1/*-
2 * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org>
3 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
4 * Copyright (c) 2004-2006 Robert N. M. Watson
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice unmodified, this list of conditions, and the following
12 * disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 */
28
29/*
30 * uma_core.c Implementation of the Universal Memory allocator
31 *
32 * This allocator is intended to replace the multitude of similar object caches
33 * in the standard FreeBSD kernel. The intent is to be flexible as well as
34 * effecient. A primary design goal is to return unused memory to the rest of
35 * the system. This will make the system as a whole more flexible due to the
36 * ability to move memory to subsystems which most need it instead of leaving
37 * pools of reserved memory unused.
38 *
39 * The basic ideas stem from similar slab/zone based allocators whose algorithms
40 * are well known.
41 *
42 */
43
44/*
45 * TODO:
46 * - Improve memory usage for large allocations
47 * - Investigate cache size adjustments
48 */
49
50#include <sys/cdefs.h>
51__FBSDID("$FreeBSD: head/sys/vm/uma_core.c 295222 2016-02-03 23:30:17Z glebius $");
52
53/* I should really use ktr.. */
54/*
55#define UMA_DEBUG 1
56#define UMA_DEBUG_ALLOC 1
57#define UMA_DEBUG_ALLOC_1 1
58*/
59
60#include "opt_ddb.h"
61#include "opt_param.h"
62#include "opt_vm.h"
63
64#include <sys/param.h>
65#include <sys/systm.h>
66#include <sys/bitset.h>
67#include <sys/kernel.h>
68#include <sys/types.h>
69#include <sys/queue.h>
70#include <sys/malloc.h>
71#include <sys/ktr.h>
72#include <sys/lock.h>
73#include <sys/sysctl.h>
74#include <sys/mutex.h>
75#include <sys/proc.h>
76#include <sys/random.h>
77#include <sys/rwlock.h>
78#include <sys/sbuf.h>
79#include <sys/sched.h>
80#include <sys/smp.h>
81#include <sys/taskqueue.h>
82#include <sys/vmmeter.h>
83
84#include <vm/vm.h>
85#include <vm/vm_object.h>
86#include <vm/vm_page.h>
87#include <vm/vm_pageout.h>
88#include <vm/vm_param.h>
89#include <vm/vm_map.h>
90#include <vm/vm_kern.h>
91#include <vm/vm_extern.h>
92#include <vm/uma.h>
93#include <vm/uma_int.h>
94#include <vm/uma_dbg.h>
95
96#include <ddb/ddb.h>
97
98#ifdef DEBUG_MEMGUARD
99#include <vm/memguard.h>
100#endif
101
102/*
103 * This is the zone and keg from which all zones are spawned. The idea is that
104 * even the zone & keg heads are allocated from the allocator, so we use the
105 * bss section to bootstrap us.
106 */
107static struct uma_keg masterkeg;
108static struct uma_zone masterzone_k;
109static struct uma_zone masterzone_z;
110static uma_zone_t kegs = &masterzone_k;
111static uma_zone_t zones = &masterzone_z;
112
113/* This is the zone from which all of uma_slab_t's are allocated. */
114static uma_zone_t slabzone;
115static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
116
117/*
118 * The initial hash tables come out of this zone so they can be allocated
119 * prior to malloc coming up.
120 */
121static uma_zone_t hashzone;
122
123/* The boot-time adjusted value for cache line alignment. */
124int uma_align_cache = 64 - 1;
125
126static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
127
128/*
129 * Are we allowed to allocate buckets?
130 */
131static int bucketdisable = 1;
132
133/* Linked list of all kegs in the system */
134static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
135
136/* Linked list of all cache-only zones in the system */
137static LIST_HEAD(,uma_zone) uma_cachezones =
138 LIST_HEAD_INITIALIZER(uma_cachezones);
139
140/* This RW lock protects the keg list */
141static struct rwlock_padalign uma_rwlock;
142
143/* Linked list of boot time pages */
144static LIST_HEAD(,uma_slab) uma_boot_pages =
145 LIST_HEAD_INITIALIZER(uma_boot_pages);
146
147/* This mutex protects the boot time pages list */
148static struct mtx_padalign uma_boot_pages_mtx;
149
150static struct sx uma_drain_lock;
151
152/* Is the VM done starting up? */
153static int booted = 0;
154#define UMA_STARTUP 1
155#define UMA_STARTUP2 2
156
157/*
158 * Only mbuf clusters use ref zones. Just provide enough references
159 * to support the one user. New code should not use the ref facility.
160 */
161static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES;
162
163/*
164 * This is the handle used to schedule events that need to happen
165 * outside of the allocation fast path.
166 */
167static struct callout uma_callout;
168#define UMA_TIMEOUT 20 /* Seconds for callout interval. */
169
170/*
171 * This structure is passed as the zone ctor arg so that I don't have to create
172 * a special allocation function just for zones.
173 */
174struct uma_zctor_args {
175 const char *name;
176 size_t size;
177 uma_ctor ctor;
178 uma_dtor dtor;
179 uma_init uminit;
180 uma_fini fini;
181 uma_import import;
182 uma_release release;
183 void *arg;
184 uma_keg_t keg;
185 int align;
186 uint32_t flags;
187};
188
189struct uma_kctor_args {
190 uma_zone_t zone;
191 size_t size;
192 uma_init uminit;
193 uma_fini fini;
194 int align;
195 uint32_t flags;
196};
197
198struct uma_bucket_zone {
199 uma_zone_t ubz_zone;
200 char *ubz_name;
201 int ubz_entries; /* Number of items it can hold. */
202 int ubz_maxsize; /* Maximum allocation size per-item. */
203};
204
205/*
206 * Compute the actual number of bucket entries to pack them in power
207 * of two sizes for more efficient space utilization.
208 */
209#define BUCKET_SIZE(n) \
210 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
211
212#define BUCKET_MAX BUCKET_SIZE(256)
213
214struct uma_bucket_zone bucket_zones[] = {
215 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
216 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
217 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
218 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
219 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
220 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
221 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
222 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
223 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
224 { NULL, NULL, 0}
225};
226
227/*
228 * Flags and enumerations to be passed to internal functions.
229 */
230enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
231
232/* Prototypes.. */
233
234static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
235static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
236static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
237static void page_free(void *, vm_size_t, uint8_t);
238static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
239static void cache_drain(uma_zone_t);
240static void bucket_drain(uma_zone_t, uma_bucket_t);
241static void bucket_cache_drain(uma_zone_t zone);
242static int keg_ctor(void *, int, void *, int);
243static void keg_dtor(void *, int, void *);
244static int zone_ctor(void *, int, void *, int);
245static void zone_dtor(void *, int, void *);
246static int zero_init(void *, int, int);
247static void keg_small_init(uma_keg_t keg);
248static void keg_large_init(uma_keg_t keg);
249static void zone_foreach(void (*zfunc)(uma_zone_t));
250static void zone_timeout(uma_zone_t zone);
251static int hash_alloc(struct uma_hash *);
252static int hash_expand(struct uma_hash *, struct uma_hash *);
253static void hash_free(struct uma_hash *hash);
254static void uma_timeout(void *);
255static void uma_startup3(void);
256static void *zone_alloc_item(uma_zone_t, void *, int);
257static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
258static void bucket_enable(void);
259static void bucket_init(void);
260static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
261static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
262static void bucket_zone_drain(void);
263static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
264static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
265static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
266static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
267static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
268static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
269 uma_fini fini, int align, uint32_t flags);
270static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
271static void zone_release(uma_zone_t zone, void **bucket, int cnt);
272static void uma_zero_item(void *item, uma_zone_t zone);
273
274void uma_print_zone(uma_zone_t);
275void uma_print_stats(void);
276static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
277static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
278
279#ifdef INVARIANTS
280static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
281static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
282#endif
283
284SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
285
286SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
287 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
288
289SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
290 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
291
292static int zone_warnings = 1;
293SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
294 "Warn when UMA zones becomes full");
295
296/*
297 * This routine checks to see whether or not it's safe to enable buckets.
298 */
299static void
300bucket_enable(void)
301{
302 bucketdisable = vm_page_count_min();
303}
304
305/*
306 * Initialize bucket_zones, the array of zones of buckets of various sizes.
307 *
308 * For each zone, calculate the memory required for each bucket, consisting
309 * of the header and an array of pointers.
310 */
311static void
312bucket_init(void)
313{
314 struct uma_bucket_zone *ubz;
315 int size;
316
317 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
318 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
319 size += sizeof(void *) * ubz->ubz_entries;
320 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
321 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
322 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
323 }
324}
325
326/*
327 * Given a desired number of entries for a bucket, return the zone from which
328 * to allocate the bucket.
329 */
330static struct uma_bucket_zone *
331bucket_zone_lookup(int entries)
332{
333 struct uma_bucket_zone *ubz;
334
335 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
336 if (ubz->ubz_entries >= entries)
337 return (ubz);
338 ubz--;
339 return (ubz);
340}
341
342static int
343bucket_select(int size)
344{
345 struct uma_bucket_zone *ubz;
346
347 ubz = &bucket_zones[0];
348 if (size > ubz->ubz_maxsize)
349 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
350
351 for (; ubz->ubz_entries != 0; ubz++)
352 if (ubz->ubz_maxsize < size)
353 break;
354 ubz--;
355 return (ubz->ubz_entries);
356}
357
358static uma_bucket_t
359bucket_alloc(uma_zone_t zone, void *udata, int flags)
360{
361 struct uma_bucket_zone *ubz;
362 uma_bucket_t bucket;
363
364 /*
365 * This is to stop us from allocating per cpu buckets while we're
366 * running out of vm.boot_pages. Otherwise, we would exhaust the
367 * boot pages. This also prevents us from allocating buckets in
368 * low memory situations.
369 */
370 if (bucketdisable)
371 return (NULL);
372 /*
373 * To limit bucket recursion we store the original zone flags
374 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
375 * NOVM flag to persist even through deep recursions. We also
376 * store ZFLAG_BUCKET once we have recursed attempting to allocate
377 * a bucket for a bucket zone so we do not allow infinite bucket
378 * recursion. This cookie will even persist to frees of unused
379 * buckets via the allocation path or bucket allocations in the
380 * free path.
381 */
382 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
383 udata = (void *)(uintptr_t)zone->uz_flags;
384 else {
385 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
386 return (NULL);
387 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
388 }
389 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
390 flags |= M_NOVM;
391 ubz = bucket_zone_lookup(zone->uz_count);
392 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
393 ubz++;
394 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
395 if (bucket) {
396#ifdef INVARIANTS
397 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
398#endif
399 bucket->ub_cnt = 0;
400 bucket->ub_entries = ubz->ubz_entries;
401 }
402
403 return (bucket);
404}
405
406static void
407bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
408{
409 struct uma_bucket_zone *ubz;
410
411 KASSERT(bucket->ub_cnt == 0,
412 ("bucket_free: Freeing a non free bucket."));
413 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
414 udata = (void *)(uintptr_t)zone->uz_flags;
415 ubz = bucket_zone_lookup(bucket->ub_entries);
416 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
417}
418
419static void
420bucket_zone_drain(void)
421{
422 struct uma_bucket_zone *ubz;
423
424 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
425 zone_drain(ubz->ubz_zone);
426}
427
428static void
429zone_log_warning(uma_zone_t zone)
430{
431 static const struct timeval warninterval = { 300, 0 };
432
433 if (!zone_warnings || zone->uz_warning == NULL)
434 return;
435
436 if (ratecheck(&zone->uz_ratecheck, &warninterval))
437 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
438}
439
440static inline void
441zone_maxaction(uma_zone_t zone)
442{
443
444 if (zone->uz_maxaction.ta_func != NULL)
445 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
446}
447
448static void
449zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
450{
451 uma_klink_t klink;
452
453 LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
454 kegfn(klink->kl_keg);
455}
456
457/*
458 * Routine called by timeout which is used to fire off some time interval
459 * based calculations. (stats, hash size, etc.)
460 *
461 * Arguments:
462 * arg Unused
463 *
464 * Returns:
465 * Nothing
466 */
467static void
468uma_timeout(void *unused)
469{
470 bucket_enable();
471 zone_foreach(zone_timeout);
472
473 /* Reschedule this event */
474 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
475}
476
477/*
478 * Routine to perform timeout driven calculations. This expands the
479 * hashes and does per cpu statistics aggregation.
480 *
481 * Returns nothing.
482 */
483static void
484keg_timeout(uma_keg_t keg)
485{
486
487 KEG_LOCK(keg);
488 /*
489 * Expand the keg hash table.
490 *
491 * This is done if the number of slabs is larger than the hash size.
492 * What I'm trying to do here is completely reduce collisions. This
493 * may be a little aggressive. Should I allow for two collisions max?
494 */
495 if (keg->uk_flags & UMA_ZONE_HASH &&
496 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
497 struct uma_hash newhash;
498 struct uma_hash oldhash;
499 int ret;
500
501 /*
502 * This is so involved because allocating and freeing
503 * while the keg lock is held will lead to deadlock.
504 * I have to do everything in stages and check for
505 * races.
506 */
507 newhash = keg->uk_hash;
508 KEG_UNLOCK(keg);
509 ret = hash_alloc(&newhash);
510 KEG_LOCK(keg);
511 if (ret) {
512 if (hash_expand(&keg->uk_hash, &newhash)) {
513 oldhash = keg->uk_hash;
514 keg->uk_hash = newhash;
515 } else
516 oldhash = newhash;
517
518 KEG_UNLOCK(keg);
519 hash_free(&oldhash);
520 return;
521 }
522 }
523 KEG_UNLOCK(keg);
524}
525
526static void
527zone_timeout(uma_zone_t zone)
528{
529
530 zone_foreach_keg(zone, &keg_timeout);
531}
532
533/*
534 * Allocate and zero fill the next sized hash table from the appropriate
535 * backing store.
536 *
537 * Arguments:
538 * hash A new hash structure with the old hash size in uh_hashsize
539 *
540 * Returns:
541 * 1 on sucess and 0 on failure.
542 */
543static int
544hash_alloc(struct uma_hash *hash)
545{
546 int oldsize;
547 int alloc;
548
549 oldsize = hash->uh_hashsize;
550
551 /* We're just going to go to a power of two greater */
552 if (oldsize) {
553 hash->uh_hashsize = oldsize * 2;
554 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
555 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
556 M_UMAHASH, M_NOWAIT);
557 } else {
558 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
559 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
560 M_WAITOK);
561 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
562 }
563 if (hash->uh_slab_hash) {
564 bzero(hash->uh_slab_hash, alloc);
565 hash->uh_hashmask = hash->uh_hashsize - 1;
566 return (1);
567 }
568
569 return (0);
570}
571
572/*
573 * Expands the hash table for HASH zones. This is done from zone_timeout
574 * to reduce collisions. This must not be done in the regular allocation
575 * path, otherwise, we can recurse on the vm while allocating pages.
576 *
577 * Arguments:
578 * oldhash The hash you want to expand
579 * newhash The hash structure for the new table
580 *
581 * Returns:
582 * Nothing
583 *
584 * Discussion:
585 */
586static int
587hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
588{
589 uma_slab_t slab;
590 int hval;
591 int i;
592
593 if (!newhash->uh_slab_hash)
594 return (0);
595
596 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
597 return (0);
598
599 /*
600 * I need to investigate hash algorithms for resizing without a
601 * full rehash.
602 */
603
604 for (i = 0; i < oldhash->uh_hashsize; i++)
605 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
606 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
607 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
608 hval = UMA_HASH(newhash, slab->us_data);
609 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
610 slab, us_hlink);
611 }
612
613 return (1);
614}
615
616/*
617 * Free the hash bucket to the appropriate backing store.
618 *
619 * Arguments:
620 * slab_hash The hash bucket we're freeing
621 * hashsize The number of entries in that hash bucket
622 *
623 * Returns:
624 * Nothing
625 */
626static void
627hash_free(struct uma_hash *hash)
628{
629 if (hash->uh_slab_hash == NULL)
630 return;
631 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
632 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
633 else
634 free(hash->uh_slab_hash, M_UMAHASH);
635}
636
637/*
638 * Frees all outstanding items in a bucket
639 *
640 * Arguments:
641 * zone The zone to free to, must be unlocked.
642 * bucket The free/alloc bucket with items, cpu queue must be locked.
643 *
644 * Returns:
645 * Nothing
646 */
647
648static void
649bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
650{
651 int i;
652
653 if (bucket == NULL)
654 return;
655
656 if (zone->uz_fini)
657 for (i = 0; i < bucket->ub_cnt; i++)
658 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
659 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
660 bucket->ub_cnt = 0;
661}
662
663/*
664 * Drains the per cpu caches for a zone.
665 *
666 * NOTE: This may only be called while the zone is being turn down, and not
667 * during normal operation. This is necessary in order that we do not have
668 * to migrate CPUs to drain the per-CPU caches.
669 *
670 * Arguments:
671 * zone The zone to drain, must be unlocked.
672 *
673 * Returns:
674 * Nothing
675 */
676static void
677cache_drain(uma_zone_t zone)
678{
679 uma_cache_t cache;
680 int cpu;
681
682 /*
683 * XXX: It is safe to not lock the per-CPU caches, because we're
684 * tearing down the zone anyway. I.e., there will be no further use
685 * of the caches at this point.
686 *
687 * XXX: It would good to be able to assert that the zone is being
688 * torn down to prevent improper use of cache_drain().
689 *
690 * XXX: We lock the zone before passing into bucket_cache_drain() as
691 * it is used elsewhere. Should the tear-down path be made special
692 * there in some form?
693 */
694 CPU_FOREACH(cpu) {
695 cache = &zone->uz_cpu[cpu];
696 bucket_drain(zone, cache->uc_allocbucket);
697 bucket_drain(zone, cache->uc_freebucket);
698 if (cache->uc_allocbucket != NULL)
699 bucket_free(zone, cache->uc_allocbucket, NULL);
700 if (cache->uc_freebucket != NULL)
701 bucket_free(zone, cache->uc_freebucket, NULL);
702 cache->uc_allocbucket = cache->uc_freebucket = NULL;
703 }
704 ZONE_LOCK(zone);
705 bucket_cache_drain(zone);
706 ZONE_UNLOCK(zone);
707}
708
709static void
710cache_shrink(uma_zone_t zone)
711{
712
713 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
714 return;
715
716 ZONE_LOCK(zone);
717 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
718 ZONE_UNLOCK(zone);
719}
720
721static void
722cache_drain_safe_cpu(uma_zone_t zone)
723{
724 uma_cache_t cache;
725 uma_bucket_t b1, b2;
726
727 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
728 return;
729
730 b1 = b2 = NULL;
731 ZONE_LOCK(zone);
732 critical_enter();
733 cache = &zone->uz_cpu[curcpu];
734 if (cache->uc_allocbucket) {
735 if (cache->uc_allocbucket->ub_cnt != 0)
736 LIST_INSERT_HEAD(&zone->uz_buckets,
737 cache->uc_allocbucket, ub_link);
738 else
739 b1 = cache->uc_allocbucket;
740 cache->uc_allocbucket = NULL;
741 }
742 if (cache->uc_freebucket) {
743 if (cache->uc_freebucket->ub_cnt != 0)
744 LIST_INSERT_HEAD(&zone->uz_buckets,
745 cache->uc_freebucket, ub_link);
746 else
747 b2 = cache->uc_freebucket;
748 cache->uc_freebucket = NULL;
749 }
750 critical_exit();
751 ZONE_UNLOCK(zone);
752 if (b1)
753 bucket_free(zone, b1, NULL);
754 if (b2)
755 bucket_free(zone, b2, NULL);
756}
757
758/*
759 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
760 * This is an expensive call because it needs to bind to all CPUs
761 * one by one and enter a critical section on each of them in order
762 * to safely access their cache buckets.
763 * Zone lock must not be held on call this function.
764 */
765static void
766cache_drain_safe(uma_zone_t zone)
767{
768 int cpu;
769
770 /*
771 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
772 */
773 if (zone)
774 cache_shrink(zone);
775 else
776 zone_foreach(cache_shrink);
777
778 CPU_FOREACH(cpu) {
779 thread_lock(curthread);
780 sched_bind(curthread, cpu);
781 thread_unlock(curthread);
782
783 if (zone)
784 cache_drain_safe_cpu(zone);
785 else
786 zone_foreach(cache_drain_safe_cpu);
787 }
788 thread_lock(curthread);
789 sched_unbind(curthread);
790 thread_unlock(curthread);
791}
792
793/*
794 * Drain the cached buckets from a zone. Expects a locked zone on entry.
795 */
796static void
797bucket_cache_drain(uma_zone_t zone)
798{
799 uma_bucket_t bucket;
800
801 /*
802 * Drain the bucket queues and free the buckets, we just keep two per
803 * cpu (alloc/free).
804 */
805 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
806 LIST_REMOVE(bucket, ub_link);
807 ZONE_UNLOCK(zone);
808 bucket_drain(zone, bucket);
809 bucket_free(zone, bucket, NULL);
810 ZONE_LOCK(zone);
811 }
812
813 /*
814 * Shrink further bucket sizes. Price of single zone lock collision
815 * is probably lower then price of global cache drain.
816 */
817 if (zone->uz_count > zone->uz_count_min)
818 zone->uz_count--;
819}
820
821static void
822keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
823{
824 uint8_t *mem;
825 int i;
826 uint8_t flags;
827
828 mem = slab->us_data;
829 flags = slab->us_flags;
830 i = start;
831 if (keg->uk_fini != NULL) {
832 for (i--; i > -1; i--)
833 keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
834 keg->uk_size);
835 }
836 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
837 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
838#ifdef UMA_DEBUG
839 printf("%s: Returning %d bytes.\n", keg->uk_name,
840 PAGE_SIZE * keg->uk_ppera);
841#endif
842 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
843}
844
845/*
846 * Frees pages from a keg back to the system. This is done on demand from
847 * the pageout daemon.
848 *
849 * Returns nothing.
850 */
851static void
852keg_drain(uma_keg_t keg)
853{
854 struct slabhead freeslabs = { 0 };
855 uma_slab_t slab;
856 uma_slab_t n;
857
858 /*
859 * We don't want to take pages from statically allocated kegs at this
860 * time
861 */
862 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
863 return;
864
865#ifdef UMA_DEBUG
866 printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
867#endif
868 KEG_LOCK(keg);
869 if (keg->uk_free == 0)
870 goto finished;
871
872 slab = LIST_FIRST(&keg->uk_free_slab);
873 while (slab) {
874 n = LIST_NEXT(slab, us_link);
875
876 /* We have no where to free these to */
877 if (slab->us_flags & UMA_SLAB_BOOT) {
878 slab = n;
879 continue;
880 }
881
882 LIST_REMOVE(slab, us_link);
883 keg->uk_pages -= keg->uk_ppera;
884 keg->uk_free -= keg->uk_ipers;
885
886 if (keg->uk_flags & UMA_ZONE_HASH)
887 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
888
889 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
890
891 slab = n;
892 }
893finished:
894 KEG_UNLOCK(keg);
895
896 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
897 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
898 keg_free_slab(keg, slab, keg->uk_ipers);
899 }
900}
901
902static void
903zone_drain_wait(uma_zone_t zone, int waitok)
904{
905
906 /*
907 * Set draining to interlock with zone_dtor() so we can release our
908 * locks as we go. Only dtor() should do a WAITOK call since it
909 * is the only call that knows the structure will still be available
910 * when it wakes up.
911 */
912 ZONE_LOCK(zone);
913 while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
914 if (waitok == M_NOWAIT)
915 goto out;
916 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
917 }
918 zone->uz_flags |= UMA_ZFLAG_DRAINING;
919 bucket_cache_drain(zone);
920 ZONE_UNLOCK(zone);
921 /*
922 * The DRAINING flag protects us from being freed while
923 * we're running. Normally the uma_rwlock would protect us but we
924 * must be able to release and acquire the right lock for each keg.
925 */
926 zone_foreach_keg(zone, &keg_drain);
927 ZONE_LOCK(zone);
928 zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
929 wakeup(zone);
930out:
931 ZONE_UNLOCK(zone);
932}
933
934void
935zone_drain(uma_zone_t zone)
936{
937
938 zone_drain_wait(zone, M_NOWAIT);
939}
940
941/*
942 * Allocate a new slab for a keg. This does not insert the slab onto a list.
943 *
944 * Arguments:
945 * wait Shall we wait?
946 *
947 * Returns:
948 * The slab that was allocated or NULL if there is no memory and the
949 * caller specified M_NOWAIT.
950 */
951static uma_slab_t
952keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
953{
954 uma_slabrefcnt_t slabref;
955 uma_alloc allocf;
956 uma_slab_t slab;
957 uint8_t *mem;
958 uint8_t flags;
959 int i;
960
961 mtx_assert(&keg->uk_lock, MA_OWNED);
962 slab = NULL;
963 mem = NULL;
964
965#ifdef UMA_DEBUG
966 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name);
967#endif
968 allocf = keg->uk_allocf;
969 KEG_UNLOCK(keg);
970
971 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
972 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
973 if (slab == NULL)
974 goto out;
975 }
976
977 /*
978 * This reproduces the old vm_zone behavior of zero filling pages the
979 * first time they are added to a zone.
980 *
981 * Malloced items are zeroed in uma_zalloc.
982 */
983
984 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
985 wait |= M_ZERO;
986 else
987 wait &= ~M_ZERO;
988
989 if (keg->uk_flags & UMA_ZONE_NODUMP)
990 wait |= M_NODUMP;
991
992 /* zone is passed for legacy reasons. */
993 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
994 if (mem == NULL) {
995 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
996 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
997 slab = NULL;
998 goto out;
999 }
1000
1001 /* Point the slab into the allocated memory */
1002 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
1003 slab = (uma_slab_t )(mem + keg->uk_pgoff);
1004
1005 if (keg->uk_flags & UMA_ZONE_VTOSLAB)
1006 for (i = 0; i < keg->uk_ppera; i++)
1007 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
1008
1009 slab->us_keg = keg;
1010 slab->us_data = mem;
1011 slab->us_freecount = keg->uk_ipers;
1012 slab->us_flags = flags;
1013 BIT_FILL(SLAB_SETSIZE, &slab->us_free);
1014#ifdef INVARIANTS
1015 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1016#endif
1017 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1018 slabref = (uma_slabrefcnt_t)slab;
1019 for (i = 0; i < keg->uk_ipers; i++)
1020 slabref->us_refcnt[i] = 0;
1021 }
1022
1023 if (keg->uk_init != NULL) {
1024 for (i = 0; i < keg->uk_ipers; i++)
1025 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1026 keg->uk_size, wait) != 0)
1027 break;
1028 if (i != keg->uk_ipers) {
1029 keg_free_slab(keg, slab, i);
1030 slab = NULL;
1031 goto out;
1032 }
1033 }
1034out:
1035 KEG_LOCK(keg);
1036
1037 if (slab != NULL) {
1038 if (keg->uk_flags & UMA_ZONE_HASH)
1039 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1040
1041 keg->uk_pages += keg->uk_ppera;
1042 keg->uk_free += keg->uk_ipers;
1043 }
1044
1045 return (slab);
1046}
1047
1048/*
1049 * This function is intended to be used early on in place of page_alloc() so
1050 * that we may use the boot time page cache to satisfy allocations before
1051 * the VM is ready.
1052 */
1053static void *
1054startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1055{
1056 uma_keg_t keg;
1057 uma_slab_t tmps;
1058 int pages, check_pages;
1059
1060 keg = zone_first_keg(zone);
1061 pages = howmany(bytes, PAGE_SIZE);
1062 check_pages = pages - 1;
1063 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1064
1065 /*
1066 * Check our small startup cache to see if it has pages remaining.
1067 */
1068 mtx_lock(&uma_boot_pages_mtx);
1069
1070 /* First check if we have enough room. */
1071 tmps = LIST_FIRST(&uma_boot_pages);
1072 while (tmps != NULL && check_pages-- > 0)
1073 tmps = LIST_NEXT(tmps, us_link);
1074 if (tmps != NULL) {
1075 /*
1076 * It's ok to lose tmps references. The last one will
1077 * have tmps->us_data pointing to the start address of
1078 * "pages" contiguous pages of memory.
1079 */
1080 while (pages-- > 0) {
1081 tmps = LIST_FIRST(&uma_boot_pages);
1082 LIST_REMOVE(tmps, us_link);
1083 }
1084 mtx_unlock(&uma_boot_pages_mtx);
1085 *pflag = tmps->us_flags;
1086 return (tmps->us_data);
1087 }
1088 mtx_unlock(&uma_boot_pages_mtx);
1089 if (booted < UMA_STARTUP2)
1090 panic("UMA: Increase vm.boot_pages");
1091 /*
1092 * Now that we've booted reset these users to their real allocator.
1093 */
1094#ifdef UMA_MD_SMALL_ALLOC
1095 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1096#else
1097 keg->uk_allocf = page_alloc;
1098#endif
1099 return keg->uk_allocf(zone, bytes, pflag, wait);
1100}
1101
1102/*
1103 * Allocates a number of pages from the system
1104 *
1105 * Arguments:
1106 * bytes The number of bytes requested
1107 * wait Shall we wait?
1108 *
1109 * Returns:
1110 * A pointer to the alloced memory or possibly
1111 * NULL if M_NOWAIT is set.
1112 */
1113static void *
1114page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1115{
1116 void *p; /* Returned page */
1117
1118 *pflag = UMA_SLAB_KMEM;
1119 p = (void *) kmem_malloc(kmem_arena, bytes, wait);
1120
1121 return (p);
1122}
1123
1124/*
1125 * Allocates a number of pages from within an object
1126 *
1127 * Arguments:
1128 * bytes The number of bytes requested
1129 * wait Shall we wait?
1130 *
1131 * Returns:
1132 * A pointer to the alloced memory or possibly
1133 * NULL if M_NOWAIT is set.
1134 */
1135static void *
1136noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait)
1137{
1138 TAILQ_HEAD(, vm_page) alloctail;
1139 u_long npages;
1140 vm_offset_t retkva, zkva;
1141 vm_page_t p, p_next;
1142 uma_keg_t keg;
1143
1144 TAILQ_INIT(&alloctail);
1145 keg = zone_first_keg(zone);
1146
1147 npages = howmany(bytes, PAGE_SIZE);
1148 while (npages > 0) {
1149 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1150 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1151 if (p != NULL) {
1152 /*
1153 * Since the page does not belong to an object, its
1154 * listq is unused.
1155 */
1156 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1157 npages--;
1158 continue;
1159 }
1160 if (wait & M_WAITOK) {
1161 VM_WAIT;
1162 continue;
1163 }
1164
1165 /*
1166 * Page allocation failed, free intermediate pages and
1167 * exit.
1168 */
1169 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1170 vm_page_unwire(p, PQ_NONE);
1171 vm_page_free(p);
1172 }
1173 return (NULL);
1174 }
1175 *flags = UMA_SLAB_PRIV;
1176 zkva = keg->uk_kva +
1177 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1178 retkva = zkva;
1179 TAILQ_FOREACH(p, &alloctail, listq) {
1180 pmap_qenter(zkva, &p, 1);
1181 zkva += PAGE_SIZE;
1182 }
1183
1184 return ((void *)retkva);
1185}
1186
1187/*
1188 * Frees a number of pages to the system
1189 *
1190 * Arguments:
1191 * mem A pointer to the memory to be freed
1192 * size The size of the memory being freed
1193 * flags The original p->us_flags field
1194 *
1195 * Returns:
1196 * Nothing
1197 */
1198static void
1199page_free(void *mem, vm_size_t size, uint8_t flags)
1200{
1201 struct vmem *vmem;
1202
1203 if (flags & UMA_SLAB_KMEM)
1204 vmem = kmem_arena;
1205 else if (flags & UMA_SLAB_KERNEL)
1206 vmem = kernel_arena;
1207 else
1208 panic("UMA: page_free used with invalid flags %d", flags);
1209
1210 kmem_free(vmem, (vm_offset_t)mem, size);
1211}
1212
1213/*
1214 * Zero fill initializer
1215 *
1216 * Arguments/Returns follow uma_init specifications
1217 */
1218static int
1219zero_init(void *mem, int size, int flags)
1220{
1221 bzero(mem, size);
1222 return (0);
1223}
1224
1225/*
1226 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1227 *
1228 * Arguments
1229 * keg The zone we should initialize
1230 *
1231 * Returns
1232 * Nothing
1233 */
1234static void
1235keg_small_init(uma_keg_t keg)
1236{
1237 u_int rsize;
1238 u_int memused;
1239 u_int wastedspace;
1240 u_int shsize;
1241
1242 if (keg->uk_flags & UMA_ZONE_PCPU) {
1243 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1244
1245 keg->uk_slabsize = sizeof(struct pcpu);
1246 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1247 PAGE_SIZE);
1248 } else {
1249 keg->uk_slabsize = UMA_SLAB_SIZE;
1250 keg->uk_ppera = 1;
1251 }
1252
1253 /*
1254 * Calculate the size of each allocation (rsize) according to
1255 * alignment. If the requested size is smaller than we have
1256 * allocation bits for we round it up.
1257 */
1258 rsize = keg->uk_size;
1259 if (rsize < keg->uk_slabsize / SLAB_SETSIZE)
1260 rsize = keg->uk_slabsize / SLAB_SETSIZE;
1261 if (rsize & keg->uk_align)
1262 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1263 keg->uk_rsize = rsize;
1264
1265 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1266 keg->uk_rsize < sizeof(struct pcpu),
1267 ("%s: size %u too large", __func__, keg->uk_rsize));
1268
1269 if (keg->uk_flags & UMA_ZONE_REFCNT)
1270 rsize += sizeof(uint32_t);
1271
1272 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1273 shsize = 0;
1274 else
1275 shsize = sizeof(struct uma_slab);
1276
1277 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1278 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1279 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1280
1281 memused = keg->uk_ipers * rsize + shsize;
1282 wastedspace = keg->uk_slabsize - memused;
1283
1284 /*
1285 * We can't do OFFPAGE if we're internal or if we've been
1286 * asked to not go to the VM for buckets. If we do this we
1287 * may end up going to the VM for slabs which we do not
1288 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1289 * of UMA_ZONE_VM, which clearly forbids it.
1290 */
1291 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1292 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1293 return;
1294
1295 /*
1296 * See if using an OFFPAGE slab will limit our waste. Only do
1297 * this if it permits more items per-slab.
1298 *
1299 * XXX We could try growing slabsize to limit max waste as well.
1300 * Historically this was not done because the VM could not
1301 * efficiently handle contiguous allocations.
1302 */
1303 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1304 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1305 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1306 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1307 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1308#ifdef UMA_DEBUG
1309 printf("UMA decided we need offpage slab headers for "
1310 "keg: %s, calculated wastedspace = %d, "
1311 "maximum wasted space allowed = %d, "
1312 "calculated ipers = %d, "
1313 "new wasted space = %d\n", keg->uk_name, wastedspace,
1314 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1315 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1316#endif
1317 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1318 }
1319
1320 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1321 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1322 keg->uk_flags |= UMA_ZONE_HASH;
1323}
1324
1325/*
1326 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1327 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1328 * more complicated.
1329 *
1330 * Arguments
1331 * keg The keg we should initialize
1332 *
1333 * Returns
1334 * Nothing
1335 */
1336static void
1337keg_large_init(uma_keg_t keg)
1338{
1339 u_int shsize;
1340
1341 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1342 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1343 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1344 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1345 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1346
1347 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1348 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1349 keg->uk_ipers = 1;
1350 keg->uk_rsize = keg->uk_size;
1351
1352 /* We can't do OFFPAGE if we're internal, bail out here. */
1353 if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1354 return;
1355
1356 /* Check whether we have enough space to not do OFFPAGE. */
1357 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1358 shsize = sizeof(struct uma_slab);
1359 if (keg->uk_flags & UMA_ZONE_REFCNT)
1360 shsize += keg->uk_ipers * sizeof(uint32_t);
1361 if (shsize & UMA_ALIGN_PTR)
1362 shsize = (shsize & ~UMA_ALIGN_PTR) +
1363 (UMA_ALIGN_PTR + 1);
1364
1365 if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize)
1366 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1367 }
1368
1369 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1370 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1371 keg->uk_flags |= UMA_ZONE_HASH;
1372}
1373
1374static void
1375keg_cachespread_init(uma_keg_t keg)
1376{
1377 int alignsize;
1378 int trailer;
1379 int pages;
1380 int rsize;
1381
1382 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1383 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1384
1385 alignsize = keg->uk_align + 1;
1386 rsize = keg->uk_size;
1387 /*
1388 * We want one item to start on every align boundary in a page. To
1389 * do this we will span pages. We will also extend the item by the
1390 * size of align if it is an even multiple of align. Otherwise, it
1391 * would fall on the same boundary every time.
1392 */
1393 if (rsize & keg->uk_align)
1394 rsize = (rsize & ~keg->uk_align) + alignsize;
1395 if ((rsize & alignsize) == 0)
1396 rsize += alignsize;
1397 trailer = rsize - keg->uk_size;
1398 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1399 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1400 keg->uk_rsize = rsize;
1401 keg->uk_ppera = pages;
1402 keg->uk_slabsize = UMA_SLAB_SIZE;
1403 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1404 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1405 KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1406 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1407 keg->uk_ipers));
1408}
1409
1410/*
1411 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1412 * the keg onto the global keg list.
1413 *
1414 * Arguments/Returns follow uma_ctor specifications
1415 * udata Actually uma_kctor_args
1416 */
1417static int
1418keg_ctor(void *mem, int size, void *udata, int flags)
1419{
1420 struct uma_kctor_args *arg = udata;
1421 uma_keg_t keg = mem;
1422 uma_zone_t zone;
1423
1424 bzero(keg, size);
1425 keg->uk_size = arg->size;
1426 keg->uk_init = arg->uminit;
1427 keg->uk_fini = arg->fini;
1428 keg->uk_align = arg->align;
1429 keg->uk_free = 0;
1430 keg->uk_reserve = 0;
1431 keg->uk_pages = 0;
1432 keg->uk_flags = arg->flags;
1433 keg->uk_allocf = page_alloc;
1434 keg->uk_freef = page_free;
1435 keg->uk_slabzone = NULL;
1436
1437 /*
1438 * The master zone is passed to us at keg-creation time.
1439 */
1440 zone = arg->zone;
1441 keg->uk_name = zone->uz_name;
1442
1443 if (arg->flags & UMA_ZONE_VM)
1444 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1445
1446 if (arg->flags & UMA_ZONE_ZINIT)
1447 keg->uk_init = zero_init;
1448
1449 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1450 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1451
1452 if (arg->flags & UMA_ZONE_PCPU)
1453#ifdef SMP
1454 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1455#else
1456 keg->uk_flags &= ~UMA_ZONE_PCPU;
1457#endif
1458
1459 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1460 keg_cachespread_init(keg);
1461 } else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1462 if (keg->uk_size >
1463 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) -
1464 sizeof(uint32_t)))
1465 keg_large_init(keg);
1466 else
1467 keg_small_init(keg);
1468 } else {
1469 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1470 keg_large_init(keg);
1471 else
1472 keg_small_init(keg);
1473 }
1474
1475 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1476 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1477 if (keg->uk_ipers > uma_max_ipers_ref)
1478 panic("Too many ref items per zone: %d > %d\n",
1479 keg->uk_ipers, uma_max_ipers_ref);
1480 keg->uk_slabzone = slabrefzone;
1481 } else
1482 keg->uk_slabzone = slabzone;
1483 }
1484
1485 /*
1486 * If we haven't booted yet we need allocations to go through the
1487 * startup cache until the vm is ready.
1488 */
1489 if (keg->uk_ppera == 1) {
1490#ifdef UMA_MD_SMALL_ALLOC
1491 keg->uk_allocf = uma_small_alloc;
1492 keg->uk_freef = uma_small_free;
1493
1494 if (booted < UMA_STARTUP)
1495 keg->uk_allocf = startup_alloc;
1496#else
1497 if (booted < UMA_STARTUP2)
1498 keg->uk_allocf = startup_alloc;
1499#endif
1500 } else if (booted < UMA_STARTUP2 &&
1501 (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1502 keg->uk_allocf = startup_alloc;
1503
1504 /*
1505 * Initialize keg's lock
1506 */
1507 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1508
1509 /*
1510 * If we're putting the slab header in the actual page we need to
1511 * figure out where in each page it goes. This calculates a right
1512 * justified offset into the memory on an ALIGN_PTR boundary.
1513 */
1514 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1515 u_int totsize;
1516
1517 /* Size of the slab struct and free list */
1518 totsize = sizeof(struct uma_slab);
1519
1520 /* Size of the reference counts. */
1521 if (keg->uk_flags & UMA_ZONE_REFCNT)
1522 totsize += keg->uk_ipers * sizeof(uint32_t);
1523
1524 if (totsize & UMA_ALIGN_PTR)
1525 totsize = (totsize & ~UMA_ALIGN_PTR) +
1526 (UMA_ALIGN_PTR + 1);
1527 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1528
1529 /*
1530 * The only way the following is possible is if with our
1531 * UMA_ALIGN_PTR adjustments we are now bigger than
1532 * UMA_SLAB_SIZE. I haven't checked whether this is
1533 * mathematically possible for all cases, so we make
1534 * sure here anyway.
1535 */
1536 totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1537 if (keg->uk_flags & UMA_ZONE_REFCNT)
1538 totsize += keg->uk_ipers * sizeof(uint32_t);
1539 if (totsize > PAGE_SIZE * keg->uk_ppera) {
1540 printf("zone %s ipers %d rsize %d size %d\n",
1541 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1542 keg->uk_size);
1543 panic("UMA slab won't fit.");
1544 }
1545 }
1546
1547 if (keg->uk_flags & UMA_ZONE_HASH)
1548 hash_alloc(&keg->uk_hash);
1549
1550#ifdef UMA_DEBUG
1551 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1552 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1553 keg->uk_ipers, keg->uk_ppera,
1554 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1555#endif
1556
1557 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1558
1559 rw_wlock(&uma_rwlock);
1560 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1561 rw_wunlock(&uma_rwlock);
1562 return (0);
1563}
1564
1565/*
1566 * Zone header ctor. This initializes all fields, locks, etc.
1567 *
1568 * Arguments/Returns follow uma_ctor specifications
1569 * udata Actually uma_zctor_args
1570 */
1571static int
1572zone_ctor(void *mem, int size, void *udata, int flags)
1573{
1574 struct uma_zctor_args *arg = udata;
1575 uma_zone_t zone = mem;
1576 uma_zone_t z;
1577 uma_keg_t keg;
1578
1579 bzero(zone, size);
1580 zone->uz_name = arg->name;
1581 zone->uz_ctor = arg->ctor;
1582 zone->uz_dtor = arg->dtor;
1583 zone->uz_slab = zone_fetch_slab;
1584 zone->uz_init = NULL;
1585 zone->uz_fini = NULL;
1586 zone->uz_allocs = 0;
1587 zone->uz_frees = 0;
1588 zone->uz_fails = 0;
1589 zone->uz_sleeps = 0;
1590 zone->uz_count = 0;
1591 zone->uz_count_min = 0;
1592 zone->uz_flags = 0;
1593 zone->uz_warning = NULL;
1594 timevalclear(&zone->uz_ratecheck);
1595 keg = arg->keg;
1596
1597 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1598
1599 /*
1600 * This is a pure cache zone, no kegs.
1601 */
1602 if (arg->import) {
1603 if (arg->flags & UMA_ZONE_VM)
1604 arg->flags |= UMA_ZFLAG_CACHEONLY;
1605 zone->uz_flags = arg->flags;
1606 zone->uz_size = arg->size;
1607 zone->uz_import = arg->import;
1608 zone->uz_release = arg->release;
1609 zone->uz_arg = arg->arg;
1610 zone->uz_lockptr = &zone->uz_lock;
1611 rw_wlock(&uma_rwlock);
1612 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1613 rw_wunlock(&uma_rwlock);
1614 goto out;
1615 }
1616
1617 /*
1618 * Use the regular zone/keg/slab allocator.
1619 */
1620 zone->uz_import = (uma_import)zone_import;
1621 zone->uz_release = (uma_release)zone_release;
1622 zone->uz_arg = zone;
1623
1624 if (arg->flags & UMA_ZONE_SECONDARY) {
1625 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1626 zone->uz_init = arg->uminit;
1627 zone->uz_fini = arg->fini;
1628 zone->uz_lockptr = &keg->uk_lock;
1629 zone->uz_flags |= UMA_ZONE_SECONDARY;
1630 rw_wlock(&uma_rwlock);
1631 ZONE_LOCK(zone);
1632 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1633 if (LIST_NEXT(z, uz_link) == NULL) {
1634 LIST_INSERT_AFTER(z, zone, uz_link);
1635 break;
1636 }
1637 }
1638 ZONE_UNLOCK(zone);
1639 rw_wunlock(&uma_rwlock);
1640 } else if (keg == NULL) {
1641 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1642 arg->align, arg->flags)) == NULL)
1643 return (ENOMEM);
1644 } else {
1645 struct uma_kctor_args karg;
1646 int error;
1647
1648 /* We should only be here from uma_startup() */
1649 karg.size = arg->size;
1650 karg.uminit = arg->uminit;
1651 karg.fini = arg->fini;
1652 karg.align = arg->align;
1653 karg.flags = arg->flags;
1654 karg.zone = zone;
1655 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1656 flags);
1657 if (error)
1658 return (error);
1659 }
1660
1661 /*
1662 * Link in the first keg.
1663 */
1664 zone->uz_klink.kl_keg = keg;
1665 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1666 zone->uz_lockptr = &keg->uk_lock;
1667 zone->uz_size = keg->uk_size;
1668 zone->uz_flags |= (keg->uk_flags &
1669 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1670
1671 /*
1672 * Some internal zones don't have room allocated for the per cpu
1673 * caches. If we're internal, bail out here.
1674 */
1675 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1676 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1677 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1678 return (0);
1679 }
1680
1681out:
1682 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1683 zone->uz_count = bucket_select(zone->uz_size);
1684 else
1685 zone->uz_count = BUCKET_MAX;
1686 zone->uz_count_min = zone->uz_count;
1687
1688 return (0);
1689}
1690
1691/*
1692 * Keg header dtor. This frees all data, destroys locks, frees the hash
1693 * table and removes the keg from the global list.
1694 *
1695 * Arguments/Returns follow uma_dtor specifications
1696 * udata unused
1697 */
1698static void
1699keg_dtor(void *arg, int size, void *udata)
1700{
1701 uma_keg_t keg;
1702
1703 keg = (uma_keg_t)arg;
1704 KEG_LOCK(keg);
1705 if (keg->uk_free != 0) {
1706 printf("Freed UMA keg (%s) was not empty (%d items). "
1707 " Lost %d pages of memory.\n",
1708 keg->uk_name ? keg->uk_name : "",
1709 keg->uk_free, keg->uk_pages);
1710 }
1711 KEG_UNLOCK(keg);
1712
1713 hash_free(&keg->uk_hash);
1714
1715 KEG_LOCK_FINI(keg);
1716}
1717
1718/*
1719 * Zone header dtor.
1720 *
1721 * Arguments/Returns follow uma_dtor specifications
1722 * udata unused
1723 */
1724static void
1725zone_dtor(void *arg, int size, void *udata)
1726{
1727 uma_klink_t klink;
1728 uma_zone_t zone;
1729 uma_keg_t keg;
1730
1731 zone = (uma_zone_t)arg;
1732 keg = zone_first_keg(zone);
1733
1734 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1735 cache_drain(zone);
1736
1737 rw_wlock(&uma_rwlock);
1738 LIST_REMOVE(zone, uz_link);
1739 rw_wunlock(&uma_rwlock);
1740 /*
1741 * XXX there are some races here where
1742 * the zone can be drained but zone lock
1743 * released and then refilled before we
1744 * remove it... we dont care for now
1745 */
1746 zone_drain_wait(zone, M_WAITOK);
1747 /*
1748 * Unlink all of our kegs.
1749 */
1750 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1751 klink->kl_keg = NULL;
1752 LIST_REMOVE(klink, kl_link);
1753 if (klink == &zone->uz_klink)
1754 continue;
1755 free(klink, M_TEMP);
1756 }
1757 /*
1758 * We only destroy kegs from non secondary zones.
1759 */
1760 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1761 rw_wlock(&uma_rwlock);
1762 LIST_REMOVE(keg, uk_link);
1763 rw_wunlock(&uma_rwlock);
1764 zone_free_item(kegs, keg, NULL, SKIP_NONE);
1765 }
1766 ZONE_LOCK_FINI(zone);
1767}
1768
1769/*
1770 * Traverses every zone in the system and calls a callback
1771 *
1772 * Arguments:
1773 * zfunc A pointer to a function which accepts a zone
1774 * as an argument.
1775 *
1776 * Returns:
1777 * Nothing
1778 */
1779static void
1780zone_foreach(void (*zfunc)(uma_zone_t))
1781{
1782 uma_keg_t keg;
1783 uma_zone_t zone;
1784
1785 rw_rlock(&uma_rwlock);
1786 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1787 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1788 zfunc(zone);
1789 }
1790 rw_runlock(&uma_rwlock);
1791}
1792
1793/* Public functions */
1794/* See uma.h */
1795void
1796uma_startup(void *bootmem, int boot_pages)
1797{
1798 struct uma_zctor_args args;
1799 uma_slab_t slab;
1800 u_int slabsize;
1801 int i;
1802
1803#ifdef UMA_DEBUG
1804 printf("Creating uma keg headers zone and keg.\n");
1805#endif
1806 rw_init(&uma_rwlock, "UMA lock");
1807
1808 /* "manually" create the initial zone */
1809 memset(&args, 0, sizeof(args));
1810 args.name = "UMA Kegs";
1811 args.size = sizeof(struct uma_keg);
1812 args.ctor = keg_ctor;
1813 args.dtor = keg_dtor;
1814 args.uminit = zero_init;
1815 args.fini = NULL;
1816 args.keg = &masterkeg;
1817 args.align = 32 - 1;
1818 args.flags = UMA_ZFLAG_INTERNAL;
1819 /* The initial zone has no Per cpu queues so it's smaller */
1820 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1821
1822#ifdef UMA_DEBUG
1823 printf("Filling boot free list.\n");
1824#endif
1825 for (i = 0; i < boot_pages; i++) {
1826 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1827 slab->us_data = (uint8_t *)slab;
1828 slab->us_flags = UMA_SLAB_BOOT;
1829 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1830 }
1831 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1832
1833#ifdef UMA_DEBUG
1834 printf("Creating uma zone headers zone and keg.\n");
1835#endif
1836 args.name = "UMA Zones";
1837 args.size = sizeof(struct uma_zone) +
1838 (sizeof(struct uma_cache) * (mp_maxid + 1));
1839 args.ctor = zone_ctor;
1840 args.dtor = zone_dtor;
1841 args.uminit = zero_init;
1842 args.fini = NULL;
1843 args.keg = NULL;
1844 args.align = 32 - 1;
1845 args.flags = UMA_ZFLAG_INTERNAL;
1846 /* The initial zone has no Per cpu queues so it's smaller */
1847 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1848
1849#ifdef UMA_DEBUG
1850 printf("Creating slab and hash zones.\n");
1851#endif
1852
1853 /* Now make a zone for slab headers */
1854 slabzone = uma_zcreate("UMA Slabs",
1855 sizeof(struct uma_slab),
1856 NULL, NULL, NULL, NULL,
1857 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1858
1859 /*
1860 * We also create a zone for the bigger slabs with reference
1861 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1862 */
1863 slabsize = sizeof(struct uma_slab_refcnt);
1864 slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1865 slabrefzone = uma_zcreate("UMA RCntSlabs",
1866 slabsize,
1867 NULL, NULL, NULL, NULL,
1868 UMA_ALIGN_PTR,
1869 UMA_ZFLAG_INTERNAL);
1870
1871 hashzone = uma_zcreate("UMA Hash",
1872 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1873 NULL, NULL, NULL, NULL,
1874 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1875
1876 bucket_init();
1877
1878 booted = UMA_STARTUP;
1879
1880#ifdef UMA_DEBUG
1881 printf("UMA startup complete.\n");
1882#endif
1883}
1884
1885/* see uma.h */
1886void
1887uma_startup2(void)
1888{
1889 booted = UMA_STARTUP2;
1890 bucket_enable();
1891 sx_init(&uma_drain_lock, "umadrain");
1892#ifdef UMA_DEBUG
1893 printf("UMA startup2 complete.\n");
1894#endif
1895}
1896
1897/*
1898 * Initialize our callout handle
1899 *
1900 */
1901
1902static void
1903uma_startup3(void)
1904{
1905#ifdef UMA_DEBUG
1906 printf("Starting callout.\n");
1907#endif
1908 callout_init(&uma_callout, 1);
1909 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1910#ifdef UMA_DEBUG
1911 printf("UMA startup3 complete.\n");
1912#endif
1913}
1914
1915static uma_keg_t
1916uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1917 int align, uint32_t flags)
1918{
1919 struct uma_kctor_args args;
1920
1921 args.size = size;
1922 args.uminit = uminit;
1923 args.fini = fini;
1924 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1925 args.flags = flags;
1926 args.zone = zone;
1927 return (zone_alloc_item(kegs, &args, M_WAITOK));
1928}
1929
1930/* See uma.h */
1931void
1932uma_set_align(int align)
1933{
1934
1935 if (align != UMA_ALIGN_CACHE)
1936 uma_align_cache = align;
1937}
1938
1939/* See uma.h */
1940uma_zone_t
1941uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1942 uma_init uminit, uma_fini fini, int align, uint32_t flags)
1943
1944{
1945 struct uma_zctor_args args;
1946 uma_zone_t res;
1947 bool locked;
1948
1949 /* This stuff is essential for the zone ctor */
1950 memset(&args, 0, sizeof(args));
1951 args.name = name;
1952 args.size = size;
1953 args.ctor = ctor;
1954 args.dtor = dtor;
1955 args.uminit = uminit;
1956 args.fini = fini;
1957#ifdef INVARIANTS
1958 /*
1959 * If a zone is being created with an empty constructor and
1960 * destructor, pass UMA constructor/destructor which checks for
1961 * memory use after free.
1962 */
1963 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) &&
1964 ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) {
1965 args.ctor = trash_ctor;
1966 args.dtor = trash_dtor;
1967 args.uminit = trash_init;
1968 args.fini = trash_fini;
1969 }
1970#endif
1971 args.align = align;
1972 args.flags = flags;
1973 args.keg = NULL;
1974
1975 if (booted < UMA_STARTUP2) {
1976 locked = false;
1977 } else {
1978 sx_slock(&uma_drain_lock);
1979 locked = true;
1980 }
1981 res = zone_alloc_item(zones, &args, M_WAITOK);
1982 if (locked)
1983 sx_sunlock(&uma_drain_lock);
1984 return (res);
1985}
1986
1987/* See uma.h */
1988uma_zone_t
1989uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1990 uma_init zinit, uma_fini zfini, uma_zone_t master)
1991{
1992 struct uma_zctor_args args;
1993 uma_keg_t keg;
1994 uma_zone_t res;
1995 bool locked;
1996
1997 keg = zone_first_keg(master);
1998 memset(&args, 0, sizeof(args));
1999 args.name = name;
2000 args.size = keg->uk_size;
2001 args.ctor = ctor;
2002 args.dtor = dtor;
2003 args.uminit = zinit;
2004 args.fini = zfini;
2005 args.align = keg->uk_align;
2006 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
2007 args.keg = keg;
2008
2009 if (booted < UMA_STARTUP2) {
2010 locked = false;
2011 } else {
2012 sx_slock(&uma_drain_lock);
2013 locked = true;
2014 }
2015 /* XXX Attaches only one keg of potentially many. */
2016 res = zone_alloc_item(zones, &args, M_WAITOK);
2017 if (locked)
2018 sx_sunlock(&uma_drain_lock);
2019 return (res);
2020}
2021
2022/* See uma.h */
2023uma_zone_t
2024uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
2025 uma_init zinit, uma_fini zfini, uma_import zimport,
2026 uma_release zrelease, void *arg, int flags)
2027{
2028 struct uma_zctor_args args;
2029
2030 memset(&args, 0, sizeof(args));
2031 args.name = name;
2032 args.size = size;
2033 args.ctor = ctor;
2034 args.dtor = dtor;
2035 args.uminit = zinit;
2036 args.fini = zfini;
2037 args.import = zimport;
2038 args.release = zrelease;
2039 args.arg = arg;
2040 args.align = 0;
2041 args.flags = flags;
2042
2043 return (zone_alloc_item(zones, &args, M_WAITOK));
2044}
2045
2046static void
2047zone_lock_pair(uma_zone_t a, uma_zone_t b)
2048{
2049 if (a < b) {
2050 ZONE_LOCK(a);
2051 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
2052 } else {
2053 ZONE_LOCK(b);
2054 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
2055 }
2056}
2057
2058static void
2059zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2060{
2061
2062 ZONE_UNLOCK(a);
2063 ZONE_UNLOCK(b);
2064}
2065
2066int
2067uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2068{
2069 uma_klink_t klink;
2070 uma_klink_t kl;
2071 int error;
2072
2073 error = 0;
2074 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2075
2076 zone_lock_pair(zone, master);
2077 /*
2078 * zone must use vtoslab() to resolve objects and must already be
2079 * a secondary.
2080 */
2081 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2082 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2083 error = EINVAL;
2084 goto out;
2085 }
2086 /*
2087 * The new master must also use vtoslab().
2088 */
2089 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2090 error = EINVAL;
2091 goto out;
2092 }
2093 /*
2094 * Both must either be refcnt, or not be refcnt.
2095 */
2096 if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
2097 (master->uz_flags & UMA_ZONE_REFCNT)) {
2098 error = EINVAL;
2099 goto out;
2100 }
2101 /*
2102 * The underlying object must be the same size. rsize
2103 * may be different.
2104 */
2105 if (master->uz_size != zone->uz_size) {
2106 error = E2BIG;
2107 goto out;
2108 }
2109 /*
2110 * Put it at the end of the list.
2111 */
2112 klink->kl_keg = zone_first_keg(master);
2113 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2114 if (LIST_NEXT(kl, kl_link) == NULL) {
2115 LIST_INSERT_AFTER(kl, klink, kl_link);
2116 break;
2117 }
2118 }
2119 klink = NULL;
2120 zone->uz_flags |= UMA_ZFLAG_MULTI;
2121 zone->uz_slab = zone_fetch_slab_multi;
2122
2123out:
2124 zone_unlock_pair(zone, master);
2125 if (klink != NULL)
2126 free(klink, M_TEMP);
2127
2128 return (error);
2129}
2130
2131
2132/* See uma.h */
2133void
2134uma_zdestroy(uma_zone_t zone)
2135{
2136
2137 sx_slock(&uma_drain_lock);
2138 zone_free_item(zones, zone, NULL, SKIP_NONE);
2139 sx_sunlock(&uma_drain_lock);
2140}
2141
2142/* See uma.h */
2143void *
2144uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2145{
2146 void *item;
2147 uma_cache_t cache;
2148 uma_bucket_t bucket;
2149 int lockfail;
2150 int cpu;
2151
2152 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2153 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2154
2155 /* This is the fast path allocation */
2156#ifdef UMA_DEBUG_ALLOC_1
2157 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2158#endif
2159 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2160 zone->uz_name, flags);
2161
2162 if (flags & M_WAITOK) {
2163 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2164 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2165 }
2166 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2167 ("uma_zalloc_arg: called with spinlock or critical section held"));
2168
2169#ifdef DEBUG_MEMGUARD
2170 if (memguard_cmp_zone(zone)) {
2171 item = memguard_alloc(zone->uz_size, flags);
2172 if (item != NULL) {
2173 /*
2174 * Avoid conflict with the use-after-free
2175 * protecting infrastructure from INVARIANTS.
2176 */
2177 if (zone->uz_init != NULL &&
2178 zone->uz_init != mtrash_init &&
2179 zone->uz_init(item, zone->uz_size, flags) != 0)
2180 return (NULL);
2181 if (zone->uz_ctor != NULL &&
2182 zone->uz_ctor != mtrash_ctor &&
2183 zone->uz_ctor(item, zone->uz_size, udata,
2184 flags) != 0) {
2185 zone->uz_fini(item, zone->uz_size);
2186 return (NULL);
2187 }
2188 return (item);
2189 }
2190 /* This is unfortunate but should not be fatal. */
2191 }
2192#endif
2193 /*
2194 * If possible, allocate from the per-CPU cache. There are two
2195 * requirements for safe access to the per-CPU cache: (1) the thread
2196 * accessing the cache must not be preempted or yield during access,
2197 * and (2) the thread must not migrate CPUs without switching which
2198 * cache it accesses. We rely on a critical section to prevent
2199 * preemption and migration. We release the critical section in
2200 * order to acquire the zone mutex if we are unable to allocate from
2201 * the current cache; when we re-acquire the critical section, we
2202 * must detect and handle migration if it has occurred.
2203 */
2204 critical_enter();
2205 cpu = curcpu;
2206 cache = &zone->uz_cpu[cpu];
2207
2208zalloc_start:
2209 bucket = cache->uc_allocbucket;
2210 if (bucket != NULL && bucket->ub_cnt > 0) {
2211 bucket->ub_cnt--;
2212 item = bucket->ub_bucket[bucket->ub_cnt];
2213#ifdef INVARIANTS
2214 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2215#endif
2216 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2217 cache->uc_allocs++;
2218 critical_exit();
2219 if (zone->uz_ctor != NULL &&
2220 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2221 atomic_add_long(&zone->uz_fails, 1);
2222 zone_free_item(zone, item, udata, SKIP_DTOR);
2223 return (NULL);
2224 }
2225#ifdef INVARIANTS
2226 uma_dbg_alloc(zone, NULL, item);
2227#endif
2228 if (flags & M_ZERO)
2229 uma_zero_item(item, zone);
2230 return (item);
2231 }
2232
2233 /*
2234 * We have run out of items in our alloc bucket.
2235 * See if we can switch with our free bucket.
2236 */
2237 bucket = cache->uc_freebucket;
2238 if (bucket != NULL && bucket->ub_cnt > 0) {
2239#ifdef UMA_DEBUG_ALLOC
2240 printf("uma_zalloc: Swapping empty with alloc.\n");
2241#endif
2242 cache->uc_freebucket = cache->uc_allocbucket;
2243 cache->uc_allocbucket = bucket;
2244 goto zalloc_start;
2245 }
2246
2247 /*
2248 * Discard any empty allocation bucket while we hold no locks.
2249 */
2250 bucket = cache->uc_allocbucket;
2251 cache->uc_allocbucket = NULL;
2252 critical_exit();
2253 if (bucket != NULL)
2254 bucket_free(zone, bucket, udata);
2255
2256 /* Short-circuit for zones without buckets and low memory. */
2257 if (zone->uz_count == 0 || bucketdisable)
2258 goto zalloc_item;
2259
2260 /*
2261 * Attempt to retrieve the item from the per-CPU cache has failed, so
2262 * we must go back to the zone. This requires the zone lock, so we
2263 * must drop the critical section, then re-acquire it when we go back
2264 * to the cache. Since the critical section is released, we may be
2265 * preempted or migrate. As such, make sure not to maintain any
2266 * thread-local state specific to the cache from prior to releasing
2267 * the critical section.
2268 */
2269 lockfail = 0;
2270 if (ZONE_TRYLOCK(zone) == 0) {
2271 /* Record contention to size the buckets. */
2272 ZONE_LOCK(zone);
2273 lockfail = 1;
2274 }
2275 critical_enter();
2276 cpu = curcpu;
2277 cache = &zone->uz_cpu[cpu];
2278
2279 /*
2280 * Since we have locked the zone we may as well send back our stats.
2281 */
2282 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2283 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2284 cache->uc_allocs = 0;
2285 cache->uc_frees = 0;
2286
2287 /* See if we lost the race to fill the cache. */
2288 if (cache->uc_allocbucket != NULL) {
2289 ZONE_UNLOCK(zone);
2290 goto zalloc_start;
2291 }
2292
2293 /*
2294 * Check the zone's cache of buckets.
2295 */
2296 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2297 KASSERT(bucket->ub_cnt != 0,
2298 ("uma_zalloc_arg: Returning an empty bucket."));
2299
2300 LIST_REMOVE(bucket, ub_link);
2301 cache->uc_allocbucket = bucket;
2302 ZONE_UNLOCK(zone);
2303 goto zalloc_start;
2304 }
2305 /* We are no longer associated with this CPU. */
2306 critical_exit();
2307
2308 /*
2309 * We bump the uz count when the cache size is insufficient to
2310 * handle the working set.
2311 */
2312 if (lockfail && zone->uz_count < BUCKET_MAX)
2313 zone->uz_count++;
2314 ZONE_UNLOCK(zone);
2315
2316 /*
2317 * Now lets just fill a bucket and put it on the free list. If that
2318 * works we'll restart the allocation from the begining and it
2319 * will use the just filled bucket.
2320 */
2321 bucket = zone_alloc_bucket(zone, udata, flags);
2322 if (bucket != NULL) {
2323 ZONE_LOCK(zone);
2324 critical_enter();
2325 cpu = curcpu;
2326 cache = &zone->uz_cpu[cpu];
2327 /*
2328 * See if we lost the race or were migrated. Cache the
2329 * initialized bucket to make this less likely or claim
2330 * the memory directly.
2331 */
2332 if (cache->uc_allocbucket == NULL)
2333 cache->uc_allocbucket = bucket;
2334 else
2335 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2336 ZONE_UNLOCK(zone);
2337 goto zalloc_start;
2338 }
2339
2340 /*
2341 * We may not be able to get a bucket so return an actual item.
2342 */
2343#ifdef UMA_DEBUG
2344 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2345#endif
2346
2347zalloc_item:
2348 item = zone_alloc_item(zone, udata, flags);
2349
2350 return (item);
2351}
2352
2353static uma_slab_t
2354keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2355{
2356 uma_slab_t slab;
2357 int reserve;
2358
2359 mtx_assert(&keg->uk_lock, MA_OWNED);
2360 slab = NULL;
2361 reserve = 0;
2362 if ((flags & M_USE_RESERVE) == 0)
2363 reserve = keg->uk_reserve;
2364
2365 for (;;) {
2366 /*
2367 * Find a slab with some space. Prefer slabs that are partially
2368 * used over those that are totally full. This helps to reduce
2369 * fragmentation.
2370 */
2371 if (keg->uk_free > reserve) {
2372 if (!LIST_EMPTY(&keg->uk_part_slab)) {
2373 slab = LIST_FIRST(&keg->uk_part_slab);
2374 } else {
2375 slab = LIST_FIRST(&keg->uk_free_slab);
2376 LIST_REMOVE(slab, us_link);
2377 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2378 us_link);
2379 }
2380 MPASS(slab->us_keg == keg);
2381 return (slab);
2382 }
2383
2384 /*
2385 * M_NOVM means don't ask at all!
2386 */
2387 if (flags & M_NOVM)
2388 break;
2389
2390 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2391 keg->uk_flags |= UMA_ZFLAG_FULL;
2392 /*
2393 * If this is not a multi-zone, set the FULL bit.
2394 * Otherwise slab_multi() takes care of it.
2395 */
2396 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2397 zone->uz_flags |= UMA_ZFLAG_FULL;
2398 zone_log_warning(zone);
2399 zone_maxaction(zone);
2400 }
2401 if (flags & M_NOWAIT)
2402 break;
2403 zone->uz_sleeps++;
2404 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2405 continue;
2406 }
2407 slab = keg_alloc_slab(keg, zone, flags);
2408 /*
2409 * If we got a slab here it's safe to mark it partially used
2410 * and return. We assume that the caller is going to remove
2411 * at least one item.
2412 */
2413 if (slab) {
2414 MPASS(slab->us_keg == keg);
2415 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2416 return (slab);
2417 }
2418 /*
2419 * We might not have been able to get a slab but another cpu
2420 * could have while we were unlocked. Check again before we
2421 * fail.
2422 */
2423 flags |= M_NOVM;
2424 }
2425 return (slab);
2426}
2427
2428static uma_slab_t
2429zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2430{
2431 uma_slab_t slab;
2432
2433 if (keg == NULL) {
2434 keg = zone_first_keg(zone);
2435 KEG_LOCK(keg);
2436 }
2437
2438 for (;;) {
2439 slab = keg_fetch_slab(keg, zone, flags);
2440 if (slab)
2441 return (slab);
2442 if (flags & (M_NOWAIT | M_NOVM))
2443 break;
2444 }
2445 KEG_UNLOCK(keg);
2446 return (NULL);
2447}
2448
2449/*
2450 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2451 * with the keg locked. On NULL no lock is held.
2452 *
2453 * The last pointer is used to seed the search. It is not required.
2454 */
2455static uma_slab_t
2456zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2457{
2458 uma_klink_t klink;
2459 uma_slab_t slab;
2460 uma_keg_t keg;
2461 int flags;
2462 int empty;
2463 int full;
2464
2465 /*
2466 * Don't wait on the first pass. This will skip limit tests
2467 * as well. We don't want to block if we can find a provider
2468 * without blocking.
2469 */
2470 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2471 /*
2472 * Use the last slab allocated as a hint for where to start
2473 * the search.
2474 */
2475 if (last != NULL) {
2476 slab = keg_fetch_slab(last, zone, flags);
2477 if (slab)
2478 return (slab);
2479 KEG_UNLOCK(last);
2480 }
2481 /*
2482 * Loop until we have a slab incase of transient failures
2483 * while M_WAITOK is specified. I'm not sure this is 100%
2484 * required but we've done it for so long now.
2485 */
2486 for (;;) {
2487 empty = 0;
2488 full = 0;
2489 /*
2490 * Search the available kegs for slabs. Be careful to hold the
2491 * correct lock while calling into the keg layer.
2492 */
2493 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2494 keg = klink->kl_keg;
2495 KEG_LOCK(keg);
2496 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2497 slab = keg_fetch_slab(keg, zone, flags);
2498 if (slab)
2499 return (slab);
2500 }
2501 if (keg->uk_flags & UMA_ZFLAG_FULL)
2502 full++;
2503 else
2504 empty++;
2505 KEG_UNLOCK(keg);
2506 }
2507 if (rflags & (M_NOWAIT | M_NOVM))
2508 break;
2509 flags = rflags;
2510 /*
2511 * All kegs are full. XXX We can't atomically check all kegs
2512 * and sleep so just sleep for a short period and retry.
2513 */
2514 if (full && !empty) {
2515 ZONE_LOCK(zone);
2516 zone->uz_flags |= UMA_ZFLAG_FULL;
2517 zone->uz_sleeps++;
2518 zone_log_warning(zone);
2519 zone_maxaction(zone);
2520 msleep(zone, zone->uz_lockptr, PVM,
2521 "zonelimit", hz/100);
2522 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2523 ZONE_UNLOCK(zone);
2524 continue;
2525 }
2526 }
2527 return (NULL);
2528}
2529
2530static void *
2531slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2532{
2533 void *item;
2534 uint8_t freei;
2535
2536 MPASS(keg == slab->us_keg);
2537 mtx_assert(&keg->uk_lock, MA_OWNED);
2538
2539 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2540 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2541 item = slab->us_data + (keg->uk_rsize * freei);
2542 slab->us_freecount--;
2543 keg->uk_free--;
2544
2545 /* Move this slab to the full list */
2546 if (slab->us_freecount == 0) {
2547 LIST_REMOVE(slab, us_link);
2548 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2549 }
2550
2551 return (item);
2552}
2553
2554static int
2555zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2556{
2557 uma_slab_t slab;
2558 uma_keg_t keg;
2559 int i;
2560
2561 slab = NULL;
2562 keg = NULL;
2563 /* Try to keep the buckets totally full */
2564 for (i = 0; i < max; ) {
2565 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2566 break;
2567 keg = slab->us_keg;
2568 while (slab->us_freecount && i < max) {
2569 bucket[i++] = slab_alloc_item(keg, slab);
2570 if (keg->uk_free <= keg->uk_reserve)
2571 break;
2572 }
2573 /* Don't grab more than one slab at a time. */
2574 flags &= ~M_WAITOK;
2575 flags |= M_NOWAIT;
2576 }
2577 if (slab != NULL)
2578 KEG_UNLOCK(keg);
2579
2580 return i;
2581}
2582
2583static uma_bucket_t
2584zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2585{
2586 uma_bucket_t bucket;
2587 int max;
2588
2589 /* Don't wait for buckets, preserve caller's NOVM setting. */
2590 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2591 if (bucket == NULL)
2592 return (NULL);
2593
2594 max = MIN(bucket->ub_entries, zone->uz_count);
2595 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2596 max, flags);
2597
2598 /*
2599 * Initialize the memory if necessary.
2600 */
2601 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2602 int i;
2603
2604 for (i = 0; i < bucket->ub_cnt; i++)
2605 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2606 flags) != 0)
2607 break;
2608 /*
2609 * If we couldn't initialize the whole bucket, put the
2610 * rest back onto the freelist.
2611 */
2612 if (i != bucket->ub_cnt) {
2613 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2614 bucket->ub_cnt - i);
2615#ifdef INVARIANTS
2616 bzero(&bucket->ub_bucket[i],
2617 sizeof(void *) * (bucket->ub_cnt - i));
2618#endif
2619 bucket->ub_cnt = i;
2620 }
2621 }
2622
2623 if (bucket->ub_cnt == 0) {
2624 bucket_free(zone, bucket, udata);
2625 atomic_add_long(&zone->uz_fails, 1);
2626 return (NULL);
2627 }
2628
2629 return (bucket);
2630}
2631
2632/*
2633 * Allocates a single item from a zone.
2634 *
2635 * Arguments
2636 * zone The zone to alloc for.
2637 * udata The data to be passed to the constructor.
2638 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2639 *
2640 * Returns
2641 * NULL if there is no memory and M_NOWAIT is set
2642 * An item if successful
2643 */
2644
2645static void *
2646zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2647{
2648 void *item;
2649
2650 item = NULL;
2651
2652#ifdef UMA_DEBUG_ALLOC
2653 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2654#endif
2655 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2656 goto fail;
2657 atomic_add_long(&zone->uz_allocs, 1);
2658
2659 /*
2660 * We have to call both the zone's init (not the keg's init)
2661 * and the zone's ctor. This is because the item is going from
2662 * a keg slab directly to the user, and the user is expecting it
2663 * to be both zone-init'd as well as zone-ctor'd.
2664 */
2665 if (zone->uz_init != NULL) {
2666 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2667 zone_free_item(zone, item, udata, SKIP_FINI);
2668 goto fail;
2669 }
2670 }
2671 if (zone->uz_ctor != NULL) {
2672 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2673 zone_free_item(zone, item, udata, SKIP_DTOR);
2674 goto fail;
2675 }
2676 }
2677#ifdef INVARIANTS
2678 uma_dbg_alloc(zone, NULL, item);
2679#endif
2680 if (flags & M_ZERO)
2681 uma_zero_item(item, zone);
2682
2683 return (item);
2684
2685fail:
2686 atomic_add_long(&zone->uz_fails, 1);
2687 return (NULL);
2688}
2689
2690/* See uma.h */
2691void
2692uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2693{
2694 uma_cache_t cache;
2695 uma_bucket_t bucket;
2696 int lockfail;
2697 int cpu;
2698
2699 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2700 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2701
2702#ifdef UMA_DEBUG_ALLOC_1
2703 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2704#endif
2705 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2706 zone->uz_name);
2707
2708 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2709 ("uma_zfree_arg: called with spinlock or critical section held"));
2710
2711 /* uma_zfree(..., NULL) does nothing, to match free(9). */
2712 if (item == NULL)
2713 return;
2714#ifdef DEBUG_MEMGUARD
2715 if (is_memguard_addr(item)) {
2716 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2717 zone->uz_dtor(item, zone->uz_size, udata);
2718 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2719 zone->uz_fini(item, zone->uz_size);
2720 memguard_free(item);
2721 return;
2722 }
2723#endif
2724#ifdef INVARIANTS
2725 if (zone->uz_flags & UMA_ZONE_MALLOC)
2726 uma_dbg_free(zone, udata, item);
2727 else
2728 uma_dbg_free(zone, NULL, item);
2729#endif
2730 if (zone->uz_dtor != NULL)
2731 zone->uz_dtor(item, zone->uz_size, udata);
2732
2733 /*
2734 * The race here is acceptable. If we miss it we'll just have to wait
2735 * a little longer for the limits to be reset.
2736 */
2737 if (zone->uz_flags & UMA_ZFLAG_FULL)
2738 goto zfree_item;
2739
2740 /*
2741 * If possible, free to the per-CPU cache. There are two
2742 * requirements for safe access to the per-CPU cache: (1) the thread
2743 * accessing the cache must not be preempted or yield during access,
2744 * and (2) the thread must not migrate CPUs without switching which
2745 * cache it accesses. We rely on a critical section to prevent
2746 * preemption and migration. We release the critical section in
2747 * order to acquire the zone mutex if we are unable to free to the
2748 * current cache; when we re-acquire the critical section, we must
2749 * detect and handle migration if it has occurred.
2750 */
2751zfree_restart:
2752 critical_enter();
2753 cpu = curcpu;
2754 cache = &zone->uz_cpu[cpu];
2755
2756zfree_start:
2757 /*
2758 * Try to free into the allocbucket first to give LIFO ordering
2759 * for cache-hot datastructures. Spill over into the freebucket
2760 * if necessary. Alloc will swap them if one runs dry.
2761 */
2762 bucket = cache->uc_allocbucket;
2763 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2764 bucket = cache->uc_freebucket;
2765 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2766 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2767 ("uma_zfree: Freeing to non free bucket index."));
2768 bucket->ub_bucket[bucket->ub_cnt] = item;
2769 bucket->ub_cnt++;
2770 cache->uc_frees++;
2771 critical_exit();
2772 return;
2773 }
2774
2775 /*
2776 * We must go back the zone, which requires acquiring the zone lock,
2777 * which in turn means we must release and re-acquire the critical
2778 * section. Since the critical section is released, we may be
2779 * preempted or migrate. As such, make sure not to maintain any
2780 * thread-local state specific to the cache from prior to releasing
2781 * the critical section.
2782 */
2783 critical_exit();
2784 if (zone->uz_count == 0 || bucketdisable)
2785 goto zfree_item;
2786
2787 lockfail = 0;
2788 if (ZONE_TRYLOCK(zone) == 0) {
2789 /* Record contention to size the buckets. */
2790 ZONE_LOCK(zone);
2791 lockfail = 1;
2792 }
2793 critical_enter();
2794 cpu = curcpu;
2795 cache = &zone->uz_cpu[cpu];
2796
2797 /*
2798 * Since we have locked the zone we may as well send back our stats.
2799 */
2800 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2801 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2802 cache->uc_allocs = 0;
2803 cache->uc_frees = 0;
2804
2805 bucket = cache->uc_freebucket;
2806 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2807 ZONE_UNLOCK(zone);
2808 goto zfree_start;
2809 }
2810 cache->uc_freebucket = NULL;
2811
2812 /* Can we throw this on the zone full list? */
2813 if (bucket != NULL) {
2814#ifdef UMA_DEBUG_ALLOC
2815 printf("uma_zfree: Putting old bucket on the free list.\n");
2816#endif
2817 /* ub_cnt is pointing to the last free item */
2818 KASSERT(bucket->ub_cnt != 0,
2819 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2820 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2821 }
2822
2823 /* We are no longer associated with this CPU. */
2824 critical_exit();
2825
2826 /*
2827 * We bump the uz count when the cache size is insufficient to
2828 * handle the working set.
2829 */
2830 if (lockfail && zone->uz_count < BUCKET_MAX)
2831 zone->uz_count++;
2832 ZONE_UNLOCK(zone);
2833
2834#ifdef UMA_DEBUG_ALLOC
2835 printf("uma_zfree: Allocating new free bucket.\n");
2836#endif
2837 bucket = bucket_alloc(zone, udata, M_NOWAIT);
2838 if (bucket) {
2839 critical_enter();
2840 cpu = curcpu;
2841 cache = &zone->uz_cpu[cpu];
2842 if (cache->uc_freebucket == NULL) {
2843 cache->uc_freebucket = bucket;
2844 goto zfree_start;
2845 }
2846 /*
2847 * We lost the race, start over. We have to drop our
2848 * critical section to free the bucket.
2849 */
2850 critical_exit();
2851 bucket_free(zone, bucket, udata);
2852 goto zfree_restart;
2853 }
2854
2855 /*
2856 * If nothing else caught this, we'll just do an internal free.
2857 */
2858zfree_item:
2859 zone_free_item(zone, item, udata, SKIP_DTOR);
2860
2861 return;
2862}
2863
2864static void
2865slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2866{
2867 uint8_t freei;
2868
2869 mtx_assert(&keg->uk_lock, MA_OWNED);
2870 MPASS(keg == slab->us_keg);
2871
2872 /* Do we need to remove from any lists? */
2873 if (slab->us_freecount+1 == keg->uk_ipers) {
2874 LIST_REMOVE(slab, us_link);
2875 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2876 } else if (slab->us_freecount == 0) {
2877 LIST_REMOVE(slab, us_link);
2878 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2879 }
2880
2881 /* Slab management. */
2882 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2883 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2884 slab->us_freecount++;
2885
2886 /* Keg statistics. */
2887 keg->uk_free++;
2888}
2889
2890static void
2891zone_release(uma_zone_t zone, void **bucket, int cnt)
2892{
2893 void *item;
2894 uma_slab_t slab;
2895 uma_keg_t keg;
2896 uint8_t *mem;
2897 int clearfull;
2898 int i;
2899
2900 clearfull = 0;
2901 keg = zone_first_keg(zone);
2902 KEG_LOCK(keg);
2903 for (i = 0; i < cnt; i++) {
2904 item = bucket[i];
2905 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2906 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2907 if (zone->uz_flags & UMA_ZONE_HASH) {
2908 slab = hash_sfind(&keg->uk_hash, mem);
2909 } else {
2910 mem += keg->uk_pgoff;
2911 slab = (uma_slab_t)mem;
2912 }
2913 } else {
2914 slab = vtoslab((vm_offset_t)item);
2915 if (slab->us_keg != keg) {
2916 KEG_UNLOCK(keg);
2917 keg = slab->us_keg;
2918 KEG_LOCK(keg);
2919 }
2920 }
2921 slab_free_item(keg, slab, item);
2922 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2923 if (keg->uk_pages < keg->uk_maxpages) {
2924 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2925 clearfull = 1;
2926 }
2927
2928 /*
2929 * We can handle one more allocation. Since we're
2930 * clearing ZFLAG_FULL, wake up all procs blocked
2931 * on pages. This should be uncommon, so keeping this
2932 * simple for now (rather than adding count of blocked
2933 * threads etc).
2934 */
2935 wakeup(keg);
2936 }
2937 }
2938 KEG_UNLOCK(keg);
2939 if (clearfull) {
2940 ZONE_LOCK(zone);
2941 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2942 wakeup(zone);
2943 ZONE_UNLOCK(zone);
2944 }
2945
2946}
2947
2948/*
2949 * Frees a single item to any zone.
2950 *
2951 * Arguments:
2952 * zone The zone to free to
2953 * item The item we're freeing
2954 * udata User supplied data for the dtor
2955 * skip Skip dtors and finis
2956 */
2957static void
2958zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2959{
2960
2961#ifdef INVARIANTS
2962 if (skip == SKIP_NONE) {
2963 if (zone->uz_flags & UMA_ZONE_MALLOC)
2964 uma_dbg_free(zone, udata, item);
2965 else
2966 uma_dbg_free(zone, NULL, item);
2967 }
2968#endif
2969 if (skip < SKIP_DTOR && zone->uz_dtor)
2970 zone->uz_dtor(item, zone->uz_size, udata);
2971
2972 if (skip < SKIP_FINI && zone->uz_fini)
2973 zone->uz_fini(item, zone->uz_size);
2974
2975 atomic_add_long(&zone->uz_frees, 1);
2976 zone->uz_release(zone->uz_arg, &item, 1);
2977}
2978
2979/* See uma.h */
2980int
2981uma_zone_set_max(uma_zone_t zone, int nitems)
2982{
2983 uma_keg_t keg;
2984
2985 keg = zone_first_keg(zone);
2986 if (keg == NULL)
2987 return (0);
2988 KEG_LOCK(keg);
2989 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2990 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2991 keg->uk_maxpages += keg->uk_ppera;
2992 nitems = keg->uk_maxpages * keg->uk_ipers;
2993 KEG_UNLOCK(keg);
2994
2995 return (nitems);
2996}
2997
2998/* See uma.h */
2999int
3000uma_zone_get_max(uma_zone_t zone)
3001{
3002 int nitems;
3003 uma_keg_t keg;
3004
3005 keg = zone_first_keg(zone);
3006 if (keg == NULL)
3007 return (0);
3008 KEG_LOCK(keg);
3009 nitems = keg->uk_maxpages * keg->uk_ipers;
3010 KEG_UNLOCK(keg);
3011
3012 return (nitems);
3013}
3014
3015/* See uma.h */
3016void
3017uma_zone_set_warning(uma_zone_t zone, const char *warning)
3018{
3019
3020 ZONE_LOCK(zone);
3021 zone->uz_warning = warning;
3022 ZONE_UNLOCK(zone);
3023}
3024
3025/* See uma.h */
3026void
3027uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
3028{
3029
3030 ZONE_LOCK(zone);
3031 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
3032 ZONE_UNLOCK(zone);
3033}
3034
3035/* See uma.h */
3036int
3037uma_zone_get_cur(uma_zone_t zone)
3038{
3039 int64_t nitems;
3040 u_int i;
3041
3042 ZONE_LOCK(zone);
3043 nitems = zone->uz_allocs - zone->uz_frees;
3044 CPU_FOREACH(i) {
3045 /*
3046 * See the comment in sysctl_vm_zone_stats() regarding the
3047 * safety of accessing the per-cpu caches. With the zone lock
3048 * held, it is safe, but can potentially result in stale data.
3049 */
3050 nitems += zone->uz_cpu[i].uc_allocs -
3051 zone->uz_cpu[i].uc_frees;
3052 }
3053 ZONE_UNLOCK(zone);
3054
3055 return (nitems < 0 ? 0 : nitems);
3056}
3057
3058/* See uma.h */
3059void
3060uma_zone_set_init(uma_zone_t zone, uma_init uminit)
3061{
3062 uma_keg_t keg;
3063
3064 keg = zone_first_keg(zone);
3065 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3066 KEG_LOCK(keg);
3067 KASSERT(keg->uk_pages == 0,
3068 ("uma_zone_set_init on non-empty keg"));
3069 keg->uk_init = uminit;
3070 KEG_UNLOCK(keg);
3071}
3072
3073/* See uma.h */
3074void
3075uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3076{
3077 uma_keg_t keg;
3078
3079 keg = zone_first_keg(zone);
3080 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
3081 KEG_LOCK(keg);
3082 KASSERT(keg->uk_pages == 0,
3083 ("uma_zone_set_fini on non-empty keg"));
3084 keg->uk_fini = fini;
3085 KEG_UNLOCK(keg);
3086}
3087
3088/* See uma.h */
3089void
3090uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3091{
3092
3093 ZONE_LOCK(zone);
3094 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3095 ("uma_zone_set_zinit on non-empty keg"));
3096 zone->uz_init = zinit;
3097 ZONE_UNLOCK(zone);
3098}
3099
3100/* See uma.h */
3101void
3102uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3103{
3104
3105 ZONE_LOCK(zone);
3106 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3107 ("uma_zone_set_zfini on non-empty keg"));
3108 zone->uz_fini = zfini;
3109 ZONE_UNLOCK(zone);
3110}
3111
3112/* See uma.h */
3113/* XXX uk_freef is not actually used with the zone locked */
3114void
3115uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3116{
3117 uma_keg_t keg;
3118
3119 keg = zone_first_keg(zone);
3120 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
3121 KEG_LOCK(keg);
3122 keg->uk_freef = freef;
3123 KEG_UNLOCK(keg);
3124}
3125
3126/* See uma.h */
3127/* XXX uk_allocf is not actually used with the zone locked */
3128void
3129uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3130{
3131 uma_keg_t keg;
3132
3133 keg = zone_first_keg(zone);
3134 KEG_LOCK(keg);
3135 keg->uk_allocf = allocf;
3136 KEG_UNLOCK(keg);
3137}
3138
3139/* See uma.h */
3140void
3141uma_zone_reserve(uma_zone_t zone, int items)
3142{
3143 uma_keg_t keg;
3144
3145 keg = zone_first_keg(zone);
3146 if (keg == NULL)
3147 return;
3148 KEG_LOCK(keg);
3149 keg->uk_reserve = items;
3150 KEG_UNLOCK(keg);
3151
3152 return;
3153}
3154
3155/* See uma.h */
3156int
3157uma_zone_reserve_kva(uma_zone_t zone, int count)
3158{
3159 uma_keg_t keg;
3160 vm_offset_t kva;
3161 u_int pages;
3162
3163 keg = zone_first_keg(zone);
3164 if (keg == NULL)
3165 return (0);
3166 pages = count / keg->uk_ipers;
3167
3168 if (pages * keg->uk_ipers < count)
3169 pages++;
3170
3171#ifdef UMA_MD_SMALL_ALLOC
3172 if (keg->uk_ppera > 1) {
3173#else
3174 if (1) {
3175#endif
3176 kva = kva_alloc((vm_size_t)pages * UMA_SLAB_SIZE);
3177 if (kva == 0)
3178 return (0);
3179 } else
3180 kva = 0;
3181 KEG_LOCK(keg);
3182 keg->uk_kva = kva;
3183 keg->uk_offset = 0;
3184 keg->uk_maxpages = pages;
3185#ifdef UMA_MD_SMALL_ALLOC
3186 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3187#else
3188 keg->uk_allocf = noobj_alloc;
3189#endif
3190 keg->uk_flags |= UMA_ZONE_NOFREE;
3191 KEG_UNLOCK(keg);
3192
3193 return (1);
3194}
3195
3196/* See uma.h */
3197void
3198uma_prealloc(uma_zone_t zone, int items)
3199{
3200 int slabs;
3201 uma_slab_t slab;
3202 uma_keg_t keg;
3203
3204 keg = zone_first_keg(zone);
3205 if (keg == NULL)
3206 return;
3207 KEG_LOCK(keg);
3208 slabs = items / keg->uk_ipers;
3209 if (slabs * keg->uk_ipers < items)
3210 slabs++;
3211 while (slabs > 0) {
3212 slab = keg_alloc_slab(keg, zone, M_WAITOK);
3213 if (slab == NULL)
3214 break;
3215 MPASS(slab->us_keg == keg);
3216 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3217 slabs--;
3218 }
3219 KEG_UNLOCK(keg);
3220}
3221
3222/* See uma.h */
3223uint32_t *
3224uma_find_refcnt(uma_zone_t zone, void *item)
3225{
3226 uma_slabrefcnt_t slabref;
3227 uma_slab_t slab;
3228 uma_keg_t keg;
3229 uint32_t *refcnt;
3230 int idx;
3231
3232 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3233 slabref = (uma_slabrefcnt_t)slab;
3234 keg = slab->us_keg;
3235 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3236 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3237 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3238 refcnt = &slabref->us_refcnt[idx];
3239 return refcnt;
3240}
3241
3242/* See uma.h */
3243static void
3244uma_reclaim_locked(bool kmem_danger)
3245{
3246
3247#ifdef UMA_DEBUG
3248 printf("UMA: vm asked us to release pages!\n");
3249#endif
3250 sx_assert(&uma_drain_lock, SA_XLOCKED);
3251 bucket_enable();
3252 zone_foreach(zone_drain);
3253 if (vm_page_count_min() || kmem_danger) {
3254 cache_drain_safe(NULL);
3255 zone_foreach(zone_drain);
3256 }
3257 /*
3258 * Some slabs may have been freed but this zone will be visited early
3259 * we visit again so that we can free pages that are empty once other
3260 * zones are drained. We have to do the same for buckets.
3261 */
3262 zone_drain(slabzone);
3263 zone_drain(slabrefzone);
3264 bucket_zone_drain();
3265}
3266
3267void
3268uma_reclaim(void)
3269{
3270
3271 sx_xlock(&uma_drain_lock);
3272 uma_reclaim_locked(false);
3273 sx_xunlock(&uma_drain_lock);
3274}
3275
3276static int uma_reclaim_needed;
3277
3278void
3279uma_reclaim_wakeup(void)
3280{
3281
3282 uma_reclaim_needed = 1;
3283 wakeup(&uma_reclaim_needed);
3284}
3285
3286void
3287uma_reclaim_worker(void *arg __unused)
3288{
3289
3290 sx_xlock(&uma_drain_lock);
3291 for (;;) {
3292 sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM,
3293 "umarcl", 0);
3294 if (uma_reclaim_needed) {
3295 uma_reclaim_needed = 0;
3296 uma_reclaim_locked(true);
3297 }
3298 }
3299}
3300
3301/* See uma.h */
3302int
3303uma_zone_exhausted(uma_zone_t zone)
3304{
3305 int full;
3306
3307 ZONE_LOCK(zone);
3308 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3309 ZONE_UNLOCK(zone);
3310 return (full);
3311}
3312
3313int
3314uma_zone_exhausted_nolock(uma_zone_t zone)
3315{
3316 return (zone->uz_flags & UMA_ZFLAG_FULL);
3317}
3318
3319void *
3320uma_large_malloc(vm_size_t size, int wait)
3321{
3322 void *mem;
3323 uma_slab_t slab;
3324 uint8_t flags;
3325
3326 slab = zone_alloc_item(slabzone, NULL, wait);
3327 if (slab == NULL)
3328 return (NULL);
3329 mem = page_alloc(NULL, size, &flags, wait);
3330 if (mem) {
3331 vsetslab((vm_offset_t)mem, slab);
3332 slab->us_data = mem;
3333 slab->us_flags = flags | UMA_SLAB_MALLOC;
3334 slab->us_size = size;
3335 } else {
3336 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3337 }
3338
3339 return (mem);
3340}
3341
3342void
3343uma_large_free(uma_slab_t slab)
3344{
3345
3346 page_free(slab->us_data, slab->us_size, slab->us_flags);
3347 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3348}
3349
3350static void
3351uma_zero_item(void *item, uma_zone_t zone)
3352{
3353
3354 if (zone->uz_flags & UMA_ZONE_PCPU) {
3355 for (int i = 0; i < mp_ncpus; i++)
3356 bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3357 } else
3358 bzero(item, zone->uz_size);
3359}
3360
3361void
3362uma_print_stats(void)
3363{
3364 zone_foreach(uma_print_zone);
3365}
3366
3367static void
3368slab_print(uma_slab_t slab)
3369{
3370 printf("slab: keg %p, data %p, freecount %d\n",
3371 slab->us_keg, slab->us_data, slab->us_freecount);
3372}
3373
3374static void
3375cache_print(uma_cache_t cache)
3376{
3377 printf("alloc: %p(%d), free: %p(%d)\n",
3378 cache->uc_allocbucket,
3379 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3380 cache->uc_freebucket,
3381 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3382}
3383
3384static void
3385uma_print_keg(uma_keg_t keg)
3386{
3387 uma_slab_t slab;
3388
3389 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3390 "out %d free %d limit %d\n",
3391 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3392 keg->uk_ipers, keg->uk_ppera,
3393 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3394 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3395 printf("Part slabs:\n");
3396 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3397 slab_print(slab);
3398 printf("Free slabs:\n");
3399 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3400 slab_print(slab);
3401 printf("Full slabs:\n");
3402 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3403 slab_print(slab);
3404}
3405
3406void
3407uma_print_zone(uma_zone_t zone)
3408{
3409 uma_cache_t cache;
3410 uma_klink_t kl;
3411 int i;
3412
3413 printf("zone: %s(%p) size %d flags %#x\n",
3414 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3415 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3416 uma_print_keg(kl->kl_keg);
3417 CPU_FOREACH(i) {
3418 cache = &zone->uz_cpu[i];
3419 printf("CPU %d Cache:\n", i);
3420 cache_print(cache);
3421 }
3422}
3423
3424#ifdef DDB
3425/*
3426 * Generate statistics across both the zone and its per-cpu cache's. Return
3427 * desired statistics if the pointer is non-NULL for that statistic.
3428 *
3429 * Note: does not update the zone statistics, as it can't safely clear the
3430 * per-CPU cache statistic.
3431 *
3432 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3433 * safe from off-CPU; we should modify the caches to track this information
3434 * directly so that we don't have to.
3435 */
3436static void
3437uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3438 uint64_t *freesp, uint64_t *sleepsp)
3439{
3440 uma_cache_t cache;
3441 uint64_t allocs, frees, sleeps;
3442 int cachefree, cpu;
3443
3444 allocs = frees = sleeps = 0;
3445 cachefree = 0;
3446 CPU_FOREACH(cpu) {
3447 cache = &z->uz_cpu[cpu];
3448 if (cache->uc_allocbucket != NULL)
3449 cachefree += cache->uc_allocbucket->ub_cnt;
3450 if (cache->uc_freebucket != NULL)
3451 cachefree += cache->uc_freebucket->ub_cnt;
3452 allocs += cache->uc_allocs;
3453 frees += cache->uc_frees;
3454 }
3455 allocs += z->uz_allocs;
3456 frees += z->uz_frees;
3457 sleeps += z->uz_sleeps;
3458 if (cachefreep != NULL)
3459 *cachefreep = cachefree;
3460 if (allocsp != NULL)
3461 *allocsp = allocs;
3462 if (freesp != NULL)
3463 *freesp = frees;
3464 if (sleepsp != NULL)
3465 *sleepsp = sleeps;
3466}
3467#endif /* DDB */
3468
3469static int
3470sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3471{
3472 uma_keg_t kz;
3473 uma_zone_t z;
3474 int count;
3475
3476 count = 0;
3477 rw_rlock(&uma_rwlock);
3478 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3479 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3480 count++;
3481 }
3482 rw_runlock(&uma_rwlock);
3483 return (sysctl_handle_int(oidp, &count, 0, req));
3484}
3485
3486static int
3487sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3488{
3489 struct uma_stream_header ush;
3490 struct uma_type_header uth;
3491 struct uma_percpu_stat ups;
3492 uma_bucket_t bucket;
3493 struct sbuf sbuf;
3494 uma_cache_t cache;
3495 uma_klink_t kl;
3496 uma_keg_t kz;
3497 uma_zone_t z;
3498 uma_keg_t k;
3499 int count, error, i;
3500
3501 error = sysctl_wire_old_buffer(req, 0);
3502 if (error != 0)
3503 return (error);
3504 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3505 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3506
3507 count = 0;
3508 rw_rlock(&uma_rwlock);
3509 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3510 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3511 count++;
3512 }
3513
3514 /*
3515 * Insert stream header.
3516 */
3517 bzero(&ush, sizeof(ush));
3518 ush.ush_version = UMA_STREAM_VERSION;
3519 ush.ush_maxcpus = (mp_maxid + 1);
3520 ush.ush_count = count;
3521 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3522
3523 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3524 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3525 bzero(&uth, sizeof(uth));
3526 ZONE_LOCK(z);
3527 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3528 uth.uth_align = kz->uk_align;
3529 uth.uth_size = kz->uk_size;
3530 uth.uth_rsize = kz->uk_rsize;
3531 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3532 k = kl->kl_keg;
3533 uth.uth_maxpages += k->uk_maxpages;
3534 uth.uth_pages += k->uk_pages;
3535 uth.uth_keg_free += k->uk_free;
3536 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3537 * k->uk_ipers;
3538 }
3539
3540 /*
3541 * A zone is secondary is it is not the first entry
3542 * on the keg's zone list.
3543 */
3544 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3545 (LIST_FIRST(&kz->uk_zones) != z))
3546 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3547
3548 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3549 uth.uth_zone_free += bucket->ub_cnt;
3550 uth.uth_allocs = z->uz_allocs;
3551 uth.uth_frees = z->uz_frees;
3552 uth.uth_fails = z->uz_fails;
3553 uth.uth_sleeps = z->uz_sleeps;
3554 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3555 /*
3556 * While it is not normally safe to access the cache
3557 * bucket pointers while not on the CPU that owns the
3558 * cache, we only allow the pointers to be exchanged
3559 * without the zone lock held, not invalidated, so
3560 * accept the possible race associated with bucket
3561 * exchange during monitoring.
3562 */
3563 for (i = 0; i < (mp_maxid + 1); i++) {
3564 bzero(&ups, sizeof(ups));
3565 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3566 goto skip;
3567 if (CPU_ABSENT(i))
3568 goto skip;
3569 cache = &z->uz_cpu[i];
3570 if (cache->uc_allocbucket != NULL)
3571 ups.ups_cache_free +=
3572 cache->uc_allocbucket->ub_cnt;
3573 if (cache->uc_freebucket != NULL)
3574 ups.ups_cache_free +=
3575 cache->uc_freebucket->ub_cnt;
3576 ups.ups_allocs = cache->uc_allocs;
3577 ups.ups_frees = cache->uc_frees;
3578skip:
3579 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3580 }
3581 ZONE_UNLOCK(z);
3582 }
3583 }
3584 rw_runlock(&uma_rwlock);
3585 error = sbuf_finish(&sbuf);
3586 sbuf_delete(&sbuf);
3587 return (error);
3588}
3589
3590int
3591sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3592{
3593 uma_zone_t zone = *(uma_zone_t *)arg1;
3594 int error, max;
3595
3596 max = uma_zone_get_max(zone);
3597 error = sysctl_handle_int(oidp, &max, 0, req);
3598 if (error || !req->newptr)
3599 return (error);
3600
3601 uma_zone_set_max(zone, max);
3602
3603 return (0);
3604}
3605
3606int
3607sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3608{
3609 uma_zone_t zone = *(uma_zone_t *)arg1;
3610 int cur;
3611
3612 cur = uma_zone_get_cur(zone);
3613 return (sysctl_handle_int(oidp, &cur, 0, req));
3614}
3615
3616#ifdef INVARIANTS
3617static uma_slab_t
3618uma_dbg_getslab(uma_zone_t zone, void *item)
3619{
3620 uma_slab_t slab;
3621 uma_keg_t keg;
3622 uint8_t *mem;
3623
3624 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
3625 if (zone->uz_flags & UMA_ZONE_VTOSLAB) {
3626 slab = vtoslab((vm_offset_t)mem);
3627 } else {
3628 /*
3629 * It is safe to return the slab here even though the
3630 * zone is unlocked because the item's allocation state
3631 * essentially holds a reference.
3632 */
3633 ZONE_LOCK(zone);
3634 keg = LIST_FIRST(&zone->uz_kegs)->kl_keg;
3635 if (keg->uk_flags & UMA_ZONE_HASH)
3636 slab = hash_sfind(&keg->uk_hash, mem);
3637 else
3638 slab = (uma_slab_t)(mem + keg->uk_pgoff);
3639 ZONE_UNLOCK(zone);
3640 }
3641
3642 return (slab);
3643}
3644
3645/*
3646 * Set up the slab's freei data such that uma_dbg_free can function.
3647 *
3648 */
3649static void
3650uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
3651{
3652 uma_keg_t keg;
3653 int freei;
3654
3655 if (zone_first_keg(zone) == NULL)
3656 return;
3657 if (slab == NULL) {
3658 slab = uma_dbg_getslab(zone, item);
3659 if (slab == NULL)
3660 panic("uma: item %p did not belong to zone %s\n",
3661 item, zone->uz_name);
3662 }
3663 keg = slab->us_keg;
3664 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3665
3666 if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3667 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
3668 item, zone, zone->uz_name, slab, freei);
3669 BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3670
3671 return;
3672}
3673
3674/*
3675 * Verifies freed addresses. Checks for alignment, valid slab membership
3676 * and duplicate frees.
3677 *
3678 */
3679static void
3680uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
3681{
3682 uma_keg_t keg;
3683 int freei;
3684
3685 if (zone_first_keg(zone) == NULL)
3686 return;
3687 if (slab == NULL) {
3688 slab = uma_dbg_getslab(zone, item);
3689 if (slab == NULL)
3690 panic("uma: Freed item %p did not belong to zone %s\n",
3691 item, zone->uz_name);
3692 }
3693 keg = slab->us_keg;
3694 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3695
3696 if (freei >= keg->uk_ipers)
3697 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
3698 item, zone, zone->uz_name, slab, freei);
3699
3700 if (((freei * keg->uk_rsize) + slab->us_data) != item)
3701 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
3702 item, zone, zone->uz_name, slab, freei);
3703
3704 if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3705 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
3706 item, zone, zone->uz_name, slab, freei);
3707
3708 BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3709}
3710#endif /* INVARIANTS */
3711
3712#ifdef DDB
3713DB_SHOW_COMMAND(uma, db_show_uma)
3714{
3715 uint64_t allocs, frees, sleeps;
3716 uma_bucket_t bucket;
3717 uma_keg_t kz;
3718 uma_zone_t z;
3719 int cachefree;
3720
3721 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3722 "Free", "Requests", "Sleeps", "Bucket");
3723 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3724 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3725 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3726 allocs = z->uz_allocs;
3727 frees = z->uz_frees;
3728 sleeps = z->uz_sleeps;
3729 cachefree = 0;
3730 } else
3731 uma_zone_sumstat(z, &cachefree, &allocs,
3732 &frees, &sleeps);
3733 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3734 (LIST_FIRST(&kz->uk_zones) != z)))
3735 cachefree += kz->uk_free;
3736 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3737 cachefree += bucket->ub_cnt;
3738 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3739 z->uz_name, (uintmax_t)kz->uk_size,
3740 (intmax_t)(allocs - frees), cachefree,
3741 (uintmax_t)allocs, sleeps, z->uz_count);
3742 if (db_pager_quit)
3743 return;
3744 }
3745 }
3746}
3747
3748DB_SHOW_COMMAND(umacache, db_show_umacache)
3749{
3750 uint64_t allocs, frees;
3751 uma_bucket_t bucket;
3752 uma_zone_t z;
3753 int cachefree;
3754
3755 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3756 "Requests", "Bucket");
3757 LIST_FOREACH(z, &uma_cachezones, uz_link) {
3758 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3759 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3760 cachefree += bucket->ub_cnt;
3761 db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3762 z->uz_name, (uintmax_t)z->uz_size,
3763 (intmax_t)(allocs - frees), cachefree,
3764 (uintmax_t)allocs, z->uz_count);
3765 if (db_pager_quit)
3766 return;
3767 }
3768}
3769#endif /* DDB */
2 * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org>
3 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
4 * Copyright (c) 2004-2006 Robert N. M. Watson
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice unmodified, this list of conditions, and the following
12 * disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 */
28
29/*
30 * uma_core.c Implementation of the Universal Memory allocator
31 *
32 * This allocator is intended to replace the multitude of similar object caches
33 * in the standard FreeBSD kernel. The intent is to be flexible as well as
34 * effecient. A primary design goal is to return unused memory to the rest of
35 * the system. This will make the system as a whole more flexible due to the
36 * ability to move memory to subsystems which most need it instead of leaving
37 * pools of reserved memory unused.
38 *
39 * The basic ideas stem from similar slab/zone based allocators whose algorithms
40 * are well known.
41 *
42 */
43
44/*
45 * TODO:
46 * - Improve memory usage for large allocations
47 * - Investigate cache size adjustments
48 */
49
50#include <sys/cdefs.h>
51__FBSDID("$FreeBSD: head/sys/vm/uma_core.c 295222 2016-02-03 23:30:17Z glebius $");
52
53/* I should really use ktr.. */
54/*
55#define UMA_DEBUG 1
56#define UMA_DEBUG_ALLOC 1
57#define UMA_DEBUG_ALLOC_1 1
58*/
59
60#include "opt_ddb.h"
61#include "opt_param.h"
62#include "opt_vm.h"
63
64#include <sys/param.h>
65#include <sys/systm.h>
66#include <sys/bitset.h>
67#include <sys/kernel.h>
68#include <sys/types.h>
69#include <sys/queue.h>
70#include <sys/malloc.h>
71#include <sys/ktr.h>
72#include <sys/lock.h>
73#include <sys/sysctl.h>
74#include <sys/mutex.h>
75#include <sys/proc.h>
76#include <sys/random.h>
77#include <sys/rwlock.h>
78#include <sys/sbuf.h>
79#include <sys/sched.h>
80#include <sys/smp.h>
81#include <sys/taskqueue.h>
82#include <sys/vmmeter.h>
83
84#include <vm/vm.h>
85#include <vm/vm_object.h>
86#include <vm/vm_page.h>
87#include <vm/vm_pageout.h>
88#include <vm/vm_param.h>
89#include <vm/vm_map.h>
90#include <vm/vm_kern.h>
91#include <vm/vm_extern.h>
92#include <vm/uma.h>
93#include <vm/uma_int.h>
94#include <vm/uma_dbg.h>
95
96#include <ddb/ddb.h>
97
98#ifdef DEBUG_MEMGUARD
99#include <vm/memguard.h>
100#endif
101
102/*
103 * This is the zone and keg from which all zones are spawned. The idea is that
104 * even the zone & keg heads are allocated from the allocator, so we use the
105 * bss section to bootstrap us.
106 */
107static struct uma_keg masterkeg;
108static struct uma_zone masterzone_k;
109static struct uma_zone masterzone_z;
110static uma_zone_t kegs = &masterzone_k;
111static uma_zone_t zones = &masterzone_z;
112
113/* This is the zone from which all of uma_slab_t's are allocated. */
114static uma_zone_t slabzone;
115static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
116
117/*
118 * The initial hash tables come out of this zone so they can be allocated
119 * prior to malloc coming up.
120 */
121static uma_zone_t hashzone;
122
123/* The boot-time adjusted value for cache line alignment. */
124int uma_align_cache = 64 - 1;
125
126static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
127
128/*
129 * Are we allowed to allocate buckets?
130 */
131static int bucketdisable = 1;
132
133/* Linked list of all kegs in the system */
134static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
135
136/* Linked list of all cache-only zones in the system */
137static LIST_HEAD(,uma_zone) uma_cachezones =
138 LIST_HEAD_INITIALIZER(uma_cachezones);
139
140/* This RW lock protects the keg list */
141static struct rwlock_padalign uma_rwlock;
142
143/* Linked list of boot time pages */
144static LIST_HEAD(,uma_slab) uma_boot_pages =
145 LIST_HEAD_INITIALIZER(uma_boot_pages);
146
147/* This mutex protects the boot time pages list */
148static struct mtx_padalign uma_boot_pages_mtx;
149
150static struct sx uma_drain_lock;
151
152/* Is the VM done starting up? */
153static int booted = 0;
154#define UMA_STARTUP 1
155#define UMA_STARTUP2 2
156
157/*
158 * Only mbuf clusters use ref zones. Just provide enough references
159 * to support the one user. New code should not use the ref facility.
160 */
161static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES;
162
163/*
164 * This is the handle used to schedule events that need to happen
165 * outside of the allocation fast path.
166 */
167static struct callout uma_callout;
168#define UMA_TIMEOUT 20 /* Seconds for callout interval. */
169
170/*
171 * This structure is passed as the zone ctor arg so that I don't have to create
172 * a special allocation function just for zones.
173 */
174struct uma_zctor_args {
175 const char *name;
176 size_t size;
177 uma_ctor ctor;
178 uma_dtor dtor;
179 uma_init uminit;
180 uma_fini fini;
181 uma_import import;
182 uma_release release;
183 void *arg;
184 uma_keg_t keg;
185 int align;
186 uint32_t flags;
187};
188
189struct uma_kctor_args {
190 uma_zone_t zone;
191 size_t size;
192 uma_init uminit;
193 uma_fini fini;
194 int align;
195 uint32_t flags;
196};
197
198struct uma_bucket_zone {
199 uma_zone_t ubz_zone;
200 char *ubz_name;
201 int ubz_entries; /* Number of items it can hold. */
202 int ubz_maxsize; /* Maximum allocation size per-item. */
203};
204
205/*
206 * Compute the actual number of bucket entries to pack them in power
207 * of two sizes for more efficient space utilization.
208 */
209#define BUCKET_SIZE(n) \
210 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
211
212#define BUCKET_MAX BUCKET_SIZE(256)
213
214struct uma_bucket_zone bucket_zones[] = {
215 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
216 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
217 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
218 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
219 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
220 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
221 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
222 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
223 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
224 { NULL, NULL, 0}
225};
226
227/*
228 * Flags and enumerations to be passed to internal functions.
229 */
230enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
231
232/* Prototypes.. */
233
234static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
235static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
236static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
237static void page_free(void *, vm_size_t, uint8_t);
238static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
239static void cache_drain(uma_zone_t);
240static void bucket_drain(uma_zone_t, uma_bucket_t);
241static void bucket_cache_drain(uma_zone_t zone);
242static int keg_ctor(void *, int, void *, int);
243static void keg_dtor(void *, int, void *);
244static int zone_ctor(void *, int, void *, int);
245static void zone_dtor(void *, int, void *);
246static int zero_init(void *, int, int);
247static void keg_small_init(uma_keg_t keg);
248static void keg_large_init(uma_keg_t keg);
249static void zone_foreach(void (*zfunc)(uma_zone_t));
250static void zone_timeout(uma_zone_t zone);
251static int hash_alloc(struct uma_hash *);
252static int hash_expand(struct uma_hash *, struct uma_hash *);
253static void hash_free(struct uma_hash *hash);
254static void uma_timeout(void *);
255static void uma_startup3(void);
256static void *zone_alloc_item(uma_zone_t, void *, int);
257static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
258static void bucket_enable(void);
259static void bucket_init(void);
260static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
261static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
262static void bucket_zone_drain(void);
263static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
264static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
265static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
266static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
267static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
268static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
269 uma_fini fini, int align, uint32_t flags);
270static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
271static void zone_release(uma_zone_t zone, void **bucket, int cnt);
272static void uma_zero_item(void *item, uma_zone_t zone);
273
274void uma_print_zone(uma_zone_t);
275void uma_print_stats(void);
276static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
277static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
278
279#ifdef INVARIANTS
280static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
281static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
282#endif
283
284SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
285
286SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
287 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
288
289SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
290 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
291
292static int zone_warnings = 1;
293SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
294 "Warn when UMA zones becomes full");
295
296/*
297 * This routine checks to see whether or not it's safe to enable buckets.
298 */
299static void
300bucket_enable(void)
301{
302 bucketdisable = vm_page_count_min();
303}
304
305/*
306 * Initialize bucket_zones, the array of zones of buckets of various sizes.
307 *
308 * For each zone, calculate the memory required for each bucket, consisting
309 * of the header and an array of pointers.
310 */
311static void
312bucket_init(void)
313{
314 struct uma_bucket_zone *ubz;
315 int size;
316
317 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
318 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
319 size += sizeof(void *) * ubz->ubz_entries;
320 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
321 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
322 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
323 }
324}
325
326/*
327 * Given a desired number of entries for a bucket, return the zone from which
328 * to allocate the bucket.
329 */
330static struct uma_bucket_zone *
331bucket_zone_lookup(int entries)
332{
333 struct uma_bucket_zone *ubz;
334
335 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
336 if (ubz->ubz_entries >= entries)
337 return (ubz);
338 ubz--;
339 return (ubz);
340}
341
342static int
343bucket_select(int size)
344{
345 struct uma_bucket_zone *ubz;
346
347 ubz = &bucket_zones[0];
348 if (size > ubz->ubz_maxsize)
349 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
350
351 for (; ubz->ubz_entries != 0; ubz++)
352 if (ubz->ubz_maxsize < size)
353 break;
354 ubz--;
355 return (ubz->ubz_entries);
356}
357
358static uma_bucket_t
359bucket_alloc(uma_zone_t zone, void *udata, int flags)
360{
361 struct uma_bucket_zone *ubz;
362 uma_bucket_t bucket;
363
364 /*
365 * This is to stop us from allocating per cpu buckets while we're
366 * running out of vm.boot_pages. Otherwise, we would exhaust the
367 * boot pages. This also prevents us from allocating buckets in
368 * low memory situations.
369 */
370 if (bucketdisable)
371 return (NULL);
372 /*
373 * To limit bucket recursion we store the original zone flags
374 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
375 * NOVM flag to persist even through deep recursions. We also
376 * store ZFLAG_BUCKET once we have recursed attempting to allocate
377 * a bucket for a bucket zone so we do not allow infinite bucket
378 * recursion. This cookie will even persist to frees of unused
379 * buckets via the allocation path or bucket allocations in the
380 * free path.
381 */
382 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
383 udata = (void *)(uintptr_t)zone->uz_flags;
384 else {
385 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
386 return (NULL);
387 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
388 }
389 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
390 flags |= M_NOVM;
391 ubz = bucket_zone_lookup(zone->uz_count);
392 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
393 ubz++;
394 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
395 if (bucket) {
396#ifdef INVARIANTS
397 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
398#endif
399 bucket->ub_cnt = 0;
400 bucket->ub_entries = ubz->ubz_entries;
401 }
402
403 return (bucket);
404}
405
406static void
407bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
408{
409 struct uma_bucket_zone *ubz;
410
411 KASSERT(bucket->ub_cnt == 0,
412 ("bucket_free: Freeing a non free bucket."));
413 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
414 udata = (void *)(uintptr_t)zone->uz_flags;
415 ubz = bucket_zone_lookup(bucket->ub_entries);
416 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
417}
418
419static void
420bucket_zone_drain(void)
421{
422 struct uma_bucket_zone *ubz;
423
424 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
425 zone_drain(ubz->ubz_zone);
426}
427
428static void
429zone_log_warning(uma_zone_t zone)
430{
431 static const struct timeval warninterval = { 300, 0 };
432
433 if (!zone_warnings || zone->uz_warning == NULL)
434 return;
435
436 if (ratecheck(&zone->uz_ratecheck, &warninterval))
437 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
438}
439
440static inline void
441zone_maxaction(uma_zone_t zone)
442{
443
444 if (zone->uz_maxaction.ta_func != NULL)
445 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
446}
447
448static void
449zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
450{
451 uma_klink_t klink;
452
453 LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
454 kegfn(klink->kl_keg);
455}
456
457/*
458 * Routine called by timeout which is used to fire off some time interval
459 * based calculations. (stats, hash size, etc.)
460 *
461 * Arguments:
462 * arg Unused
463 *
464 * Returns:
465 * Nothing
466 */
467static void
468uma_timeout(void *unused)
469{
470 bucket_enable();
471 zone_foreach(zone_timeout);
472
473 /* Reschedule this event */
474 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
475}
476
477/*
478 * Routine to perform timeout driven calculations. This expands the
479 * hashes and does per cpu statistics aggregation.
480 *
481 * Returns nothing.
482 */
483static void
484keg_timeout(uma_keg_t keg)
485{
486
487 KEG_LOCK(keg);
488 /*
489 * Expand the keg hash table.
490 *
491 * This is done if the number of slabs is larger than the hash size.
492 * What I'm trying to do here is completely reduce collisions. This
493 * may be a little aggressive. Should I allow for two collisions max?
494 */
495 if (keg->uk_flags & UMA_ZONE_HASH &&
496 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
497 struct uma_hash newhash;
498 struct uma_hash oldhash;
499 int ret;
500
501 /*
502 * This is so involved because allocating and freeing
503 * while the keg lock is held will lead to deadlock.
504 * I have to do everything in stages and check for
505 * races.
506 */
507 newhash = keg->uk_hash;
508 KEG_UNLOCK(keg);
509 ret = hash_alloc(&newhash);
510 KEG_LOCK(keg);
511 if (ret) {
512 if (hash_expand(&keg->uk_hash, &newhash)) {
513 oldhash = keg->uk_hash;
514 keg->uk_hash = newhash;
515 } else
516 oldhash = newhash;
517
518 KEG_UNLOCK(keg);
519 hash_free(&oldhash);
520 return;
521 }
522 }
523 KEG_UNLOCK(keg);
524}
525
526static void
527zone_timeout(uma_zone_t zone)
528{
529
530 zone_foreach_keg(zone, &keg_timeout);
531}
532
533/*
534 * Allocate and zero fill the next sized hash table from the appropriate
535 * backing store.
536 *
537 * Arguments:
538 * hash A new hash structure with the old hash size in uh_hashsize
539 *
540 * Returns:
541 * 1 on sucess and 0 on failure.
542 */
543static int
544hash_alloc(struct uma_hash *hash)
545{
546 int oldsize;
547 int alloc;
548
549 oldsize = hash->uh_hashsize;
550
551 /* We're just going to go to a power of two greater */
552 if (oldsize) {
553 hash->uh_hashsize = oldsize * 2;
554 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
555 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
556 M_UMAHASH, M_NOWAIT);
557 } else {
558 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
559 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
560 M_WAITOK);
561 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
562 }
563 if (hash->uh_slab_hash) {
564 bzero(hash->uh_slab_hash, alloc);
565 hash->uh_hashmask = hash->uh_hashsize - 1;
566 return (1);
567 }
568
569 return (0);
570}
571
572/*
573 * Expands the hash table for HASH zones. This is done from zone_timeout
574 * to reduce collisions. This must not be done in the regular allocation
575 * path, otherwise, we can recurse on the vm while allocating pages.
576 *
577 * Arguments:
578 * oldhash The hash you want to expand
579 * newhash The hash structure for the new table
580 *
581 * Returns:
582 * Nothing
583 *
584 * Discussion:
585 */
586static int
587hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
588{
589 uma_slab_t slab;
590 int hval;
591 int i;
592
593 if (!newhash->uh_slab_hash)
594 return (0);
595
596 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
597 return (0);
598
599 /*
600 * I need to investigate hash algorithms for resizing without a
601 * full rehash.
602 */
603
604 for (i = 0; i < oldhash->uh_hashsize; i++)
605 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
606 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
607 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
608 hval = UMA_HASH(newhash, slab->us_data);
609 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
610 slab, us_hlink);
611 }
612
613 return (1);
614}
615
616/*
617 * Free the hash bucket to the appropriate backing store.
618 *
619 * Arguments:
620 * slab_hash The hash bucket we're freeing
621 * hashsize The number of entries in that hash bucket
622 *
623 * Returns:
624 * Nothing
625 */
626static void
627hash_free(struct uma_hash *hash)
628{
629 if (hash->uh_slab_hash == NULL)
630 return;
631 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
632 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
633 else
634 free(hash->uh_slab_hash, M_UMAHASH);
635}
636
637/*
638 * Frees all outstanding items in a bucket
639 *
640 * Arguments:
641 * zone The zone to free to, must be unlocked.
642 * bucket The free/alloc bucket with items, cpu queue must be locked.
643 *
644 * Returns:
645 * Nothing
646 */
647
648static void
649bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
650{
651 int i;
652
653 if (bucket == NULL)
654 return;
655
656 if (zone->uz_fini)
657 for (i = 0; i < bucket->ub_cnt; i++)
658 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
659 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
660 bucket->ub_cnt = 0;
661}
662
663/*
664 * Drains the per cpu caches for a zone.
665 *
666 * NOTE: This may only be called while the zone is being turn down, and not
667 * during normal operation. This is necessary in order that we do not have
668 * to migrate CPUs to drain the per-CPU caches.
669 *
670 * Arguments:
671 * zone The zone to drain, must be unlocked.
672 *
673 * Returns:
674 * Nothing
675 */
676static void
677cache_drain(uma_zone_t zone)
678{
679 uma_cache_t cache;
680 int cpu;
681
682 /*
683 * XXX: It is safe to not lock the per-CPU caches, because we're
684 * tearing down the zone anyway. I.e., there will be no further use
685 * of the caches at this point.
686 *
687 * XXX: It would good to be able to assert that the zone is being
688 * torn down to prevent improper use of cache_drain().
689 *
690 * XXX: We lock the zone before passing into bucket_cache_drain() as
691 * it is used elsewhere. Should the tear-down path be made special
692 * there in some form?
693 */
694 CPU_FOREACH(cpu) {
695 cache = &zone->uz_cpu[cpu];
696 bucket_drain(zone, cache->uc_allocbucket);
697 bucket_drain(zone, cache->uc_freebucket);
698 if (cache->uc_allocbucket != NULL)
699 bucket_free(zone, cache->uc_allocbucket, NULL);
700 if (cache->uc_freebucket != NULL)
701 bucket_free(zone, cache->uc_freebucket, NULL);
702 cache->uc_allocbucket = cache->uc_freebucket = NULL;
703 }
704 ZONE_LOCK(zone);
705 bucket_cache_drain(zone);
706 ZONE_UNLOCK(zone);
707}
708
709static void
710cache_shrink(uma_zone_t zone)
711{
712
713 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
714 return;
715
716 ZONE_LOCK(zone);
717 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
718 ZONE_UNLOCK(zone);
719}
720
721static void
722cache_drain_safe_cpu(uma_zone_t zone)
723{
724 uma_cache_t cache;
725 uma_bucket_t b1, b2;
726
727 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
728 return;
729
730 b1 = b2 = NULL;
731 ZONE_LOCK(zone);
732 critical_enter();
733 cache = &zone->uz_cpu[curcpu];
734 if (cache->uc_allocbucket) {
735 if (cache->uc_allocbucket->ub_cnt != 0)
736 LIST_INSERT_HEAD(&zone->uz_buckets,
737 cache->uc_allocbucket, ub_link);
738 else
739 b1 = cache->uc_allocbucket;
740 cache->uc_allocbucket = NULL;
741 }
742 if (cache->uc_freebucket) {
743 if (cache->uc_freebucket->ub_cnt != 0)
744 LIST_INSERT_HEAD(&zone->uz_buckets,
745 cache->uc_freebucket, ub_link);
746 else
747 b2 = cache->uc_freebucket;
748 cache->uc_freebucket = NULL;
749 }
750 critical_exit();
751 ZONE_UNLOCK(zone);
752 if (b1)
753 bucket_free(zone, b1, NULL);
754 if (b2)
755 bucket_free(zone, b2, NULL);
756}
757
758/*
759 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
760 * This is an expensive call because it needs to bind to all CPUs
761 * one by one and enter a critical section on each of them in order
762 * to safely access their cache buckets.
763 * Zone lock must not be held on call this function.
764 */
765static void
766cache_drain_safe(uma_zone_t zone)
767{
768 int cpu;
769
770 /*
771 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
772 */
773 if (zone)
774 cache_shrink(zone);
775 else
776 zone_foreach(cache_shrink);
777
778 CPU_FOREACH(cpu) {
779 thread_lock(curthread);
780 sched_bind(curthread, cpu);
781 thread_unlock(curthread);
782
783 if (zone)
784 cache_drain_safe_cpu(zone);
785 else
786 zone_foreach(cache_drain_safe_cpu);
787 }
788 thread_lock(curthread);
789 sched_unbind(curthread);
790 thread_unlock(curthread);
791}
792
793/*
794 * Drain the cached buckets from a zone. Expects a locked zone on entry.
795 */
796static void
797bucket_cache_drain(uma_zone_t zone)
798{
799 uma_bucket_t bucket;
800
801 /*
802 * Drain the bucket queues and free the buckets, we just keep two per
803 * cpu (alloc/free).
804 */
805 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
806 LIST_REMOVE(bucket, ub_link);
807 ZONE_UNLOCK(zone);
808 bucket_drain(zone, bucket);
809 bucket_free(zone, bucket, NULL);
810 ZONE_LOCK(zone);
811 }
812
813 /*
814 * Shrink further bucket sizes. Price of single zone lock collision
815 * is probably lower then price of global cache drain.
816 */
817 if (zone->uz_count > zone->uz_count_min)
818 zone->uz_count--;
819}
820
821static void
822keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
823{
824 uint8_t *mem;
825 int i;
826 uint8_t flags;
827
828 mem = slab->us_data;
829 flags = slab->us_flags;
830 i = start;
831 if (keg->uk_fini != NULL) {
832 for (i--; i > -1; i--)
833 keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
834 keg->uk_size);
835 }
836 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
837 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
838#ifdef UMA_DEBUG
839 printf("%s: Returning %d bytes.\n", keg->uk_name,
840 PAGE_SIZE * keg->uk_ppera);
841#endif
842 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
843}
844
845/*
846 * Frees pages from a keg back to the system. This is done on demand from
847 * the pageout daemon.
848 *
849 * Returns nothing.
850 */
851static void
852keg_drain(uma_keg_t keg)
853{
854 struct slabhead freeslabs = { 0 };
855 uma_slab_t slab;
856 uma_slab_t n;
857
858 /*
859 * We don't want to take pages from statically allocated kegs at this
860 * time
861 */
862 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
863 return;
864
865#ifdef UMA_DEBUG
866 printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
867#endif
868 KEG_LOCK(keg);
869 if (keg->uk_free == 0)
870 goto finished;
871
872 slab = LIST_FIRST(&keg->uk_free_slab);
873 while (slab) {
874 n = LIST_NEXT(slab, us_link);
875
876 /* We have no where to free these to */
877 if (slab->us_flags & UMA_SLAB_BOOT) {
878 slab = n;
879 continue;
880 }
881
882 LIST_REMOVE(slab, us_link);
883 keg->uk_pages -= keg->uk_ppera;
884 keg->uk_free -= keg->uk_ipers;
885
886 if (keg->uk_flags & UMA_ZONE_HASH)
887 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
888
889 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
890
891 slab = n;
892 }
893finished:
894 KEG_UNLOCK(keg);
895
896 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
897 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
898 keg_free_slab(keg, slab, keg->uk_ipers);
899 }
900}
901
902static void
903zone_drain_wait(uma_zone_t zone, int waitok)
904{
905
906 /*
907 * Set draining to interlock with zone_dtor() so we can release our
908 * locks as we go. Only dtor() should do a WAITOK call since it
909 * is the only call that knows the structure will still be available
910 * when it wakes up.
911 */
912 ZONE_LOCK(zone);
913 while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
914 if (waitok == M_NOWAIT)
915 goto out;
916 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
917 }
918 zone->uz_flags |= UMA_ZFLAG_DRAINING;
919 bucket_cache_drain(zone);
920 ZONE_UNLOCK(zone);
921 /*
922 * The DRAINING flag protects us from being freed while
923 * we're running. Normally the uma_rwlock would protect us but we
924 * must be able to release and acquire the right lock for each keg.
925 */
926 zone_foreach_keg(zone, &keg_drain);
927 ZONE_LOCK(zone);
928 zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
929 wakeup(zone);
930out:
931 ZONE_UNLOCK(zone);
932}
933
934void
935zone_drain(uma_zone_t zone)
936{
937
938 zone_drain_wait(zone, M_NOWAIT);
939}
940
941/*
942 * Allocate a new slab for a keg. This does not insert the slab onto a list.
943 *
944 * Arguments:
945 * wait Shall we wait?
946 *
947 * Returns:
948 * The slab that was allocated or NULL if there is no memory and the
949 * caller specified M_NOWAIT.
950 */
951static uma_slab_t
952keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
953{
954 uma_slabrefcnt_t slabref;
955 uma_alloc allocf;
956 uma_slab_t slab;
957 uint8_t *mem;
958 uint8_t flags;
959 int i;
960
961 mtx_assert(&keg->uk_lock, MA_OWNED);
962 slab = NULL;
963 mem = NULL;
964
965#ifdef UMA_DEBUG
966 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name);
967#endif
968 allocf = keg->uk_allocf;
969 KEG_UNLOCK(keg);
970
971 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
972 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
973 if (slab == NULL)
974 goto out;
975 }
976
977 /*
978 * This reproduces the old vm_zone behavior of zero filling pages the
979 * first time they are added to a zone.
980 *
981 * Malloced items are zeroed in uma_zalloc.
982 */
983
984 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
985 wait |= M_ZERO;
986 else
987 wait &= ~M_ZERO;
988
989 if (keg->uk_flags & UMA_ZONE_NODUMP)
990 wait |= M_NODUMP;
991
992 /* zone is passed for legacy reasons. */
993 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
994 if (mem == NULL) {
995 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
996 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
997 slab = NULL;
998 goto out;
999 }
1000
1001 /* Point the slab into the allocated memory */
1002 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
1003 slab = (uma_slab_t )(mem + keg->uk_pgoff);
1004
1005 if (keg->uk_flags & UMA_ZONE_VTOSLAB)
1006 for (i = 0; i < keg->uk_ppera; i++)
1007 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
1008
1009 slab->us_keg = keg;
1010 slab->us_data = mem;
1011 slab->us_freecount = keg->uk_ipers;
1012 slab->us_flags = flags;
1013 BIT_FILL(SLAB_SETSIZE, &slab->us_free);
1014#ifdef INVARIANTS
1015 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1016#endif
1017 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1018 slabref = (uma_slabrefcnt_t)slab;
1019 for (i = 0; i < keg->uk_ipers; i++)
1020 slabref->us_refcnt[i] = 0;
1021 }
1022
1023 if (keg->uk_init != NULL) {
1024 for (i = 0; i < keg->uk_ipers; i++)
1025 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1026 keg->uk_size, wait) != 0)
1027 break;
1028 if (i != keg->uk_ipers) {
1029 keg_free_slab(keg, slab, i);
1030 slab = NULL;
1031 goto out;
1032 }
1033 }
1034out:
1035 KEG_LOCK(keg);
1036
1037 if (slab != NULL) {
1038 if (keg->uk_flags & UMA_ZONE_HASH)
1039 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1040
1041 keg->uk_pages += keg->uk_ppera;
1042 keg->uk_free += keg->uk_ipers;
1043 }
1044
1045 return (slab);
1046}
1047
1048/*
1049 * This function is intended to be used early on in place of page_alloc() so
1050 * that we may use the boot time page cache to satisfy allocations before
1051 * the VM is ready.
1052 */
1053static void *
1054startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1055{
1056 uma_keg_t keg;
1057 uma_slab_t tmps;
1058 int pages, check_pages;
1059
1060 keg = zone_first_keg(zone);
1061 pages = howmany(bytes, PAGE_SIZE);
1062 check_pages = pages - 1;
1063 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1064
1065 /*
1066 * Check our small startup cache to see if it has pages remaining.
1067 */
1068 mtx_lock(&uma_boot_pages_mtx);
1069
1070 /* First check if we have enough room. */
1071 tmps = LIST_FIRST(&uma_boot_pages);
1072 while (tmps != NULL && check_pages-- > 0)
1073 tmps = LIST_NEXT(tmps, us_link);
1074 if (tmps != NULL) {
1075 /*
1076 * It's ok to lose tmps references. The last one will
1077 * have tmps->us_data pointing to the start address of
1078 * "pages" contiguous pages of memory.
1079 */
1080 while (pages-- > 0) {
1081 tmps = LIST_FIRST(&uma_boot_pages);
1082 LIST_REMOVE(tmps, us_link);
1083 }
1084 mtx_unlock(&uma_boot_pages_mtx);
1085 *pflag = tmps->us_flags;
1086 return (tmps->us_data);
1087 }
1088 mtx_unlock(&uma_boot_pages_mtx);
1089 if (booted < UMA_STARTUP2)
1090 panic("UMA: Increase vm.boot_pages");
1091 /*
1092 * Now that we've booted reset these users to their real allocator.
1093 */
1094#ifdef UMA_MD_SMALL_ALLOC
1095 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1096#else
1097 keg->uk_allocf = page_alloc;
1098#endif
1099 return keg->uk_allocf(zone, bytes, pflag, wait);
1100}
1101
1102/*
1103 * Allocates a number of pages from the system
1104 *
1105 * Arguments:
1106 * bytes The number of bytes requested
1107 * wait Shall we wait?
1108 *
1109 * Returns:
1110 * A pointer to the alloced memory or possibly
1111 * NULL if M_NOWAIT is set.
1112 */
1113static void *
1114page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1115{
1116 void *p; /* Returned page */
1117
1118 *pflag = UMA_SLAB_KMEM;
1119 p = (void *) kmem_malloc(kmem_arena, bytes, wait);
1120
1121 return (p);
1122}
1123
1124/*
1125 * Allocates a number of pages from within an object
1126 *
1127 * Arguments:
1128 * bytes The number of bytes requested
1129 * wait Shall we wait?
1130 *
1131 * Returns:
1132 * A pointer to the alloced memory or possibly
1133 * NULL if M_NOWAIT is set.
1134 */
1135static void *
1136noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait)
1137{
1138 TAILQ_HEAD(, vm_page) alloctail;
1139 u_long npages;
1140 vm_offset_t retkva, zkva;
1141 vm_page_t p, p_next;
1142 uma_keg_t keg;
1143
1144 TAILQ_INIT(&alloctail);
1145 keg = zone_first_keg(zone);
1146
1147 npages = howmany(bytes, PAGE_SIZE);
1148 while (npages > 0) {
1149 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1150 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1151 if (p != NULL) {
1152 /*
1153 * Since the page does not belong to an object, its
1154 * listq is unused.
1155 */
1156 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1157 npages--;
1158 continue;
1159 }
1160 if (wait & M_WAITOK) {
1161 VM_WAIT;
1162 continue;
1163 }
1164
1165 /*
1166 * Page allocation failed, free intermediate pages and
1167 * exit.
1168 */
1169 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1170 vm_page_unwire(p, PQ_NONE);
1171 vm_page_free(p);
1172 }
1173 return (NULL);
1174 }
1175 *flags = UMA_SLAB_PRIV;
1176 zkva = keg->uk_kva +
1177 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1178 retkva = zkva;
1179 TAILQ_FOREACH(p, &alloctail, listq) {
1180 pmap_qenter(zkva, &p, 1);
1181 zkva += PAGE_SIZE;
1182 }
1183
1184 return ((void *)retkva);
1185}
1186
1187/*
1188 * Frees a number of pages to the system
1189 *
1190 * Arguments:
1191 * mem A pointer to the memory to be freed
1192 * size The size of the memory being freed
1193 * flags The original p->us_flags field
1194 *
1195 * Returns:
1196 * Nothing
1197 */
1198static void
1199page_free(void *mem, vm_size_t size, uint8_t flags)
1200{
1201 struct vmem *vmem;
1202
1203 if (flags & UMA_SLAB_KMEM)
1204 vmem = kmem_arena;
1205 else if (flags & UMA_SLAB_KERNEL)
1206 vmem = kernel_arena;
1207 else
1208 panic("UMA: page_free used with invalid flags %d", flags);
1209
1210 kmem_free(vmem, (vm_offset_t)mem, size);
1211}
1212
1213/*
1214 * Zero fill initializer
1215 *
1216 * Arguments/Returns follow uma_init specifications
1217 */
1218static int
1219zero_init(void *mem, int size, int flags)
1220{
1221 bzero(mem, size);
1222 return (0);
1223}
1224
1225/*
1226 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1227 *
1228 * Arguments
1229 * keg The zone we should initialize
1230 *
1231 * Returns
1232 * Nothing
1233 */
1234static void
1235keg_small_init(uma_keg_t keg)
1236{
1237 u_int rsize;
1238 u_int memused;
1239 u_int wastedspace;
1240 u_int shsize;
1241
1242 if (keg->uk_flags & UMA_ZONE_PCPU) {
1243 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1244
1245 keg->uk_slabsize = sizeof(struct pcpu);
1246 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1247 PAGE_SIZE);
1248 } else {
1249 keg->uk_slabsize = UMA_SLAB_SIZE;
1250 keg->uk_ppera = 1;
1251 }
1252
1253 /*
1254 * Calculate the size of each allocation (rsize) according to
1255 * alignment. If the requested size is smaller than we have
1256 * allocation bits for we round it up.
1257 */
1258 rsize = keg->uk_size;
1259 if (rsize < keg->uk_slabsize / SLAB_SETSIZE)
1260 rsize = keg->uk_slabsize / SLAB_SETSIZE;
1261 if (rsize & keg->uk_align)
1262 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1263 keg->uk_rsize = rsize;
1264
1265 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1266 keg->uk_rsize < sizeof(struct pcpu),
1267 ("%s: size %u too large", __func__, keg->uk_rsize));
1268
1269 if (keg->uk_flags & UMA_ZONE_REFCNT)
1270 rsize += sizeof(uint32_t);
1271
1272 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1273 shsize = 0;
1274 else
1275 shsize = sizeof(struct uma_slab);
1276
1277 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1278 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1279 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1280
1281 memused = keg->uk_ipers * rsize + shsize;
1282 wastedspace = keg->uk_slabsize - memused;
1283
1284 /*
1285 * We can't do OFFPAGE if we're internal or if we've been
1286 * asked to not go to the VM for buckets. If we do this we
1287 * may end up going to the VM for slabs which we do not
1288 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1289 * of UMA_ZONE_VM, which clearly forbids it.
1290 */
1291 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1292 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1293 return;
1294
1295 /*
1296 * See if using an OFFPAGE slab will limit our waste. Only do
1297 * this if it permits more items per-slab.
1298 *
1299 * XXX We could try growing slabsize to limit max waste as well.
1300 * Historically this was not done because the VM could not
1301 * efficiently handle contiguous allocations.
1302 */
1303 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1304 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1305 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1306 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1307 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1308#ifdef UMA_DEBUG
1309 printf("UMA decided we need offpage slab headers for "
1310 "keg: %s, calculated wastedspace = %d, "
1311 "maximum wasted space allowed = %d, "
1312 "calculated ipers = %d, "
1313 "new wasted space = %d\n", keg->uk_name, wastedspace,
1314 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1315 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1316#endif
1317 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1318 }
1319
1320 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1321 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1322 keg->uk_flags |= UMA_ZONE_HASH;
1323}
1324
1325/*
1326 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1327 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1328 * more complicated.
1329 *
1330 * Arguments
1331 * keg The keg we should initialize
1332 *
1333 * Returns
1334 * Nothing
1335 */
1336static void
1337keg_large_init(uma_keg_t keg)
1338{
1339 u_int shsize;
1340
1341 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1342 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1343 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1344 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1345 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1346
1347 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1348 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1349 keg->uk_ipers = 1;
1350 keg->uk_rsize = keg->uk_size;
1351
1352 /* We can't do OFFPAGE if we're internal, bail out here. */
1353 if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1354 return;
1355
1356 /* Check whether we have enough space to not do OFFPAGE. */
1357 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1358 shsize = sizeof(struct uma_slab);
1359 if (keg->uk_flags & UMA_ZONE_REFCNT)
1360 shsize += keg->uk_ipers * sizeof(uint32_t);
1361 if (shsize & UMA_ALIGN_PTR)
1362 shsize = (shsize & ~UMA_ALIGN_PTR) +
1363 (UMA_ALIGN_PTR + 1);
1364
1365 if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize)
1366 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1367 }
1368
1369 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1370 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1371 keg->uk_flags |= UMA_ZONE_HASH;
1372}
1373
1374static void
1375keg_cachespread_init(uma_keg_t keg)
1376{
1377 int alignsize;
1378 int trailer;
1379 int pages;
1380 int rsize;
1381
1382 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1383 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1384
1385 alignsize = keg->uk_align + 1;
1386 rsize = keg->uk_size;
1387 /*
1388 * We want one item to start on every align boundary in a page. To
1389 * do this we will span pages. We will also extend the item by the
1390 * size of align if it is an even multiple of align. Otherwise, it
1391 * would fall on the same boundary every time.
1392 */
1393 if (rsize & keg->uk_align)
1394 rsize = (rsize & ~keg->uk_align) + alignsize;
1395 if ((rsize & alignsize) == 0)
1396 rsize += alignsize;
1397 trailer = rsize - keg->uk_size;
1398 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1399 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1400 keg->uk_rsize = rsize;
1401 keg->uk_ppera = pages;
1402 keg->uk_slabsize = UMA_SLAB_SIZE;
1403 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1404 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1405 KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1406 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1407 keg->uk_ipers));
1408}
1409
1410/*
1411 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1412 * the keg onto the global keg list.
1413 *
1414 * Arguments/Returns follow uma_ctor specifications
1415 * udata Actually uma_kctor_args
1416 */
1417static int
1418keg_ctor(void *mem, int size, void *udata, int flags)
1419{
1420 struct uma_kctor_args *arg = udata;
1421 uma_keg_t keg = mem;
1422 uma_zone_t zone;
1423
1424 bzero(keg, size);
1425 keg->uk_size = arg->size;
1426 keg->uk_init = arg->uminit;
1427 keg->uk_fini = arg->fini;
1428 keg->uk_align = arg->align;
1429 keg->uk_free = 0;
1430 keg->uk_reserve = 0;
1431 keg->uk_pages = 0;
1432 keg->uk_flags = arg->flags;
1433 keg->uk_allocf = page_alloc;
1434 keg->uk_freef = page_free;
1435 keg->uk_slabzone = NULL;
1436
1437 /*
1438 * The master zone is passed to us at keg-creation time.
1439 */
1440 zone = arg->zone;
1441 keg->uk_name = zone->uz_name;
1442
1443 if (arg->flags & UMA_ZONE_VM)
1444 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1445
1446 if (arg->flags & UMA_ZONE_ZINIT)
1447 keg->uk_init = zero_init;
1448
1449 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1450 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1451
1452 if (arg->flags & UMA_ZONE_PCPU)
1453#ifdef SMP
1454 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1455#else
1456 keg->uk_flags &= ~UMA_ZONE_PCPU;
1457#endif
1458
1459 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1460 keg_cachespread_init(keg);
1461 } else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1462 if (keg->uk_size >
1463 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) -
1464 sizeof(uint32_t)))
1465 keg_large_init(keg);
1466 else
1467 keg_small_init(keg);
1468 } else {
1469 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1470 keg_large_init(keg);
1471 else
1472 keg_small_init(keg);
1473 }
1474
1475 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1476 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1477 if (keg->uk_ipers > uma_max_ipers_ref)
1478 panic("Too many ref items per zone: %d > %d\n",
1479 keg->uk_ipers, uma_max_ipers_ref);
1480 keg->uk_slabzone = slabrefzone;
1481 } else
1482 keg->uk_slabzone = slabzone;
1483 }
1484
1485 /*
1486 * If we haven't booted yet we need allocations to go through the
1487 * startup cache until the vm is ready.
1488 */
1489 if (keg->uk_ppera == 1) {
1490#ifdef UMA_MD_SMALL_ALLOC
1491 keg->uk_allocf = uma_small_alloc;
1492 keg->uk_freef = uma_small_free;
1493
1494 if (booted < UMA_STARTUP)
1495 keg->uk_allocf = startup_alloc;
1496#else
1497 if (booted < UMA_STARTUP2)
1498 keg->uk_allocf = startup_alloc;
1499#endif
1500 } else if (booted < UMA_STARTUP2 &&
1501 (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1502 keg->uk_allocf = startup_alloc;
1503
1504 /*
1505 * Initialize keg's lock
1506 */
1507 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1508
1509 /*
1510 * If we're putting the slab header in the actual page we need to
1511 * figure out where in each page it goes. This calculates a right
1512 * justified offset into the memory on an ALIGN_PTR boundary.
1513 */
1514 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1515 u_int totsize;
1516
1517 /* Size of the slab struct and free list */
1518 totsize = sizeof(struct uma_slab);
1519
1520 /* Size of the reference counts. */
1521 if (keg->uk_flags & UMA_ZONE_REFCNT)
1522 totsize += keg->uk_ipers * sizeof(uint32_t);
1523
1524 if (totsize & UMA_ALIGN_PTR)
1525 totsize = (totsize & ~UMA_ALIGN_PTR) +
1526 (UMA_ALIGN_PTR + 1);
1527 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1528
1529 /*
1530 * The only way the following is possible is if with our
1531 * UMA_ALIGN_PTR adjustments we are now bigger than
1532 * UMA_SLAB_SIZE. I haven't checked whether this is
1533 * mathematically possible for all cases, so we make
1534 * sure here anyway.
1535 */
1536 totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1537 if (keg->uk_flags & UMA_ZONE_REFCNT)
1538 totsize += keg->uk_ipers * sizeof(uint32_t);
1539 if (totsize > PAGE_SIZE * keg->uk_ppera) {
1540 printf("zone %s ipers %d rsize %d size %d\n",
1541 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1542 keg->uk_size);
1543 panic("UMA slab won't fit.");
1544 }
1545 }
1546
1547 if (keg->uk_flags & UMA_ZONE_HASH)
1548 hash_alloc(&keg->uk_hash);
1549
1550#ifdef UMA_DEBUG
1551 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1552 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1553 keg->uk_ipers, keg->uk_ppera,
1554 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1555#endif
1556
1557 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1558
1559 rw_wlock(&uma_rwlock);
1560 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1561 rw_wunlock(&uma_rwlock);
1562 return (0);
1563}
1564
1565/*
1566 * Zone header ctor. This initializes all fields, locks, etc.
1567 *
1568 * Arguments/Returns follow uma_ctor specifications
1569 * udata Actually uma_zctor_args
1570 */
1571static int
1572zone_ctor(void *mem, int size, void *udata, int flags)
1573{
1574 struct uma_zctor_args *arg = udata;
1575 uma_zone_t zone = mem;
1576 uma_zone_t z;
1577 uma_keg_t keg;
1578
1579 bzero(zone, size);
1580 zone->uz_name = arg->name;
1581 zone->uz_ctor = arg->ctor;
1582 zone->uz_dtor = arg->dtor;
1583 zone->uz_slab = zone_fetch_slab;
1584 zone->uz_init = NULL;
1585 zone->uz_fini = NULL;
1586 zone->uz_allocs = 0;
1587 zone->uz_frees = 0;
1588 zone->uz_fails = 0;
1589 zone->uz_sleeps = 0;
1590 zone->uz_count = 0;
1591 zone->uz_count_min = 0;
1592 zone->uz_flags = 0;
1593 zone->uz_warning = NULL;
1594 timevalclear(&zone->uz_ratecheck);
1595 keg = arg->keg;
1596
1597 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1598
1599 /*
1600 * This is a pure cache zone, no kegs.
1601 */
1602 if (arg->import) {
1603 if (arg->flags & UMA_ZONE_VM)
1604 arg->flags |= UMA_ZFLAG_CACHEONLY;
1605 zone->uz_flags = arg->flags;
1606 zone->uz_size = arg->size;
1607 zone->uz_import = arg->import;
1608 zone->uz_release = arg->release;
1609 zone->uz_arg = arg->arg;
1610 zone->uz_lockptr = &zone->uz_lock;
1611 rw_wlock(&uma_rwlock);
1612 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1613 rw_wunlock(&uma_rwlock);
1614 goto out;
1615 }
1616
1617 /*
1618 * Use the regular zone/keg/slab allocator.
1619 */
1620 zone->uz_import = (uma_import)zone_import;
1621 zone->uz_release = (uma_release)zone_release;
1622 zone->uz_arg = zone;
1623
1624 if (arg->flags & UMA_ZONE_SECONDARY) {
1625 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1626 zone->uz_init = arg->uminit;
1627 zone->uz_fini = arg->fini;
1628 zone->uz_lockptr = &keg->uk_lock;
1629 zone->uz_flags |= UMA_ZONE_SECONDARY;
1630 rw_wlock(&uma_rwlock);
1631 ZONE_LOCK(zone);
1632 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1633 if (LIST_NEXT(z, uz_link) == NULL) {
1634 LIST_INSERT_AFTER(z, zone, uz_link);
1635 break;
1636 }
1637 }
1638 ZONE_UNLOCK(zone);
1639 rw_wunlock(&uma_rwlock);
1640 } else if (keg == NULL) {
1641 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1642 arg->align, arg->flags)) == NULL)
1643 return (ENOMEM);
1644 } else {
1645 struct uma_kctor_args karg;
1646 int error;
1647
1648 /* We should only be here from uma_startup() */
1649 karg.size = arg->size;
1650 karg.uminit = arg->uminit;
1651 karg.fini = arg->fini;
1652 karg.align = arg->align;
1653 karg.flags = arg->flags;
1654 karg.zone = zone;
1655 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1656 flags);
1657 if (error)
1658 return (error);
1659 }
1660
1661 /*
1662 * Link in the first keg.
1663 */
1664 zone->uz_klink.kl_keg = keg;
1665 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1666 zone->uz_lockptr = &keg->uk_lock;
1667 zone->uz_size = keg->uk_size;
1668 zone->uz_flags |= (keg->uk_flags &
1669 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1670
1671 /*
1672 * Some internal zones don't have room allocated for the per cpu
1673 * caches. If we're internal, bail out here.
1674 */
1675 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1676 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1677 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1678 return (0);
1679 }
1680
1681out:
1682 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1683 zone->uz_count = bucket_select(zone->uz_size);
1684 else
1685 zone->uz_count = BUCKET_MAX;
1686 zone->uz_count_min = zone->uz_count;
1687
1688 return (0);
1689}
1690
1691/*
1692 * Keg header dtor. This frees all data, destroys locks, frees the hash
1693 * table and removes the keg from the global list.
1694 *
1695 * Arguments/Returns follow uma_dtor specifications
1696 * udata unused
1697 */
1698static void
1699keg_dtor(void *arg, int size, void *udata)
1700{
1701 uma_keg_t keg;
1702
1703 keg = (uma_keg_t)arg;
1704 KEG_LOCK(keg);
1705 if (keg->uk_free != 0) {
1706 printf("Freed UMA keg (%s) was not empty (%d items). "
1707 " Lost %d pages of memory.\n",
1708 keg->uk_name ? keg->uk_name : "",
1709 keg->uk_free, keg->uk_pages);
1710 }
1711 KEG_UNLOCK(keg);
1712
1713 hash_free(&keg->uk_hash);
1714
1715 KEG_LOCK_FINI(keg);
1716}
1717
1718/*
1719 * Zone header dtor.
1720 *
1721 * Arguments/Returns follow uma_dtor specifications
1722 * udata unused
1723 */
1724static void
1725zone_dtor(void *arg, int size, void *udata)
1726{
1727 uma_klink_t klink;
1728 uma_zone_t zone;
1729 uma_keg_t keg;
1730
1731 zone = (uma_zone_t)arg;
1732 keg = zone_first_keg(zone);
1733
1734 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1735 cache_drain(zone);
1736
1737 rw_wlock(&uma_rwlock);
1738 LIST_REMOVE(zone, uz_link);
1739 rw_wunlock(&uma_rwlock);
1740 /*
1741 * XXX there are some races here where
1742 * the zone can be drained but zone lock
1743 * released and then refilled before we
1744 * remove it... we dont care for now
1745 */
1746 zone_drain_wait(zone, M_WAITOK);
1747 /*
1748 * Unlink all of our kegs.
1749 */
1750 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1751 klink->kl_keg = NULL;
1752 LIST_REMOVE(klink, kl_link);
1753 if (klink == &zone->uz_klink)
1754 continue;
1755 free(klink, M_TEMP);
1756 }
1757 /*
1758 * We only destroy kegs from non secondary zones.
1759 */
1760 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1761 rw_wlock(&uma_rwlock);
1762 LIST_REMOVE(keg, uk_link);
1763 rw_wunlock(&uma_rwlock);
1764 zone_free_item(kegs, keg, NULL, SKIP_NONE);
1765 }
1766 ZONE_LOCK_FINI(zone);
1767}
1768
1769/*
1770 * Traverses every zone in the system and calls a callback
1771 *
1772 * Arguments:
1773 * zfunc A pointer to a function which accepts a zone
1774 * as an argument.
1775 *
1776 * Returns:
1777 * Nothing
1778 */
1779static void
1780zone_foreach(void (*zfunc)(uma_zone_t))
1781{
1782 uma_keg_t keg;
1783 uma_zone_t zone;
1784
1785 rw_rlock(&uma_rwlock);
1786 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1787 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1788 zfunc(zone);
1789 }
1790 rw_runlock(&uma_rwlock);
1791}
1792
1793/* Public functions */
1794/* See uma.h */
1795void
1796uma_startup(void *bootmem, int boot_pages)
1797{
1798 struct uma_zctor_args args;
1799 uma_slab_t slab;
1800 u_int slabsize;
1801 int i;
1802
1803#ifdef UMA_DEBUG
1804 printf("Creating uma keg headers zone and keg.\n");
1805#endif
1806 rw_init(&uma_rwlock, "UMA lock");
1807
1808 /* "manually" create the initial zone */
1809 memset(&args, 0, sizeof(args));
1810 args.name = "UMA Kegs";
1811 args.size = sizeof(struct uma_keg);
1812 args.ctor = keg_ctor;
1813 args.dtor = keg_dtor;
1814 args.uminit = zero_init;
1815 args.fini = NULL;
1816 args.keg = &masterkeg;
1817 args.align = 32 - 1;
1818 args.flags = UMA_ZFLAG_INTERNAL;
1819 /* The initial zone has no Per cpu queues so it's smaller */
1820 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1821
1822#ifdef UMA_DEBUG
1823 printf("Filling boot free list.\n");
1824#endif
1825 for (i = 0; i < boot_pages; i++) {
1826 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1827 slab->us_data = (uint8_t *)slab;
1828 slab->us_flags = UMA_SLAB_BOOT;
1829 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1830 }
1831 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1832
1833#ifdef UMA_DEBUG
1834 printf("Creating uma zone headers zone and keg.\n");
1835#endif
1836 args.name = "UMA Zones";
1837 args.size = sizeof(struct uma_zone) +
1838 (sizeof(struct uma_cache) * (mp_maxid + 1));
1839 args.ctor = zone_ctor;
1840 args.dtor = zone_dtor;
1841 args.uminit = zero_init;
1842 args.fini = NULL;
1843 args.keg = NULL;
1844 args.align = 32 - 1;
1845 args.flags = UMA_ZFLAG_INTERNAL;
1846 /* The initial zone has no Per cpu queues so it's smaller */
1847 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1848
1849#ifdef UMA_DEBUG
1850 printf("Creating slab and hash zones.\n");
1851#endif
1852
1853 /* Now make a zone for slab headers */
1854 slabzone = uma_zcreate("UMA Slabs",
1855 sizeof(struct uma_slab),
1856 NULL, NULL, NULL, NULL,
1857 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1858
1859 /*
1860 * We also create a zone for the bigger slabs with reference
1861 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1862 */
1863 slabsize = sizeof(struct uma_slab_refcnt);
1864 slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1865 slabrefzone = uma_zcreate("UMA RCntSlabs",
1866 slabsize,
1867 NULL, NULL, NULL, NULL,
1868 UMA_ALIGN_PTR,
1869 UMA_ZFLAG_INTERNAL);
1870
1871 hashzone = uma_zcreate("UMA Hash",
1872 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1873 NULL, NULL, NULL, NULL,
1874 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1875
1876 bucket_init();
1877
1878 booted = UMA_STARTUP;
1879
1880#ifdef UMA_DEBUG
1881 printf("UMA startup complete.\n");
1882#endif
1883}
1884
1885/* see uma.h */
1886void
1887uma_startup2(void)
1888{
1889 booted = UMA_STARTUP2;
1890 bucket_enable();
1891 sx_init(&uma_drain_lock, "umadrain");
1892#ifdef UMA_DEBUG
1893 printf("UMA startup2 complete.\n");
1894#endif
1895}
1896
1897/*
1898 * Initialize our callout handle
1899 *
1900 */
1901
1902static void
1903uma_startup3(void)
1904{
1905#ifdef UMA_DEBUG
1906 printf("Starting callout.\n");
1907#endif
1908 callout_init(&uma_callout, 1);
1909 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1910#ifdef UMA_DEBUG
1911 printf("UMA startup3 complete.\n");
1912#endif
1913}
1914
1915static uma_keg_t
1916uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1917 int align, uint32_t flags)
1918{
1919 struct uma_kctor_args args;
1920
1921 args.size = size;
1922 args.uminit = uminit;
1923 args.fini = fini;
1924 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1925 args.flags = flags;
1926 args.zone = zone;
1927 return (zone_alloc_item(kegs, &args, M_WAITOK));
1928}
1929
1930/* See uma.h */
1931void
1932uma_set_align(int align)
1933{
1934
1935 if (align != UMA_ALIGN_CACHE)
1936 uma_align_cache = align;
1937}
1938
1939/* See uma.h */
1940uma_zone_t
1941uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1942 uma_init uminit, uma_fini fini, int align, uint32_t flags)
1943
1944{
1945 struct uma_zctor_args args;
1946 uma_zone_t res;
1947 bool locked;
1948
1949 /* This stuff is essential for the zone ctor */
1950 memset(&args, 0, sizeof(args));
1951 args.name = name;
1952 args.size = size;
1953 args.ctor = ctor;
1954 args.dtor = dtor;
1955 args.uminit = uminit;
1956 args.fini = fini;
1957#ifdef INVARIANTS
1958 /*
1959 * If a zone is being created with an empty constructor and
1960 * destructor, pass UMA constructor/destructor which checks for
1961 * memory use after free.
1962 */
1963 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) &&
1964 ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) {
1965 args.ctor = trash_ctor;
1966 args.dtor = trash_dtor;
1967 args.uminit = trash_init;
1968 args.fini = trash_fini;
1969 }
1970#endif
1971 args.align = align;
1972 args.flags = flags;
1973 args.keg = NULL;
1974
1975 if (booted < UMA_STARTUP2) {
1976 locked = false;
1977 } else {
1978 sx_slock(&uma_drain_lock);
1979 locked = true;
1980 }
1981 res = zone_alloc_item(zones, &args, M_WAITOK);
1982 if (locked)
1983 sx_sunlock(&uma_drain_lock);
1984 return (res);
1985}
1986
1987/* See uma.h */
1988uma_zone_t
1989uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1990 uma_init zinit, uma_fini zfini, uma_zone_t master)
1991{
1992 struct uma_zctor_args args;
1993 uma_keg_t keg;
1994 uma_zone_t res;
1995 bool locked;
1996
1997 keg = zone_first_keg(master);
1998 memset(&args, 0, sizeof(args));
1999 args.name = name;
2000 args.size = keg->uk_size;
2001 args.ctor = ctor;
2002 args.dtor = dtor;
2003 args.uminit = zinit;
2004 args.fini = zfini;
2005 args.align = keg->uk_align;
2006 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
2007 args.keg = keg;
2008
2009 if (booted < UMA_STARTUP2) {
2010 locked = false;
2011 } else {
2012 sx_slock(&uma_drain_lock);
2013 locked = true;
2014 }
2015 /* XXX Attaches only one keg of potentially many. */
2016 res = zone_alloc_item(zones, &args, M_WAITOK);
2017 if (locked)
2018 sx_sunlock(&uma_drain_lock);
2019 return (res);
2020}
2021
2022/* See uma.h */
2023uma_zone_t
2024uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
2025 uma_init zinit, uma_fini zfini, uma_import zimport,
2026 uma_release zrelease, void *arg, int flags)
2027{
2028 struct uma_zctor_args args;
2029
2030 memset(&args, 0, sizeof(args));
2031 args.name = name;
2032 args.size = size;
2033 args.ctor = ctor;
2034 args.dtor = dtor;
2035 args.uminit = zinit;
2036 args.fini = zfini;
2037 args.import = zimport;
2038 args.release = zrelease;
2039 args.arg = arg;
2040 args.align = 0;
2041 args.flags = flags;
2042
2043 return (zone_alloc_item(zones, &args, M_WAITOK));
2044}
2045
2046static void
2047zone_lock_pair(uma_zone_t a, uma_zone_t b)
2048{
2049 if (a < b) {
2050 ZONE_LOCK(a);
2051 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
2052 } else {
2053 ZONE_LOCK(b);
2054 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
2055 }
2056}
2057
2058static void
2059zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2060{
2061
2062 ZONE_UNLOCK(a);
2063 ZONE_UNLOCK(b);
2064}
2065
2066int
2067uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2068{
2069 uma_klink_t klink;
2070 uma_klink_t kl;
2071 int error;
2072
2073 error = 0;
2074 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2075
2076 zone_lock_pair(zone, master);
2077 /*
2078 * zone must use vtoslab() to resolve objects and must already be
2079 * a secondary.
2080 */
2081 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2082 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2083 error = EINVAL;
2084 goto out;
2085 }
2086 /*
2087 * The new master must also use vtoslab().
2088 */
2089 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2090 error = EINVAL;
2091 goto out;
2092 }
2093 /*
2094 * Both must either be refcnt, or not be refcnt.
2095 */
2096 if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
2097 (master->uz_flags & UMA_ZONE_REFCNT)) {
2098 error = EINVAL;
2099 goto out;
2100 }
2101 /*
2102 * The underlying object must be the same size. rsize
2103 * may be different.
2104 */
2105 if (master->uz_size != zone->uz_size) {
2106 error = E2BIG;
2107 goto out;
2108 }
2109 /*
2110 * Put it at the end of the list.
2111 */
2112 klink->kl_keg = zone_first_keg(master);
2113 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2114 if (LIST_NEXT(kl, kl_link) == NULL) {
2115 LIST_INSERT_AFTER(kl, klink, kl_link);
2116 break;
2117 }
2118 }
2119 klink = NULL;
2120 zone->uz_flags |= UMA_ZFLAG_MULTI;
2121 zone->uz_slab = zone_fetch_slab_multi;
2122
2123out:
2124 zone_unlock_pair(zone, master);
2125 if (klink != NULL)
2126 free(klink, M_TEMP);
2127
2128 return (error);
2129}
2130
2131
2132/* See uma.h */
2133void
2134uma_zdestroy(uma_zone_t zone)
2135{
2136
2137 sx_slock(&uma_drain_lock);
2138 zone_free_item(zones, zone, NULL, SKIP_NONE);
2139 sx_sunlock(&uma_drain_lock);
2140}
2141
2142/* See uma.h */
2143void *
2144uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2145{
2146 void *item;
2147 uma_cache_t cache;
2148 uma_bucket_t bucket;
2149 int lockfail;
2150 int cpu;
2151
2152 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2153 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2154
2155 /* This is the fast path allocation */
2156#ifdef UMA_DEBUG_ALLOC_1
2157 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2158#endif
2159 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2160 zone->uz_name, flags);
2161
2162 if (flags & M_WAITOK) {
2163 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2164 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2165 }
2166 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2167 ("uma_zalloc_arg: called with spinlock or critical section held"));
2168
2169#ifdef DEBUG_MEMGUARD
2170 if (memguard_cmp_zone(zone)) {
2171 item = memguard_alloc(zone->uz_size, flags);
2172 if (item != NULL) {
2173 /*
2174 * Avoid conflict with the use-after-free
2175 * protecting infrastructure from INVARIANTS.
2176 */
2177 if (zone->uz_init != NULL &&
2178 zone->uz_init != mtrash_init &&
2179 zone->uz_init(item, zone->uz_size, flags) != 0)
2180 return (NULL);
2181 if (zone->uz_ctor != NULL &&
2182 zone->uz_ctor != mtrash_ctor &&
2183 zone->uz_ctor(item, zone->uz_size, udata,
2184 flags) != 0) {
2185 zone->uz_fini(item, zone->uz_size);
2186 return (NULL);
2187 }
2188 return (item);
2189 }
2190 /* This is unfortunate but should not be fatal. */
2191 }
2192#endif
2193 /*
2194 * If possible, allocate from the per-CPU cache. There are two
2195 * requirements for safe access to the per-CPU cache: (1) the thread
2196 * accessing the cache must not be preempted or yield during access,
2197 * and (2) the thread must not migrate CPUs without switching which
2198 * cache it accesses. We rely on a critical section to prevent
2199 * preemption and migration. We release the critical section in
2200 * order to acquire the zone mutex if we are unable to allocate from
2201 * the current cache; when we re-acquire the critical section, we
2202 * must detect and handle migration if it has occurred.
2203 */
2204 critical_enter();
2205 cpu = curcpu;
2206 cache = &zone->uz_cpu[cpu];
2207
2208zalloc_start:
2209 bucket = cache->uc_allocbucket;
2210 if (bucket != NULL && bucket->ub_cnt > 0) {
2211 bucket->ub_cnt--;
2212 item = bucket->ub_bucket[bucket->ub_cnt];
2213#ifdef INVARIANTS
2214 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2215#endif
2216 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2217 cache->uc_allocs++;
2218 critical_exit();
2219 if (zone->uz_ctor != NULL &&
2220 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2221 atomic_add_long(&zone->uz_fails, 1);
2222 zone_free_item(zone, item, udata, SKIP_DTOR);
2223 return (NULL);
2224 }
2225#ifdef INVARIANTS
2226 uma_dbg_alloc(zone, NULL, item);
2227#endif
2228 if (flags & M_ZERO)
2229 uma_zero_item(item, zone);
2230 return (item);
2231 }
2232
2233 /*
2234 * We have run out of items in our alloc bucket.
2235 * See if we can switch with our free bucket.
2236 */
2237 bucket = cache->uc_freebucket;
2238 if (bucket != NULL && bucket->ub_cnt > 0) {
2239#ifdef UMA_DEBUG_ALLOC
2240 printf("uma_zalloc: Swapping empty with alloc.\n");
2241#endif
2242 cache->uc_freebucket = cache->uc_allocbucket;
2243 cache->uc_allocbucket = bucket;
2244 goto zalloc_start;
2245 }
2246
2247 /*
2248 * Discard any empty allocation bucket while we hold no locks.
2249 */
2250 bucket = cache->uc_allocbucket;
2251 cache->uc_allocbucket = NULL;
2252 critical_exit();
2253 if (bucket != NULL)
2254 bucket_free(zone, bucket, udata);
2255
2256 /* Short-circuit for zones without buckets and low memory. */
2257 if (zone->uz_count == 0 || bucketdisable)
2258 goto zalloc_item;
2259
2260 /*
2261 * Attempt to retrieve the item from the per-CPU cache has failed, so
2262 * we must go back to the zone. This requires the zone lock, so we
2263 * must drop the critical section, then re-acquire it when we go back
2264 * to the cache. Since the critical section is released, we may be
2265 * preempted or migrate. As such, make sure not to maintain any
2266 * thread-local state specific to the cache from prior to releasing
2267 * the critical section.
2268 */
2269 lockfail = 0;
2270 if (ZONE_TRYLOCK(zone) == 0) {
2271 /* Record contention to size the buckets. */
2272 ZONE_LOCK(zone);
2273 lockfail = 1;
2274 }
2275 critical_enter();
2276 cpu = curcpu;
2277 cache = &zone->uz_cpu[cpu];
2278
2279 /*
2280 * Since we have locked the zone we may as well send back our stats.
2281 */
2282 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2283 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2284 cache->uc_allocs = 0;
2285 cache->uc_frees = 0;
2286
2287 /* See if we lost the race to fill the cache. */
2288 if (cache->uc_allocbucket != NULL) {
2289 ZONE_UNLOCK(zone);
2290 goto zalloc_start;
2291 }
2292
2293 /*
2294 * Check the zone's cache of buckets.
2295 */
2296 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2297 KASSERT(bucket->ub_cnt != 0,
2298 ("uma_zalloc_arg: Returning an empty bucket."));
2299
2300 LIST_REMOVE(bucket, ub_link);
2301 cache->uc_allocbucket = bucket;
2302 ZONE_UNLOCK(zone);
2303 goto zalloc_start;
2304 }
2305 /* We are no longer associated with this CPU. */
2306 critical_exit();
2307
2308 /*
2309 * We bump the uz count when the cache size is insufficient to
2310 * handle the working set.
2311 */
2312 if (lockfail && zone->uz_count < BUCKET_MAX)
2313 zone->uz_count++;
2314 ZONE_UNLOCK(zone);
2315
2316 /*
2317 * Now lets just fill a bucket and put it on the free list. If that
2318 * works we'll restart the allocation from the begining and it
2319 * will use the just filled bucket.
2320 */
2321 bucket = zone_alloc_bucket(zone, udata, flags);
2322 if (bucket != NULL) {
2323 ZONE_LOCK(zone);
2324 critical_enter();
2325 cpu = curcpu;
2326 cache = &zone->uz_cpu[cpu];
2327 /*
2328 * See if we lost the race or were migrated. Cache the
2329 * initialized bucket to make this less likely or claim
2330 * the memory directly.
2331 */
2332 if (cache->uc_allocbucket == NULL)
2333 cache->uc_allocbucket = bucket;
2334 else
2335 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2336 ZONE_UNLOCK(zone);
2337 goto zalloc_start;
2338 }
2339
2340 /*
2341 * We may not be able to get a bucket so return an actual item.
2342 */
2343#ifdef UMA_DEBUG
2344 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2345#endif
2346
2347zalloc_item:
2348 item = zone_alloc_item(zone, udata, flags);
2349
2350 return (item);
2351}
2352
2353static uma_slab_t
2354keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2355{
2356 uma_slab_t slab;
2357 int reserve;
2358
2359 mtx_assert(&keg->uk_lock, MA_OWNED);
2360 slab = NULL;
2361 reserve = 0;
2362 if ((flags & M_USE_RESERVE) == 0)
2363 reserve = keg->uk_reserve;
2364
2365 for (;;) {
2366 /*
2367 * Find a slab with some space. Prefer slabs that are partially
2368 * used over those that are totally full. This helps to reduce
2369 * fragmentation.
2370 */
2371 if (keg->uk_free > reserve) {
2372 if (!LIST_EMPTY(&keg->uk_part_slab)) {
2373 slab = LIST_FIRST(&keg->uk_part_slab);
2374 } else {
2375 slab = LIST_FIRST(&keg->uk_free_slab);
2376 LIST_REMOVE(slab, us_link);
2377 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2378 us_link);
2379 }
2380 MPASS(slab->us_keg == keg);
2381 return (slab);
2382 }
2383
2384 /*
2385 * M_NOVM means don't ask at all!
2386 */
2387 if (flags & M_NOVM)
2388 break;
2389
2390 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2391 keg->uk_flags |= UMA_ZFLAG_FULL;
2392 /*
2393 * If this is not a multi-zone, set the FULL bit.
2394 * Otherwise slab_multi() takes care of it.
2395 */
2396 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2397 zone->uz_flags |= UMA_ZFLAG_FULL;
2398 zone_log_warning(zone);
2399 zone_maxaction(zone);
2400 }
2401 if (flags & M_NOWAIT)
2402 break;
2403 zone->uz_sleeps++;
2404 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2405 continue;
2406 }
2407 slab = keg_alloc_slab(keg, zone, flags);
2408 /*
2409 * If we got a slab here it's safe to mark it partially used
2410 * and return. We assume that the caller is going to remove
2411 * at least one item.
2412 */
2413 if (slab) {
2414 MPASS(slab->us_keg == keg);
2415 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2416 return (slab);
2417 }
2418 /*
2419 * We might not have been able to get a slab but another cpu
2420 * could have while we were unlocked. Check again before we
2421 * fail.
2422 */
2423 flags |= M_NOVM;
2424 }
2425 return (slab);
2426}
2427
2428static uma_slab_t
2429zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2430{
2431 uma_slab_t slab;
2432
2433 if (keg == NULL) {
2434 keg = zone_first_keg(zone);
2435 KEG_LOCK(keg);
2436 }
2437
2438 for (;;) {
2439 slab = keg_fetch_slab(keg, zone, flags);
2440 if (slab)
2441 return (slab);
2442 if (flags & (M_NOWAIT | M_NOVM))
2443 break;
2444 }
2445 KEG_UNLOCK(keg);
2446 return (NULL);
2447}
2448
2449/*
2450 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2451 * with the keg locked. On NULL no lock is held.
2452 *
2453 * The last pointer is used to seed the search. It is not required.
2454 */
2455static uma_slab_t
2456zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2457{
2458 uma_klink_t klink;
2459 uma_slab_t slab;
2460 uma_keg_t keg;
2461 int flags;
2462 int empty;
2463 int full;
2464
2465 /*
2466 * Don't wait on the first pass. This will skip limit tests
2467 * as well. We don't want to block if we can find a provider
2468 * without blocking.
2469 */
2470 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2471 /*
2472 * Use the last slab allocated as a hint for where to start
2473 * the search.
2474 */
2475 if (last != NULL) {
2476 slab = keg_fetch_slab(last, zone, flags);
2477 if (slab)
2478 return (slab);
2479 KEG_UNLOCK(last);
2480 }
2481 /*
2482 * Loop until we have a slab incase of transient failures
2483 * while M_WAITOK is specified. I'm not sure this is 100%
2484 * required but we've done it for so long now.
2485 */
2486 for (;;) {
2487 empty = 0;
2488 full = 0;
2489 /*
2490 * Search the available kegs for slabs. Be careful to hold the
2491 * correct lock while calling into the keg layer.
2492 */
2493 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2494 keg = klink->kl_keg;
2495 KEG_LOCK(keg);
2496 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2497 slab = keg_fetch_slab(keg, zone, flags);
2498 if (slab)
2499 return (slab);
2500 }
2501 if (keg->uk_flags & UMA_ZFLAG_FULL)
2502 full++;
2503 else
2504 empty++;
2505 KEG_UNLOCK(keg);
2506 }
2507 if (rflags & (M_NOWAIT | M_NOVM))
2508 break;
2509 flags = rflags;
2510 /*
2511 * All kegs are full. XXX We can't atomically check all kegs
2512 * and sleep so just sleep for a short period and retry.
2513 */
2514 if (full && !empty) {
2515 ZONE_LOCK(zone);
2516 zone->uz_flags |= UMA_ZFLAG_FULL;
2517 zone->uz_sleeps++;
2518 zone_log_warning(zone);
2519 zone_maxaction(zone);
2520 msleep(zone, zone->uz_lockptr, PVM,
2521 "zonelimit", hz/100);
2522 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2523 ZONE_UNLOCK(zone);
2524 continue;
2525 }
2526 }
2527 return (NULL);
2528}
2529
2530static void *
2531slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2532{
2533 void *item;
2534 uint8_t freei;
2535
2536 MPASS(keg == slab->us_keg);
2537 mtx_assert(&keg->uk_lock, MA_OWNED);
2538
2539 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2540 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2541 item = slab->us_data + (keg->uk_rsize * freei);
2542 slab->us_freecount--;
2543 keg->uk_free--;
2544
2545 /* Move this slab to the full list */
2546 if (slab->us_freecount == 0) {
2547 LIST_REMOVE(slab, us_link);
2548 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2549 }
2550
2551 return (item);
2552}
2553
2554static int
2555zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2556{
2557 uma_slab_t slab;
2558 uma_keg_t keg;
2559 int i;
2560
2561 slab = NULL;
2562 keg = NULL;
2563 /* Try to keep the buckets totally full */
2564 for (i = 0; i < max; ) {
2565 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2566 break;
2567 keg = slab->us_keg;
2568 while (slab->us_freecount && i < max) {
2569 bucket[i++] = slab_alloc_item(keg, slab);
2570 if (keg->uk_free <= keg->uk_reserve)
2571 break;
2572 }
2573 /* Don't grab more than one slab at a time. */
2574 flags &= ~M_WAITOK;
2575 flags |= M_NOWAIT;
2576 }
2577 if (slab != NULL)
2578 KEG_UNLOCK(keg);
2579
2580 return i;
2581}
2582
2583static uma_bucket_t
2584zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2585{
2586 uma_bucket_t bucket;
2587 int max;
2588
2589 /* Don't wait for buckets, preserve caller's NOVM setting. */
2590 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2591 if (bucket == NULL)
2592 return (NULL);
2593
2594 max = MIN(bucket->ub_entries, zone->uz_count);
2595 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2596 max, flags);
2597
2598 /*
2599 * Initialize the memory if necessary.
2600 */
2601 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2602 int i;
2603
2604 for (i = 0; i < bucket->ub_cnt; i++)
2605 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2606 flags) != 0)
2607 break;
2608 /*
2609 * If we couldn't initialize the whole bucket, put the
2610 * rest back onto the freelist.
2611 */
2612 if (i != bucket->ub_cnt) {
2613 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2614 bucket->ub_cnt - i);
2615#ifdef INVARIANTS
2616 bzero(&bucket->ub_bucket[i],
2617 sizeof(void *) * (bucket->ub_cnt - i));
2618#endif
2619 bucket->ub_cnt = i;
2620 }
2621 }
2622
2623 if (bucket->ub_cnt == 0) {
2624 bucket_free(zone, bucket, udata);
2625 atomic_add_long(&zone->uz_fails, 1);
2626 return (NULL);
2627 }
2628
2629 return (bucket);
2630}
2631
2632/*
2633 * Allocates a single item from a zone.
2634 *
2635 * Arguments
2636 * zone The zone to alloc for.
2637 * udata The data to be passed to the constructor.
2638 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2639 *
2640 * Returns
2641 * NULL if there is no memory and M_NOWAIT is set
2642 * An item if successful
2643 */
2644
2645static void *
2646zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2647{
2648 void *item;
2649
2650 item = NULL;
2651
2652#ifdef UMA_DEBUG_ALLOC
2653 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2654#endif
2655 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2656 goto fail;
2657 atomic_add_long(&zone->uz_allocs, 1);
2658
2659 /*
2660 * We have to call both the zone's init (not the keg's init)
2661 * and the zone's ctor. This is because the item is going from
2662 * a keg slab directly to the user, and the user is expecting it
2663 * to be both zone-init'd as well as zone-ctor'd.
2664 */
2665 if (zone->uz_init != NULL) {
2666 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2667 zone_free_item(zone, item, udata, SKIP_FINI);
2668 goto fail;
2669 }
2670 }
2671 if (zone->uz_ctor != NULL) {
2672 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2673 zone_free_item(zone, item, udata, SKIP_DTOR);
2674 goto fail;
2675 }
2676 }
2677#ifdef INVARIANTS
2678 uma_dbg_alloc(zone, NULL, item);
2679#endif
2680 if (flags & M_ZERO)
2681 uma_zero_item(item, zone);
2682
2683 return (item);
2684
2685fail:
2686 atomic_add_long(&zone->uz_fails, 1);
2687 return (NULL);
2688}
2689
2690/* See uma.h */
2691void
2692uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2693{
2694 uma_cache_t cache;
2695 uma_bucket_t bucket;
2696 int lockfail;
2697 int cpu;
2698
2699 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2700 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2701
2702#ifdef UMA_DEBUG_ALLOC_1
2703 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2704#endif
2705 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2706 zone->uz_name);
2707
2708 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2709 ("uma_zfree_arg: called with spinlock or critical section held"));
2710
2711 /* uma_zfree(..., NULL) does nothing, to match free(9). */
2712 if (item == NULL)
2713 return;
2714#ifdef DEBUG_MEMGUARD
2715 if (is_memguard_addr(item)) {
2716 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2717 zone->uz_dtor(item, zone->uz_size, udata);
2718 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2719 zone->uz_fini(item, zone->uz_size);
2720 memguard_free(item);
2721 return;
2722 }
2723#endif
2724#ifdef INVARIANTS
2725 if (zone->uz_flags & UMA_ZONE_MALLOC)
2726 uma_dbg_free(zone, udata, item);
2727 else
2728 uma_dbg_free(zone, NULL, item);
2729#endif
2730 if (zone->uz_dtor != NULL)
2731 zone->uz_dtor(item, zone->uz_size, udata);
2732
2733 /*
2734 * The race here is acceptable. If we miss it we'll just have to wait
2735 * a little longer for the limits to be reset.
2736 */
2737 if (zone->uz_flags & UMA_ZFLAG_FULL)
2738 goto zfree_item;
2739
2740 /*
2741 * If possible, free to the per-CPU cache. There are two
2742 * requirements for safe access to the per-CPU cache: (1) the thread
2743 * accessing the cache must not be preempted or yield during access,
2744 * and (2) the thread must not migrate CPUs without switching which
2745 * cache it accesses. We rely on a critical section to prevent
2746 * preemption and migration. We release the critical section in
2747 * order to acquire the zone mutex if we are unable to free to the
2748 * current cache; when we re-acquire the critical section, we must
2749 * detect and handle migration if it has occurred.
2750 */
2751zfree_restart:
2752 critical_enter();
2753 cpu = curcpu;
2754 cache = &zone->uz_cpu[cpu];
2755
2756zfree_start:
2757 /*
2758 * Try to free into the allocbucket first to give LIFO ordering
2759 * for cache-hot datastructures. Spill over into the freebucket
2760 * if necessary. Alloc will swap them if one runs dry.
2761 */
2762 bucket = cache->uc_allocbucket;
2763 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2764 bucket = cache->uc_freebucket;
2765 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2766 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2767 ("uma_zfree: Freeing to non free bucket index."));
2768 bucket->ub_bucket[bucket->ub_cnt] = item;
2769 bucket->ub_cnt++;
2770 cache->uc_frees++;
2771 critical_exit();
2772 return;
2773 }
2774
2775 /*
2776 * We must go back the zone, which requires acquiring the zone lock,
2777 * which in turn means we must release and re-acquire the critical
2778 * section. Since the critical section is released, we may be
2779 * preempted or migrate. As such, make sure not to maintain any
2780 * thread-local state specific to the cache from prior to releasing
2781 * the critical section.
2782 */
2783 critical_exit();
2784 if (zone->uz_count == 0 || bucketdisable)
2785 goto zfree_item;
2786
2787 lockfail = 0;
2788 if (ZONE_TRYLOCK(zone) == 0) {
2789 /* Record contention to size the buckets. */
2790 ZONE_LOCK(zone);
2791 lockfail = 1;
2792 }
2793 critical_enter();
2794 cpu = curcpu;
2795 cache = &zone->uz_cpu[cpu];
2796
2797 /*
2798 * Since we have locked the zone we may as well send back our stats.
2799 */
2800 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2801 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2802 cache->uc_allocs = 0;
2803 cache->uc_frees = 0;
2804
2805 bucket = cache->uc_freebucket;
2806 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2807 ZONE_UNLOCK(zone);
2808 goto zfree_start;
2809 }
2810 cache->uc_freebucket = NULL;
2811
2812 /* Can we throw this on the zone full list? */
2813 if (bucket != NULL) {
2814#ifdef UMA_DEBUG_ALLOC
2815 printf("uma_zfree: Putting old bucket on the free list.\n");
2816#endif
2817 /* ub_cnt is pointing to the last free item */
2818 KASSERT(bucket->ub_cnt != 0,
2819 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2820 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2821 }
2822
2823 /* We are no longer associated with this CPU. */
2824 critical_exit();
2825
2826 /*
2827 * We bump the uz count when the cache size is insufficient to
2828 * handle the working set.
2829 */
2830 if (lockfail && zone->uz_count < BUCKET_MAX)
2831 zone->uz_count++;
2832 ZONE_UNLOCK(zone);
2833
2834#ifdef UMA_DEBUG_ALLOC
2835 printf("uma_zfree: Allocating new free bucket.\n");
2836#endif
2837 bucket = bucket_alloc(zone, udata, M_NOWAIT);
2838 if (bucket) {
2839 critical_enter();
2840 cpu = curcpu;
2841 cache = &zone->uz_cpu[cpu];
2842 if (cache->uc_freebucket == NULL) {
2843 cache->uc_freebucket = bucket;
2844 goto zfree_start;
2845 }
2846 /*
2847 * We lost the race, start over. We have to drop our
2848 * critical section to free the bucket.
2849 */
2850 critical_exit();
2851 bucket_free(zone, bucket, udata);
2852 goto zfree_restart;
2853 }
2854
2855 /*
2856 * If nothing else caught this, we'll just do an internal free.
2857 */
2858zfree_item:
2859 zone_free_item(zone, item, udata, SKIP_DTOR);
2860
2861 return;
2862}
2863
2864static void
2865slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2866{
2867 uint8_t freei;
2868
2869 mtx_assert(&keg->uk_lock, MA_OWNED);
2870 MPASS(keg == slab->us_keg);
2871
2872 /* Do we need to remove from any lists? */
2873 if (slab->us_freecount+1 == keg->uk_ipers) {
2874 LIST_REMOVE(slab, us_link);
2875 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2876 } else if (slab->us_freecount == 0) {
2877 LIST_REMOVE(slab, us_link);
2878 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2879 }
2880
2881 /* Slab management. */
2882 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2883 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2884 slab->us_freecount++;
2885
2886 /* Keg statistics. */
2887 keg->uk_free++;
2888}
2889
2890static void
2891zone_release(uma_zone_t zone, void **bucket, int cnt)
2892{
2893 void *item;
2894 uma_slab_t slab;
2895 uma_keg_t keg;
2896 uint8_t *mem;
2897 int clearfull;
2898 int i;
2899
2900 clearfull = 0;
2901 keg = zone_first_keg(zone);
2902 KEG_LOCK(keg);
2903 for (i = 0; i < cnt; i++) {
2904 item = bucket[i];
2905 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2906 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2907 if (zone->uz_flags & UMA_ZONE_HASH) {
2908 slab = hash_sfind(&keg->uk_hash, mem);
2909 } else {
2910 mem += keg->uk_pgoff;
2911 slab = (uma_slab_t)mem;
2912 }
2913 } else {
2914 slab = vtoslab((vm_offset_t)item);
2915 if (slab->us_keg != keg) {
2916 KEG_UNLOCK(keg);
2917 keg = slab->us_keg;
2918 KEG_LOCK(keg);
2919 }
2920 }
2921 slab_free_item(keg, slab, item);
2922 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2923 if (keg->uk_pages < keg->uk_maxpages) {
2924 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2925 clearfull = 1;
2926 }
2927
2928 /*
2929 * We can handle one more allocation. Since we're
2930 * clearing ZFLAG_FULL, wake up all procs blocked
2931 * on pages. This should be uncommon, so keeping this
2932 * simple for now (rather than adding count of blocked
2933 * threads etc).
2934 */
2935 wakeup(keg);
2936 }
2937 }
2938 KEG_UNLOCK(keg);
2939 if (clearfull) {
2940 ZONE_LOCK(zone);
2941 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2942 wakeup(zone);
2943 ZONE_UNLOCK(zone);
2944 }
2945
2946}
2947
2948/*
2949 * Frees a single item to any zone.
2950 *
2951 * Arguments:
2952 * zone The zone to free to
2953 * item The item we're freeing
2954 * udata User supplied data for the dtor
2955 * skip Skip dtors and finis
2956 */
2957static void
2958zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2959{
2960
2961#ifdef INVARIANTS
2962 if (skip == SKIP_NONE) {
2963 if (zone->uz_flags & UMA_ZONE_MALLOC)
2964 uma_dbg_free(zone, udata, item);
2965 else
2966 uma_dbg_free(zone, NULL, item);
2967 }
2968#endif
2969 if (skip < SKIP_DTOR && zone->uz_dtor)
2970 zone->uz_dtor(item, zone->uz_size, udata);
2971
2972 if (skip < SKIP_FINI && zone->uz_fini)
2973 zone->uz_fini(item, zone->uz_size);
2974
2975 atomic_add_long(&zone->uz_frees, 1);
2976 zone->uz_release(zone->uz_arg, &item, 1);
2977}
2978
2979/* See uma.h */
2980int
2981uma_zone_set_max(uma_zone_t zone, int nitems)
2982{
2983 uma_keg_t keg;
2984
2985 keg = zone_first_keg(zone);
2986 if (keg == NULL)
2987 return (0);
2988 KEG_LOCK(keg);
2989 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2990 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2991 keg->uk_maxpages += keg->uk_ppera;
2992 nitems = keg->uk_maxpages * keg->uk_ipers;
2993 KEG_UNLOCK(keg);
2994
2995 return (nitems);
2996}
2997
2998/* See uma.h */
2999int
3000uma_zone_get_max(uma_zone_t zone)
3001{
3002 int nitems;
3003 uma_keg_t keg;
3004
3005 keg = zone_first_keg(zone);
3006 if (keg == NULL)
3007 return (0);
3008 KEG_LOCK(keg);
3009 nitems = keg->uk_maxpages * keg->uk_ipers;
3010 KEG_UNLOCK(keg);
3011
3012 return (nitems);
3013}
3014
3015/* See uma.h */
3016void
3017uma_zone_set_warning(uma_zone_t zone, const char *warning)
3018{
3019
3020 ZONE_LOCK(zone);
3021 zone->uz_warning = warning;
3022 ZONE_UNLOCK(zone);
3023}
3024
3025/* See uma.h */
3026void
3027uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
3028{
3029
3030 ZONE_LOCK(zone);
3031 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
3032 ZONE_UNLOCK(zone);
3033}
3034
3035/* See uma.h */
3036int
3037uma_zone_get_cur(uma_zone_t zone)
3038{
3039 int64_t nitems;
3040 u_int i;
3041
3042 ZONE_LOCK(zone);
3043 nitems = zone->uz_allocs - zone->uz_frees;
3044 CPU_FOREACH(i) {
3045 /*
3046 * See the comment in sysctl_vm_zone_stats() regarding the
3047 * safety of accessing the per-cpu caches. With the zone lock
3048 * held, it is safe, but can potentially result in stale data.
3049 */
3050 nitems += zone->uz_cpu[i].uc_allocs -
3051 zone->uz_cpu[i].uc_frees;
3052 }
3053 ZONE_UNLOCK(zone);
3054
3055 return (nitems < 0 ? 0 : nitems);
3056}
3057
3058/* See uma.h */
3059void
3060uma_zone_set_init(uma_zone_t zone, uma_init uminit)
3061{
3062 uma_keg_t keg;
3063
3064 keg = zone_first_keg(zone);
3065 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3066 KEG_LOCK(keg);
3067 KASSERT(keg->uk_pages == 0,
3068 ("uma_zone_set_init on non-empty keg"));
3069 keg->uk_init = uminit;
3070 KEG_UNLOCK(keg);
3071}
3072
3073/* See uma.h */
3074void
3075uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3076{
3077 uma_keg_t keg;
3078
3079 keg = zone_first_keg(zone);
3080 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
3081 KEG_LOCK(keg);
3082 KASSERT(keg->uk_pages == 0,
3083 ("uma_zone_set_fini on non-empty keg"));
3084 keg->uk_fini = fini;
3085 KEG_UNLOCK(keg);
3086}
3087
3088/* See uma.h */
3089void
3090uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3091{
3092
3093 ZONE_LOCK(zone);
3094 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3095 ("uma_zone_set_zinit on non-empty keg"));
3096 zone->uz_init = zinit;
3097 ZONE_UNLOCK(zone);
3098}
3099
3100/* See uma.h */
3101void
3102uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3103{
3104
3105 ZONE_LOCK(zone);
3106 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3107 ("uma_zone_set_zfini on non-empty keg"));
3108 zone->uz_fini = zfini;
3109 ZONE_UNLOCK(zone);
3110}
3111
3112/* See uma.h */
3113/* XXX uk_freef is not actually used with the zone locked */
3114void
3115uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3116{
3117 uma_keg_t keg;
3118
3119 keg = zone_first_keg(zone);
3120 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
3121 KEG_LOCK(keg);
3122 keg->uk_freef = freef;
3123 KEG_UNLOCK(keg);
3124}
3125
3126/* See uma.h */
3127/* XXX uk_allocf is not actually used with the zone locked */
3128void
3129uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3130{
3131 uma_keg_t keg;
3132
3133 keg = zone_first_keg(zone);
3134 KEG_LOCK(keg);
3135 keg->uk_allocf = allocf;
3136 KEG_UNLOCK(keg);
3137}
3138
3139/* See uma.h */
3140void
3141uma_zone_reserve(uma_zone_t zone, int items)
3142{
3143 uma_keg_t keg;
3144
3145 keg = zone_first_keg(zone);
3146 if (keg == NULL)
3147 return;
3148 KEG_LOCK(keg);
3149 keg->uk_reserve = items;
3150 KEG_UNLOCK(keg);
3151
3152 return;
3153}
3154
3155/* See uma.h */
3156int
3157uma_zone_reserve_kva(uma_zone_t zone, int count)
3158{
3159 uma_keg_t keg;
3160 vm_offset_t kva;
3161 u_int pages;
3162
3163 keg = zone_first_keg(zone);
3164 if (keg == NULL)
3165 return (0);
3166 pages = count / keg->uk_ipers;
3167
3168 if (pages * keg->uk_ipers < count)
3169 pages++;
3170
3171#ifdef UMA_MD_SMALL_ALLOC
3172 if (keg->uk_ppera > 1) {
3173#else
3174 if (1) {
3175#endif
3176 kva = kva_alloc((vm_size_t)pages * UMA_SLAB_SIZE);
3177 if (kva == 0)
3178 return (0);
3179 } else
3180 kva = 0;
3181 KEG_LOCK(keg);
3182 keg->uk_kva = kva;
3183 keg->uk_offset = 0;
3184 keg->uk_maxpages = pages;
3185#ifdef UMA_MD_SMALL_ALLOC
3186 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3187#else
3188 keg->uk_allocf = noobj_alloc;
3189#endif
3190 keg->uk_flags |= UMA_ZONE_NOFREE;
3191 KEG_UNLOCK(keg);
3192
3193 return (1);
3194}
3195
3196/* See uma.h */
3197void
3198uma_prealloc(uma_zone_t zone, int items)
3199{
3200 int slabs;
3201 uma_slab_t slab;
3202 uma_keg_t keg;
3203
3204 keg = zone_first_keg(zone);
3205 if (keg == NULL)
3206 return;
3207 KEG_LOCK(keg);
3208 slabs = items / keg->uk_ipers;
3209 if (slabs * keg->uk_ipers < items)
3210 slabs++;
3211 while (slabs > 0) {
3212 slab = keg_alloc_slab(keg, zone, M_WAITOK);
3213 if (slab == NULL)
3214 break;
3215 MPASS(slab->us_keg == keg);
3216 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3217 slabs--;
3218 }
3219 KEG_UNLOCK(keg);
3220}
3221
3222/* See uma.h */
3223uint32_t *
3224uma_find_refcnt(uma_zone_t zone, void *item)
3225{
3226 uma_slabrefcnt_t slabref;
3227 uma_slab_t slab;
3228 uma_keg_t keg;
3229 uint32_t *refcnt;
3230 int idx;
3231
3232 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3233 slabref = (uma_slabrefcnt_t)slab;
3234 keg = slab->us_keg;
3235 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3236 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3237 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3238 refcnt = &slabref->us_refcnt[idx];
3239 return refcnt;
3240}
3241
3242/* See uma.h */
3243static void
3244uma_reclaim_locked(bool kmem_danger)
3245{
3246
3247#ifdef UMA_DEBUG
3248 printf("UMA: vm asked us to release pages!\n");
3249#endif
3250 sx_assert(&uma_drain_lock, SA_XLOCKED);
3251 bucket_enable();
3252 zone_foreach(zone_drain);
3253 if (vm_page_count_min() || kmem_danger) {
3254 cache_drain_safe(NULL);
3255 zone_foreach(zone_drain);
3256 }
3257 /*
3258 * Some slabs may have been freed but this zone will be visited early
3259 * we visit again so that we can free pages that are empty once other
3260 * zones are drained. We have to do the same for buckets.
3261 */
3262 zone_drain(slabzone);
3263 zone_drain(slabrefzone);
3264 bucket_zone_drain();
3265}
3266
3267void
3268uma_reclaim(void)
3269{
3270
3271 sx_xlock(&uma_drain_lock);
3272 uma_reclaim_locked(false);
3273 sx_xunlock(&uma_drain_lock);
3274}
3275
3276static int uma_reclaim_needed;
3277
3278void
3279uma_reclaim_wakeup(void)
3280{
3281
3282 uma_reclaim_needed = 1;
3283 wakeup(&uma_reclaim_needed);
3284}
3285
3286void
3287uma_reclaim_worker(void *arg __unused)
3288{
3289
3290 sx_xlock(&uma_drain_lock);
3291 for (;;) {
3292 sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM,
3293 "umarcl", 0);
3294 if (uma_reclaim_needed) {
3295 uma_reclaim_needed = 0;
3296 uma_reclaim_locked(true);
3297 }
3298 }
3299}
3300
3301/* See uma.h */
3302int
3303uma_zone_exhausted(uma_zone_t zone)
3304{
3305 int full;
3306
3307 ZONE_LOCK(zone);
3308 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3309 ZONE_UNLOCK(zone);
3310 return (full);
3311}
3312
3313int
3314uma_zone_exhausted_nolock(uma_zone_t zone)
3315{
3316 return (zone->uz_flags & UMA_ZFLAG_FULL);
3317}
3318
3319void *
3320uma_large_malloc(vm_size_t size, int wait)
3321{
3322 void *mem;
3323 uma_slab_t slab;
3324 uint8_t flags;
3325
3326 slab = zone_alloc_item(slabzone, NULL, wait);
3327 if (slab == NULL)
3328 return (NULL);
3329 mem = page_alloc(NULL, size, &flags, wait);
3330 if (mem) {
3331 vsetslab((vm_offset_t)mem, slab);
3332 slab->us_data = mem;
3333 slab->us_flags = flags | UMA_SLAB_MALLOC;
3334 slab->us_size = size;
3335 } else {
3336 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3337 }
3338
3339 return (mem);
3340}
3341
3342void
3343uma_large_free(uma_slab_t slab)
3344{
3345
3346 page_free(slab->us_data, slab->us_size, slab->us_flags);
3347 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3348}
3349
3350static void
3351uma_zero_item(void *item, uma_zone_t zone)
3352{
3353
3354 if (zone->uz_flags & UMA_ZONE_PCPU) {
3355 for (int i = 0; i < mp_ncpus; i++)
3356 bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3357 } else
3358 bzero(item, zone->uz_size);
3359}
3360
3361void
3362uma_print_stats(void)
3363{
3364 zone_foreach(uma_print_zone);
3365}
3366
3367static void
3368slab_print(uma_slab_t slab)
3369{
3370 printf("slab: keg %p, data %p, freecount %d\n",
3371 slab->us_keg, slab->us_data, slab->us_freecount);
3372}
3373
3374static void
3375cache_print(uma_cache_t cache)
3376{
3377 printf("alloc: %p(%d), free: %p(%d)\n",
3378 cache->uc_allocbucket,
3379 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3380 cache->uc_freebucket,
3381 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3382}
3383
3384static void
3385uma_print_keg(uma_keg_t keg)
3386{
3387 uma_slab_t slab;
3388
3389 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3390 "out %d free %d limit %d\n",
3391 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3392 keg->uk_ipers, keg->uk_ppera,
3393 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3394 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3395 printf("Part slabs:\n");
3396 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3397 slab_print(slab);
3398 printf("Free slabs:\n");
3399 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3400 slab_print(slab);
3401 printf("Full slabs:\n");
3402 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3403 slab_print(slab);
3404}
3405
3406void
3407uma_print_zone(uma_zone_t zone)
3408{
3409 uma_cache_t cache;
3410 uma_klink_t kl;
3411 int i;
3412
3413 printf("zone: %s(%p) size %d flags %#x\n",
3414 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3415 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3416 uma_print_keg(kl->kl_keg);
3417 CPU_FOREACH(i) {
3418 cache = &zone->uz_cpu[i];
3419 printf("CPU %d Cache:\n", i);
3420 cache_print(cache);
3421 }
3422}
3423
3424#ifdef DDB
3425/*
3426 * Generate statistics across both the zone and its per-cpu cache's. Return
3427 * desired statistics if the pointer is non-NULL for that statistic.
3428 *
3429 * Note: does not update the zone statistics, as it can't safely clear the
3430 * per-CPU cache statistic.
3431 *
3432 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3433 * safe from off-CPU; we should modify the caches to track this information
3434 * directly so that we don't have to.
3435 */
3436static void
3437uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3438 uint64_t *freesp, uint64_t *sleepsp)
3439{
3440 uma_cache_t cache;
3441 uint64_t allocs, frees, sleeps;
3442 int cachefree, cpu;
3443
3444 allocs = frees = sleeps = 0;
3445 cachefree = 0;
3446 CPU_FOREACH(cpu) {
3447 cache = &z->uz_cpu[cpu];
3448 if (cache->uc_allocbucket != NULL)
3449 cachefree += cache->uc_allocbucket->ub_cnt;
3450 if (cache->uc_freebucket != NULL)
3451 cachefree += cache->uc_freebucket->ub_cnt;
3452 allocs += cache->uc_allocs;
3453 frees += cache->uc_frees;
3454 }
3455 allocs += z->uz_allocs;
3456 frees += z->uz_frees;
3457 sleeps += z->uz_sleeps;
3458 if (cachefreep != NULL)
3459 *cachefreep = cachefree;
3460 if (allocsp != NULL)
3461 *allocsp = allocs;
3462 if (freesp != NULL)
3463 *freesp = frees;
3464 if (sleepsp != NULL)
3465 *sleepsp = sleeps;
3466}
3467#endif /* DDB */
3468
3469static int
3470sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3471{
3472 uma_keg_t kz;
3473 uma_zone_t z;
3474 int count;
3475
3476 count = 0;
3477 rw_rlock(&uma_rwlock);
3478 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3479 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3480 count++;
3481 }
3482 rw_runlock(&uma_rwlock);
3483 return (sysctl_handle_int(oidp, &count, 0, req));
3484}
3485
3486static int
3487sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3488{
3489 struct uma_stream_header ush;
3490 struct uma_type_header uth;
3491 struct uma_percpu_stat ups;
3492 uma_bucket_t bucket;
3493 struct sbuf sbuf;
3494 uma_cache_t cache;
3495 uma_klink_t kl;
3496 uma_keg_t kz;
3497 uma_zone_t z;
3498 uma_keg_t k;
3499 int count, error, i;
3500
3501 error = sysctl_wire_old_buffer(req, 0);
3502 if (error != 0)
3503 return (error);
3504 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3505 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3506
3507 count = 0;
3508 rw_rlock(&uma_rwlock);
3509 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3510 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3511 count++;
3512 }
3513
3514 /*
3515 * Insert stream header.
3516 */
3517 bzero(&ush, sizeof(ush));
3518 ush.ush_version = UMA_STREAM_VERSION;
3519 ush.ush_maxcpus = (mp_maxid + 1);
3520 ush.ush_count = count;
3521 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3522
3523 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3524 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3525 bzero(&uth, sizeof(uth));
3526 ZONE_LOCK(z);
3527 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3528 uth.uth_align = kz->uk_align;
3529 uth.uth_size = kz->uk_size;
3530 uth.uth_rsize = kz->uk_rsize;
3531 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3532 k = kl->kl_keg;
3533 uth.uth_maxpages += k->uk_maxpages;
3534 uth.uth_pages += k->uk_pages;
3535 uth.uth_keg_free += k->uk_free;
3536 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3537 * k->uk_ipers;
3538 }
3539
3540 /*
3541 * A zone is secondary is it is not the first entry
3542 * on the keg's zone list.
3543 */
3544 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3545 (LIST_FIRST(&kz->uk_zones) != z))
3546 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3547
3548 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3549 uth.uth_zone_free += bucket->ub_cnt;
3550 uth.uth_allocs = z->uz_allocs;
3551 uth.uth_frees = z->uz_frees;
3552 uth.uth_fails = z->uz_fails;
3553 uth.uth_sleeps = z->uz_sleeps;
3554 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3555 /*
3556 * While it is not normally safe to access the cache
3557 * bucket pointers while not on the CPU that owns the
3558 * cache, we only allow the pointers to be exchanged
3559 * without the zone lock held, not invalidated, so
3560 * accept the possible race associated with bucket
3561 * exchange during monitoring.
3562 */
3563 for (i = 0; i < (mp_maxid + 1); i++) {
3564 bzero(&ups, sizeof(ups));
3565 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3566 goto skip;
3567 if (CPU_ABSENT(i))
3568 goto skip;
3569 cache = &z->uz_cpu[i];
3570 if (cache->uc_allocbucket != NULL)
3571 ups.ups_cache_free +=
3572 cache->uc_allocbucket->ub_cnt;
3573 if (cache->uc_freebucket != NULL)
3574 ups.ups_cache_free +=
3575 cache->uc_freebucket->ub_cnt;
3576 ups.ups_allocs = cache->uc_allocs;
3577 ups.ups_frees = cache->uc_frees;
3578skip:
3579 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3580 }
3581 ZONE_UNLOCK(z);
3582 }
3583 }
3584 rw_runlock(&uma_rwlock);
3585 error = sbuf_finish(&sbuf);
3586 sbuf_delete(&sbuf);
3587 return (error);
3588}
3589
3590int
3591sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3592{
3593 uma_zone_t zone = *(uma_zone_t *)arg1;
3594 int error, max;
3595
3596 max = uma_zone_get_max(zone);
3597 error = sysctl_handle_int(oidp, &max, 0, req);
3598 if (error || !req->newptr)
3599 return (error);
3600
3601 uma_zone_set_max(zone, max);
3602
3603 return (0);
3604}
3605
3606int
3607sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3608{
3609 uma_zone_t zone = *(uma_zone_t *)arg1;
3610 int cur;
3611
3612 cur = uma_zone_get_cur(zone);
3613 return (sysctl_handle_int(oidp, &cur, 0, req));
3614}
3615
3616#ifdef INVARIANTS
3617static uma_slab_t
3618uma_dbg_getslab(uma_zone_t zone, void *item)
3619{
3620 uma_slab_t slab;
3621 uma_keg_t keg;
3622 uint8_t *mem;
3623
3624 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
3625 if (zone->uz_flags & UMA_ZONE_VTOSLAB) {
3626 slab = vtoslab((vm_offset_t)mem);
3627 } else {
3628 /*
3629 * It is safe to return the slab here even though the
3630 * zone is unlocked because the item's allocation state
3631 * essentially holds a reference.
3632 */
3633 ZONE_LOCK(zone);
3634 keg = LIST_FIRST(&zone->uz_kegs)->kl_keg;
3635 if (keg->uk_flags & UMA_ZONE_HASH)
3636 slab = hash_sfind(&keg->uk_hash, mem);
3637 else
3638 slab = (uma_slab_t)(mem + keg->uk_pgoff);
3639 ZONE_UNLOCK(zone);
3640 }
3641
3642 return (slab);
3643}
3644
3645/*
3646 * Set up the slab's freei data such that uma_dbg_free can function.
3647 *
3648 */
3649static void
3650uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
3651{
3652 uma_keg_t keg;
3653 int freei;
3654
3655 if (zone_first_keg(zone) == NULL)
3656 return;
3657 if (slab == NULL) {
3658 slab = uma_dbg_getslab(zone, item);
3659 if (slab == NULL)
3660 panic("uma: item %p did not belong to zone %s\n",
3661 item, zone->uz_name);
3662 }
3663 keg = slab->us_keg;
3664 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3665
3666 if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3667 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
3668 item, zone, zone->uz_name, slab, freei);
3669 BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3670
3671 return;
3672}
3673
3674/*
3675 * Verifies freed addresses. Checks for alignment, valid slab membership
3676 * and duplicate frees.
3677 *
3678 */
3679static void
3680uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
3681{
3682 uma_keg_t keg;
3683 int freei;
3684
3685 if (zone_first_keg(zone) == NULL)
3686 return;
3687 if (slab == NULL) {
3688 slab = uma_dbg_getslab(zone, item);
3689 if (slab == NULL)
3690 panic("uma: Freed item %p did not belong to zone %s\n",
3691 item, zone->uz_name);
3692 }
3693 keg = slab->us_keg;
3694 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3695
3696 if (freei >= keg->uk_ipers)
3697 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
3698 item, zone, zone->uz_name, slab, freei);
3699
3700 if (((freei * keg->uk_rsize) + slab->us_data) != item)
3701 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
3702 item, zone, zone->uz_name, slab, freei);
3703
3704 if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3705 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
3706 item, zone, zone->uz_name, slab, freei);
3707
3708 BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3709}
3710#endif /* INVARIANTS */
3711
3712#ifdef DDB
3713DB_SHOW_COMMAND(uma, db_show_uma)
3714{
3715 uint64_t allocs, frees, sleeps;
3716 uma_bucket_t bucket;
3717 uma_keg_t kz;
3718 uma_zone_t z;
3719 int cachefree;
3720
3721 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3722 "Free", "Requests", "Sleeps", "Bucket");
3723 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3724 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3725 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3726 allocs = z->uz_allocs;
3727 frees = z->uz_frees;
3728 sleeps = z->uz_sleeps;
3729 cachefree = 0;
3730 } else
3731 uma_zone_sumstat(z, &cachefree, &allocs,
3732 &frees, &sleeps);
3733 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3734 (LIST_FIRST(&kz->uk_zones) != z)))
3735 cachefree += kz->uk_free;
3736 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3737 cachefree += bucket->ub_cnt;
3738 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3739 z->uz_name, (uintmax_t)kz->uk_size,
3740 (intmax_t)(allocs - frees), cachefree,
3741 (uintmax_t)allocs, sleeps, z->uz_count);
3742 if (db_pager_quit)
3743 return;
3744 }
3745 }
3746}
3747
3748DB_SHOW_COMMAND(umacache, db_show_umacache)
3749{
3750 uint64_t allocs, frees;
3751 uma_bucket_t bucket;
3752 uma_zone_t z;
3753 int cachefree;
3754
3755 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3756 "Requests", "Bucket");
3757 LIST_FOREACH(z, &uma_cachezones, uz_link) {
3758 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3759 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3760 cachefree += bucket->ub_cnt;
3761 db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3762 z->uz_name, (uintmax_t)z->uz_size,
3763 (intmax_t)(allocs - frees), cachefree,
3764 (uintmax_t)allocs, z->uz_count);
3765 if (db_pager_quit)
3766 return;
3767 }
3768}
3769#endif /* DDB */