma_keg_t keg)
410{
411
412 KEG_LOCK(keg);
413 /*
414 * Expand the keg hash table.
415 *
416 * This is done if the number of slabs is larger than the hash size.
417 * What I'm trying to do here is completely reduce collisions. This
418 * may be a little aggressive. Should I allow for two collisions max?
419 */
420 if (keg->uk_flags & UMA_ZONE_HASH &&
421 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
422 struct uma_hash newhash;
423 struct uma_hash oldhash;
424 int ret;
425
426 /*
427 * This is so involved because allocating and freeing
428 * while the keg lock is held will lead to deadlock.
429 * I have to do everything in stages and check for
430 * races.
431 */
432 newhash = keg->uk_hash;
433 KEG_UNLOCK(keg);
434 ret = hash_alloc(&newhash);
435 KEG_LOCK(keg);
436 if (ret) {
437 if (hash_expand(&keg->uk_hash, &newhash)) {
438 oldhash = keg->uk_hash;
439 keg->uk_hash = newhash;
440 } else
441 oldhash = newhash;
442
443 KEG_UNLOCK(keg);
444 hash_free(&oldhash);
445 KEG_LOCK(keg);
446 }
447 }
448 KEG_UNLOCK(keg);
449}
450
451static void
452zone_timeout(uma_zone_t zone)
453{
454
455 zone_foreach_keg(zone, &keg_timeout);
456}
457
458/*
459 * Allocate and zero fill the next sized hash table from the appropriate
460 * backing store.
461 *
462 * Arguments:
463 * hash A new hash structure with the old hash size in uh_hashsize
464 *
465 * Returns:
466 * 1 on sucess and 0 on failure.
467 */
468static int
469hash_alloc(struct uma_hash *hash)
470{
471 int oldsize;
472 int alloc;
473
474 oldsize = hash->uh_hashsize;
475
476 /* We're just going to go to a power of two greater */
477 if (oldsize) {
478 hash->uh_hashsize = oldsize * 2;
479 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
480 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
481 M_UMAHASH, M_NOWAIT);
482 } else {
483 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
484 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
485 M_WAITOK);
486 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
487 }
488 if (hash->uh_slab_hash) {
489 bzero(hash->uh_slab_hash, alloc);
490 hash->uh_hashmask = hash->uh_hashsize - 1;
491 return (1);
492 }
493
494 return (0);
495}
496
497/*
498 * Expands the hash table for HASH zones. This is done from zone_timeout
499 * to reduce collisions. This must not be done in the regular allocation
500 * path, otherwise, we can recurse on the vm while allocating pages.
501 *
502 * Arguments:
503 * oldhash The hash you want to expand
504 * newhash The hash structure for the new table
505 *
506 * Returns:
507 * Nothing
508 *
509 * Discussion:
510 */
511static int
512hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
513{
514 uma_slab_t slab;
515 int hval;
516 int i;
517
518 if (!newhash->uh_slab_hash)
519 return (0);
520
521 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
522 return (0);
523
524 /*
525 * I need to investigate hash algorithms for resizing without a
526 * full rehash.
527 */
528
529 for (i = 0; i < oldhash->uh_hashsize; i++)
530 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
531 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
532 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
533 hval = UMA_HASH(newhash, slab->us_data);
534 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
535 slab, us_hlink);
536 }
537
538 return (1);
539}
540
541/*
542 * Free the hash bucket to the appropriate backing store.
543 *
544 * Arguments:
545 * slab_hash The hash bucket we're freeing
546 * hashsize The number of entries in that hash bucket
547 *
548 * Returns:
549 * Nothing
550 */
551static void
552hash_free(struct uma_hash *hash)
553{
554 if (hash->uh_slab_hash == NULL)
555 return;
556 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
557 zone_free_item(hashzone,
558 hash->uh_slab_hash, NULL, SKIP_NONE, ZFREE_STATFREE);
559 else
560 free(hash->uh_slab_hash, M_UMAHASH);
561}
562
563/*
564 * Frees all outstanding items in a bucket
565 *
566 * Arguments:
567 * zone The zone to free to, must be unlocked.
568 * bucket The free/alloc bucket with items, cpu queue must be locked.
569 *
570 * Returns:
571 * Nothing
572 */
573
574static void
575bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
576{
577 void *item;
578
579 if (bucket == NULL)
580 return;
581
582 while (bucket->ub_cnt > 0) {
583 bucket->ub_cnt--;
584 item = bucket->ub_bucket[bucket->ub_cnt];
585#ifdef INVARIANTS
586 bucket->ub_bucket[bucket->ub_cnt] = NULL;
587 KASSERT(item != NULL,
588 ("bucket_drain: botched ptr, item is NULL"));
589#endif
590 zone_free_item(zone, item, NULL, SKIP_DTOR, 0);
591 }
592}
593
594/*
595 * Drains the per cpu caches for a zone.
596 *
597 * NOTE: This may only be called while the zone is being turn down, and not
598 * during normal operation. This is necessary in order that we do not have
599 * to migrate CPUs to drain the per-CPU caches.
600 *
601 * Arguments:
602 * zone The zone to drain, must be unlocked.
603 *
604 * Returns:
605 * Nothing
606 */
607static void
608cache_drain(uma_zone_t zone)
609{
610 uma_cache_t cache;
611 int cpu;
612
613 /*
614 * XXX: It is safe to not lock the per-CPU caches, because we're
615 * tearing down the zone anyway. I.e., there will be no further use
616 * of the caches at this point.
617 *
618 * XXX: It would good to be able to assert that the zone is being
619 * torn down to prevent improper use of cache_drain().
620 *
621 * XXX: We lock the zone before passing into bucket_cache_drain() as
622 * it is used elsewhere. Should the tear-down path be made special
623 * there in some form?
624 */
625 CPU_FOREACH(cpu) {
626 cache = &zone->uz_cpu[cpu];
627 bucket_drain(zone, cache->uc_allocbucket);
628 bucket_drain(zone, cache->uc_freebucket);
629 if (cache->uc_allocbucket != NULL)
630 bucket_free(cache->uc_allocbucket);
631 if (cache->uc_freebucket != NULL)
632 bucket_free(cache->uc_freebucket);
633 cache->uc_allocbucket = cache->uc_freebucket = NULL;
634 }
635 ZONE_LOCK(zone);
636 bucket_cache_drain(zone);
637 ZONE_UNLOCK(zone);
638}
639
640/*
641 * Drain the cached buckets from a zone. Expects a locked zone on entry.
642 */
643static void
644bucket_cache_drain(uma_zone_t zone)
645{
646 uma_bucket_t bucket;
647
648 /*
649 * Drain the bucket queues and free the buckets, we just keep two per
650 * cpu (alloc/free).
651 */
652 while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
653 LIST_REMOVE(bucket, ub_link);
654 ZONE_UNLOCK(zone);
655 bucket_drain(zone, bucket);
656 bucket_free(bucket);
657 ZONE_LOCK(zone);
658 }
659
660 /* Now we do the free queue.. */
661 while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
662 LIST_REMOVE(bucket, ub_link);
663 bucket_free(bucket);
664 }
665}
666
667/*
668 * Frees pages from a keg back to the system. This is done on demand from
669 * the pageout daemon.
670 *
671 * Returns nothing.
672 */
673static void
674keg_drain(uma_keg_t keg)
675{
676 struct slabhead freeslabs = { 0 };
677 uma_slab_t slab;
678 uma_slab_t n;
679 u_int8_t flags;
680 u_int8_t *mem;
681 int i;
682
683 /*
684 * We don't want to take pages from statically allocated kegs at this
685 * time
686 */
687 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
688 return;
689
690#ifdef UMA_DEBUG
691 printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
692#endif
693 KEG_LOCK(keg);
694 if (keg->uk_free == 0)
695 goto finished;
696
697 slab = LIST_FIRST(&keg->uk_free_slab);
698 while (slab) {
699 n = LIST_NEXT(slab, us_link);
700
701 /* We have no where to free these to */
702 if (slab->us_flags & UMA_SLAB_BOOT) {
703 slab = n;
704 continue;
705 }
706
707 LIST_REMOVE(slab, us_link);
708 keg->uk_pages -= keg->uk_ppera;
709 keg->uk_free -= keg->uk_ipers;
710
711 if (keg->uk_flags & UMA_ZONE_HASH)
712 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
713
714 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
715
716 slab = n;
717 }
718finished:
719 KEG_UNLOCK(keg);
720
721 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
722 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
723 if (keg->uk_fini)
724 for (i = 0; i < keg->uk_ipers; i++)
725 keg->uk_fini(
726 slab->us_data + (keg->uk_rsize * i),
727 keg->uk_size);
728 flags = slab->us_flags;
729 mem = slab->us_data;
730
731 if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
732 vm_object_t obj;
733
734 if (flags & UMA_SLAB_KMEM)
735 obj = kmem_object;
736 else if (flags & UMA_SLAB_KERNEL)
737 obj = kernel_object;
738 else
739 obj = NULL;
740 for (i = 0; i < keg->uk_ppera; i++)
741 vsetobj((vm_offset_t)mem + (i * PAGE_SIZE),
742 obj);
743 }
744 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
745 zone_free_item(keg->uk_slabzone, slab, NULL,
746 SKIP_NONE, ZFREE_STATFREE);
747#ifdef UMA_DEBUG
748 printf("%s: Returning %d bytes.\n",
749 keg->uk_name, UMA_SLAB_SIZE * keg->uk_ppera);
750#endif
751 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags);
752 }
753}
754
755static void
756zone_drain_wait(uma_zone_t zone, int waitok)
757{
758
759 /*
760 * Set draining to interlock with zone_dtor() so we can release our
761 * locks as we go. Only dtor() should do a WAITOK call since it
762 * is the only call that knows the structure will still be available
763 * when it wakes up.
764 */
765 ZONE_LOCK(zone);
766 while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
767 if (waitok == M_NOWAIT)
768 goto out;
769 mtx_unlock(&uma_mtx);
770 msleep(zone, zone->uz_lock, PVM, "zonedrain", 1);
771 mtx_lock(&uma_mtx);
772 }
773 zone->uz_flags |= UMA_ZFLAG_DRAINING;
774 bucket_cache_drain(zone);
775 ZONE_UNLOCK(zone);
776 /*
777 * The DRAINING flag protects us from being freed while
778 * we're running. Normally the uma_mtx would protect us but we
779 * must be able to release and acquire the right lock for each keg.
780 */
781 zone_foreach_keg(zone, &keg_drain);
782 ZONE_LOCK(zone);
783 zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
784 wakeup(zone);
785out:
786 ZONE_UNLOCK(zone);
787}
788
789void
790zone_drain(uma_zone_t zone)
791{
792
793 zone_drain_wait(zone, M_NOWAIT);
794}
795
796/*
797 * Allocate a new slab for a keg. This does not insert the slab onto a list.
798 *
799 * Arguments:
800 * wait Shall we wait?
801 *
802 * Returns:
803 * The slab that was allocated or NULL if there is no memory and the
804 * caller specified M_NOWAIT.
805 */
806static uma_slab_t
807keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
808{
809 uma_slabrefcnt_t slabref;
810 uma_alloc allocf;
811 uma_slab_t slab;
812 u_int8_t *mem;
813 u_int8_t flags;
814 int i;
815
816 mtx_assert(&keg->uk_lock, MA_OWNED);
817 slab = NULL;
818
819#ifdef UMA_DEBUG
820 printf("slab_zalloc: Allocating a new slab for %s\n", keg->uk_name);
821#endif
822 allocf = keg->uk_allocf;
823 KEG_UNLOCK(keg);
824
825 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
826 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
827 if (slab == NULL) {
828 KEG_LOCK(keg);
829 return NULL;
830 }
831 }
832
833 /*
834 * This reproduces the old vm_zone behavior of zero filling pages the
835 * first time they are added to a zone.
836 *
837 * Malloced items are zeroed in uma_zalloc.
838 */
839
840 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
841 wait |= M_ZERO;
842 else
843 wait &= ~M_ZERO;
844
845 if (keg->uk_flags & UMA_ZONE_NODUMP)
846 wait |= M_NODUMP;
847
848 /* zone is passed for legacy reasons. */
849 mem = allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE, &flags, wait);
850 if (mem == NULL) {
851 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
852 zone_free_item(keg->uk_slabzone, slab, NULL,
853 SKIP_NONE, ZFREE_STATFREE);
854 KEG_LOCK(keg);
855 return (NULL);
856 }
857
858 /* Point the slab into the allocated memory */
859 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
860 slab = (uma_slab_t )(mem + keg->uk_pgoff);
861
862 if (keg->uk_flags & UMA_ZONE_VTOSLAB)
863 for (i = 0; i < keg->uk_ppera; i++)
864 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
865
866 slab->us_keg = keg;
867 slab->us_data = mem;
868 slab->us_freecount = keg->uk_ipers;
869 slab->us_firstfree = 0;
870 slab->us_flags = flags;
871
872 if (keg->uk_flags & UMA_ZONE_REFCNT) {
873 slabref = (uma_slabrefcnt_t)slab;
874 for (i = 0; i < keg->uk_ipers; i++) {
875 slabref->us_freelist[i].us_refcnt = 0;
876 slabref->us_freelist[i].us_item = i+1;
877 }
878 } else {
879 for (i = 0; i < keg->uk_ipers; i++)
880 slab->us_freelist[i].us_item = i+1;
881 }
882
883 if (keg->uk_init != NULL) {
884 for (i = 0; i < keg->uk_ipers; i++)
885 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
886 keg->uk_size, wait) != 0)
887 break;
888 if (i != keg->uk_ipers) {
889 if (keg->uk_fini != NULL) {
890 for (i--; i > -1; i--)
891 keg->uk_fini(slab->us_data +
892 (keg->uk_rsize * i),
893 keg->uk_size);
894 }
895 if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
896 vm_object_t obj;
897
898 if (flags & UMA_SLAB_KMEM)
899 obj = kmem_object;
900 else if (flags & UMA_SLAB_KERNEL)
901 obj = kernel_object;
902 else
903 obj = NULL;
904 for (i = 0; i < keg->uk_ppera; i++)
905 vsetobj((vm_offset_t)mem +
906 (i * PAGE_SIZE), obj);
907 }
908 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
909 zone_free_item(keg->uk_slabzone, slab,
910 NULL, SKIP_NONE, ZFREE_STATFREE);
911 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera,
912 flags);
913 KEG_LOCK(keg);
914 return (NULL);
915 }
916 }
917 KEG_LOCK(keg);
918
919 if (keg->uk_flags & UMA_ZONE_HASH)
920 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
921
922 keg->uk_pages += keg->uk_ppera;
923 keg->uk_free += keg->uk_ipers;
924
925 return (slab);
926}
927
928/*
929 * This function is intended to be used early on in place of page_alloc() so
930 * that we may use the boot time page cache to satisfy allocations before
931 * the VM is ready.
932 */
933static void *
934startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
935{
936 uma_keg_t keg;
937 uma_slab_t tmps;
938 int pages, check_pages;
939
940 keg = zone_first_keg(zone);
941 pages = howmany(bytes, PAGE_SIZE);
942 check_pages = pages - 1;
943 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
944
945 /*
946 * Check our small startup cache to see if it has pages remaining.
947 */
948 mtx_lock(&uma_boot_pages_mtx);
949
950 /* First check if we have enough room. */
951 tmps = LIST_FIRST(&uma_boot_pages);
952 while (tmps != NULL && check_pages-- > 0)
953 tmps = LIST_NEXT(tmps, us_link);
954 if (tmps != NULL) {
955 /*
956 * It's ok to lose tmps references. The last one will
957 * have tmps->us_data pointing to the start address of
958 * "pages" contiguous pages of memory.
959 */
960 while (pages-- > 0) {
961 tmps = LIST_FIRST(&uma_boot_pages);
962 LIST_REMOVE(tmps, us_link);
963 }
964 mtx_unlock(&uma_boot_pages_mtx);
965 *pflag = tmps->us_flags;
966 return (tmps->us_data);
967 }
968 mtx_unlock(&uma_boot_pages_mtx);
969 if (booted < UMA_STARTUP2)
970 panic("UMA: Increase vm.boot_pages");
971 /*
972 * Now that we've booted reset these users to their real allocator.
973 */
974#ifdef UMA_MD_SMALL_ALLOC
975 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
976#else
977 keg->uk_allocf = page_alloc;
978#endif
979 return keg->uk_allocf(zone, bytes, pflag, wait);
980}
981
982/*
983 * Allocates a number of pages from the system
984 *
985 * Arguments:
986 * bytes The number of bytes requested
987 * wait Shall we wait?
988 *
989 * Returns:
990 * A pointer to the alloced memory or possibly
991 * NULL if M_NOWAIT is set.
992 */
993static void *
994page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
995{
996 void *p; /* Returned page */
997
998 *pflag = UMA_SLAB_KMEM;
999 p = (void *) kmem_malloc(kmem_map, bytes, wait);
1000
1001 return (p);
1002}
1003
1004/*
1005 * Allocates a number of pages from within an object
1006 *
1007 * Arguments:
1008 * bytes The number of bytes requested
1009 * wait Shall we wait?
1010 *
1011 * Returns:
1012 * A pointer to the alloced memory or possibly
1013 * NULL if M_NOWAIT is set.
1014 */
1015static void *
1016obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
1017{
1018 vm_object_t object;
1019 vm_offset_t retkva, zkva;
1020 vm_page_t p;
1021 int pages, startpages;
1022 uma_keg_t keg;
1023
1024 keg = zone_first_keg(zone);
1025 object = keg->uk_obj;
1026 retkva = 0;
1027
1028 /*
1029 * This looks a little weird since we're getting one page at a time.
1030 */
1031 VM_OBJECT_LOCK(object);
1032 p = TAILQ_LAST(&object->memq, pglist);
1033 pages = p != NULL ? p->pindex + 1 : 0;
1034 startpages = pages;
1035 zkva = keg->uk_kva + pages * PAGE_SIZE;
1036 for (; bytes > 0; bytes -= PAGE_SIZE) {
1037 p = vm_page_alloc(object, pages,
1038 VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED);
1039 if (p == NULL) {
1040 if (pages != startpages)
1041 pmap_qremove(retkva, pages - startpages);
1042 while (pages != startpages) {
1043 pages--;
1044 p = TAILQ_LAST(&object->memq, pglist);
1045 vm_page_unwire(p, 0);
1046 vm_page_free(p);
1047 }
1048 retkva = 0;
1049 goto done;
1050 }
1051 pmap_qenter(zkva, &p, 1);
1052 if (retkva == 0)
1053 retkva = zkva;
1054 zkva += PAGE_SIZE;
1055 pages += 1;
1056 }
1057done:
1058 VM_OBJECT_UNLOCK(object);
1059 *flags = UMA_SLAB_PRIV;
1060
1061 return ((void *)retkva);
1062}
1063
1064/*
1065 * Frees a number of pages to the system
1066 *
1067 * Arguments:
1068 * mem A pointer to the memory to be freed
1069 * size The size of the memory being freed
1070 * flags The original p->us_flags field
1071 *
1072 * Returns:
1073 * Nothing
1074 */
1075static void
1076page_free(void *mem, int size, u_int8_t flags)
1077{
1078 vm_map_t map;
1079
1080 if (flags & UMA_SLAB_KMEM)
1081 map = kmem_map;
1082 else if (flags & UMA_SLAB_KERNEL)
1083 map = kernel_map;
1084 else
1085 panic("UMA: page_free used with invalid flags %d", flags);
1086
1087 kmem_free(map, (vm_offset_t)mem, size);
1088}
1089
1090/*
1091 * Zero fill initializer
1092 *
1093 * Arguments/Returns follow uma_init specifications
1094 */
1095static int
1096zero_init(void *mem, int size, int flags)
1097{
1098 bzero(mem, size);
1099 return (0);
1100}
1101
1102/*
1103 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1104 *
1105 * Arguments
1106 * keg The zone we should initialize
1107 *
1108 * Returns
1109 * Nothing
1110 */
1111static void
1112keg_small_init(uma_keg_t keg)
1113{
1114 u_int rsize;
1115 u_int memused;
1116 u_int wastedspace;
1117 u_int shsize;
1118
1119 KASSERT(keg != NULL, ("Keg is null in keg_small_init"));
1120 rsize = keg->uk_size;
1121
1122 if (rsize < UMA_SMALLEST_UNIT)
1123 rsize = UMA_SMALLEST_UNIT;
1124 if (rsize & keg->uk_align)
1125 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1126
1127 keg->uk_rsize = rsize;
1128 keg->uk_ppera = 1;
1129
1130 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1131 shsize = 0;
1132 } else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1133 rsize += UMA_FRITMREF_SZ; /* linkage & refcnt */
1134 shsize = sizeof(struct uma_slab_refcnt);
1135 } else {
1136 rsize += UMA_FRITM_SZ; /* Account for linkage */
1137 shsize = sizeof(struct uma_slab);
1138 }
1139
1140 keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize;
1141 KASSERT(keg->uk_ipers != 0, ("keg_small_init: ipers is 0"));
1142 memused = keg->uk_ipers * rsize + shsize;
1143 wastedspace = UMA_SLAB_SIZE - memused;
1144
1145 /*
1146 * We can't do OFFPAGE if we're internal or if we've been
1147 * asked to not go to the VM for buckets. If we do this we
1148 * may end up going to the VM (kmem_map) for slabs which we
1149 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1150 * result of UMA_ZONE_VM, which clearly forbids it.
1151 */
1152 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1153 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1154 return;
1155
1156 if ((wastedspace >= UMA_MAX_WASTE) &&
1157 (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) {
1158 keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize;
1159 KASSERT(keg->uk_ipers <= 255,
1160 ("keg_small_init: keg->uk_ipers too high!"));
1161#ifdef UMA_DEBUG
1162 printf("UMA decided we need offpage slab headers for "
1163 "keg: %s, calculated wastedspace = %d, "
1164 "maximum wasted space allowed = %d, "
1165 "calculated ipers = %d, "
1166 "new wasted space = %d\n", keg->uk_name, wastedspace,
1167 UMA_MAX_WASTE, keg->uk_ipers,
1168 UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize);
1169#endif
1170 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1171 if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1172 keg->uk_flags |= UMA_ZONE_HASH;
1173 }
1174}
1175
1176/*
1177 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1178 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1179 * more complicated.
1180 *
1181 * Arguments
1182 * keg The keg we should initialize
1183 *
1184 * Returns
1185 * Nothing
1186 */
1187static void
1188keg_large_init(uma_keg_t keg)
1189{
1190 int pages;
1191
1192 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1193 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1194 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1195
1196 pages = keg->uk_size / UMA_SLAB_SIZE;
1197
1198 /* Account for remainder */
1199 if ((pages * UMA_SLAB_SIZE) < keg->uk_size)
1200 pages++;
1201
1202 keg->uk_ppera = pages;
1203 keg->uk_ipers = 1;
1204 keg->uk_rsize = keg->uk_size;
1205
1206 /* We can't do OFFPAGE if we're internal, bail out here. */
1207 if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1208 return;
1209
1210 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1211 if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1212 keg->uk_flags |= UMA_ZONE_HASH;
1213}
1214
1215static void
1216keg_cachespread_init(uma_keg_t keg)
1217{
1218 int alignsize;
1219 int trailer;
1220 int pages;
1221 int rsize;
1222
1223 alignsize = keg->uk_align + 1;
1224 rsize = keg->uk_size;
1225 /*
1226 * We want one item to start on every align boundary in a page. To
1227 * do this we will span pages. We will also extend the item by the
1228 * size of align if it is an even multiple of align. Otherwise, it
1229 * would fall on the same boundary every time.
1230 */
1231 if (rsize & keg->uk_align)
1232 rsize = (rsize & ~keg->uk_align) + alignsize;
1233 if ((rsize & alignsize) == 0)
1234 rsize += alignsize;
1235 trailer = rsize - keg->uk_size;
1236 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1237 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1238 keg->uk_rsize = rsize;
1239 keg->uk_ppera = pages;
1240 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1241 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1242 KASSERT(keg->uk_ipers <= uma_max_ipers,
1243 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1244 keg->uk_ipers));
1245}
1246
1247/*
1248 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1249 * the keg onto the global keg list.
1250 *
1251 * Arguments/Returns follow uma_ctor specifications
1252 * udata Actually uma_kctor_args
1253 */
1254static int
1255keg_ctor(void *mem, int size, void *udata, int flags)
1256{
1257 struct uma_kctor_args *arg = udata;
1258 uma_keg_t keg = mem;
1259 uma_zone_t zone;
1260
1261 bzero(keg, size);
1262 keg->uk_size = arg->size;
1263 keg->uk_init = arg->uminit;
1264 keg->uk_fini = arg->fini;
1265 keg->uk_align = arg->align;
1266 keg->uk_free = 0;
1267 keg->uk_pages = 0;
1268 keg->uk_flags = arg->flags;
1269 keg->uk_allocf = page_alloc;
1270 keg->uk_freef = page_free;
1271 keg->uk_recurse = 0;
1272 keg->uk_slabzone = NULL;
1273
1274 /*
1275 * The master zone is passed to us at keg-creation time.
1276 */
1277 zone = arg->zone;
1278 keg->uk_name = zone->uz_name;
1279
1280 if (arg->flags & UMA_ZONE_VM)
1281 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1282
1283 if (arg->flags & UMA_ZONE_ZINIT)
1284 keg->uk_init = zero_init;
1285
1286 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1287 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1288
1289 /*
1290 * The +UMA_FRITM_SZ added to uk_size is to account for the
1291 * linkage that is added to the size in keg_small_init(). If
1292 * we don't account for this here then we may end up in
1293 * keg_small_init() with a calculated 'ipers' of 0.
1294 */
1295 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1296 if (keg->uk_flags & UMA_ZONE_CACHESPREAD)
1297 keg_cachespread_init(keg);
1298 else if ((keg->uk_size+UMA_FRITMREF_SZ) >
1299 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)))
1300 keg_large_init(keg);
1301 else
1302 keg_small_init(keg);
1303 } else {
1304 if (keg->uk_flags & UMA_ZONE_CACHESPREAD)
1305 keg_cachespread_init(keg);
1306 else if ((keg->uk_size+UMA_FRITM_SZ) >
1307 (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1308 keg_large_init(keg);
1309 else
1310 keg_small_init(keg);
1311 }
1312
1313 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1314 if (keg->uk_flags & UMA_ZONE_REFCNT)
1315 keg->uk_slabzone = slabrefzone;
1316 else
1317 keg->uk_slabzone = slabzone;
1318 }
1319
1320 /*
1321 * If we haven't booted yet we need allocations to go through the
1322 * startup cache until the vm is ready.
1323 */
1324 if (keg->uk_ppera == 1) {
1325#ifdef UMA_MD_SMALL_ALLOC
1326 keg->uk_allocf = uma_small_alloc;
1327 keg->uk_freef = uma_small_free;
1328
1329 if (booted < UMA_STARTUP)
1330 keg->uk_allocf = startup_alloc;
1331#else
1332 if (booted < UMA_STARTUP2)
1333 keg->uk_allocf = startup_alloc;
1334#endif
1335 } else if (booted < UMA_STARTUP2 &&
1336 (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1337 keg->uk_allocf = startup_alloc;
1338
1339 /*
1340 * Initialize keg's lock (shared among zones).
1341 */
1342 if (arg->flags & UMA_ZONE_MTXCLASS)
1343 KEG_LOCK_INIT(keg, 1);
1344 else
1345 KEG_LOCK_INIT(keg, 0);
1346
1347 /*
1348 * If we're putting the slab header in the actual page we need to
1349 * figure out where in each page it goes. This calculates a right
1350 * justified offset into the memory on an ALIGN_PTR boundary.
1351 */
1352 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1353 u_int totsize;
1354
1355 /* Size of the slab struct and free list */
1356 if (keg->uk_flags & UMA_ZONE_REFCNT)
1357 totsize = sizeof(struct uma_slab_refcnt) +
1358 keg->uk_ipers * UMA_FRITMREF_SZ;
1359 else
1360 totsize = sizeof(struct uma_slab) +
1361 keg->uk_ipers * UMA_FRITM_SZ;
1362
1363 if (totsize & UMA_ALIGN_PTR)
1364 totsize = (totsize & ~UMA_ALIGN_PTR) +
1365 (UMA_ALIGN_PTR + 1);
1366 keg->uk_pgoff = (UMA_SLAB_SIZE * keg->uk_ppera) - totsize;
1367
1368 if (keg->uk_flags & UMA_ZONE_REFCNT)
1369 totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt)
1370 + keg->uk_ipers * UMA_FRITMREF_SZ;
1371 else
1372 totsize = keg->uk_pgoff + sizeof(struct uma_slab)
1373 + keg->uk_ipers * UMA_FRITM_SZ;
1374
1375 /*
1376 * The only way the following is possible is if with our
1377 * UMA_ALIGN_PTR adjustments we are now bigger than
1378 * UMA_SLAB_SIZE. I haven't checked whether this is
1379 * mathematically possible for all cases, so we make
1380 * sure here anyway.
1381 */
1382 if (totsize > UMA_SLAB_SIZE * keg->uk_ppera) {
1383 printf("zone %s ipers %d rsize %d size %d\n",
1384 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1385 keg->uk_size);
1386 panic("UMA slab won't fit.");
1387 }
1388 }
1389
1390 if (keg->uk_flags & UMA_ZONE_HASH)
1391 hash_alloc(&keg->uk_hash);
1392
1393#ifdef UMA_DEBUG
1394 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1395 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1396 keg->uk_ipers, keg->uk_ppera,
1397 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1398#endif
1399
1400 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1401
1402 mtx_lock(&uma_mtx);
1403 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1404 mtx_unlock(&uma_mtx);
1405 return (0);
1406}
1407
1408/*
1409 * Zone header ctor. This initializes all fields, locks, etc.
1410 *
1411 * Arguments/Returns follow uma_ctor specifications
1412 * udata Actually uma_zctor_args
1413 */
1414static int
1415zone_ctor(void *mem, int size, void *udata, int flags)
1416{
1417 struct uma_zctor_args *arg = udata;
1418 uma_zone_t zone = mem;
1419 uma_zone_t z;
1420 uma_keg_t keg;
1421
1422 bzero(zone, size);
1423 zone->uz_name = arg->name;
1424 zone->uz_ctor = arg->ctor;
1425 zone->uz_dtor = arg->dtor;
1426 zone->uz_slab = zone_fetch_slab;
1427 zone->uz_init = NULL;
1428 zone->uz_fini = NULL;
1429 zone->uz_allocs = 0;
1430 zone->uz_frees = 0;
1431 zone->uz_fails = 0;
1432 zone->uz_sleeps = 0;
1433 zone->uz_fills = zone->uz_count = 0;
1434 zone->uz_flags = 0;
1435 keg = arg->keg;
1436
1437 if (arg->flags & UMA_ZONE_SECONDARY) {
1438 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1439 zone->uz_init = arg->uminit;
1440 zone->uz_fini = arg->fini;
1441 zone->uz_lock = &keg->uk_lock;
1442 zone->uz_flags |= UMA_ZONE_SECONDARY;
1443 mtx_lock(&uma_mtx);
1444 ZONE_LOCK(zone);
1445 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1446 if (LIST_NEXT(z, uz_link) == NULL) {
1447 LIST_INSERT_AFTER(z, zone, uz_link);
1448 break;
1449 }
1450 }
1451 ZONE_UNLOCK(zone);
1452 mtx_unlock(&uma_mtx);
1453 } else if (keg == NULL) {
1454 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1455 arg->align, arg->flags)) == NULL)
1456 return (ENOMEM);
1457 } else {
1458 struct uma_kctor_args karg;
1459 int error;
1460
1461 /* We should only be here from uma_startup() */
1462 karg.size = arg->size;
1463 karg.uminit = arg->uminit;
1464 karg.fini = arg->fini;
1465 karg.align = arg->align;
1466 karg.flags = arg->flags;
1467 karg.zone = zone;
1468 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1469 flags);
1470 if (error)
1471 return (error);
1472 }
1473 /*
1474 * Link in the first keg.
1475 */
1476 zone->uz_klink.kl_keg = keg;
1477 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1478 zone->uz_lock = &keg->uk_lock;
1479 zone->uz_size = keg->uk_size;
1480 zone->uz_flags |= (keg->uk_flags &
1481 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1482
1483 /*
1484 * Some internal zones don't have room allocated for the per cpu
1485 * caches. If we're internal, bail out here.
1486 */
1487 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1488 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1489 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1490 return (0);
1491 }
1492
1493 if (keg->uk_flags & UMA_ZONE_MAXBUCKET)
1494 zone->uz_count = BUCKET_MAX;
1495 else if (keg->uk_ipers <= BUCKET_MAX)
1496 zone->uz_count = keg->uk_ipers;
1497 else
1498 zone->uz_count = BUCKET_MAX;
1499 return (0);
1500}
1501
1502/*
1503 * Keg header dtor. This frees all data, destroys locks, frees the hash
1504 * table and removes the keg from the global list.
1505 *
1506 * Arguments/Returns follow uma_dtor specifications
1507 * udata unused
1508 */
1509static void
1510keg_dtor(void *arg, int size, void *udata)
1511{
1512 uma_keg_t keg;
1513
1514 keg = (uma_keg_t)arg;
1515 KEG_LOCK(keg);
1516 if (keg->uk_free != 0) {
1517 printf("Freed UMA keg was not empty (%d items). "
1518 " Lost %d pages of memory.\n",
1519 keg->uk_free, keg->uk_pages);
1520 }
1521 KEG_UNLOCK(keg);
1522
1523 hash_free(&keg->uk_hash);
1524
1525 KEG_LOCK_FINI(keg);
1526}
1527
1528/*
1529 * Zone header dtor.
1530 *
1531 * Arguments/Returns follow uma_dtor specifications
1532 * udata unused
1533 */
1534static void
1535zone_dtor(void *arg, int size, void *udata)
1536{
1537 uma_klink_t klink;
1538 uma_zone_t zone;
1539 uma_keg_t keg;
1540
1541 zone = (uma_zone_t)arg;
1542 keg = zone_first_keg(zone);
1543
1544 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1545 cache_drain(zone);
1546
1547 mtx_lock(&uma_mtx);
1548 LIST_REMOVE(zone, uz_link);
1549 mtx_unlock(&uma_mtx);
1550 /*
1551 * XXX there are some races here where
1552 * the zone can be drained but zone lock
1553 * released and then refilled before we
1554 * remove it... we dont care for now
1555 */
1556 zone_drain_wait(zone, M_WAITOK);
1557 /*
1558 * Unlink all of our kegs.
1559 */
1560 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1561 klink->kl_keg = NULL;
1562 LIST_REMOVE(klink, kl_link);
1563 if (klink == &zone->uz_klink)
1564 continue;
1565 free(klink, M_TEMP);
1566 }
1567 /*
1568 * We only destroy kegs from non secondary zones.
1569 */
1570 if ((zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1571 mtx_lock(&uma_mtx);
1572 LIST_REMOVE(keg, uk_link);
1573 mtx_unlock(&uma_mtx);
1574 zone_free_item(kegs, keg, NULL, SKIP_NONE,
1575 ZFREE_STATFREE);
1576 }
1577}
1578
1579/*
1580 * Traverses every zone in the system and calls a callback
1581 *
1582 * Arguments:
1583 * zfunc A pointer to a function which accepts a zone
1584 * as an argument.
1585 *
1586 * Returns:
1587 * Nothing
1588 */
1589static void
1590zone_foreach(void (*zfunc)(uma_zone_t))
1591{
1592 uma_keg_t keg;
1593 uma_zone_t zone;
1594
1595 mtx_lock(&uma_mtx);
1596 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1597 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1598 zfunc(zone);
1599 }
1600 mtx_unlock(&uma_mtx);
1601}
1602
1603/* Public functions */
1604/* See uma.h */
1605void
1606uma_startup(void *bootmem, int boot_pages)
1607{
1608 struct uma_zctor_args args;
1609 uma_slab_t slab;
1610 u_int slabsize;
1611 u_int objsize, totsize, wsize;
1612 int i;
1613
1614#ifdef UMA_DEBUG
1615 printf("Creating uma keg headers zone and keg.\n");
1616#endif
1617 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1618
1619 /*
1620 * Figure out the maximum number of items-per-slab we'll have if
1621 * we're using the OFFPAGE slab header to track free items, given
1622 * all possible object sizes and the maximum desired wastage
1623 * (UMA_MAX_WASTE).
1624 *
1625 * We iterate until we find an object size for
1626 * which the calculated wastage in keg_small_init() will be
1627 * enough to warrant OFFPAGE. Since wastedspace versus objsize
1628 * is an overall increasing see-saw function, we find the smallest
1629 * objsize such that the wastage is always acceptable for objects
1630 * with that objsize or smaller. Since a smaller objsize always
1631 * generates a larger possible uma_max_ipers, we use this computed
1632 * objsize to calculate the largest ipers possible. Since the
1633 * ipers calculated for OFFPAGE slab headers is always larger than
1634 * the ipers initially calculated in keg_small_init(), we use
1635 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to
1636 * obtain the maximum ipers possible for offpage slab headers.
1637 *
1638 * It should be noted that ipers versus objsize is an inversly
1639 * proportional function which drops off rather quickly so as
1640 * long as our UMA_MAX_WASTE is such that the objsize we calculate
1641 * falls into the portion of the inverse relation AFTER the steep
1642 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386).
1643 *
1644 * Note that we have 8-bits (1 byte) to use as a freelist index
1645 * inside the actual slab header itself and this is enough to
1646 * accomodate us. In the worst case, a UMA_SMALLEST_UNIT sized
1647 * object with offpage slab header would have ipers =
1648 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is
1649 * 1 greater than what our byte-integer freelist index can
1650 * accomodate, but we know that this situation never occurs as
1651 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate
1652 * that we need to go to offpage slab headers. Or, if we do,
1653 * then we trap that condition below and panic in the INVARIANTS case.
1654 */
1655 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE;
1656 totsize = wsize;
1657 objsize = UMA_SMALLEST_UNIT;
1658 while (totsize >= wsize) {
1659 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) /
1660 (objsize + UMA_FRITM_SZ);
1661 totsize *= (UMA_FRITM_SZ + objsize);
1662 objsize++;
1663 }
1664 if (objsize > UMA_SMALLEST_UNIT)
1665 objsize--;
1666 uma_max_ipers = MAX(UMA_SLAB_SIZE / objsize, 64);
1667
1668 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE;
1669 totsize = wsize;
1670 objsize = UMA_SMALLEST_UNIT;
1671 while (totsize >= wsize) {
1672 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) /
1673 (objsize + UMA_FRITMREF_SZ);
1674 totsize *= (UMA_FRITMREF_SZ + objsize);
1675 objsize++;
1676 }
1677 if (objsize > UMA_SMALLEST_UNIT)
1678 objsize--;
1679 uma_max_ipers_ref = MAX(UMA_SLAB_SIZE / objsize, 64);
1680
1681 KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255),
1682 ("uma_startup: calculated uma_max_ipers values too large!"));
1683
1684#ifdef UMA_DEBUG
1685 printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers);
1686 printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n",
1687 uma_max_ipers_ref);
1688#endif
1689
1690 /* "manually" create the initial zone */
1691 args.name = "UMA Kegs";
1692 args.size = sizeof(struct uma_keg);
1693 args.ctor = keg_ctor;
1694 args.dtor = keg_dtor;
1695 args.uminit = zero_init;
1696 args.fini = NULL;
1697 args.keg = &masterkeg;
1698 args.align = 32 - 1;
1699 args.flags = UMA_ZFLAG_INTERNAL;
1700 /* The initial zone has no Per cpu queues so it's smaller */
1701 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1702
1703#ifdef UMA_DEBUG
1704 printf("Filling boot free list.\n");
1705#endif
1706 for (i = 0; i < boot_pages; i++) {
1707 slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE));
1708 slab->us_data = (u_int8_t *)slab;
1709 slab->us_flags = UMA_SLAB_BOOT;
1710 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1711 }
1712 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1713
1714#ifdef UMA_DEBUG
1715 printf("Creating uma zone headers zone and keg.\n");
1716#endif
1717 args.name = "UMA Zones";
1718 args.size = sizeof(struct uma_zone) +
1719 (sizeof(struct uma_cache) * (mp_maxid + 1));
1720 args.ctor = zone_ctor;
1721 args.dtor = zone_dtor;
1722 args.uminit = zero_init;
1723 args.fini = NULL;
1724 args.keg = NULL;
1725 args.align = 32 - 1;
1726 args.flags = UMA_ZFLAG_INTERNAL;
1727 /* The initial zone has no Per cpu queues so it's smaller */
1728 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1729
1730#ifdef UMA_DEBUG
1731 printf("Initializing pcpu cache locks.\n");
1732#endif
1733#ifdef UMA_DEBUG
1734 printf("Creating slab and hash zones.\n");
1735#endif
1736
1737 /*
1738 * This is the max number of free list items we'll have with
1739 * offpage slabs.
1740 */
1741 slabsize = uma_max_ipers * UMA_FRITM_SZ;
1742 slabsize += sizeof(struct uma_slab);
1743
1744 /* Now make a zone for slab headers */
1745 slabzone = uma_zcreate("UMA Slabs",
1746 slabsize,
1747 NULL, NULL, NULL, NULL,
1748 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1749
1750 /*
1751 * We also create a zone for the bigger slabs with reference
1752 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1753 */
1754 slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ;
1755 slabsize += sizeof(struct uma_slab_refcnt);
1756 slabrefzone = uma_zcreate("UMA RCntSlabs",
1757 slabsize,
1758 NULL, NULL, NULL, NULL,
1759 UMA_ALIGN_PTR,
1760 UMA_ZFLAG_INTERNAL);
1761
1762 hashzone = uma_zcreate("UMA Hash",
1763 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1764 NULL, NULL, NULL, NULL,
1765 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1766
1767 bucket_init();
1768
1769 booted = UMA_STARTUP;
1770
1771#ifdef UMA_DEBUG
1772 printf("UMA startup complete.\n");
1773#endif
1774}
1775
1776/* see uma.h */
1777void
1778uma_startup2(void)
1779{
1780 booted = UMA_STARTUP2;
1781 bucket_enable();
1782#ifdef UMA_DEBUG
1783 printf("UMA startup2 complete.\n");
1784#endif
1785}
1786
1787/*
1788 * Initialize our callout handle
1789 *
1790 */
1791
1792static void
1793uma_startup3(void)
1794{
1795#ifdef UMA_DEBUG
1796 printf("Starting callout.\n");
1797#endif
1798 callout_init(&uma_callout, CALLOUT_MPSAFE);
1799 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1800#ifdef UMA_DEBUG
1801 printf("UMA startup3 complete.\n");
1802#endif
1803}
1804
1805static uma_keg_t
1806uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1807 int align, u_int32_t flags)
1808{
1809 struct uma_kctor_args args;
1810
1811 args.size = size;
1812 args.uminit = uminit;
1813 args.fini = fini;
1814 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1815 args.flags = flags;
1816 args.zone = zone;
1817 return (zone_alloc_item(kegs, &args, M_WAITOK));
1818}
1819
1820/* See uma.h */
1821void
1822uma_set_align(int align)
1823{
1824
1825 if (align != UMA_ALIGN_CACHE)
1826 uma_align_cache = align;
1827}
1828
1829/* See uma.h */
1830uma_zone_t
1831uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1832 uma_init uminit, uma_fini fini, int align, u_int32_t flags)
1833
1834{
1835 struct uma_zctor_args args;
1836
1837 /* This stuff is essential for the zone ctor */
1838 args.name = name;
1839 args.size = size;
1840 args.ctor = ctor;
1841 args.dtor = dtor;
1842 args.uminit = uminit;
1843 args.fini = fini;
1844 args.align = align;
1845 args.flags = flags;
1846 args.keg = NULL;
1847
1848 return (zone_alloc_item(zones, &args, M_WAITOK));
1849}
1850
1851/* See uma.h */
1852uma_zone_t
1853uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1854 uma_init zinit, uma_fini zfini, uma_zone_t master)
1855{
1856 struct uma_zctor_args args;
1857 uma_keg_t keg;
1858
1859 keg = zone_first_keg(master);
1860 args.name = name;
1861 args.size = keg->uk_size;
1862 args.ctor = ctor;
1863 args.dtor = dtor;
1864 args.uminit = zinit;
1865 args.fini = zfini;
1866 args.align = keg->uk_align;
1867 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1868 args.keg = keg;
1869
1870 /* XXX Attaches only one keg of potentially many. */
1871 return (zone_alloc_item(zones, &args, M_WAITOK));
1872}
1873
1874static void
1875zone_lock_pair(uma_zone_t a, uma_zone_t b)
1876{
1877 if (a < b) {
1878 ZONE_LOCK(a);
1879 mtx_lock_flags(b->uz_lock, MTX_DUPOK);
1880 } else {
1881 ZONE_LOCK(b);
1882 mtx_lock_flags(a->uz_lock, MTX_DUPOK);
1883 }
1884}
1885
1886static void
1887zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1888{
1889
1890 ZONE_UNLOCK(a);
1891 ZONE_UNLOCK(b);
1892}
1893
1894int
1895uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1896{
1897 uma_klink_t klink;
1898 uma_klink_t kl;
1899 int error;
1900
1901 error = 0;
1902 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1903
1904 zone_lock_pair(zone, master);
1905 /*
1906 * zone must use vtoslab() to resolve objects and must already be
1907 * a secondary.
1908 */
1909 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1910 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1911 error = EINVAL;
1912 goto out;
1913 }
1914 /*
1915 * The new master must also use vtoslab().
1916 */
1917 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
1918 error = EINVAL;
1919 goto out;
1920 }
1921 /*
1922 * Both must either be refcnt, or not be refcnt.
1923 */
1924 if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
1925 (master->uz_flags & UMA_ZONE_REFCNT)) {
1926 error = EINVAL;
1927 goto out;
1928 }
1929 /*
1930 * The underlying object must be the same size. rsize
1931 * may be different.
1932 */
1933 if (master->uz_size != zone->uz_size) {
1934 error = E2BIG;
1935 goto out;
1936 }
1937 /*
1938 * Put it at the end of the list.
1939 */
1940 klink->kl_keg = zone_first_keg(master);
1941 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
1942 if (LIST_NEXT(kl, kl_link) == NULL) {
1943 LIST_INSERT_AFTER(kl, klink, kl_link);
1944 break;
1945 }
1946 }
1947 klink = NULL;
1948 zone->uz_flags |= UMA_ZFLAG_MULTI;
1949 zone->uz_slab = zone_fetch_slab_multi;
1950
1951out:
1952 zone_unlock_pair(zone, master);
1953 if (klink != NULL)
1954 free(klink, M_TEMP);
1955
1956 return (error);
1957}
1958
1959
1960/* See uma.h */
1961void
1962uma_zdestroy(uma_zone_t zone)
1963{
1964
1965 zone_free_item(zones, zone, NULL, SKIP_NONE, ZFREE_STATFREE);
1966}
1967
1968/* See uma.h */
1969void *
1970uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1971{
1972 void *item;
1973 uma_cache_t cache;
1974 uma_bucket_t bucket;
1975 int cpu;
1976
1977 /* This is the fast path allocation */
1978#ifdef UMA_DEBUG_ALLOC_1
1979 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1980#endif
1981 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1982 zone->uz_name, flags);
1983
1984 if (flags & M_WAITOK) {
1985 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1986 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
1987 }
1988#ifdef DEBUG_MEMGUARD
1989 if (memguard_cmp_zone(zone)) {
1990 item = memguard_alloc(zone->uz_size, flags);
1991 if (item != NULL) {
1992 /*
1993 * Avoid conflict with the use-after-free
1994 * protecting infrastructure from INVARIANTS.
1995 */
1996 if (zone->uz_init != NULL &&
1997 zone->uz_init != mtrash_init &&
1998 zone->uz_init(item, zone->uz_size, flags) != 0)
1999 return (NULL);
2000 if (zone->uz_ctor != NULL &&
2001 zone->uz_ctor != mtrash_ctor &&
2002 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2003 zone->uz_fini(item, zone->uz_size);
2004 return (NULL);
2005 }
2006 return (item);
2007 }
2008 /* This is unfortunate but should not be fatal. */
2009 }
2010#endif
2011 /*
2012 * If possible, allocate from the per-CPU cache. There are two
2013 * requirements for safe access to the per-CPU cache: (1) the thread
2014 * accessing the cache must not be preempted or yield during access,
2015 * and (2) the thread must not migrate CPUs without switching which
2016 * cache it accesses. We rely on a critical section to prevent
2017 * preemption and migration. We release the critical section in
2018 * order to acquire the zone mutex if we are unable to allocate from
2019 * the current cache; when we re-acquire the critical section, we
2020 * must detect and handle migration if it has occurred.
2021 */
2022zalloc_restart:
2023 critical_enter();
2024 cpu = curcpu;
2025 cache = &zone->uz_cpu[cpu];
2026
2027zalloc_start:
2028 bucket = cache->uc_allocbucket;
2029
2030 if (bucket) {
2031 if (bucket->ub_cnt > 0) {
2032 bucket->ub_cnt--;
2033 item = bucket->ub_bucket[bucket->ub_cnt];
2034#ifdef INVARIANTS
2035 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2036#endif
2037 KASSERT(item != NULL,
2038 ("uma_zalloc: Bucket pointer mangled."));
2039 cache->uc_allocs++;
2040 critical_exit();
2041#ifdef INVARIANTS
2042 ZONE_LOCK(zone);
2043 uma_dbg_alloc(zone, NULL, item);
2044 ZONE_UNLOCK(zone);
2045#endif
2046 if (zone->uz_ctor != NULL) {
2047 if (zone->uz_ctor(item, zone->uz_size,
2048 udata, flags) != 0) {
2049 zone_free_item(zone, item, udata,
2050 SKIP_DTOR, ZFREE_STATFAIL |
2051 ZFREE_STATFREE);
2052 return (NULL);
2053 }
2054 }
2055 if (flags & M_ZERO)
2056 bzero(item, zone->uz_size);
2057 return (item);
2058 } else if (cache->uc_freebucket) {
2059 /*
2060 * We have run out of items in our allocbucket.
2061 * See if we can switch with our free bucket.
2062 */
2063 if (cache->uc_freebucket->ub_cnt > 0) {
2064#ifdef UMA_DEBUG_ALLOC
2065 printf("uma_zalloc: Swapping empty with"
2066 " alloc.\n");
2067#endif
2068 bucket = cache->uc_freebucket;
2069 cache->uc_freebucket = cache->uc_allocbucket;
2070 cache->uc_allocbucket = bucket;
2071
2072 goto zalloc_start;
2073 }
2074 }
2075 }
2076 /*
2077 * Attempt to retrieve the item from the per-CPU cache has failed, so
2078 * we must go back to the zone. This requires the zone lock, so we
2079 * must drop the critical section, then re-acquire it when we go back
2080 * to the cache. Since the critical section is released, we may be
2081 * preempted or migrate. As such, make sure not to maintain any
2082 * thread-local state specific to the cache from prior to releasing
2083 * the critical section.
2084 */
2085 critical_exit();
2086 ZONE_LOCK(zone);
2087 critical_enter();
2088 cpu = curcpu;
2089 cache = &zone->uz_cpu[cpu];
2090 bucket = cache->uc_allocbucket;
2091 if (bucket != NULL) {
2092 if (bucket->ub_cnt > 0) {
2093 ZONE_UNLOCK(zone);
2094 goto zalloc_start;
2095 }
2096 bucket = cache->uc_freebucket;
2097 if (bucket != NULL && bucket->ub_cnt > 0) {
2098 ZONE_UNLOCK(zone);
2099 goto zalloc_start;
2100 }
2101 }
2102
2103 /* Since we have locked the zone we may as well send back our stats */
2104 zone->uz_allocs += cache->uc_allocs;
2105 cache->uc_allocs = 0;
2106 zone->uz_frees += cache->uc_frees;
2107 cache->uc_frees = 0;
2108
2109 /* Our old one is now a free bucket */
2110 if (cache->uc_allocbucket) {
2111 KASSERT(cache->uc_allocbucket->ub_cnt == 0,
2112 ("uma_zalloc_arg: Freeing a non free bucket."));
2113 LIST_INSERT_HEAD(&zone->uz_free_bucket,
2114 cache->uc_allocbucket, ub_link);
2115 cache->uc_allocbucket = NULL;
2116 }
2117
2118 /* Check the free list for a new alloc bucket */
2119 if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
2120 KASSERT(bucket->ub_cnt != 0,
2121 ("uma_zalloc_arg: Returning an empty bucket."));
2122
2123 LIST_REMOVE(bucket, ub_link);
2124 cache->uc_allocbucket = bucket;
2125 ZONE_UNLOCK(zone);
2126 goto zalloc_start;
2127 }
2128 /* We are no longer associated with this CPU. */
2129 critical_exit();
2130
2131 /* Bump up our uz_count so we get here less */
2132 if (zone->uz_count < BUCKET_MAX)
2133 zone->uz_count++;
2134
2135 /*
2136 * Now lets just fill a bucket and put it on the free list. If that
2137 * works we'll restart the allocation from the begining.
2138 */
2139 if (zone_alloc_bucket(zone, flags)) {
2140 ZONE_UNLOCK(zone);
2141 goto zalloc_restart;
2142 }
2143 ZONE_UNLOCK(zone);
2144 /*
2145 * We may not be able to get a bucket so return an actual item.
2146 */
2147#ifdef UMA_DEBUG
2148 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2149#endif
2150
2151 item = zone_alloc_item(zone, udata, flags);
2152 return (item);
2153}
2154
2155static uma_slab_t
2156keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2157{
2158 uma_slab_t slab;
2159
2160 mtx_assert(&keg->uk_lock, MA_OWNED);
2161 slab = NULL;
2162
2163 for (;;) {
2164 /*
2165 * Find a slab with some space. Prefer slabs that are partially
2166 * used over those that are totally full. This helps to reduce
2167 * fragmentation.
2168 */
2169 if (keg->uk_free != 0) {
2170 if (!LIST_EMPTY(&keg->uk_part_slab)) {
2171 slab = LIST_FIRST(&keg->uk_part_slab);
2172 } else {
2173 slab = LIST_FIRST(&keg->uk_free_slab);
2174 LIST_REMOVE(slab, us_link);
2175 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2176 us_link);
2177 }
2178 MPASS(slab->us_keg == keg);
2179 return (slab);
2180 }
2181
2182 /*
2183 * M_NOVM means don't ask at all!
2184 */
2185 if (flags & M_NOVM)
2186 break;
2187
2188 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2189 keg->uk_flags |= UMA_ZFLAG_FULL;
2190 /*
2191 * If this is not a multi-zone, set the FULL bit.
2192 * Otherwise slab_multi() takes care of it.
2193 */
2194 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0)
2195 zone->uz_flags |= UMA_ZFLAG_FULL;
2196 if (flags & M_NOWAIT)
2197 break;
2198 zone->uz_sleeps++;
2199 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2200 continue;
2201 }
2202 keg->uk_recurse++;
2203 slab = keg_alloc_slab(keg, zone, flags);
2204 keg->uk_recurse--;
2205 /*
2206 * If we got a slab here it's safe to mark it partially used
2207 * and return. We assume that the caller is going to remove
2208 * at least one item.
2209 */
2210 if (slab) {
2211 MPASS(slab->us_keg == keg);
2212 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2213 return (slab);
2214 }
2215 /*
2216 * We might not have been able to get a slab but another cpu
2217 * could have while we were unlocked. Check again before we
2218 * fail.
2219 */
2220 flags |= M_NOVM;
2221 }
2222 return (slab);
2223}
2224
2225static inline void
2226zone_relock(uma_zone_t zone, uma_keg_t keg)
2227{
2228 if (zone->uz_lock != &keg->uk_lock) {
2229 KEG_UNLOCK(keg);
2230 ZONE_LOCK(zone);
2231 }
2232}
2233
2234static inline void
2235keg_relock(uma_keg_t keg, uma_zone_t zone)
2236{
2237 if (zone->uz_lock != &keg->uk_lock) {
2238 ZONE_UNLOCK(zone);
2239 KEG_LOCK(keg);
2240 }
2241}
2242
2243static uma_slab_t
2244zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2245{
2246 uma_slab_t slab;
2247
2248 if (keg == NULL)
2249 keg = zone_first_keg(zone);
2250 /*
2251 * This is to prevent us from recursively trying to allocate
2252 * buckets. The problem is that if an allocation forces us to
2253 * grab a new bucket we will call page_alloc, which will go off
2254 * and cause the vm to allocate vm_map_entries. If we need new
2255 * buckets there too we will recurse in kmem_alloc and bad
2256 * things happen. So instead we return a NULL bucket, and make
2257 * the code that allocates buckets smart enough to deal with it
2258 */
2259 if (keg->uk_flags & UMA_ZFLAG_BUCKET && keg->uk_recurse != 0)
2260 return (NULL);
2261
2262 for (;;) {
2263 slab = keg_fetch_slab(keg, zone, flags);
2264 if (slab)
2265 return (slab);
2266 if (flags & (M_NOWAIT | M_NOVM))
2267 break;
2268 }
2269 return (NULL);
2270}
2271
2272/*
2273 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2274 * with the keg locked. Caller must call zone_relock() afterwards if the
2275 * zone lock is required. On NULL the zone lock is held.
2276 *
2277 * The last pointer is used to seed the search. It is not required.
2278 */
2279static uma_slab_t
2280zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2281{
2282 uma_klink_t klink;
2283 uma_slab_t slab;
2284 uma_keg_t keg;
2285 int flags;
2286 int empty;
2287 int full;
2288
2289 /*
2290 * Don't wait on the first pass. This will skip limit tests
2291 * as well. We don't want to block if we can find a provider
2292 * without blocking.
2293 */
2294 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2295 /*
2296 * Use the last slab allocated as a hint for where to start
2297 * the search.
2298 */
2299 if (last) {
2300 slab = keg_fetch_slab(last, zone, flags);
2301 if (slab)
2302 return (slab);
2303 zone_relock(zone, last);
2304 last = NULL;
2305 }
2306 /*
2307 * Loop until we have a slab incase of transient failures
2308 * while M_WAITOK is specified. I'm not sure this is 100%
2309 * required but we've done it for so long now.
2310 */
2311 for (;;) {
2312 empty = 0;
2313 full = 0;
2314 /*
2315 * Search the available kegs for slabs. Be careful to hold the
2316 * correct lock while calling into the keg layer.
2317 */
2318 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2319 keg = klink->kl_keg;
2320 keg_relock(keg, zone);
2321 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2322 slab = keg_fetch_slab(keg, zone, flags);
2323 if (slab)
2324 return (slab);
2325 }
2326 if (keg->uk_flags & UMA_ZFLAG_FULL)
2327 full++;
2328 else
2329 empty++;
2330 zone_relock(zone, keg);
2331 }
2332 if (rflags & (M_NOWAIT | M_NOVM))
2333 break;
2334 flags = rflags;
2335 /*
2336 * All kegs are full. XXX We can't atomically check all kegs
2337 * and sleep so just sleep for a short period and retry.
2338 */
2339 if (full && !empty) {
2340 zone->uz_flags |= UMA_ZFLAG_FULL;
2341 zone->uz_sleeps++;
2342 msleep(zone, zone->uz_lock, PVM, "zonelimit", hz/100);
2343 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2344 continue;
2345 }
2346 }
2347 return (NULL);
2348}
2349
2350static void *
2351slab_alloc_item(uma_zone_t zone, uma_slab_t slab)
2352{
2353 uma_keg_t keg;
2354 uma_slabrefcnt_t slabref;
2355 void *item;
2356 u_int8_t freei;
2357
2358 keg = slab->us_keg;
2359 mtx_assert(&keg->uk_lock, MA_OWNED);
2360
2361 freei = slab->us_firstfree;
2362 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2363 slabref = (uma_slabrefcnt_t)slab;
2364 slab->us_firstfree = slabref->us_freelist[freei].us_item;
2365 } else {
2366 slab->us_firstfree = slab->us_freelist[freei].us_item;
2367 }
2368 item = slab->us_data + (keg->uk_rsize * freei);
2369
2370 slab->us_freecount--;
2371 keg->uk_free--;
2372#ifdef INVARIANTS
2373 uma_dbg_alloc(zone, slab, item);
2374#endif
2375 /* Move this slab to the full list */
2376 if (slab->us_freecount == 0) {
2377 LIST_REMOVE(slab, us_link);
2378 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2379 }
2380
2381 return (item);
2382}
2383
2384static int
2385zone_alloc_bucket(uma_zone_t zone, int flags)
2386{
2387 uma_bucket_t bucket;
2388 uma_slab_t slab;
2389 uma_keg_t keg;
2390 int16_t saved;
2391 int max, origflags = flags;
2392
2393 /*
2394 * Try this zone's free list first so we don't allocate extra buckets.
2395 */
2396 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2397 KASSERT(bucket->ub_cnt == 0,
2398 ("zone_alloc_bucket: Bucket on free list is not empty."));
2399 LIST_REMOVE(bucket, ub_link);
2400 } else {
2401 int bflags;
2402
2403 bflags = (flags & ~M_ZERO);
2404 if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2405 bflags |= M_NOVM;
2406
2407 ZONE_UNLOCK(zone);
2408 bucket = bucket_alloc(zone->uz_count, bflags);
2409 ZONE_LOCK(zone);
2410 }
2411
2412 if (bucket == NULL) {
2413 return (0);
2414 }
2415
2416#ifdef SMP
2417 /*
2418 * This code is here to limit the number of simultaneous bucket fills
2419 * for any given zone to the number of per cpu caches in this zone. This
2420 * is done so that we don't allocate more memory than we really need.
2421 */
2422 if (zone->uz_fills >= mp_ncpus)
2423 goto done;
2424
2425#endif
2426 zone->uz_fills++;
2427
2428 max = MIN(bucket->ub_entries, zone->uz_count);
2429 /* Try to keep the buckets totally full */
2430 saved = bucket->ub_cnt;
2431 slab = NULL;
2432 keg = NULL;
2433 while (bucket->ub_cnt < max &&
2434 (slab = zone->uz_slab(zone, keg, flags)) != NULL) {
2435 keg = slab->us_keg;
2436 while (slab->us_freecount && bucket->ub_cnt < max) {
2437 bucket->ub_bucket[bucket->ub_cnt++] =
2438 slab_alloc_item(zone, slab);
2439 }
2440
2441 /* Don't block on the next fill */
2442 flags |= M_NOWAIT;
2443 }
2444 if (slab)
2445 zone_relock(zone, keg);
2446
2447 /*
2448 * We unlock here because we need to call the zone's init.
2449 * It should be safe to unlock because the slab dealt with
2450 * above is already on the appropriate list within the keg
2451 * and the bucket we filled is not yet on any list, so we
2452 * own it.
2453 */
2454 if (zone->uz_init != NULL) {
2455 int i;
2456
2457 ZONE_UNLOCK(zone);
2458 for (i = saved; i < bucket->ub_cnt; i++)
2459 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2460 origflags) != 0)
2461 break;
2462 /*
2463 * If we couldn't initialize the whole bucket, put the
2464 * rest back onto the freelist.
2465 */
2466 if (i != bucket->ub_cnt) {
2467 int j;
2468
2469 for (j = i; j < bucket->ub_cnt; j++) {
2470 zone_free_item(zone, bucket->ub_bucket[j],
2471 NULL, SKIP_FINI, 0);
2472#ifdef INVARIANTS
2473 bucket->ub_bucket[j] = NULL;
2474#endif
2475 }
2476 bucket->ub_cnt = i;
2477 }
2478 ZONE_LOCK(zone);
2479 }
2480
2481 zone->uz_fills--;
2482 if (bucket->ub_cnt != 0) {
2483 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2484 bucket, ub_link);
2485 return (1);
2486 }
2487#ifdef SMP
2488done:
2489#endif
2490 bucket_free(bucket);
2491
2492 return (0);
2493}
2494/*
2495 * Allocates an item for an internal zone
2496 *
2497 * Arguments
2498 * zone The zone to alloc for.
2499 * udata The data to be passed to the constructor.
2500 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2501 *
2502 * Returns
2503 * NULL if there is no memory and M_NOWAIT is set
2504 * An item if successful
2505 */
2506
2507static void *
2508zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2509{
2510 uma_slab_t slab;
2511 void *item;
2512
2513 item = NULL;
2514
2515#ifdef UMA_DEBUG_ALLOC
2516 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2517#endif
2518 ZONE_LOCK(zone);
2519
2520 slab = zone->uz_slab(zone, NULL, flags);
2521 if (slab == NULL) {
2522 zone->uz_fails++;
2523 ZONE_UNLOCK(zone);
2524 return (NULL);
2525 }
2526
2527 item = slab_alloc_item(zone, slab);
2528
2529 zone_relock(zone, slab->us_keg);
2530 zone->uz_allocs++;
2531 ZONE_UNLOCK(zone);
2532
2533 /*
2534 * We have to call both the zone's init (not the keg's init)
2535 * and the zone's ctor. This is because the item is going from
2536 * a keg slab directly to the user, and the user is expecting it
2537 * to be both zone-init'd as well as zone-ctor'd.
2538 */
2539 if (zone->uz_init != NULL) {
2540 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2541 zone_free_item(zone, item, udata, SKIP_FINI,
2542 ZFREE_STATFAIL | ZFREE_STATFREE);
2543 return (NULL);
2544 }
2545 }
2546 if (zone->uz_ctor != NULL) {
2547 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2548 zone_free_item(zone, item, udata, SKIP_DTOR,
2549 ZFREE_STATFAIL | ZFREE_STATFREE);
2550 return (NULL);
2551 }
2552 }
2553 if (flags & M_ZERO)
2554 bzero(item, zone->uz_size);
2555
2556 return (item);
2557}
2558
2559/* See uma.h */
2560void
2561uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2562{
2563 uma_cache_t cache;
2564 uma_bucket_t bucket;
2565 int bflags;
2566 int cpu;
2567
2568#ifdef UMA_DEBUG_ALLOC_1
2569 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2570#endif
2571 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2572 zone->uz_name);
2573
2574 /* uma_zfree(..., NULL) does nothing, to match free(9). */
2575 if (item == NULL)
2576 return;
2577#ifdef DEBUG_MEMGUARD
2578 if (is_memguard_addr(item)) {
2579 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2580 zone->uz_dtor(item, zone->uz_size, udata);
2581 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2582 zone->uz_fini(item, zone->uz_size);
2583 memguard_free(item);
2584 return;
2585 }
2586#endif
2587 if (zone->uz_dtor)
2588 zone->uz_dtor(item, zone->uz_size, udata);
2589
2590#ifdef INVARIANTS
2591 ZONE_LOCK(zone);
2592 if (zone->uz_flags & UMA_ZONE_MALLOC)
2593 uma_dbg_free(zone, udata, item);
2594 else
2595 uma_dbg_free(zone, NULL, item);
2596 ZONE_UNLOCK(zone);
2597#endif
2598 /*
2599 * The race here is acceptable. If we miss it we'll just have to wait
2600 * a little longer for the limits to be reset.
2601 */
2602 if (zone->uz_flags & UMA_ZFLAG_FULL)
2603 goto zfree_internal;
2604
2605 /*
2606 * If possible, free to the per-CPU cache. There are two
2607 * requirements for safe access to the per-CPU cache: (1) the thread
2608 * accessing the cache must not be preempted or yield during access,
2609 * and (2) the thread must not migrate CPUs without switching which
2610 * cache it accesses. We rely on a critical section to prevent
2611 * preemption and migration. We release the critical section in
2612 * order to acquire the zone mutex if we are unable to free to the
2613 * current cache; when we re-acquire the critical section, we must
2614 * detect and handle migration if it has occurred.
2615 */
2616zfree_restart:
2617 critical_enter();
2618 cpu = curcpu;
2619 cache = &zone->uz_cpu[cpu];
2620
2621zfree_start:
2622 bucket = cache->uc_freebucket;
2623
2624 if (bucket) {
2625 /*
2626 * Do we have room in our bucket? It is OK for this uz count
2627 * check to be slightly out of sync.
2628 */
2629
2630 if (bucket->ub_cnt < bucket->ub_entries) {
2631 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2632 ("uma_zfree: Freeing to non free bucket index."));
2633 bucket->ub_bucket[bucket->ub_cnt] = item;
2634 bucket->ub_cnt++;
2635 cache->uc_frees++;
2636 critical_exit();
2637 return;
2638 } else if (cache->uc_allocbucket) {
2639#ifdef UMA_DEBUG_ALLOC
2640 printf("uma_zfree: Swapping buckets.\n");
2641#endif
2642 /*
2643 * We have run out of space in our freebucket.
2644 * See if we can switch with our alloc bucket.
2645 */
2646 if (cache->uc_allocbucket->ub_cnt <
2647 cache->uc_freebucket->ub_cnt) {
2648 bucket = cache->uc_freebucket;
2649 cache->uc_freebucket = cache->uc_allocbucket;
2650 cache->uc_allocbucket = bucket;
2651 goto zfree_start;
2652 }
2653 }
2654 }
2655 /*
2656 * We can get here for two reasons:
2657 *
2658 * 1) The buckets are NULL
2659 * 2) The alloc and free buckets are both somewhat full.
2660 *
2661 * We must go back the zone, which requires acquiring the zone lock,
2662 * which in turn means we must release and re-acquire the critical
2663 * section. Since the critical section is released, we may be
2664 * preempted or migrate. As such, make sure not to maintain any
2665 * thread-local state specific to the cache from prior to releasing
2666 * the critical section.
2667 */
2668 critical_exit();
2669 ZONE_LOCK(zone);
2670 critical_enter();
2671 cpu = curcpu;
2672 cache = &zone->uz_cpu[cpu];
2673 if (cache->uc_freebucket != NULL) {
2674 if (cache->uc_freebucket->ub_cnt <
2675 cache->uc_freebucket->ub_entries) {
2676 ZONE_UNLOCK(zone);
2677 goto zfree_start;
2678 }
2679 if (cache->uc_allocbucket != NULL &&
2680 (cache->uc_allocbucket->ub_cnt <
2681 cache->uc_freebucket->ub_cnt)) {
2682 ZONE_UNLOCK(zone);
2683 goto zfree_start;
2684 }
2685 }
2686
2687 /* Since we have locked the zone we may as well send back our stats */
2688 zone->uz_allocs += cache->uc_allocs;
2689 cache->uc_allocs = 0;
2690 zone->uz_frees += cache->uc_frees;
2691 cache->uc_frees = 0;
2692
2693 bucket = cache->uc_freebucket;
2694 cache->uc_freebucket = NULL;
2695
2696 /* Can we throw this on the zone full list? */
2697 if (bucket != NULL) {
2698#ifdef UMA_DEBUG_ALLOC
2699 printf("uma_zfree: Putting old bucket on the free list.\n");
2700#endif
2701 /* ub_cnt is pointing to the last free item */
2702 KASSERT(bucket->ub_cnt != 0,
2703 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2704 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2705 bucket, ub_link);
2706 }
2707 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2708 LIST_REMOVE(bucket, ub_link);
2709 ZONE_UNLOCK(zone);
2710 cache->uc_freebucket = bucket;
2711 goto zfree_start;
2712 }
2713 /* We are no longer associated with this CPU. */
2714 critical_exit();
2715
2716 /* And the zone.. */
2717 ZONE_UNLOCK(zone);
2718
2719#ifdef UMA_DEBUG_ALLOC
2720 printf("uma_zfree: Allocating new free bucket.\n");
2721#endif
2722 bflags = M_NOWAIT;
2723
2724 if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2725 bflags |= M_NOVM;
2726 bucket = bucket_alloc(zone->uz_count, bflags);
2727 if (bucket) {
2728 ZONE_LOCK(zone);
2729 LIST_INSERT_HEAD(&zone->uz_free_bucket,
2730 bucket, ub_link);
2731 ZONE_UNLOCK(zone);
2732 goto zfree_restart;
2733 }
2734
2735 /*
2736 * If nothing else caught this, we'll just do an internal free.
2737 */
2738zfree_internal:
2739 zone_free_item(zone, item, udata, SKIP_DTOR, ZFREE_STATFREE);
2740
2741 return;
2742}
2743
2744/*
2745 * Frees an item to an INTERNAL zone or allocates a free bucket
2746 *
2747 * Arguments:
2748 * zone The zone to free to
2749 * item The item we're freeing
2750 * udata User supplied data for the dtor
2751 * skip Skip dtors and finis
2752 */
2753static void
2754zone_free_item(uma_zone_t zone, void *item, void *udata,
2755 enum zfreeskip skip, int flags)
2756{
2757 uma_slab_t slab;
2758 uma_slabrefcnt_t slabref;
2759 uma_keg_t keg;
2760 u_int8_t *mem;
2761 u_int8_t freei;
2762 int clearfull;
2763
2764 if (skip < SKIP_DTOR && zone->uz_dtor)
2765 zone->uz_dtor(item, zone->uz_size, udata);
2766
2767 if (skip < SKIP_FINI && zone->uz_fini)
2768 zone->uz_fini(item, zone->uz_size);
2769
2770 ZONE_LOCK(zone);
2771
2772 if (flags & ZFREE_STATFAIL)
2773 zone->uz_fails++;
2774 if (flags & ZFREE_STATFREE)
2775 zone->uz_frees++;
2776
2777 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2778 mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK));
2779 keg = zone_first_keg(zone); /* Must only be one. */
2780 if (zone->uz_flags & UMA_ZONE_HASH) {
2781 slab = hash_sfind(&keg->uk_hash, mem);
2782 } else {
2783 mem += keg->uk_pgoff;
2784 slab = (uma_slab_t)mem;
2785 }
2786 } else {
2787 /* This prevents redundant lookups via free(). */
2788 if ((zone->uz_flags & UMA_ZONE_MALLOC) && udata != NULL)
2789 slab = (uma_slab_t)udata;
2790 else
2791 slab = vtoslab((vm_offset_t)item);
2792 keg = slab->us_keg;
2793 keg_relock(keg, zone);
2794 }
2795 MPASS(keg == slab->us_keg);
2796
2797 /* Do we need to remove from any lists? */
2798 if (slab->us_freecount+1 == keg->uk_ipers) {
2799 LIST_REMOVE(slab, us_link);
2800 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2801 } else if (slab->us_freecount == 0) {
2802 LIST_REMOVE(slab, us_link);
2803 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2804 }
2805
2806 /* Slab management stuff */
2807 freei = ((unsigned long)item - (unsigned long)slab->us_data)
2808 / keg->uk_rsize;
2809
2810#ifdef INVARIANTS
2811 if (!skip)
2812 uma_dbg_free(zone, slab, item);
2813#endif
2814
2815 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2816 slabref = (uma_slabrefcnt_t)slab;
2817 slabref->us_freelist[freei].us_item = slab->us_firstfree;
2818 } else {
2819 slab->us_freelist[freei].us_item = slab->us_firstfree;
2820 }
2821 slab->us_firstfree = freei;
2822 slab->us_freecount++;
2823
2824 /* Zone statistics */
2825 keg->uk_free++;
2826
2827 clearfull = 0;
2828 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2829 if (keg->uk_pages < keg->uk_maxpages) {
2830 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2831 clearfull = 1;
2832 }
2833
2834 /*
2835 * We can handle one more allocation. Since we're clearing ZFLAG_FULL,
2836 * wake up all procs blocked on pages. This should be uncommon, so
2837 * keeping this simple for now (rather than adding count of blocked
2838 * threads etc).
2839 */
2840 wakeup(keg);
2841 }
2842 if (clearfull) {
2843 zone_relock(zone, keg);
2844 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2845 wakeup(zone);
2846 ZONE_UNLOCK(zone);
2847 } else
2848 KEG_UNLOCK(keg);
2849}
2850
2851/* See uma.h */
2852int
2853uma_zone_set_max(uma_zone_t zone, int nitems)
2854{
2855 uma_keg_t keg;
2856
2857 ZONE_LOCK(zone);
2858 keg = zone_first_keg(zone);
2859 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2860 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2861 keg->uk_maxpages += keg->uk_ppera;
2862 nitems = keg->uk_maxpages * keg->uk_ipers;
2863 ZONE_UNLOCK(zone);
2864
2865 return (nitems);
2866}
2867
2868/* See uma.h */
2869int
2870uma_zone_get_max(uma_zone_t zone)
2871{
2872 int nitems;
2873 uma_keg_t keg;
2874
2875 ZONE_LOCK(zone);
2876 keg = zone_first_keg(zone);
2877 nitems = keg->uk_maxpages * keg->uk_ipers;
2878 ZONE_UNLOCK(zone);
2879
2880 return (nitems);
2881}
2882
2883/* See uma.h */
2884int
2885uma_zone_get_cur(uma_zone_t zone)
2886{
2887 int64_t nitems;
2888 u_int i;
2889
2890 ZONE_LOCK(zone);
2891 nitems = zone->uz_allocs - zone->uz_frees;
2892 CPU_FOREACH(i) {
2893 /*
2894 * See the comment in sysctl_vm_zone_stats() regarding the
2895 * safety of accessing the per-cpu caches. With the zone lock
2896 * held, it is safe, but can potentially result in stale data.
2897 */
2898 nitems += zone->uz_cpu[i].uc_allocs -
2899 zone->uz_cpu[i].uc_frees;
2900 }
2901 ZONE_UNLOCK(zone);
2902
2903 return (nitems < 0 ? 0 : nitems);
2904}
2905
2906/* See uma.h */
2907void
2908uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2909{
2910 uma_keg_t keg;
2911
2912 ZONE_LOCK(zone);
2913 keg = zone_first_keg(zone);
2914 KASSERT(keg->uk_pages == 0,
2915 ("uma_zone_set_init on non-empty keg"));
2916 keg->uk_init = uminit;
2917 ZONE_UNLOCK(zone);
2918}
2919
2920/* See uma.h */
2921void
2922uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2923{
2924 uma_keg_t keg;
2925
2926 ZONE_LOCK(zone);
2927 keg = zone_first_keg(zone);
2928 KASSERT(keg->uk_pages == 0,
2929 ("uma_zone_set_fini on non-empty keg"));
2930 keg->uk_fini = fini;
2931 ZONE_UNLOCK(zone);
2932}
2933
2934/* See uma.h */
2935void
2936uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2937{
2938 ZONE_LOCK(zone);
2939 KASSERT(zone_first_keg(zone)->uk_pages == 0,
2940 ("uma_zone_set_zinit on non-empty keg"));
2941 zone->uz_init = zinit;
2942 ZONE_UNLOCK(zone);
2943}
2944
2945/* See uma.h */
2946void
2947uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2948{
2949 ZONE_LOCK(zone);
2950 KASSERT(zone_first_keg(zone)->uk_pages == 0,
2951 ("uma_zone_set_zfini on non-empty keg"));
2952 zone->uz_fini = zfini;
2953 ZONE_UNLOCK(zone);
2954}
2955
2956/* See uma.h */
2957/* XXX uk_freef is not actually used with the zone locked */
2958void
2959uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2960{
2961
2962 ZONE_LOCK(zone);
2963 zone_first_keg(zone)->uk_freef = freef;
2964 ZONE_UNLOCK(zone);
2965}
2966
2967/* See uma.h */
2968/* XXX uk_allocf is not actually used with the zone locked */
2969void
2970uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2971{
2972 uma_keg_t keg;
2973
2974 ZONE_LOCK(zone);
2975 keg = zone_first_keg(zone);
2976 keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2977 keg->uk_allocf = allocf;
2978 ZONE_UNLOCK(zone);
2979}
2980
2981/* See uma.h */
2982int
2983uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count)
2984{
2985 uma_keg_t keg;
2986 vm_offset_t kva;
2987 int pages;
2988
2989 keg = zone_first_keg(zone);
2990 pages = count / keg->uk_ipers;
2991
2992 if (pages * keg->uk_ipers < count)
2993 pages++;
2994
2995 kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2996
2997 if (kva == 0)
2998 return (0);
2999 if (obj == NULL)
3000 obj = vm_object_allocate(OBJT_PHYS, pages);
3001 else {
3002 VM_OBJECT_LOCK_INIT(obj, "uma object");
3003 _vm_object_allocate(OBJT_PHYS, pages, obj);
3004 }
3005 ZONE_LOCK(zone);
3006 keg->uk_kva = kva;
3007 keg->uk_obj = obj;
3008 keg->uk_maxpages = pages;
3009 keg->uk_allocf = obj_alloc;
3010 keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
3011 ZONE_UNLOCK(zone);
3012 return (1);
3013}
3014
3015/* See uma.h */
3016void
3017uma_prealloc(uma_zone_t zone, int items)
3018{
3019 int slabs;
3020 uma_slab_t slab;
3021 uma_keg_t keg;
3022
3023 keg = zone_first_keg(zone);
3024 ZONE_LOCK(zone);
3025 slabs = items / keg->uk_ipers;
3026 if (slabs * keg->uk_ipers < items)
3027 slabs++;
3028 while (slabs > 0) {
3029 slab = keg_alloc_slab(keg, zone, M_WAITOK);
3030 if (slab == NULL)
3031 break;
3032 MPASS(slab->us_keg == keg);
3033 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3034 slabs--;
3035 }
3036 ZONE_UNLOCK(zone);
3037}
3038
3039/* See uma.h */
3040u_int32_t *
3041uma_find_refcnt(uma_zone_t zone, void *item)
3042{
3043 uma_slabrefcnt_t slabref;
3044 uma_keg_t keg;
3045 u_int32_t *refcnt;
3046 int idx;
3047
3048 slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item &
3049 (~UMA_SLAB_MASK));
3050 keg = slabref->us_keg;
3051 KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT,
3052 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3053 idx = ((unsigned long)item - (unsigned long)slabref->us_data)
3054 / keg->uk_rsize;
3055 refcnt = &slabref->us_freelist[idx].us_refcnt;
3056 return refcnt;
3057}
3058
3059/* See uma.h */
3060void
3061uma_reclaim(void)
3062{
3063#ifdef UMA_DEBUG
3064 printf("UMA: vm asked us to release pages!\n");
3065#endif
3066 bucket_enable();
3067 zone_foreach(zone_drain);
3068 /*
3069 * Some slabs may have been freed but this zone will be visited early
3070 * we visit again so that we can free pages that are empty once other
3071 * zones are drained. We have to do the same for buckets.
3072 */
3073 zone_drain(slabzone);
3074 zone_drain(slabrefzone);
3075 bucket_zone_drain();
3076}
3077
3078/* See uma.h */
3079int
3080uma_zone_exhausted(uma_zone_t zone)
3081{
3082 int full;
3083
3084 ZONE_LOCK(zone);
3085 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3086 ZONE_UNLOCK(zone);
3087 return (full);
3088}
3089
3090int
3091uma_zone_exhausted_nolock(uma_zone_t zone)
3092{
3093 return (zone->uz_flags & UMA_ZFLAG_FULL);
3094}
3095
3096void *
3097uma_large_malloc(int size, int wait)
3098{
3099 void *mem;
3100 uma_slab_t slab;
3101 u_int8_t flags;
3102
3103 slab = zone_alloc_item(slabzone, NULL, wait);
3104 if (slab == NULL)
3105 return (NULL);
3106 mem = page_alloc(NULL, size, &flags, wait);
3107 if (mem) {
3108 vsetslab((vm_offset_t)mem, slab);
3109 slab->us_data = mem;
3110 slab->us_flags = flags | UMA_SLAB_MALLOC;
3111 slab->us_size = size;
3112 } else {
3113 zone_free_item(slabzone, slab, NULL, SKIP_NONE,
3114 ZFREE_STATFAIL | ZFREE_STATFREE);
3115 }
3116
3117 return (mem);
3118}
3119
3120void
3121uma_large_free(uma_slab_t slab)
3122{
3123 vsetobj((vm_offset_t)slab->us_data, kmem_object);
3124 page_free(slab->us_data, slab->us_size, slab->us_flags);
3125 zone_free_item(slabzone, slab, NULL, SKIP_NONE, ZFREE_STATFREE);
3126}
3127
3128void
3129uma_print_stats(void)
3130{
3131 zone_foreach(uma_print_zone);
3132}
3133
3134static void
3135slab_print(uma_slab_t slab)
3136{
3137 printf("slab: keg %p, data %p, freecount %d, firstfree %d\n",
3138 slab->us_keg, slab->us_data, slab->us_freecount,
3139 slab->us_firstfree);
3140}
3141
3142static void
3143cache_print(uma_cache_t cache)
3144{
3145 printf("alloc: %p(%d), free: %p(%d)\n",
3146 cache->uc_allocbucket,
3147 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3148 cache->uc_freebucket,
3149 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3150}
3151
3152static void
3153uma_print_keg(uma_keg_t keg)
3154{
3155 uma_slab_t slab;
3156
3157 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3158 "out %d free %d limit %d\n",
3159 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3160 keg->uk_ipers, keg->uk_ppera,
3161 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3162 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3163 printf("Part slabs:\n");
3164 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3165 slab_print(slab);
3166 printf("Free slabs:\n");
3167 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3168 slab_print(slab);
3169 printf("Full slabs:\n");
3170 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3171 slab_print(slab);
3172}
3173
3174void
3175uma_print_zone(uma_zone_t zone)
3176{
3177 uma_cache_t cache;
3178 uma_klink_t kl;
3179 int i;
3180
3181 printf("zone: %s(%p) size %d flags %#x\n",
3182 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3183 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3184 uma_print_keg(kl->kl_keg);
3185 CPU_FOREACH(i) {
3186 cache = &zone->uz_cpu[i];
3187 printf("CPU %d Cache:\n", i);
3188 cache_print(cache);
3189 }
3190}
3191
3192#ifdef DDB
3193/*
3194 * Generate statistics across both the zone and its per-cpu cache's. Return
3195 * desired statistics if the pointer is non-NULL for that statistic.
3196 *
3197 * Note: does not update the zone statistics, as it can't safely clear the
3198 * per-CPU cache statistic.
3199 *
3200 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3201 * safe from off-CPU; we should modify the caches to track this information
3202 * directly so that we don't have to.
3203 */
3204static void
3205uma_zone_sumstat(uma_zone_t z, int *cachefreep, u_int64_t *allocsp,
3206 u_int64_t *freesp, u_int64_t *sleepsp)
3207{
3208 uma_cache_t cache;
3209 u_int64_t allocs, frees, sleeps;
3210 int cachefree, cpu;
3211
3212 allocs = frees = sleeps = 0;
3213 cachefree = 0;
3214 CPU_FOREACH(cpu) {
3215 cache = &z->uz_cpu[cpu];
3216 if (cache->uc_allocbucket != NULL)
3217 cachefree += cache->uc_allocbucket->ub_cnt;
3218 if (cache->uc_freebucket != NULL)
3219 cachefree += cache->uc_freebucket->ub_cnt;
3220 allocs += cache->uc_allocs;
3221 frees += cache->uc_frees;
3222 }
3223 allocs += z->uz_allocs;
3224 frees += z->uz_frees;
3225 sleeps += z->uz_sleeps;
3226 if (cachefreep != NULL)
3227 *cachefreep = cachefree;
3228 if (allocsp != NULL)
3229 *allocsp = allocs;
3230 if (freesp != NULL)
3231 *freesp = frees;
3232 if (sleepsp != NULL)
3233 *sleepsp = sleeps;
3234}
3235#endif /* DDB */
3236
3237static int
3238sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3239{
3240 uma_keg_t kz;
3241 uma_zone_t z;
3242 int count;
3243
3244 count = 0;
3245 mtx_lock(&uma_mtx);
3246 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3247 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3248 count++;
3249 }
3250 mtx_unlock(&uma_mtx);
3251 return (sysctl_handle_int(oidp, &count, 0, req));
3252}
3253
3254static int
3255sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3256{
3257 struct uma_stream_header ush;
3258 struct uma_type_header uth;
3259 struct uma_percpu_stat ups;
3260 uma_bucket_t bucket;
3261 struct sbuf sbuf;
3262 uma_cache_t cache;
3263 uma_klink_t kl;
3264 uma_keg_t kz;
3265 uma_zone_t z;
3266 uma_keg_t k;
3267 int count, error, i;
3268
3269 error = sysctl_wire_old_buffer(req, 0);
3270 if (error != 0)
3271 return (error);
3272 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3273
3274 count = 0;
3275 mtx_lock(&uma_mtx);
3276 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3277 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3278 count++;
3279 }
3280
3281 /*
3282 * Insert stream header.
3283 */
3284 bzero(&ush, sizeof(ush));
3285 ush.ush_version = UMA_STREAM_VERSION;
3286 ush.ush_maxcpus = (mp_maxid + 1);
3287 ush.ush_count = count;
3288 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3289
3290 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3291 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3292 bzero(&uth, sizeof(uth));
3293 ZONE_LOCK(z);
3294 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3295 uth.uth_align = kz->uk_align;
3296 uth.uth_size = kz->uk_size;
3297 uth.uth_rsize = kz->uk_rsize;
3298 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3299 k = kl->kl_keg;
3300 uth.uth_maxpages += k->uk_maxpages;
3301 uth.uth_pages += k->uk_pages;
3302 uth.uth_keg_free += k->uk_free;
3303 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3304 * k->uk_ipers;
3305 }
3306
3307 /*
3308 * A zone is secondary is it is not the first entry
3309 * on the keg's zone list.
3310 */
3311 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3312 (LIST_FIRST(&kz->uk_zones) != z))
3313 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3314
3315 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3316 uth.uth_zone_free += bucket->ub_cnt;
3317 uth.uth_allocs = z->uz_allocs;
3318 uth.uth_frees = z->uz_frees;
3319 uth.uth_fails = z->uz_fails;
3320 uth.uth_sleeps = z->uz_sleeps;
3321 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3322 /*
3323 * While it is not normally safe to access the cache
3324 * bucket pointers while not on the CPU that owns the
3325 * cache, we only allow the pointers to be exchanged
3326 * without the zone lock held, not invalidated, so
3327 * accept the possible race associated with bucket
3328 * exchange during monitoring.
3329 */
3330 for (i = 0; i < (mp_maxid + 1); i++) {
3331 bzero(&ups, sizeof(ups));
3332 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3333 goto skip;
3334 if (CPU_ABSENT(i))
3335 goto skip;
3336 cache = &z->uz_cpu[i];
3337 if (cache->uc_allocbucket != NULL)
3338 ups.ups_cache_free +=
3339 cache->uc_allocbucket->ub_cnt;
3340 if (cache->uc_freebucket != NULL)
3341 ups.ups_cache_free +=
3342 cache->uc_freebucket->ub_cnt;
3343 ups.ups_allocs = cache->uc_allocs;
3344 ups.ups_frees = cache->uc_frees;
3345skip:
3346 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3347 }
3348 ZONE_UNLOCK(z);
3349 }
3350 }
3351 mtx_unlock(&uma_mtx);
3352 error = sbuf_finish(&sbuf);
3353 sbuf_delete(&sbuf);
3354 return (error);
3355}
3356
3357#ifdef DDB
3358DB_SHOW_COMMAND(uma, db_show_uma)
3359{
3360 u_int64_t allocs, frees, sleeps;
3361 uma_bucket_t bucket;
3362 uma_keg_t kz;
3363 uma_zone_t z;
3364 int cachefree;
3365
3366 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3367 "Requests", "Sleeps");
3368 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3369 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3370 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3371 allocs = z->uz_allocs;
3372 frees = z->uz_frees;
3373 sleeps = z->uz_sleeps;
3374 cachefree = 0;
3375 } else
3376 uma_zone_sumstat(z, &cachefree, &allocs,
3377 &frees, &sleeps);
3378 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3379 (LIST_FIRST(&kz->uk_zones) != z)))
3380 cachefree += kz->uk_free;
3381 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3382 cachefree += bucket->ub_cnt;
3383 db_printf("%18s %8ju %8jd %8d %12ju %8ju\n", z->uz_name,
3384 (uintmax_t)kz->uk_size,
3385 (intmax_t)(allocs - frees), cachefree,
3386 (uintmax_t)allocs, sleeps);
3387 if (db_pager_quit)
3388 return;
3389 }
3390 }
3391}
3392#endif
2 * Copyright (c) 2002-2005, 2009 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 242152 2012-10-26 17:51:05Z mdf $");
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#include "opt_vm.h"
63
64#include <sys/param.h>
65#include <sys/systm.h>
66#include <sys/kernel.h>
67#include <sys/types.h>
68#include <sys/queue.h>
69#include <sys/malloc.h>
70#include <sys/ktr.h>
71#include <sys/lock.h>
72#include <sys/sysctl.h>
73#include <sys/mutex.h>
74#include <sys/proc.h>
75#include <sys/sbuf.h>
76#include <sys/smp.h>
77#include <sys/vmmeter.h>
78
79#include <vm/vm.h>
80#include <vm/vm_object.h>
81#include <vm/vm_page.h>
82#include <vm/vm_param.h>
83#include <vm/vm_map.h>
84#include <vm/vm_kern.h>
85#include <vm/vm_extern.h>
86#include <vm/uma.h>
87#include <vm/uma_int.h>
88#include <vm/uma_dbg.h>
89
90#include <ddb/ddb.h>
91
92#ifdef DEBUG_MEMGUARD
93#include <vm/memguard.h>
94#endif
95
96/*
97 * This is the zone and keg from which all zones are spawned. The idea is that
98 * even the zone & keg heads are allocated from the allocator, so we use the
99 * bss section to bootstrap us.
100 */
101static struct uma_keg masterkeg;
102static struct uma_zone masterzone_k;
103static struct uma_zone masterzone_z;
104static uma_zone_t kegs = &masterzone_k;
105static uma_zone_t zones = &masterzone_z;
106
107/* This is the zone from which all of uma_slab_t's are allocated. */
108static uma_zone_t slabzone;
109static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
110
111/*
112 * The initial hash tables come out of this zone so they can be allocated
113 * prior to malloc coming up.
114 */
115static uma_zone_t hashzone;
116
117/* The boot-time adjusted value for cache line alignment. */
118int uma_align_cache = 64 - 1;
119
120static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
121
122/*
123 * Are we allowed to allocate buckets?
124 */
125static int bucketdisable = 1;
126
127/* Linked list of all kegs in the system */
128static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
129
130/* This mutex protects the keg list */
131static struct mtx uma_mtx;
132
133/* Linked list of boot time pages */
134static LIST_HEAD(,uma_slab) uma_boot_pages =
135 LIST_HEAD_INITIALIZER(uma_boot_pages);
136
137/* This mutex protects the boot time pages list */
138static struct mtx uma_boot_pages_mtx;
139
140/* Is the VM done starting up? */
141static int booted = 0;
142#define UMA_STARTUP 1
143#define UMA_STARTUP2 2
144
145/* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
146static u_int uma_max_ipers;
147static u_int uma_max_ipers_ref;
148
149/*
150 * This is the handle used to schedule events that need to happen
151 * outside of the allocation fast path.
152 */
153static struct callout uma_callout;
154#define UMA_TIMEOUT 20 /* Seconds for callout interval. */
155
156/*
157 * This structure is passed as the zone ctor arg so that I don't have to create
158 * a special allocation function just for zones.
159 */
160struct uma_zctor_args {
161 const char *name;
162 size_t size;
163 uma_ctor ctor;
164 uma_dtor dtor;
165 uma_init uminit;
166 uma_fini fini;
167 uma_keg_t keg;
168 int align;
169 u_int32_t flags;
170};
171
172struct uma_kctor_args {
173 uma_zone_t zone;
174 size_t size;
175 uma_init uminit;
176 uma_fini fini;
177 int align;
178 u_int32_t flags;
179};
180
181struct uma_bucket_zone {
182 uma_zone_t ubz_zone;
183 char *ubz_name;
184 int ubz_entries;
185};
186
187#define BUCKET_MAX 128
188
189struct uma_bucket_zone bucket_zones[] = {
190 { NULL, "16 Bucket", 16 },
191 { NULL, "32 Bucket", 32 },
192 { NULL, "64 Bucket", 64 },
193 { NULL, "128 Bucket", 128 },
194 { NULL, NULL, 0}
195};
196
197#define BUCKET_SHIFT 4
198#define BUCKET_ZONES ((BUCKET_MAX >> BUCKET_SHIFT) + 1)
199
200/*
201 * bucket_size[] maps requested bucket sizes to zones that allocate a bucket
202 * of approximately the right size.
203 */
204static uint8_t bucket_size[BUCKET_ZONES];
205
206/*
207 * Flags and enumerations to be passed to internal functions.
208 */
209enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI };
210
211#define ZFREE_STATFAIL 0x00000001 /* Update zone failure statistic. */
212#define ZFREE_STATFREE 0x00000002 /* Update zone free statistic. */
213
214/* Prototypes.. */
215
216static void *obj_alloc(uma_zone_t, int, u_int8_t *, int);
217static void *page_alloc(uma_zone_t, int, u_int8_t *, int);
218static void *startup_alloc(uma_zone_t, int, u_int8_t *, int);
219static void page_free(void *, int, u_int8_t);
220static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
221static void cache_drain(uma_zone_t);
222static void bucket_drain(uma_zone_t, uma_bucket_t);
223static void bucket_cache_drain(uma_zone_t zone);
224static int keg_ctor(void *, int, void *, int);
225static void keg_dtor(void *, int, void *);
226static int zone_ctor(void *, int, void *, int);
227static void zone_dtor(void *, int, void *);
228static int zero_init(void *, int, int);
229static void keg_small_init(uma_keg_t keg);
230static void keg_large_init(uma_keg_t keg);
231static void zone_foreach(void (*zfunc)(uma_zone_t));
232static void zone_timeout(uma_zone_t zone);
233static int hash_alloc(struct uma_hash *);
234static int hash_expand(struct uma_hash *, struct uma_hash *);
235static void hash_free(struct uma_hash *hash);
236static void uma_timeout(void *);
237static void uma_startup3(void);
238static void *zone_alloc_item(uma_zone_t, void *, int);
239static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip,
240 int);
241static void bucket_enable(void);
242static void bucket_init(void);
243static uma_bucket_t bucket_alloc(int, int);
244static void bucket_free(uma_bucket_t);
245static void bucket_zone_drain(void);
246static int zone_alloc_bucket(uma_zone_t zone, int flags);
247static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
248static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
249static void *slab_alloc_item(uma_zone_t zone, uma_slab_t slab);
250static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
251 uma_fini fini, int align, u_int32_t flags);
252static inline void zone_relock(uma_zone_t zone, uma_keg_t keg);
253static inline void keg_relock(uma_keg_t keg, uma_zone_t zone);
254
255void uma_print_zone(uma_zone_t);
256void uma_print_stats(void);
257static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
258static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
259
260SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
261
262SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
263 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
264
265SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
266 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
267
268/*
269 * This routine checks to see whether or not it's safe to enable buckets.
270 */
271
272static void
273bucket_enable(void)
274{
275 bucketdisable = vm_page_count_min();
276}
277
278/*
279 * Initialize bucket_zones, the array of zones of buckets of various sizes.
280 *
281 * For each zone, calculate the memory required for each bucket, consisting
282 * of the header and an array of pointers. Initialize bucket_size[] to point
283 * the range of appropriate bucket sizes at the zone.
284 */
285static void
286bucket_init(void)
287{
288 struct uma_bucket_zone *ubz;
289 int i;
290 int j;
291
292 for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) {
293 int size;
294
295 ubz = &bucket_zones[j];
296 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
297 size += sizeof(void *) * ubz->ubz_entries;
298 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
299 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
300 UMA_ZFLAG_INTERNAL | UMA_ZFLAG_BUCKET);
301 for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT))
302 bucket_size[i >> BUCKET_SHIFT] = j;
303 }
304}
305
306/*
307 * Given a desired number of entries for a bucket, return the zone from which
308 * to allocate the bucket.
309 */
310static struct uma_bucket_zone *
311bucket_zone_lookup(int entries)
312{
313 int idx;
314
315 idx = howmany(entries, 1 << BUCKET_SHIFT);
316 return (&bucket_zones[bucket_size[idx]]);
317}
318
319static uma_bucket_t
320bucket_alloc(int entries, int bflags)
321{
322 struct uma_bucket_zone *ubz;
323 uma_bucket_t bucket;
324
325 /*
326 * This is to stop us from allocating per cpu buckets while we're
327 * running out of vm.boot_pages. Otherwise, we would exhaust the
328 * boot pages. This also prevents us from allocating buckets in
329 * low memory situations.
330 */
331 if (bucketdisable)
332 return (NULL);
333
334 ubz = bucket_zone_lookup(entries);
335 bucket = zone_alloc_item(ubz->ubz_zone, NULL, bflags);
336 if (bucket) {
337#ifdef INVARIANTS
338 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
339#endif
340 bucket->ub_cnt = 0;
341 bucket->ub_entries = ubz->ubz_entries;
342 }
343
344 return (bucket);
345}
346
347static void
348bucket_free(uma_bucket_t bucket)
349{
350 struct uma_bucket_zone *ubz;
351
352 ubz = bucket_zone_lookup(bucket->ub_entries);
353 zone_free_item(ubz->ubz_zone, bucket, NULL, SKIP_NONE,
354 ZFREE_STATFREE);
355}
356
357static void
358bucket_zone_drain(void)
359{
360 struct uma_bucket_zone *ubz;
361
362 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
363 zone_drain(ubz->ubz_zone);
364}
365
366static inline uma_keg_t
367zone_first_keg(uma_zone_t zone)
368{
369
370 return (LIST_FIRST(&zone->uz_kegs)->kl_keg);
371}
372
373static void
374zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
375{
376 uma_klink_t klink;
377
378 LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
379 kegfn(klink->kl_keg);
380}
381
382/*
383 * Routine called by timeout which is used to fire off some time interval
384 * based calculations. (stats, hash size, etc.)
385 *
386 * Arguments:
387 * arg Unused
388 *
389 * Returns:
390 * Nothing
391 */
392static void
393uma_timeout(void *unused)
394{
395 bucket_enable();
396 zone_foreach(zone_timeout);
397
398 /* Reschedule this event */
399 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
400}
401
402/*
403 * Routine to perform timeout driven calculations. This expands the
404 * hashes and does per cpu statistics aggregation.
405 *
406 * Returns nothing.
407 */
408static void
409keg_timeout(uma_keg_t keg)
410{
411
412 KEG_LOCK(keg);
413 /*
414 * Expand the keg hash table.
415 *
416 * This is done if the number of slabs is larger than the hash size.
417 * What I'm trying to do here is completely reduce collisions. This
418 * may be a little aggressive. Should I allow for two collisions max?
419 */
420 if (keg->uk_flags & UMA_ZONE_HASH &&
421 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
422 struct uma_hash newhash;
423 struct uma_hash oldhash;
424 int ret;
425
426 /*
427 * This is so involved because allocating and freeing
428 * while the keg lock is held will lead to deadlock.
429 * I have to do everything in stages and check for
430 * races.
431 */
432 newhash = keg->uk_hash;
433 KEG_UNLOCK(keg);
434 ret = hash_alloc(&newhash);
435 KEG_LOCK(keg);
436 if (ret) {
437 if (hash_expand(&keg->uk_hash, &newhash)) {
438 oldhash = keg->uk_hash;
439 keg->uk_hash = newhash;
440 } else
441 oldhash = newhash;
442
443 KEG_UNLOCK(keg);
444 hash_free(&oldhash);
445 KEG_LOCK(keg);
446 }
447 }
448 KEG_UNLOCK(keg);
449}
450
451static void
452zone_timeout(uma_zone_t zone)
453{
454
455 zone_foreach_keg(zone, &keg_timeout);
456}
457
458/*
459 * Allocate and zero fill the next sized hash table from the appropriate
460 * backing store.
461 *
462 * Arguments:
463 * hash A new hash structure with the old hash size in uh_hashsize
464 *
465 * Returns:
466 * 1 on sucess and 0 on failure.
467 */
468static int
469hash_alloc(struct uma_hash *hash)
470{
471 int oldsize;
472 int alloc;
473
474 oldsize = hash->uh_hashsize;
475
476 /* We're just going to go to a power of two greater */
477 if (oldsize) {
478 hash->uh_hashsize = oldsize * 2;
479 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
480 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
481 M_UMAHASH, M_NOWAIT);
482 } else {
483 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
484 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
485 M_WAITOK);
486 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
487 }
488 if (hash->uh_slab_hash) {
489 bzero(hash->uh_slab_hash, alloc);
490 hash->uh_hashmask = hash->uh_hashsize - 1;
491 return (1);
492 }
493
494 return (0);
495}
496
497/*
498 * Expands the hash table for HASH zones. This is done from zone_timeout
499 * to reduce collisions. This must not be done in the regular allocation
500 * path, otherwise, we can recurse on the vm while allocating pages.
501 *
502 * Arguments:
503 * oldhash The hash you want to expand
504 * newhash The hash structure for the new table
505 *
506 * Returns:
507 * Nothing
508 *
509 * Discussion:
510 */
511static int
512hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
513{
514 uma_slab_t slab;
515 int hval;
516 int i;
517
518 if (!newhash->uh_slab_hash)
519 return (0);
520
521 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
522 return (0);
523
524 /*
525 * I need to investigate hash algorithms for resizing without a
526 * full rehash.
527 */
528
529 for (i = 0; i < oldhash->uh_hashsize; i++)
530 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
531 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
532 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
533 hval = UMA_HASH(newhash, slab->us_data);
534 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
535 slab, us_hlink);
536 }
537
538 return (1);
539}
540
541/*
542 * Free the hash bucket to the appropriate backing store.
543 *
544 * Arguments:
545 * slab_hash The hash bucket we're freeing
546 * hashsize The number of entries in that hash bucket
547 *
548 * Returns:
549 * Nothing
550 */
551static void
552hash_free(struct uma_hash *hash)
553{
554 if (hash->uh_slab_hash == NULL)
555 return;
556 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
557 zone_free_item(hashzone,
558 hash->uh_slab_hash, NULL, SKIP_NONE, ZFREE_STATFREE);
559 else
560 free(hash->uh_slab_hash, M_UMAHASH);
561}
562
563/*
564 * Frees all outstanding items in a bucket
565 *
566 * Arguments:
567 * zone The zone to free to, must be unlocked.
568 * bucket The free/alloc bucket with items, cpu queue must be locked.
569 *
570 * Returns:
571 * Nothing
572 */
573
574static void
575bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
576{
577 void *item;
578
579 if (bucket == NULL)
580 return;
581
582 while (bucket->ub_cnt > 0) {
583 bucket->ub_cnt--;
584 item = bucket->ub_bucket[bucket->ub_cnt];
585#ifdef INVARIANTS
586 bucket->ub_bucket[bucket->ub_cnt] = NULL;
587 KASSERT(item != NULL,
588 ("bucket_drain: botched ptr, item is NULL"));
589#endif
590 zone_free_item(zone, item, NULL, SKIP_DTOR, 0);
591 }
592}
593
594/*
595 * Drains the per cpu caches for a zone.
596 *
597 * NOTE: This may only be called while the zone is being turn down, and not
598 * during normal operation. This is necessary in order that we do not have
599 * to migrate CPUs to drain the per-CPU caches.
600 *
601 * Arguments:
602 * zone The zone to drain, must be unlocked.
603 *
604 * Returns:
605 * Nothing
606 */
607static void
608cache_drain(uma_zone_t zone)
609{
610 uma_cache_t cache;
611 int cpu;
612
613 /*
614 * XXX: It is safe to not lock the per-CPU caches, because we're
615 * tearing down the zone anyway. I.e., there will be no further use
616 * of the caches at this point.
617 *
618 * XXX: It would good to be able to assert that the zone is being
619 * torn down to prevent improper use of cache_drain().
620 *
621 * XXX: We lock the zone before passing into bucket_cache_drain() as
622 * it is used elsewhere. Should the tear-down path be made special
623 * there in some form?
624 */
625 CPU_FOREACH(cpu) {
626 cache = &zone->uz_cpu[cpu];
627 bucket_drain(zone, cache->uc_allocbucket);
628 bucket_drain(zone, cache->uc_freebucket);
629 if (cache->uc_allocbucket != NULL)
630 bucket_free(cache->uc_allocbucket);
631 if (cache->uc_freebucket != NULL)
632 bucket_free(cache->uc_freebucket);
633 cache->uc_allocbucket = cache->uc_freebucket = NULL;
634 }
635 ZONE_LOCK(zone);
636 bucket_cache_drain(zone);
637 ZONE_UNLOCK(zone);
638}
639
640/*
641 * Drain the cached buckets from a zone. Expects a locked zone on entry.
642 */
643static void
644bucket_cache_drain(uma_zone_t zone)
645{
646 uma_bucket_t bucket;
647
648 /*
649 * Drain the bucket queues and free the buckets, we just keep two per
650 * cpu (alloc/free).
651 */
652 while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
653 LIST_REMOVE(bucket, ub_link);
654 ZONE_UNLOCK(zone);
655 bucket_drain(zone, bucket);
656 bucket_free(bucket);
657 ZONE_LOCK(zone);
658 }
659
660 /* Now we do the free queue.. */
661 while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
662 LIST_REMOVE(bucket, ub_link);
663 bucket_free(bucket);
664 }
665}
666
667/*
668 * Frees pages from a keg back to the system. This is done on demand from
669 * the pageout daemon.
670 *
671 * Returns nothing.
672 */
673static void
674keg_drain(uma_keg_t keg)
675{
676 struct slabhead freeslabs = { 0 };
677 uma_slab_t slab;
678 uma_slab_t n;
679 u_int8_t flags;
680 u_int8_t *mem;
681 int i;
682
683 /*
684 * We don't want to take pages from statically allocated kegs at this
685 * time
686 */
687 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
688 return;
689
690#ifdef UMA_DEBUG
691 printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
692#endif
693 KEG_LOCK(keg);
694 if (keg->uk_free == 0)
695 goto finished;
696
697 slab = LIST_FIRST(&keg->uk_free_slab);
698 while (slab) {
699 n = LIST_NEXT(slab, us_link);
700
701 /* We have no where to free these to */
702 if (slab->us_flags & UMA_SLAB_BOOT) {
703 slab = n;
704 continue;
705 }
706
707 LIST_REMOVE(slab, us_link);
708 keg->uk_pages -= keg->uk_ppera;
709 keg->uk_free -= keg->uk_ipers;
710
711 if (keg->uk_flags & UMA_ZONE_HASH)
712 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
713
714 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
715
716 slab = n;
717 }
718finished:
719 KEG_UNLOCK(keg);
720
721 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
722 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
723 if (keg->uk_fini)
724 for (i = 0; i < keg->uk_ipers; i++)
725 keg->uk_fini(
726 slab->us_data + (keg->uk_rsize * i),
727 keg->uk_size);
728 flags = slab->us_flags;
729 mem = slab->us_data;
730
731 if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
732 vm_object_t obj;
733
734 if (flags & UMA_SLAB_KMEM)
735 obj = kmem_object;
736 else if (flags & UMA_SLAB_KERNEL)
737 obj = kernel_object;
738 else
739 obj = NULL;
740 for (i = 0; i < keg->uk_ppera; i++)
741 vsetobj((vm_offset_t)mem + (i * PAGE_SIZE),
742 obj);
743 }
744 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
745 zone_free_item(keg->uk_slabzone, slab, NULL,
746 SKIP_NONE, ZFREE_STATFREE);
747#ifdef UMA_DEBUG
748 printf("%s: Returning %d bytes.\n",
749 keg->uk_name, UMA_SLAB_SIZE * keg->uk_ppera);
750#endif
751 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags);
752 }
753}
754
755static void
756zone_drain_wait(uma_zone_t zone, int waitok)
757{
758
759 /*
760 * Set draining to interlock with zone_dtor() so we can release our
761 * locks as we go. Only dtor() should do a WAITOK call since it
762 * is the only call that knows the structure will still be available
763 * when it wakes up.
764 */
765 ZONE_LOCK(zone);
766 while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
767 if (waitok == M_NOWAIT)
768 goto out;
769 mtx_unlock(&uma_mtx);
770 msleep(zone, zone->uz_lock, PVM, "zonedrain", 1);
771 mtx_lock(&uma_mtx);
772 }
773 zone->uz_flags |= UMA_ZFLAG_DRAINING;
774 bucket_cache_drain(zone);
775 ZONE_UNLOCK(zone);
776 /*
777 * The DRAINING flag protects us from being freed while
778 * we're running. Normally the uma_mtx would protect us but we
779 * must be able to release and acquire the right lock for each keg.
780 */
781 zone_foreach_keg(zone, &keg_drain);
782 ZONE_LOCK(zone);
783 zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
784 wakeup(zone);
785out:
786 ZONE_UNLOCK(zone);
787}
788
789void
790zone_drain(uma_zone_t zone)
791{
792
793 zone_drain_wait(zone, M_NOWAIT);
794}
795
796/*
797 * Allocate a new slab for a keg. This does not insert the slab onto a list.
798 *
799 * Arguments:
800 * wait Shall we wait?
801 *
802 * Returns:
803 * The slab that was allocated or NULL if there is no memory and the
804 * caller specified M_NOWAIT.
805 */
806static uma_slab_t
807keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
808{
809 uma_slabrefcnt_t slabref;
810 uma_alloc allocf;
811 uma_slab_t slab;
812 u_int8_t *mem;
813 u_int8_t flags;
814 int i;
815
816 mtx_assert(&keg->uk_lock, MA_OWNED);
817 slab = NULL;
818
819#ifdef UMA_DEBUG
820 printf("slab_zalloc: Allocating a new slab for %s\n", keg->uk_name);
821#endif
822 allocf = keg->uk_allocf;
823 KEG_UNLOCK(keg);
824
825 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
826 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
827 if (slab == NULL) {
828 KEG_LOCK(keg);
829 return NULL;
830 }
831 }
832
833 /*
834 * This reproduces the old vm_zone behavior of zero filling pages the
835 * first time they are added to a zone.
836 *
837 * Malloced items are zeroed in uma_zalloc.
838 */
839
840 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
841 wait |= M_ZERO;
842 else
843 wait &= ~M_ZERO;
844
845 if (keg->uk_flags & UMA_ZONE_NODUMP)
846 wait |= M_NODUMP;
847
848 /* zone is passed for legacy reasons. */
849 mem = allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE, &flags, wait);
850 if (mem == NULL) {
851 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
852 zone_free_item(keg->uk_slabzone, slab, NULL,
853 SKIP_NONE, ZFREE_STATFREE);
854 KEG_LOCK(keg);
855 return (NULL);
856 }
857
858 /* Point the slab into the allocated memory */
859 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
860 slab = (uma_slab_t )(mem + keg->uk_pgoff);
861
862 if (keg->uk_flags & UMA_ZONE_VTOSLAB)
863 for (i = 0; i < keg->uk_ppera; i++)
864 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
865
866 slab->us_keg = keg;
867 slab->us_data = mem;
868 slab->us_freecount = keg->uk_ipers;
869 slab->us_firstfree = 0;
870 slab->us_flags = flags;
871
872 if (keg->uk_flags & UMA_ZONE_REFCNT) {
873 slabref = (uma_slabrefcnt_t)slab;
874 for (i = 0; i < keg->uk_ipers; i++) {
875 slabref->us_freelist[i].us_refcnt = 0;
876 slabref->us_freelist[i].us_item = i+1;
877 }
878 } else {
879 for (i = 0; i < keg->uk_ipers; i++)
880 slab->us_freelist[i].us_item = i+1;
881 }
882
883 if (keg->uk_init != NULL) {
884 for (i = 0; i < keg->uk_ipers; i++)
885 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
886 keg->uk_size, wait) != 0)
887 break;
888 if (i != keg->uk_ipers) {
889 if (keg->uk_fini != NULL) {
890 for (i--; i > -1; i--)
891 keg->uk_fini(slab->us_data +
892 (keg->uk_rsize * i),
893 keg->uk_size);
894 }
895 if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
896 vm_object_t obj;
897
898 if (flags & UMA_SLAB_KMEM)
899 obj = kmem_object;
900 else if (flags & UMA_SLAB_KERNEL)
901 obj = kernel_object;
902 else
903 obj = NULL;
904 for (i = 0; i < keg->uk_ppera; i++)
905 vsetobj((vm_offset_t)mem +
906 (i * PAGE_SIZE), obj);
907 }
908 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
909 zone_free_item(keg->uk_slabzone, slab,
910 NULL, SKIP_NONE, ZFREE_STATFREE);
911 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera,
912 flags);
913 KEG_LOCK(keg);
914 return (NULL);
915 }
916 }
917 KEG_LOCK(keg);
918
919 if (keg->uk_flags & UMA_ZONE_HASH)
920 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
921
922 keg->uk_pages += keg->uk_ppera;
923 keg->uk_free += keg->uk_ipers;
924
925 return (slab);
926}
927
928/*
929 * This function is intended to be used early on in place of page_alloc() so
930 * that we may use the boot time page cache to satisfy allocations before
931 * the VM is ready.
932 */
933static void *
934startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
935{
936 uma_keg_t keg;
937 uma_slab_t tmps;
938 int pages, check_pages;
939
940 keg = zone_first_keg(zone);
941 pages = howmany(bytes, PAGE_SIZE);
942 check_pages = pages - 1;
943 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
944
945 /*
946 * Check our small startup cache to see if it has pages remaining.
947 */
948 mtx_lock(&uma_boot_pages_mtx);
949
950 /* First check if we have enough room. */
951 tmps = LIST_FIRST(&uma_boot_pages);
952 while (tmps != NULL && check_pages-- > 0)
953 tmps = LIST_NEXT(tmps, us_link);
954 if (tmps != NULL) {
955 /*
956 * It's ok to lose tmps references. The last one will
957 * have tmps->us_data pointing to the start address of
958 * "pages" contiguous pages of memory.
959 */
960 while (pages-- > 0) {
961 tmps = LIST_FIRST(&uma_boot_pages);
962 LIST_REMOVE(tmps, us_link);
963 }
964 mtx_unlock(&uma_boot_pages_mtx);
965 *pflag = tmps->us_flags;
966 return (tmps->us_data);
967 }
968 mtx_unlock(&uma_boot_pages_mtx);
969 if (booted < UMA_STARTUP2)
970 panic("UMA: Increase vm.boot_pages");
971 /*
972 * Now that we've booted reset these users to their real allocator.
973 */
974#ifdef UMA_MD_SMALL_ALLOC
975 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
976#else
977 keg->uk_allocf = page_alloc;
978#endif
979 return keg->uk_allocf(zone, bytes, pflag, wait);
980}
981
982/*
983 * Allocates a number of pages from the system
984 *
985 * Arguments:
986 * bytes The number of bytes requested
987 * wait Shall we wait?
988 *
989 * Returns:
990 * A pointer to the alloced memory or possibly
991 * NULL if M_NOWAIT is set.
992 */
993static void *
994page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
995{
996 void *p; /* Returned page */
997
998 *pflag = UMA_SLAB_KMEM;
999 p = (void *) kmem_malloc(kmem_map, bytes, wait);
1000
1001 return (p);
1002}
1003
1004/*
1005 * Allocates a number of pages from within an object
1006 *
1007 * Arguments:
1008 * bytes The number of bytes requested
1009 * wait Shall we wait?
1010 *
1011 * Returns:
1012 * A pointer to the alloced memory or possibly
1013 * NULL if M_NOWAIT is set.
1014 */
1015static void *
1016obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
1017{
1018 vm_object_t object;
1019 vm_offset_t retkva, zkva;
1020 vm_page_t p;
1021 int pages, startpages;
1022 uma_keg_t keg;
1023
1024 keg = zone_first_keg(zone);
1025 object = keg->uk_obj;
1026 retkva = 0;
1027
1028 /*
1029 * This looks a little weird since we're getting one page at a time.
1030 */
1031 VM_OBJECT_LOCK(object);
1032 p = TAILQ_LAST(&object->memq, pglist);
1033 pages = p != NULL ? p->pindex + 1 : 0;
1034 startpages = pages;
1035 zkva = keg->uk_kva + pages * PAGE_SIZE;
1036 for (; bytes > 0; bytes -= PAGE_SIZE) {
1037 p = vm_page_alloc(object, pages,
1038 VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED);
1039 if (p == NULL) {
1040 if (pages != startpages)
1041 pmap_qremove(retkva, pages - startpages);
1042 while (pages != startpages) {
1043 pages--;
1044 p = TAILQ_LAST(&object->memq, pglist);
1045 vm_page_unwire(p, 0);
1046 vm_page_free(p);
1047 }
1048 retkva = 0;
1049 goto done;
1050 }
1051 pmap_qenter(zkva, &p, 1);
1052 if (retkva == 0)
1053 retkva = zkva;
1054 zkva += PAGE_SIZE;
1055 pages += 1;
1056 }
1057done:
1058 VM_OBJECT_UNLOCK(object);
1059 *flags = UMA_SLAB_PRIV;
1060
1061 return ((void *)retkva);
1062}
1063
1064/*
1065 * Frees a number of pages to the system
1066 *
1067 * Arguments:
1068 * mem A pointer to the memory to be freed
1069 * size The size of the memory being freed
1070 * flags The original p->us_flags field
1071 *
1072 * Returns:
1073 * Nothing
1074 */
1075static void
1076page_free(void *mem, int size, u_int8_t flags)
1077{
1078 vm_map_t map;
1079
1080 if (flags & UMA_SLAB_KMEM)
1081 map = kmem_map;
1082 else if (flags & UMA_SLAB_KERNEL)
1083 map = kernel_map;
1084 else
1085 panic("UMA: page_free used with invalid flags %d", flags);
1086
1087 kmem_free(map, (vm_offset_t)mem, size);
1088}
1089
1090/*
1091 * Zero fill initializer
1092 *
1093 * Arguments/Returns follow uma_init specifications
1094 */
1095static int
1096zero_init(void *mem, int size, int flags)
1097{
1098 bzero(mem, size);
1099 return (0);
1100}
1101
1102/*
1103 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1104 *
1105 * Arguments
1106 * keg The zone we should initialize
1107 *
1108 * Returns
1109 * Nothing
1110 */
1111static void
1112keg_small_init(uma_keg_t keg)
1113{
1114 u_int rsize;
1115 u_int memused;
1116 u_int wastedspace;
1117 u_int shsize;
1118
1119 KASSERT(keg != NULL, ("Keg is null in keg_small_init"));
1120 rsize = keg->uk_size;
1121
1122 if (rsize < UMA_SMALLEST_UNIT)
1123 rsize = UMA_SMALLEST_UNIT;
1124 if (rsize & keg->uk_align)
1125 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1126
1127 keg->uk_rsize = rsize;
1128 keg->uk_ppera = 1;
1129
1130 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1131 shsize = 0;
1132 } else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1133 rsize += UMA_FRITMREF_SZ; /* linkage & refcnt */
1134 shsize = sizeof(struct uma_slab_refcnt);
1135 } else {
1136 rsize += UMA_FRITM_SZ; /* Account for linkage */
1137 shsize = sizeof(struct uma_slab);
1138 }
1139
1140 keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize;
1141 KASSERT(keg->uk_ipers != 0, ("keg_small_init: ipers is 0"));
1142 memused = keg->uk_ipers * rsize + shsize;
1143 wastedspace = UMA_SLAB_SIZE - memused;
1144
1145 /*
1146 * We can't do OFFPAGE if we're internal or if we've been
1147 * asked to not go to the VM for buckets. If we do this we
1148 * may end up going to the VM (kmem_map) for slabs which we
1149 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1150 * result of UMA_ZONE_VM, which clearly forbids it.
1151 */
1152 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1153 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1154 return;
1155
1156 if ((wastedspace >= UMA_MAX_WASTE) &&
1157 (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) {
1158 keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize;
1159 KASSERT(keg->uk_ipers <= 255,
1160 ("keg_small_init: keg->uk_ipers too high!"));
1161#ifdef UMA_DEBUG
1162 printf("UMA decided we need offpage slab headers for "
1163 "keg: %s, calculated wastedspace = %d, "
1164 "maximum wasted space allowed = %d, "
1165 "calculated ipers = %d, "
1166 "new wasted space = %d\n", keg->uk_name, wastedspace,
1167 UMA_MAX_WASTE, keg->uk_ipers,
1168 UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize);
1169#endif
1170 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1171 if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1172 keg->uk_flags |= UMA_ZONE_HASH;
1173 }
1174}
1175
1176/*
1177 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1178 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1179 * more complicated.
1180 *
1181 * Arguments
1182 * keg The keg we should initialize
1183 *
1184 * Returns
1185 * Nothing
1186 */
1187static void
1188keg_large_init(uma_keg_t keg)
1189{
1190 int pages;
1191
1192 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1193 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1194 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1195
1196 pages = keg->uk_size / UMA_SLAB_SIZE;
1197
1198 /* Account for remainder */
1199 if ((pages * UMA_SLAB_SIZE) < keg->uk_size)
1200 pages++;
1201
1202 keg->uk_ppera = pages;
1203 keg->uk_ipers = 1;
1204 keg->uk_rsize = keg->uk_size;
1205
1206 /* We can't do OFFPAGE if we're internal, bail out here. */
1207 if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1208 return;
1209
1210 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1211 if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1212 keg->uk_flags |= UMA_ZONE_HASH;
1213}
1214
1215static void
1216keg_cachespread_init(uma_keg_t keg)
1217{
1218 int alignsize;
1219 int trailer;
1220 int pages;
1221 int rsize;
1222
1223 alignsize = keg->uk_align + 1;
1224 rsize = keg->uk_size;
1225 /*
1226 * We want one item to start on every align boundary in a page. To
1227 * do this we will span pages. We will also extend the item by the
1228 * size of align if it is an even multiple of align. Otherwise, it
1229 * would fall on the same boundary every time.
1230 */
1231 if (rsize & keg->uk_align)
1232 rsize = (rsize & ~keg->uk_align) + alignsize;
1233 if ((rsize & alignsize) == 0)
1234 rsize += alignsize;
1235 trailer = rsize - keg->uk_size;
1236 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1237 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1238 keg->uk_rsize = rsize;
1239 keg->uk_ppera = pages;
1240 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1241 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1242 KASSERT(keg->uk_ipers <= uma_max_ipers,
1243 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1244 keg->uk_ipers));
1245}
1246
1247/*
1248 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1249 * the keg onto the global keg list.
1250 *
1251 * Arguments/Returns follow uma_ctor specifications
1252 * udata Actually uma_kctor_args
1253 */
1254static int
1255keg_ctor(void *mem, int size, void *udata, int flags)
1256{
1257 struct uma_kctor_args *arg = udata;
1258 uma_keg_t keg = mem;
1259 uma_zone_t zone;
1260
1261 bzero(keg, size);
1262 keg->uk_size = arg->size;
1263 keg->uk_init = arg->uminit;
1264 keg->uk_fini = arg->fini;
1265 keg->uk_align = arg->align;
1266 keg->uk_free = 0;
1267 keg->uk_pages = 0;
1268 keg->uk_flags = arg->flags;
1269 keg->uk_allocf = page_alloc;
1270 keg->uk_freef = page_free;
1271 keg->uk_recurse = 0;
1272 keg->uk_slabzone = NULL;
1273
1274 /*
1275 * The master zone is passed to us at keg-creation time.
1276 */
1277 zone = arg->zone;
1278 keg->uk_name = zone->uz_name;
1279
1280 if (arg->flags & UMA_ZONE_VM)
1281 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1282
1283 if (arg->flags & UMA_ZONE_ZINIT)
1284 keg->uk_init = zero_init;
1285
1286 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1287 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1288
1289 /*
1290 * The +UMA_FRITM_SZ added to uk_size is to account for the
1291 * linkage that is added to the size in keg_small_init(). If
1292 * we don't account for this here then we may end up in
1293 * keg_small_init() with a calculated 'ipers' of 0.
1294 */
1295 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1296 if (keg->uk_flags & UMA_ZONE_CACHESPREAD)
1297 keg_cachespread_init(keg);
1298 else if ((keg->uk_size+UMA_FRITMREF_SZ) >
1299 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)))
1300 keg_large_init(keg);
1301 else
1302 keg_small_init(keg);
1303 } else {
1304 if (keg->uk_flags & UMA_ZONE_CACHESPREAD)
1305 keg_cachespread_init(keg);
1306 else if ((keg->uk_size+UMA_FRITM_SZ) >
1307 (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1308 keg_large_init(keg);
1309 else
1310 keg_small_init(keg);
1311 }
1312
1313 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1314 if (keg->uk_flags & UMA_ZONE_REFCNT)
1315 keg->uk_slabzone = slabrefzone;
1316 else
1317 keg->uk_slabzone = slabzone;
1318 }
1319
1320 /*
1321 * If we haven't booted yet we need allocations to go through the
1322 * startup cache until the vm is ready.
1323 */
1324 if (keg->uk_ppera == 1) {
1325#ifdef UMA_MD_SMALL_ALLOC
1326 keg->uk_allocf = uma_small_alloc;
1327 keg->uk_freef = uma_small_free;
1328
1329 if (booted < UMA_STARTUP)
1330 keg->uk_allocf = startup_alloc;
1331#else
1332 if (booted < UMA_STARTUP2)
1333 keg->uk_allocf = startup_alloc;
1334#endif
1335 } else if (booted < UMA_STARTUP2 &&
1336 (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1337 keg->uk_allocf = startup_alloc;
1338
1339 /*
1340 * Initialize keg's lock (shared among zones).
1341 */
1342 if (arg->flags & UMA_ZONE_MTXCLASS)
1343 KEG_LOCK_INIT(keg, 1);
1344 else
1345 KEG_LOCK_INIT(keg, 0);
1346
1347 /*
1348 * If we're putting the slab header in the actual page we need to
1349 * figure out where in each page it goes. This calculates a right
1350 * justified offset into the memory on an ALIGN_PTR boundary.
1351 */
1352 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1353 u_int totsize;
1354
1355 /* Size of the slab struct and free list */
1356 if (keg->uk_flags & UMA_ZONE_REFCNT)
1357 totsize = sizeof(struct uma_slab_refcnt) +
1358 keg->uk_ipers * UMA_FRITMREF_SZ;
1359 else
1360 totsize = sizeof(struct uma_slab) +
1361 keg->uk_ipers * UMA_FRITM_SZ;
1362
1363 if (totsize & UMA_ALIGN_PTR)
1364 totsize = (totsize & ~UMA_ALIGN_PTR) +
1365 (UMA_ALIGN_PTR + 1);
1366 keg->uk_pgoff = (UMA_SLAB_SIZE * keg->uk_ppera) - totsize;
1367
1368 if (keg->uk_flags & UMA_ZONE_REFCNT)
1369 totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt)
1370 + keg->uk_ipers * UMA_FRITMREF_SZ;
1371 else
1372 totsize = keg->uk_pgoff + sizeof(struct uma_slab)
1373 + keg->uk_ipers * UMA_FRITM_SZ;
1374
1375 /*
1376 * The only way the following is possible is if with our
1377 * UMA_ALIGN_PTR adjustments we are now bigger than
1378 * UMA_SLAB_SIZE. I haven't checked whether this is
1379 * mathematically possible for all cases, so we make
1380 * sure here anyway.
1381 */
1382 if (totsize > UMA_SLAB_SIZE * keg->uk_ppera) {
1383 printf("zone %s ipers %d rsize %d size %d\n",
1384 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1385 keg->uk_size);
1386 panic("UMA slab won't fit.");
1387 }
1388 }
1389
1390 if (keg->uk_flags & UMA_ZONE_HASH)
1391 hash_alloc(&keg->uk_hash);
1392
1393#ifdef UMA_DEBUG
1394 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1395 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1396 keg->uk_ipers, keg->uk_ppera,
1397 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1398#endif
1399
1400 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1401
1402 mtx_lock(&uma_mtx);
1403 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1404 mtx_unlock(&uma_mtx);
1405 return (0);
1406}
1407
1408/*
1409 * Zone header ctor. This initializes all fields, locks, etc.
1410 *
1411 * Arguments/Returns follow uma_ctor specifications
1412 * udata Actually uma_zctor_args
1413 */
1414static int
1415zone_ctor(void *mem, int size, void *udata, int flags)
1416{
1417 struct uma_zctor_args *arg = udata;
1418 uma_zone_t zone = mem;
1419 uma_zone_t z;
1420 uma_keg_t keg;
1421
1422 bzero(zone, size);
1423 zone->uz_name = arg->name;
1424 zone->uz_ctor = arg->ctor;
1425 zone->uz_dtor = arg->dtor;
1426 zone->uz_slab = zone_fetch_slab;
1427 zone->uz_init = NULL;
1428 zone->uz_fini = NULL;
1429 zone->uz_allocs = 0;
1430 zone->uz_frees = 0;
1431 zone->uz_fails = 0;
1432 zone->uz_sleeps = 0;
1433 zone->uz_fills = zone->uz_count = 0;
1434 zone->uz_flags = 0;
1435 keg = arg->keg;
1436
1437 if (arg->flags & UMA_ZONE_SECONDARY) {
1438 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1439 zone->uz_init = arg->uminit;
1440 zone->uz_fini = arg->fini;
1441 zone->uz_lock = &keg->uk_lock;
1442 zone->uz_flags |= UMA_ZONE_SECONDARY;
1443 mtx_lock(&uma_mtx);
1444 ZONE_LOCK(zone);
1445 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1446 if (LIST_NEXT(z, uz_link) == NULL) {
1447 LIST_INSERT_AFTER(z, zone, uz_link);
1448 break;
1449 }
1450 }
1451 ZONE_UNLOCK(zone);
1452 mtx_unlock(&uma_mtx);
1453 } else if (keg == NULL) {
1454 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1455 arg->align, arg->flags)) == NULL)
1456 return (ENOMEM);
1457 } else {
1458 struct uma_kctor_args karg;
1459 int error;
1460
1461 /* We should only be here from uma_startup() */
1462 karg.size = arg->size;
1463 karg.uminit = arg->uminit;
1464 karg.fini = arg->fini;
1465 karg.align = arg->align;
1466 karg.flags = arg->flags;
1467 karg.zone = zone;
1468 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1469 flags);
1470 if (error)
1471 return (error);
1472 }
1473 /*
1474 * Link in the first keg.
1475 */
1476 zone->uz_klink.kl_keg = keg;
1477 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1478 zone->uz_lock = &keg->uk_lock;
1479 zone->uz_size = keg->uk_size;
1480 zone->uz_flags |= (keg->uk_flags &
1481 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1482
1483 /*
1484 * Some internal zones don't have room allocated for the per cpu
1485 * caches. If we're internal, bail out here.
1486 */
1487 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1488 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1489 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1490 return (0);
1491 }
1492
1493 if (keg->uk_flags & UMA_ZONE_MAXBUCKET)
1494 zone->uz_count = BUCKET_MAX;
1495 else if (keg->uk_ipers <= BUCKET_MAX)
1496 zone->uz_count = keg->uk_ipers;
1497 else
1498 zone->uz_count = BUCKET_MAX;
1499 return (0);
1500}
1501
1502/*
1503 * Keg header dtor. This frees all data, destroys locks, frees the hash
1504 * table and removes the keg from the global list.
1505 *
1506 * Arguments/Returns follow uma_dtor specifications
1507 * udata unused
1508 */
1509static void
1510keg_dtor(void *arg, int size, void *udata)
1511{
1512 uma_keg_t keg;
1513
1514 keg = (uma_keg_t)arg;
1515 KEG_LOCK(keg);
1516 if (keg->uk_free != 0) {
1517 printf("Freed UMA keg was not empty (%d items). "
1518 " Lost %d pages of memory.\n",
1519 keg->uk_free, keg->uk_pages);
1520 }
1521 KEG_UNLOCK(keg);
1522
1523 hash_free(&keg->uk_hash);
1524
1525 KEG_LOCK_FINI(keg);
1526}
1527
1528/*
1529 * Zone header dtor.
1530 *
1531 * Arguments/Returns follow uma_dtor specifications
1532 * udata unused
1533 */
1534static void
1535zone_dtor(void *arg, int size, void *udata)
1536{
1537 uma_klink_t klink;
1538 uma_zone_t zone;
1539 uma_keg_t keg;
1540
1541 zone = (uma_zone_t)arg;
1542 keg = zone_first_keg(zone);
1543
1544 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1545 cache_drain(zone);
1546
1547 mtx_lock(&uma_mtx);
1548 LIST_REMOVE(zone, uz_link);
1549 mtx_unlock(&uma_mtx);
1550 /*
1551 * XXX there are some races here where
1552 * the zone can be drained but zone lock
1553 * released and then refilled before we
1554 * remove it... we dont care for now
1555 */
1556 zone_drain_wait(zone, M_WAITOK);
1557 /*
1558 * Unlink all of our kegs.
1559 */
1560 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1561 klink->kl_keg = NULL;
1562 LIST_REMOVE(klink, kl_link);
1563 if (klink == &zone->uz_klink)
1564 continue;
1565 free(klink, M_TEMP);
1566 }
1567 /*
1568 * We only destroy kegs from non secondary zones.
1569 */
1570 if ((zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1571 mtx_lock(&uma_mtx);
1572 LIST_REMOVE(keg, uk_link);
1573 mtx_unlock(&uma_mtx);
1574 zone_free_item(kegs, keg, NULL, SKIP_NONE,
1575 ZFREE_STATFREE);
1576 }
1577}
1578
1579/*
1580 * Traverses every zone in the system and calls a callback
1581 *
1582 * Arguments:
1583 * zfunc A pointer to a function which accepts a zone
1584 * as an argument.
1585 *
1586 * Returns:
1587 * Nothing
1588 */
1589static void
1590zone_foreach(void (*zfunc)(uma_zone_t))
1591{
1592 uma_keg_t keg;
1593 uma_zone_t zone;
1594
1595 mtx_lock(&uma_mtx);
1596 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1597 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1598 zfunc(zone);
1599 }
1600 mtx_unlock(&uma_mtx);
1601}
1602
1603/* Public functions */
1604/* See uma.h */
1605void
1606uma_startup(void *bootmem, int boot_pages)
1607{
1608 struct uma_zctor_args args;
1609 uma_slab_t slab;
1610 u_int slabsize;
1611 u_int objsize, totsize, wsize;
1612 int i;
1613
1614#ifdef UMA_DEBUG
1615 printf("Creating uma keg headers zone and keg.\n");
1616#endif
1617 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1618
1619 /*
1620 * Figure out the maximum number of items-per-slab we'll have if
1621 * we're using the OFFPAGE slab header to track free items, given
1622 * all possible object sizes and the maximum desired wastage
1623 * (UMA_MAX_WASTE).
1624 *
1625 * We iterate until we find an object size for
1626 * which the calculated wastage in keg_small_init() will be
1627 * enough to warrant OFFPAGE. Since wastedspace versus objsize
1628 * is an overall increasing see-saw function, we find the smallest
1629 * objsize such that the wastage is always acceptable for objects
1630 * with that objsize or smaller. Since a smaller objsize always
1631 * generates a larger possible uma_max_ipers, we use this computed
1632 * objsize to calculate the largest ipers possible. Since the
1633 * ipers calculated for OFFPAGE slab headers is always larger than
1634 * the ipers initially calculated in keg_small_init(), we use
1635 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to
1636 * obtain the maximum ipers possible for offpage slab headers.
1637 *
1638 * It should be noted that ipers versus objsize is an inversly
1639 * proportional function which drops off rather quickly so as
1640 * long as our UMA_MAX_WASTE is such that the objsize we calculate
1641 * falls into the portion of the inverse relation AFTER the steep
1642 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386).
1643 *
1644 * Note that we have 8-bits (1 byte) to use as a freelist index
1645 * inside the actual slab header itself and this is enough to
1646 * accomodate us. In the worst case, a UMA_SMALLEST_UNIT sized
1647 * object with offpage slab header would have ipers =
1648 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is
1649 * 1 greater than what our byte-integer freelist index can
1650 * accomodate, but we know that this situation never occurs as
1651 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate
1652 * that we need to go to offpage slab headers. Or, if we do,
1653 * then we trap that condition below and panic in the INVARIANTS case.
1654 */
1655 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE;
1656 totsize = wsize;
1657 objsize = UMA_SMALLEST_UNIT;
1658 while (totsize >= wsize) {
1659 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) /
1660 (objsize + UMA_FRITM_SZ);
1661 totsize *= (UMA_FRITM_SZ + objsize);
1662 objsize++;
1663 }
1664 if (objsize > UMA_SMALLEST_UNIT)
1665 objsize--;
1666 uma_max_ipers = MAX(UMA_SLAB_SIZE / objsize, 64);
1667
1668 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE;
1669 totsize = wsize;
1670 objsize = UMA_SMALLEST_UNIT;
1671 while (totsize >= wsize) {
1672 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) /
1673 (objsize + UMA_FRITMREF_SZ);
1674 totsize *= (UMA_FRITMREF_SZ + objsize);
1675 objsize++;
1676 }
1677 if (objsize > UMA_SMALLEST_UNIT)
1678 objsize--;
1679 uma_max_ipers_ref = MAX(UMA_SLAB_SIZE / objsize, 64);
1680
1681 KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255),
1682 ("uma_startup: calculated uma_max_ipers values too large!"));
1683
1684#ifdef UMA_DEBUG
1685 printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers);
1686 printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n",
1687 uma_max_ipers_ref);
1688#endif
1689
1690 /* "manually" create the initial zone */
1691 args.name = "UMA Kegs";
1692 args.size = sizeof(struct uma_keg);
1693 args.ctor = keg_ctor;
1694 args.dtor = keg_dtor;
1695 args.uminit = zero_init;
1696 args.fini = NULL;
1697 args.keg = &masterkeg;
1698 args.align = 32 - 1;
1699 args.flags = UMA_ZFLAG_INTERNAL;
1700 /* The initial zone has no Per cpu queues so it's smaller */
1701 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1702
1703#ifdef UMA_DEBUG
1704 printf("Filling boot free list.\n");
1705#endif
1706 for (i = 0; i < boot_pages; i++) {
1707 slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE));
1708 slab->us_data = (u_int8_t *)slab;
1709 slab->us_flags = UMA_SLAB_BOOT;
1710 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1711 }
1712 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1713
1714#ifdef UMA_DEBUG
1715 printf("Creating uma zone headers zone and keg.\n");
1716#endif
1717 args.name = "UMA Zones";
1718 args.size = sizeof(struct uma_zone) +
1719 (sizeof(struct uma_cache) * (mp_maxid + 1));
1720 args.ctor = zone_ctor;
1721 args.dtor = zone_dtor;
1722 args.uminit = zero_init;
1723 args.fini = NULL;
1724 args.keg = NULL;
1725 args.align = 32 - 1;
1726 args.flags = UMA_ZFLAG_INTERNAL;
1727 /* The initial zone has no Per cpu queues so it's smaller */
1728 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1729
1730#ifdef UMA_DEBUG
1731 printf("Initializing pcpu cache locks.\n");
1732#endif
1733#ifdef UMA_DEBUG
1734 printf("Creating slab and hash zones.\n");
1735#endif
1736
1737 /*
1738 * This is the max number of free list items we'll have with
1739 * offpage slabs.
1740 */
1741 slabsize = uma_max_ipers * UMA_FRITM_SZ;
1742 slabsize += sizeof(struct uma_slab);
1743
1744 /* Now make a zone for slab headers */
1745 slabzone = uma_zcreate("UMA Slabs",
1746 slabsize,
1747 NULL, NULL, NULL, NULL,
1748 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1749
1750 /*
1751 * We also create a zone for the bigger slabs with reference
1752 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1753 */
1754 slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ;
1755 slabsize += sizeof(struct uma_slab_refcnt);
1756 slabrefzone = uma_zcreate("UMA RCntSlabs",
1757 slabsize,
1758 NULL, NULL, NULL, NULL,
1759 UMA_ALIGN_PTR,
1760 UMA_ZFLAG_INTERNAL);
1761
1762 hashzone = uma_zcreate("UMA Hash",
1763 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1764 NULL, NULL, NULL, NULL,
1765 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1766
1767 bucket_init();
1768
1769 booted = UMA_STARTUP;
1770
1771#ifdef UMA_DEBUG
1772 printf("UMA startup complete.\n");
1773#endif
1774}
1775
1776/* see uma.h */
1777void
1778uma_startup2(void)
1779{
1780 booted = UMA_STARTUP2;
1781 bucket_enable();
1782#ifdef UMA_DEBUG
1783 printf("UMA startup2 complete.\n");
1784#endif
1785}
1786
1787/*
1788 * Initialize our callout handle
1789 *
1790 */
1791
1792static void
1793uma_startup3(void)
1794{
1795#ifdef UMA_DEBUG
1796 printf("Starting callout.\n");
1797#endif
1798 callout_init(&uma_callout, CALLOUT_MPSAFE);
1799 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1800#ifdef UMA_DEBUG
1801 printf("UMA startup3 complete.\n");
1802#endif
1803}
1804
1805static uma_keg_t
1806uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1807 int align, u_int32_t flags)
1808{
1809 struct uma_kctor_args args;
1810
1811 args.size = size;
1812 args.uminit = uminit;
1813 args.fini = fini;
1814 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1815 args.flags = flags;
1816 args.zone = zone;
1817 return (zone_alloc_item(kegs, &args, M_WAITOK));
1818}
1819
1820/* See uma.h */
1821void
1822uma_set_align(int align)
1823{
1824
1825 if (align != UMA_ALIGN_CACHE)
1826 uma_align_cache = align;
1827}
1828
1829/* See uma.h */
1830uma_zone_t
1831uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1832 uma_init uminit, uma_fini fini, int align, u_int32_t flags)
1833
1834{
1835 struct uma_zctor_args args;
1836
1837 /* This stuff is essential for the zone ctor */
1838 args.name = name;
1839 args.size = size;
1840 args.ctor = ctor;
1841 args.dtor = dtor;
1842 args.uminit = uminit;
1843 args.fini = fini;
1844 args.align = align;
1845 args.flags = flags;
1846 args.keg = NULL;
1847
1848 return (zone_alloc_item(zones, &args, M_WAITOK));
1849}
1850
1851/* See uma.h */
1852uma_zone_t
1853uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1854 uma_init zinit, uma_fini zfini, uma_zone_t master)
1855{
1856 struct uma_zctor_args args;
1857 uma_keg_t keg;
1858
1859 keg = zone_first_keg(master);
1860 args.name = name;
1861 args.size = keg->uk_size;
1862 args.ctor = ctor;
1863 args.dtor = dtor;
1864 args.uminit = zinit;
1865 args.fini = zfini;
1866 args.align = keg->uk_align;
1867 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1868 args.keg = keg;
1869
1870 /* XXX Attaches only one keg of potentially many. */
1871 return (zone_alloc_item(zones, &args, M_WAITOK));
1872}
1873
1874static void
1875zone_lock_pair(uma_zone_t a, uma_zone_t b)
1876{
1877 if (a < b) {
1878 ZONE_LOCK(a);
1879 mtx_lock_flags(b->uz_lock, MTX_DUPOK);
1880 } else {
1881 ZONE_LOCK(b);
1882 mtx_lock_flags(a->uz_lock, MTX_DUPOK);
1883 }
1884}
1885
1886static void
1887zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1888{
1889
1890 ZONE_UNLOCK(a);
1891 ZONE_UNLOCK(b);
1892}
1893
1894int
1895uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1896{
1897 uma_klink_t klink;
1898 uma_klink_t kl;
1899 int error;
1900
1901 error = 0;
1902 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1903
1904 zone_lock_pair(zone, master);
1905 /*
1906 * zone must use vtoslab() to resolve objects and must already be
1907 * a secondary.
1908 */
1909 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1910 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1911 error = EINVAL;
1912 goto out;
1913 }
1914 /*
1915 * The new master must also use vtoslab().
1916 */
1917 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
1918 error = EINVAL;
1919 goto out;
1920 }
1921 /*
1922 * Both must either be refcnt, or not be refcnt.
1923 */
1924 if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
1925 (master->uz_flags & UMA_ZONE_REFCNT)) {
1926 error = EINVAL;
1927 goto out;
1928 }
1929 /*
1930 * The underlying object must be the same size. rsize
1931 * may be different.
1932 */
1933 if (master->uz_size != zone->uz_size) {
1934 error = E2BIG;
1935 goto out;
1936 }
1937 /*
1938 * Put it at the end of the list.
1939 */
1940 klink->kl_keg = zone_first_keg(master);
1941 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
1942 if (LIST_NEXT(kl, kl_link) == NULL) {
1943 LIST_INSERT_AFTER(kl, klink, kl_link);
1944 break;
1945 }
1946 }
1947 klink = NULL;
1948 zone->uz_flags |= UMA_ZFLAG_MULTI;
1949 zone->uz_slab = zone_fetch_slab_multi;
1950
1951out:
1952 zone_unlock_pair(zone, master);
1953 if (klink != NULL)
1954 free(klink, M_TEMP);
1955
1956 return (error);
1957}
1958
1959
1960/* See uma.h */
1961void
1962uma_zdestroy(uma_zone_t zone)
1963{
1964
1965 zone_free_item(zones, zone, NULL, SKIP_NONE, ZFREE_STATFREE);
1966}
1967
1968/* See uma.h */
1969void *
1970uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1971{
1972 void *item;
1973 uma_cache_t cache;
1974 uma_bucket_t bucket;
1975 int cpu;
1976
1977 /* This is the fast path allocation */
1978#ifdef UMA_DEBUG_ALLOC_1
1979 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1980#endif
1981 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1982 zone->uz_name, flags);
1983
1984 if (flags & M_WAITOK) {
1985 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1986 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
1987 }
1988#ifdef DEBUG_MEMGUARD
1989 if (memguard_cmp_zone(zone)) {
1990 item = memguard_alloc(zone->uz_size, flags);
1991 if (item != NULL) {
1992 /*
1993 * Avoid conflict with the use-after-free
1994 * protecting infrastructure from INVARIANTS.
1995 */
1996 if (zone->uz_init != NULL &&
1997 zone->uz_init != mtrash_init &&
1998 zone->uz_init(item, zone->uz_size, flags) != 0)
1999 return (NULL);
2000 if (zone->uz_ctor != NULL &&
2001 zone->uz_ctor != mtrash_ctor &&
2002 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2003 zone->uz_fini(item, zone->uz_size);
2004 return (NULL);
2005 }
2006 return (item);
2007 }
2008 /* This is unfortunate but should not be fatal. */
2009 }
2010#endif
2011 /*
2012 * If possible, allocate from the per-CPU cache. There are two
2013 * requirements for safe access to the per-CPU cache: (1) the thread
2014 * accessing the cache must not be preempted or yield during access,
2015 * and (2) the thread must not migrate CPUs without switching which
2016 * cache it accesses. We rely on a critical section to prevent
2017 * preemption and migration. We release the critical section in
2018 * order to acquire the zone mutex if we are unable to allocate from
2019 * the current cache; when we re-acquire the critical section, we
2020 * must detect and handle migration if it has occurred.
2021 */
2022zalloc_restart:
2023 critical_enter();
2024 cpu = curcpu;
2025 cache = &zone->uz_cpu[cpu];
2026
2027zalloc_start:
2028 bucket = cache->uc_allocbucket;
2029
2030 if (bucket) {
2031 if (bucket->ub_cnt > 0) {
2032 bucket->ub_cnt--;
2033 item = bucket->ub_bucket[bucket->ub_cnt];
2034#ifdef INVARIANTS
2035 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2036#endif
2037 KASSERT(item != NULL,
2038 ("uma_zalloc: Bucket pointer mangled."));
2039 cache->uc_allocs++;
2040 critical_exit();
2041#ifdef INVARIANTS
2042 ZONE_LOCK(zone);
2043 uma_dbg_alloc(zone, NULL, item);
2044 ZONE_UNLOCK(zone);
2045#endif
2046 if (zone->uz_ctor != NULL) {
2047 if (zone->uz_ctor(item, zone->uz_size,
2048 udata, flags) != 0) {
2049 zone_free_item(zone, item, udata,
2050 SKIP_DTOR, ZFREE_STATFAIL |
2051 ZFREE_STATFREE);
2052 return (NULL);
2053 }
2054 }
2055 if (flags & M_ZERO)
2056 bzero(item, zone->uz_size);
2057 return (item);
2058 } else if (cache->uc_freebucket) {
2059 /*
2060 * We have run out of items in our allocbucket.
2061 * See if we can switch with our free bucket.
2062 */
2063 if (cache->uc_freebucket->ub_cnt > 0) {
2064#ifdef UMA_DEBUG_ALLOC
2065 printf("uma_zalloc: Swapping empty with"
2066 " alloc.\n");
2067#endif
2068 bucket = cache->uc_freebucket;
2069 cache->uc_freebucket = cache->uc_allocbucket;
2070 cache->uc_allocbucket = bucket;
2071
2072 goto zalloc_start;
2073 }
2074 }
2075 }
2076 /*
2077 * Attempt to retrieve the item from the per-CPU cache has failed, so
2078 * we must go back to the zone. This requires the zone lock, so we
2079 * must drop the critical section, then re-acquire it when we go back
2080 * to the cache. Since the critical section is released, we may be
2081 * preempted or migrate. As such, make sure not to maintain any
2082 * thread-local state specific to the cache from prior to releasing
2083 * the critical section.
2084 */
2085 critical_exit();
2086 ZONE_LOCK(zone);
2087 critical_enter();
2088 cpu = curcpu;
2089 cache = &zone->uz_cpu[cpu];
2090 bucket = cache->uc_allocbucket;
2091 if (bucket != NULL) {
2092 if (bucket->ub_cnt > 0) {
2093 ZONE_UNLOCK(zone);
2094 goto zalloc_start;
2095 }
2096 bucket = cache->uc_freebucket;
2097 if (bucket != NULL && bucket->ub_cnt > 0) {
2098 ZONE_UNLOCK(zone);
2099 goto zalloc_start;
2100 }
2101 }
2102
2103 /* Since we have locked the zone we may as well send back our stats */
2104 zone->uz_allocs += cache->uc_allocs;
2105 cache->uc_allocs = 0;
2106 zone->uz_frees += cache->uc_frees;
2107 cache->uc_frees = 0;
2108
2109 /* Our old one is now a free bucket */
2110 if (cache->uc_allocbucket) {
2111 KASSERT(cache->uc_allocbucket->ub_cnt == 0,
2112 ("uma_zalloc_arg: Freeing a non free bucket."));
2113 LIST_INSERT_HEAD(&zone->uz_free_bucket,
2114 cache->uc_allocbucket, ub_link);
2115 cache->uc_allocbucket = NULL;
2116 }
2117
2118 /* Check the free list for a new alloc bucket */
2119 if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
2120 KASSERT(bucket->ub_cnt != 0,
2121 ("uma_zalloc_arg: Returning an empty bucket."));
2122
2123 LIST_REMOVE(bucket, ub_link);
2124 cache->uc_allocbucket = bucket;
2125 ZONE_UNLOCK(zone);
2126 goto zalloc_start;
2127 }
2128 /* We are no longer associated with this CPU. */
2129 critical_exit();
2130
2131 /* Bump up our uz_count so we get here less */
2132 if (zone->uz_count < BUCKET_MAX)
2133 zone->uz_count++;
2134
2135 /*
2136 * Now lets just fill a bucket and put it on the free list. If that
2137 * works we'll restart the allocation from the begining.
2138 */
2139 if (zone_alloc_bucket(zone, flags)) {
2140 ZONE_UNLOCK(zone);
2141 goto zalloc_restart;
2142 }
2143 ZONE_UNLOCK(zone);
2144 /*
2145 * We may not be able to get a bucket so return an actual item.
2146 */
2147#ifdef UMA_DEBUG
2148 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2149#endif
2150
2151 item = zone_alloc_item(zone, udata, flags);
2152 return (item);
2153}
2154
2155static uma_slab_t
2156keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2157{
2158 uma_slab_t slab;
2159
2160 mtx_assert(&keg->uk_lock, MA_OWNED);
2161 slab = NULL;
2162
2163 for (;;) {
2164 /*
2165 * Find a slab with some space. Prefer slabs that are partially
2166 * used over those that are totally full. This helps to reduce
2167 * fragmentation.
2168 */
2169 if (keg->uk_free != 0) {
2170 if (!LIST_EMPTY(&keg->uk_part_slab)) {
2171 slab = LIST_FIRST(&keg->uk_part_slab);
2172 } else {
2173 slab = LIST_FIRST(&keg->uk_free_slab);
2174 LIST_REMOVE(slab, us_link);
2175 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2176 us_link);
2177 }
2178 MPASS(slab->us_keg == keg);
2179 return (slab);
2180 }
2181
2182 /*
2183 * M_NOVM means don't ask at all!
2184 */
2185 if (flags & M_NOVM)
2186 break;
2187
2188 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2189 keg->uk_flags |= UMA_ZFLAG_FULL;
2190 /*
2191 * If this is not a multi-zone, set the FULL bit.
2192 * Otherwise slab_multi() takes care of it.
2193 */
2194 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0)
2195 zone->uz_flags |= UMA_ZFLAG_FULL;
2196 if (flags & M_NOWAIT)
2197 break;
2198 zone->uz_sleeps++;
2199 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2200 continue;
2201 }
2202 keg->uk_recurse++;
2203 slab = keg_alloc_slab(keg, zone, flags);
2204 keg->uk_recurse--;
2205 /*
2206 * If we got a slab here it's safe to mark it partially used
2207 * and return. We assume that the caller is going to remove
2208 * at least one item.
2209 */
2210 if (slab) {
2211 MPASS(slab->us_keg == keg);
2212 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2213 return (slab);
2214 }
2215 /*
2216 * We might not have been able to get a slab but another cpu
2217 * could have while we were unlocked. Check again before we
2218 * fail.
2219 */
2220 flags |= M_NOVM;
2221 }
2222 return (slab);
2223}
2224
2225static inline void
2226zone_relock(uma_zone_t zone, uma_keg_t keg)
2227{
2228 if (zone->uz_lock != &keg->uk_lock) {
2229 KEG_UNLOCK(keg);
2230 ZONE_LOCK(zone);
2231 }
2232}
2233
2234static inline void
2235keg_relock(uma_keg_t keg, uma_zone_t zone)
2236{
2237 if (zone->uz_lock != &keg->uk_lock) {
2238 ZONE_UNLOCK(zone);
2239 KEG_LOCK(keg);
2240 }
2241}
2242
2243static uma_slab_t
2244zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2245{
2246 uma_slab_t slab;
2247
2248 if (keg == NULL)
2249 keg = zone_first_keg(zone);
2250 /*
2251 * This is to prevent us from recursively trying to allocate
2252 * buckets. The problem is that if an allocation forces us to
2253 * grab a new bucket we will call page_alloc, which will go off
2254 * and cause the vm to allocate vm_map_entries. If we need new
2255 * buckets there too we will recurse in kmem_alloc and bad
2256 * things happen. So instead we return a NULL bucket, and make
2257 * the code that allocates buckets smart enough to deal with it
2258 */
2259 if (keg->uk_flags & UMA_ZFLAG_BUCKET && keg->uk_recurse != 0)
2260 return (NULL);
2261
2262 for (;;) {
2263 slab = keg_fetch_slab(keg, zone, flags);
2264 if (slab)
2265 return (slab);
2266 if (flags & (M_NOWAIT | M_NOVM))
2267 break;
2268 }
2269 return (NULL);
2270}
2271
2272/*
2273 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2274 * with the keg locked. Caller must call zone_relock() afterwards if the
2275 * zone lock is required. On NULL the zone lock is held.
2276 *
2277 * The last pointer is used to seed the search. It is not required.
2278 */
2279static uma_slab_t
2280zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2281{
2282 uma_klink_t klink;
2283 uma_slab_t slab;
2284 uma_keg_t keg;
2285 int flags;
2286 int empty;
2287 int full;
2288
2289 /*
2290 * Don't wait on the first pass. This will skip limit tests
2291 * as well. We don't want to block if we can find a provider
2292 * without blocking.
2293 */
2294 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2295 /*
2296 * Use the last slab allocated as a hint for where to start
2297 * the search.
2298 */
2299 if (last) {
2300 slab = keg_fetch_slab(last, zone, flags);
2301 if (slab)
2302 return (slab);
2303 zone_relock(zone, last);
2304 last = NULL;
2305 }
2306 /*
2307 * Loop until we have a slab incase of transient failures
2308 * while M_WAITOK is specified. I'm not sure this is 100%
2309 * required but we've done it for so long now.
2310 */
2311 for (;;) {
2312 empty = 0;
2313 full = 0;
2314 /*
2315 * Search the available kegs for slabs. Be careful to hold the
2316 * correct lock while calling into the keg layer.
2317 */
2318 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2319 keg = klink->kl_keg;
2320 keg_relock(keg, zone);
2321 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2322 slab = keg_fetch_slab(keg, zone, flags);
2323 if (slab)
2324 return (slab);
2325 }
2326 if (keg->uk_flags & UMA_ZFLAG_FULL)
2327 full++;
2328 else
2329 empty++;
2330 zone_relock(zone, keg);
2331 }
2332 if (rflags & (M_NOWAIT | M_NOVM))
2333 break;
2334 flags = rflags;
2335 /*
2336 * All kegs are full. XXX We can't atomically check all kegs
2337 * and sleep so just sleep for a short period and retry.
2338 */
2339 if (full && !empty) {
2340 zone->uz_flags |= UMA_ZFLAG_FULL;
2341 zone->uz_sleeps++;
2342 msleep(zone, zone->uz_lock, PVM, "zonelimit", hz/100);
2343 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2344 continue;
2345 }
2346 }
2347 return (NULL);
2348}
2349
2350static void *
2351slab_alloc_item(uma_zone_t zone, uma_slab_t slab)
2352{
2353 uma_keg_t keg;
2354 uma_slabrefcnt_t slabref;
2355 void *item;
2356 u_int8_t freei;
2357
2358 keg = slab->us_keg;
2359 mtx_assert(&keg->uk_lock, MA_OWNED);
2360
2361 freei = slab->us_firstfree;
2362 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2363 slabref = (uma_slabrefcnt_t)slab;
2364 slab->us_firstfree = slabref->us_freelist[freei].us_item;
2365 } else {
2366 slab->us_firstfree = slab->us_freelist[freei].us_item;
2367 }
2368 item = slab->us_data + (keg->uk_rsize * freei);
2369
2370 slab->us_freecount--;
2371 keg->uk_free--;
2372#ifdef INVARIANTS
2373 uma_dbg_alloc(zone, slab, item);
2374#endif
2375 /* Move this slab to the full list */
2376 if (slab->us_freecount == 0) {
2377 LIST_REMOVE(slab, us_link);
2378 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2379 }
2380
2381 return (item);
2382}
2383
2384static int
2385zone_alloc_bucket(uma_zone_t zone, int flags)
2386{
2387 uma_bucket_t bucket;
2388 uma_slab_t slab;
2389 uma_keg_t keg;
2390 int16_t saved;
2391 int max, origflags = flags;
2392
2393 /*
2394 * Try this zone's free list first so we don't allocate extra buckets.
2395 */
2396 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2397 KASSERT(bucket->ub_cnt == 0,
2398 ("zone_alloc_bucket: Bucket on free list is not empty."));
2399 LIST_REMOVE(bucket, ub_link);
2400 } else {
2401 int bflags;
2402
2403 bflags = (flags & ~M_ZERO);
2404 if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2405 bflags |= M_NOVM;
2406
2407 ZONE_UNLOCK(zone);
2408 bucket = bucket_alloc(zone->uz_count, bflags);
2409 ZONE_LOCK(zone);
2410 }
2411
2412 if (bucket == NULL) {
2413 return (0);
2414 }
2415
2416#ifdef SMP
2417 /*
2418 * This code is here to limit the number of simultaneous bucket fills
2419 * for any given zone to the number of per cpu caches in this zone. This
2420 * is done so that we don't allocate more memory than we really need.
2421 */
2422 if (zone->uz_fills >= mp_ncpus)
2423 goto done;
2424
2425#endif
2426 zone->uz_fills++;
2427
2428 max = MIN(bucket->ub_entries, zone->uz_count);
2429 /* Try to keep the buckets totally full */
2430 saved = bucket->ub_cnt;
2431 slab = NULL;
2432 keg = NULL;
2433 while (bucket->ub_cnt < max &&
2434 (slab = zone->uz_slab(zone, keg, flags)) != NULL) {
2435 keg = slab->us_keg;
2436 while (slab->us_freecount && bucket->ub_cnt < max) {
2437 bucket->ub_bucket[bucket->ub_cnt++] =
2438 slab_alloc_item(zone, slab);
2439 }
2440
2441 /* Don't block on the next fill */
2442 flags |= M_NOWAIT;
2443 }
2444 if (slab)
2445 zone_relock(zone, keg);
2446
2447 /*
2448 * We unlock here because we need to call the zone's init.
2449 * It should be safe to unlock because the slab dealt with
2450 * above is already on the appropriate list within the keg
2451 * and the bucket we filled is not yet on any list, so we
2452 * own it.
2453 */
2454 if (zone->uz_init != NULL) {
2455 int i;
2456
2457 ZONE_UNLOCK(zone);
2458 for (i = saved; i < bucket->ub_cnt; i++)
2459 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2460 origflags) != 0)
2461 break;
2462 /*
2463 * If we couldn't initialize the whole bucket, put the
2464 * rest back onto the freelist.
2465 */
2466 if (i != bucket->ub_cnt) {
2467 int j;
2468
2469 for (j = i; j < bucket->ub_cnt; j++) {
2470 zone_free_item(zone, bucket->ub_bucket[j],
2471 NULL, SKIP_FINI, 0);
2472#ifdef INVARIANTS
2473 bucket->ub_bucket[j] = NULL;
2474#endif
2475 }
2476 bucket->ub_cnt = i;
2477 }
2478 ZONE_LOCK(zone);
2479 }
2480
2481 zone->uz_fills--;
2482 if (bucket->ub_cnt != 0) {
2483 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2484 bucket, ub_link);
2485 return (1);
2486 }
2487#ifdef SMP
2488done:
2489#endif
2490 bucket_free(bucket);
2491
2492 return (0);
2493}
2494/*
2495 * Allocates an item for an internal zone
2496 *
2497 * Arguments
2498 * zone The zone to alloc for.
2499 * udata The data to be passed to the constructor.
2500 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2501 *
2502 * Returns
2503 * NULL if there is no memory and M_NOWAIT is set
2504 * An item if successful
2505 */
2506
2507static void *
2508zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2509{
2510 uma_slab_t slab;
2511 void *item;
2512
2513 item = NULL;
2514
2515#ifdef UMA_DEBUG_ALLOC
2516 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2517#endif
2518 ZONE_LOCK(zone);
2519
2520 slab = zone->uz_slab(zone, NULL, flags);
2521 if (slab == NULL) {
2522 zone->uz_fails++;
2523 ZONE_UNLOCK(zone);
2524 return (NULL);
2525 }
2526
2527 item = slab_alloc_item(zone, slab);
2528
2529 zone_relock(zone, slab->us_keg);
2530 zone->uz_allocs++;
2531 ZONE_UNLOCK(zone);
2532
2533 /*
2534 * We have to call both the zone's init (not the keg's init)
2535 * and the zone's ctor. This is because the item is going from
2536 * a keg slab directly to the user, and the user is expecting it
2537 * to be both zone-init'd as well as zone-ctor'd.
2538 */
2539 if (zone->uz_init != NULL) {
2540 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2541 zone_free_item(zone, item, udata, SKIP_FINI,
2542 ZFREE_STATFAIL | ZFREE_STATFREE);
2543 return (NULL);
2544 }
2545 }
2546 if (zone->uz_ctor != NULL) {
2547 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2548 zone_free_item(zone, item, udata, SKIP_DTOR,
2549 ZFREE_STATFAIL | ZFREE_STATFREE);
2550 return (NULL);
2551 }
2552 }
2553 if (flags & M_ZERO)
2554 bzero(item, zone->uz_size);
2555
2556 return (item);
2557}
2558
2559/* See uma.h */
2560void
2561uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2562{
2563 uma_cache_t cache;
2564 uma_bucket_t bucket;
2565 int bflags;
2566 int cpu;
2567
2568#ifdef UMA_DEBUG_ALLOC_1
2569 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2570#endif
2571 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2572 zone->uz_name);
2573
2574 /* uma_zfree(..., NULL) does nothing, to match free(9). */
2575 if (item == NULL)
2576 return;
2577#ifdef DEBUG_MEMGUARD
2578 if (is_memguard_addr(item)) {
2579 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2580 zone->uz_dtor(item, zone->uz_size, udata);
2581 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2582 zone->uz_fini(item, zone->uz_size);
2583 memguard_free(item);
2584 return;
2585 }
2586#endif
2587 if (zone->uz_dtor)
2588 zone->uz_dtor(item, zone->uz_size, udata);
2589
2590#ifdef INVARIANTS
2591 ZONE_LOCK(zone);
2592 if (zone->uz_flags & UMA_ZONE_MALLOC)
2593 uma_dbg_free(zone, udata, item);
2594 else
2595 uma_dbg_free(zone, NULL, item);
2596 ZONE_UNLOCK(zone);
2597#endif
2598 /*
2599 * The race here is acceptable. If we miss it we'll just have to wait
2600 * a little longer for the limits to be reset.
2601 */
2602 if (zone->uz_flags & UMA_ZFLAG_FULL)
2603 goto zfree_internal;
2604
2605 /*
2606 * If possible, free to the per-CPU cache. There are two
2607 * requirements for safe access to the per-CPU cache: (1) the thread
2608 * accessing the cache must not be preempted or yield during access,
2609 * and (2) the thread must not migrate CPUs without switching which
2610 * cache it accesses. We rely on a critical section to prevent
2611 * preemption and migration. We release the critical section in
2612 * order to acquire the zone mutex if we are unable to free to the
2613 * current cache; when we re-acquire the critical section, we must
2614 * detect and handle migration if it has occurred.
2615 */
2616zfree_restart:
2617 critical_enter();
2618 cpu = curcpu;
2619 cache = &zone->uz_cpu[cpu];
2620
2621zfree_start:
2622 bucket = cache->uc_freebucket;
2623
2624 if (bucket) {
2625 /*
2626 * Do we have room in our bucket? It is OK for this uz count
2627 * check to be slightly out of sync.
2628 */
2629
2630 if (bucket->ub_cnt < bucket->ub_entries) {
2631 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2632 ("uma_zfree: Freeing to non free bucket index."));
2633 bucket->ub_bucket[bucket->ub_cnt] = item;
2634 bucket->ub_cnt++;
2635 cache->uc_frees++;
2636 critical_exit();
2637 return;
2638 } else if (cache->uc_allocbucket) {
2639#ifdef UMA_DEBUG_ALLOC
2640 printf("uma_zfree: Swapping buckets.\n");
2641#endif
2642 /*
2643 * We have run out of space in our freebucket.
2644 * See if we can switch with our alloc bucket.
2645 */
2646 if (cache->uc_allocbucket->ub_cnt <
2647 cache->uc_freebucket->ub_cnt) {
2648 bucket = cache->uc_freebucket;
2649 cache->uc_freebucket = cache->uc_allocbucket;
2650 cache->uc_allocbucket = bucket;
2651 goto zfree_start;
2652 }
2653 }
2654 }
2655 /*
2656 * We can get here for two reasons:
2657 *
2658 * 1) The buckets are NULL
2659 * 2) The alloc and free buckets are both somewhat full.
2660 *
2661 * We must go back the zone, which requires acquiring the zone lock,
2662 * which in turn means we must release and re-acquire the critical
2663 * section. Since the critical section is released, we may be
2664 * preempted or migrate. As such, make sure not to maintain any
2665 * thread-local state specific to the cache from prior to releasing
2666 * the critical section.
2667 */
2668 critical_exit();
2669 ZONE_LOCK(zone);
2670 critical_enter();
2671 cpu = curcpu;
2672 cache = &zone->uz_cpu[cpu];
2673 if (cache->uc_freebucket != NULL) {
2674 if (cache->uc_freebucket->ub_cnt <
2675 cache->uc_freebucket->ub_entries) {
2676 ZONE_UNLOCK(zone);
2677 goto zfree_start;
2678 }
2679 if (cache->uc_allocbucket != NULL &&
2680 (cache->uc_allocbucket->ub_cnt <
2681 cache->uc_freebucket->ub_cnt)) {
2682 ZONE_UNLOCK(zone);
2683 goto zfree_start;
2684 }
2685 }
2686
2687 /* Since we have locked the zone we may as well send back our stats */
2688 zone->uz_allocs += cache->uc_allocs;
2689 cache->uc_allocs = 0;
2690 zone->uz_frees += cache->uc_frees;
2691 cache->uc_frees = 0;
2692
2693 bucket = cache->uc_freebucket;
2694 cache->uc_freebucket = NULL;
2695
2696 /* Can we throw this on the zone full list? */
2697 if (bucket != NULL) {
2698#ifdef UMA_DEBUG_ALLOC
2699 printf("uma_zfree: Putting old bucket on the free list.\n");
2700#endif
2701 /* ub_cnt is pointing to the last free item */
2702 KASSERT(bucket->ub_cnt != 0,
2703 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2704 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2705 bucket, ub_link);
2706 }
2707 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2708 LIST_REMOVE(bucket, ub_link);
2709 ZONE_UNLOCK(zone);
2710 cache->uc_freebucket = bucket;
2711 goto zfree_start;
2712 }
2713 /* We are no longer associated with this CPU. */
2714 critical_exit();
2715
2716 /* And the zone.. */
2717 ZONE_UNLOCK(zone);
2718
2719#ifdef UMA_DEBUG_ALLOC
2720 printf("uma_zfree: Allocating new free bucket.\n");
2721#endif
2722 bflags = M_NOWAIT;
2723
2724 if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2725 bflags |= M_NOVM;
2726 bucket = bucket_alloc(zone->uz_count, bflags);
2727 if (bucket) {
2728 ZONE_LOCK(zone);
2729 LIST_INSERT_HEAD(&zone->uz_free_bucket,
2730 bucket, ub_link);
2731 ZONE_UNLOCK(zone);
2732 goto zfree_restart;
2733 }
2734
2735 /*
2736 * If nothing else caught this, we'll just do an internal free.
2737 */
2738zfree_internal:
2739 zone_free_item(zone, item, udata, SKIP_DTOR, ZFREE_STATFREE);
2740
2741 return;
2742}
2743
2744/*
2745 * Frees an item to an INTERNAL zone or allocates a free bucket
2746 *
2747 * Arguments:
2748 * zone The zone to free to
2749 * item The item we're freeing
2750 * udata User supplied data for the dtor
2751 * skip Skip dtors and finis
2752 */
2753static void
2754zone_free_item(uma_zone_t zone, void *item, void *udata,
2755 enum zfreeskip skip, int flags)
2756{
2757 uma_slab_t slab;
2758 uma_slabrefcnt_t slabref;
2759 uma_keg_t keg;
2760 u_int8_t *mem;
2761 u_int8_t freei;
2762 int clearfull;
2763
2764 if (skip < SKIP_DTOR && zone->uz_dtor)
2765 zone->uz_dtor(item, zone->uz_size, udata);
2766
2767 if (skip < SKIP_FINI && zone->uz_fini)
2768 zone->uz_fini(item, zone->uz_size);
2769
2770 ZONE_LOCK(zone);
2771
2772 if (flags & ZFREE_STATFAIL)
2773 zone->uz_fails++;
2774 if (flags & ZFREE_STATFREE)
2775 zone->uz_frees++;
2776
2777 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2778 mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK));
2779 keg = zone_first_keg(zone); /* Must only be one. */
2780 if (zone->uz_flags & UMA_ZONE_HASH) {
2781 slab = hash_sfind(&keg->uk_hash, mem);
2782 } else {
2783 mem += keg->uk_pgoff;
2784 slab = (uma_slab_t)mem;
2785 }
2786 } else {
2787 /* This prevents redundant lookups via free(). */
2788 if ((zone->uz_flags & UMA_ZONE_MALLOC) && udata != NULL)
2789 slab = (uma_slab_t)udata;
2790 else
2791 slab = vtoslab((vm_offset_t)item);
2792 keg = slab->us_keg;
2793 keg_relock(keg, zone);
2794 }
2795 MPASS(keg == slab->us_keg);
2796
2797 /* Do we need to remove from any lists? */
2798 if (slab->us_freecount+1 == keg->uk_ipers) {
2799 LIST_REMOVE(slab, us_link);
2800 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2801 } else if (slab->us_freecount == 0) {
2802 LIST_REMOVE(slab, us_link);
2803 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2804 }
2805
2806 /* Slab management stuff */
2807 freei = ((unsigned long)item - (unsigned long)slab->us_data)
2808 / keg->uk_rsize;
2809
2810#ifdef INVARIANTS
2811 if (!skip)
2812 uma_dbg_free(zone, slab, item);
2813#endif
2814
2815 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2816 slabref = (uma_slabrefcnt_t)slab;
2817 slabref->us_freelist[freei].us_item = slab->us_firstfree;
2818 } else {
2819 slab->us_freelist[freei].us_item = slab->us_firstfree;
2820 }
2821 slab->us_firstfree = freei;
2822 slab->us_freecount++;
2823
2824 /* Zone statistics */
2825 keg->uk_free++;
2826
2827 clearfull = 0;
2828 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2829 if (keg->uk_pages < keg->uk_maxpages) {
2830 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2831 clearfull = 1;
2832 }
2833
2834 /*
2835 * We can handle one more allocation. Since we're clearing ZFLAG_FULL,
2836 * wake up all procs blocked on pages. This should be uncommon, so
2837 * keeping this simple for now (rather than adding count of blocked
2838 * threads etc).
2839 */
2840 wakeup(keg);
2841 }
2842 if (clearfull) {
2843 zone_relock(zone, keg);
2844 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2845 wakeup(zone);
2846 ZONE_UNLOCK(zone);
2847 } else
2848 KEG_UNLOCK(keg);
2849}
2850
2851/* See uma.h */
2852int
2853uma_zone_set_max(uma_zone_t zone, int nitems)
2854{
2855 uma_keg_t keg;
2856
2857 ZONE_LOCK(zone);
2858 keg = zone_first_keg(zone);
2859 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2860 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2861 keg->uk_maxpages += keg->uk_ppera;
2862 nitems = keg->uk_maxpages * keg->uk_ipers;
2863 ZONE_UNLOCK(zone);
2864
2865 return (nitems);
2866}
2867
2868/* See uma.h */
2869int
2870uma_zone_get_max(uma_zone_t zone)
2871{
2872 int nitems;
2873 uma_keg_t keg;
2874
2875 ZONE_LOCK(zone);
2876 keg = zone_first_keg(zone);
2877 nitems = keg->uk_maxpages * keg->uk_ipers;
2878 ZONE_UNLOCK(zone);
2879
2880 return (nitems);
2881}
2882
2883/* See uma.h */
2884int
2885uma_zone_get_cur(uma_zone_t zone)
2886{
2887 int64_t nitems;
2888 u_int i;
2889
2890 ZONE_LOCK(zone);
2891 nitems = zone->uz_allocs - zone->uz_frees;
2892 CPU_FOREACH(i) {
2893 /*
2894 * See the comment in sysctl_vm_zone_stats() regarding the
2895 * safety of accessing the per-cpu caches. With the zone lock
2896 * held, it is safe, but can potentially result in stale data.
2897 */
2898 nitems += zone->uz_cpu[i].uc_allocs -
2899 zone->uz_cpu[i].uc_frees;
2900 }
2901 ZONE_UNLOCK(zone);
2902
2903 return (nitems < 0 ? 0 : nitems);
2904}
2905
2906/* See uma.h */
2907void
2908uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2909{
2910 uma_keg_t keg;
2911
2912 ZONE_LOCK(zone);
2913 keg = zone_first_keg(zone);
2914 KASSERT(keg->uk_pages == 0,
2915 ("uma_zone_set_init on non-empty keg"));
2916 keg->uk_init = uminit;
2917 ZONE_UNLOCK(zone);
2918}
2919
2920/* See uma.h */
2921void
2922uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2923{
2924 uma_keg_t keg;
2925
2926 ZONE_LOCK(zone);
2927 keg = zone_first_keg(zone);
2928 KASSERT(keg->uk_pages == 0,
2929 ("uma_zone_set_fini on non-empty keg"));
2930 keg->uk_fini = fini;
2931 ZONE_UNLOCK(zone);
2932}
2933
2934/* See uma.h */
2935void
2936uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2937{
2938 ZONE_LOCK(zone);
2939 KASSERT(zone_first_keg(zone)->uk_pages == 0,
2940 ("uma_zone_set_zinit on non-empty keg"));
2941 zone->uz_init = zinit;
2942 ZONE_UNLOCK(zone);
2943}
2944
2945/* See uma.h */
2946void
2947uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2948{
2949 ZONE_LOCK(zone);
2950 KASSERT(zone_first_keg(zone)->uk_pages == 0,
2951 ("uma_zone_set_zfini on non-empty keg"));
2952 zone->uz_fini = zfini;
2953 ZONE_UNLOCK(zone);
2954}
2955
2956/* See uma.h */
2957/* XXX uk_freef is not actually used with the zone locked */
2958void
2959uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2960{
2961
2962 ZONE_LOCK(zone);
2963 zone_first_keg(zone)->uk_freef = freef;
2964 ZONE_UNLOCK(zone);
2965}
2966
2967/* See uma.h */
2968/* XXX uk_allocf is not actually used with the zone locked */
2969void
2970uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2971{
2972 uma_keg_t keg;
2973
2974 ZONE_LOCK(zone);
2975 keg = zone_first_keg(zone);
2976 keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2977 keg->uk_allocf = allocf;
2978 ZONE_UNLOCK(zone);
2979}
2980
2981/* See uma.h */
2982int
2983uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count)
2984{
2985 uma_keg_t keg;
2986 vm_offset_t kva;
2987 int pages;
2988
2989 keg = zone_first_keg(zone);
2990 pages = count / keg->uk_ipers;
2991
2992 if (pages * keg->uk_ipers < count)
2993 pages++;
2994
2995 kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2996
2997 if (kva == 0)
2998 return (0);
2999 if (obj == NULL)
3000 obj = vm_object_allocate(OBJT_PHYS, pages);
3001 else {
3002 VM_OBJECT_LOCK_INIT(obj, "uma object");
3003 _vm_object_allocate(OBJT_PHYS, pages, obj);
3004 }
3005 ZONE_LOCK(zone);
3006 keg->uk_kva = kva;
3007 keg->uk_obj = obj;
3008 keg->uk_maxpages = pages;
3009 keg->uk_allocf = obj_alloc;
3010 keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
3011 ZONE_UNLOCK(zone);
3012 return (1);
3013}
3014
3015/* See uma.h */
3016void
3017uma_prealloc(uma_zone_t zone, int items)
3018{
3019 int slabs;
3020 uma_slab_t slab;
3021 uma_keg_t keg;
3022
3023 keg = zone_first_keg(zone);
3024 ZONE_LOCK(zone);
3025 slabs = items / keg->uk_ipers;
3026 if (slabs * keg->uk_ipers < items)
3027 slabs++;
3028 while (slabs > 0) {
3029 slab = keg_alloc_slab(keg, zone, M_WAITOK);
3030 if (slab == NULL)
3031 break;
3032 MPASS(slab->us_keg == keg);
3033 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3034 slabs--;
3035 }
3036 ZONE_UNLOCK(zone);
3037}
3038
3039/* See uma.h */
3040u_int32_t *
3041uma_find_refcnt(uma_zone_t zone, void *item)
3042{
3043 uma_slabrefcnt_t slabref;
3044 uma_keg_t keg;
3045 u_int32_t *refcnt;
3046 int idx;
3047
3048 slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item &
3049 (~UMA_SLAB_MASK));
3050 keg = slabref->us_keg;
3051 KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT,
3052 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3053 idx = ((unsigned long)item - (unsigned long)slabref->us_data)
3054 / keg->uk_rsize;
3055 refcnt = &slabref->us_freelist[idx].us_refcnt;
3056 return refcnt;
3057}
3058
3059/* See uma.h */
3060void
3061uma_reclaim(void)
3062{
3063#ifdef UMA_DEBUG
3064 printf("UMA: vm asked us to release pages!\n");
3065#endif
3066 bucket_enable();
3067 zone_foreach(zone_drain);
3068 /*
3069 * Some slabs may have been freed but this zone will be visited early
3070 * we visit again so that we can free pages that are empty once other
3071 * zones are drained. We have to do the same for buckets.
3072 */
3073 zone_drain(slabzone);
3074 zone_drain(slabrefzone);
3075 bucket_zone_drain();
3076}
3077
3078/* See uma.h */
3079int
3080uma_zone_exhausted(uma_zone_t zone)
3081{
3082 int full;
3083
3084 ZONE_LOCK(zone);
3085 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3086 ZONE_UNLOCK(zone);
3087 return (full);
3088}
3089
3090int
3091uma_zone_exhausted_nolock(uma_zone_t zone)
3092{
3093 return (zone->uz_flags & UMA_ZFLAG_FULL);
3094}
3095
3096void *
3097uma_large_malloc(int size, int wait)
3098{
3099 void *mem;
3100 uma_slab_t slab;
3101 u_int8_t flags;
3102
3103 slab = zone_alloc_item(slabzone, NULL, wait);
3104 if (slab == NULL)
3105 return (NULL);
3106 mem = page_alloc(NULL, size, &flags, wait);
3107 if (mem) {
3108 vsetslab((vm_offset_t)mem, slab);
3109 slab->us_data = mem;
3110 slab->us_flags = flags | UMA_SLAB_MALLOC;
3111 slab->us_size = size;
3112 } else {
3113 zone_free_item(slabzone, slab, NULL, SKIP_NONE,
3114 ZFREE_STATFAIL | ZFREE_STATFREE);
3115 }
3116
3117 return (mem);
3118}
3119
3120void
3121uma_large_free(uma_slab_t slab)
3122{
3123 vsetobj((vm_offset_t)slab->us_data, kmem_object);
3124 page_free(slab->us_data, slab->us_size, slab->us_flags);
3125 zone_free_item(slabzone, slab, NULL, SKIP_NONE, ZFREE_STATFREE);
3126}
3127
3128void
3129uma_print_stats(void)
3130{
3131 zone_foreach(uma_print_zone);
3132}
3133
3134static void
3135slab_print(uma_slab_t slab)
3136{
3137 printf("slab: keg %p, data %p, freecount %d, firstfree %d\n",
3138 slab->us_keg, slab->us_data, slab->us_freecount,
3139 slab->us_firstfree);
3140}
3141
3142static void
3143cache_print(uma_cache_t cache)
3144{
3145 printf("alloc: %p(%d), free: %p(%d)\n",
3146 cache->uc_allocbucket,
3147 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3148 cache->uc_freebucket,
3149 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3150}
3151
3152static void
3153uma_print_keg(uma_keg_t keg)
3154{
3155 uma_slab_t slab;
3156
3157 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3158 "out %d free %d limit %d\n",
3159 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3160 keg->uk_ipers, keg->uk_ppera,
3161 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3162 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3163 printf("Part slabs:\n");
3164 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3165 slab_print(slab);
3166 printf("Free slabs:\n");
3167 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3168 slab_print(slab);
3169 printf("Full slabs:\n");
3170 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3171 slab_print(slab);
3172}
3173
3174void
3175uma_print_zone(uma_zone_t zone)
3176{
3177 uma_cache_t cache;
3178 uma_klink_t kl;
3179 int i;
3180
3181 printf("zone: %s(%p) size %d flags %#x\n",
3182 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3183 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3184 uma_print_keg(kl->kl_keg);
3185 CPU_FOREACH(i) {
3186 cache = &zone->uz_cpu[i];
3187 printf("CPU %d Cache:\n", i);
3188 cache_print(cache);
3189 }
3190}
3191
3192#ifdef DDB
3193/*
3194 * Generate statistics across both the zone and its per-cpu cache's. Return
3195 * desired statistics if the pointer is non-NULL for that statistic.
3196 *
3197 * Note: does not update the zone statistics, as it can't safely clear the
3198 * per-CPU cache statistic.
3199 *
3200 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3201 * safe from off-CPU; we should modify the caches to track this information
3202 * directly so that we don't have to.
3203 */
3204static void
3205uma_zone_sumstat(uma_zone_t z, int *cachefreep, u_int64_t *allocsp,
3206 u_int64_t *freesp, u_int64_t *sleepsp)
3207{
3208 uma_cache_t cache;
3209 u_int64_t allocs, frees, sleeps;
3210 int cachefree, cpu;
3211
3212 allocs = frees = sleeps = 0;
3213 cachefree = 0;
3214 CPU_FOREACH(cpu) {
3215 cache = &z->uz_cpu[cpu];
3216 if (cache->uc_allocbucket != NULL)
3217 cachefree += cache->uc_allocbucket->ub_cnt;
3218 if (cache->uc_freebucket != NULL)
3219 cachefree += cache->uc_freebucket->ub_cnt;
3220 allocs += cache->uc_allocs;
3221 frees += cache->uc_frees;
3222 }
3223 allocs += z->uz_allocs;
3224 frees += z->uz_frees;
3225 sleeps += z->uz_sleeps;
3226 if (cachefreep != NULL)
3227 *cachefreep = cachefree;
3228 if (allocsp != NULL)
3229 *allocsp = allocs;
3230 if (freesp != NULL)
3231 *freesp = frees;
3232 if (sleepsp != NULL)
3233 *sleepsp = sleeps;
3234}
3235#endif /* DDB */
3236
3237static int
3238sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3239{
3240 uma_keg_t kz;
3241 uma_zone_t z;
3242 int count;
3243
3244 count = 0;
3245 mtx_lock(&uma_mtx);
3246 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3247 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3248 count++;
3249 }
3250 mtx_unlock(&uma_mtx);
3251 return (sysctl_handle_int(oidp, &count, 0, req));
3252}
3253
3254static int
3255sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3256{
3257 struct uma_stream_header ush;
3258 struct uma_type_header uth;
3259 struct uma_percpu_stat ups;
3260 uma_bucket_t bucket;
3261 struct sbuf sbuf;
3262 uma_cache_t cache;
3263 uma_klink_t kl;
3264 uma_keg_t kz;
3265 uma_zone_t z;
3266 uma_keg_t k;
3267 int count, error, i;
3268
3269 error = sysctl_wire_old_buffer(req, 0);
3270 if (error != 0)
3271 return (error);
3272 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3273
3274 count = 0;
3275 mtx_lock(&uma_mtx);
3276 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3277 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3278 count++;
3279 }
3280
3281 /*
3282 * Insert stream header.
3283 */
3284 bzero(&ush, sizeof(ush));
3285 ush.ush_version = UMA_STREAM_VERSION;
3286 ush.ush_maxcpus = (mp_maxid + 1);
3287 ush.ush_count = count;
3288 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3289
3290 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3291 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3292 bzero(&uth, sizeof(uth));
3293 ZONE_LOCK(z);
3294 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3295 uth.uth_align = kz->uk_align;
3296 uth.uth_size = kz->uk_size;
3297 uth.uth_rsize = kz->uk_rsize;
3298 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3299 k = kl->kl_keg;
3300 uth.uth_maxpages += k->uk_maxpages;
3301 uth.uth_pages += k->uk_pages;
3302 uth.uth_keg_free += k->uk_free;
3303 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3304 * k->uk_ipers;
3305 }
3306
3307 /*
3308 * A zone is secondary is it is not the first entry
3309 * on the keg's zone list.
3310 */
3311 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3312 (LIST_FIRST(&kz->uk_zones) != z))
3313 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3314
3315 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3316 uth.uth_zone_free += bucket->ub_cnt;
3317 uth.uth_allocs = z->uz_allocs;
3318 uth.uth_frees = z->uz_frees;
3319 uth.uth_fails = z->uz_fails;
3320 uth.uth_sleeps = z->uz_sleeps;
3321 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3322 /*
3323 * While it is not normally safe to access the cache
3324 * bucket pointers while not on the CPU that owns the
3325 * cache, we only allow the pointers to be exchanged
3326 * without the zone lock held, not invalidated, so
3327 * accept the possible race associated with bucket
3328 * exchange during monitoring.
3329 */
3330 for (i = 0; i < (mp_maxid + 1); i++) {
3331 bzero(&ups, sizeof(ups));
3332 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3333 goto skip;
3334 if (CPU_ABSENT(i))
3335 goto skip;
3336 cache = &z->uz_cpu[i];
3337 if (cache->uc_allocbucket != NULL)
3338 ups.ups_cache_free +=
3339 cache->uc_allocbucket->ub_cnt;
3340 if (cache->uc_freebucket != NULL)
3341 ups.ups_cache_free +=
3342 cache->uc_freebucket->ub_cnt;
3343 ups.ups_allocs = cache->uc_allocs;
3344 ups.ups_frees = cache->uc_frees;
3345skip:
3346 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3347 }
3348 ZONE_UNLOCK(z);
3349 }
3350 }
3351 mtx_unlock(&uma_mtx);
3352 error = sbuf_finish(&sbuf);
3353 sbuf_delete(&sbuf);
3354 return (error);
3355}
3356
3357#ifdef DDB
3358DB_SHOW_COMMAND(uma, db_show_uma)
3359{
3360 u_int64_t allocs, frees, sleeps;
3361 uma_bucket_t bucket;
3362 uma_keg_t kz;
3363 uma_zone_t z;
3364 int cachefree;
3365
3366 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3367 "Requests", "Sleeps");
3368 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3369 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3370 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3371 allocs = z->uz_allocs;
3372 frees = z->uz_frees;
3373 sleeps = z->uz_sleeps;
3374 cachefree = 0;
3375 } else
3376 uma_zone_sumstat(z, &cachefree, &allocs,
3377 &frees, &sleeps);
3378 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3379 (LIST_FIRST(&kz->uk_zones) != z)))
3380 cachefree += kz->uk_free;
3381 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3382 cachefree += bucket->ub_cnt;
3383 db_printf("%18s %8ju %8jd %8d %12ju %8ju\n", z->uz_name,
3384 (uintmax_t)kz->uk_size,
3385 (intmax_t)(allocs - frees), cachefree,
3386 (uintmax_t)allocs, sleeps);
3387 if (db_pager_quit)
3388 return;
3389 }
3390 }
3391}
3392#endif