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 147996 2005-07-14 16:35:13Z 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/sbuf.h>
74#include <sys/smp.h>
75#include <sys/vmmeter.h>
76
77#include <vm/vm.h>
78#include <vm/vm_object.h>
79#include <vm/vm_page.h>
80#include <vm/vm_param.h>
81#include <vm/vm_map.h>
82#include <vm/vm_kern.h>
83#include <vm/vm_extern.h>
84#include <vm/uma.h>
85#include <vm/uma_int.h>
86#include <vm/uma_dbg.h>
87
88#include <machine/vmparam.h>
89
90/*
91 * This is the zone and keg from which all zones are spawned. The idea is that
92 * even the zone & keg heads are allocated from the allocator, so we use the
93 * bss section to bootstrap us.
94 */
95static struct uma_keg masterkeg;
96static struct uma_zone masterzone_k;
97static struct uma_zone masterzone_z;
98static uma_zone_t kegs = &masterzone_k;
99static uma_zone_t zones = &masterzone_z;
100
101/* This is the zone from which all of uma_slab_t's are allocated. */
102static uma_zone_t slabzone;
103static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
104
105/*
106 * The initial hash tables come out of this zone so they can be allocated
107 * prior to malloc coming up.
108 */
109static uma_zone_t hashzone;
110
111static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
112
113/*
114 * Are we allowed to allocate buckets?
115 */
116static int bucketdisable = 1;
117
118/* Linked list of all kegs in the system */
119static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(&uma_kegs);
120
121/* This mutex protects the keg list */
122static struct mtx uma_mtx;
123
124/* Linked list of boot time pages */
125static LIST_HEAD(,uma_slab) uma_boot_pages =
126 LIST_HEAD_INITIALIZER(&uma_boot_pages);
127
128/* Count of free boottime pages */
129static int uma_boot_free = 0;
130
131/* Is the VM done starting up? */
132static int booted = 0;
133
134/* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
135static u_int uma_max_ipers;
136static u_int uma_max_ipers_ref;
137
138/*
139 * This is the handle used to schedule events that need to happen
140 * outside of the allocation fast path.
141 */
142static struct callout uma_callout;
143#define UMA_TIMEOUT 20 /* Seconds for callout interval. */
144
145/*
146 * This structure is passed as the zone ctor arg so that I don't have to create
147 * a special allocation function just for zones.
148 */
149struct uma_zctor_args {
150 char *name;
151 size_t size;
152 uma_ctor ctor;
153 uma_dtor dtor;
154 uma_init uminit;
155 uma_fini fini;
156 uma_keg_t keg;
157 int align;
158 u_int16_t flags;
159};
160
161struct uma_kctor_args {
162 uma_zone_t zone;
163 size_t size;
164 uma_init uminit;
165 uma_fini fini;
166 int align;
167 u_int16_t flags;
168};
169
170struct uma_bucket_zone {
171 uma_zone_t ubz_zone;
172 char *ubz_name;
173 int ubz_entries;
174};
175
176#define BUCKET_MAX 128
177
178struct uma_bucket_zone bucket_zones[] = {
179 { NULL, "16 Bucket", 16 },
180 { NULL, "32 Bucket", 32 },
181 { NULL, "64 Bucket", 64 },
182 { NULL, "128 Bucket", 128 },
183 { NULL, NULL, 0}
184};
185
186#define BUCKET_SHIFT 4
187#define BUCKET_ZONES ((BUCKET_MAX >> BUCKET_SHIFT) + 1)
188
189/*
190 * bucket_size[] maps requested bucket sizes to zones that allocate a bucket
191 * of approximately the right size.
192 */
193static uint8_t bucket_size[BUCKET_ZONES];
194
195enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI };
196
197/* Prototypes.. */
198
199static void *obj_alloc(uma_zone_t, int, u_int8_t *, int);
200static void *page_alloc(uma_zone_t, int, u_int8_t *, int);
201static void *startup_alloc(uma_zone_t, int, u_int8_t *, int);
202static void page_free(void *, int, u_int8_t);
203static uma_slab_t slab_zalloc(uma_zone_t, int);
204static void cache_drain(uma_zone_t);
205static void bucket_drain(uma_zone_t, uma_bucket_t);
206static void bucket_cache_drain(uma_zone_t zone);
207static int keg_ctor(void *, int, void *, int);
208static void keg_dtor(void *, int, void *);
209static int zone_ctor(void *, int, void *, int);
210static void zone_dtor(void *, int, void *);
211static int zero_init(void *, int, int);
212static void zone_small_init(uma_zone_t zone);
213static void zone_large_init(uma_zone_t zone);
214static void zone_foreach(void (*zfunc)(uma_zone_t));
215static void zone_timeout(uma_zone_t zone);
216static int hash_alloc(struct uma_hash *);
217static int hash_expand(struct uma_hash *, struct uma_hash *);
218static void hash_free(struct uma_hash *hash);
219static void uma_timeout(void *);
220static void uma_startup3(void);
221static void *uma_zalloc_internal(uma_zone_t, void *, int);
222static void uma_zfree_internal(uma_zone_t, void *, void *, enum zfreeskip);
223static void bucket_enable(void);
224static void bucket_init(void);
225static uma_bucket_t bucket_alloc(int, int);
226static void bucket_free(uma_bucket_t);
227static void bucket_zone_drain(void);
228static int uma_zalloc_bucket(uma_zone_t zone, int flags);
229static uma_slab_t uma_zone_slab(uma_zone_t zone, int flags);
230static void *uma_slab_alloc(uma_zone_t zone, uma_slab_t slab);
231static void zone_drain(uma_zone_t);
232static uma_zone_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
233 uma_fini fini, int align, u_int16_t flags);
234
235void uma_print_zone(uma_zone_t);
236void uma_print_stats(void);
237static int sysctl_vm_zone(SYSCTL_HANDLER_ARGS);
238static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
239static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
240
241#ifdef WITNESS
242static int nosleepwithlocks = 1;
243SYSCTL_INT(_debug, OID_AUTO, nosleepwithlocks, CTLFLAG_RW, &nosleepwithlocks,
244 0, "Convert M_WAITOK to M_NOWAIT to avoid lock-held-across-sleep paths");
245#else
246static int nosleepwithlocks = 0;
247SYSCTL_INT(_debug, OID_AUTO, nosleepwithlocks, CTLFLAG_RW, &nosleepwithlocks,
248 0, "Convert M_WAITOK to M_NOWAIT to avoid lock-held-across-sleep paths");
249#endif
250SYSCTL_OID(_vm, OID_AUTO, zone, CTLTYPE_STRING|CTLFLAG_RD,
251 NULL, 0, sysctl_vm_zone, "A", "Zone Info");
252SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
253
254SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
255 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
256
257SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
258 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
259
260/*
261 * This routine checks to see whether or not it's safe to enable buckets.
262 */
263
264static void
265bucket_enable(void)
266{
267 if (cnt.v_free_count < cnt.v_free_min)
268 bucketdisable = 1;
269 else
270 bucketdisable = 0;
271}
272
273/*
274 * Initialize bucket_zones, the array of zones of buckets of various sizes.
275 *
276 * For each zone, calculate the memory required for each bucket, consisting
277 * of the header and an array of pointers. Initialize bucket_size[] to point
278 * the range of appropriate bucket sizes at the zone.
279 */
280static void
281bucket_init(void)
282{
283 struct uma_bucket_zone *ubz;
284 int i;
285 int j;
286
287 for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) {
288 int size;
289
290 ubz = &bucket_zones[j];
291 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
292 size += sizeof(void *) * ubz->ubz_entries;
293 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
294 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
295 for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT))
296 bucket_size[i >> BUCKET_SHIFT] = j;
297 }
298}
299
300/*
301 * Given a desired number of entries for a bucket, return the zone from which
302 * to allocate the bucket.
303 */
304static struct uma_bucket_zone *
305bucket_zone_lookup(int entries)
306{
307 int idx;
308
309 idx = howmany(entries, 1 << BUCKET_SHIFT);
310 return (&bucket_zones[bucket_size[idx]]);
311}
312
313static uma_bucket_t
314bucket_alloc(int entries, int bflags)
315{
316 struct uma_bucket_zone *ubz;
317 uma_bucket_t bucket;
318
319 /*
320 * This is to stop us from allocating per cpu buckets while we're
321 * running out of UMA_BOOT_PAGES. Otherwise, we would exhaust the
322 * boot pages. This also prevents us from allocating buckets in
323 * low memory situations.
324 */
325 if (bucketdisable)
326 return (NULL);
327
328 ubz = bucket_zone_lookup(entries);
329 bucket = uma_zalloc_internal(ubz->ubz_zone, NULL, bflags);
330 if (bucket) {
331#ifdef INVARIANTS
332 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
333#endif
334 bucket->ub_cnt = 0;
335 bucket->ub_entries = ubz->ubz_entries;
336 }
337
338 return (bucket);
339}
340
341static void
342bucket_free(uma_bucket_t bucket)
343{
344 struct uma_bucket_zone *ubz;
345
346 ubz = bucket_zone_lookup(bucket->ub_entries);
347 uma_zfree_internal(ubz->ubz_zone, bucket, NULL, SKIP_NONE);
348}
349
350static void
351bucket_zone_drain(void)
352{
353 struct uma_bucket_zone *ubz;
354
355 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
356 zone_drain(ubz->ubz_zone);
357}
358
359
360/*
361 * Routine called by timeout which is used to fire off some time interval
362 * based calculations. (stats, hash size, etc.)
363 *
364 * Arguments:
365 * arg Unused
366 *
367 * Returns:
368 * Nothing
369 */
370static void
371uma_timeout(void *unused)
372{
373 bucket_enable();
374 zone_foreach(zone_timeout);
375
376 /* Reschedule this event */
377 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
378}
379
380/*
381 * Routine to perform timeout driven calculations. This expands the
382 * hashes and does per cpu statistics aggregation.
383 *
384 * Arguments:
385 * zone The zone to operate on
386 *
387 * Returns:
388 * Nothing
389 */
390static void
391zone_timeout(uma_zone_t zone)
392{
393 uma_keg_t keg;
394 u_int64_t alloc;
395
396 keg = zone->uz_keg;
397 alloc = 0;
398
399 /*
400 * Expand the zone hash table.
401 *
402 * This is done if the number of slabs is larger than the hash size.
403 * What I'm trying to do here is completely reduce collisions. This
404 * may be a little aggressive. Should I allow for two collisions max?
405 */
406 ZONE_LOCK(zone);
407 if (keg->uk_flags & UMA_ZONE_HASH &&
408 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
409 struct uma_hash newhash;
410 struct uma_hash oldhash;
411 int ret;
412
413 /*
414 * This is so involved because allocating and freeing
415 * while the zone lock is held will lead to deadlock.
416 * I have to do everything in stages and check for
417 * races.
418 */
419 newhash = keg->uk_hash;
420 ZONE_UNLOCK(zone);
421 ret = hash_alloc(&newhash);
422 ZONE_LOCK(zone);
423 if (ret) {
424 if (hash_expand(&keg->uk_hash, &newhash)) {
425 oldhash = keg->uk_hash;
426 keg->uk_hash = newhash;
427 } else
428 oldhash = newhash;
429
430 ZONE_UNLOCK(zone);
431 hash_free(&oldhash);
432 ZONE_LOCK(zone);
433 }
434 }
435 ZONE_UNLOCK(zone);
436}
437
438/*
439 * Allocate and zero fill the next sized hash table from the appropriate
440 * backing store.
441 *
442 * Arguments:
443 * hash A new hash structure with the old hash size in uh_hashsize
444 *
445 * Returns:
446 * 1 on sucess and 0 on failure.
447 */
448static int
449hash_alloc(struct uma_hash *hash)
450{
451 int oldsize;
452 int alloc;
453
454 oldsize = hash->uh_hashsize;
455
456 /* We're just going to go to a power of two greater */
457 if (oldsize) {
458 hash->uh_hashsize = oldsize * 2;
459 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
460 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
461 M_UMAHASH, M_NOWAIT);
462 } else {
463 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
464 hash->uh_slab_hash = uma_zalloc_internal(hashzone, NULL,
465 M_WAITOK);
466 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
467 }
468 if (hash->uh_slab_hash) {
469 bzero(hash->uh_slab_hash, alloc);
470 hash->uh_hashmask = hash->uh_hashsize - 1;
471 return (1);
472 }
473
474 return (0);
475}
476
477/*
478 * Expands the hash table for HASH zones. This is done from zone_timeout
479 * to reduce collisions. This must not be done in the regular allocation
480 * path, otherwise, we can recurse on the vm while allocating pages.
481 *
482 * Arguments:
483 * oldhash The hash you want to expand
484 * newhash The hash structure for the new table
485 *
486 * Returns:
487 * Nothing
488 *
489 * Discussion:
490 */
491static int
492hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
493{
494 uma_slab_t slab;
495 int hval;
496 int i;
497
498 if (!newhash->uh_slab_hash)
499 return (0);
500
501 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
502 return (0);
503
504 /*
505 * I need to investigate hash algorithms for resizing without a
506 * full rehash.
507 */
508
509 for (i = 0; i < oldhash->uh_hashsize; i++)
510 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
511 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
512 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
513 hval = UMA_HASH(newhash, slab->us_data);
514 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
515 slab, us_hlink);
516 }
517
518 return (1);
519}
520
521/*
522 * Free the hash bucket to the appropriate backing store.
523 *
524 * Arguments:
525 * slab_hash The hash bucket we're freeing
526 * hashsize The number of entries in that hash bucket
527 *
528 * Returns:
529 * Nothing
530 */
531static void
532hash_free(struct uma_hash *hash)
533{
534 if (hash->uh_slab_hash == NULL)
535 return;
536 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
537 uma_zfree_internal(hashzone,
538 hash->uh_slab_hash, NULL, SKIP_NONE);
539 else
540 free(hash->uh_slab_hash, M_UMAHASH);
541}
542
543/*
544 * Frees all outstanding items in a bucket
545 *
546 * Arguments:
547 * zone The zone to free to, must be unlocked.
548 * bucket The free/alloc bucket with items, cpu queue must be locked.
549 *
550 * Returns:
551 * Nothing
552 */
553
554static void
555bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
556{
557 uma_slab_t slab;
558 int mzone;
559 void *item;
560
561 if (bucket == NULL)
562 return;
563
564 slab = NULL;
565 mzone = 0;
566
567 /* We have to lookup the slab again for malloc.. */
568 if (zone->uz_keg->uk_flags & UMA_ZONE_MALLOC)
569 mzone = 1;
570
571 while (bucket->ub_cnt > 0) {
572 bucket->ub_cnt--;
573 item = bucket->ub_bucket[bucket->ub_cnt];
574#ifdef INVARIANTS
575 bucket->ub_bucket[bucket->ub_cnt] = NULL;
576 KASSERT(item != NULL,
577 ("bucket_drain: botched ptr, item is NULL"));
578#endif
579 /*
580 * This is extremely inefficient. The slab pointer was passed
581 * to uma_zfree_arg, but we lost it because the buckets don't
582 * hold them. This will go away when free() gets a size passed
583 * to it.
584 */
585 if (mzone)
586 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
587 uma_zfree_internal(zone, item, slab, SKIP_DTOR);
588 }
589}
590
591/*
592 * Drains the per cpu caches for a zone.
593 *
594 * NOTE: This may only be called while the zone is being turn down, and not
595 * during normal operation. This is necessary in order that we do not have
596 * to migrate CPUs to drain the per-CPU caches.
597 *
598 * Arguments:
599 * zone The zone to drain, must be unlocked.
600 *
601 * Returns:
602 * Nothing
603 */
604static void
605cache_drain(uma_zone_t zone)
606{
607 uma_cache_t cache;
608 int cpu;
609
610 /*
611 * XXX: It is safe to not lock the per-CPU caches, because we're
612 * tearing down the zone anyway. I.e., there will be no further use
613 * of the caches at this point.
614 *
615 * XXX: It would good to be able to assert that the zone is being
616 * torn down to prevent improper use of cache_drain().
617 *
618 * XXX: We lock the zone before passing into bucket_cache_drain() as
619 * it is used elsewhere. Should the tear-down path be made special
620 * there in some form?
621 */
622 for (cpu = 0; cpu <= mp_maxid; cpu++) {
623 if (CPU_ABSENT(cpu))
624 continue;
625 cache = &zone->uz_cpu[cpu];
626 bucket_drain(zone, cache->uc_allocbucket);
627 bucket_drain(zone, cache->uc_freebucket);
628 if (cache->uc_allocbucket != NULL)
629 bucket_free(cache->uc_allocbucket);
630 if (cache->uc_freebucket != NULL)
631 bucket_free(cache->uc_freebucket);
632 cache->uc_allocbucket = cache->uc_freebucket = NULL;
633 }
634 ZONE_LOCK(zone);
635 bucket_cache_drain(zone);
636 ZONE_UNLOCK(zone);
637}
638
639/*
640 * Drain the cached buckets from a zone. Expects a locked zone on entry.
641 */
642static void
643bucket_cache_drain(uma_zone_t zone)
644{
645 uma_bucket_t bucket;
646
647 /*
648 * Drain the bucket queues and free the buckets, we just keep two per
649 * cpu (alloc/free).
650 */
651 while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
652 LIST_REMOVE(bucket, ub_link);
653 ZONE_UNLOCK(zone);
654 bucket_drain(zone, bucket);
655 bucket_free(bucket);
656 ZONE_LOCK(zone);
657 }
658
659 /* Now we do the free queue.. */
660 while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
661 LIST_REMOVE(bucket, ub_link);
662 bucket_free(bucket);
663 }
664}
665
666/*
667 * Frees pages from a zone back to the system. This is done on demand from
668 * the pageout daemon.
669 *
670 * Arguments:
671 * zone The zone to free pages from
672 * all Should we drain all items?
673 *
674 * Returns:
675 * Nothing.
676 */
677static void
678zone_drain(uma_zone_t zone)
679{
680 struct slabhead freeslabs = { 0 };
681 uma_keg_t keg;
682 uma_slab_t slab;
683 uma_slab_t n;
684 u_int8_t flags;
685 u_int8_t *mem;
686 int i;
687
688 keg = zone->uz_keg;
689
690 /*
691 * We don't want to take pages from statically allocated zones at this
692 * time
693 */
694 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
695 return;
696
697 ZONE_LOCK(zone);
698
699#ifdef UMA_DEBUG
700 printf("%s free items: %u\n", zone->uz_name, keg->uk_free);
701#endif
702 bucket_cache_drain(zone);
703 if (keg->uk_free == 0)
704 goto finished;
705
706 slab = LIST_FIRST(&keg->uk_free_slab);
707 while (slab) {
708 n = LIST_NEXT(slab, us_link);
709
710 /* We have no where to free these to */
711 if (slab->us_flags & UMA_SLAB_BOOT) {
712 slab = n;
713 continue;
714 }
715
716 LIST_REMOVE(slab, us_link);
717 keg->uk_pages -= keg->uk_ppera;
718 keg->uk_free -= keg->uk_ipers;
719
720 if (keg->uk_flags & UMA_ZONE_HASH)
721 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
722
723 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
724
725 slab = n;
726 }
727finished:
728 ZONE_UNLOCK(zone);
729
730 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
731 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
732 if (keg->uk_fini)
733 for (i = 0; i < keg->uk_ipers; i++)
734 keg->uk_fini(
735 slab->us_data + (keg->uk_rsize * i),
736 keg->uk_size);
737 flags = slab->us_flags;
738 mem = slab->us_data;
739
740 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
741 (keg->uk_flags & UMA_ZONE_REFCNT)) {
742 vm_object_t obj;
743
744 if (flags & UMA_SLAB_KMEM)
745 obj = kmem_object;
746 else
747 obj = NULL;
748 for (i = 0; i < keg->uk_ppera; i++)
749 vsetobj((vm_offset_t)mem + (i * PAGE_SIZE),
750 obj);
751 }
752 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
753 uma_zfree_internal(keg->uk_slabzone, slab, NULL,
754 SKIP_NONE);
755#ifdef UMA_DEBUG
756 printf("%s: Returning %d bytes.\n",
757 zone->uz_name, UMA_SLAB_SIZE * keg->uk_ppera);
758#endif
759 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags);
760 }
761}
762
763/*
764 * Allocate a new slab for a zone. This does not insert the slab onto a list.
765 *
766 * Arguments:
767 * zone The zone to allocate slabs for
768 * wait Shall we wait?
769 *
770 * Returns:
771 * The slab that was allocated or NULL if there is no memory and the
772 * caller specified M_NOWAIT.
773 */
774static uma_slab_t
775slab_zalloc(uma_zone_t zone, int wait)
776{
777 uma_slabrefcnt_t slabref;
778 uma_slab_t slab;
779 uma_keg_t keg;
780 u_int8_t *mem;
781 u_int8_t flags;
782 int i;
783
784 slab = NULL;
785 keg = zone->uz_keg;
786
787#ifdef UMA_DEBUG
788 printf("slab_zalloc: Allocating a new slab for %s\n", zone->uz_name);
789#endif
790 ZONE_UNLOCK(zone);
791
792 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
793 slab = uma_zalloc_internal(keg->uk_slabzone, NULL, wait);
794 if (slab == NULL) {
795 ZONE_LOCK(zone);
796 return NULL;
797 }
798 }
799
800 /*
801 * This reproduces the old vm_zone behavior of zero filling pages the
802 * first time they are added to a zone.
803 *
804 * Malloced items are zeroed in uma_zalloc.
805 */
806
807 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
808 wait |= M_ZERO;
809 else
810 wait &= ~M_ZERO;
811
812 mem = keg->uk_allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE,
813 &flags, wait);
814 if (mem == NULL) {
815 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
816 uma_zfree_internal(keg->uk_slabzone, slab, NULL,
817 SKIP_NONE);
818 ZONE_LOCK(zone);
819 return (NULL);
820 }
821
822 /* Point the slab into the allocated memory */
823 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
824 slab = (uma_slab_t )(mem + keg->uk_pgoff);
825
826 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
827 (keg->uk_flags & UMA_ZONE_REFCNT))
828 for (i = 0; i < keg->uk_ppera; i++)
829 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
830
831 slab->us_keg = keg;
832 slab->us_data = mem;
833 slab->us_freecount = keg->uk_ipers;
834 slab->us_firstfree = 0;
835 slab->us_flags = flags;
836
837 if (keg->uk_flags & UMA_ZONE_REFCNT) {
838 slabref = (uma_slabrefcnt_t)slab;
839 for (i = 0; i < keg->uk_ipers; i++) {
840 slabref->us_freelist[i].us_refcnt = 0;
841 slabref->us_freelist[i].us_item = i+1;
842 }
843 } else {
844 for (i = 0; i < keg->uk_ipers; i++)
845 slab->us_freelist[i].us_item = i+1;
846 }
847
848 if (keg->uk_init != NULL) {
849 for (i = 0; i < keg->uk_ipers; i++)
850 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
851 keg->uk_size, wait) != 0)
852 break;
853 if (i != keg->uk_ipers) {
854 if (keg->uk_fini != NULL) {
855 for (i--; i > -1; i--)
856 keg->uk_fini(slab->us_data +
857 (keg->uk_rsize * i),
858 keg->uk_size);
859 }
860 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
861 (keg->uk_flags & UMA_ZONE_REFCNT)) {
862 vm_object_t obj;
863
864 if (flags & UMA_SLAB_KMEM)
865 obj = kmem_object;
866 else
867 obj = NULL;
868 for (i = 0; i < keg->uk_ppera; i++)
869 vsetobj((vm_offset_t)mem +
870 (i * PAGE_SIZE), obj);
871 }
872 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
873 uma_zfree_internal(keg->uk_slabzone, slab,
874 NULL, SKIP_NONE);
875 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera,
876 flags);
877 ZONE_LOCK(zone);
878 return (NULL);
879 }
880 }
881 ZONE_LOCK(zone);
882
883 if (keg->uk_flags & UMA_ZONE_HASH)
884 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
885
886 keg->uk_pages += keg->uk_ppera;
887 keg->uk_free += keg->uk_ipers;
888
889 return (slab);
890}
891
892/*
893 * This function is intended to be used early on in place of page_alloc() so
894 * that we may use the boot time page cache to satisfy allocations before
895 * the VM is ready.
896 */
897static void *
898startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
899{
900 uma_keg_t keg;
901
902 keg = zone->uz_keg;
903
904 /*
905 * Check our small startup cache to see if it has pages remaining.
906 */
907 mtx_lock(&uma_mtx);
908 if (uma_boot_free != 0) {
909 uma_slab_t tmps;
910
911 tmps = LIST_FIRST(&uma_boot_pages);
912 LIST_REMOVE(tmps, us_link);
913 uma_boot_free--;
914 mtx_unlock(&uma_mtx);
915 *pflag = tmps->us_flags;
916 return (tmps->us_data);
917 }
918 mtx_unlock(&uma_mtx);
919 if (booted == 0)
920 panic("UMA: Increase UMA_BOOT_PAGES");
921 /*
922 * Now that we've booted reset these users to their real allocator.
923 */
924#ifdef UMA_MD_SMALL_ALLOC
925 keg->uk_allocf = uma_small_alloc;
926#else
927 keg->uk_allocf = page_alloc;
928#endif
929 return keg->uk_allocf(zone, bytes, pflag, wait);
930}
931
932/*
933 * Allocates a number of pages from the system
934 *
935 * Arguments:
936 * zone Unused
937 * bytes The number of bytes requested
938 * wait Shall we wait?
939 *
940 * Returns:
941 * A pointer to the alloced memory or possibly
942 * NULL if M_NOWAIT is set.
943 */
944static void *
945page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
946{
947 void *p; /* Returned page */
948
949 *pflag = UMA_SLAB_KMEM;
950 p = (void *) kmem_malloc(kmem_map, bytes, wait);
951
952 return (p);
953}
954
955/*
956 * Allocates a number of pages from within an object
957 *
958 * Arguments:
959 * zone Unused
960 * bytes The number of bytes requested
961 * wait Shall we wait?
962 *
963 * Returns:
964 * A pointer to the alloced memory or possibly
965 * NULL if M_NOWAIT is set.
966 */
967static void *
968obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
969{
970 vm_object_t object;
971 vm_offset_t retkva, zkva;
972 vm_page_t p;
973 int pages, startpages;
974
975 object = zone->uz_keg->uk_obj;
976 retkva = 0;
977
978 /*
979 * This looks a little weird since we're getting one page at a time.
980 */
981 VM_OBJECT_LOCK(object);
982 p = TAILQ_LAST(&object->memq, pglist);
983 pages = p != NULL ? p->pindex + 1 : 0;
984 startpages = pages;
985 zkva = zone->uz_keg->uk_kva + pages * PAGE_SIZE;
986 for (; bytes > 0; bytes -= PAGE_SIZE) {
987 p = vm_page_alloc(object, pages,
988 VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED);
989 if (p == NULL) {
990 if (pages != startpages)
991 pmap_qremove(retkva, pages - startpages);
992 while (pages != startpages) {
993 pages--;
994 p = TAILQ_LAST(&object->memq, pglist);
995 vm_page_lock_queues();
996 vm_page_unwire(p, 0);
997 vm_page_free(p);
998 vm_page_unlock_queues();
999 }
1000 retkva = 0;
1001 goto done;
1002 }
1003 pmap_qenter(zkva, &p, 1);
1004 if (retkva == 0)
1005 retkva = zkva;
1006 zkva += PAGE_SIZE;
1007 pages += 1;
1008 }
1009done:
1010 VM_OBJECT_UNLOCK(object);
1011 *flags = UMA_SLAB_PRIV;
1012
1013 return ((void *)retkva);
1014}
1015
1016/*
1017 * Frees a number of pages to the system
1018 *
1019 * Arguments:
1020 * mem A pointer to the memory to be freed
1021 * size The size of the memory being freed
1022 * flags The original p->us_flags field
1023 *
1024 * Returns:
1025 * Nothing
1026 */
1027static void
1028page_free(void *mem, int size, u_int8_t flags)
1029{
1030 vm_map_t map;
1031
1032 if (flags & UMA_SLAB_KMEM)
1033 map = kmem_map;
1034 else
1035 panic("UMA: page_free used with invalid flags %d\n", flags);
1036
1037 kmem_free(map, (vm_offset_t)mem, size);
1038}
1039
1040/*
1041 * Zero fill initializer
1042 *
1043 * Arguments/Returns follow uma_init specifications
1044 */
1045static int
1046zero_init(void *mem, int size, int flags)
1047{
1048 bzero(mem, size);
1049 return (0);
1050}
1051
1052/*
1053 * Finish creating a small uma zone. This calculates ipers, and the zone size.
1054 *
1055 * Arguments
1056 * zone The zone we should initialize
1057 *
1058 * Returns
1059 * Nothing
1060 */
1061static void
1062zone_small_init(uma_zone_t zone)
1063{
1064 uma_keg_t keg;
1065 u_int rsize;
1066 u_int memused;
1067 u_int wastedspace;
1068 u_int shsize;
1069
1070 keg = zone->uz_keg;
1071 KASSERT(keg != NULL, ("Keg is null in zone_small_init"));
1072 rsize = keg->uk_size;
1073
1074 if (rsize < UMA_SMALLEST_UNIT)
1075 rsize = UMA_SMALLEST_UNIT;
1076 if (rsize & keg->uk_align)
1077 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1078
1079 keg->uk_rsize = rsize;
1080 keg->uk_ppera = 1;
1081
1082 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1083 rsize += UMA_FRITMREF_SZ; /* linkage & refcnt */
1084 shsize = sizeof(struct uma_slab_refcnt);
1085 } else {
1086 rsize += UMA_FRITM_SZ; /* Account for linkage */
1087 shsize = sizeof(struct uma_slab);
1088 }
1089
1090 keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize;
1091 KASSERT(keg->uk_ipers != 0, ("zone_small_init: ipers is 0"));
1092 memused = keg->uk_ipers * rsize + shsize;
1093 wastedspace = UMA_SLAB_SIZE - memused;
1094
1095 /*
1096 * We can't do OFFPAGE if we're internal or if we've been
1097 * asked to not go to the VM for buckets. If we do this we
1098 * may end up going to the VM (kmem_map) for slabs which we
1099 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1100 * result of UMA_ZONE_VM, which clearly forbids it.
1101 */
1102 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1103 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1104 return;
1105
1106 if ((wastedspace >= UMA_MAX_WASTE) &&
1107 (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) {
1108 keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize;
1109 KASSERT(keg->uk_ipers <= 255,
1110 ("zone_small_init: keg->uk_ipers too high!"));
1111#ifdef UMA_DEBUG
1112 printf("UMA decided we need offpage slab headers for "
1113 "zone: %s, calculated wastedspace = %d, "
1114 "maximum wasted space allowed = %d, "
1115 "calculated ipers = %d, "
1116 "new wasted space = %d\n", zone->uz_name, wastedspace,
1117 UMA_MAX_WASTE, keg->uk_ipers,
1118 UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize);
1119#endif
1120 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1121 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1122 keg->uk_flags |= UMA_ZONE_HASH;
1123 }
1124}
1125
1126/*
1127 * Finish creating a large (> UMA_SLAB_SIZE) uma zone. Just give in and do
1128 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1129 * more complicated.
1130 *
1131 * Arguments
1132 * zone The zone we should initialize
1133 *
1134 * Returns
1135 * Nothing
1136 */
1137static void
1138zone_large_init(uma_zone_t zone)
1139{
1140 uma_keg_t keg;
1141 int pages;
1142
1143 keg = zone->uz_keg;
1144
1145 KASSERT(keg != NULL, ("Keg is null in zone_large_init"));
1146 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1147 ("zone_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY zone"));
1148
1149 pages = keg->uk_size / UMA_SLAB_SIZE;
1150
1151 /* Account for remainder */
1152 if ((pages * UMA_SLAB_SIZE) < keg->uk_size)
1153 pages++;
1154
1155 keg->uk_ppera = pages;
1156 keg->uk_ipers = 1;
1157
1158 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1159 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1160 keg->uk_flags |= UMA_ZONE_HASH;
1161
1162 keg->uk_rsize = keg->uk_size;
1163}
1164
1165/*
1166 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1167 * the keg onto the global keg list.
1168 *
1169 * Arguments/Returns follow uma_ctor specifications
1170 * udata Actually uma_kctor_args
1171 */
1172static int
1173keg_ctor(void *mem, int size, void *udata, int flags)
1174{
1175 struct uma_kctor_args *arg = udata;
1176 uma_keg_t keg = mem;
1177 uma_zone_t zone;
1178
1179 bzero(keg, size);
1180 keg->uk_size = arg->size;
1181 keg->uk_init = arg->uminit;
1182 keg->uk_fini = arg->fini;
1183 keg->uk_align = arg->align;
1184 keg->uk_free = 0;
1185 keg->uk_pages = 0;
1186 keg->uk_flags = arg->flags;
1187 keg->uk_allocf = page_alloc;
1188 keg->uk_freef = page_free;
1189 keg->uk_recurse = 0;
1190 keg->uk_slabzone = NULL;
1191
1192 /*
1193 * The master zone is passed to us at keg-creation time.
1194 */
1195 zone = arg->zone;
1196 zone->uz_keg = keg;
1197
1198 if (arg->flags & UMA_ZONE_VM)
1199 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1200
1201 if (arg->flags & UMA_ZONE_ZINIT)
1202 keg->uk_init = zero_init;
1203
1204 /*
1205 * The +UMA_FRITM_SZ added to uk_size is to account for the
1206 * linkage that is added to the size in zone_small_init(). If
1207 * we don't account for this here then we may end up in
1208 * zone_small_init() with a calculated 'ipers' of 0.
1209 */
1210 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1211 if ((keg->uk_size+UMA_FRITMREF_SZ) >
1212 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)))
1213 zone_large_init(zone);
1214 else
1215 zone_small_init(zone);
1216 } else {
1217 if ((keg->uk_size+UMA_FRITM_SZ) >
1218 (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1219 zone_large_init(zone);
1220 else
1221 zone_small_init(zone);
1222 }
1223
1224 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1225 if (keg->uk_flags & UMA_ZONE_REFCNT)
1226 keg->uk_slabzone = slabrefzone;
1227 else
1228 keg->uk_slabzone = slabzone;
1229 }
1230
1231 /*
1232 * If we haven't booted yet we need allocations to go through the
1233 * startup cache until the vm is ready.
1234 */
1235 if (keg->uk_ppera == 1) {
1236#ifdef UMA_MD_SMALL_ALLOC
1237 keg->uk_allocf = uma_small_alloc;
1238 keg->uk_freef = uma_small_free;
1239#endif
1240 if (booted == 0)
1241 keg->uk_allocf = startup_alloc;
1242 }
1243
1244 /*
1245 * Initialize keg's lock (shared among zones) through
1246 * Master zone
1247 */
1248 zone->uz_lock = &keg->uk_lock;
1249 if (arg->flags & UMA_ZONE_MTXCLASS)
1250 ZONE_LOCK_INIT(zone, 1);
1251 else
1252 ZONE_LOCK_INIT(zone, 0);
1253
1254 /*
1255 * If we're putting the slab header in the actual page we need to
1256 * figure out where in each page it goes. This calculates a right
1257 * justified offset into the memory on an ALIGN_PTR boundary.
1258 */
1259 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1260 u_int totsize;
1261
1262 /* Size of the slab struct and free list */
1263 if (keg->uk_flags & UMA_ZONE_REFCNT)
1264 totsize = sizeof(struct uma_slab_refcnt) +
1265 keg->uk_ipers * UMA_FRITMREF_SZ;
1266 else
1267 totsize = sizeof(struct uma_slab) +
1268 keg->uk_ipers * UMA_FRITM_SZ;
1269
1270 if (totsize & UMA_ALIGN_PTR)
1271 totsize = (totsize & ~UMA_ALIGN_PTR) +
1272 (UMA_ALIGN_PTR + 1);
1273 keg->uk_pgoff = UMA_SLAB_SIZE - totsize;
1274
1275 if (keg->uk_flags & UMA_ZONE_REFCNT)
1276 totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt)
1277 + keg->uk_ipers * UMA_FRITMREF_SZ;
1278 else
1279 totsize = keg->uk_pgoff + sizeof(struct uma_slab)
1280 + keg->uk_ipers * UMA_FRITM_SZ;
1281
1282 /*
1283 * The only way the following is possible is if with our
1284 * UMA_ALIGN_PTR adjustments we are now bigger than
1285 * UMA_SLAB_SIZE. I haven't checked whether this is
1286 * mathematically possible for all cases, so we make
1287 * sure here anyway.
1288 */
1289 if (totsize > UMA_SLAB_SIZE) {
1290 printf("zone %s ipers %d rsize %d size %d\n",
1291 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1292 keg->uk_size);
1293 panic("UMA slab won't fit.\n");
1294 }
1295 }
1296
1297 if (keg->uk_flags & UMA_ZONE_HASH)
1298 hash_alloc(&keg->uk_hash);
1299
1300#ifdef UMA_DEBUG
1301 printf("%s(%p) size = %d ipers = %d ppera = %d pgoff = %d\n",
1302 zone->uz_name, zone,
1303 keg->uk_size, keg->uk_ipers,
1304 keg->uk_ppera, keg->uk_pgoff);
1305#endif
1306
1307 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1308
1309 mtx_lock(&uma_mtx);
1310 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1311 mtx_unlock(&uma_mtx);
1312 return (0);
1313}
1314
1315/*
1316 * Zone header ctor. This initializes all fields, locks, etc.
1317 *
1318 * Arguments/Returns follow uma_ctor specifications
1319 * udata Actually uma_zctor_args
1320 */
1321
1322static int
1323zone_ctor(void *mem, int size, void *udata, int flags)
1324{
1325 struct uma_zctor_args *arg = udata;
1326 uma_zone_t zone = mem;
1327 uma_zone_t z;
1328 uma_keg_t keg;
1329
1330 bzero(zone, size);
1331 zone->uz_name = arg->name;
1332 zone->uz_ctor = arg->ctor;
1333 zone->uz_dtor = arg->dtor;
1334 zone->uz_init = NULL;
1335 zone->uz_fini = NULL;
1336 zone->uz_allocs = 0;
1337 zone->uz_frees = 0;
1338 zone->uz_fills = zone->uz_count = 0;
1339
1340 if (arg->flags & UMA_ZONE_SECONDARY) {
1341 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1342 keg = arg->keg;
1343 zone->uz_keg = keg;
1344 zone->uz_init = arg->uminit;
1345 zone->uz_fini = arg->fini;
1346 zone->uz_lock = &keg->uk_lock;
1347 mtx_lock(&uma_mtx);
1348 ZONE_LOCK(zone);
1349 keg->uk_flags |= UMA_ZONE_SECONDARY;
1350 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1351 if (LIST_NEXT(z, uz_link) == NULL) {
1352 LIST_INSERT_AFTER(z, zone, uz_link);
1353 break;
1354 }
1355 }
1356 ZONE_UNLOCK(zone);
1357 mtx_unlock(&uma_mtx);
1358 } else if (arg->keg == NULL) {
1359 if (uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1360 arg->align, arg->flags) == NULL)
1361 return (ENOMEM);
1362 } else {
1363 struct uma_kctor_args karg;
1364 int error;
1365
1366 /* We should only be here from uma_startup() */
1367 karg.size = arg->size;
1368 karg.uminit = arg->uminit;
1369 karg.fini = arg->fini;
1370 karg.align = arg->align;
1371 karg.flags = arg->flags;
1372 karg.zone = zone;
1373 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1374 flags);
1375 if (error)
1376 return (error);
1377 }
1378 keg = zone->uz_keg;
1379 zone->uz_lock = &keg->uk_lock;
1380
1381 /*
1382 * Some internal zones don't have room allocated for the per cpu
1383 * caches. If we're internal, bail out here.
1384 */
1385 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1386 KASSERT((keg->uk_flags & UMA_ZONE_SECONDARY) == 0,
1387 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1388 return (0);
1389 }
1390
1391 if (keg->uk_flags & UMA_ZONE_MAXBUCKET)
1392 zone->uz_count = BUCKET_MAX;
1393 else if (keg->uk_ipers <= BUCKET_MAX)
1394 zone->uz_count = keg->uk_ipers;
1395 else
1396 zone->uz_count = BUCKET_MAX;
1397 return (0);
1398}
1399
1400/*
1401 * Keg header dtor. This frees all data, destroys locks, frees the hash
1402 * table and removes the keg from the global list.
1403 *
1404 * Arguments/Returns follow uma_dtor specifications
1405 * udata unused
1406 */
1407static void
1408keg_dtor(void *arg, int size, void *udata)
1409{
1410 uma_keg_t keg;
1411
1412 keg = (uma_keg_t)arg;
1413 mtx_lock(&keg->uk_lock);
1414 if (keg->uk_free != 0) {
1415 printf("Freed UMA keg was not empty (%d items). "
1416 " Lost %d pages of memory.\n",
1417 keg->uk_free, keg->uk_pages);
1418 }
1419 mtx_unlock(&keg->uk_lock);
1420
1421 if (keg->uk_flags & UMA_ZONE_HASH)
1422 hash_free(&keg->uk_hash);
1423
1424 mtx_destroy(&keg->uk_lock);
1425}
1426
1427/*
1428 * Zone header dtor.
1429 *
1430 * Arguments/Returns follow uma_dtor specifications
1431 * udata unused
1432 */
1433static void
1434zone_dtor(void *arg, int size, void *udata)
1435{
1436 uma_zone_t zone;
1437 uma_keg_t keg;
1438
1439 zone = (uma_zone_t)arg;
1440 keg = zone->uz_keg;
1441
1442 if (!(keg->uk_flags & UMA_ZFLAG_INTERNAL))
1443 cache_drain(zone);
1444
1445 mtx_lock(&uma_mtx);
1446 zone_drain(zone);
1447 if (keg->uk_flags & UMA_ZONE_SECONDARY) {
1448 LIST_REMOVE(zone, uz_link);
1449 /*
1450 * XXX there are some races here where
1451 * the zone can be drained but zone lock
1452 * released and then refilled before we
1453 * remove it... we dont care for now
1454 */
1455 ZONE_LOCK(zone);
1456 if (LIST_EMPTY(&keg->uk_zones))
1457 keg->uk_flags &= ~UMA_ZONE_SECONDARY;
1458 ZONE_UNLOCK(zone);
1459 mtx_unlock(&uma_mtx);
1460 } else {
1461 LIST_REMOVE(keg, uk_link);
1462 LIST_REMOVE(zone, uz_link);
1463 mtx_unlock(&uma_mtx);
1464 uma_zfree_internal(kegs, keg, NULL, SKIP_NONE);
1465 }
1466 zone->uz_keg = NULL;
1467}
1468
1469/*
1470 * Traverses every zone in the system and calls a callback
1471 *
1472 * Arguments:
1473 * zfunc A pointer to a function which accepts a zone
1474 * as an argument.
1475 *
1476 * Returns:
1477 * Nothing
1478 */
1479static void
1480zone_foreach(void (*zfunc)(uma_zone_t))
1481{
1482 uma_keg_t keg;
1483 uma_zone_t zone;
1484
1485 mtx_lock(&uma_mtx);
1486 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1487 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1488 zfunc(zone);
1489 }
1490 mtx_unlock(&uma_mtx);
1491}
1492
1493/* Public functions */
1494/* See uma.h */
1495void
1496uma_startup(void *bootmem)
1497{
1498 struct uma_zctor_args args;
1499 uma_slab_t slab;
1500 u_int slabsize;
1501 u_int objsize, totsize, wsize;
1502 int i;
1503
1504#ifdef UMA_DEBUG
1505 printf("Creating uma keg headers zone and keg.\n");
1506#endif
1507 /*
1508 * The general UMA lock is a recursion-allowed lock because
1509 * there is a code path where, while we're still configured
1510 * to use startup_alloc() for backend page allocations, we
1511 * may end up in uma_reclaim() which calls zone_foreach(zone_drain),
1512 * which grabs uma_mtx, only to later call into startup_alloc()
1513 * because while freeing we needed to allocate a bucket. Since
1514 * startup_alloc() also takes uma_mtx, we need to be able to
1515 * recurse on it.
1516 */
1517 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF | MTX_RECURSE);
1518
1519 /*
1520 * Figure out the maximum number of items-per-slab we'll have if
1521 * we're using the OFFPAGE slab header to track free items, given
1522 * all possible object sizes and the maximum desired wastage
1523 * (UMA_MAX_WASTE).
1524 *
1525 * We iterate until we find an object size for
1526 * which the calculated wastage in zone_small_init() will be
1527 * enough to warrant OFFPAGE. Since wastedspace versus objsize
1528 * is an overall increasing see-saw function, we find the smallest
1529 * objsize such that the wastage is always acceptable for objects
1530 * with that objsize or smaller. Since a smaller objsize always
1531 * generates a larger possible uma_max_ipers, we use this computed
1532 * objsize to calculate the largest ipers possible. Since the
1533 * ipers calculated for OFFPAGE slab headers is always larger than
1534 * the ipers initially calculated in zone_small_init(), we use
1535 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to
1536 * obtain the maximum ipers possible for offpage slab headers.
1537 *
1538 * It should be noted that ipers versus objsize is an inversly
1539 * proportional function which drops off rather quickly so as
1540 * long as our UMA_MAX_WASTE is such that the objsize we calculate
1541 * falls into the portion of the inverse relation AFTER the steep
1542 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386).
1543 *
1544 * Note that we have 8-bits (1 byte) to use as a freelist index
1545 * inside the actual slab header itself and this is enough to
1546 * accomodate us. In the worst case, a UMA_SMALLEST_UNIT sized
1547 * object with offpage slab header would have ipers =
1548 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is
1549 * 1 greater than what our byte-integer freelist index can
1550 * accomodate, but we know that this situation never occurs as
1551 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate
1552 * that we need to go to offpage slab headers. Or, if we do,
1553 * then we trap that condition below and panic in the INVARIANTS case.
1554 */
1555 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE;
1556 totsize = wsize;
1557 objsize = UMA_SMALLEST_UNIT;
1558 while (totsize >= wsize) {
1559 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) /
1560 (objsize + UMA_FRITM_SZ);
1561 totsize *= (UMA_FRITM_SZ + objsize);
1562 objsize++;
1563 }
1564 if (objsize > UMA_SMALLEST_UNIT)
1565 objsize--;
1566 uma_max_ipers = UMA_SLAB_SIZE / objsize;
1567
1568 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE;
1569 totsize = wsize;
1570 objsize = UMA_SMALLEST_UNIT;
1571 while (totsize >= wsize) {
1572 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) /
1573 (objsize + UMA_FRITMREF_SZ);
1574 totsize *= (UMA_FRITMREF_SZ + objsize);
1575 objsize++;
1576 }
1577 if (objsize > UMA_SMALLEST_UNIT)
1578 objsize--;
1579 uma_max_ipers_ref = UMA_SLAB_SIZE / objsize;
1580
1581 KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255),
1582 ("uma_startup: calculated uma_max_ipers values too large!"));
1583
1584#ifdef UMA_DEBUG
1585 printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers);
1586 printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n",
1587 uma_max_ipers_ref);
1588#endif
1589
1590 /* "manually" create the initial zone */
1591 args.name = "UMA Kegs";
1592 args.size = sizeof(struct uma_keg);
1593 args.ctor = keg_ctor;
1594 args.dtor = keg_dtor;
1595 args.uminit = zero_init;
1596 args.fini = NULL;
1597 args.keg = &masterkeg;
1598 args.align = 32 - 1;
1599 args.flags = UMA_ZFLAG_INTERNAL;
1600 /* The initial zone has no Per cpu queues so it's smaller */
1601 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1602
1603#ifdef UMA_DEBUG
1604 printf("Filling boot free list.\n");
1605#endif
1606 for (i = 0; i < UMA_BOOT_PAGES; i++) {
1607 slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE));
1608 slab->us_data = (u_int8_t *)slab;
1609 slab->us_flags = UMA_SLAB_BOOT;
1610 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1611 uma_boot_free++;
1612 }
1613
1614#ifdef UMA_DEBUG
1615 printf("Creating uma zone headers zone and keg.\n");
1616#endif
1617 args.name = "UMA Zones";
1618 args.size = sizeof(struct uma_zone) +
1619 (sizeof(struct uma_cache) * (mp_maxid + 1));
1620 args.ctor = zone_ctor;
1621 args.dtor = zone_dtor;
1622 args.uminit = zero_init;
1623 args.fini = NULL;
1624 args.keg = NULL;
1625 args.align = 32 - 1;
1626 args.flags = UMA_ZFLAG_INTERNAL;
1627 /* The initial zone has no Per cpu queues so it's smaller */
1628 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1629
1630#ifdef UMA_DEBUG
1631 printf("Initializing pcpu cache locks.\n");
1632#endif
1633#ifdef UMA_DEBUG
1634 printf("Creating slab and hash zones.\n");
1635#endif
1636
1637 /*
1638 * This is the max number of free list items we'll have with
1639 * offpage slabs.
1640 */
1641 slabsize = uma_max_ipers * UMA_FRITM_SZ;
1642 slabsize += sizeof(struct uma_slab);
1643
1644 /* Now make a zone for slab headers */
1645 slabzone = uma_zcreate("UMA Slabs",
1646 slabsize,
1647 NULL, NULL, NULL, NULL,
1648 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1649
1650 /*
1651 * We also create a zone for the bigger slabs with reference
1652 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1653 */
1654 slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ;
1655 slabsize += sizeof(struct uma_slab_refcnt);
1656 slabrefzone = uma_zcreate("UMA RCntSlabs",
1657 slabsize,
1658 NULL, NULL, NULL, NULL,
1659 UMA_ALIGN_PTR,
1660 UMA_ZFLAG_INTERNAL);
1661
1662 hashzone = uma_zcreate("UMA Hash",
1663 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1664 NULL, NULL, NULL, NULL,
1665 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1666
1667 bucket_init();
1668
1669#ifdef UMA_MD_SMALL_ALLOC
1670 booted = 1;
1671#endif
1672
1673#ifdef UMA_DEBUG
1674 printf("UMA startup complete.\n");
1675#endif
1676}
1677
1678/* see uma.h */
1679void
1680uma_startup2(void)
1681{
1682 booted = 1;
1683 bucket_enable();
1684#ifdef UMA_DEBUG
1685 printf("UMA startup2 complete.\n");
1686#endif
1687}
1688
1689/*
1690 * Initialize our callout handle
1691 *
1692 */
1693
1694static void
1695uma_startup3(void)
1696{
1697#ifdef UMA_DEBUG
1698 printf("Starting callout.\n");
1699#endif
1700 callout_init(&uma_callout, CALLOUT_MPSAFE);
1701 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1702#ifdef UMA_DEBUG
1703 printf("UMA startup3 complete.\n");
1704#endif
1705}
1706
1707static uma_zone_t
1708uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1709 int align, u_int16_t flags)
1710{
1711 struct uma_kctor_args args;
1712
1713 args.size = size;
1714 args.uminit = uminit;
1715 args.fini = fini;
1716 args.align = align;
1717 args.flags = flags;
1718 args.zone = zone;
1719 return (uma_zalloc_internal(kegs, &args, M_WAITOK));
1720}
1721
1722/* See uma.h */
1723uma_zone_t
1724uma_zcreate(char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1725 uma_init uminit, uma_fini fini, int align, u_int16_t flags)
1726
1727{
1728 struct uma_zctor_args args;
1729
1730 /* This stuff is essential for the zone ctor */
1731 args.name = name;
1732 args.size = size;
1733 args.ctor = ctor;
1734 args.dtor = dtor;
1735 args.uminit = uminit;
1736 args.fini = fini;
1737 args.align = align;
1738 args.flags = flags;
1739 args.keg = NULL;
1740
1741 return (uma_zalloc_internal(zones, &args, M_WAITOK));
1742}
1743
1744/* See uma.h */
1745uma_zone_t
1746uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1747 uma_init zinit, uma_fini zfini, uma_zone_t master)
1748{
1749 struct uma_zctor_args args;
1750
1751 args.name = name;
1752 args.size = master->uz_keg->uk_size;
1753 args.ctor = ctor;
1754 args.dtor = dtor;
1755 args.uminit = zinit;
1756 args.fini = zfini;
1757 args.align = master->uz_keg->uk_align;
1758 args.flags = master->uz_keg->uk_flags | UMA_ZONE_SECONDARY;
1759 args.keg = master->uz_keg;
1760
1761 return (uma_zalloc_internal(zones, &args, M_WAITOK));
1762}
1763
1764/* See uma.h */
1765void
1766uma_zdestroy(uma_zone_t zone)
1767{
1768 uma_zfree_internal(zones, zone, NULL, SKIP_NONE);
1769}
1770
1771/* See uma.h */
1772void *
1773uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1774{
1775 void *item;
1776 uma_cache_t cache;
1777 uma_bucket_t bucket;
1778 int cpu;
1779 int badness;
1780
1781 /* This is the fast path allocation */
1782#ifdef UMA_DEBUG_ALLOC_1
1783 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1784#endif
1785 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1786 zone->uz_name, flags);
1787
1788 if (!(flags & M_NOWAIT)) {
1789 KASSERT(curthread->td_intr_nesting_level == 0,
1790 ("malloc(M_WAITOK) in interrupt context"));
1791 if (nosleepwithlocks) {
1792#ifdef WITNESS
1793 badness = WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK,
1794 NULL,
1795 "malloc(M_WAITOK) of \"%s\", forcing M_NOWAIT",
1796 zone->uz_name);
1797#else
1798 badness = 1;
1799#endif
1800 } else {
1801 badness = 0;
1802#ifdef WITNESS
1803 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1804 "malloc(M_WAITOK) of \"%s\"", zone->uz_name);
1805#endif
1806 }
1807 if (badness) {
1808 flags &= ~M_WAITOK;
1809 flags |= M_NOWAIT;
1810 }
1811 }
1812
1813 /*
1814 * If possible, allocate from the per-CPU cache. There are two
1815 * requirements for safe access to the per-CPU cache: (1) the thread
1816 * accessing the cache must not be preempted or yield during access,
1817 * and (2) the thread must not migrate CPUs without switching which
1818 * cache it accesses. We rely on a critical section to prevent
1819 * preemption and migration. We release the critical section in
1820 * order to acquire the zone mutex if we are unable to allocate from
1821 * the current cache; when we re-acquire the critical section, we
1822 * must detect and handle migration if it has occurred.
1823 */
1824zalloc_restart:
1825 critical_enter();
1826 cpu = curcpu;
1827 cache = &zone->uz_cpu[cpu];
1828
1829zalloc_start:
1830 bucket = cache->uc_allocbucket;
1831
1832 if (bucket) {
1833 if (bucket->ub_cnt > 0) {
1834 bucket->ub_cnt--;
1835 item = bucket->ub_bucket[bucket->ub_cnt];
1836#ifdef INVARIANTS
1837 bucket->ub_bucket[bucket->ub_cnt] = NULL;
1838#endif
1839 KASSERT(item != NULL,
1840 ("uma_zalloc: Bucket pointer mangled."));
1841 cache->uc_allocs++;
1842 critical_exit();
1843#ifdef INVARIANTS
1844 ZONE_LOCK(zone);
1845 uma_dbg_alloc(zone, NULL, item);
1846 ZONE_UNLOCK(zone);
1847#endif
1848 if (zone->uz_ctor != NULL) {
1849 if (zone->uz_ctor(item, zone->uz_keg->uk_size,
1850 udata, flags) != 0) {
1851 uma_zfree_internal(zone, item, udata,
1852 SKIP_DTOR);
1853 return (NULL);
1854 }
1855 }
1856 if (flags & M_ZERO)
1857 bzero(item, zone->uz_keg->uk_size);
1858 return (item);
1859 } else if (cache->uc_freebucket) {
1860 /*
1861 * We have run out of items in our allocbucket.
1862 * See if we can switch with our free bucket.
1863 */
1864 if (cache->uc_freebucket->ub_cnt > 0) {
1865#ifdef UMA_DEBUG_ALLOC
1866 printf("uma_zalloc: Swapping empty with"
1867 " alloc.\n");
1868#endif
1869 bucket = cache->uc_freebucket;
1870 cache->uc_freebucket = cache->uc_allocbucket;
1871 cache->uc_allocbucket = bucket;
1872
1873 goto zalloc_start;
1874 }
1875 }
1876 }
1877 /*
1878 * Attempt to retrieve the item from the per-CPU cache has failed, so
1879 * we must go back to the zone. This requires the zone lock, so we
1880 * must drop the critical section, then re-acquire it when we go back
1881 * to the cache. Since the critical section is released, we may be
1882 * preempted or migrate. As such, make sure not to maintain any
1883 * thread-local state specific to the cache from prior to releasing
1884 * the critical section.
1885 */
1886 critical_exit();
1887 ZONE_LOCK(zone);
1888 critical_enter();
1889 cpu = curcpu;
1890 cache = &zone->uz_cpu[cpu];
1891 bucket = cache->uc_allocbucket;
1892 if (bucket != NULL) {
1893 if (bucket->ub_cnt > 0) {
1894 ZONE_UNLOCK(zone);
1895 goto zalloc_start;
1896 }
1897 bucket = cache->uc_freebucket;
1898 if (bucket != NULL && bucket->ub_cnt > 0) {
1899 ZONE_UNLOCK(zone);
1900 goto zalloc_start;
1901 }
1902 }
1903
1904 /* Since we have locked the zone we may as well send back our stats */
1905 zone->uz_allocs += cache->uc_allocs;
1906 cache->uc_allocs = 0;
1907 zone->uz_frees += cache->uc_frees;
1908 cache->uc_frees = 0;
1909
1910 /* Our old one is now a free bucket */
1911 if (cache->uc_allocbucket) {
1912 KASSERT(cache->uc_allocbucket->ub_cnt == 0,
1913 ("uma_zalloc_arg: Freeing a non free bucket."));
1914 LIST_INSERT_HEAD(&zone->uz_free_bucket,
1915 cache->uc_allocbucket, ub_link);
1916 cache->uc_allocbucket = NULL;
1917 }
1918
1919 /* Check the free list for a new alloc bucket */
1920 if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
1921 KASSERT(bucket->ub_cnt != 0,
1922 ("uma_zalloc_arg: Returning an empty bucket."));
1923
1924 LIST_REMOVE(bucket, ub_link);
1925 cache->uc_allocbucket = bucket;
1926 ZONE_UNLOCK(zone);
1927 goto zalloc_start;
1928 }
1929 /* We are no longer associated with this CPU. */
1930 critical_exit();
1931
1932 /* Bump up our uz_count so we get here less */
1933 if (zone->uz_count < BUCKET_MAX)
1934 zone->uz_count++;
1935
1936 /*
1937 * Now lets just fill a bucket and put it on the free list. If that
1938 * works we'll restart the allocation from the begining.
1939 */
1940 if (uma_zalloc_bucket(zone, flags)) {
1941 ZONE_UNLOCK(zone);
1942 goto zalloc_restart;
1943 }
1944 ZONE_UNLOCK(zone);
1945 /*
1946 * We may not be able to get a bucket so return an actual item.
1947 */
1948#ifdef UMA_DEBUG
1949 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
1950#endif
1951
1952 return (uma_zalloc_internal(zone, udata, flags));
1953}
1954
1955static uma_slab_t
1956uma_zone_slab(uma_zone_t zone, int flags)
1957{
1958 uma_slab_t slab;
1959 uma_keg_t keg;
1960
1961 keg = zone->uz_keg;
1962
1963 /*
1964 * This is to prevent us from recursively trying to allocate
1965 * buckets. The problem is that if an allocation forces us to
1966 * grab a new bucket we will call page_alloc, which will go off
1967 * and cause the vm to allocate vm_map_entries. If we need new
1968 * buckets there too we will recurse in kmem_alloc and bad
1969 * things happen. So instead we return a NULL bucket, and make
1970 * the code that allocates buckets smart enough to deal with it
1971 *
1972 * XXX: While we want this protection for the bucket zones so that
1973 * recursion from the VM is handled (and the calling code that
1974 * allocates buckets knows how to deal with it), we do not want
1975 * to prevent allocation from the slab header zones (slabzone
1976 * and slabrefzone) if uk_recurse is not zero for them. The
1977 * reason is that it could lead to NULL being returned for
1978 * slab header allocations even in the M_WAITOK case, and the
1979 * caller can't handle that.
1980 */
1981 if (keg->uk_flags & UMA_ZFLAG_INTERNAL && keg->uk_recurse != 0)
1982 if ((zone != slabzone) && (zone != slabrefzone))
1983 return (NULL);
1984
1985 slab = NULL;
1986
1987 for (;;) {
1988 /*
1989 * Find a slab with some space. Prefer slabs that are partially
1990 * used over those that are totally full. This helps to reduce
1991 * fragmentation.
1992 */
1993 if (keg->uk_free != 0) {
1994 if (!LIST_EMPTY(&keg->uk_part_slab)) {
1995 slab = LIST_FIRST(&keg->uk_part_slab);
1996 } else {
1997 slab = LIST_FIRST(&keg->uk_free_slab);
1998 LIST_REMOVE(slab, us_link);
1999 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2000 us_link);
2001 }
2002 return (slab);
2003 }
2004
2005 /*
2006 * M_NOVM means don't ask at all!
2007 */
2008 if (flags & M_NOVM)
2009 break;
2010
2011 if (keg->uk_maxpages &&
2012 keg->uk_pages >= keg->uk_maxpages) {
2013 keg->uk_flags |= UMA_ZFLAG_FULL;
2014
2015 if (flags & M_NOWAIT)
2016 break;
2017 else
2018 msleep(keg, &keg->uk_lock, PVM,
2019 "zonelimit", 0);
2020 continue;
2021 }
2022 keg->uk_recurse++;
2023 slab = slab_zalloc(zone, flags);
2024 keg->uk_recurse--;
2025
2026 /*
2027 * If we got a slab here it's safe to mark it partially used
2028 * and return. We assume that the caller is going to remove
2029 * at least one item.
2030 */
2031 if (slab) {
2032 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2033 return (slab);
2034 }
2035 /*
2036 * We might not have been able to get a slab but another cpu
2037 * could have while we were unlocked. Check again before we
2038 * fail.
2039 */
2040 if (flags & M_NOWAIT)
2041 flags |= M_NOVM;
2042 }
2043 return (slab);
2044}
2045
2046static void *
2047uma_slab_alloc(uma_zone_t zone, uma_slab_t slab)
2048{
2049 uma_keg_t keg;
2050 uma_slabrefcnt_t slabref;
2051 void *item;
2052 u_int8_t freei;
2053
2054 keg = zone->uz_keg;
2055
2056 freei = slab->us_firstfree;
2057 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2058 slabref = (uma_slabrefcnt_t)slab;
2059 slab->us_firstfree = slabref->us_freelist[freei].us_item;
2060 } else {
2061 slab->us_firstfree = slab->us_freelist[freei].us_item;
2062 }
2063 item = slab->us_data + (keg->uk_rsize * freei);
2064
2065 slab->us_freecount--;
2066 keg->uk_free--;
2067#ifdef INVARIANTS
2068 uma_dbg_alloc(zone, slab, item);
2069#endif
2070 /* Move this slab to the full list */
2071 if (slab->us_freecount == 0) {
2072 LIST_REMOVE(slab, us_link);
2073 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2074 }
2075
2076 return (item);
2077}
2078
2079static int
2080uma_zalloc_bucket(uma_zone_t zone, int flags)
2081{
2082 uma_bucket_t bucket;
2083 uma_slab_t slab;
2084 int16_t saved;
2085 int max, origflags = flags;
2086
2087 /*
2088 * Try this zone's free list first so we don't allocate extra buckets.
2089 */
2090 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2091 KASSERT(bucket->ub_cnt == 0,
2092 ("uma_zalloc_bucket: Bucket on free list is not empty."));
2093 LIST_REMOVE(bucket, ub_link);
2094 } else {
2095 int bflags;
2096
2097 bflags = (flags & ~M_ZERO);
2098 if (zone->uz_keg->uk_flags & UMA_ZFLAG_CACHEONLY)
2099 bflags |= M_NOVM;
2100
2101 ZONE_UNLOCK(zone);
2102 bucket = bucket_alloc(zone->uz_count, bflags);
2103 ZONE_LOCK(zone);
2104 }
2105
2106 if (bucket == NULL)
2107 return (0);
2108
2109#ifdef SMP
2110 /*
2111 * This code is here to limit the number of simultaneous bucket fills
2112 * for any given zone to the number of per cpu caches in this zone. This
2113 * is done so that we don't allocate more memory than we really need.
2114 */
2115 if (zone->uz_fills >= mp_ncpus)
2116 goto done;
2117
2118#endif
2119 zone->uz_fills++;
2120
2121 max = MIN(bucket->ub_entries, zone->uz_count);
2122 /* Try to keep the buckets totally full */
2123 saved = bucket->ub_cnt;
2124 while (bucket->ub_cnt < max &&
2125 (slab = uma_zone_slab(zone, flags)) != NULL) {
2126 while (slab->us_freecount && bucket->ub_cnt < max) {
2127 bucket->ub_bucket[bucket->ub_cnt++] =
2128 uma_slab_alloc(zone, slab);
2129 }
2130
2131 /* Don't block on the next fill */
2132 flags |= M_NOWAIT;
2133 }
2134
2135 /*
2136 * We unlock here because we need to call the zone's init.
2137 * It should be safe to unlock because the slab dealt with
2138 * above is already on the appropriate list within the keg
2139 * and the bucket we filled is not yet on any list, so we
2140 * own it.
2141 */
2142 if (zone->uz_init != NULL) {
2143 int i;
2144
2145 ZONE_UNLOCK(zone);
2146 for (i = saved; i < bucket->ub_cnt; i++)
2147 if (zone->uz_init(bucket->ub_bucket[i],
2148 zone->uz_keg->uk_size, origflags) != 0)
2149 break;
2150 /*
2151 * If we couldn't initialize the whole bucket, put the
2152 * rest back onto the freelist.
2153 */
2154 if (i != bucket->ub_cnt) {
2155 int j;
2156
2157 for (j = i; j < bucket->ub_cnt; j++) {
2158 uma_zfree_internal(zone, bucket->ub_bucket[j],
2159 NULL, SKIP_FINI);
2160#ifdef INVARIANTS
2161 bucket->ub_bucket[j] = NULL;
2162#endif
2163 }
2164 bucket->ub_cnt = i;
2165 }
2166 ZONE_LOCK(zone);
2167 }
2168
2169 zone->uz_fills--;
2170 if (bucket->ub_cnt != 0) {
2171 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2172 bucket, ub_link);
2173 return (1);
2174 }
2175#ifdef SMP
2176done:
2177#endif
2178 bucket_free(bucket);
2179
2180 return (0);
2181}
2182/*
2183 * Allocates an item for an internal zone
2184 *
2185 * Arguments
2186 * zone The zone to alloc for.
2187 * udata The data to be passed to the constructor.
2188 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2189 *
2190 * Returns
2191 * NULL if there is no memory and M_NOWAIT is set
2192 * An item if successful
2193 */
2194
2195static void *
2196uma_zalloc_internal(uma_zone_t zone, void *udata, int flags)
2197{
2198 uma_keg_t keg;
2199 uma_slab_t slab;
2200 void *item;
2201
2202 item = NULL;
2203 keg = zone->uz_keg;
2204
2205#ifdef UMA_DEBUG_ALLOC
2206 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2207#endif
2208 ZONE_LOCK(zone);
2209
2210 slab = uma_zone_slab(zone, flags);
2211 if (slab == NULL) {
2212 ZONE_UNLOCK(zone);
2213 return (NULL);
2214 }
2215
2216 item = uma_slab_alloc(zone, slab);
2217
2218 zone->uz_allocs++;
2219
2220 ZONE_UNLOCK(zone);
2221
2222 /*
2223 * We have to call both the zone's init (not the keg's init)
2224 * and the zone's ctor. This is because the item is going from
2225 * a keg slab directly to the user, and the user is expecting it
2226 * to be both zone-init'd as well as zone-ctor'd.
2227 */
2228 if (zone->uz_init != NULL) {
2229 if (zone->uz_init(item, keg->uk_size, flags) != 0) {
2230 uma_zfree_internal(zone, item, udata, SKIP_FINI);
2231 return (NULL);
2232 }
2233 }
2234 if (zone->uz_ctor != NULL) {
2235 if (zone->uz_ctor(item, keg->uk_size, udata, flags) != 0) {
2236 uma_zfree_internal(zone, item, udata, SKIP_DTOR);
2237 return (NULL);
2238 }
2239 }
2240 if (flags & M_ZERO)
2241 bzero(item, keg->uk_size);
2242
2243 return (item);
2244}
2245
2246/* See uma.h */
2247void
2248uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2249{
2250 uma_keg_t keg;
2251 uma_cache_t cache;
2252 uma_bucket_t bucket;
2253 int bflags;
2254 int cpu;
2255
2256 keg = zone->uz_keg;
2257
2258#ifdef UMA_DEBUG_ALLOC_1
2259 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2260#endif
2261 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2262 zone->uz_name);
2263
2264 if (zone->uz_dtor)
2265 zone->uz_dtor(item, keg->uk_size, udata);
2266#ifdef INVARIANTS
2267 ZONE_LOCK(zone);
2268 if (keg->uk_flags & UMA_ZONE_MALLOC)
2269 uma_dbg_free(zone, udata, item);
2270 else
2271 uma_dbg_free(zone, NULL, item);
2272 ZONE_UNLOCK(zone);
2273#endif
2274 /*
2275 * The race here is acceptable. If we miss it we'll just have to wait
2276 * a little longer for the limits to be reset.
2277 */
2278 if (keg->uk_flags & UMA_ZFLAG_FULL)
2279 goto zfree_internal;
2280
2281 /*
2282 * If possible, free to the per-CPU cache. There are two
2283 * requirements for safe access to the per-CPU cache: (1) the thread
2284 * accessing the cache must not be preempted or yield during access,
2285 * and (2) the thread must not migrate CPUs without switching which
2286 * cache it accesses. We rely on a critical section to prevent
2287 * preemption and migration. We release the critical section in
2288 * order to acquire the zone mutex if we are unable to free to the
2289 * current cache; when we re-acquire the critical section, we must
2290 * detect and handle migration if it has occurred.
2291 */
2292zfree_restart:
2293 critical_enter();
2294 cpu = curcpu;
2295 cache = &zone->uz_cpu[cpu];
2296
2297zfree_start:
2298 bucket = cache->uc_freebucket;
2299
2300 if (bucket) {
2301 /*
2302 * Do we have room in our bucket? It is OK for this uz count
2303 * check to be slightly out of sync.
2304 */
2305
2306 if (bucket->ub_cnt < bucket->ub_entries) {
2307 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2308 ("uma_zfree: Freeing to non free bucket index."));
2309 bucket->ub_bucket[bucket->ub_cnt] = item;
2310 bucket->ub_cnt++;
2311 cache->uc_frees++;
2312 critical_exit();
2313 return;
2314 } else if (cache->uc_allocbucket) {
2315#ifdef UMA_DEBUG_ALLOC
2316 printf("uma_zfree: Swapping buckets.\n");
2317#endif
2318 /*
2319 * We have run out of space in our freebucket.
2320 * See if we can switch with our alloc bucket.
2321 */
2322 if (cache->uc_allocbucket->ub_cnt <
2323 cache->uc_freebucket->ub_cnt) {
2324 bucket = cache->uc_freebucket;
2325 cache->uc_freebucket = cache->uc_allocbucket;
2326 cache->uc_allocbucket = bucket;
2327 goto zfree_start;
2328 }
2329 }
2330 }
2331 /*
2332 * We can get here for two reasons:
2333 *
2334 * 1) The buckets are NULL
2335 * 2) The alloc and free buckets are both somewhat full.
2336 *
2337 * We must go back the zone, which requires acquiring the zone lock,
2338 * which in turn means we must release and re-acquire the critical
2339 * section. Since the critical section is released, we may be
2340 * preempted or migrate. As such, make sure not to maintain any
2341 * thread-local state specific to the cache from prior to releasing
2342 * the critical section.
2343 */
2344 critical_exit();
2345 ZONE_LOCK(zone);
2346 critical_enter();
2347 cpu = curcpu;
2348 cache = &zone->uz_cpu[cpu];
2349 if (cache->uc_freebucket != NULL) {
2350 if (cache->uc_freebucket->ub_cnt <
2351 cache->uc_freebucket->ub_entries) {
2352 ZONE_UNLOCK(zone);
2353 goto zfree_start;
2354 }
2355 if (cache->uc_allocbucket != NULL &&
2356 (cache->uc_allocbucket->ub_cnt <
2357 cache->uc_freebucket->ub_cnt)) {
2358 ZONE_UNLOCK(zone);
2359 goto zfree_start;
2360 }
2361 }
2362
2363 bucket = cache->uc_freebucket;
2364 cache->uc_freebucket = NULL;
2365
2366 /* Can we throw this on the zone full list? */
2367 if (bucket != NULL) {
2368#ifdef UMA_DEBUG_ALLOC
2369 printf("uma_zfree: Putting old bucket on the free list.\n");
2370#endif
2371 /* ub_cnt is pointing to the last free item */
2372 KASSERT(bucket->ub_cnt != 0,
2373 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2374 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2375 bucket, ub_link);
2376 }
2377 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2378 LIST_REMOVE(bucket, ub_link);
2379 ZONE_UNLOCK(zone);
2380 cache->uc_freebucket = bucket;
2381 goto zfree_start;
2382 }
2383 /* We are no longer associated with this CPU. */
2384 critical_exit();
2385
2386 /* And the zone.. */
2387 ZONE_UNLOCK(zone);
2388
2389#ifdef UMA_DEBUG_ALLOC
2390 printf("uma_zfree: Allocating new free bucket.\n");
2391#endif
2392 bflags = M_NOWAIT;
2393
2394 if (keg->uk_flags & UMA_ZFLAG_CACHEONLY)
2395 bflags |= M_NOVM;
2396 bucket = bucket_alloc(zone->uz_count, bflags);
2397 if (bucket) {
2398 ZONE_LOCK(zone);
2399 LIST_INSERT_HEAD(&zone->uz_free_bucket,
2400 bucket, ub_link);
2401 ZONE_UNLOCK(zone);
2402 goto zfree_restart;
2403 }
2404
2405 /*
2406 * If nothing else caught this, we'll just do an internal free.
2407 */
2408zfree_internal:
2409 uma_zfree_internal(zone, item, udata, SKIP_DTOR);
2410
2411 return;
2412}
2413
2414/*
2415 * Frees an item to an INTERNAL zone or allocates a free bucket
2416 *
2417 * Arguments:
2418 * zone The zone to free to
2419 * item The item we're freeing
2420 * udata User supplied data for the dtor
2421 * skip Skip dtors and finis
2422 */
2423static void
2424uma_zfree_internal(uma_zone_t zone, void *item, void *udata,
2425 enum zfreeskip skip)
2426{
2427 uma_slab_t slab;
2428 uma_slabrefcnt_t slabref;
2429 uma_keg_t keg;
2430 u_int8_t *mem;
2431 u_int8_t freei;
2432
2433 keg = zone->uz_keg;
2434
2435 if (skip < SKIP_DTOR && zone->uz_dtor)
2436 zone->uz_dtor(item, keg->uk_size, udata);
2437 if (skip < SKIP_FINI && zone->uz_fini)
2438 zone->uz_fini(item, keg->uk_size);
2439
2440 ZONE_LOCK(zone);
2441
2442 if (!(keg->uk_flags & UMA_ZONE_MALLOC)) {
2443 mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK));
2444 if (keg->uk_flags & UMA_ZONE_HASH)
2445 slab = hash_sfind(&keg->uk_hash, mem);
2446 else {
2447 mem += keg->uk_pgoff;
2448 slab = (uma_slab_t)mem;
2449 }
2450 } else {
2451 slab = (uma_slab_t)udata;
2452 }
2453
2454 /* Do we need to remove from any lists? */
2455 if (slab->us_freecount+1 == keg->uk_ipers) {
2456 LIST_REMOVE(slab, us_link);
2457 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2458 } else if (slab->us_freecount == 0) {
2459 LIST_REMOVE(slab, us_link);
2460 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2461 }
2462
2463 /* Slab management stuff */
2464 freei = ((unsigned long)item - (unsigned long)slab->us_data)
2465 / keg->uk_rsize;
2466
2467#ifdef INVARIANTS
2468 if (!skip)
2469 uma_dbg_free(zone, slab, item);
2470#endif
2471
2472 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2473 slabref = (uma_slabrefcnt_t)slab;
2474 slabref->us_freelist[freei].us_item = slab->us_firstfree;
2475 } else {
2476 slab->us_freelist[freei].us_item = slab->us_firstfree;
2477 }
2478 slab->us_firstfree = freei;
2479 slab->us_freecount++;
2480
2481 /* Zone statistics */
2482 keg->uk_free++;
2483 zone->uz_frees++;
2484
2485 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2486 if (keg->uk_pages < keg->uk_maxpages)
2487 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2488
2489 /* We can handle one more allocation */
2490 wakeup_one(keg);
2491 }
2492
2493 ZONE_UNLOCK(zone);
2494}
2495
2496/* See uma.h */
2497void
2498uma_zone_set_max(uma_zone_t zone, int nitems)
2499{
2500 uma_keg_t keg;
2501
2502 keg = zone->uz_keg;
2503 ZONE_LOCK(zone);
2504 if (keg->uk_ppera > 1)
2505 keg->uk_maxpages = nitems * keg->uk_ppera;
2506 else
2507 keg->uk_maxpages = nitems / keg->uk_ipers;
2508
2509 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2510 keg->uk_maxpages++;
2511
2512 ZONE_UNLOCK(zone);
2513}
2514
2515/* See uma.h */
2516void
2517uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2518{
2519 ZONE_LOCK(zone);
2520 KASSERT(zone->uz_keg->uk_pages == 0,
2521 ("uma_zone_set_init on non-empty keg"));
2522 zone->uz_keg->uk_init = uminit;
2523 ZONE_UNLOCK(zone);
2524}
2525
2526/* See uma.h */
2527void
2528uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2529{
2530 ZONE_LOCK(zone);
2531 KASSERT(zone->uz_keg->uk_pages == 0,
2532 ("uma_zone_set_fini on non-empty keg"));
2533 zone->uz_keg->uk_fini = fini;
2534 ZONE_UNLOCK(zone);
2535}
2536
2537/* See uma.h */
2538void
2539uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2540{
2541 ZONE_LOCK(zone);
2542 KASSERT(zone->uz_keg->uk_pages == 0,
2543 ("uma_zone_set_zinit on non-empty keg"));
2544 zone->uz_init = zinit;
2545 ZONE_UNLOCK(zone);
2546}
2547
2548/* See uma.h */
2549void
2550uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2551{
2552 ZONE_LOCK(zone);
2553 KASSERT(zone->uz_keg->uk_pages == 0,
2554 ("uma_zone_set_zfini on non-empty keg"));
2555 zone->uz_fini = zfini;
2556 ZONE_UNLOCK(zone);
2557}
2558
2559/* See uma.h */
2560/* XXX uk_freef is not actually used with the zone locked */
2561void
2562uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2563{
2564 ZONE_LOCK(zone);
2565 zone->uz_keg->uk_freef = freef;
2566 ZONE_UNLOCK(zone);
2567}
2568
2569/* See uma.h */
2570/* XXX uk_allocf is not actually used with the zone locked */
2571void
2572uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2573{
2574 ZONE_LOCK(zone);
2575 zone->uz_keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2576 zone->uz_keg->uk_allocf = allocf;
2577 ZONE_UNLOCK(zone);
2578}
2579
2580/* See uma.h */
2581int
2582uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count)
2583{
2584 uma_keg_t keg;
2585 vm_offset_t kva;
2586 int pages;
2587
2588 keg = zone->uz_keg;
2589 pages = count / keg->uk_ipers;
2590
2591 if (pages * keg->uk_ipers < count)
2592 pages++;
2593
2594 kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2595
2596 if (kva == 0)
2597 return (0);
2598 if (obj == NULL) {
2599 obj = vm_object_allocate(OBJT_DEFAULT,
2600 pages);
2601 } else {
2602 VM_OBJECT_LOCK_INIT(obj, "uma object");
2603 _vm_object_allocate(OBJT_DEFAULT,
2604 pages, obj);
2605 }
2606 ZONE_LOCK(zone);
2607 keg->uk_kva = kva;
2608 keg->uk_obj = obj;
2609 keg->uk_maxpages = pages;
2610 keg->uk_allocf = obj_alloc;
2611 keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
2612 ZONE_UNLOCK(zone);
2613 return (1);
2614}
2615
2616/* See uma.h */
2617void
2618uma_prealloc(uma_zone_t zone, int items)
2619{
2620 int slabs;
2621 uma_slab_t slab;
2622 uma_keg_t keg;
2623
2624 keg = zone->uz_keg;
2625 ZONE_LOCK(zone);
2626 slabs = items / keg->uk_ipers;
2627 if (slabs * keg->uk_ipers < items)
2628 slabs++;
2629 while (slabs > 0) {
2630 slab = slab_zalloc(zone, M_WAITOK);
2631 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2632 slabs--;
2633 }
2634 ZONE_UNLOCK(zone);
2635}
2636
2637/* See uma.h */
2638u_int32_t *
2639uma_find_refcnt(uma_zone_t zone, void *item)
2640{
2641 uma_slabrefcnt_t slabref;
2642 uma_keg_t keg;
2643 u_int32_t *refcnt;
2644 int idx;
2645
2646 keg = zone->uz_keg;
2647 slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item &
2648 (~UMA_SLAB_MASK));
2649 KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT,
2650 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
2651 idx = ((unsigned long)item - (unsigned long)slabref->us_data)
2652 / keg->uk_rsize;
2653 refcnt = &slabref->us_freelist[idx].us_refcnt;
2654 return refcnt;
2655}
2656
2657/* See uma.h */
2658void
2659uma_reclaim(void)
2660{
2661#ifdef UMA_DEBUG
2662 printf("UMA: vm asked us to release pages!\n");
2663#endif
2664 bucket_enable();
2665 zone_foreach(zone_drain);
2666 /*
2667 * Some slabs may have been freed but this zone will be visited early
2668 * we visit again so that we can free pages that are empty once other
2669 * zones are drained. We have to do the same for buckets.
2670 */
2671 zone_drain(slabzone);
2672 zone_drain(slabrefzone);
2673 bucket_zone_drain();
2674}
2675
2676void *
2677uma_large_malloc(int size, int wait)
2678{
2679 void *mem;
2680 uma_slab_t slab;
2681 u_int8_t flags;
2682
2683 slab = uma_zalloc_internal(slabzone, NULL, wait);
2684 if (slab == NULL)
2685 return (NULL);
2686 mem = page_alloc(NULL, size, &flags, wait);
2687 if (mem) {
2688 vsetslab((vm_offset_t)mem, slab);
2689 slab->us_data = mem;
2690 slab->us_flags = flags | UMA_SLAB_MALLOC;
2691 slab->us_size = size;
2692 } else {
2693 uma_zfree_internal(slabzone, slab, NULL, SKIP_NONE);
2694 }
2695
2696 return (mem);
2697}
2698
2699void
2700uma_large_free(uma_slab_t slab)
2701{
2702 vsetobj((vm_offset_t)slab->us_data, kmem_object);
2703 page_free(slab->us_data, slab->us_size, slab->us_flags);
2704 uma_zfree_internal(slabzone, slab, NULL, SKIP_NONE);
2705}
2706
2707void
2708uma_print_stats(void)
2709{
2710 zone_foreach(uma_print_zone);
2711}
2712
2713static void
2714slab_print(uma_slab_t slab)
2715{
2716 printf("slab: keg %p, data %p, freecount %d, firstfree %d\n",
2717 slab->us_keg, slab->us_data, slab->us_freecount,
2718 slab->us_firstfree);
2719}
2720
2721static void
2722cache_print(uma_cache_t cache)
2723{
2724 printf("alloc: %p(%d), free: %p(%d)\n",
2725 cache->uc_allocbucket,
2726 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
2727 cache->uc_freebucket,
2728 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
2729}
2730
2731void
2732uma_print_zone(uma_zone_t zone)
2733{
2734 uma_cache_t cache;
2735 uma_keg_t keg;
2736 uma_slab_t slab;
2737 int i;
2738
2739 keg = zone->uz_keg;
2740 printf("%s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n",
2741 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
2742 keg->uk_ipers, keg->uk_ppera,
2743 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
2744 printf("Part slabs:\n");
2745 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
2746 slab_print(slab);
2747 printf("Free slabs:\n");
2748 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
2749 slab_print(slab);
2750 printf("Full slabs:\n");
2751 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
2752 slab_print(slab);
2753 for (i = 0; i <= mp_maxid; i++) {
2754 if (CPU_ABSENT(i))
2755 continue;
2756 cache = &zone->uz_cpu[i];
2757 printf("CPU %d Cache:\n", i);
2758 cache_print(cache);
2759 }
2760}
2761
2762/*
2763 * Generate statistics across both the zone and its per-cpu cache's. Return
2764 * desired statistics if the pointer is non-NULL for that statistic.
2765 *
2766 * Note: does not update the zone statistics, as it can't safely clear the
2767 * per-CPU cache statistic.
2768 *
2769 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
2770 * safe from off-CPU; we should modify the caches to track this information
2771 * directly so that we don't have to.
2772 */
2773static void
2774uma_zone_sumstat(uma_zone_t z, int *cachefreep, u_int64_t *allocsp,
2775 u_int64_t *freesp)
2776{
2777 uma_cache_t cache;
2778 u_int64_t allocs, frees;
2779 int cachefree, cpu;
2780
2781 allocs = frees = 0;
2782 cachefree = 0;
2783 for (cpu = 0; cpu <= mp_maxid; cpu++) {
2784 if (CPU_ABSENT(cpu))
2785 continue;
2786 cache = &z->uz_cpu[cpu];
2787 if (cache->uc_allocbucket != NULL)
2788 cachefree += cache->uc_allocbucket->ub_cnt;
2789 if (cache->uc_freebucket != NULL)
2790 cachefree += cache->uc_freebucket->ub_cnt;
2791 allocs += cache->uc_allocs;
2792 frees += cache->uc_frees;
2793 }
2794 allocs += z->uz_allocs;
2795 frees += z->uz_frees;
2796 if (cachefreep != NULL)
2797 *cachefreep = cachefree;
2798 if (allocsp != NULL)
2799 *allocsp = allocs;
2800 if (freesp != NULL)
2801 *freesp = frees;
2802}
2803
2804/*
2805 * Sysctl handler for vm.zone
2806 *
2807 * stolen from vm_zone.c
2808 */
2809static int
2810sysctl_vm_zone(SYSCTL_HANDLER_ARGS)
2811{
2812 int error, len, cnt;
2813 const int linesize = 128; /* conservative */
2814 int totalfree;
2815 char *tmpbuf, *offset;
2816 uma_zone_t z;
2817 uma_keg_t zk;
2818 char *p;
2819 int cachefree;
2820 uma_bucket_t bucket;
2821 u_int64_t allocs, frees;
2822
2823 cnt = 0;
2824 mtx_lock(&uma_mtx);
2825 LIST_FOREACH(zk, &uma_kegs, uk_link) {
2826 LIST_FOREACH(z, &zk->uk_zones, uz_link)
2827 cnt++;
2828 }
2829 mtx_unlock(&uma_mtx);
2830 MALLOC(tmpbuf, char *, (cnt == 0 ? 1 : cnt) * linesize,
2831 M_TEMP, M_WAITOK);
2832 len = snprintf(tmpbuf, linesize,
2833 "\nITEM SIZE LIMIT USED FREE REQUESTS\n\n");
2834 if (cnt == 0)
2835 tmpbuf[len - 1] = '\0';
2836 error = SYSCTL_OUT(req, tmpbuf, cnt == 0 ? len-1 : len);
2837 if (error || cnt == 0)
2838 goto out;
2839 offset = tmpbuf;
2840 mtx_lock(&uma_mtx);
2841 LIST_FOREACH(zk, &uma_kegs, uk_link) {
2842 LIST_FOREACH(z, &zk->uk_zones, uz_link) {
2843 if (cnt == 0) /* list may have changed size */
2844 break;
2845 ZONE_LOCK(z);
2846 cachefree = 0;
2847 if (!(zk->uk_flags & UMA_ZFLAG_INTERNAL)) {
2848 uma_zone_sumstat(z, &cachefree, &allocs, &frees);
2849 } else {
2850 allocs = z->uz_allocs;
2851 frees = z->uz_frees;
2852 }
2853
2854 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link) {
2855 cachefree += bucket->ub_cnt;
2856 }
2857 totalfree = zk->uk_free + cachefree;
2858 len = snprintf(offset, linesize,
2859 "%-12.12s %6.6u, %8.8u, %6.6u, %6.6u, %8.8llu\n",
2860 z->uz_name, zk->uk_size,
2861 zk->uk_maxpages * zk->uk_ipers,
2862 (zk->uk_ipers * (zk->uk_pages / zk->uk_ppera)) - totalfree,
2863 totalfree,
2864 (unsigned long long)allocs);
2865 ZONE_UNLOCK(z);
2866 for (p = offset + 12; p > offset && *p == ' '; --p)
2867 /* nothing */ ;
2868 p[1] = ':';
2869 cnt--;
2870 offset += len;
2871 }
2872 }
2873 mtx_unlock(&uma_mtx);
2874 *offset++ = '\0';
2875 error = SYSCTL_OUT(req, tmpbuf, offset - tmpbuf);
2876out:
2877 FREE(tmpbuf, M_TEMP);
2878 return (error);
2879}
2880
2881static int
2882sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
2883{
2884 uma_keg_t kz;
2885 uma_zone_t z;
2886 int count;
2887
2888 count = 0;
2889 mtx_lock(&uma_mtx);
2890 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2891 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2892 count++;
2893 }
2894 mtx_unlock(&uma_mtx);
2895 return (sysctl_handle_int(oidp, &count, 0, req));
2896}
2897
2898static int
2899sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
2900{
2901 struct uma_stream_header ush;
2902 struct uma_type_header uth;
2903 struct uma_percpu_stat ups;
2904 uma_bucket_t bucket;
2905 struct sbuf sbuf;
2906 uma_cache_t cache;
2907 uma_keg_t kz;
2908 uma_zone_t z;
2909 char *buffer;
2910 int buflen, count, error, i;
2911
2912 mtx_lock(&uma_mtx);
2913restart:
2914 mtx_assert(&uma_mtx, MA_OWNED);
2915 count = 0;
2916 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2917 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2918 count++;
2919 }
2920 mtx_unlock(&uma_mtx);
2921
2922 buflen = sizeof(ush) + count * (sizeof(uth) + sizeof(ups) *
2923 MAXCPU) + 1;
2924 buffer = malloc(buflen, M_TEMP, M_WAITOK | M_ZERO);
2925
2926 mtx_lock(&uma_mtx);
2927 i = 0;
2928 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2929 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2930 i++;
2931 }
2932 if (i > count) {
2933 free(buffer, M_TEMP);
2934 goto restart;
2935 }
2936 count = i;
2937
2938 sbuf_new(&sbuf, buffer, buflen, SBUF_FIXEDLEN);
2939
2940 /*
2941 * Insert stream header.
2942 */
2943 bzero(&ush, sizeof(ush));
2944 ush.ush_version = UMA_STREAM_VERSION;
2945 ush.ush_maxcpus = MAXCPU;
2946 ush.ush_count = count;
2947 if (sbuf_bcat(&sbuf, &ush, sizeof(ush)) < 0) {
2948 mtx_unlock(&uma_mtx);
2949 error = ENOMEM;
2950 goto out;
2951 }
2952
2953 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2954 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
2955 bzero(&uth, sizeof(uth));
2956 ZONE_LOCK(z);
2957 strlcpy(uth.uth_name, z->uz_name, UMA_MAX_NAME);
2958 uth.uth_align = kz->uk_align;
2959 uth.uth_pages = kz->uk_pages;
2960 uth.uth_keg_free = kz->uk_free;
2961 uth.uth_size = kz->uk_size;
2962 uth.uth_rsize = kz->uk_rsize;
2963 uth.uth_maxpages = kz->uk_maxpages;
2964 if (kz->uk_ppera > 1)
2965 uth.uth_limit = kz->uk_maxpages /
2966 kz->uk_ppera;
2967 else
2968 uth.uth_limit = kz->uk_maxpages *
2969 kz->uk_ipers;
2970 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
2971 uth.uth_zone_free += bucket->ub_cnt;
2972 uth.uth_allocs = z->uz_allocs;
2973 uth.uth_frees = z->uz_frees;
2974 ZONE_UNLOCK(z);
2975 if (sbuf_bcat(&sbuf, &uth, sizeof(uth)) < 0) {
2976 mtx_unlock(&uma_mtx);
2977 error = ENOMEM;
2978 goto out;
2979 }
2980 /*
2981 * XXXRW: Should not access bucket fields from
2982 * non-local CPU. Instead need to modify the caches
2983 * to directly maintain these statistics so we don't
2984 * have to.
2985 */
2986 for (i = 0; i < MAXCPU; i++) {
2987 bzero(&ups, sizeof(ups));
2988 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
2989 goto skip;
2990 cache = &z->uz_cpu[i];
2991 if (cache->uc_allocbucket != NULL)
2992 ups.ups_cache_free +=
2993 cache->uc_allocbucket->ub_cnt;
2994 if (cache->uc_freebucket != NULL)
2995 ups.ups_cache_free +=
2996 cache->uc_freebucket->ub_cnt;
2997 ups.ups_allocs = cache->uc_allocs;
2998 ups.ups_frees = cache->uc_frees;
2999skip:
3000 if (sbuf_bcat(&sbuf, &ups, sizeof(ups)) < 0) {
3001 mtx_unlock(&uma_mtx);
3002 error = ENOMEM;
3003 goto out;
3004 }
3005 }
3006 }
3007 }
3008 mtx_unlock(&uma_mtx);
3009 sbuf_finish(&sbuf);
3010 error = SYSCTL_OUT(req, sbuf_data(&sbuf), sbuf_len(&sbuf));
3011out:
3012 free(buffer, M_TEMP);
3013 return (error);
3014}