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
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice unmodified, this list of conditions, and the following
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.
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.
30 * uma_core.c Implementation of the Universal Memory allocator
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.
39 * The basic ideas stem from similar slab/zone based allocators whose algorithms
46 * - Improve memory usage for large allocations
47 * - Investigate cache size adjustments
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
53 /* I should really use ktr.. */
56 #define UMA_DEBUG_ALLOC 1
57 #define UMA_DEBUG_ALLOC_1 1
61 #include "opt_param.h"
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>
73 #include <sys/sysctl.h>
74 #include <sys/mutex.h>
76 #include <sys/rwlock.h>
78 #include <sys/sched.h>
80 #include <sys/vmmeter.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_param.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_extern.h>
91 #include <vm/uma_int.h>
92 #include <vm/uma_dbg.h>
97 #include <vm/memguard.h>
101 * This is the zone and keg from which all zones are spawned. The idea is that
102 * even the zone & keg heads are allocated from the allocator, so we use the
103 * bss section to bootstrap us.
105 static struct uma_keg masterkeg;
106 static struct uma_zone masterzone_k;
107 static struct uma_zone masterzone_z;
108 static uma_zone_t kegs = &masterzone_k;
109 static uma_zone_t zones = &masterzone_z;
111 /* This is the zone from which all of uma_slab_t's are allocated. */
112 static uma_zone_t slabzone;
113 static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
116 * The initial hash tables come out of this zone so they can be allocated
117 * prior to malloc coming up.
119 static uma_zone_t hashzone;
121 /* The boot-time adjusted value for cache line alignment. */
122 int uma_align_cache = 64 - 1;
124 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
127 * Are we allowed to allocate buckets?
129 static int bucketdisable = 1;
131 /* Linked list of all kegs in the system */
132 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
134 /* This mutex protects the keg list */
135 static struct mtx_padalign uma_mtx;
137 /* Linked list of boot time pages */
138 static LIST_HEAD(,uma_slab) uma_boot_pages =
139 LIST_HEAD_INITIALIZER(uma_boot_pages);
141 /* This mutex protects the boot time pages list */
142 static struct mtx_padalign uma_boot_pages_mtx;
144 /* Is the VM done starting up? */
145 static int booted = 0;
146 #define UMA_STARTUP 1
147 #define UMA_STARTUP2 2
149 /* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
150 static const u_int uma_max_ipers = SLAB_SETSIZE;
153 * Only mbuf clusters use ref zones. Just provide enough references
154 * to support the one user. New code should not use the ref facility.
156 static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES;
159 * This is the handle used to schedule events that need to happen
160 * outside of the allocation fast path.
162 static struct callout uma_callout;
163 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
166 * This structure is passed as the zone ctor arg so that I don't have to create
167 * a special allocation function just for zones.
169 struct uma_zctor_args {
184 struct uma_kctor_args {
193 struct uma_bucket_zone {
196 int ubz_entries; /* Number of items it can hold. */
197 int ubz_maxsize; /* Maximum allocation size per-item. */
201 * Compute the actual number of bucket entries to pack them in power
202 * of two sizes for more efficient space utilization.
204 #define BUCKET_SIZE(n) \
205 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
207 #define BUCKET_MAX BUCKET_SIZE(128)
209 struct uma_bucket_zone bucket_zones[] = {
210 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
211 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
212 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
213 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
214 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
215 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
216 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
217 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
222 * Flags and enumerations to be passed to internal functions.
224 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
228 static void *noobj_alloc(uma_zone_t, int, uint8_t *, int);
229 static void *page_alloc(uma_zone_t, int, uint8_t *, int);
230 static void *startup_alloc(uma_zone_t, int, uint8_t *, int);
231 static void page_free(void *, int, uint8_t);
232 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
233 static void cache_drain(uma_zone_t);
234 static void bucket_drain(uma_zone_t, uma_bucket_t);
235 static void bucket_cache_drain(uma_zone_t zone);
236 static int keg_ctor(void *, int, void *, int);
237 static void keg_dtor(void *, int, void *);
238 static int zone_ctor(void *, int, void *, int);
239 static void zone_dtor(void *, int, void *);
240 static int zero_init(void *, int, int);
241 static void keg_small_init(uma_keg_t keg);
242 static void keg_large_init(uma_keg_t keg);
243 static void zone_foreach(void (*zfunc)(uma_zone_t));
244 static void zone_timeout(uma_zone_t zone);
245 static int hash_alloc(struct uma_hash *);
246 static int hash_expand(struct uma_hash *, struct uma_hash *);
247 static void hash_free(struct uma_hash *hash);
248 static void uma_timeout(void *);
249 static void uma_startup3(void);
250 static void *zone_alloc_item(uma_zone_t, void *, int);
251 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
252 static void bucket_enable(void);
253 static void bucket_init(void);
254 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
255 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
256 static void bucket_zone_drain(void);
257 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
258 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
259 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
260 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
261 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
262 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
263 uma_fini fini, int align, uint32_t flags);
264 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
265 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
267 void uma_print_zone(uma_zone_t);
268 void uma_print_stats(void);
269 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
270 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
272 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
274 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
275 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
277 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
278 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
280 static int zone_warnings = 1;
281 TUNABLE_INT("vm.zone_warnings", &zone_warnings);
282 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RW, &zone_warnings, 0,
283 "Warn when UMA zones becomes full");
286 * This routine checks to see whether or not it's safe to enable buckets.
291 bucketdisable = vm_page_count_min();
295 * Initialize bucket_zones, the array of zones of buckets of various sizes.
297 * For each zone, calculate the memory required for each bucket, consisting
298 * of the header and an array of pointers.
303 struct uma_bucket_zone *ubz;
307 for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
308 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
309 size += sizeof(void *) * ubz->ubz_entries;
310 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
311 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
312 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
317 * Given a desired number of entries for a bucket, return the zone from which
318 * to allocate the bucket.
320 static struct uma_bucket_zone *
321 bucket_zone_lookup(int entries)
323 struct uma_bucket_zone *ubz;
325 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
326 if (ubz->ubz_entries >= entries)
333 bucket_select(int size)
335 struct uma_bucket_zone *ubz;
337 ubz = &bucket_zones[0];
338 if (size > ubz->ubz_maxsize)
339 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
341 for (; ubz->ubz_entries != 0; ubz++)
342 if (ubz->ubz_maxsize < size)
345 return (ubz->ubz_entries);
349 bucket_alloc(uma_zone_t zone, void *udata, int flags)
351 struct uma_bucket_zone *ubz;
355 * This is to stop us from allocating per cpu buckets while we're
356 * running out of vm.boot_pages. Otherwise, we would exhaust the
357 * boot pages. This also prevents us from allocating buckets in
358 * low memory situations.
363 * To limit bucket recursion we store the original zone flags
364 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
365 * NOVM flag to persist even through deep recursions. We also
366 * store ZFLAG_BUCKET once we have recursed attempting to allocate
367 * a bucket for a bucket zone so we do not allow infinite bucket
368 * recursion. This cookie will even persist to frees of unused
369 * buckets via the allocation path or bucket allocations in the
372 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
373 udata = (void *)(uintptr_t)zone->uz_flags;
375 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
377 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
379 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
381 ubz = bucket_zone_lookup(zone->uz_count);
382 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
385 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
388 bucket->ub_entries = ubz->ubz_entries;
395 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
397 struct uma_bucket_zone *ubz;
399 KASSERT(bucket->ub_cnt == 0,
400 ("bucket_free: Freeing a non free bucket."));
401 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
402 udata = (void *)(uintptr_t)zone->uz_flags;
403 ubz = bucket_zone_lookup(bucket->ub_entries);
404 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
408 bucket_zone_drain(void)
410 struct uma_bucket_zone *ubz;
412 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
413 zone_drain(ubz->ubz_zone);
417 zone_log_warning(uma_zone_t zone)
419 static const struct timeval warninterval = { 300, 0 };
421 if (!zone_warnings || zone->uz_warning == NULL)
424 if (ratecheck(&zone->uz_ratecheck, &warninterval))
425 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
429 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
433 LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
434 kegfn(klink->kl_keg);
438 * Routine called by timeout which is used to fire off some time interval
439 * based calculations. (stats, hash size, etc.)
448 uma_timeout(void *unused)
451 zone_foreach(zone_timeout);
453 /* Reschedule this event */
454 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
458 * Routine to perform timeout driven calculations. This expands the
459 * hashes and does per cpu statistics aggregation.
464 keg_timeout(uma_keg_t keg)
469 * Expand the keg hash table.
471 * This is done if the number of slabs is larger than the hash size.
472 * What I'm trying to do here is completely reduce collisions. This
473 * may be a little aggressive. Should I allow for two collisions max?
475 if (keg->uk_flags & UMA_ZONE_HASH &&
476 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
477 struct uma_hash newhash;
478 struct uma_hash oldhash;
482 * This is so involved because allocating and freeing
483 * while the keg lock is held will lead to deadlock.
484 * I have to do everything in stages and check for
487 newhash = keg->uk_hash;
489 ret = hash_alloc(&newhash);
492 if (hash_expand(&keg->uk_hash, &newhash)) {
493 oldhash = keg->uk_hash;
494 keg->uk_hash = newhash;
507 zone_timeout(uma_zone_t zone)
510 zone_foreach_keg(zone, &keg_timeout);
514 * Allocate and zero fill the next sized hash table from the appropriate
518 * hash A new hash structure with the old hash size in uh_hashsize
521 * 1 on sucess and 0 on failure.
524 hash_alloc(struct uma_hash *hash)
529 oldsize = hash->uh_hashsize;
531 /* We're just going to go to a power of two greater */
533 hash->uh_hashsize = oldsize * 2;
534 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
535 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
536 M_UMAHASH, M_NOWAIT);
538 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
539 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
541 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
543 if (hash->uh_slab_hash) {
544 bzero(hash->uh_slab_hash, alloc);
545 hash->uh_hashmask = hash->uh_hashsize - 1;
553 * Expands the hash table for HASH zones. This is done from zone_timeout
554 * to reduce collisions. This must not be done in the regular allocation
555 * path, otherwise, we can recurse on the vm while allocating pages.
558 * oldhash The hash you want to expand
559 * newhash The hash structure for the new table
567 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
573 if (!newhash->uh_slab_hash)
576 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
580 * I need to investigate hash algorithms for resizing without a
584 for (i = 0; i < oldhash->uh_hashsize; i++)
585 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
586 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
587 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
588 hval = UMA_HASH(newhash, slab->us_data);
589 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
597 * Free the hash bucket to the appropriate backing store.
600 * slab_hash The hash bucket we're freeing
601 * hashsize The number of entries in that hash bucket
607 hash_free(struct uma_hash *hash)
609 if (hash->uh_slab_hash == NULL)
611 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
612 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
614 free(hash->uh_slab_hash, M_UMAHASH);
618 * Frees all outstanding items in a bucket
621 * zone The zone to free to, must be unlocked.
622 * bucket The free/alloc bucket with items, cpu queue must be locked.
629 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
637 for (i = 0; i < bucket->ub_cnt; i++)
638 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
639 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
644 * Drains the per cpu caches for a zone.
646 * NOTE: This may only be called while the zone is being turn down, and not
647 * during normal operation. This is necessary in order that we do not have
648 * to migrate CPUs to drain the per-CPU caches.
651 * zone The zone to drain, must be unlocked.
657 cache_drain(uma_zone_t zone)
663 * XXX: It is safe to not lock the per-CPU caches, because we're
664 * tearing down the zone anyway. I.e., there will be no further use
665 * of the caches at this point.
667 * XXX: It would good to be able to assert that the zone is being
668 * torn down to prevent improper use of cache_drain().
670 * XXX: We lock the zone before passing into bucket_cache_drain() as
671 * it is used elsewhere. Should the tear-down path be made special
672 * there in some form?
675 cache = &zone->uz_cpu[cpu];
676 bucket_drain(zone, cache->uc_allocbucket);
677 bucket_drain(zone, cache->uc_freebucket);
678 if (cache->uc_allocbucket != NULL)
679 bucket_free(zone, cache->uc_allocbucket, NULL);
680 if (cache->uc_freebucket != NULL)
681 bucket_free(zone, cache->uc_freebucket, NULL);
682 cache->uc_allocbucket = cache->uc_freebucket = NULL;
685 bucket_cache_drain(zone);
690 cache_shrink(uma_zone_t zone)
693 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
697 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
702 cache_drain_safe_cpu(uma_zone_t zone)
707 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
713 cache = &zone->uz_cpu[curcpu];
714 if (cache->uc_allocbucket) {
715 if (cache->uc_allocbucket->ub_cnt != 0)
716 LIST_INSERT_HEAD(&zone->uz_buckets,
717 cache->uc_allocbucket, ub_link);
719 b1 = cache->uc_allocbucket;
720 cache->uc_allocbucket = NULL;
722 if (cache->uc_freebucket) {
723 if (cache->uc_freebucket->ub_cnt != 0)
724 LIST_INSERT_HEAD(&zone->uz_buckets,
725 cache->uc_freebucket, ub_link);
727 b2 = cache->uc_freebucket;
728 cache->uc_freebucket = NULL;
733 bucket_free(zone, b1, NULL);
735 bucket_free(zone, b2, NULL);
739 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
740 * This is an expensive call because it needs to bind to all CPUs
741 * one by one and enter a critical section on each of them in order
742 * to safely access their cache buckets.
743 * Zone lock must not be held on call this function.
746 cache_drain_safe(uma_zone_t zone)
751 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
756 zone_foreach(cache_shrink);
759 thread_lock(curthread);
760 sched_bind(curthread, cpu);
761 thread_unlock(curthread);
764 cache_drain_safe_cpu(zone);
766 zone_foreach(cache_drain_safe_cpu);
768 thread_lock(curthread);
769 sched_unbind(curthread);
770 thread_unlock(curthread);
774 * Drain the cached buckets from a zone. Expects a locked zone on entry.
777 bucket_cache_drain(uma_zone_t zone)
782 * Drain the bucket queues and free the buckets, we just keep two per
785 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
786 LIST_REMOVE(bucket, ub_link);
788 bucket_drain(zone, bucket);
789 bucket_free(zone, bucket, NULL);
794 * Shrink further bucket sizes. Price of single zone lock collision
795 * is probably lower then price of global cache drain.
797 if (zone->uz_count > zone->uz_count_min)
802 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
809 flags = slab->us_flags;
811 if (keg->uk_fini != NULL) {
812 for (i--; i > -1; i--)
813 keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
816 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
817 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
819 printf("%s: Returning %d bytes.\n", keg->uk_name,
820 PAGE_SIZE * keg->uk_ppera);
822 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
826 * Frees pages from a keg back to the system. This is done on demand from
827 * the pageout daemon.
832 keg_drain(uma_keg_t keg)
834 struct slabhead freeslabs = { 0 };
839 * We don't want to take pages from statically allocated kegs at this
842 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
846 printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
849 if (keg->uk_free == 0)
852 slab = LIST_FIRST(&keg->uk_free_slab);
854 n = LIST_NEXT(slab, us_link);
856 /* We have no where to free these to */
857 if (slab->us_flags & UMA_SLAB_BOOT) {
862 LIST_REMOVE(slab, us_link);
863 keg->uk_pages -= keg->uk_ppera;
864 keg->uk_free -= keg->uk_ipers;
866 if (keg->uk_flags & UMA_ZONE_HASH)
867 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
869 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
876 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
877 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
878 keg_free_slab(keg, slab, keg->uk_ipers);
883 zone_drain_wait(uma_zone_t zone, int waitok)
887 * Set draining to interlock with zone_dtor() so we can release our
888 * locks as we go. Only dtor() should do a WAITOK call since it
889 * is the only call that knows the structure will still be available
893 while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
894 if (waitok == M_NOWAIT)
896 mtx_unlock(&uma_mtx);
897 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
900 zone->uz_flags |= UMA_ZFLAG_DRAINING;
901 bucket_cache_drain(zone);
904 * The DRAINING flag protects us from being freed while
905 * we're running. Normally the uma_mtx would protect us but we
906 * must be able to release and acquire the right lock for each keg.
908 zone_foreach_keg(zone, &keg_drain);
910 zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
917 zone_drain(uma_zone_t zone)
920 zone_drain_wait(zone, M_NOWAIT);
924 * Allocate a new slab for a keg. This does not insert the slab onto a list.
927 * wait Shall we wait?
930 * The slab that was allocated or NULL if there is no memory and the
931 * caller specified M_NOWAIT.
934 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
936 uma_slabrefcnt_t slabref;
943 mtx_assert(&keg->uk_lock, MA_OWNED);
948 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name);
950 allocf = keg->uk_allocf;
953 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
954 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
960 * This reproduces the old vm_zone behavior of zero filling pages the
961 * first time they are added to a zone.
963 * Malloced items are zeroed in uma_zalloc.
966 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
971 if (keg->uk_flags & UMA_ZONE_NODUMP)
974 /* zone is passed for legacy reasons. */
975 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
977 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
978 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
983 /* Point the slab into the allocated memory */
984 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
985 slab = (uma_slab_t )(mem + keg->uk_pgoff);
987 if (keg->uk_flags & UMA_ZONE_VTOSLAB)
988 for (i = 0; i < keg->uk_ppera; i++)
989 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
993 slab->us_freecount = keg->uk_ipers;
994 slab->us_flags = flags;
995 BIT_FILL(SLAB_SETSIZE, &slab->us_free);
997 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
999 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1000 slabref = (uma_slabrefcnt_t)slab;
1001 for (i = 0; i < keg->uk_ipers; i++)
1002 slabref->us_refcnt[i] = 0;
1005 if (keg->uk_init != NULL) {
1006 for (i = 0; i < keg->uk_ipers; i++)
1007 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1008 keg->uk_size, wait) != 0)
1010 if (i != keg->uk_ipers) {
1011 keg_free_slab(keg, slab, i);
1020 if (keg->uk_flags & UMA_ZONE_HASH)
1021 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1023 keg->uk_pages += keg->uk_ppera;
1024 keg->uk_free += keg->uk_ipers;
1031 * This function is intended to be used early on in place of page_alloc() so
1032 * that we may use the boot time page cache to satisfy allocations before
1036 startup_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
1040 int pages, check_pages;
1042 keg = zone_first_keg(zone);
1043 pages = howmany(bytes, PAGE_SIZE);
1044 check_pages = pages - 1;
1045 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1048 * Check our small startup cache to see if it has pages remaining.
1050 mtx_lock(&uma_boot_pages_mtx);
1052 /* First check if we have enough room. */
1053 tmps = LIST_FIRST(&uma_boot_pages);
1054 while (tmps != NULL && check_pages-- > 0)
1055 tmps = LIST_NEXT(tmps, us_link);
1058 * It's ok to lose tmps references. The last one will
1059 * have tmps->us_data pointing to the start address of
1060 * "pages" contiguous pages of memory.
1062 while (pages-- > 0) {
1063 tmps = LIST_FIRST(&uma_boot_pages);
1064 LIST_REMOVE(tmps, us_link);
1066 mtx_unlock(&uma_boot_pages_mtx);
1067 *pflag = tmps->us_flags;
1068 return (tmps->us_data);
1070 mtx_unlock(&uma_boot_pages_mtx);
1071 if (booted < UMA_STARTUP2)
1072 panic("UMA: Increase vm.boot_pages");
1074 * Now that we've booted reset these users to their real allocator.
1076 #ifdef UMA_MD_SMALL_ALLOC
1077 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1079 keg->uk_allocf = page_alloc;
1081 return keg->uk_allocf(zone, bytes, pflag, wait);
1085 * Allocates a number of pages from the system
1088 * bytes The number of bytes requested
1089 * wait Shall we wait?
1092 * A pointer to the alloced memory or possibly
1093 * NULL if M_NOWAIT is set.
1096 page_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
1098 void *p; /* Returned page */
1100 *pflag = UMA_SLAB_KMEM;
1101 p = (void *) kmem_malloc(kmem_arena, bytes, wait);
1107 * Allocates a number of pages from within an object
1110 * bytes The number of bytes requested
1111 * wait Shall we wait?
1114 * A pointer to the alloced memory or possibly
1115 * NULL if M_NOWAIT is set.
1118 noobj_alloc(uma_zone_t zone, int bytes, uint8_t *flags, int wait)
1120 TAILQ_HEAD(, vm_page) alloctail;
1122 vm_offset_t retkva, zkva;
1123 vm_page_t p, p_next;
1126 TAILQ_INIT(&alloctail);
1127 keg = zone_first_keg(zone);
1129 npages = howmany(bytes, PAGE_SIZE);
1130 while (npages > 0) {
1131 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1132 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1135 * Since the page does not belong to an object, its
1138 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1142 if (wait & M_WAITOK) {
1148 * Page allocation failed, free intermediate pages and
1151 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1152 vm_page_unwire(p, 0);
1157 *flags = UMA_SLAB_PRIV;
1158 zkva = keg->uk_kva +
1159 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1161 TAILQ_FOREACH(p, &alloctail, listq) {
1162 pmap_qenter(zkva, &p, 1);
1166 return ((void *)retkva);
1170 * Frees a number of pages to the system
1173 * mem A pointer to the memory to be freed
1174 * size The size of the memory being freed
1175 * flags The original p->us_flags field
1181 page_free(void *mem, int size, uint8_t flags)
1185 if (flags & UMA_SLAB_KMEM)
1187 else if (flags & UMA_SLAB_KERNEL)
1188 vmem = kernel_arena;
1190 panic("UMA: page_free used with invalid flags %d", flags);
1192 kmem_free(vmem, (vm_offset_t)mem, size);
1196 * Zero fill initializer
1198 * Arguments/Returns follow uma_init specifications
1201 zero_init(void *mem, int size, int flags)
1208 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1211 * keg The zone we should initialize
1217 keg_small_init(uma_keg_t keg)
1224 if (keg->uk_flags & UMA_ZONE_PCPU) {
1225 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1227 keg->uk_slabsize = sizeof(struct pcpu);
1228 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1231 keg->uk_slabsize = UMA_SLAB_SIZE;
1236 * Calculate the size of each allocation (rsize) according to
1237 * alignment. If the requested size is smaller than we have
1238 * allocation bits for we round it up.
1240 rsize = keg->uk_size;
1241 if (rsize < keg->uk_slabsize / SLAB_SETSIZE)
1242 rsize = keg->uk_slabsize / SLAB_SETSIZE;
1243 if (rsize & keg->uk_align)
1244 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1245 keg->uk_rsize = rsize;
1247 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1248 keg->uk_rsize < sizeof(struct pcpu),
1249 ("%s: size %u too large", __func__, keg->uk_rsize));
1251 if (keg->uk_flags & UMA_ZONE_REFCNT)
1252 rsize += sizeof(uint32_t);
1254 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1257 shsize = sizeof(struct uma_slab);
1259 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1260 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1261 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1263 memused = keg->uk_ipers * rsize + shsize;
1264 wastedspace = keg->uk_slabsize - memused;
1267 * We can't do OFFPAGE if we're internal or if we've been
1268 * asked to not go to the VM for buckets. If we do this we
1269 * may end up going to the VM for slabs which we do not
1270 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1271 * of UMA_ZONE_VM, which clearly forbids it.
1273 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1274 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1278 * See if using an OFFPAGE slab will limit our waste. Only do
1279 * this if it permits more items per-slab.
1281 * XXX We could try growing slabsize to limit max waste as well.
1282 * Historically this was not done because the VM could not
1283 * efficiently handle contiguous allocations.
1285 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1286 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1287 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1288 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1289 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1291 printf("UMA decided we need offpage slab headers for "
1292 "keg: %s, calculated wastedspace = %d, "
1293 "maximum wasted space allowed = %d, "
1294 "calculated ipers = %d, "
1295 "new wasted space = %d\n", keg->uk_name, wastedspace,
1296 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1297 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1299 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1302 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1303 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1304 keg->uk_flags |= UMA_ZONE_HASH;
1308 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1309 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1313 * keg The keg we should initialize
1319 keg_large_init(uma_keg_t keg)
1322 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1323 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1324 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1325 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1326 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1328 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1329 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1331 keg->uk_rsize = keg->uk_size;
1333 /* We can't do OFFPAGE if we're internal, bail out here. */
1334 if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1337 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1338 if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1339 keg->uk_flags |= UMA_ZONE_HASH;
1343 keg_cachespread_init(uma_keg_t keg)
1350 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1351 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1353 alignsize = keg->uk_align + 1;
1354 rsize = keg->uk_size;
1356 * We want one item to start on every align boundary in a page. To
1357 * do this we will span pages. We will also extend the item by the
1358 * size of align if it is an even multiple of align. Otherwise, it
1359 * would fall on the same boundary every time.
1361 if (rsize & keg->uk_align)
1362 rsize = (rsize & ~keg->uk_align) + alignsize;
1363 if ((rsize & alignsize) == 0)
1365 trailer = rsize - keg->uk_size;
1366 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1367 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1368 keg->uk_rsize = rsize;
1369 keg->uk_ppera = pages;
1370 keg->uk_slabsize = UMA_SLAB_SIZE;
1371 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1372 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1373 KASSERT(keg->uk_ipers <= uma_max_ipers,
1374 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1379 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1380 * the keg onto the global keg list.
1382 * Arguments/Returns follow uma_ctor specifications
1383 * udata Actually uma_kctor_args
1386 keg_ctor(void *mem, int size, void *udata, int flags)
1388 struct uma_kctor_args *arg = udata;
1389 uma_keg_t keg = mem;
1393 keg->uk_size = arg->size;
1394 keg->uk_init = arg->uminit;
1395 keg->uk_fini = arg->fini;
1396 keg->uk_align = arg->align;
1398 keg->uk_reserve = 0;
1400 keg->uk_flags = arg->flags;
1401 keg->uk_allocf = page_alloc;
1402 keg->uk_freef = page_free;
1403 keg->uk_slabzone = NULL;
1406 * The master zone is passed to us at keg-creation time.
1409 keg->uk_name = zone->uz_name;
1411 if (arg->flags & UMA_ZONE_VM)
1412 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1414 if (arg->flags & UMA_ZONE_ZINIT)
1415 keg->uk_init = zero_init;
1417 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1418 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1420 if (arg->flags & UMA_ZONE_PCPU)
1422 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1424 keg->uk_flags &= ~UMA_ZONE_PCPU;
1427 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1428 keg_cachespread_init(keg);
1429 } else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1431 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) -
1433 keg_large_init(keg);
1435 keg_small_init(keg);
1437 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1438 keg_large_init(keg);
1440 keg_small_init(keg);
1443 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1444 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1445 if (keg->uk_ipers > uma_max_ipers_ref)
1446 panic("Too many ref items per zone: %d > %d\n",
1447 keg->uk_ipers, uma_max_ipers_ref);
1448 keg->uk_slabzone = slabrefzone;
1450 keg->uk_slabzone = slabzone;
1454 * If we haven't booted yet we need allocations to go through the
1455 * startup cache until the vm is ready.
1457 if (keg->uk_ppera == 1) {
1458 #ifdef UMA_MD_SMALL_ALLOC
1459 keg->uk_allocf = uma_small_alloc;
1460 keg->uk_freef = uma_small_free;
1462 if (booted < UMA_STARTUP)
1463 keg->uk_allocf = startup_alloc;
1465 if (booted < UMA_STARTUP2)
1466 keg->uk_allocf = startup_alloc;
1468 } else if (booted < UMA_STARTUP2 &&
1469 (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1470 keg->uk_allocf = startup_alloc;
1473 * Initialize keg's lock
1475 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1478 * If we're putting the slab header in the actual page we need to
1479 * figure out where in each page it goes. This calculates a right
1480 * justified offset into the memory on an ALIGN_PTR boundary.
1482 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1485 /* Size of the slab struct and free list */
1486 totsize = sizeof(struct uma_slab);
1488 /* Size of the reference counts. */
1489 if (keg->uk_flags & UMA_ZONE_REFCNT)
1490 totsize += keg->uk_ipers * sizeof(uint32_t);
1492 if (totsize & UMA_ALIGN_PTR)
1493 totsize = (totsize & ~UMA_ALIGN_PTR) +
1494 (UMA_ALIGN_PTR + 1);
1495 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1498 * The only way the following is possible is if with our
1499 * UMA_ALIGN_PTR adjustments we are now bigger than
1500 * UMA_SLAB_SIZE. I haven't checked whether this is
1501 * mathematically possible for all cases, so we make
1504 totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1505 if (keg->uk_flags & UMA_ZONE_REFCNT)
1506 totsize += keg->uk_ipers * sizeof(uint32_t);
1507 if (totsize > PAGE_SIZE * keg->uk_ppera) {
1508 printf("zone %s ipers %d rsize %d size %d\n",
1509 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1511 panic("UMA slab won't fit.");
1515 if (keg->uk_flags & UMA_ZONE_HASH)
1516 hash_alloc(&keg->uk_hash);
1519 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1520 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1521 keg->uk_ipers, keg->uk_ppera,
1522 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1525 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1528 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1529 mtx_unlock(&uma_mtx);
1534 * Zone header ctor. This initializes all fields, locks, etc.
1536 * Arguments/Returns follow uma_ctor specifications
1537 * udata Actually uma_zctor_args
1540 zone_ctor(void *mem, int size, void *udata, int flags)
1542 struct uma_zctor_args *arg = udata;
1543 uma_zone_t zone = mem;
1548 zone->uz_name = arg->name;
1549 zone->uz_ctor = arg->ctor;
1550 zone->uz_dtor = arg->dtor;
1551 zone->uz_slab = zone_fetch_slab;
1552 zone->uz_init = NULL;
1553 zone->uz_fini = NULL;
1554 zone->uz_allocs = 0;
1557 zone->uz_sleeps = 0;
1559 zone->uz_count_min = 0;
1561 zone->uz_warning = NULL;
1562 timevalclear(&zone->uz_ratecheck);
1565 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1568 * This is a pure cache zone, no kegs.
1571 if (arg->flags & UMA_ZONE_VM)
1572 arg->flags |= UMA_ZFLAG_CACHEONLY;
1573 zone->uz_flags = arg->flags;
1574 zone->uz_size = arg->size;
1575 zone->uz_import = arg->import;
1576 zone->uz_release = arg->release;
1577 zone->uz_arg = arg->arg;
1578 zone->uz_lockptr = &zone->uz_lock;
1583 * Use the regular zone/keg/slab allocator.
1585 zone->uz_import = (uma_import)zone_import;
1586 zone->uz_release = (uma_release)zone_release;
1587 zone->uz_arg = zone;
1589 if (arg->flags & UMA_ZONE_SECONDARY) {
1590 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1591 zone->uz_init = arg->uminit;
1592 zone->uz_fini = arg->fini;
1593 zone->uz_lockptr = &keg->uk_lock;
1594 zone->uz_flags |= UMA_ZONE_SECONDARY;
1597 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1598 if (LIST_NEXT(z, uz_link) == NULL) {
1599 LIST_INSERT_AFTER(z, zone, uz_link);
1604 mtx_unlock(&uma_mtx);
1605 } else if (keg == NULL) {
1606 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1607 arg->align, arg->flags)) == NULL)
1610 struct uma_kctor_args karg;
1613 /* We should only be here from uma_startup() */
1614 karg.size = arg->size;
1615 karg.uminit = arg->uminit;
1616 karg.fini = arg->fini;
1617 karg.align = arg->align;
1618 karg.flags = arg->flags;
1620 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1627 * Link in the first keg.
1629 zone->uz_klink.kl_keg = keg;
1630 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1631 zone->uz_lockptr = &keg->uk_lock;
1632 zone->uz_size = keg->uk_size;
1633 zone->uz_flags |= (keg->uk_flags &
1634 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1637 * Some internal zones don't have room allocated for the per cpu
1638 * caches. If we're internal, bail out here.
1640 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1641 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1642 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1647 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1648 zone->uz_count = bucket_select(zone->uz_size);
1650 zone->uz_count = BUCKET_MAX;
1651 zone->uz_count_min = zone->uz_count;
1657 * Keg header dtor. This frees all data, destroys locks, frees the hash
1658 * table and removes the keg from the global list.
1660 * Arguments/Returns follow uma_dtor specifications
1664 keg_dtor(void *arg, int size, void *udata)
1668 keg = (uma_keg_t)arg;
1670 if (keg->uk_free != 0) {
1671 printf("Freed UMA keg (%s) was not empty (%d items). "
1672 " Lost %d pages of memory.\n",
1673 keg->uk_name ? keg->uk_name : "",
1674 keg->uk_free, keg->uk_pages);
1678 hash_free(&keg->uk_hash);
1686 * Arguments/Returns follow uma_dtor specifications
1690 zone_dtor(void *arg, int size, void *udata)
1696 zone = (uma_zone_t)arg;
1697 keg = zone_first_keg(zone);
1699 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1703 LIST_REMOVE(zone, uz_link);
1704 mtx_unlock(&uma_mtx);
1706 * XXX there are some races here where
1707 * the zone can be drained but zone lock
1708 * released and then refilled before we
1709 * remove it... we dont care for now
1711 zone_drain_wait(zone, M_WAITOK);
1713 * Unlink all of our kegs.
1715 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1716 klink->kl_keg = NULL;
1717 LIST_REMOVE(klink, kl_link);
1718 if (klink == &zone->uz_klink)
1720 free(klink, M_TEMP);
1723 * We only destroy kegs from non secondary zones.
1725 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1727 LIST_REMOVE(keg, uk_link);
1728 mtx_unlock(&uma_mtx);
1729 zone_free_item(kegs, keg, NULL, SKIP_NONE);
1731 ZONE_LOCK_FINI(zone);
1735 * Traverses every zone in the system and calls a callback
1738 * zfunc A pointer to a function which accepts a zone
1745 zone_foreach(void (*zfunc)(uma_zone_t))
1751 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1752 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1755 mtx_unlock(&uma_mtx);
1758 /* Public functions */
1761 uma_startup(void *bootmem, int boot_pages)
1763 struct uma_zctor_args args;
1769 printf("Creating uma keg headers zone and keg.\n");
1771 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1773 /* "manually" create the initial zone */
1774 memset(&args, 0, sizeof(args));
1775 args.name = "UMA Kegs";
1776 args.size = sizeof(struct uma_keg);
1777 args.ctor = keg_ctor;
1778 args.dtor = keg_dtor;
1779 args.uminit = zero_init;
1781 args.keg = &masterkeg;
1782 args.align = 32 - 1;
1783 args.flags = UMA_ZFLAG_INTERNAL;
1784 /* The initial zone has no Per cpu queues so it's smaller */
1785 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1788 printf("Filling boot free list.\n");
1790 for (i = 0; i < boot_pages; i++) {
1791 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1792 slab->us_data = (uint8_t *)slab;
1793 slab->us_flags = UMA_SLAB_BOOT;
1794 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1796 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1799 printf("Creating uma zone headers zone and keg.\n");
1801 args.name = "UMA Zones";
1802 args.size = sizeof(struct uma_zone) +
1803 (sizeof(struct uma_cache) * (mp_maxid + 1));
1804 args.ctor = zone_ctor;
1805 args.dtor = zone_dtor;
1806 args.uminit = zero_init;
1809 args.align = 32 - 1;
1810 args.flags = UMA_ZFLAG_INTERNAL;
1811 /* The initial zone has no Per cpu queues so it's smaller */
1812 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1815 printf("Initializing pcpu cache locks.\n");
1818 printf("Creating slab and hash zones.\n");
1821 /* Now make a zone for slab headers */
1822 slabzone = uma_zcreate("UMA Slabs",
1823 sizeof(struct uma_slab),
1824 NULL, NULL, NULL, NULL,
1825 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1828 * We also create a zone for the bigger slabs with reference
1829 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1831 slabsize = sizeof(struct uma_slab_refcnt);
1832 slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1833 slabrefzone = uma_zcreate("UMA RCntSlabs",
1835 NULL, NULL, NULL, NULL,
1837 UMA_ZFLAG_INTERNAL);
1839 hashzone = uma_zcreate("UMA Hash",
1840 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1841 NULL, NULL, NULL, NULL,
1842 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1846 booted = UMA_STARTUP;
1849 printf("UMA startup complete.\n");
1857 booted = UMA_STARTUP2;
1860 printf("UMA startup2 complete.\n");
1865 * Initialize our callout handle
1873 printf("Starting callout.\n");
1875 callout_init(&uma_callout, CALLOUT_MPSAFE);
1876 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1878 printf("UMA startup3 complete.\n");
1883 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1884 int align, uint32_t flags)
1886 struct uma_kctor_args args;
1889 args.uminit = uminit;
1891 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1894 return (zone_alloc_item(kegs, &args, M_WAITOK));
1899 uma_set_align(int align)
1902 if (align != UMA_ALIGN_CACHE)
1903 uma_align_cache = align;
1908 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1909 uma_init uminit, uma_fini fini, int align, uint32_t flags)
1912 struct uma_zctor_args args;
1914 /* This stuff is essential for the zone ctor */
1915 memset(&args, 0, sizeof(args));
1920 args.uminit = uminit;
1926 return (zone_alloc_item(zones, &args, M_WAITOK));
1931 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1932 uma_init zinit, uma_fini zfini, uma_zone_t master)
1934 struct uma_zctor_args args;
1937 keg = zone_first_keg(master);
1938 memset(&args, 0, sizeof(args));
1940 args.size = keg->uk_size;
1943 args.uminit = zinit;
1945 args.align = keg->uk_align;
1946 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1949 /* XXX Attaches only one keg of potentially many. */
1950 return (zone_alloc_item(zones, &args, M_WAITOK));
1955 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1956 uma_init zinit, uma_fini zfini, uma_import zimport,
1957 uma_release zrelease, void *arg, int flags)
1959 struct uma_zctor_args args;
1961 memset(&args, 0, sizeof(args));
1966 args.uminit = zinit;
1968 args.import = zimport;
1969 args.release = zrelease;
1974 return (zone_alloc_item(zones, &args, M_WAITOK));
1978 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1982 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
1985 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
1990 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1998 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2005 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2007 zone_lock_pair(zone, master);
2009 * zone must use vtoslab() to resolve objects and must already be
2012 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2013 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2018 * The new master must also use vtoslab().
2020 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2025 * Both must either be refcnt, or not be refcnt.
2027 if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
2028 (master->uz_flags & UMA_ZONE_REFCNT)) {
2033 * The underlying object must be the same size. rsize
2036 if (master->uz_size != zone->uz_size) {
2041 * Put it at the end of the list.
2043 klink->kl_keg = zone_first_keg(master);
2044 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2045 if (LIST_NEXT(kl, kl_link) == NULL) {
2046 LIST_INSERT_AFTER(kl, klink, kl_link);
2051 zone->uz_flags |= UMA_ZFLAG_MULTI;
2052 zone->uz_slab = zone_fetch_slab_multi;
2055 zone_unlock_pair(zone, master);
2057 free(klink, M_TEMP);
2065 uma_zdestroy(uma_zone_t zone)
2068 zone_free_item(zones, zone, NULL, SKIP_NONE);
2073 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2077 uma_bucket_t bucket;
2081 /* This is the fast path allocation */
2082 #ifdef UMA_DEBUG_ALLOC_1
2083 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2085 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2086 zone->uz_name, flags);
2088 if (flags & M_WAITOK) {
2089 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2090 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2092 #ifdef DEBUG_MEMGUARD
2093 if (memguard_cmp_zone(zone)) {
2094 item = memguard_alloc(zone->uz_size, flags);
2097 * Avoid conflict with the use-after-free
2098 * protecting infrastructure from INVARIANTS.
2100 if (zone->uz_init != NULL &&
2101 zone->uz_init != mtrash_init &&
2102 zone->uz_init(item, zone->uz_size, flags) != 0)
2104 if (zone->uz_ctor != NULL &&
2105 zone->uz_ctor != mtrash_ctor &&
2106 zone->uz_ctor(item, zone->uz_size, udata,
2108 zone->uz_fini(item, zone->uz_size);
2113 /* This is unfortunate but should not be fatal. */
2117 * If possible, allocate from the per-CPU cache. There are two
2118 * requirements for safe access to the per-CPU cache: (1) the thread
2119 * accessing the cache must not be preempted or yield during access,
2120 * and (2) the thread must not migrate CPUs without switching which
2121 * cache it accesses. We rely on a critical section to prevent
2122 * preemption and migration. We release the critical section in
2123 * order to acquire the zone mutex if we are unable to allocate from
2124 * the current cache; when we re-acquire the critical section, we
2125 * must detect and handle migration if it has occurred.
2129 cache = &zone->uz_cpu[cpu];
2132 bucket = cache->uc_allocbucket;
2133 if (bucket != NULL && bucket->ub_cnt > 0) {
2135 item = bucket->ub_bucket[bucket->ub_cnt];
2137 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2139 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2142 if (zone->uz_ctor != NULL &&
2143 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2144 atomic_add_long(&zone->uz_fails, 1);
2145 zone_free_item(zone, item, udata, SKIP_DTOR);
2149 uma_dbg_alloc(zone, NULL, item);
2152 bzero(item, zone->uz_size);
2157 * We have run out of items in our alloc bucket.
2158 * See if we can switch with our free bucket.
2160 bucket = cache->uc_freebucket;
2161 if (bucket != NULL && bucket->ub_cnt > 0) {
2162 #ifdef UMA_DEBUG_ALLOC
2163 printf("uma_zalloc: Swapping empty with alloc.\n");
2165 cache->uc_freebucket = cache->uc_allocbucket;
2166 cache->uc_allocbucket = bucket;
2171 * Discard any empty allocation bucket while we hold no locks.
2173 bucket = cache->uc_allocbucket;
2174 cache->uc_allocbucket = NULL;
2177 bucket_free(zone, bucket, udata);
2179 /* Short-circuit for zones without buckets and low memory. */
2180 if (zone->uz_count == 0 || bucketdisable)
2184 * Attempt to retrieve the item from the per-CPU cache has failed, so
2185 * we must go back to the zone. This requires the zone lock, so we
2186 * must drop the critical section, then re-acquire it when we go back
2187 * to the cache. Since the critical section is released, we may be
2188 * preempted or migrate. As such, make sure not to maintain any
2189 * thread-local state specific to the cache from prior to releasing
2190 * the critical section.
2193 if (ZONE_TRYLOCK(zone) == 0) {
2194 /* Record contention to size the buckets. */
2200 cache = &zone->uz_cpu[cpu];
2203 * Since we have locked the zone we may as well send back our stats.
2205 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2206 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2207 cache->uc_allocs = 0;
2208 cache->uc_frees = 0;
2210 /* See if we lost the race to fill the cache. */
2211 if (cache->uc_allocbucket != NULL) {
2217 * Check the zone's cache of buckets.
2219 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2220 KASSERT(bucket->ub_cnt != 0,
2221 ("uma_zalloc_arg: Returning an empty bucket."));
2223 LIST_REMOVE(bucket, ub_link);
2224 cache->uc_allocbucket = bucket;
2228 /* We are no longer associated with this CPU. */
2232 * We bump the uz count when the cache size is insufficient to
2233 * handle the working set.
2235 if (lockfail && zone->uz_count < BUCKET_MAX)
2240 * Now lets just fill a bucket and put it on the free list. If that
2241 * works we'll restart the allocation from the begining and it
2242 * will use the just filled bucket.
2244 bucket = zone_alloc_bucket(zone, udata, flags);
2245 if (bucket != NULL) {
2249 cache = &zone->uz_cpu[cpu];
2251 * See if we lost the race or were migrated. Cache the
2252 * initialized bucket to make this less likely or claim
2253 * the memory directly.
2255 if (cache->uc_allocbucket == NULL)
2256 cache->uc_allocbucket = bucket;
2258 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2264 * We may not be able to get a bucket so return an actual item.
2267 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2271 item = zone_alloc_item(zone, udata, flags);
2277 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2282 mtx_assert(&keg->uk_lock, MA_OWNED);
2285 if ((flags & M_USE_RESERVE) == 0)
2286 reserve = keg->uk_reserve;
2290 * Find a slab with some space. Prefer slabs that are partially
2291 * used over those that are totally full. This helps to reduce
2294 if (keg->uk_free > reserve) {
2295 if (!LIST_EMPTY(&keg->uk_part_slab)) {
2296 slab = LIST_FIRST(&keg->uk_part_slab);
2298 slab = LIST_FIRST(&keg->uk_free_slab);
2299 LIST_REMOVE(slab, us_link);
2300 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2303 MPASS(slab->us_keg == keg);
2308 * M_NOVM means don't ask at all!
2313 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2314 keg->uk_flags |= UMA_ZFLAG_FULL;
2316 * If this is not a multi-zone, set the FULL bit.
2317 * Otherwise slab_multi() takes care of it.
2319 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2320 zone->uz_flags |= UMA_ZFLAG_FULL;
2321 zone_log_warning(zone);
2323 if (flags & M_NOWAIT)
2326 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2329 slab = keg_alloc_slab(keg, zone, flags);
2331 * If we got a slab here it's safe to mark it partially used
2332 * and return. We assume that the caller is going to remove
2333 * at least one item.
2336 MPASS(slab->us_keg == keg);
2337 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2341 * We might not have been able to get a slab but another cpu
2342 * could have while we were unlocked. Check again before we
2351 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2356 keg = zone_first_keg(zone);
2361 slab = keg_fetch_slab(keg, zone, flags);
2364 if (flags & (M_NOWAIT | M_NOVM))
2372 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2373 * with the keg locked. On NULL no lock is held.
2375 * The last pointer is used to seed the search. It is not required.
2378 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2388 * Don't wait on the first pass. This will skip limit tests
2389 * as well. We don't want to block if we can find a provider
2392 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2394 * Use the last slab allocated as a hint for where to start
2398 slab = keg_fetch_slab(last, zone, flags);
2404 * Loop until we have a slab incase of transient failures
2405 * while M_WAITOK is specified. I'm not sure this is 100%
2406 * required but we've done it for so long now.
2412 * Search the available kegs for slabs. Be careful to hold the
2413 * correct lock while calling into the keg layer.
2415 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2416 keg = klink->kl_keg;
2418 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2419 slab = keg_fetch_slab(keg, zone, flags);
2423 if (keg->uk_flags & UMA_ZFLAG_FULL)
2429 if (rflags & (M_NOWAIT | M_NOVM))
2433 * All kegs are full. XXX We can't atomically check all kegs
2434 * and sleep so just sleep for a short period and retry.
2436 if (full && !empty) {
2438 zone->uz_flags |= UMA_ZFLAG_FULL;
2440 zone_log_warning(zone);
2441 msleep(zone, zone->uz_lockptr, PVM,
2442 "zonelimit", hz/100);
2443 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2452 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2457 MPASS(keg == slab->us_keg);
2458 mtx_assert(&keg->uk_lock, MA_OWNED);
2460 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2461 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2462 item = slab->us_data + (keg->uk_rsize * freei);
2463 slab->us_freecount--;
2466 /* Move this slab to the full list */
2467 if (slab->us_freecount == 0) {
2468 LIST_REMOVE(slab, us_link);
2469 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2476 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2484 /* Try to keep the buckets totally full */
2485 for (i = 0; i < max; ) {
2486 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2489 while (slab->us_freecount && i < max) {
2490 bucket[i++] = slab_alloc_item(keg, slab);
2491 if (keg->uk_free <= keg->uk_reserve)
2494 /* Don't grab more than one slab at a time. */
2505 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2507 uma_bucket_t bucket;
2510 /* Don't wait for buckets, preserve caller's NOVM setting. */
2511 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2515 max = MIN(bucket->ub_entries, zone->uz_count);
2516 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2520 * Initialize the memory if necessary.
2522 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2525 for (i = 0; i < bucket->ub_cnt; i++)
2526 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2530 * If we couldn't initialize the whole bucket, put the
2531 * rest back onto the freelist.
2533 if (i != bucket->ub_cnt) {
2534 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2535 bucket->ub_cnt - i);
2537 bzero(&bucket->ub_bucket[i],
2538 sizeof(void *) * (bucket->ub_cnt - i));
2545 if (bucket == NULL || bucket->ub_cnt == 0) {
2547 bucket_free(zone, bucket, udata);
2548 atomic_add_long(&zone->uz_fails, 1);
2556 * Allocates a single item from a zone.
2559 * zone The zone to alloc for.
2560 * udata The data to be passed to the constructor.
2561 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2564 * NULL if there is no memory and M_NOWAIT is set
2565 * An item if successful
2569 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2575 #ifdef UMA_DEBUG_ALLOC
2576 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2578 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2580 atomic_add_long(&zone->uz_allocs, 1);
2583 * We have to call both the zone's init (not the keg's init)
2584 * and the zone's ctor. This is because the item is going from
2585 * a keg slab directly to the user, and the user is expecting it
2586 * to be both zone-init'd as well as zone-ctor'd.
2588 if (zone->uz_init != NULL) {
2589 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2590 zone_free_item(zone, item, udata, SKIP_FINI);
2594 if (zone->uz_ctor != NULL) {
2595 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2596 zone_free_item(zone, item, udata, SKIP_DTOR);
2601 uma_dbg_alloc(zone, NULL, item);
2604 bzero(item, zone->uz_size);
2609 atomic_add_long(&zone->uz_fails, 1);
2615 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2618 uma_bucket_t bucket;
2622 #ifdef UMA_DEBUG_ALLOC_1
2623 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2625 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2628 /* uma_zfree(..., NULL) does nothing, to match free(9). */
2631 #ifdef DEBUG_MEMGUARD
2632 if (is_memguard_addr(item)) {
2633 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2634 zone->uz_dtor(item, zone->uz_size, udata);
2635 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2636 zone->uz_fini(item, zone->uz_size);
2637 memguard_free(item);
2642 if (zone->uz_flags & UMA_ZONE_MALLOC)
2643 uma_dbg_free(zone, udata, item);
2645 uma_dbg_free(zone, NULL, item);
2647 if (zone->uz_dtor != NULL)
2648 zone->uz_dtor(item, zone->uz_size, udata);
2651 * The race here is acceptable. If we miss it we'll just have to wait
2652 * a little longer for the limits to be reset.
2654 if (zone->uz_flags & UMA_ZFLAG_FULL)
2658 * If possible, free to the per-CPU cache. There are two
2659 * requirements for safe access to the per-CPU cache: (1) the thread
2660 * accessing the cache must not be preempted or yield during access,
2661 * and (2) the thread must not migrate CPUs without switching which
2662 * cache it accesses. We rely on a critical section to prevent
2663 * preemption and migration. We release the critical section in
2664 * order to acquire the zone mutex if we are unable to free to the
2665 * current cache; when we re-acquire the critical section, we must
2666 * detect and handle migration if it has occurred.
2671 cache = &zone->uz_cpu[cpu];
2675 * Try to free into the allocbucket first to give LIFO ordering
2676 * for cache-hot datastructures. Spill over into the freebucket
2677 * if necessary. Alloc will swap them if one runs dry.
2679 bucket = cache->uc_allocbucket;
2680 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2681 bucket = cache->uc_freebucket;
2682 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2683 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2684 ("uma_zfree: Freeing to non free bucket index."));
2685 bucket->ub_bucket[bucket->ub_cnt] = item;
2693 * We must go back the zone, which requires acquiring the zone lock,
2694 * which in turn means we must release and re-acquire the critical
2695 * section. Since the critical section is released, we may be
2696 * preempted or migrate. As such, make sure not to maintain any
2697 * thread-local state specific to the cache from prior to releasing
2698 * the critical section.
2701 if (zone->uz_count == 0 || bucketdisable)
2705 if (ZONE_TRYLOCK(zone) == 0) {
2706 /* Record contention to size the buckets. */
2712 cache = &zone->uz_cpu[cpu];
2715 * Since we have locked the zone we may as well send back our stats.
2717 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2718 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2719 cache->uc_allocs = 0;
2720 cache->uc_frees = 0;
2722 bucket = cache->uc_freebucket;
2723 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2727 cache->uc_freebucket = NULL;
2729 /* Can we throw this on the zone full list? */
2730 if (bucket != NULL) {
2731 #ifdef UMA_DEBUG_ALLOC
2732 printf("uma_zfree: Putting old bucket on the free list.\n");
2734 /* ub_cnt is pointing to the last free item */
2735 KASSERT(bucket->ub_cnt != 0,
2736 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2737 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2740 /* We are no longer associated with this CPU. */
2744 * We bump the uz count when the cache size is insufficient to
2745 * handle the working set.
2747 if (lockfail && zone->uz_count < BUCKET_MAX)
2751 #ifdef UMA_DEBUG_ALLOC
2752 printf("uma_zfree: Allocating new free bucket.\n");
2754 bucket = bucket_alloc(zone, udata, M_NOWAIT);
2758 cache = &zone->uz_cpu[cpu];
2759 if (cache->uc_freebucket == NULL) {
2760 cache->uc_freebucket = bucket;
2764 * We lost the race, start over. We have to drop our
2765 * critical section to free the bucket.
2768 bucket_free(zone, bucket, udata);
2773 * If nothing else caught this, we'll just do an internal free.
2776 zone_free_item(zone, item, udata, SKIP_DTOR);
2782 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2786 mtx_assert(&keg->uk_lock, MA_OWNED);
2787 MPASS(keg == slab->us_keg);
2789 /* Do we need to remove from any lists? */
2790 if (slab->us_freecount+1 == keg->uk_ipers) {
2791 LIST_REMOVE(slab, us_link);
2792 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2793 } else if (slab->us_freecount == 0) {
2794 LIST_REMOVE(slab, us_link);
2795 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2798 /* Slab management. */
2799 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2800 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2801 slab->us_freecount++;
2803 /* Keg statistics. */
2808 zone_release(uma_zone_t zone, void **bucket, int cnt)
2818 keg = zone_first_keg(zone);
2820 for (i = 0; i < cnt; i++) {
2822 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2823 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2824 if (zone->uz_flags & UMA_ZONE_HASH) {
2825 slab = hash_sfind(&keg->uk_hash, mem);
2827 mem += keg->uk_pgoff;
2828 slab = (uma_slab_t)mem;
2831 slab = vtoslab((vm_offset_t)item);
2832 if (slab->us_keg != keg) {
2838 slab_free_item(keg, slab, item);
2839 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2840 if (keg->uk_pages < keg->uk_maxpages) {
2841 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2846 * We can handle one more allocation. Since we're
2847 * clearing ZFLAG_FULL, wake up all procs blocked
2848 * on pages. This should be uncommon, so keeping this
2849 * simple for now (rather than adding count of blocked
2858 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2866 * Frees a single item to any zone.
2869 * zone The zone to free to
2870 * item The item we're freeing
2871 * udata User supplied data for the dtor
2872 * skip Skip dtors and finis
2875 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2879 if (skip == SKIP_NONE) {
2880 if (zone->uz_flags & UMA_ZONE_MALLOC)
2881 uma_dbg_free(zone, udata, item);
2883 uma_dbg_free(zone, NULL, item);
2886 if (skip < SKIP_DTOR && zone->uz_dtor)
2887 zone->uz_dtor(item, zone->uz_size, udata);
2889 if (skip < SKIP_FINI && zone->uz_fini)
2890 zone->uz_fini(item, zone->uz_size);
2892 atomic_add_long(&zone->uz_frees, 1);
2893 zone->uz_release(zone->uz_arg, &item, 1);
2898 uma_zone_set_max(uma_zone_t zone, int nitems)
2902 keg = zone_first_keg(zone);
2906 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2907 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2908 keg->uk_maxpages += keg->uk_ppera;
2909 nitems = keg->uk_maxpages * keg->uk_ipers;
2917 uma_zone_get_max(uma_zone_t zone)
2922 keg = zone_first_keg(zone);
2926 nitems = keg->uk_maxpages * keg->uk_ipers;
2934 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2938 zone->uz_warning = warning;
2944 uma_zone_get_cur(uma_zone_t zone)
2950 nitems = zone->uz_allocs - zone->uz_frees;
2953 * See the comment in sysctl_vm_zone_stats() regarding the
2954 * safety of accessing the per-cpu caches. With the zone lock
2955 * held, it is safe, but can potentially result in stale data.
2957 nitems += zone->uz_cpu[i].uc_allocs -
2958 zone->uz_cpu[i].uc_frees;
2962 return (nitems < 0 ? 0 : nitems);
2967 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2971 keg = zone_first_keg(zone);
2972 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2974 KASSERT(keg->uk_pages == 0,
2975 ("uma_zone_set_init on non-empty keg"));
2976 keg->uk_init = uminit;
2982 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2986 keg = zone_first_keg(zone);
2987 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2989 KASSERT(keg->uk_pages == 0,
2990 ("uma_zone_set_fini on non-empty keg"));
2991 keg->uk_fini = fini;
2997 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3001 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3002 ("uma_zone_set_zinit on non-empty keg"));
3003 zone->uz_init = zinit;
3009 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3013 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3014 ("uma_zone_set_zfini on non-empty keg"));
3015 zone->uz_fini = zfini;
3020 /* XXX uk_freef is not actually used with the zone locked */
3022 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3026 keg = zone_first_keg(zone);
3027 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3029 keg->uk_freef = freef;
3034 /* XXX uk_allocf is not actually used with the zone locked */
3036 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3040 keg = zone_first_keg(zone);
3042 keg->uk_allocf = allocf;
3048 uma_zone_reserve(uma_zone_t zone, int items)
3052 keg = zone_first_keg(zone);
3056 keg->uk_reserve = items;
3064 uma_zone_reserve_kva(uma_zone_t zone, int count)
3070 keg = zone_first_keg(zone);
3073 pages = count / keg->uk_ipers;
3075 if (pages * keg->uk_ipers < count)
3078 #ifdef UMA_MD_SMALL_ALLOC
3079 if (keg->uk_ppera > 1) {
3083 kva = kva_alloc(pages * UMA_SLAB_SIZE);
3091 keg->uk_maxpages = pages;
3092 #ifdef UMA_MD_SMALL_ALLOC
3093 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3095 keg->uk_allocf = noobj_alloc;
3097 keg->uk_flags |= UMA_ZONE_NOFREE;
3105 uma_prealloc(uma_zone_t zone, int items)
3111 keg = zone_first_keg(zone);
3115 slabs = items / keg->uk_ipers;
3116 if (slabs * keg->uk_ipers < items)
3119 slab = keg_alloc_slab(keg, zone, M_WAITOK);
3122 MPASS(slab->us_keg == keg);
3123 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3131 uma_find_refcnt(uma_zone_t zone, void *item)
3133 uma_slabrefcnt_t slabref;
3139 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3140 slabref = (uma_slabrefcnt_t)slab;
3142 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3143 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3144 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3145 refcnt = &slabref->us_refcnt[idx];
3154 printf("UMA: vm asked us to release pages!\n");
3157 zone_foreach(zone_drain);
3158 if (vm_page_count_min()) {
3159 cache_drain_safe(NULL);
3160 zone_foreach(zone_drain);
3163 * Some slabs may have been freed but this zone will be visited early
3164 * we visit again so that we can free pages that are empty once other
3165 * zones are drained. We have to do the same for buckets.
3167 zone_drain(slabzone);
3168 zone_drain(slabrefzone);
3169 bucket_zone_drain();
3174 uma_zone_exhausted(uma_zone_t zone)
3179 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3185 uma_zone_exhausted_nolock(uma_zone_t zone)
3187 return (zone->uz_flags & UMA_ZFLAG_FULL);
3191 uma_large_malloc(int size, int wait)
3197 slab = zone_alloc_item(slabzone, NULL, wait);
3200 mem = page_alloc(NULL, size, &flags, wait);
3202 vsetslab((vm_offset_t)mem, slab);
3203 slab->us_data = mem;
3204 slab->us_flags = flags | UMA_SLAB_MALLOC;
3205 slab->us_size = size;
3207 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3214 uma_large_free(uma_slab_t slab)
3217 page_free(slab->us_data, slab->us_size, slab->us_flags);
3218 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3222 uma_print_stats(void)
3224 zone_foreach(uma_print_zone);
3228 slab_print(uma_slab_t slab)
3230 printf("slab: keg %p, data %p, freecount %d\n",
3231 slab->us_keg, slab->us_data, slab->us_freecount);
3235 cache_print(uma_cache_t cache)
3237 printf("alloc: %p(%d), free: %p(%d)\n",
3238 cache->uc_allocbucket,
3239 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3240 cache->uc_freebucket,
3241 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3245 uma_print_keg(uma_keg_t keg)
3249 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3250 "out %d free %d limit %d\n",
3251 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3252 keg->uk_ipers, keg->uk_ppera,
3253 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3254 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3255 printf("Part slabs:\n");
3256 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3258 printf("Free slabs:\n");
3259 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3261 printf("Full slabs:\n");
3262 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3267 uma_print_zone(uma_zone_t zone)
3273 printf("zone: %s(%p) size %d flags %#x\n",
3274 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3275 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3276 uma_print_keg(kl->kl_keg);
3278 cache = &zone->uz_cpu[i];
3279 printf("CPU %d Cache:\n", i);
3286 * Generate statistics across both the zone and its per-cpu cache's. Return
3287 * desired statistics if the pointer is non-NULL for that statistic.
3289 * Note: does not update the zone statistics, as it can't safely clear the
3290 * per-CPU cache statistic.
3292 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3293 * safe from off-CPU; we should modify the caches to track this information
3294 * directly so that we don't have to.
3297 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3298 uint64_t *freesp, uint64_t *sleepsp)
3301 uint64_t allocs, frees, sleeps;
3304 allocs = frees = sleeps = 0;
3307 cache = &z->uz_cpu[cpu];
3308 if (cache->uc_allocbucket != NULL)
3309 cachefree += cache->uc_allocbucket->ub_cnt;
3310 if (cache->uc_freebucket != NULL)
3311 cachefree += cache->uc_freebucket->ub_cnt;
3312 allocs += cache->uc_allocs;
3313 frees += cache->uc_frees;
3315 allocs += z->uz_allocs;
3316 frees += z->uz_frees;
3317 sleeps += z->uz_sleeps;
3318 if (cachefreep != NULL)
3319 *cachefreep = cachefree;
3320 if (allocsp != NULL)
3324 if (sleepsp != NULL)
3330 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3338 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3339 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3342 mtx_unlock(&uma_mtx);
3343 return (sysctl_handle_int(oidp, &count, 0, req));
3347 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3349 struct uma_stream_header ush;
3350 struct uma_type_header uth;
3351 struct uma_percpu_stat ups;
3352 uma_bucket_t bucket;
3359 int count, error, i;
3361 error = sysctl_wire_old_buffer(req, 0);
3364 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3368 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3369 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3374 * Insert stream header.
3376 bzero(&ush, sizeof(ush));
3377 ush.ush_version = UMA_STREAM_VERSION;
3378 ush.ush_maxcpus = (mp_maxid + 1);
3379 ush.ush_count = count;
3380 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3382 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3383 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3384 bzero(&uth, sizeof(uth));
3386 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3387 uth.uth_align = kz->uk_align;
3388 uth.uth_size = kz->uk_size;
3389 uth.uth_rsize = kz->uk_rsize;
3390 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3392 uth.uth_maxpages += k->uk_maxpages;
3393 uth.uth_pages += k->uk_pages;
3394 uth.uth_keg_free += k->uk_free;
3395 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3400 * A zone is secondary is it is not the first entry
3401 * on the keg's zone list.
3403 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3404 (LIST_FIRST(&kz->uk_zones) != z))
3405 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3407 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3408 uth.uth_zone_free += bucket->ub_cnt;
3409 uth.uth_allocs = z->uz_allocs;
3410 uth.uth_frees = z->uz_frees;
3411 uth.uth_fails = z->uz_fails;
3412 uth.uth_sleeps = z->uz_sleeps;
3413 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3415 * While it is not normally safe to access the cache
3416 * bucket pointers while not on the CPU that owns the
3417 * cache, we only allow the pointers to be exchanged
3418 * without the zone lock held, not invalidated, so
3419 * accept the possible race associated with bucket
3420 * exchange during monitoring.
3422 for (i = 0; i < (mp_maxid + 1); i++) {
3423 bzero(&ups, sizeof(ups));
3424 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3428 cache = &z->uz_cpu[i];
3429 if (cache->uc_allocbucket != NULL)
3430 ups.ups_cache_free +=
3431 cache->uc_allocbucket->ub_cnt;
3432 if (cache->uc_freebucket != NULL)
3433 ups.ups_cache_free +=
3434 cache->uc_freebucket->ub_cnt;
3435 ups.ups_allocs = cache->uc_allocs;
3436 ups.ups_frees = cache->uc_frees;
3438 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3443 mtx_unlock(&uma_mtx);
3444 error = sbuf_finish(&sbuf);
3450 DB_SHOW_COMMAND(uma, db_show_uma)
3452 uint64_t allocs, frees, sleeps;
3453 uma_bucket_t bucket;
3458 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3459 "Requests", "Sleeps");
3460 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3461 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3462 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3463 allocs = z->uz_allocs;
3464 frees = z->uz_frees;
3465 sleeps = z->uz_sleeps;
3468 uma_zone_sumstat(z, &cachefree, &allocs,
3470 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3471 (LIST_FIRST(&kz->uk_zones) != z)))
3472 cachefree += kz->uk_free;
3473 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3474 cachefree += bucket->ub_cnt;
3475 db_printf("%18s %8ju %8jd %8d %12ju %8ju\n", z->uz_name,
3476 (uintmax_t)kz->uk_size,
3477 (intmax_t)(allocs - frees), cachefree,
3478 (uintmax_t)allocs, sleeps);