1/*-
2 * Copyright (c) 2004-2005 Robert N. M. Watson
3 * Copyright (c) 2004, 2005,
4 * Bosko Milekic <bmilekic@FreeBSD.org>. All rights reserved.
5 * Copyright (c) 2002, 2003, 2004, 2005,
6 * Jeffrey Roberson <jeff@FreeBSD.org>. All rights reserved.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice unmodified, this list of conditions, and the following
13 * disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 *
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
23 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
24 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
27 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28 */
29
30/*
31 * uma_core.c Implementation of the Universal Memory allocator
32 *
33 * This allocator is intended to replace the multitude of similar object caches
34 * in the standard FreeBSD kernel. The intent is to be flexible as well as
35 * effecient. A primary design goal is to return unused memory to the rest of
36 * the system. This will make the system as a whole more flexible due to the
37 * ability to move memory to subsystems which most need it instead of leaving
38 * pools of reserved memory unused.
39 *
40 * The basic ideas stem from similar slab/zone based allocators whose algorithms
41 * are well known.
42 *
43 */
44
45/*
46 * TODO:
47 * - Improve memory usage for large allocations
48 * - Investigate cache size adjustments
49 */
50
51#include <sys/cdefs.h>
52__FBSDID("$FreeBSD: head/sys/vm/uma_core.c 147995 2005-07-14 16:17:21Z rwatson $");
53
54/* I should really use ktr.. */
55/*
56#define UMA_DEBUG 1
57#define UMA_DEBUG_ALLOC 1
58#define UMA_DEBUG_ALLOC_1 1
59*/
60
61#include "opt_param.h"
62#include <sys/param.h>
63#include <sys/systm.h>
64#include <sys/kernel.h>
65#include <sys/types.h>
66#include <sys/queue.h>
67#include <sys/malloc.h>
68#include <sys/ktr.h>
69#include <sys/lock.h>
70#include <sys/sysctl.h>
71#include <sys/mutex.h>
72#include <sys/proc.h>
73#include <sys/smp.h>
74#include <sys/vmmeter.h>
75
76#include <vm/vm.h>
77#include <vm/vm_object.h>
78#include <vm/vm_page.h>
79#include <vm/vm_param.h>
80#include <vm/vm_map.h>
81#include <vm/vm_kern.h>
82#include <vm/vm_extern.h>
83#include <vm/uma.h>
84#include <vm/uma_int.h>
85#include <vm/uma_dbg.h>
86
87#include <machine/vmparam.h>
88
89/*
90 * This is the zone and keg from which all zones are spawned. The idea is that
91 * even the zone & keg heads are allocated from the allocator, so we use the
92 * bss section to bootstrap us.
93 */
94static struct uma_keg masterkeg;
95static struct uma_zone masterzone_k;
96static struct uma_zone masterzone_z;
97static uma_zone_t kegs = &masterzone_k;
98static uma_zone_t zones = &masterzone_z;
99
100/* This is the zone from which all of uma_slab_t's are allocated. */
101static uma_zone_t slabzone;
102static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
103
104/*
105 * The initial hash tables come out of this zone so they can be allocated
106 * prior to malloc coming up.
107 */
108static uma_zone_t hashzone;
109
110static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
111
112/*
113 * Are we allowed to allocate buckets?
114 */
115static int bucketdisable = 1;
116
117/* Linked list of all kegs in the system */
118static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(&uma_kegs);
119
120/* This mutex protects the keg list */
121static struct mtx uma_mtx;
122
123/* Linked list of boot time pages */
124static LIST_HEAD(,uma_slab) uma_boot_pages =
125 LIST_HEAD_INITIALIZER(&uma_boot_pages);
126
127/* Count of free boottime pages */
128static int uma_boot_free = 0;
129
130/* Is the VM done starting up? */
131static int booted = 0;
132
133/* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
134static u_int uma_max_ipers;
135static u_int uma_max_ipers_ref;
136
137/*
138 * This is the handle used to schedule events that need to happen
139 * outside of the allocation fast path.
140 */
141static struct callout uma_callout;
142#define UMA_TIMEOUT 20 /* Seconds for callout interval. */
143
144/*
145 * This structure is passed as the zone ctor arg so that I don't have to create
146 * a special allocation function just for zones.
147 */
148struct uma_zctor_args {
149 char *name;
150 size_t size;
151 uma_ctor ctor;
152 uma_dtor dtor;
153 uma_init uminit;
154 uma_fini fini;
155 uma_keg_t keg;
156 int align;
157 u_int16_t flags;
158};
159
160struct uma_kctor_args {
161 uma_zone_t zone;
162 size_t size;
163 uma_init uminit;
164 uma_fini fini;
165 int align;
166 u_int16_t flags;
167};
168
169struct uma_bucket_zone {
170 uma_zone_t ubz_zone;
171 char *ubz_name;
172 int ubz_entries;
173};
174
175#define BUCKET_MAX 128
176
177struct uma_bucket_zone bucket_zones[] = {
178 { NULL, "16 Bucket", 16 },
179 { NULL, "32 Bucket", 32 },
180 { NULL, "64 Bucket", 64 },
181 { NULL, "128 Bucket", 128 },
182 { NULL, NULL, 0}
183};
184
185#define BUCKET_SHIFT 4
186#define BUCKET_ZONES ((BUCKET_MAX >> BUCKET_SHIFT) + 1)
187
188/*
189 * bucket_size[] maps requested bucket sizes to zones that allocate a bucket
190 * of approximately the right size.
191 */
192static uint8_t bucket_size[BUCKET_ZONES];
193
194enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI };
195
196/* Prototypes.. */
197
198static void *obj_alloc(uma_zone_t, int, u_int8_t *, int);
199static void *page_alloc(uma_zone_t, int, u_int8_t *, int);
200static void *startup_alloc(uma_zone_t, int, u_int8_t *, int);
201static void page_free(void *, int, u_int8_t);
202static uma_slab_t slab_zalloc(uma_zone_t, int);
203static void cache_drain(uma_zone_t);
204static void bucket_drain(uma_zone_t, uma_bucket_t);
205static void bucket_cache_drain(uma_zone_t zone);
206static int keg_ctor(void *, int, void *, int);
207static void keg_dtor(void *, int, void *);
208static int zone_ctor(void *, int, void *, int);
209static void zone_dtor(void *, int, void *);
210static int zero_init(void *, int, int);
211static void zone_small_init(uma_zone_t zone);
212static void zone_large_init(uma_zone_t zone);
213static void zone_foreach(void (*zfunc)(uma_zone_t));
214static void zone_timeout(uma_zone_t zone);
215static int hash_alloc(struct uma_hash *);
216static int hash_expand(struct uma_hash *, struct uma_hash *);
217static void hash_free(struct uma_hash *hash);
218static void uma_timeout(void *);
219static void uma_startup3(void);
220static void *uma_zalloc_internal(uma_zone_t, void *, int);
221static void uma_zfree_internal(uma_zone_t, void *, void *, enum zfreeskip);
222static void bucket_enable(void);
223static void bucket_init(void);
224static uma_bucket_t bucket_alloc(int, int);
225static void bucket_free(uma_bucket_t);
226static void bucket_zone_drain(void);
227static int uma_zalloc_bucket(uma_zone_t zone, int flags);
228static uma_slab_t uma_zone_slab(uma_zone_t zone, int flags);
229static void *uma_slab_alloc(uma_zone_t zone, uma_slab_t slab);
230static void zone_drain(uma_zone_t);
231static uma_zone_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
232 uma_fini fini, int align, u_int16_t flags);
233
234void uma_print_zone(uma_zone_t);
235void uma_print_stats(void);
236static int sysctl_vm_zone(SYSCTL_HANDLER_ARGS);
237
238#ifdef WITNESS
239static int nosleepwithlocks = 1;
240SYSCTL_INT(_debug, OID_AUTO, nosleepwithlocks, CTLFLAG_RW, &nosleepwithlocks,
241 0, "Convert M_WAITOK to M_NOWAIT to avoid lock-held-across-sleep paths");
242#else
243static int nosleepwithlocks = 0;
244SYSCTL_INT(_debug, OID_AUTO, nosleepwithlocks, CTLFLAG_RW, &nosleepwithlocks,
245 0, "Convert M_WAITOK to M_NOWAIT to avoid lock-held-across-sleep paths");
246#endif
247SYSCTL_OID(_vm, OID_AUTO, zone, CTLTYPE_STRING|CTLFLAG_RD,
248 NULL, 0, sysctl_vm_zone, "A", "Zone Info");
249SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
250
251/*
252 * This routine checks to see whether or not it's safe to enable buckets.
253 */
254
255static void
256bucket_enable(void)
257{
258 if (cnt.v_free_count < cnt.v_free_min)
259 bucketdisable = 1;
260 else
261 bucketdisable = 0;
262}
263
264/*
265 * Initialize bucket_zones, the array of zones of buckets of various sizes.
266 *
267 * For each zone, calculate the memory required for each bucket, consisting
268 * of the header and an array of pointers. Initialize bucket_size[] to point
269 * the range of appropriate bucket sizes at the zone.
270 */
271static void
272bucket_init(void)
273{
274 struct uma_bucket_zone *ubz;
275 int i;
276 int j;
277
278 for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) {
279 int size;
280
281 ubz = &bucket_zones[j];
282 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
283 size += sizeof(void *) * ubz->ubz_entries;
284 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
285 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
286 for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT))
287 bucket_size[i >> BUCKET_SHIFT] = j;
288 }
289}
290
291/*
292 * Given a desired number of entries for a bucket, return the zone from which
293 * to allocate the bucket.
294 */
295static struct uma_bucket_zone *
296bucket_zone_lookup(int entries)
297{
298 int idx;
299
300 idx = howmany(entries, 1 << BUCKET_SHIFT);
301 return (&bucket_zones[bucket_size[idx]]);
302}
303
304static uma_bucket_t
305bucket_alloc(int entries, int bflags)
306{
307 struct uma_bucket_zone *ubz;
308 uma_bucket_t bucket;
309
310 /*
311 * This is to stop us from allocating per cpu buckets while we're
312 * running out of UMA_BOOT_PAGES. Otherwise, we would exhaust the
313 * boot pages. This also prevents us from allocating buckets in
314 * low memory situations.
315 */
316 if (bucketdisable)
317 return (NULL);
318
319 ubz = bucket_zone_lookup(entries);
320 bucket = uma_zalloc_internal(ubz->ubz_zone, NULL, bflags);
321 if (bucket) {
322#ifdef INVARIANTS
323 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
324#endif
325 bucket->ub_cnt = 0;
326 bucket->ub_entries = ubz->ubz_entries;
327 }
328
329 return (bucket);
330}
331
332static void
333bucket_free(uma_bucket_t bucket)
334{
335 struct uma_bucket_zone *ubz;
336
337 ubz = bucket_zone_lookup(bucket->ub_entries);
338 uma_zfree_internal(ubz->ubz_zone, bucket, NULL, SKIP_NONE);
339}
340
341static void
342bucket_zone_drain(void)
343{
344 struct uma_bucket_zone *ubz;
345
346 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
347 zone_drain(ubz->ubz_zone);
348}
349
350
351/*
352 * Routine called by timeout which is used to fire off some time interval
353 * based calculations. (stats, hash size, etc.)
354 *
355 * Arguments:
356 * arg Unused
357 *
358 * Returns:
359 * Nothing
360 */
361static void
362uma_timeout(void *unused)
363{
364 bucket_enable();
365 zone_foreach(zone_timeout);
366
367 /* Reschedule this event */
368 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
369}
370
371/*
372 * Routine to perform timeout driven calculations. This expands the
373 * hashes and does per cpu statistics aggregation.
374 *
375 * Arguments:
376 * zone The zone to operate on
377 *
378 * Returns:
379 * Nothing
380 */
381static void
382zone_timeout(uma_zone_t zone)
383{
384 uma_keg_t keg;
385 u_int64_t alloc;
386
387 keg = zone->uz_keg;
388 alloc = 0;
389
390 /*
391 * Expand the zone hash table.
392 *
393 * This is done if the number of slabs is larger than the hash size.
394 * What I'm trying to do here is completely reduce collisions. This
395 * may be a little aggressive. Should I allow for two collisions max?
396 */
397 ZONE_LOCK(zone);
398 if (keg->uk_flags & UMA_ZONE_HASH &&
399 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
400 struct uma_hash newhash;
401 struct uma_hash oldhash;
402 int ret;
403
404 /*
405 * This is so involved because allocating and freeing
406 * while the zone lock is held will lead to deadlock.
407 * I have to do everything in stages and check for
408 * races.
409 */
410 newhash = keg->uk_hash;
411 ZONE_UNLOCK(zone);
412 ret = hash_alloc(&newhash);
413 ZONE_LOCK(zone);
414 if (ret) {
415 if (hash_expand(&keg->uk_hash, &newhash)) {
416 oldhash = keg->uk_hash;
417 keg->uk_hash = newhash;
418 } else
419 oldhash = newhash;
420
421 ZONE_UNLOCK(zone);
422 hash_free(&oldhash);
423 ZONE_LOCK(zone);
424 }
425 }
426 ZONE_UNLOCK(zone);
427}
428
429/*
430 * Allocate and zero fill the next sized hash table from the appropriate
431 * backing store.
432 *
433 * Arguments:
434 * hash A new hash structure with the old hash size in uh_hashsize
435 *
436 * Returns:
437 * 1 on sucess and 0 on failure.
438 */
439static int
440hash_alloc(struct uma_hash *hash)
441{
442 int oldsize;
443 int alloc;
444
445 oldsize = hash->uh_hashsize;
446
447 /* We're just going to go to a power of two greater */
448 if (oldsize) {
449 hash->uh_hashsize = oldsize * 2;
450 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
451 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
452 M_UMAHASH, M_NOWAIT);
453 } else {
454 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
455 hash->uh_slab_hash = uma_zalloc_internal(hashzone, NULL,
456 M_WAITOK);
457 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
458 }
459 if (hash->uh_slab_hash) {
460 bzero(hash->uh_slab_hash, alloc);
461 hash->uh_hashmask = hash->uh_hashsize - 1;
462 return (1);
463 }
464
465 return (0);
466}
467
468/*
469 * Expands the hash table for HASH zones. This is done from zone_timeout
470 * to reduce collisions. This must not be done in the regular allocation
471 * path, otherwise, we can recurse on the vm while allocating pages.
472 *
473 * Arguments:
474 * oldhash The hash you want to expand
475 * newhash The hash structure for the new table
476 *
477 * Returns:
478 * Nothing
479 *
480 * Discussion:
481 */
482static int
483hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
484{
485 uma_slab_t slab;
486 int hval;
487 int i;
488
489 if (!newhash->uh_slab_hash)
490 return (0);
491
492 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
493 return (0);
494
495 /*
496 * I need to investigate hash algorithms for resizing without a
497 * full rehash.
498 */
499
500 for (i = 0; i < oldhash->uh_hashsize; i++)
501 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
502 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
503 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
504 hval = UMA_HASH(newhash, slab->us_data);
505 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
506 slab, us_hlink);
507 }
508
509 return (1);
510}
511
512/*
513 * Free the hash bucket to the appropriate backing store.
514 *
515 * Arguments:
516 * slab_hash The hash bucket we're freeing
517 * hashsize The number of entries in that hash bucket
518 *
519 * Returns:
520 * Nothing
521 */
522static void
523hash_free(struct uma_hash *hash)
524{
525 if (hash->uh_slab_hash == NULL)
526 return;
527 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
528 uma_zfree_internal(hashzone,
529 hash->uh_slab_hash, NULL, SKIP_NONE);
530 else
531 free(hash->uh_slab_hash, M_UMAHASH);
532}
533
534/*
535 * Frees all outstanding items in a bucket
536 *
537 * Arguments:
538 * zone The zone to free to, must be unlocked.
539 * bucket The free/alloc bucket with items, cpu queue must be locked.
540 *
541 * Returns:
542 * Nothing
543 */
544
545static void
546bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
547{
548 uma_slab_t slab;
549 int mzone;
550 void *item;
551
552 if (bucket == NULL)
553 return;
554
555 slab = NULL;
556 mzone = 0;
557
558 /* We have to lookup the slab again for malloc.. */
559 if (zone->uz_keg->uk_flags & UMA_ZONE_MALLOC)
560 mzone = 1;
561
562 while (bucket->ub_cnt > 0) {
563 bucket->ub_cnt--;
564 item = bucket->ub_bucket[bucket->ub_cnt];
565#ifdef INVARIANTS
566 bucket->ub_bucket[bucket->ub_cnt] = NULL;
567 KASSERT(item != NULL,
568 ("bucket_drain: botched ptr, item is NULL"));
569#endif
570 /*
571 * This is extremely inefficient. The slab pointer was passed
572 * to uma_zfree_arg, but we lost it because the buckets don't
573 * hold them. This will go away when free() gets a size passed
574 * to it.
575 */
576 if (mzone)
577 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
578 uma_zfree_internal(zone, item, slab, SKIP_DTOR);
579 }
580}
581
582/*
583 * Drains the per cpu caches for a zone.
584 *
585 * NOTE: This may only be called while the zone is being turn down, and not
586 * during normal operation. This is necessary in order that we do not have
587 * to migrate CPUs to drain the per-CPU caches.
588 *
589 * Arguments:
590 * zone The zone to drain, must be unlocked.
591 *
592 * Returns:
593 * Nothing
594 */
595static void
596cache_drain(uma_zone_t zone)
597{
598 uma_cache_t cache;
599 int cpu;
600
601 /*
602 * XXX: It is safe to not lock the per-CPU caches, because we're
603 * tearing down the zone anyway. I.e., there will be no further use
604 * of the caches at this point.
605 *
606 * XXX: It would good to be able to assert that the zone is being
607 * torn down to prevent improper use of cache_drain().
608 *
609 * XXX: We lock the zone before passing into bucket_cache_drain() as
610 * it is used elsewhere. Should the tear-down path be made special
611 * there in some form?
612 */
613 for (cpu = 0; cpu <= mp_maxid; cpu++) {
614 if (CPU_ABSENT(cpu))
615 continue;
616 cache = &zone->uz_cpu[cpu];
617 bucket_drain(zone, cache->uc_allocbucket);
618 bucket_drain(zone, cache->uc_freebucket);
619 if (cache->uc_allocbucket != NULL)
620 bucket_free(cache->uc_allocbucket);
621 if (cache->uc_freebucket != NULL)
622 bucket_free(cache->uc_freebucket);
623 cache->uc_allocbucket = cache->uc_freebucket = NULL;
624 }
625 ZONE_LOCK(zone);
626 bucket_cache_drain(zone);
627 ZONE_UNLOCK(zone);
628}
629
630/*
631 * Drain the cached buckets from a zone. Expects a locked zone on entry.
632 */
633static void
634bucket_cache_drain(uma_zone_t zone)
635{
636 uma_bucket_t bucket;
637
638 /*
639 * Drain the bucket queues and free the buckets, we just keep two per
640 * cpu (alloc/free).
641 */
642 while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
643 LIST_REMOVE(bucket, ub_link);
644 ZONE_UNLOCK(zone);
645 bucket_drain(zone, bucket);
646 bucket_free(bucket);
647 ZONE_LOCK(zone);
648 }
649
650 /* Now we do the free queue.. */
651 while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
652 LIST_REMOVE(bucket, ub_link);
653 bucket_free(bucket);
654 }
655}
656
657/*
658 * Frees pages from a zone back to the system. This is done on demand from
659 * the pageout daemon.
660 *
661 * Arguments:
662 * zone The zone to free pages from
663 * all Should we drain all items?
664 *
665 * Returns:
666 * Nothing.
667 */
668static void
669zone_drain(uma_zone_t zone)
670{
671 struct slabhead freeslabs = { 0 };
672 uma_keg_t keg;
673 uma_slab_t slab;
674 uma_slab_t n;
675 u_int8_t flags;
676 u_int8_t *mem;
677 int i;
678
679 keg = zone->uz_keg;
680
681 /*
682 * We don't want to take pages from statically allocated zones at this
683 * time
684 */
685 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
686 return;
687
688 ZONE_LOCK(zone);
689
690#ifdef UMA_DEBUG
691 printf("%s free items: %u\n", zone->uz_name, keg->uk_free);
692#endif
693 bucket_cache_drain(zone);
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 ZONE_UNLOCK(zone);
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_MALLOC) ||
732 (keg->uk_flags & UMA_ZONE_REFCNT)) {
733 vm_object_t obj;
734
735 if (flags & UMA_SLAB_KMEM)
736 obj = kmem_object;
737 else
738 obj = NULL;
739 for (i = 0; i < keg->uk_ppera; i++)
740 vsetobj((vm_offset_t)mem + (i * PAGE_SIZE),
741 obj);
742 }
743 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
744 uma_zfree_internal(keg->uk_slabzone, slab, NULL,
745 SKIP_NONE);
746#ifdef UMA_DEBUG
747 printf("%s: Returning %d bytes.\n",
748 zone->uz_name, UMA_SLAB_SIZE * keg->uk_ppera);
749#endif
750 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags);
751 }
752}
753
754/*
755 * Allocate a new slab for a zone. This does not insert the slab onto a list.
756 *
757 * Arguments:
758 * zone The zone to allocate slabs for
759 * wait Shall we wait?
760 *
761 * Returns:
762 * The slab that was allocated or NULL if there is no memory and the
763 * caller specified M_NOWAIT.
764 */
765static uma_slab_t
766slab_zalloc(uma_zone_t zone, int wait)
767{
768 uma_slabrefcnt_t slabref;
769 uma_slab_t slab;
770 uma_keg_t keg;
771 u_int8_t *mem;
772 u_int8_t flags;
773 int i;
774
775 slab = NULL;
776 keg = zone->uz_keg;
777
778#ifdef UMA_DEBUG
779 printf("slab_zalloc: Allocating a new slab for %s\n", zone->uz_name);
780#endif
781 ZONE_UNLOCK(zone);
782
783 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
784 slab = uma_zalloc_internal(keg->uk_slabzone, NULL, wait);
785 if (slab == NULL) {
786 ZONE_LOCK(zone);
787 return NULL;
788 }
789 }
790
791 /*
792 * This reproduces the old vm_zone behavior of zero filling pages the
793 * first time they are added to a zone.
794 *
795 * Malloced items are zeroed in uma_zalloc.
796 */
797
798 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
799 wait |= M_ZERO;
800 else
801 wait &= ~M_ZERO;
802
803 mem = keg->uk_allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE,
804 &flags, wait);
805 if (mem == NULL) {
806 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
807 uma_zfree_internal(keg->uk_slabzone, slab, NULL,
808 SKIP_NONE);
809 ZONE_LOCK(zone);
810 return (NULL);
811 }
812
813 /* Point the slab into the allocated memory */
814 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
815 slab = (uma_slab_t )(mem + keg->uk_pgoff);
816
817 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
818 (keg->uk_flags & UMA_ZONE_REFCNT))
819 for (i = 0; i < keg->uk_ppera; i++)
820 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
821
822 slab->us_keg = keg;
823 slab->us_data = mem;
824 slab->us_freecount = keg->uk_ipers;
825 slab->us_firstfree = 0;
826 slab->us_flags = flags;
827
828 if (keg->uk_flags & UMA_ZONE_REFCNT) {
829 slabref = (uma_slabrefcnt_t)slab;
830 for (i = 0; i < keg->uk_ipers; i++) {
831 slabref->us_freelist[i].us_refcnt = 0;
832 slabref->us_freelist[i].us_item = i+1;
833 }
834 } else {
835 for (i = 0; i < keg->uk_ipers; i++)
836 slab->us_freelist[i].us_item = i+1;
837 }
838
839 if (keg->uk_init != NULL) {
840 for (i = 0; i < keg->uk_ipers; i++)
841 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
842 keg->uk_size, wait) != 0)
843 break;
844 if (i != keg->uk_ipers) {
845 if (keg->uk_fini != NULL) {
846 for (i--; i > -1; i--)
847 keg->uk_fini(slab->us_data +
848 (keg->uk_rsize * i),
849 keg->uk_size);
850 }
851 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
852 (keg->uk_flags & UMA_ZONE_REFCNT)) {
853 vm_object_t obj;
854
855 if (flags & UMA_SLAB_KMEM)
856 obj = kmem_object;
857 else
858 obj = NULL;
859 for (i = 0; i < keg->uk_ppera; i++)
860 vsetobj((vm_offset_t)mem +
861 (i * PAGE_SIZE), obj);
862 }
863 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
864 uma_zfree_internal(keg->uk_slabzone, slab,
865 NULL, SKIP_NONE);
866 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera,
867 flags);
868 ZONE_LOCK(zone);
869 return (NULL);
870 }
871 }
872 ZONE_LOCK(zone);
873
874 if (keg->uk_flags & UMA_ZONE_HASH)
875 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
876
877 keg->uk_pages += keg->uk_ppera;
878 keg->uk_free += keg->uk_ipers;
879
880 return (slab);
881}
882
883/*
884 * This function is intended to be used early on in place of page_alloc() so
885 * that we may use the boot time page cache to satisfy allocations before
886 * the VM is ready.
887 */
888static void *
889startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
890{
891 uma_keg_t keg;
892
893 keg = zone->uz_keg;
894
895 /*
896 * Check our small startup cache to see if it has pages remaining.
897 */
898 mtx_lock(&uma_mtx);
899 if (uma_boot_free != 0) {
900 uma_slab_t tmps;
901
902 tmps = LIST_FIRST(&uma_boot_pages);
903 LIST_REMOVE(tmps, us_link);
904 uma_boot_free--;
905 mtx_unlock(&uma_mtx);
906 *pflag = tmps->us_flags;
907 return (tmps->us_data);
908 }
909 mtx_unlock(&uma_mtx);
910 if (booted == 0)
911 panic("UMA: Increase UMA_BOOT_PAGES");
912 /*
913 * Now that we've booted reset these users to their real allocator.
914 */
915#ifdef UMA_MD_SMALL_ALLOC
916 keg->uk_allocf = uma_small_alloc;
917#else
918 keg->uk_allocf = page_alloc;
919#endif
920 return keg->uk_allocf(zone, bytes, pflag, wait);
921}
922
923/*
924 * Allocates a number of pages from the system
925 *
926 * Arguments:
927 * zone Unused
928 * bytes The number of bytes requested
929 * wait Shall we wait?
930 *
931 * Returns:
932 * A pointer to the alloced memory or possibly
933 * NULL if M_NOWAIT is set.
934 */
935static void *
936page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
937{
938 void *p; /* Returned page */
939
940 *pflag = UMA_SLAB_KMEM;
941 p = (void *) kmem_malloc(kmem_map, bytes, wait);
942
943 return (p);
944}
945
946/*
947 * Allocates a number of pages from within an object
948 *
949 * Arguments:
950 * zone Unused
951 * bytes The number of bytes requested
952 * wait Shall we wait?
953 *
954 * Returns:
955 * A pointer to the alloced memory or possibly
956 * NULL if M_NOWAIT is set.
957 */
958static void *
959obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
960{
961 vm_object_t object;
962 vm_offset_t retkva, zkva;
963 vm_page_t p;
964 int pages, startpages;
965
966 object = zone->uz_keg->uk_obj;
967 retkva = 0;
968
969 /*
970 * This looks a little weird since we're getting one page at a time.
971 */
972 VM_OBJECT_LOCK(object);
973 p = TAILQ_LAST(&object->memq, pglist);
974 pages = p != NULL ? p->pindex + 1 : 0;
975 startpages = pages;
976 zkva = zone->uz_keg->uk_kva + pages * PAGE_SIZE;
977 for (; bytes > 0; bytes -= PAGE_SIZE) {
978 p = vm_page_alloc(object, pages,
979 VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED);
980 if (p == NULL) {
981 if (pages != startpages)
982 pmap_qremove(retkva, pages - startpages);
983 while (pages != startpages) {
984 pages--;
985 p = TAILQ_LAST(&object->memq, pglist);
986 vm_page_lock_queues();
987 vm_page_unwire(p, 0);
988 vm_page_free(p);
989 vm_page_unlock_queues();
990 }
991 retkva = 0;
992 goto done;
993 }
994 pmap_qenter(zkva, &p, 1);
995 if (retkva == 0)
996 retkva = zkva;
997 zkva += PAGE_SIZE;
998 pages += 1;
999 }
1000done:
1001 VM_OBJECT_UNLOCK(object);
1002 *flags = UMA_SLAB_PRIV;
1003
1004 return ((void *)retkva);
1005}
1006
1007/*
1008 * Frees a number of pages to the system
1009 *
1010 * Arguments:
1011 * mem A pointer to the memory to be freed
1012 * size The size of the memory being freed
1013 * flags The original p->us_flags field
1014 *
1015 * Returns:
1016 * Nothing
1017 */
1018static void
1019page_free(void *mem, int size, u_int8_t flags)
1020{
1021 vm_map_t map;
1022
1023 if (flags & UMA_SLAB_KMEM)
1024 map = kmem_map;
1025 else
1026 panic("UMA: page_free used with invalid flags %d\n", flags);
1027
1028 kmem_free(map, (vm_offset_t)mem, size);
1029}
1030
1031/*
1032 * Zero fill initializer
1033 *
1034 * Arguments/Returns follow uma_init specifications
1035 */
1036static int
1037zero_init(void *mem, int size, int flags)
1038{
1039 bzero(mem, size);
1040 return (0);
1041}
1042
1043/*
1044 * Finish creating a small uma zone. This calculates ipers, and the zone size.
1045 *
1046 * Arguments
1047 * zone The zone we should initialize
1048 *
1049 * Returns
1050 * Nothing
1051 */
1052static void
1053zone_small_init(uma_zone_t zone)
1054{
1055 uma_keg_t keg;
1056 u_int rsize;
1057 u_int memused;
1058 u_int wastedspace;
1059 u_int shsize;
1060
1061 keg = zone->uz_keg;
1062 KASSERT(keg != NULL, ("Keg is null in zone_small_init"));
1063 rsize = keg->uk_size;
1064
1065 if (rsize < UMA_SMALLEST_UNIT)
1066 rsize = UMA_SMALLEST_UNIT;
1067 if (rsize & keg->uk_align)
1068 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1069
1070 keg->uk_rsize = rsize;
1071 keg->uk_ppera = 1;
1072
1073 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1074 rsize += UMA_FRITMREF_SZ; /* linkage & refcnt */
1075 shsize = sizeof(struct uma_slab_refcnt);
1076 } else {
1077 rsize += UMA_FRITM_SZ; /* Account for linkage */
1078 shsize = sizeof(struct uma_slab);
1079 }
1080
1081 keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize;
1082 KASSERT(keg->uk_ipers != 0, ("zone_small_init: ipers is 0"));
1083 memused = keg->uk_ipers * rsize + shsize;
1084 wastedspace = UMA_SLAB_SIZE - memused;
1085
1086 /*
1087 * We can't do OFFPAGE if we're internal or if we've been
1088 * asked to not go to the VM for buckets. If we do this we
1089 * may end up going to the VM (kmem_map) for slabs which we
1090 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1091 * result of UMA_ZONE_VM, which clearly forbids it.
1092 */
1093 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1094 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1095 return;
1096
1097 if ((wastedspace >= UMA_MAX_WASTE) &&
1098 (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) {
1099 keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize;
1100 KASSERT(keg->uk_ipers <= 255,
1101 ("zone_small_init: keg->uk_ipers too high!"));
1102#ifdef UMA_DEBUG
1103 printf("UMA decided we need offpage slab headers for "
1104 "zone: %s, calculated wastedspace = %d, "
1105 "maximum wasted space allowed = %d, "
1106 "calculated ipers = %d, "
1107 "new wasted space = %d\n", zone->uz_name, wastedspace,
1108 UMA_MAX_WASTE, keg->uk_ipers,
1109 UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize);
1110#endif
1111 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1112 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1113 keg->uk_flags |= UMA_ZONE_HASH;
1114 }
1115}
1116
1117/*
1118 * Finish creating a large (> UMA_SLAB_SIZE) uma zone. Just give in and do
1119 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1120 * more complicated.
1121 *
1122 * Arguments
1123 * zone The zone we should initialize
1124 *
1125 * Returns
1126 * Nothing
1127 */
1128static void
1129zone_large_init(uma_zone_t zone)
1130{
1131 uma_keg_t keg;
1132 int pages;
1133
1134 keg = zone->uz_keg;
1135
1136 KASSERT(keg != NULL, ("Keg is null in zone_large_init"));
1137 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1138 ("zone_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY zone"));
1139
1140 pages = keg->uk_size / UMA_SLAB_SIZE;
1141
1142 /* Account for remainder */
1143 if ((pages * UMA_SLAB_SIZE) < keg->uk_size)
1144 pages++;
1145
1146 keg->uk_ppera = pages;
1147 keg->uk_ipers = 1;
1148
1149 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1150 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1151 keg->uk_flags |= UMA_ZONE_HASH;
1152
1153 keg->uk_rsize = keg->uk_size;
1154}
1155
1156/*
1157 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1158 * the keg onto the global keg list.
1159 *
1160 * Arguments/Returns follow uma_ctor specifications
1161 * udata Actually uma_kctor_args
1162 */
1163static int
1164keg_ctor(void *mem, int size, void *udata, int flags)
1165{
1166 struct uma_kctor_args *arg = udata;
1167 uma_keg_t keg = mem;
1168 uma_zone_t zone;
1169
1170 bzero(keg, size);
1171 keg->uk_size = arg->size;
1172 keg->uk_init = arg->uminit;
1173 keg->uk_fini = arg->fini;
1174 keg->uk_align = arg->align;
1175 keg->uk_free = 0;
1176 keg->uk_pages = 0;
1177 keg->uk_flags = arg->flags;
1178 keg->uk_allocf = page_alloc;
1179 keg->uk_freef = page_free;
1180 keg->uk_recurse = 0;
1181 keg->uk_slabzone = NULL;
1182
1183 /*
1184 * The master zone is passed to us at keg-creation time.
1185 */
1186 zone = arg->zone;
1187 zone->uz_keg = keg;
1188
1189 if (arg->flags & UMA_ZONE_VM)
1190 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1191
1192 if (arg->flags & UMA_ZONE_ZINIT)
1193 keg->uk_init = zero_init;
1194
1195 /*
1196 * The +UMA_FRITM_SZ added to uk_size is to account for the
1197 * linkage that is added to the size in zone_small_init(). If
1198 * we don't account for this here then we may end up in
1199 * zone_small_init() with a calculated 'ipers' of 0.
1200 */
1201 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1202 if ((keg->uk_size+UMA_FRITMREF_SZ) >
1203 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)))
1204 zone_large_init(zone);
1205 else
1206 zone_small_init(zone);
1207 } else {
1208 if ((keg->uk_size+UMA_FRITM_SZ) >
1209 (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1210 zone_large_init(zone);
1211 else
1212 zone_small_init(zone);
1213 }
1214
1215 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1216 if (keg->uk_flags & UMA_ZONE_REFCNT)
1217 keg->uk_slabzone = slabrefzone;
1218 else
1219 keg->uk_slabzone = slabzone;
1220 }
1221
1222 /*
1223 * If we haven't booted yet we need allocations to go through the
1224 * startup cache until the vm is ready.
1225 */
1226 if (keg->uk_ppera == 1) {
1227#ifdef UMA_MD_SMALL_ALLOC
1228 keg->uk_allocf = uma_small_alloc;
1229 keg->uk_freef = uma_small_free;
1230#endif
1231 if (booted == 0)
1232 keg->uk_allocf = startup_alloc;
1233 }
1234
1235 /*
1236 * Initialize keg's lock (shared among zones) through
1237 * Master zone
1238 */
1239 zone->uz_lock = &keg->uk_lock;
1240 if (arg->flags & UMA_ZONE_MTXCLASS)
1241 ZONE_LOCK_INIT(zone, 1);
1242 else
1243 ZONE_LOCK_INIT(zone, 0);
1244
1245 /*
1246 * If we're putting the slab header in the actual page we need to
1247 * figure out where in each page it goes. This calculates a right
1248 * justified offset into the memory on an ALIGN_PTR boundary.
1249 */
1250 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1251 u_int totsize;
1252
1253 /* Size of the slab struct and free list */
1254 if (keg->uk_flags & UMA_ZONE_REFCNT)
1255 totsize = sizeof(struct uma_slab_refcnt) +
1256 keg->uk_ipers * UMA_FRITMREF_SZ;
1257 else
1258 totsize = sizeof(struct uma_slab) +
1259 keg->uk_ipers * UMA_FRITM_SZ;
1260
1261 if (totsize & UMA_ALIGN_PTR)
1262 totsize = (totsize & ~UMA_ALIGN_PTR) +
1263 (UMA_ALIGN_PTR + 1);
1264 keg->uk_pgoff = UMA_SLAB_SIZE - totsize;
1265
1266 if (keg->uk_flags & UMA_ZONE_REFCNT)
1267 totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt)
1268 + keg->uk_ipers * UMA_FRITMREF_SZ;
1269 else
1270 totsize = keg->uk_pgoff + sizeof(struct uma_slab)
1271 + keg->uk_ipers * UMA_FRITM_SZ;
1272
1273 /*
1274 * The only way the following is possible is if with our
1275 * UMA_ALIGN_PTR adjustments we are now bigger than
1276 * UMA_SLAB_SIZE. I haven't checked whether this is
1277 * mathematically possible for all cases, so we make
1278 * sure here anyway.
1279 */
1280 if (totsize > UMA_SLAB_SIZE) {
1281 printf("zone %s ipers %d rsize %d size %d\n",
1282 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1283 keg->uk_size);
1284 panic("UMA slab won't fit.\n");
1285 }
1286 }
1287
1288 if (keg->uk_flags & UMA_ZONE_HASH)
1289 hash_alloc(&keg->uk_hash);
1290
1291#ifdef UMA_DEBUG
1292 printf("%s(%p) size = %d ipers = %d ppera = %d pgoff = %d\n",
1293 zone->uz_name, zone,
1294 keg->uk_size, keg->uk_ipers,
1295 keg->uk_ppera, keg->uk_pgoff);
1296#endif
1297
1298 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1299
1300 mtx_lock(&uma_mtx);
1301 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1302 mtx_unlock(&uma_mtx);
1303 return (0);
1304}
1305
1306/*
1307 * Zone header ctor. This initializes all fields, locks, etc.
1308 *
1309 * Arguments/Returns follow uma_ctor specifications
1310 * udata Actually uma_zctor_args
1311 */
1312
1313static int
1314zone_ctor(void *mem, int size, void *udata, int flags)
1315{
1316 struct uma_zctor_args *arg = udata;
1317 uma_zone_t zone = mem;
1318 uma_zone_t z;
1319 uma_keg_t keg;
1320
1321 bzero(zone, size);
1322 zone->uz_name = arg->name;
1323 zone->uz_ctor = arg->ctor;
1324 zone->uz_dtor = arg->dtor;
1325 zone->uz_init = NULL;
1326 zone->uz_fini = NULL;
1327 zone->uz_allocs = 0;
1328 zone->uz_frees = 0;
1329 zone->uz_fills = zone->uz_count = 0;
1330
1331 if (arg->flags & UMA_ZONE_SECONDARY) {
1332 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1333 keg = arg->keg;
1334 zone->uz_keg = keg;
1335 zone->uz_init = arg->uminit;
1336 zone->uz_fini = arg->fini;
1337 zone->uz_lock = &keg->uk_lock;
1338 mtx_lock(&uma_mtx);
1339 ZONE_LOCK(zone);
1340 keg->uk_flags |= UMA_ZONE_SECONDARY;
1341 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1342 if (LIST_NEXT(z, uz_link) == NULL) {
1343 LIST_INSERT_AFTER(z, zone, uz_link);
1344 break;
1345 }
1346 }
1347 ZONE_UNLOCK(zone);
1348 mtx_unlock(&uma_mtx);
1349 } else if (arg->keg == NULL) {
1350 if (uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1351 arg->align, arg->flags) == NULL)
1352 return (ENOMEM);
1353 } else {
1354 struct uma_kctor_args karg;
1355 int error;
1356
1357 /* We should only be here from uma_startup() */
1358 karg.size = arg->size;
1359 karg.uminit = arg->uminit;
1360 karg.fini = arg->fini;
1361 karg.align = arg->align;
1362 karg.flags = arg->flags;
1363 karg.zone = zone;
1364 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1365 flags);
1366 if (error)
1367 return (error);
1368 }
1369 keg = zone->uz_keg;
1370 zone->uz_lock = &keg->uk_lock;
1371
1372 /*
1373 * Some internal zones don't have room allocated for the per cpu
1374 * caches. If we're internal, bail out here.
1375 */
1376 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1377 KASSERT((keg->uk_flags & UMA_ZONE_SECONDARY) == 0,
1378 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1379 return (0);
1380 }
1381
1382 if (keg->uk_flags & UMA_ZONE_MAXBUCKET)
1383 zone->uz_count = BUCKET_MAX;
1384 else if (keg->uk_ipers <= BUCKET_MAX)
1385 zone->uz_count = keg->uk_ipers;
1386 else
1387 zone->uz_count = BUCKET_MAX;
1388 return (0);
1389}
1390
1391/*
1392 * Keg header dtor. This frees all data, destroys locks, frees the hash
1393 * table and removes the keg from the global list.
1394 *
1395 * Arguments/Returns follow uma_dtor specifications
1396 * udata unused
1397 */
1398static void
1399keg_dtor(void *arg, int size, void *udata)
1400{
1401 uma_keg_t keg;
1402
1403 keg = (uma_keg_t)arg;
1404 mtx_lock(&keg->uk_lock);
1405 if (keg->uk_free != 0) {
1406 printf("Freed UMA keg was not empty (%d items). "
1407 " Lost %d pages of memory.\n",
1408 keg->uk_free, keg->uk_pages);
1409 }
1410 mtx_unlock(&keg->uk_lock);
1411
1412 if (keg->uk_flags & UMA_ZONE_HASH)
1413 hash_free(&keg->uk_hash);
1414
1415 mtx_destroy(&keg->uk_lock);
1416}
1417
1418/*
1419 * Zone header dtor.
1420 *
1421 * Arguments/Returns follow uma_dtor specifications
1422 * udata unused
1423 */
1424static void
1425zone_dtor(void *arg, int size, void *udata)
1426{
1427 uma_zone_t zone;
1428 uma_keg_t keg;
1429
1430 zone = (uma_zone_t)arg;
1431 keg = zone->uz_keg;
1432
1433 if (!(keg->uk_flags & UMA_ZFLAG_INTERNAL))
1434 cache_drain(zone);
1435
1436 mtx_lock(&uma_mtx);
1437 zone_drain(zone);
1438 if (keg->uk_flags & UMA_ZONE_SECONDARY) {
1439 LIST_REMOVE(zone, uz_link);
1440 /*
1441 * XXX there are some races here where
1442 * the zone can be drained but zone lock
1443 * released and then refilled before we
1444 * remove it... we dont care for now
1445 */
1446 ZONE_LOCK(zone);
1447 if (LIST_EMPTY(&keg->uk_zones))
1448 keg->uk_flags &= ~UMA_ZONE_SECONDARY;
1449 ZONE_UNLOCK(zone);
1450 mtx_unlock(&uma_mtx);
1451 } else {
1452 LIST_REMOVE(keg, uk_link);
1453 LIST_REMOVE(zone, uz_link);
1454 mtx_unlock(&uma_mtx);
1455 uma_zfree_internal(kegs, keg, NULL, SKIP_NONE);
1456 }
1457 zone->uz_keg = NULL;
1458}
1459
1460/*
1461 * Traverses every zone in the system and calls a callback
1462 *
1463 * Arguments:
1464 * zfunc A pointer to a function which accepts a zone
1465 * as an argument.
1466 *
1467 * Returns:
1468 * Nothing
1469 */
1470static void
1471zone_foreach(void (*zfunc)(uma_zone_t))
1472{
1473 uma_keg_t keg;
1474 uma_zone_t zone;
1475
1476 mtx_lock(&uma_mtx);
1477 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1478 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1479 zfunc(zone);
1480 }
1481 mtx_unlock(&uma_mtx);
1482}
1483
1484/* Public functions */
1485/* See uma.h */
1486void
1487uma_startup(void *bootmem)
1488{
1489 struct uma_zctor_args args;
1490 uma_slab_t slab;
1491 u_int slabsize;
1492 u_int objsize, totsize, wsize;
1493 int i;
1494
1495#ifdef UMA_DEBUG
1496 printf("Creating uma keg headers zone and keg.\n");
1497#endif
1498 /*
1499 * The general UMA lock is a recursion-allowed lock because
1500 * there is a code path where, while we're still configured
1501 * to use startup_alloc() for backend page allocations, we
1502 * may end up in uma_reclaim() which calls zone_foreach(zone_drain),
1503 * which grabs uma_mtx, only to later call into startup_alloc()
1504 * because while freeing we needed to allocate a bucket. Since
1505 * startup_alloc() also takes uma_mtx, we need to be able to
1506 * recurse on it.
1507 */
1508 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF | MTX_RECURSE);
1509
1510 /*
1511 * Figure out the maximum number of items-per-slab we'll have if
1512 * we're using the OFFPAGE slab header to track free items, given
1513 * all possible object sizes and the maximum desired wastage
1514 * (UMA_MAX_WASTE).
1515 *
1516 * We iterate until we find an object size for
1517 * which the calculated wastage in zone_small_init() will be
1518 * enough to warrant OFFPAGE. Since wastedspace versus objsize
1519 * is an overall increasing see-saw function, we find the smallest
1520 * objsize such that the wastage is always acceptable for objects
1521 * with that objsize or smaller. Since a smaller objsize always
1522 * generates a larger possible uma_max_ipers, we use this computed
1523 * objsize to calculate the largest ipers possible. Since the
1524 * ipers calculated for OFFPAGE slab headers is always larger than
1525 * the ipers initially calculated in zone_small_init(), we use
1526 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to
1527 * obtain the maximum ipers possible for offpage slab headers.
1528 *
1529 * It should be noted that ipers versus objsize is an inversly
1530 * proportional function which drops off rather quickly so as
1531 * long as our UMA_MAX_WASTE is such that the objsize we calculate
1532 * falls into the portion of the inverse relation AFTER the steep
1533 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386).
1534 *
1535 * Note that we have 8-bits (1 byte) to use as a freelist index
1536 * inside the actual slab header itself and this is enough to
1537 * accomodate us. In the worst case, a UMA_SMALLEST_UNIT sized
1538 * object with offpage slab header would have ipers =
1539 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is
1540 * 1 greater than what our byte-integer freelist index can
1541 * accomodate, but we know that this situation never occurs as
1542 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate
1543 * that we need to go to offpage slab headers. Or, if we do,
1544 * then we trap that condition below and panic in the INVARIANTS case.
1545 */
1546 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE;
1547 totsize = wsize;
1548 objsize = UMA_SMALLEST_UNIT;
1549 while (totsize >= wsize) {
1550 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) /
1551 (objsize + UMA_FRITM_SZ);
1552 totsize *= (UMA_FRITM_SZ + objsize);
1553 objsize++;
1554 }
1555 if (objsize > UMA_SMALLEST_UNIT)
1556 objsize--;
1557 uma_max_ipers = UMA_SLAB_SIZE / objsize;
1558
1559 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE;
1560 totsize = wsize;
1561 objsize = UMA_SMALLEST_UNIT;
1562 while (totsize >= wsize) {
1563 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) /
1564 (objsize + UMA_FRITMREF_SZ);
1565 totsize *= (UMA_FRITMREF_SZ + objsize);
1566 objsize++;
1567 }
1568 if (objsize > UMA_SMALLEST_UNIT)
1569 objsize--;
1570 uma_max_ipers_ref = UMA_SLAB_SIZE / objsize;
1571
1572 KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255),
1573 ("uma_startup: calculated uma_max_ipers values too large!"));
1574
1575#ifdef UMA_DEBUG
1576 printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers);
1577 printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n",
1578 uma_max_ipers_ref);
1579#endif
1580
1581 /* "manually" create the initial zone */
1582 args.name = "UMA Kegs";
1583 args.size = sizeof(struct uma_keg);
1584 args.ctor = keg_ctor;
1585 args.dtor = keg_dtor;
1586 args.uminit = zero_init;
1587 args.fini = NULL;
1588 args.keg = &masterkeg;
1589 args.align = 32 - 1;
1590 args.flags = UMA_ZFLAG_INTERNAL;
1591 /* The initial zone has no Per cpu queues so it's smaller */
1592 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1593
1594#ifdef UMA_DEBUG
1595 printf("Filling boot free list.\n");
1596#endif
1597 for (i = 0; i < UMA_BOOT_PAGES; i++) {
1598 slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE));
1599 slab->us_data = (u_int8_t *)slab;
1600 slab->us_flags = UMA_SLAB_BOOT;
1601 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1602 uma_boot_free++;
1603 }
1604
1605#ifdef UMA_DEBUG
1606 printf("Creating uma zone headers zone and keg.\n");
1607#endif
1608 args.name = "UMA Zones";
1609 args.size = sizeof(struct uma_zone) +
1610 (sizeof(struct uma_cache) * (mp_maxid + 1));
1611 args.ctor = zone_ctor;
1612 args.dtor = zone_dtor;
1613 args.uminit = zero_init;
1614 args.fini = NULL;
1615 args.keg = NULL;
1616 args.align = 32 - 1;
1617 args.flags = UMA_ZFLAG_INTERNAL;
1618 /* The initial zone has no Per cpu queues so it's smaller */
1619 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1620
1621#ifdef UMA_DEBUG
1622 printf("Initializing pcpu cache locks.\n");
1623#endif
1624#ifdef UMA_DEBUG
1625 printf("Creating slab and hash zones.\n");
1626#endif
1627
1628 /*
1629 * This is the max number of free list items we'll have with
1630 * offpage slabs.
1631 */
1632 slabsize = uma_max_ipers * UMA_FRITM_SZ;
1633 slabsize += sizeof(struct uma_slab);
1634
1635 /* Now make a zone for slab headers */
1636 slabzone = uma_zcreate("UMA Slabs",
1637 slabsize,
1638 NULL, NULL, NULL, NULL,
1639 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1640
1641 /*
1642 * We also create a zone for the bigger slabs with reference
1643 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1644 */
1645 slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ;
1646 slabsize += sizeof(struct uma_slab_refcnt);
1647 slabrefzone = uma_zcreate("UMA RCntSlabs",
1648 slabsize,
1649 NULL, NULL, NULL, NULL,
1650 UMA_ALIGN_PTR,
1651 UMA_ZFLAG_INTERNAL);
1652
1653 hashzone = uma_zcreate("UMA Hash",
1654 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1655 NULL, NULL, NULL, NULL,
1656 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1657
1658 bucket_init();
1659
1660#ifdef UMA_MD_SMALL_ALLOC
1661 booted = 1;
1662#endif
1663
1664#ifdef UMA_DEBUG
1665 printf("UMA startup complete.\n");
1666#endif
1667}
1668
1669/* see uma.h */
1670void
1671uma_startup2(void)
1672{
1673 booted = 1;
1674 bucket_enable();
1675#ifdef UMA_DEBUG
1676 printf("UMA startup2 complete.\n");
1677#endif
1678}
1679
1680/*
1681 * Initialize our callout handle
1682 *
1683 */
1684
1685static void
1686uma_startup3(void)
1687{
1688#ifdef UMA_DEBUG
1689 printf("Starting callout.\n");
1690#endif
1691 callout_init(&uma_callout, CALLOUT_MPSAFE);
1692 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1693#ifdef UMA_DEBUG
1694 printf("UMA startup3 complete.\n");
1695#endif
1696}
1697
1698static uma_zone_t
1699uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1700 int align, u_int16_t flags)
1701{
1702 struct uma_kctor_args args;
1703
1704 args.size = size;
1705 args.uminit = uminit;
1706 args.fini = fini;
1707 args.align = align;
1708 args.flags = flags;
1709 args.zone = zone;
1710 return (uma_zalloc_internal(kegs, &args, M_WAITOK));
1711}
1712
1713/* See uma.h */
1714uma_zone_t
1715uma_zcreate(char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1716 uma_init uminit, uma_fini fini, int align, u_int16_t flags)
1717
1718{
1719 struct uma_zctor_args args;
1720
1721 /* This stuff is essential for the zone ctor */
1722 args.name = name;
1723 args.size = size;
1724 args.ctor = ctor;
1725 args.dtor = dtor;
1726 args.uminit = uminit;
1727 args.fini = fini;
1728 args.align = align;
1729 args.flags = flags;
1730 args.keg = NULL;
1731
1732 return (uma_zalloc_internal(zones, &args, M_WAITOK));
1733}
1734
1735/* See uma.h */
1736uma_zone_t
1737uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1738 uma_init zinit, uma_fini zfini, uma_zone_t master)
1739{
1740 struct uma_zctor_args args;
1741
1742 args.name = name;
1743 args.size = master->uz_keg->uk_size;
1744 args.ctor = ctor;
1745 args.dtor = dtor;
1746 args.uminit = zinit;
1747 args.fini = zfini;
1748 args.align = master->uz_keg->uk_align;
1749 args.flags = master->uz_keg->uk_flags | UMA_ZONE_SECONDARY;
1750 args.keg = master->uz_keg;
1751
1752 return (uma_zalloc_internal(zones, &args, M_WAITOK));
1753}
1754
1755/* See uma.h */
1756void
1757uma_zdestroy(uma_zone_t zone)
1758{
1759 uma_zfree_internal(zones, zone, NULL, SKIP_NONE);
1760}
1761
1762/* See uma.h */
1763void *
1764uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1765{
1766 void *item;
1767 uma_cache_t cache;
1768 uma_bucket_t bucket;
1769 int cpu;
1770 int badness;
1771
1772 /* This is the fast path allocation */
1773#ifdef UMA_DEBUG_ALLOC_1
1774 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1775#endif
1776 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1777 zone->uz_name, flags);
1778
1779 if (!(flags & M_NOWAIT)) {
1780 KASSERT(curthread->td_intr_nesting_level == 0,
1781 ("malloc(M_WAITOK) in interrupt context"));
1782 if (nosleepwithlocks) {
1783#ifdef WITNESS
1784 badness = WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK,
1785 NULL,
1786 "malloc(M_WAITOK) of \"%s\", forcing M_NOWAIT",
1787 zone->uz_name);
1788#else
1789 badness = 1;
1790#endif
1791 } else {
1792 badness = 0;
1793#ifdef WITNESS
1794 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1795 "malloc(M_WAITOK) of \"%s\"", zone->uz_name);
1796#endif
1797 }
1798 if (badness) {
1799 flags &= ~M_WAITOK;
1800 flags |= M_NOWAIT;
1801 }
1802 }
1803
1804 /*
1805 * If possible, allocate from the per-CPU cache. There are two
1806 * requirements for safe access to the per-CPU cache: (1) the thread
1807 * accessing the cache must not be preempted or yield during access,
1808 * and (2) the thread must not migrate CPUs without switching which
1809 * cache it accesses. We rely on a critical section to prevent
1810 * preemption and migration. We release the critical section in
1811 * order to acquire the zone mutex if we are unable to allocate from
1812 * the current cache; when we re-acquire the critical section, we
1813 * must detect and handle migration if it has occurred.
1814 */
1815zalloc_restart:
1816 critical_enter();
1817 cpu = curcpu;
1818 cache = &zone->uz_cpu[cpu];
1819
1820zalloc_start:
1821 bucket = cache->uc_allocbucket;
1822
1823 if (bucket) {
1824 if (bucket->ub_cnt > 0) {
1825 bucket->ub_cnt--;
1826 item = bucket->ub_bucket[bucket->ub_cnt];
1827#ifdef INVARIANTS
1828 bucket->ub_bucket[bucket->ub_cnt] = NULL;
1829#endif
1830 KASSERT(item != NULL,
1831 ("uma_zalloc: Bucket pointer mangled."));
1832 cache->uc_allocs++;
1833 critical_exit();
1834#ifdef INVARIANTS
1835 ZONE_LOCK(zone);
1836 uma_dbg_alloc(zone, NULL, item);
1837 ZONE_UNLOCK(zone);
1838#endif
1839 if (zone->uz_ctor != NULL) {
1840 if (zone->uz_ctor(item, zone->uz_keg->uk_size,
1841 udata, flags) != 0) {
1842 uma_zfree_internal(zone, item, udata,
1843 SKIP_DTOR);
1844 return (NULL);
1845 }
1846 }
1847 if (flags & M_ZERO)
1848 bzero(item, zone->uz_keg->uk_size);
1849 return (item);
1850 } else if (cache->uc_freebucket) {
1851 /*
1852 * We have run out of items in our allocbucket.
1853 * See if we can switch with our free bucket.
1854 */
1855 if (cache->uc_freebucket->ub_cnt > 0) {
1856#ifdef UMA_DEBUG_ALLOC
1857 printf("uma_zalloc: Swapping empty with"
1858 " alloc.\n");
1859#endif
1860 bucket = cache->uc_freebucket;
1861 cache->uc_freebucket = cache->uc_allocbucket;
1862 cache->uc_allocbucket = bucket;
1863
1864 goto zalloc_start;
1865 }
1866 }
1867 }
1868 /*
1869 * Attempt to retrieve the item from the per-CPU cache has failed, so
1870 * we must go back to the zone. This requires the zone lock, so we
1871 * must drop the critical section, then re-acquire it when we go back
1872 * to the cache. Since the critical section is released, we may be
1873 * preempted or migrate. As such, make sure not to maintain any
1874 * thread-local state specific to the cache from prior to releasing
1875 * the critical section.
1876 */
1877 critical_exit();
1878 ZONE_LOCK(zone);
1879 critical_enter();
1880 cpu = curcpu;
1881 cache = &zone->uz_cpu[cpu];
1882 bucket = cache->uc_allocbucket;
1883 if (bucket != NULL) {
1884 if (bucket->ub_cnt > 0) {
1885 ZONE_UNLOCK(zone);
1886 goto zalloc_start;
1887 }
1888 bucket = cache->uc_freebucket;
1889 if (bucket != NULL && bucket->ub_cnt > 0) {
1890 ZONE_UNLOCK(zone);
1891 goto zalloc_start;
1892 }
1893 }
1894
1895 /* Since we have locked the zone we may as well send back our stats */
1896 zone->uz_allocs += cache->uc_allocs;
1897 cache->uc_allocs = 0;
1898 zone->uz_frees += cache->uc_frees;
1899 cache->uc_frees = 0;
1900
1901 /* Our old one is now a free bucket */
1902 if (cache->uc_allocbucket) {
1903 KASSERT(cache->uc_allocbucket->ub_cnt == 0,
1904 ("uma_zalloc_arg: Freeing a non free bucket."));
1905 LIST_INSERT_HEAD(&zone->uz_free_bucket,
1906 cache->uc_allocbucket, ub_link);
1907 cache->uc_allocbucket = NULL;
1908 }
1909
1910 /* Check the free list for a new alloc bucket */
1911 if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
1912 KASSERT(bucket->ub_cnt != 0,
1913 ("uma_zalloc_arg: Returning an empty bucket."));
1914
1915 LIST_REMOVE(bucket, ub_link);
1916 cache->uc_allocbucket = bucket;
1917 ZONE_UNLOCK(zone);
1918 goto zalloc_start;
1919 }
1920 /* We are no longer associated with this CPU. */
1921 critical_exit();
1922
1923 /* Bump up our uz_count so we get here less */
1924 if (zone->uz_count < BUCKET_MAX)
1925 zone->uz_count++;
1926
1927 /*
1928 * Now lets just fill a bucket and put it on the free list. If that
1929 * works we'll restart the allocation from the begining.
1930 */
1931 if (uma_zalloc_bucket(zone, flags)) {
1932 ZONE_UNLOCK(zone);
1933 goto zalloc_restart;
1934 }
1935 ZONE_UNLOCK(zone);
1936 /*
1937 * We may not be able to get a bucket so return an actual item.
1938 */
1939#ifdef UMA_DEBUG
1940 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
1941#endif
1942
1943 return (uma_zalloc_internal(zone, udata, flags));
1944}
1945
1946static uma_slab_t
1947uma_zone_slab(uma_zone_t zone, int flags)
1948{
1949 uma_slab_t slab;
1950 uma_keg_t keg;
1951
1952 keg = zone->uz_keg;
1953
1954 /*
1955 * This is to prevent us from recursively trying to allocate
1956 * buckets. The problem is that if an allocation forces us to
1957 * grab a new bucket we will call page_alloc, which will go off
1958 * and cause the vm to allocate vm_map_entries. If we need new
1959 * buckets there too we will recurse in kmem_alloc and bad
1960 * things happen. So instead we return a NULL bucket, and make
1961 * the code that allocates buckets smart enough to deal with it
1962 *
1963 * XXX: While we want this protection for the bucket zones so that
1964 * recursion from the VM is handled (and the calling code that
1965 * allocates buckets knows how to deal with it), we do not want
1966 * to prevent allocation from the slab header zones (slabzone
1967 * and slabrefzone) if uk_recurse is not zero for them. The
1968 * reason is that it could lead to NULL being returned for
1969 * slab header allocations even in the M_WAITOK case, and the
1970 * caller can't handle that.
1971 */
1972 if (keg->uk_flags & UMA_ZFLAG_INTERNAL && keg->uk_recurse != 0)
1973 if ((zone != slabzone) && (zone != slabrefzone))
1974 return (NULL);
1975
1976 slab = NULL;
1977
1978 for (;;) {
1979 /*
1980 * Find a slab with some space. Prefer slabs that are partially
1981 * used over those that are totally full. This helps to reduce
1982 * fragmentation.
1983 */
1984 if (keg->uk_free != 0) {
1985 if (!LIST_EMPTY(&keg->uk_part_slab)) {
1986 slab = LIST_FIRST(&keg->uk_part_slab);
1987 } else {
1988 slab = LIST_FIRST(&keg->uk_free_slab);
1989 LIST_REMOVE(slab, us_link);
1990 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
1991 us_link);
1992 }
1993 return (slab);
1994 }
1995
1996 /*
1997 * M_NOVM means don't ask at all!
1998 */
1999 if (flags & M_NOVM)
2000 break;
2001
2002 if (keg->uk_maxpages &&
2003 keg->uk_pages >= keg->uk_maxpages) {
2004 keg->uk_flags |= UMA_ZFLAG_FULL;
2005
2006 if (flags & M_NOWAIT)
2007 break;
2008 else
2009 msleep(keg, &keg->uk_lock, PVM,
2010 "zonelimit", 0);
2011 continue;
2012 }
2013 keg->uk_recurse++;
2014 slab = slab_zalloc(zone, flags);
2015 keg->uk_recurse--;
2016
2017 /*
2018 * If we got a slab here it's safe to mark it partially used
2019 * and return. We assume that the caller is going to remove
2020 * at least one item.
2021 */
2022 if (slab) {
2023 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2024 return (slab);
2025 }
2026 /*
2027 * We might not have been able to get a slab but another cpu
2028 * could have while we were unlocked. Check again before we
2029 * fail.
2030 */
2031 if (flags & M_NOWAIT)
2032 flags |= M_NOVM;
2033 }
2034 return (slab);
2035}
2036
2037static void *
2038uma_slab_alloc(uma_zone_t zone, uma_slab_t slab)
2039{
2040 uma_keg_t keg;
2041 uma_slabrefcnt_t slabref;
2042 void *item;
2043 u_int8_t freei;
2044
2045 keg = zone->uz_keg;
2046
2047 freei = slab->us_firstfree;
2048 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2049 slabref = (uma_slabrefcnt_t)slab;
2050 slab->us_firstfree = slabref->us_freelist[freei].us_item;
2051 } else {
2052 slab->us_firstfree = slab->us_freelist[freei].us_item;
2053 }
2054 item = slab->us_data + (keg->uk_rsize * freei);
2055
2056 slab->us_freecount--;
2057 keg->uk_free--;
2058#ifdef INVARIANTS
2059 uma_dbg_alloc(zone, slab, item);
2060#endif
2061 /* Move this slab to the full list */
2062 if (slab->us_freecount == 0) {
2063 LIST_REMOVE(slab, us_link);
2064 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2065 }
2066
2067 return (item);
2068}
2069
2070static int
2071uma_zalloc_bucket(uma_zone_t zone, int flags)
2072{
2073 uma_bucket_t bucket;
2074 uma_slab_t slab;
2075 int16_t saved;
2076 int max, origflags = flags;
2077
2078 /*
2079 * Try this zone's free list first so we don't allocate extra buckets.
2080 */
2081 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2082 KASSERT(bucket->ub_cnt == 0,
2083 ("uma_zalloc_bucket: Bucket on free list is not empty."));
2084 LIST_REMOVE(bucket, ub_link);
2085 } else {
2086 int bflags;
2087
2088 bflags = (flags & ~M_ZERO);
2089 if (zone->uz_keg->uk_flags & UMA_ZFLAG_CACHEONLY)
2090 bflags |= M_NOVM;
2091
2092 ZONE_UNLOCK(zone);
2093 bucket = bucket_alloc(zone->uz_count, bflags);
2094 ZONE_LOCK(zone);
2095 }
2096
2097 if (bucket == NULL)
2098 return (0);
2099
2100#ifdef SMP
2101 /*
2102 * This code is here to limit the number of simultaneous bucket fills
2103 * for any given zone to the number of per cpu caches in this zone. This
2104 * is done so that we don't allocate more memory than we really need.
2105 */
2106 if (zone->uz_fills >= mp_ncpus)
2107 goto done;
2108
2109#endif
2110 zone->uz_fills++;
2111
2112 max = MIN(bucket->ub_entries, zone->uz_count);
2113 /* Try to keep the buckets totally full */
2114 saved = bucket->ub_cnt;
2115 while (bucket->ub_cnt < max &&
2116 (slab = uma_zone_slab(zone, flags)) != NULL) {
2117 while (slab->us_freecount && bucket->ub_cnt < max) {
2118 bucket->ub_bucket[bucket->ub_cnt++] =
2119 uma_slab_alloc(zone, slab);
2120 }
2121
2122 /* Don't block on the next fill */
2123 flags |= M_NOWAIT;
2124 }
2125
2126 /*
2127 * We unlock here because we need to call the zone's init.
2128 * It should be safe to unlock because the slab dealt with
2129 * above is already on the appropriate list within the keg
2130 * and the bucket we filled is not yet on any list, so we
2131 * own it.
2132 */
2133 if (zone->uz_init != NULL) {
2134 int i;
2135
2136 ZONE_UNLOCK(zone);
2137 for (i = saved; i < bucket->ub_cnt; i++)
2138 if (zone->uz_init(bucket->ub_bucket[i],
2139 zone->uz_keg->uk_size, origflags) != 0)
2140 break;
2141 /*
2142 * If we couldn't initialize the whole bucket, put the
2143 * rest back onto the freelist.
2144 */
2145 if (i != bucket->ub_cnt) {
2146 int j;
2147
2148 for (j = i; j < bucket->ub_cnt; j++) {
2149 uma_zfree_internal(zone, bucket->ub_bucket[j],
2150 NULL, SKIP_FINI);
2151#ifdef INVARIANTS
2152 bucket->ub_bucket[j] = NULL;
2153#endif
2154 }
2155 bucket->ub_cnt = i;
2156 }
2157 ZONE_LOCK(zone);
2158 }
2159
2160 zone->uz_fills--;
2161 if (bucket->ub_cnt != 0) {
2162 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2163 bucket, ub_link);
2164 return (1);
2165 }
2166#ifdef SMP
2167done:
2168#endif
2169 bucket_free(bucket);
2170
2171 return (0);
2172}
2173/*
2174 * Allocates an item for an internal zone
2175 *
2176 * Arguments
2177 * zone The zone to alloc for.
2178 * udata The data to be passed to the constructor.
2179 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2180 *
2181 * Returns
2182 * NULL if there is no memory and M_NOWAIT is set
2183 * An item if successful
2184 */
2185
2186static void *
2187uma_zalloc_internal(uma_zone_t zone, void *udata, int flags)
2188{
2189 uma_keg_t keg;
2190 uma_slab_t slab;
2191 void *item;
2192
2193 item = NULL;
2194 keg = zone->uz_keg;
2195
2196#ifdef UMA_DEBUG_ALLOC
2197 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2198#endif
2199 ZONE_LOCK(zone);
2200
2201 slab = uma_zone_slab(zone, flags);
2202 if (slab == NULL) {
2203 ZONE_UNLOCK(zone);
2204 return (NULL);
2205 }
2206
2207 item = uma_slab_alloc(zone, slab);
2208
2209 zone->uz_allocs++;
2210
2211 ZONE_UNLOCK(zone);
2212
2213 /*
2214 * We have to call both the zone's init (not the keg's init)
2215 * and the zone's ctor. This is because the item is going from
2216 * a keg slab directly to the user, and the user is expecting it
2217 * to be both zone-init'd as well as zone-ctor'd.
2218 */
2219 if (zone->uz_init != NULL) {
2220 if (zone->uz_init(item, keg->uk_size, flags) != 0) {
2221 uma_zfree_internal(zone, item, udata, SKIP_FINI);
2222 return (NULL);
2223 }
2224 }
2225 if (zone->uz_ctor != NULL) {
2226 if (zone->uz_ctor(item, keg->uk_size, udata, flags) != 0) {
2227 uma_zfree_internal(zone, item, udata, SKIP_DTOR);
2228 return (NULL);
2229 }
2230 }
2231 if (flags & M_ZERO)
2232 bzero(item, keg->uk_size);
2233
2234 return (item);
2235}
2236
2237/* See uma.h */
2238void
2239uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2240{
2241 uma_keg_t keg;
2242 uma_cache_t cache;
2243 uma_bucket_t bucket;
2244 int bflags;
2245 int cpu;
2246
2247 keg = zone->uz_keg;
2248
2249#ifdef UMA_DEBUG_ALLOC_1
2250 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2251#endif
2252 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2253 zone->uz_name);
2254
2255 if (zone->uz_dtor)
2256 zone->uz_dtor(item, keg->uk_size, udata);
2257#ifdef INVARIANTS
2258 ZONE_LOCK(zone);
2259 if (keg->uk_flags & UMA_ZONE_MALLOC)
2260 uma_dbg_free(zone, udata, item);
2261 else
2262 uma_dbg_free(zone, NULL, item);
2263 ZONE_UNLOCK(zone);
2264#endif
2265 /*
2266 * The race here is acceptable. If we miss it we'll just have to wait
2267 * a little longer for the limits to be reset.
2268 */
2269 if (keg->uk_flags & UMA_ZFLAG_FULL)
2270 goto zfree_internal;
2271
2272 /*
2273 * If possible, free to the per-CPU cache. There are two
2274 * requirements for safe access to the per-CPU cache: (1) the thread
2275 * accessing the cache must not be preempted or yield during access,
2276 * and (2) the thread must not migrate CPUs without switching which
2277 * cache it accesses. We rely on a critical section to prevent
2278 * preemption and migration. We release the critical section in
2279 * order to acquire the zone mutex if we are unable to free to the
2280 * current cache; when we re-acquire the critical section, we must
2281 * detect and handle migration if it has occurred.
2282 */
2283zfree_restart:
2284 critical_enter();
2285 cpu = curcpu;
2286 cache = &zone->uz_cpu[cpu];
2287
2288zfree_start:
2289 bucket = cache->uc_freebucket;
2290
2291 if (bucket) {
2292 /*
2293 * Do we have room in our bucket? It is OK for this uz count
2294 * check to be slightly out of sync.
2295 */
2296
2297 if (bucket->ub_cnt < bucket->ub_entries) {
2298 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2299 ("uma_zfree: Freeing to non free bucket index."));
2300 bucket->ub_bucket[bucket->ub_cnt] = item;
2301 bucket->ub_cnt++;
2302 cache->uc_frees++;
2303 critical_exit();
2304 return;
2305 } else if (cache->uc_allocbucket) {
2306#ifdef UMA_DEBUG_ALLOC
2307 printf("uma_zfree: Swapping buckets.\n");
2308#endif
2309 /*
2310 * We have run out of space in our freebucket.
2311 * See if we can switch with our alloc bucket.
2312 */
2313 if (cache->uc_allocbucket->ub_cnt <
2314 cache->uc_freebucket->ub_cnt) {
2315 bucket = cache->uc_freebucket;
2316 cache->uc_freebucket = cache->uc_allocbucket;
2317 cache->uc_allocbucket = bucket;
2318 goto zfree_start;
2319 }
2320 }
2321 }
2322 /*
2323 * We can get here for two reasons:
2324 *
2325 * 1) The buckets are NULL
2326 * 2) The alloc and free buckets are both somewhat full.
2327 *
2328 * We must go back the zone, which requires acquiring the zone lock,
2329 * which in turn means we must release and re-acquire the critical
2330 * section. Since the critical section is released, we may be
2331 * preempted or migrate. As such, make sure not to maintain any
2332 * thread-local state specific to the cache from prior to releasing
2333 * the critical section.
2334 */
2335 critical_exit();
2336 ZONE_LOCK(zone);
2337 critical_enter();
2338 cpu = curcpu;
2339 cache = &zone->uz_cpu[cpu];
2340 if (cache->uc_freebucket != NULL) {
2341 if (cache->uc_freebucket->ub_cnt <
2342 cache->uc_freebucket->ub_entries) {
2343 ZONE_UNLOCK(zone);
2344 goto zfree_start;
2345 }
2346 if (cache->uc_allocbucket != NULL &&
2347 (cache->uc_allocbucket->ub_cnt <
2348 cache->uc_freebucket->ub_cnt)) {
2349 ZONE_UNLOCK(zone);
2350 goto zfree_start;
2351 }
2352 }
2353
2354 bucket = cache->uc_freebucket;
2355 cache->uc_freebucket = NULL;
2356
2357 /* Can we throw this on the zone full list? */
2358 if (bucket != NULL) {
2359#ifdef UMA_DEBUG_ALLOC
2360 printf("uma_zfree: Putting old bucket on the free list.\n");
2361#endif
2362 /* ub_cnt is pointing to the last free item */
2363 KASSERT(bucket->ub_cnt != 0,
2364 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2365 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2366 bucket, ub_link);
2367 }
2368 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2369 LIST_REMOVE(bucket, ub_link);
2370 ZONE_UNLOCK(zone);
2371 cache->uc_freebucket = bucket;
2372 goto zfree_start;
2373 }
2374 /* We are no longer associated with this CPU. */
2375 critical_exit();
2376
2377 /* And the zone.. */
2378 ZONE_UNLOCK(zone);
2379
2380#ifdef UMA_DEBUG_ALLOC
2381 printf("uma_zfree: Allocating new free bucket.\n");
2382#endif
2383 bflags = M_NOWAIT;
2384
2385 if (keg->uk_flags & UMA_ZFLAG_CACHEONLY)
2386 bflags |= M_NOVM;
2387 bucket = bucket_alloc(zone->uz_count, bflags);
2388 if (bucket) {
2389 ZONE_LOCK(zone);
2390 LIST_INSERT_HEAD(&zone->uz_free_bucket,
2391 bucket, ub_link);
2392 ZONE_UNLOCK(zone);
2393 goto zfree_restart;
2394 }
2395
2396 /*
2397 * If nothing else caught this, we'll just do an internal free.
2398 */
2399zfree_internal:
2400 uma_zfree_internal(zone, item, udata, SKIP_DTOR);
2401
2402 return;
2403}
2404
2405/*
2406 * Frees an item to an INTERNAL zone or allocates a free bucket
2407 *
2408 * Arguments:
2409 * zone The zone to free to
2410 * item The item we're freeing
2411 * udata User supplied data for the dtor
2412 * skip Skip dtors and finis
2413 */
2414static void
2415uma_zfree_internal(uma_zone_t zone, void *item, void *udata,
2416 enum zfreeskip skip)
2417{
2418 uma_slab_t slab;
2419 uma_slabrefcnt_t slabref;
2420 uma_keg_t keg;
2421 u_int8_t *mem;
2422 u_int8_t freei;
2423
2424 keg = zone->uz_keg;
2425
2426 if (skip < SKIP_DTOR && zone->uz_dtor)
2427 zone->uz_dtor(item, keg->uk_size, udata);
2428 if (skip < SKIP_FINI && zone->uz_fini)
2429 zone->uz_fini(item, keg->uk_size);
2430
2431 ZONE_LOCK(zone);
2432
2433 if (!(keg->uk_flags & UMA_ZONE_MALLOC)) {
2434 mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK));
2435 if (keg->uk_flags & UMA_ZONE_HASH)
2436 slab = hash_sfind(&keg->uk_hash, mem);
2437 else {
2438 mem += keg->uk_pgoff;
2439 slab = (uma_slab_t)mem;
2440 }
2441 } else {
2442 slab = (uma_slab_t)udata;
2443 }
2444
2445 /* Do we need to remove from any lists? */
2446 if (slab->us_freecount+1 == keg->uk_ipers) {
2447 LIST_REMOVE(slab, us_link);
2448 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2449 } else if (slab->us_freecount == 0) {
2450 LIST_REMOVE(slab, us_link);
2451 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2452 }
2453
2454 /* Slab management stuff */
2455 freei = ((unsigned long)item - (unsigned long)slab->us_data)
2456 / keg->uk_rsize;
2457
2458#ifdef INVARIANTS
2459 if (!skip)
2460 uma_dbg_free(zone, slab, item);
2461#endif
2462
2463 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2464 slabref = (uma_slabrefcnt_t)slab;
2465 slabref->us_freelist[freei].us_item = slab->us_firstfree;
2466 } else {
2467 slab->us_freelist[freei].us_item = slab->us_firstfree;
2468 }
2469 slab->us_firstfree = freei;
2470 slab->us_freecount++;
2471
2472 /* Zone statistics */
2473 keg->uk_free++;
2474 zone->uz_frees++;
2475
2476 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2477 if (keg->uk_pages < keg->uk_maxpages)
2478 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2479
2480 /* We can handle one more allocation */
2481 wakeup_one(keg);
2482 }
2483
2484 ZONE_UNLOCK(zone);
2485}
2486
2487/* See uma.h */
2488void
2489uma_zone_set_max(uma_zone_t zone, int nitems)
2490{
2491 uma_keg_t keg;
2492
2493 keg = zone->uz_keg;
2494 ZONE_LOCK(zone);
2495 if (keg->uk_ppera > 1)
2496 keg->uk_maxpages = nitems * keg->uk_ppera;
2497 else
2498 keg->uk_maxpages = nitems / keg->uk_ipers;
2499
2500 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2501 keg->uk_maxpages++;
2502
2503 ZONE_UNLOCK(zone);
2504}
2505
2506/* See uma.h */
2507void
2508uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2509{
2510 ZONE_LOCK(zone);
2511 KASSERT(zone->uz_keg->uk_pages == 0,
2512 ("uma_zone_set_init on non-empty keg"));
2513 zone->uz_keg->uk_init = uminit;
2514 ZONE_UNLOCK(zone);
2515}
2516
2517/* See uma.h */
2518void
2519uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2520{
2521 ZONE_LOCK(zone);
2522 KASSERT(zone->uz_keg->uk_pages == 0,
2523 ("uma_zone_set_fini on non-empty keg"));
2524 zone->uz_keg->uk_fini = fini;
2525 ZONE_UNLOCK(zone);
2526}
2527
2528/* See uma.h */
2529void
2530uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2531{
2532 ZONE_LOCK(zone);
2533 KASSERT(zone->uz_keg->uk_pages == 0,
2534 ("uma_zone_set_zinit on non-empty keg"));
2535 zone->uz_init = zinit;
2536 ZONE_UNLOCK(zone);
2537}
2538
2539/* See uma.h */
2540void
2541uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2542{
2543 ZONE_LOCK(zone);
2544 KASSERT(zone->uz_keg->uk_pages == 0,
2545 ("uma_zone_set_zfini on non-empty keg"));
2546 zone->uz_fini = zfini;
2547 ZONE_UNLOCK(zone);
2548}
2549
2550/* See uma.h */
2551/* XXX uk_freef is not actually used with the zone locked */
2552void
2553uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2554{
2555 ZONE_LOCK(zone);
2556 zone->uz_keg->uk_freef = freef;
2557 ZONE_UNLOCK(zone);
2558}
2559
2560/* See uma.h */
2561/* XXX uk_allocf is not actually used with the zone locked */
2562void
2563uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2564{
2565 ZONE_LOCK(zone);
2566 zone->uz_keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2567 zone->uz_keg->uk_allocf = allocf;
2568 ZONE_UNLOCK(zone);
2569}
2570
2571/* See uma.h */
2572int
2573uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count)
2574{
2575 uma_keg_t keg;
2576 vm_offset_t kva;
2577 int pages;
2578
2579 keg = zone->uz_keg;
2580 pages = count / keg->uk_ipers;
2581
2582 if (pages * keg->uk_ipers < count)
2583 pages++;
2584
2585 kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2586
2587 if (kva == 0)
2588 return (0);
2589 if (obj == NULL) {
2590 obj = vm_object_allocate(OBJT_DEFAULT,
2591 pages);
2592 } else {
2593 VM_OBJECT_LOCK_INIT(obj, "uma object");
2594 _vm_object_allocate(OBJT_DEFAULT,
2595 pages, obj);
2596 }
2597 ZONE_LOCK(zone);
2598 keg->uk_kva = kva;
2599 keg->uk_obj = obj;
2600 keg->uk_maxpages = pages;
2601 keg->uk_allocf = obj_alloc;
2602 keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
2603 ZONE_UNLOCK(zone);
2604 return (1);
2605}
2606
2607/* See uma.h */
2608void
2609uma_prealloc(uma_zone_t zone, int items)
2610{
2611 int slabs;
2612 uma_slab_t slab;
2613 uma_keg_t keg;
2614
2615 keg = zone->uz_keg;
2616 ZONE_LOCK(zone);
2617 slabs = items / keg->uk_ipers;
2618 if (slabs * keg->uk_ipers < items)
2619 slabs++;
2620 while (slabs > 0) {
2621 slab = slab_zalloc(zone, M_WAITOK);
2622 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2623 slabs--;
2624 }
2625 ZONE_UNLOCK(zone);
2626}
2627
2628/* See uma.h */
2629u_int32_t *
2630uma_find_refcnt(uma_zone_t zone, void *item)
2631{
2632 uma_slabrefcnt_t slabref;
2633 uma_keg_t keg;
2634 u_int32_t *refcnt;
2635 int idx;
2636
2637 keg = zone->uz_keg;
2638 slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item &
2639 (~UMA_SLAB_MASK));
2640 KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT,
2641 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
2642 idx = ((unsigned long)item - (unsigned long)slabref->us_data)
2643 / keg->uk_rsize;
2644 refcnt = &slabref->us_freelist[idx].us_refcnt;
2645 return refcnt;
2646}
2647
2648/* See uma.h */
2649void
2650uma_reclaim(void)
2651{
2652#ifdef UMA_DEBUG
2653 printf("UMA: vm asked us to release pages!\n");
2654#endif
2655 bucket_enable();
2656 zone_foreach(zone_drain);
2657 /*
2658 * Some slabs may have been freed but this zone will be visited early
2659 * we visit again so that we can free pages that are empty once other
2660 * zones are drained. We have to do the same for buckets.
2661 */
2662 zone_drain(slabzone);
2663 zone_drain(slabrefzone);
2664 bucket_zone_drain();
2665}
2666
2667void *
2668uma_large_malloc(int size, int wait)
2669{
2670 void *mem;
2671 uma_slab_t slab;
2672 u_int8_t flags;
2673
2674 slab = uma_zalloc_internal(slabzone, NULL, wait);
2675 if (slab == NULL)
2676 return (NULL);
2677 mem = page_alloc(NULL, size, &flags, wait);
2678 if (mem) {
2679 vsetslab((vm_offset_t)mem, slab);
2680 slab->us_data = mem;
2681 slab->us_flags = flags | UMA_SLAB_MALLOC;
2682 slab->us_size = size;
2683 } else {
2684 uma_zfree_internal(slabzone, slab, NULL, SKIP_NONE);
2685 }
2686
2687 return (mem);
2688}
2689
2690void
2691uma_large_free(uma_slab_t slab)
2692{
2693 vsetobj((vm_offset_t)slab->us_data, kmem_object);
2694 page_free(slab->us_data, slab->us_size, slab->us_flags);
2695 uma_zfree_internal(slabzone, slab, NULL, SKIP_NONE);
2696}
2697
2698void
2699uma_print_stats(void)
2700{
2701 zone_foreach(uma_print_zone);
2702}
2703
2704static void
2705slab_print(uma_slab_t slab)
2706{
2707 printf("slab: keg %p, data %p, freecount %d, firstfree %d\n",
2708 slab->us_keg, slab->us_data, slab->us_freecount,
2709 slab->us_firstfree);
2710}
2711
2712static void
2713cache_print(uma_cache_t cache)
2714{
2715 printf("alloc: %p(%d), free: %p(%d)\n",
2716 cache->uc_allocbucket,
2717 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
2718 cache->uc_freebucket,
2719 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
2720}
2721
2722void
2723uma_print_zone(uma_zone_t zone)
2724{
2725 uma_cache_t cache;
2726 uma_keg_t keg;
2727 uma_slab_t slab;
2728 int i;
2729
2730 keg = zone->uz_keg;
2731 printf("%s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n",
2732 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
2733 keg->uk_ipers, keg->uk_ppera,
2734 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
2735 printf("Part slabs:\n");
2736 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
2737 slab_print(slab);
2738 printf("Free slabs:\n");
2739 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
2740 slab_print(slab);
2741 printf("Full slabs:\n");
2742 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
2743 slab_print(slab);
2744 for (i = 0; i <= mp_maxid; i++) {
2745 if (CPU_ABSENT(i))
2746 continue;
2747 cache = &zone->uz_cpu[i];
2748 printf("CPU %d Cache:\n", i);
2749 cache_print(cache);
2750 }
2751}
2752
2753/*
2754 * Sysctl handler for vm.zone
2755 *
2756 * stolen from vm_zone.c
2757 */
2758static int
2759sysctl_vm_zone(SYSCTL_HANDLER_ARGS)
2760{
2761 int error, len, cnt;
2762 const int linesize = 128; /* conservative */
2763 int totalfree;
2764 char *tmpbuf, *offset;
2765 uma_zone_t z;
2766 uma_keg_t zk;
2767 char *p;
2768 int cpu;
2769 int cachefree;
2770 uma_bucket_t bucket;
2771 uma_cache_t cache;
2772 u_int64_t alloc;
2773
2774 cnt = 0;
2775 mtx_lock(&uma_mtx);
2776 LIST_FOREACH(zk, &uma_kegs, uk_link) {
2777 LIST_FOREACH(z, &zk->uk_zones, uz_link)
2778 cnt++;
2779 }
2780 mtx_unlock(&uma_mtx);
2781 MALLOC(tmpbuf, char *, (cnt == 0 ? 1 : cnt) * linesize,
2782 M_TEMP, M_WAITOK);
2783 len = snprintf(tmpbuf, linesize,
2784 "\nITEM SIZE LIMIT USED FREE REQUESTS\n\n");
2785 if (cnt == 0)
2786 tmpbuf[len - 1] = '\0';
2787 error = SYSCTL_OUT(req, tmpbuf, cnt == 0 ? len-1 : len);
2788 if (error || cnt == 0)
2789 goto out;
2790 offset = tmpbuf;
2791 mtx_lock(&uma_mtx);
2792 LIST_FOREACH(zk, &uma_kegs, uk_link) {
2793 LIST_FOREACH(z, &zk->uk_zones, uz_link) {
2794 if (cnt == 0) /* list may have changed size */
2795 break;
2796 ZONE_LOCK(z);
2797 cachefree = 0;
2798 alloc = 0;
2799 if (!(zk->uk_flags & UMA_ZFLAG_INTERNAL)) {
2800 for (cpu = 0; cpu <= mp_maxid; cpu++) {
2801 if (CPU_ABSENT(cpu))
2802 continue;
2803 cache = &z->uz_cpu[cpu];
2804 if (cache->uc_allocbucket != NULL)
2805 cachefree += cache->uc_allocbucket->ub_cnt;
2806 if (cache->uc_freebucket != NULL)
2807 cachefree += cache->uc_freebucket->ub_cnt;
2808 alloc += cache->uc_allocs;
2809 }
2810 }
2811 alloc += z->uz_allocs;
2812
2813 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link) {
2814 cachefree += bucket->ub_cnt;
2815 }
2816 totalfree = zk->uk_free + cachefree;
2817 len = snprintf(offset, linesize,
2818 "%-12.12s %6.6u, %8.8u, %6.6u, %6.6u, %8.8llu\n",
2819 z->uz_name, zk->uk_size,
2820 zk->uk_maxpages * zk->uk_ipers,
2821 (zk->uk_ipers * (zk->uk_pages / zk->uk_ppera)) - totalfree,
2822 totalfree,
2823 (unsigned long long)alloc);
2824 ZONE_UNLOCK(z);
2825 for (p = offset + 12; p > offset && *p == ' '; --p)
2826 /* nothing */ ;
2827 p[1] = ':';
2828 cnt--;
2829 offset += len;
2830 }
2831 }
2832 mtx_unlock(&uma_mtx);
2833 *offset++ = '\0';
2834 error = SYSCTL_OUT(req, tmpbuf, offset - tmpbuf);
2835out:
2836 FREE(tmpbuf, M_TEMP);
2837 return (error);
2838}