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/random.h>
77 #include <sys/rwlock.h>
79 #include <sys/sched.h>
81 #include <sys/vmmeter.h>
84 #include <vm/vm_object.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_pageout.h>
87 #include <vm/vm_param.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_kern.h>
90 #include <vm/vm_extern.h>
92 #include <vm/uma_int.h>
93 #include <vm/uma_dbg.h>
98 #include <vm/memguard.h>
102 * This is the zone and keg from which all zones are spawned. The idea is that
103 * even the zone & keg heads are allocated from the allocator, so we use the
104 * bss section to bootstrap us.
106 static struct uma_keg masterkeg;
107 static struct uma_zone masterzone_k;
108 static struct uma_zone masterzone_z;
109 static uma_zone_t kegs = &masterzone_k;
110 static uma_zone_t zones = &masterzone_z;
112 /* This is the zone from which all of uma_slab_t's are allocated. */
113 static uma_zone_t slabzone;
114 static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
117 * The initial hash tables come out of this zone so they can be allocated
118 * prior to malloc coming up.
120 static uma_zone_t hashzone;
122 /* The boot-time adjusted value for cache line alignment. */
123 int uma_align_cache = 64 - 1;
125 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
128 * Are we allowed to allocate buckets?
130 static int bucketdisable = 1;
132 /* Linked list of all kegs in the system */
133 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
135 /* Linked list of all cache-only zones in the system */
136 static LIST_HEAD(,uma_zone) uma_cachezones =
137 LIST_HEAD_INITIALIZER(uma_cachezones);
139 /* This RW lock protects the keg list */
140 static struct rwlock_padalign uma_rwlock;
142 /* Linked list of boot time pages */
143 static LIST_HEAD(,uma_slab) uma_boot_pages =
144 LIST_HEAD_INITIALIZER(uma_boot_pages);
146 /* This mutex protects the boot time pages list */
147 static struct mtx_padalign uma_boot_pages_mtx;
149 static struct sx uma_drain_lock;
151 /* Is the VM done starting up? */
152 static int booted = 0;
153 #define UMA_STARTUP 1
154 #define UMA_STARTUP2 2
157 * Only mbuf clusters use ref zones. Just provide enough references
158 * to support the one user. New code should not use the ref facility.
160 static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES;
163 * This is the handle used to schedule events that need to happen
164 * outside of the allocation fast path.
166 static struct callout uma_callout;
167 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
170 * This structure is passed as the zone ctor arg so that I don't have to create
171 * a special allocation function just for zones.
173 struct uma_zctor_args {
188 struct uma_kctor_args {
197 struct uma_bucket_zone {
200 int ubz_entries; /* Number of items it can hold. */
201 int ubz_maxsize; /* Maximum allocation size per-item. */
205 * Compute the actual number of bucket entries to pack them in power
206 * of two sizes for more efficient space utilization.
208 #define BUCKET_SIZE(n) \
209 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
211 #define BUCKET_MAX BUCKET_SIZE(256)
213 struct uma_bucket_zone bucket_zones[] = {
214 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
215 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
216 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
217 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
218 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
219 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
220 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
221 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
222 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
227 * Flags and enumerations to be passed to internal functions.
229 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
233 static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
234 static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
235 static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
236 static void page_free(void *, vm_size_t, uint8_t);
237 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
238 static void cache_drain(uma_zone_t);
239 static void bucket_drain(uma_zone_t, uma_bucket_t);
240 static void bucket_cache_drain(uma_zone_t zone);
241 static int keg_ctor(void *, int, void *, int);
242 static void keg_dtor(void *, int, void *);
243 static int zone_ctor(void *, int, void *, int);
244 static void zone_dtor(void *, int, void *);
245 static int zero_init(void *, int, int);
246 static void keg_small_init(uma_keg_t keg);
247 static void keg_large_init(uma_keg_t keg);
248 static void zone_foreach(void (*zfunc)(uma_zone_t));
249 static void zone_timeout(uma_zone_t zone);
250 static int hash_alloc(struct uma_hash *);
251 static int hash_expand(struct uma_hash *, struct uma_hash *);
252 static void hash_free(struct uma_hash *hash);
253 static void uma_timeout(void *);
254 static void uma_startup3(void);
255 static void *zone_alloc_item(uma_zone_t, void *, int);
256 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
257 static void bucket_enable(void);
258 static void bucket_init(void);
259 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
260 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
261 static void bucket_zone_drain(void);
262 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
263 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
264 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
265 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
266 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
267 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
268 uma_fini fini, int align, uint32_t flags);
269 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
270 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
271 static void uma_zero_item(void *item, uma_zone_t zone);
273 void uma_print_zone(uma_zone_t);
274 void uma_print_stats(void);
275 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
276 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
278 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
280 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
281 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
283 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
284 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
286 static int zone_warnings = 1;
287 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
288 "Warn when UMA zones becomes full");
291 * This routine checks to see whether or not it's safe to enable buckets.
296 bucketdisable = vm_page_count_min();
300 * Initialize bucket_zones, the array of zones of buckets of various sizes.
302 * For each zone, calculate the memory required for each bucket, consisting
303 * of the header and an array of pointers.
308 struct uma_bucket_zone *ubz;
311 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
312 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
313 size += sizeof(void *) * ubz->ubz_entries;
314 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
315 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
316 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
321 * Given a desired number of entries for a bucket, return the zone from which
322 * to allocate the bucket.
324 static struct uma_bucket_zone *
325 bucket_zone_lookup(int entries)
327 struct uma_bucket_zone *ubz;
329 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
330 if (ubz->ubz_entries >= entries)
337 bucket_select(int size)
339 struct uma_bucket_zone *ubz;
341 ubz = &bucket_zones[0];
342 if (size > ubz->ubz_maxsize)
343 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
345 for (; ubz->ubz_entries != 0; ubz++)
346 if (ubz->ubz_maxsize < size)
349 return (ubz->ubz_entries);
353 bucket_alloc(uma_zone_t zone, void *udata, int flags)
355 struct uma_bucket_zone *ubz;
359 * This is to stop us from allocating per cpu buckets while we're
360 * running out of vm.boot_pages. Otherwise, we would exhaust the
361 * boot pages. This also prevents us from allocating buckets in
362 * low memory situations.
367 * To limit bucket recursion we store the original zone flags
368 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
369 * NOVM flag to persist even through deep recursions. We also
370 * store ZFLAG_BUCKET once we have recursed attempting to allocate
371 * a bucket for a bucket zone so we do not allow infinite bucket
372 * recursion. This cookie will even persist to frees of unused
373 * buckets via the allocation path or bucket allocations in the
376 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
377 udata = (void *)(uintptr_t)zone->uz_flags;
379 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
381 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
383 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
385 ubz = bucket_zone_lookup(zone->uz_count);
386 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
388 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
391 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
394 bucket->ub_entries = ubz->ubz_entries;
401 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
403 struct uma_bucket_zone *ubz;
405 KASSERT(bucket->ub_cnt == 0,
406 ("bucket_free: Freeing a non free bucket."));
407 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
408 udata = (void *)(uintptr_t)zone->uz_flags;
409 ubz = bucket_zone_lookup(bucket->ub_entries);
410 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
414 bucket_zone_drain(void)
416 struct uma_bucket_zone *ubz;
418 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
419 zone_drain(ubz->ubz_zone);
423 zone_log_warning(uma_zone_t zone)
425 static const struct timeval warninterval = { 300, 0 };
427 if (!zone_warnings || zone->uz_warning == NULL)
430 if (ratecheck(&zone->uz_ratecheck, &warninterval))
431 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
435 zone_maxaction(uma_zone_t zone)
437 if (zone->uz_maxaction)
438 (*zone->uz_maxaction)(zone);
442 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
446 LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
447 kegfn(klink->kl_keg);
451 * Routine called by timeout which is used to fire off some time interval
452 * based calculations. (stats, hash size, etc.)
461 uma_timeout(void *unused)
464 zone_foreach(zone_timeout);
466 /* Reschedule this event */
467 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
471 * Routine to perform timeout driven calculations. This expands the
472 * hashes and does per cpu statistics aggregation.
477 keg_timeout(uma_keg_t keg)
482 * Expand the keg hash table.
484 * This is done if the number of slabs is larger than the hash size.
485 * What I'm trying to do here is completely reduce collisions. This
486 * may be a little aggressive. Should I allow for two collisions max?
488 if (keg->uk_flags & UMA_ZONE_HASH &&
489 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
490 struct uma_hash newhash;
491 struct uma_hash oldhash;
495 * This is so involved because allocating and freeing
496 * while the keg lock is held will lead to deadlock.
497 * I have to do everything in stages and check for
500 newhash = keg->uk_hash;
502 ret = hash_alloc(&newhash);
505 if (hash_expand(&keg->uk_hash, &newhash)) {
506 oldhash = keg->uk_hash;
507 keg->uk_hash = newhash;
520 zone_timeout(uma_zone_t zone)
523 zone_foreach_keg(zone, &keg_timeout);
527 * Allocate and zero fill the next sized hash table from the appropriate
531 * hash A new hash structure with the old hash size in uh_hashsize
534 * 1 on sucess and 0 on failure.
537 hash_alloc(struct uma_hash *hash)
542 oldsize = hash->uh_hashsize;
544 /* We're just going to go to a power of two greater */
546 hash->uh_hashsize = oldsize * 2;
547 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
548 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
549 M_UMAHASH, M_NOWAIT);
551 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
552 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
554 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
556 if (hash->uh_slab_hash) {
557 bzero(hash->uh_slab_hash, alloc);
558 hash->uh_hashmask = hash->uh_hashsize - 1;
566 * Expands the hash table for HASH zones. This is done from zone_timeout
567 * to reduce collisions. This must not be done in the regular allocation
568 * path, otherwise, we can recurse on the vm while allocating pages.
571 * oldhash The hash you want to expand
572 * newhash The hash structure for the new table
580 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
586 if (!newhash->uh_slab_hash)
589 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
593 * I need to investigate hash algorithms for resizing without a
597 for (i = 0; i < oldhash->uh_hashsize; i++)
598 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
599 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
600 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
601 hval = UMA_HASH(newhash, slab->us_data);
602 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
610 * Free the hash bucket to the appropriate backing store.
613 * slab_hash The hash bucket we're freeing
614 * hashsize The number of entries in that hash bucket
620 hash_free(struct uma_hash *hash)
622 if (hash->uh_slab_hash == NULL)
624 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
625 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
627 free(hash->uh_slab_hash, M_UMAHASH);
631 * Frees all outstanding items in a bucket
634 * zone The zone to free to, must be unlocked.
635 * bucket The free/alloc bucket with items, cpu queue must be locked.
642 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
650 for (i = 0; i < bucket->ub_cnt; i++)
651 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
652 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
657 * Drains the per cpu caches for a zone.
659 * NOTE: This may only be called while the zone is being turn down, and not
660 * during normal operation. This is necessary in order that we do not have
661 * to migrate CPUs to drain the per-CPU caches.
664 * zone The zone to drain, must be unlocked.
670 cache_drain(uma_zone_t zone)
676 * XXX: It is safe to not lock the per-CPU caches, because we're
677 * tearing down the zone anyway. I.e., there will be no further use
678 * of the caches at this point.
680 * XXX: It would good to be able to assert that the zone is being
681 * torn down to prevent improper use of cache_drain().
683 * XXX: We lock the zone before passing into bucket_cache_drain() as
684 * it is used elsewhere. Should the tear-down path be made special
685 * there in some form?
688 cache = &zone->uz_cpu[cpu];
689 bucket_drain(zone, cache->uc_allocbucket);
690 bucket_drain(zone, cache->uc_freebucket);
691 if (cache->uc_allocbucket != NULL)
692 bucket_free(zone, cache->uc_allocbucket, NULL);
693 if (cache->uc_freebucket != NULL)
694 bucket_free(zone, cache->uc_freebucket, NULL);
695 cache->uc_allocbucket = cache->uc_freebucket = NULL;
698 bucket_cache_drain(zone);
703 cache_shrink(uma_zone_t zone)
706 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
710 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
715 cache_drain_safe_cpu(uma_zone_t zone)
720 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
726 cache = &zone->uz_cpu[curcpu];
727 if (cache->uc_allocbucket) {
728 if (cache->uc_allocbucket->ub_cnt != 0)
729 LIST_INSERT_HEAD(&zone->uz_buckets,
730 cache->uc_allocbucket, ub_link);
732 b1 = cache->uc_allocbucket;
733 cache->uc_allocbucket = NULL;
735 if (cache->uc_freebucket) {
736 if (cache->uc_freebucket->ub_cnt != 0)
737 LIST_INSERT_HEAD(&zone->uz_buckets,
738 cache->uc_freebucket, ub_link);
740 b2 = cache->uc_freebucket;
741 cache->uc_freebucket = NULL;
746 bucket_free(zone, b1, NULL);
748 bucket_free(zone, b2, NULL);
752 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
753 * This is an expensive call because it needs to bind to all CPUs
754 * one by one and enter a critical section on each of them in order
755 * to safely access their cache buckets.
756 * Zone lock must not be held on call this function.
759 cache_drain_safe(uma_zone_t zone)
764 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
769 zone_foreach(cache_shrink);
772 thread_lock(curthread);
773 sched_bind(curthread, cpu);
774 thread_unlock(curthread);
777 cache_drain_safe_cpu(zone);
779 zone_foreach(cache_drain_safe_cpu);
781 thread_lock(curthread);
782 sched_unbind(curthread);
783 thread_unlock(curthread);
787 * Drain the cached buckets from a zone. Expects a locked zone on entry.
790 bucket_cache_drain(uma_zone_t zone)
795 * Drain the bucket queues and free the buckets, we just keep two per
798 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
799 LIST_REMOVE(bucket, ub_link);
801 bucket_drain(zone, bucket);
802 bucket_free(zone, bucket, NULL);
807 * Shrink further bucket sizes. Price of single zone lock collision
808 * is probably lower then price of global cache drain.
810 if (zone->uz_count > zone->uz_count_min)
815 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
822 flags = slab->us_flags;
824 if (keg->uk_fini != NULL) {
825 for (i--; i > -1; i--)
826 keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
829 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
830 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
832 printf("%s: Returning %d bytes.\n", keg->uk_name,
833 PAGE_SIZE * keg->uk_ppera);
835 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
839 * Frees pages from a keg back to the system. This is done on demand from
840 * the pageout daemon.
845 keg_drain(uma_keg_t keg)
847 struct slabhead freeslabs = { 0 };
852 * We don't want to take pages from statically allocated kegs at this
855 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
859 printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
862 if (keg->uk_free == 0)
865 slab = LIST_FIRST(&keg->uk_free_slab);
867 n = LIST_NEXT(slab, us_link);
869 /* We have no where to free these to */
870 if (slab->us_flags & UMA_SLAB_BOOT) {
875 LIST_REMOVE(slab, us_link);
876 keg->uk_pages -= keg->uk_ppera;
877 keg->uk_free -= keg->uk_ipers;
879 if (keg->uk_flags & UMA_ZONE_HASH)
880 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
882 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
889 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
890 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
891 keg_free_slab(keg, slab, keg->uk_ipers);
896 zone_drain_wait(uma_zone_t zone, int waitok)
900 * Set draining to interlock with zone_dtor() so we can release our
901 * locks as we go. Only dtor() should do a WAITOK call since it
902 * is the only call that knows the structure will still be available
906 while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
907 if (waitok == M_NOWAIT)
909 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
911 zone->uz_flags |= UMA_ZFLAG_DRAINING;
912 bucket_cache_drain(zone);
915 * The DRAINING flag protects us from being freed while
916 * we're running. Normally the uma_rwlock would protect us but we
917 * must be able to release and acquire the right lock for each keg.
919 zone_foreach_keg(zone, &keg_drain);
921 zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
928 zone_drain(uma_zone_t zone)
931 zone_drain_wait(zone, M_NOWAIT);
935 * Allocate a new slab for a keg. This does not insert the slab onto a list.
938 * wait Shall we wait?
941 * The slab that was allocated or NULL if there is no memory and the
942 * caller specified M_NOWAIT.
945 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
947 uma_slabrefcnt_t slabref;
954 mtx_assert(&keg->uk_lock, MA_OWNED);
959 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name);
961 allocf = keg->uk_allocf;
964 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
965 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
971 * This reproduces the old vm_zone behavior of zero filling pages the
972 * first time they are added to a zone.
974 * Malloced items are zeroed in uma_zalloc.
977 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
982 if (keg->uk_flags & UMA_ZONE_NODUMP)
985 /* zone is passed for legacy reasons. */
986 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
988 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
989 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
994 /* Point the slab into the allocated memory */
995 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
996 slab = (uma_slab_t )(mem + keg->uk_pgoff);
998 if (keg->uk_flags & UMA_ZONE_VTOSLAB)
999 for (i = 0; i < keg->uk_ppera; i++)
1000 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
1003 slab->us_data = mem;
1004 slab->us_freecount = keg->uk_ipers;
1005 slab->us_flags = flags;
1006 BIT_FILL(SLAB_SETSIZE, &slab->us_free);
1008 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1010 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1011 slabref = (uma_slabrefcnt_t)slab;
1012 for (i = 0; i < keg->uk_ipers; i++)
1013 slabref->us_refcnt[i] = 0;
1016 if (keg->uk_init != NULL) {
1017 for (i = 0; i < keg->uk_ipers; i++)
1018 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1019 keg->uk_size, wait) != 0)
1021 if (i != keg->uk_ipers) {
1022 keg_free_slab(keg, slab, i);
1031 if (keg->uk_flags & UMA_ZONE_HASH)
1032 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1034 keg->uk_pages += keg->uk_ppera;
1035 keg->uk_free += keg->uk_ipers;
1042 * This function is intended to be used early on in place of page_alloc() so
1043 * that we may use the boot time page cache to satisfy allocations before
1047 startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1051 int pages, check_pages;
1053 keg = zone_first_keg(zone);
1054 pages = howmany(bytes, PAGE_SIZE);
1055 check_pages = pages - 1;
1056 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1059 * Check our small startup cache to see if it has pages remaining.
1061 mtx_lock(&uma_boot_pages_mtx);
1063 /* First check if we have enough room. */
1064 tmps = LIST_FIRST(&uma_boot_pages);
1065 while (tmps != NULL && check_pages-- > 0)
1066 tmps = LIST_NEXT(tmps, us_link);
1069 * It's ok to lose tmps references. The last one will
1070 * have tmps->us_data pointing to the start address of
1071 * "pages" contiguous pages of memory.
1073 while (pages-- > 0) {
1074 tmps = LIST_FIRST(&uma_boot_pages);
1075 LIST_REMOVE(tmps, us_link);
1077 mtx_unlock(&uma_boot_pages_mtx);
1078 *pflag = tmps->us_flags;
1079 return (tmps->us_data);
1081 mtx_unlock(&uma_boot_pages_mtx);
1082 if (booted < UMA_STARTUP2)
1083 panic("UMA: Increase vm.boot_pages");
1085 * Now that we've booted reset these users to their real allocator.
1087 #ifdef UMA_MD_SMALL_ALLOC
1088 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1090 keg->uk_allocf = page_alloc;
1092 return keg->uk_allocf(zone, bytes, pflag, wait);
1096 * Allocates a number of pages from the system
1099 * bytes The number of bytes requested
1100 * wait Shall we wait?
1103 * A pointer to the alloced memory or possibly
1104 * NULL if M_NOWAIT is set.
1107 page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1109 void *p; /* Returned page */
1111 *pflag = UMA_SLAB_KMEM;
1112 p = (void *) kmem_malloc(kmem_arena, bytes, wait);
1118 * Allocates a number of pages from within an object
1121 * bytes The number of bytes requested
1122 * wait Shall we wait?
1125 * A pointer to the alloced memory or possibly
1126 * NULL if M_NOWAIT is set.
1129 noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait)
1131 TAILQ_HEAD(, vm_page) alloctail;
1133 vm_offset_t retkva, zkva;
1134 vm_page_t p, p_next;
1137 TAILQ_INIT(&alloctail);
1138 keg = zone_first_keg(zone);
1140 npages = howmany(bytes, PAGE_SIZE);
1141 while (npages > 0) {
1142 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1143 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1146 * Since the page does not belong to an object, its
1149 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1153 if (wait & M_WAITOK) {
1159 * Page allocation failed, free intermediate pages and
1162 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1163 vm_page_unwire(p, PQ_NONE);
1168 *flags = UMA_SLAB_PRIV;
1169 zkva = keg->uk_kva +
1170 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1172 TAILQ_FOREACH(p, &alloctail, listq) {
1173 pmap_qenter(zkva, &p, 1);
1177 return ((void *)retkva);
1181 * Frees a number of pages to the system
1184 * mem A pointer to the memory to be freed
1185 * size The size of the memory being freed
1186 * flags The original p->us_flags field
1192 page_free(void *mem, vm_size_t size, uint8_t flags)
1196 if (flags & UMA_SLAB_KMEM)
1198 else if (flags & UMA_SLAB_KERNEL)
1199 vmem = kernel_arena;
1201 panic("UMA: page_free used with invalid flags %d", flags);
1203 kmem_free(vmem, (vm_offset_t)mem, size);
1207 * Zero fill initializer
1209 * Arguments/Returns follow uma_init specifications
1212 zero_init(void *mem, int size, int flags)
1219 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1222 * keg The zone we should initialize
1228 keg_small_init(uma_keg_t keg)
1235 if (keg->uk_flags & UMA_ZONE_PCPU) {
1236 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1238 keg->uk_slabsize = sizeof(struct pcpu);
1239 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1242 keg->uk_slabsize = UMA_SLAB_SIZE;
1247 * Calculate the size of each allocation (rsize) according to
1248 * alignment. If the requested size is smaller than we have
1249 * allocation bits for we round it up.
1251 rsize = keg->uk_size;
1252 if (rsize < keg->uk_slabsize / SLAB_SETSIZE)
1253 rsize = keg->uk_slabsize / SLAB_SETSIZE;
1254 if (rsize & keg->uk_align)
1255 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1256 keg->uk_rsize = rsize;
1258 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1259 keg->uk_rsize < sizeof(struct pcpu),
1260 ("%s: size %u too large", __func__, keg->uk_rsize));
1262 if (keg->uk_flags & UMA_ZONE_REFCNT)
1263 rsize += sizeof(uint32_t);
1265 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1268 shsize = sizeof(struct uma_slab);
1270 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1271 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1272 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1274 memused = keg->uk_ipers * rsize + shsize;
1275 wastedspace = keg->uk_slabsize - memused;
1278 * We can't do OFFPAGE if we're internal or if we've been
1279 * asked to not go to the VM for buckets. If we do this we
1280 * may end up going to the VM for slabs which we do not
1281 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1282 * of UMA_ZONE_VM, which clearly forbids it.
1284 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1285 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1289 * See if using an OFFPAGE slab will limit our waste. Only do
1290 * this if it permits more items per-slab.
1292 * XXX We could try growing slabsize to limit max waste as well.
1293 * Historically this was not done because the VM could not
1294 * efficiently handle contiguous allocations.
1296 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1297 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1298 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1299 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1300 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1302 printf("UMA decided we need offpage slab headers for "
1303 "keg: %s, calculated wastedspace = %d, "
1304 "maximum wasted space allowed = %d, "
1305 "calculated ipers = %d, "
1306 "new wasted space = %d\n", keg->uk_name, wastedspace,
1307 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1308 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1310 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1313 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1314 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1315 keg->uk_flags |= UMA_ZONE_HASH;
1319 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1320 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1324 * keg The keg we should initialize
1330 keg_large_init(uma_keg_t keg)
1334 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1335 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1336 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1337 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1338 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1340 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1341 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1343 keg->uk_rsize = keg->uk_size;
1345 /* We can't do OFFPAGE if we're internal, bail out here. */
1346 if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1349 /* Check whether we have enough space to not do OFFPAGE. */
1350 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1351 shsize = sizeof(struct uma_slab);
1352 if (keg->uk_flags & UMA_ZONE_REFCNT)
1353 shsize += keg->uk_ipers * sizeof(uint32_t);
1354 if (shsize & UMA_ALIGN_PTR)
1355 shsize = (shsize & ~UMA_ALIGN_PTR) +
1356 (UMA_ALIGN_PTR + 1);
1358 if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize)
1359 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1362 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1363 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1364 keg->uk_flags |= UMA_ZONE_HASH;
1368 keg_cachespread_init(uma_keg_t keg)
1375 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1376 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1378 alignsize = keg->uk_align + 1;
1379 rsize = keg->uk_size;
1381 * We want one item to start on every align boundary in a page. To
1382 * do this we will span pages. We will also extend the item by the
1383 * size of align if it is an even multiple of align. Otherwise, it
1384 * would fall on the same boundary every time.
1386 if (rsize & keg->uk_align)
1387 rsize = (rsize & ~keg->uk_align) + alignsize;
1388 if ((rsize & alignsize) == 0)
1390 trailer = rsize - keg->uk_size;
1391 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1392 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1393 keg->uk_rsize = rsize;
1394 keg->uk_ppera = pages;
1395 keg->uk_slabsize = UMA_SLAB_SIZE;
1396 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1397 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1398 KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1399 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1404 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1405 * the keg onto the global keg list.
1407 * Arguments/Returns follow uma_ctor specifications
1408 * udata Actually uma_kctor_args
1411 keg_ctor(void *mem, int size, void *udata, int flags)
1413 struct uma_kctor_args *arg = udata;
1414 uma_keg_t keg = mem;
1418 keg->uk_size = arg->size;
1419 keg->uk_init = arg->uminit;
1420 keg->uk_fini = arg->fini;
1421 keg->uk_align = arg->align;
1423 keg->uk_reserve = 0;
1425 keg->uk_flags = arg->flags;
1426 keg->uk_allocf = page_alloc;
1427 keg->uk_freef = page_free;
1428 keg->uk_slabzone = NULL;
1431 * The master zone is passed to us at keg-creation time.
1434 keg->uk_name = zone->uz_name;
1436 if (arg->flags & UMA_ZONE_VM)
1437 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1439 if (arg->flags & UMA_ZONE_ZINIT)
1440 keg->uk_init = zero_init;
1442 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1443 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1445 if (arg->flags & UMA_ZONE_PCPU)
1447 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1449 keg->uk_flags &= ~UMA_ZONE_PCPU;
1452 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1453 keg_cachespread_init(keg);
1454 } else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1456 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) -
1458 keg_large_init(keg);
1460 keg_small_init(keg);
1462 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1463 keg_large_init(keg);
1465 keg_small_init(keg);
1468 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1469 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1470 if (keg->uk_ipers > uma_max_ipers_ref)
1471 panic("Too many ref items per zone: %d > %d\n",
1472 keg->uk_ipers, uma_max_ipers_ref);
1473 keg->uk_slabzone = slabrefzone;
1475 keg->uk_slabzone = slabzone;
1479 * If we haven't booted yet we need allocations to go through the
1480 * startup cache until the vm is ready.
1482 if (keg->uk_ppera == 1) {
1483 #ifdef UMA_MD_SMALL_ALLOC
1484 keg->uk_allocf = uma_small_alloc;
1485 keg->uk_freef = uma_small_free;
1487 if (booted < UMA_STARTUP)
1488 keg->uk_allocf = startup_alloc;
1490 if (booted < UMA_STARTUP2)
1491 keg->uk_allocf = startup_alloc;
1493 } else if (booted < UMA_STARTUP2 &&
1494 (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1495 keg->uk_allocf = startup_alloc;
1498 * Initialize keg's lock
1500 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1503 * If we're putting the slab header in the actual page we need to
1504 * figure out where in each page it goes. This calculates a right
1505 * justified offset into the memory on an ALIGN_PTR boundary.
1507 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1510 /* Size of the slab struct and free list */
1511 totsize = sizeof(struct uma_slab);
1513 /* Size of the reference counts. */
1514 if (keg->uk_flags & UMA_ZONE_REFCNT)
1515 totsize += keg->uk_ipers * sizeof(uint32_t);
1517 if (totsize & UMA_ALIGN_PTR)
1518 totsize = (totsize & ~UMA_ALIGN_PTR) +
1519 (UMA_ALIGN_PTR + 1);
1520 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1523 * The only way the following is possible is if with our
1524 * UMA_ALIGN_PTR adjustments we are now bigger than
1525 * UMA_SLAB_SIZE. I haven't checked whether this is
1526 * mathematically possible for all cases, so we make
1529 totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1530 if (keg->uk_flags & UMA_ZONE_REFCNT)
1531 totsize += keg->uk_ipers * sizeof(uint32_t);
1532 if (totsize > PAGE_SIZE * keg->uk_ppera) {
1533 printf("zone %s ipers %d rsize %d size %d\n",
1534 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1536 panic("UMA slab won't fit.");
1540 if (keg->uk_flags & UMA_ZONE_HASH)
1541 hash_alloc(&keg->uk_hash);
1544 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1545 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1546 keg->uk_ipers, keg->uk_ppera,
1547 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1550 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1552 rw_wlock(&uma_rwlock);
1553 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1554 rw_wunlock(&uma_rwlock);
1559 * Zone header ctor. This initializes all fields, locks, etc.
1561 * Arguments/Returns follow uma_ctor specifications
1562 * udata Actually uma_zctor_args
1565 zone_ctor(void *mem, int size, void *udata, int flags)
1567 struct uma_zctor_args *arg = udata;
1568 uma_zone_t zone = mem;
1573 zone->uz_name = arg->name;
1574 zone->uz_ctor = arg->ctor;
1575 zone->uz_dtor = arg->dtor;
1576 zone->uz_slab = zone_fetch_slab;
1577 zone->uz_init = NULL;
1578 zone->uz_fini = NULL;
1579 zone->uz_allocs = 0;
1582 zone->uz_sleeps = 0;
1584 zone->uz_count_min = 0;
1586 zone->uz_warning = NULL;
1587 timevalclear(&zone->uz_ratecheck);
1588 zone->uz_maxaction = NULL;
1591 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1594 * This is a pure cache zone, no kegs.
1597 if (arg->flags & UMA_ZONE_VM)
1598 arg->flags |= UMA_ZFLAG_CACHEONLY;
1599 zone->uz_flags = arg->flags;
1600 zone->uz_size = arg->size;
1601 zone->uz_import = arg->import;
1602 zone->uz_release = arg->release;
1603 zone->uz_arg = arg->arg;
1604 zone->uz_lockptr = &zone->uz_lock;
1605 rw_wlock(&uma_rwlock);
1606 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1607 rw_wunlock(&uma_rwlock);
1612 * Use the regular zone/keg/slab allocator.
1614 zone->uz_import = (uma_import)zone_import;
1615 zone->uz_release = (uma_release)zone_release;
1616 zone->uz_arg = zone;
1618 if (arg->flags & UMA_ZONE_SECONDARY) {
1619 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1620 zone->uz_init = arg->uminit;
1621 zone->uz_fini = arg->fini;
1622 zone->uz_lockptr = &keg->uk_lock;
1623 zone->uz_flags |= UMA_ZONE_SECONDARY;
1624 rw_wlock(&uma_rwlock);
1626 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1627 if (LIST_NEXT(z, uz_link) == NULL) {
1628 LIST_INSERT_AFTER(z, zone, uz_link);
1633 rw_wunlock(&uma_rwlock);
1634 } else if (keg == NULL) {
1635 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1636 arg->align, arg->flags)) == NULL)
1639 struct uma_kctor_args karg;
1642 /* We should only be here from uma_startup() */
1643 karg.size = arg->size;
1644 karg.uminit = arg->uminit;
1645 karg.fini = arg->fini;
1646 karg.align = arg->align;
1647 karg.flags = arg->flags;
1649 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1656 * Link in the first keg.
1658 zone->uz_klink.kl_keg = keg;
1659 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1660 zone->uz_lockptr = &keg->uk_lock;
1661 zone->uz_size = keg->uk_size;
1662 zone->uz_flags |= (keg->uk_flags &
1663 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1666 * Some internal zones don't have room allocated for the per cpu
1667 * caches. If we're internal, bail out here.
1669 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1670 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1671 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1676 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1677 zone->uz_count = bucket_select(zone->uz_size);
1679 zone->uz_count = BUCKET_MAX;
1680 zone->uz_count_min = zone->uz_count;
1686 * Keg header dtor. This frees all data, destroys locks, frees the hash
1687 * table and removes the keg from the global list.
1689 * Arguments/Returns follow uma_dtor specifications
1693 keg_dtor(void *arg, int size, void *udata)
1697 keg = (uma_keg_t)arg;
1699 if (keg->uk_free != 0) {
1700 printf("Freed UMA keg (%s) was not empty (%d items). "
1701 " Lost %d pages of memory.\n",
1702 keg->uk_name ? keg->uk_name : "",
1703 keg->uk_free, keg->uk_pages);
1707 hash_free(&keg->uk_hash);
1715 * Arguments/Returns follow uma_dtor specifications
1719 zone_dtor(void *arg, int size, void *udata)
1725 zone = (uma_zone_t)arg;
1726 keg = zone_first_keg(zone);
1728 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1731 rw_wlock(&uma_rwlock);
1732 LIST_REMOVE(zone, uz_link);
1733 rw_wunlock(&uma_rwlock);
1735 * XXX there are some races here where
1736 * the zone can be drained but zone lock
1737 * released and then refilled before we
1738 * remove it... we dont care for now
1740 zone_drain_wait(zone, M_WAITOK);
1742 * Unlink all of our kegs.
1744 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1745 klink->kl_keg = NULL;
1746 LIST_REMOVE(klink, kl_link);
1747 if (klink == &zone->uz_klink)
1749 free(klink, M_TEMP);
1752 * We only destroy kegs from non secondary zones.
1754 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1755 rw_wlock(&uma_rwlock);
1756 LIST_REMOVE(keg, uk_link);
1757 rw_wunlock(&uma_rwlock);
1758 zone_free_item(kegs, keg, NULL, SKIP_NONE);
1760 ZONE_LOCK_FINI(zone);
1764 * Traverses every zone in the system and calls a callback
1767 * zfunc A pointer to a function which accepts a zone
1774 zone_foreach(void (*zfunc)(uma_zone_t))
1779 rw_rlock(&uma_rwlock);
1780 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1781 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1784 rw_runlock(&uma_rwlock);
1787 /* Public functions */
1790 uma_startup(void *bootmem, int boot_pages)
1792 struct uma_zctor_args args;
1798 printf("Creating uma keg headers zone and keg.\n");
1800 rw_init(&uma_rwlock, "UMA lock");
1802 /* "manually" create the initial zone */
1803 memset(&args, 0, sizeof(args));
1804 args.name = "UMA Kegs";
1805 args.size = sizeof(struct uma_keg);
1806 args.ctor = keg_ctor;
1807 args.dtor = keg_dtor;
1808 args.uminit = zero_init;
1810 args.keg = &masterkeg;
1811 args.align = 32 - 1;
1812 args.flags = UMA_ZFLAG_INTERNAL;
1813 /* The initial zone has no Per cpu queues so it's smaller */
1814 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1817 printf("Filling boot free list.\n");
1819 for (i = 0; i < boot_pages; i++) {
1820 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1821 slab->us_data = (uint8_t *)slab;
1822 slab->us_flags = UMA_SLAB_BOOT;
1823 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1825 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1828 printf("Creating uma zone headers zone and keg.\n");
1830 args.name = "UMA Zones";
1831 args.size = sizeof(struct uma_zone) +
1832 (sizeof(struct uma_cache) * (mp_maxid + 1));
1833 args.ctor = zone_ctor;
1834 args.dtor = zone_dtor;
1835 args.uminit = zero_init;
1838 args.align = 32 - 1;
1839 args.flags = UMA_ZFLAG_INTERNAL;
1840 /* The initial zone has no Per cpu queues so it's smaller */
1841 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1844 printf("Creating slab and hash zones.\n");
1847 /* Now make a zone for slab headers */
1848 slabzone = uma_zcreate("UMA Slabs",
1849 sizeof(struct uma_slab),
1850 NULL, NULL, NULL, NULL,
1851 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1854 * We also create a zone for the bigger slabs with reference
1855 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1857 slabsize = sizeof(struct uma_slab_refcnt);
1858 slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1859 slabrefzone = uma_zcreate("UMA RCntSlabs",
1861 NULL, NULL, NULL, NULL,
1863 UMA_ZFLAG_INTERNAL);
1865 hashzone = uma_zcreate("UMA Hash",
1866 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1867 NULL, NULL, NULL, NULL,
1868 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1872 booted = UMA_STARTUP;
1875 printf("UMA startup complete.\n");
1883 booted = UMA_STARTUP2;
1885 sx_init(&uma_drain_lock, "umadrain");
1887 printf("UMA startup2 complete.\n");
1892 * Initialize our callout handle
1900 printf("Starting callout.\n");
1902 callout_init(&uma_callout, 1);
1903 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1905 printf("UMA startup3 complete.\n");
1910 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1911 int align, uint32_t flags)
1913 struct uma_kctor_args args;
1916 args.uminit = uminit;
1918 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1921 return (zone_alloc_item(kegs, &args, M_WAITOK));
1926 uma_set_align(int align)
1929 if (align != UMA_ALIGN_CACHE)
1930 uma_align_cache = align;
1935 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1936 uma_init uminit, uma_fini fini, int align, uint32_t flags)
1939 struct uma_zctor_args args;
1943 /* This stuff is essential for the zone ctor */
1944 memset(&args, 0, sizeof(args));
1949 args.uminit = uminit;
1953 * If a zone is being created with an empty constructor and
1954 * destructor, pass UMA constructor/destructor which checks for
1955 * memory use after free.
1957 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) &&
1958 ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) {
1959 args.ctor = trash_ctor;
1960 args.dtor = trash_dtor;
1961 args.uminit = trash_init;
1962 args.fini = trash_fini;
1969 if (booted < UMA_STARTUP2) {
1972 sx_slock(&uma_drain_lock);
1975 res = zone_alloc_item(zones, &args, M_WAITOK);
1977 sx_sunlock(&uma_drain_lock);
1983 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1984 uma_init zinit, uma_fini zfini, uma_zone_t master)
1986 struct uma_zctor_args args;
1991 keg = zone_first_keg(master);
1992 memset(&args, 0, sizeof(args));
1994 args.size = keg->uk_size;
1997 args.uminit = zinit;
1999 args.align = keg->uk_align;
2000 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
2003 if (booted < UMA_STARTUP2) {
2006 sx_slock(&uma_drain_lock);
2009 /* XXX Attaches only one keg of potentially many. */
2010 res = zone_alloc_item(zones, &args, M_WAITOK);
2012 sx_sunlock(&uma_drain_lock);
2018 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
2019 uma_init zinit, uma_fini zfini, uma_import zimport,
2020 uma_release zrelease, void *arg, int flags)
2022 struct uma_zctor_args args;
2024 memset(&args, 0, sizeof(args));
2029 args.uminit = zinit;
2031 args.import = zimport;
2032 args.release = zrelease;
2037 return (zone_alloc_item(zones, &args, M_WAITOK));
2041 zone_lock_pair(uma_zone_t a, uma_zone_t b)
2045 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
2048 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
2053 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2061 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2068 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2070 zone_lock_pair(zone, master);
2072 * zone must use vtoslab() to resolve objects and must already be
2075 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2076 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2081 * The new master must also use vtoslab().
2083 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2088 * Both must either be refcnt, or not be refcnt.
2090 if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
2091 (master->uz_flags & UMA_ZONE_REFCNT)) {
2096 * The underlying object must be the same size. rsize
2099 if (master->uz_size != zone->uz_size) {
2104 * Put it at the end of the list.
2106 klink->kl_keg = zone_first_keg(master);
2107 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2108 if (LIST_NEXT(kl, kl_link) == NULL) {
2109 LIST_INSERT_AFTER(kl, klink, kl_link);
2114 zone->uz_flags |= UMA_ZFLAG_MULTI;
2115 zone->uz_slab = zone_fetch_slab_multi;
2118 zone_unlock_pair(zone, master);
2120 free(klink, M_TEMP);
2128 uma_zdestroy(uma_zone_t zone)
2131 sx_slock(&uma_drain_lock);
2132 zone_free_item(zones, zone, NULL, SKIP_NONE);
2133 sx_sunlock(&uma_drain_lock);
2138 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2142 uma_bucket_t bucket;
2146 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2147 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2149 /* This is the fast path allocation */
2150 #ifdef UMA_DEBUG_ALLOC_1
2151 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2153 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2154 zone->uz_name, flags);
2156 if (flags & M_WAITOK) {
2157 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2158 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2160 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2161 ("uma_zalloc_arg: called with spinlock or critical section held"));
2163 #ifdef DEBUG_MEMGUARD
2164 if (memguard_cmp_zone(zone)) {
2165 item = memguard_alloc(zone->uz_size, flags);
2168 * Avoid conflict with the use-after-free
2169 * protecting infrastructure from INVARIANTS.
2171 if (zone->uz_init != NULL &&
2172 zone->uz_init != mtrash_init &&
2173 zone->uz_init(item, zone->uz_size, flags) != 0)
2175 if (zone->uz_ctor != NULL &&
2176 zone->uz_ctor != mtrash_ctor &&
2177 zone->uz_ctor(item, zone->uz_size, udata,
2179 zone->uz_fini(item, zone->uz_size);
2184 /* This is unfortunate but should not be fatal. */
2188 * If possible, allocate from the per-CPU cache. There are two
2189 * requirements for safe access to the per-CPU cache: (1) the thread
2190 * accessing the cache must not be preempted or yield during access,
2191 * and (2) the thread must not migrate CPUs without switching which
2192 * cache it accesses. We rely on a critical section to prevent
2193 * preemption and migration. We release the critical section in
2194 * order to acquire the zone mutex if we are unable to allocate from
2195 * the current cache; when we re-acquire the critical section, we
2196 * must detect and handle migration if it has occurred.
2200 cache = &zone->uz_cpu[cpu];
2203 bucket = cache->uc_allocbucket;
2204 if (bucket != NULL && bucket->ub_cnt > 0) {
2206 item = bucket->ub_bucket[bucket->ub_cnt];
2208 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2210 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2213 if (zone->uz_ctor != NULL &&
2214 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2215 atomic_add_long(&zone->uz_fails, 1);
2216 zone_free_item(zone, item, udata, SKIP_DTOR);
2220 uma_dbg_alloc(zone, NULL, item);
2223 uma_zero_item(item, zone);
2228 * We have run out of items in our alloc bucket.
2229 * See if we can switch with our free bucket.
2231 bucket = cache->uc_freebucket;
2232 if (bucket != NULL && bucket->ub_cnt > 0) {
2233 #ifdef UMA_DEBUG_ALLOC
2234 printf("uma_zalloc: Swapping empty with alloc.\n");
2236 cache->uc_freebucket = cache->uc_allocbucket;
2237 cache->uc_allocbucket = bucket;
2242 * Discard any empty allocation bucket while we hold no locks.
2244 bucket = cache->uc_allocbucket;
2245 cache->uc_allocbucket = NULL;
2248 bucket_free(zone, bucket, udata);
2250 /* Short-circuit for zones without buckets and low memory. */
2251 if (zone->uz_count == 0 || bucketdisable)
2255 * Attempt to retrieve the item from the per-CPU cache has failed, so
2256 * we must go back to the zone. This requires the zone lock, so we
2257 * must drop the critical section, then re-acquire it when we go back
2258 * to the cache. Since the critical section is released, we may be
2259 * preempted or migrate. As such, make sure not to maintain any
2260 * thread-local state specific to the cache from prior to releasing
2261 * the critical section.
2264 if (ZONE_TRYLOCK(zone) == 0) {
2265 /* Record contention to size the buckets. */
2271 cache = &zone->uz_cpu[cpu];
2274 * Since we have locked the zone we may as well send back our stats.
2276 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2277 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2278 cache->uc_allocs = 0;
2279 cache->uc_frees = 0;
2281 /* See if we lost the race to fill the cache. */
2282 if (cache->uc_allocbucket != NULL) {
2288 * Check the zone's cache of buckets.
2290 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2291 KASSERT(bucket->ub_cnt != 0,
2292 ("uma_zalloc_arg: Returning an empty bucket."));
2294 LIST_REMOVE(bucket, ub_link);
2295 cache->uc_allocbucket = bucket;
2299 /* We are no longer associated with this CPU. */
2303 * We bump the uz count when the cache size is insufficient to
2304 * handle the working set.
2306 if (lockfail && zone->uz_count < BUCKET_MAX)
2311 * Now lets just fill a bucket and put it on the free list. If that
2312 * works we'll restart the allocation from the begining and it
2313 * will use the just filled bucket.
2315 bucket = zone_alloc_bucket(zone, udata, flags);
2316 if (bucket != NULL) {
2320 cache = &zone->uz_cpu[cpu];
2322 * See if we lost the race or were migrated. Cache the
2323 * initialized bucket to make this less likely or claim
2324 * the memory directly.
2326 if (cache->uc_allocbucket == NULL)
2327 cache->uc_allocbucket = bucket;
2329 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2335 * We may not be able to get a bucket so return an actual item.
2338 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2342 item = zone_alloc_item(zone, udata, flags);
2348 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2353 mtx_assert(&keg->uk_lock, MA_OWNED);
2356 if ((flags & M_USE_RESERVE) == 0)
2357 reserve = keg->uk_reserve;
2361 * Find a slab with some space. Prefer slabs that are partially
2362 * used over those that are totally full. This helps to reduce
2365 if (keg->uk_free > reserve) {
2366 if (!LIST_EMPTY(&keg->uk_part_slab)) {
2367 slab = LIST_FIRST(&keg->uk_part_slab);
2369 slab = LIST_FIRST(&keg->uk_free_slab);
2370 LIST_REMOVE(slab, us_link);
2371 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2374 MPASS(slab->us_keg == keg);
2379 * M_NOVM means don't ask at all!
2384 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2385 keg->uk_flags |= UMA_ZFLAG_FULL;
2387 * If this is not a multi-zone, set the FULL bit.
2388 * Otherwise slab_multi() takes care of it.
2390 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2391 zone->uz_flags |= UMA_ZFLAG_FULL;
2392 zone_log_warning(zone);
2393 zone_maxaction(zone);
2395 if (flags & M_NOWAIT)
2398 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2401 slab = keg_alloc_slab(keg, zone, flags);
2403 * If we got a slab here it's safe to mark it partially used
2404 * and return. We assume that the caller is going to remove
2405 * at least one item.
2408 MPASS(slab->us_keg == keg);
2409 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2413 * We might not have been able to get a slab but another cpu
2414 * could have while we were unlocked. Check again before we
2423 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2428 keg = zone_first_keg(zone);
2433 slab = keg_fetch_slab(keg, zone, flags);
2436 if (flags & (M_NOWAIT | M_NOVM))
2444 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2445 * with the keg locked. On NULL no lock is held.
2447 * The last pointer is used to seed the search. It is not required.
2450 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2460 * Don't wait on the first pass. This will skip limit tests
2461 * as well. We don't want to block if we can find a provider
2464 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2466 * Use the last slab allocated as a hint for where to start
2470 slab = keg_fetch_slab(last, zone, flags);
2476 * Loop until we have a slab incase of transient failures
2477 * while M_WAITOK is specified. I'm not sure this is 100%
2478 * required but we've done it for so long now.
2484 * Search the available kegs for slabs. Be careful to hold the
2485 * correct lock while calling into the keg layer.
2487 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2488 keg = klink->kl_keg;
2490 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2491 slab = keg_fetch_slab(keg, zone, flags);
2495 if (keg->uk_flags & UMA_ZFLAG_FULL)
2501 if (rflags & (M_NOWAIT | M_NOVM))
2505 * All kegs are full. XXX We can't atomically check all kegs
2506 * and sleep so just sleep for a short period and retry.
2508 if (full && !empty) {
2510 zone->uz_flags |= UMA_ZFLAG_FULL;
2512 zone_log_warning(zone);
2513 zone_maxaction(zone);
2514 msleep(zone, zone->uz_lockptr, PVM,
2515 "zonelimit", hz/100);
2516 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2525 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2530 MPASS(keg == slab->us_keg);
2531 mtx_assert(&keg->uk_lock, MA_OWNED);
2533 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2534 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2535 item = slab->us_data + (keg->uk_rsize * freei);
2536 slab->us_freecount--;
2539 /* Move this slab to the full list */
2540 if (slab->us_freecount == 0) {
2541 LIST_REMOVE(slab, us_link);
2542 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2549 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2557 /* Try to keep the buckets totally full */
2558 for (i = 0; i < max; ) {
2559 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2562 while (slab->us_freecount && i < max) {
2563 bucket[i++] = slab_alloc_item(keg, slab);
2564 if (keg->uk_free <= keg->uk_reserve)
2567 /* Don't grab more than one slab at a time. */
2578 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2580 uma_bucket_t bucket;
2583 /* Don't wait for buckets, preserve caller's NOVM setting. */
2584 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2588 max = MIN(bucket->ub_entries, zone->uz_count);
2589 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2593 * Initialize the memory if necessary.
2595 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2598 for (i = 0; i < bucket->ub_cnt; i++)
2599 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2603 * If we couldn't initialize the whole bucket, put the
2604 * rest back onto the freelist.
2606 if (i != bucket->ub_cnt) {
2607 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2608 bucket->ub_cnt - i);
2610 bzero(&bucket->ub_bucket[i],
2611 sizeof(void *) * (bucket->ub_cnt - i));
2617 if (bucket->ub_cnt == 0) {
2618 bucket_free(zone, bucket, udata);
2619 atomic_add_long(&zone->uz_fails, 1);
2627 * Allocates a single item from a zone.
2630 * zone The zone to alloc for.
2631 * udata The data to be passed to the constructor.
2632 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2635 * NULL if there is no memory and M_NOWAIT is set
2636 * An item if successful
2640 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2646 #ifdef UMA_DEBUG_ALLOC
2647 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2649 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2651 atomic_add_long(&zone->uz_allocs, 1);
2654 * We have to call both the zone's init (not the keg's init)
2655 * and the zone's ctor. This is because the item is going from
2656 * a keg slab directly to the user, and the user is expecting it
2657 * to be both zone-init'd as well as zone-ctor'd.
2659 if (zone->uz_init != NULL) {
2660 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2661 zone_free_item(zone, item, udata, SKIP_FINI);
2665 if (zone->uz_ctor != NULL) {
2666 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2667 zone_free_item(zone, item, udata, SKIP_DTOR);
2672 uma_dbg_alloc(zone, NULL, item);
2675 uma_zero_item(item, zone);
2680 atomic_add_long(&zone->uz_fails, 1);
2686 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2689 uma_bucket_t bucket;
2693 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2694 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2696 #ifdef UMA_DEBUG_ALLOC_1
2697 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2699 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2702 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2703 ("uma_zfree_arg: called with spinlock or critical section held"));
2705 /* uma_zfree(..., NULL) does nothing, to match free(9). */
2708 #ifdef DEBUG_MEMGUARD
2709 if (is_memguard_addr(item)) {
2710 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2711 zone->uz_dtor(item, zone->uz_size, udata);
2712 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2713 zone->uz_fini(item, zone->uz_size);
2714 memguard_free(item);
2719 if (zone->uz_flags & UMA_ZONE_MALLOC)
2720 uma_dbg_free(zone, udata, item);
2722 uma_dbg_free(zone, NULL, item);
2724 if (zone->uz_dtor != NULL)
2725 zone->uz_dtor(item, zone->uz_size, udata);
2728 * The race here is acceptable. If we miss it we'll just have to wait
2729 * a little longer for the limits to be reset.
2731 if (zone->uz_flags & UMA_ZFLAG_FULL)
2735 * If possible, free to the per-CPU cache. There are two
2736 * requirements for safe access to the per-CPU cache: (1) the thread
2737 * accessing the cache must not be preempted or yield during access,
2738 * and (2) the thread must not migrate CPUs without switching which
2739 * cache it accesses. We rely on a critical section to prevent
2740 * preemption and migration. We release the critical section in
2741 * order to acquire the zone mutex if we are unable to free to the
2742 * current cache; when we re-acquire the critical section, we must
2743 * detect and handle migration if it has occurred.
2748 cache = &zone->uz_cpu[cpu];
2752 * Try to free into the allocbucket first to give LIFO ordering
2753 * for cache-hot datastructures. Spill over into the freebucket
2754 * if necessary. Alloc will swap them if one runs dry.
2756 bucket = cache->uc_allocbucket;
2757 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2758 bucket = cache->uc_freebucket;
2759 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2760 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2761 ("uma_zfree: Freeing to non free bucket index."));
2762 bucket->ub_bucket[bucket->ub_cnt] = item;
2770 * We must go back the zone, which requires acquiring the zone lock,
2771 * which in turn means we must release and re-acquire the critical
2772 * section. Since the critical section is released, we may be
2773 * preempted or migrate. As such, make sure not to maintain any
2774 * thread-local state specific to the cache from prior to releasing
2775 * the critical section.
2778 if (zone->uz_count == 0 || bucketdisable)
2782 if (ZONE_TRYLOCK(zone) == 0) {
2783 /* Record contention to size the buckets. */
2789 cache = &zone->uz_cpu[cpu];
2792 * Since we have locked the zone we may as well send back our stats.
2794 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2795 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2796 cache->uc_allocs = 0;
2797 cache->uc_frees = 0;
2799 bucket = cache->uc_freebucket;
2800 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2804 cache->uc_freebucket = NULL;
2806 /* Can we throw this on the zone full list? */
2807 if (bucket != NULL) {
2808 #ifdef UMA_DEBUG_ALLOC
2809 printf("uma_zfree: Putting old bucket on the free list.\n");
2811 /* ub_cnt is pointing to the last free item */
2812 KASSERT(bucket->ub_cnt != 0,
2813 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2814 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2817 /* We are no longer associated with this CPU. */
2821 * We bump the uz count when the cache size is insufficient to
2822 * handle the working set.
2824 if (lockfail && zone->uz_count < BUCKET_MAX)
2828 #ifdef UMA_DEBUG_ALLOC
2829 printf("uma_zfree: Allocating new free bucket.\n");
2831 bucket = bucket_alloc(zone, udata, M_NOWAIT);
2835 cache = &zone->uz_cpu[cpu];
2836 if (cache->uc_freebucket == NULL) {
2837 cache->uc_freebucket = bucket;
2841 * We lost the race, start over. We have to drop our
2842 * critical section to free the bucket.
2845 bucket_free(zone, bucket, udata);
2850 * If nothing else caught this, we'll just do an internal free.
2853 zone_free_item(zone, item, udata, SKIP_DTOR);
2859 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2863 mtx_assert(&keg->uk_lock, MA_OWNED);
2864 MPASS(keg == slab->us_keg);
2866 /* Do we need to remove from any lists? */
2867 if (slab->us_freecount+1 == keg->uk_ipers) {
2868 LIST_REMOVE(slab, us_link);
2869 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2870 } else if (slab->us_freecount == 0) {
2871 LIST_REMOVE(slab, us_link);
2872 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2875 /* Slab management. */
2876 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2877 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2878 slab->us_freecount++;
2880 /* Keg statistics. */
2885 zone_release(uma_zone_t zone, void **bucket, int cnt)
2895 keg = zone_first_keg(zone);
2897 for (i = 0; i < cnt; i++) {
2899 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2900 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2901 if (zone->uz_flags & UMA_ZONE_HASH) {
2902 slab = hash_sfind(&keg->uk_hash, mem);
2904 mem += keg->uk_pgoff;
2905 slab = (uma_slab_t)mem;
2908 slab = vtoslab((vm_offset_t)item);
2909 if (slab->us_keg != keg) {
2915 slab_free_item(keg, slab, item);
2916 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2917 if (keg->uk_pages < keg->uk_maxpages) {
2918 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2923 * We can handle one more allocation. Since we're
2924 * clearing ZFLAG_FULL, wake up all procs blocked
2925 * on pages. This should be uncommon, so keeping this
2926 * simple for now (rather than adding count of blocked
2935 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2943 * Frees a single item to any zone.
2946 * zone The zone to free to
2947 * item The item we're freeing
2948 * udata User supplied data for the dtor
2949 * skip Skip dtors and finis
2952 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2956 if (skip == SKIP_NONE) {
2957 if (zone->uz_flags & UMA_ZONE_MALLOC)
2958 uma_dbg_free(zone, udata, item);
2960 uma_dbg_free(zone, NULL, item);
2963 if (skip < SKIP_DTOR && zone->uz_dtor)
2964 zone->uz_dtor(item, zone->uz_size, udata);
2966 if (skip < SKIP_FINI && zone->uz_fini)
2967 zone->uz_fini(item, zone->uz_size);
2969 atomic_add_long(&zone->uz_frees, 1);
2970 zone->uz_release(zone->uz_arg, &item, 1);
2975 uma_zone_set_max(uma_zone_t zone, int nitems)
2979 keg = zone_first_keg(zone);
2983 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2984 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2985 keg->uk_maxpages += keg->uk_ppera;
2986 nitems = keg->uk_maxpages * keg->uk_ipers;
2994 uma_zone_get_max(uma_zone_t zone)
2999 keg = zone_first_keg(zone);
3003 nitems = keg->uk_maxpages * keg->uk_ipers;
3011 uma_zone_set_warning(uma_zone_t zone, const char *warning)
3015 zone->uz_warning = warning;
3021 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
3025 zone->uz_maxaction = maxaction;
3031 uma_zone_get_cur(uma_zone_t zone)
3037 nitems = zone->uz_allocs - zone->uz_frees;
3040 * See the comment in sysctl_vm_zone_stats() regarding the
3041 * safety of accessing the per-cpu caches. With the zone lock
3042 * held, it is safe, but can potentially result in stale data.
3044 nitems += zone->uz_cpu[i].uc_allocs -
3045 zone->uz_cpu[i].uc_frees;
3049 return (nitems < 0 ? 0 : nitems);
3054 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
3058 keg = zone_first_keg(zone);
3059 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3061 KASSERT(keg->uk_pages == 0,
3062 ("uma_zone_set_init on non-empty keg"));
3063 keg->uk_init = uminit;
3069 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3073 keg = zone_first_keg(zone);
3074 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
3076 KASSERT(keg->uk_pages == 0,
3077 ("uma_zone_set_fini on non-empty keg"));
3078 keg->uk_fini = fini;
3084 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3088 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3089 ("uma_zone_set_zinit on non-empty keg"));
3090 zone->uz_init = zinit;
3096 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3100 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3101 ("uma_zone_set_zfini on non-empty keg"));
3102 zone->uz_fini = zfini;
3107 /* XXX uk_freef is not actually used with the zone locked */
3109 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3113 keg = zone_first_keg(zone);
3114 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
3116 keg->uk_freef = freef;
3121 /* XXX uk_allocf is not actually used with the zone locked */
3123 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3127 keg = zone_first_keg(zone);
3129 keg->uk_allocf = allocf;
3135 uma_zone_reserve(uma_zone_t zone, int items)
3139 keg = zone_first_keg(zone);
3143 keg->uk_reserve = items;
3151 uma_zone_reserve_kva(uma_zone_t zone, int count)
3157 keg = zone_first_keg(zone);
3160 pages = count / keg->uk_ipers;
3162 if (pages * keg->uk_ipers < count)
3165 #ifdef UMA_MD_SMALL_ALLOC
3166 if (keg->uk_ppera > 1) {
3170 kva = kva_alloc((vm_size_t)pages * UMA_SLAB_SIZE);
3178 keg->uk_maxpages = pages;
3179 #ifdef UMA_MD_SMALL_ALLOC
3180 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3182 keg->uk_allocf = noobj_alloc;
3184 keg->uk_flags |= UMA_ZONE_NOFREE;
3192 uma_prealloc(uma_zone_t zone, int items)
3198 keg = zone_first_keg(zone);
3202 slabs = items / keg->uk_ipers;
3203 if (slabs * keg->uk_ipers < items)
3206 slab = keg_alloc_slab(keg, zone, M_WAITOK);
3209 MPASS(slab->us_keg == keg);
3210 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3218 uma_find_refcnt(uma_zone_t zone, void *item)
3220 uma_slabrefcnt_t slabref;
3226 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3227 slabref = (uma_slabrefcnt_t)slab;
3229 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3230 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3231 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3232 refcnt = &slabref->us_refcnt[idx];
3238 uma_reclaim_locked(bool kmem_danger)
3242 printf("UMA: vm asked us to release pages!\n");
3244 sx_assert(&uma_drain_lock, SA_XLOCKED);
3246 zone_foreach(zone_drain);
3247 if (vm_page_count_min() || kmem_danger) {
3248 cache_drain_safe(NULL);
3249 zone_foreach(zone_drain);
3252 * Some slabs may have been freed but this zone will be visited early
3253 * we visit again so that we can free pages that are empty once other
3254 * zones are drained. We have to do the same for buckets.
3256 zone_drain(slabzone);
3257 zone_drain(slabrefzone);
3258 bucket_zone_drain();
3265 sx_xlock(&uma_drain_lock);
3266 uma_reclaim_locked(false);
3267 sx_xunlock(&uma_drain_lock);
3270 static int uma_reclaim_needed;
3273 uma_reclaim_wakeup(void)
3276 uma_reclaim_needed = 1;
3277 wakeup(&uma_reclaim_needed);
3281 uma_reclaim_worker(void *arg __unused)
3284 sx_xlock(&uma_drain_lock);
3286 sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM,
3288 if (uma_reclaim_needed) {
3289 uma_reclaim_needed = 0;
3290 uma_reclaim_locked(true);
3297 uma_zone_exhausted(uma_zone_t zone)
3302 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3308 uma_zone_exhausted_nolock(uma_zone_t zone)
3310 return (zone->uz_flags & UMA_ZFLAG_FULL);
3314 uma_large_malloc(vm_size_t size, int wait)
3320 slab = zone_alloc_item(slabzone, NULL, wait);
3323 mem = page_alloc(NULL, size, &flags, wait);
3325 vsetslab((vm_offset_t)mem, slab);
3326 slab->us_data = mem;
3327 slab->us_flags = flags | UMA_SLAB_MALLOC;
3328 slab->us_size = size;
3330 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3337 uma_large_free(uma_slab_t slab)
3340 page_free(slab->us_data, slab->us_size, slab->us_flags);
3341 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3345 uma_zero_item(void *item, uma_zone_t zone)
3348 if (zone->uz_flags & UMA_ZONE_PCPU) {
3349 for (int i = 0; i < mp_ncpus; i++)
3350 bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3352 bzero(item, zone->uz_size);
3356 uma_print_stats(void)
3358 zone_foreach(uma_print_zone);
3362 slab_print(uma_slab_t slab)
3364 printf("slab: keg %p, data %p, freecount %d\n",
3365 slab->us_keg, slab->us_data, slab->us_freecount);
3369 cache_print(uma_cache_t cache)
3371 printf("alloc: %p(%d), free: %p(%d)\n",
3372 cache->uc_allocbucket,
3373 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3374 cache->uc_freebucket,
3375 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3379 uma_print_keg(uma_keg_t keg)
3383 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3384 "out %d free %d limit %d\n",
3385 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3386 keg->uk_ipers, keg->uk_ppera,
3387 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3388 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3389 printf("Part slabs:\n");
3390 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3392 printf("Free slabs:\n");
3393 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3395 printf("Full slabs:\n");
3396 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3401 uma_print_zone(uma_zone_t zone)
3407 printf("zone: %s(%p) size %d flags %#x\n",
3408 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3409 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3410 uma_print_keg(kl->kl_keg);
3412 cache = &zone->uz_cpu[i];
3413 printf("CPU %d Cache:\n", i);
3420 * Generate statistics across both the zone and its per-cpu cache's. Return
3421 * desired statistics if the pointer is non-NULL for that statistic.
3423 * Note: does not update the zone statistics, as it can't safely clear the
3424 * per-CPU cache statistic.
3426 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3427 * safe from off-CPU; we should modify the caches to track this information
3428 * directly so that we don't have to.
3431 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3432 uint64_t *freesp, uint64_t *sleepsp)
3435 uint64_t allocs, frees, sleeps;
3438 allocs = frees = sleeps = 0;
3441 cache = &z->uz_cpu[cpu];
3442 if (cache->uc_allocbucket != NULL)
3443 cachefree += cache->uc_allocbucket->ub_cnt;
3444 if (cache->uc_freebucket != NULL)
3445 cachefree += cache->uc_freebucket->ub_cnt;
3446 allocs += cache->uc_allocs;
3447 frees += cache->uc_frees;
3449 allocs += z->uz_allocs;
3450 frees += z->uz_frees;
3451 sleeps += z->uz_sleeps;
3452 if (cachefreep != NULL)
3453 *cachefreep = cachefree;
3454 if (allocsp != NULL)
3458 if (sleepsp != NULL)
3464 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3471 rw_rlock(&uma_rwlock);
3472 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3473 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3476 rw_runlock(&uma_rwlock);
3477 return (sysctl_handle_int(oidp, &count, 0, req));
3481 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3483 struct uma_stream_header ush;
3484 struct uma_type_header uth;
3485 struct uma_percpu_stat ups;
3486 uma_bucket_t bucket;
3493 int count, error, i;
3495 error = sysctl_wire_old_buffer(req, 0);
3498 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3499 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3502 rw_rlock(&uma_rwlock);
3503 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3504 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3509 * Insert stream header.
3511 bzero(&ush, sizeof(ush));
3512 ush.ush_version = UMA_STREAM_VERSION;
3513 ush.ush_maxcpus = (mp_maxid + 1);
3514 ush.ush_count = count;
3515 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3517 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3518 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3519 bzero(&uth, sizeof(uth));
3521 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3522 uth.uth_align = kz->uk_align;
3523 uth.uth_size = kz->uk_size;
3524 uth.uth_rsize = kz->uk_rsize;
3525 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3527 uth.uth_maxpages += k->uk_maxpages;
3528 uth.uth_pages += k->uk_pages;
3529 uth.uth_keg_free += k->uk_free;
3530 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3535 * A zone is secondary is it is not the first entry
3536 * on the keg's zone list.
3538 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3539 (LIST_FIRST(&kz->uk_zones) != z))
3540 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3542 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3543 uth.uth_zone_free += bucket->ub_cnt;
3544 uth.uth_allocs = z->uz_allocs;
3545 uth.uth_frees = z->uz_frees;
3546 uth.uth_fails = z->uz_fails;
3547 uth.uth_sleeps = z->uz_sleeps;
3548 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3550 * While it is not normally safe to access the cache
3551 * bucket pointers while not on the CPU that owns the
3552 * cache, we only allow the pointers to be exchanged
3553 * without the zone lock held, not invalidated, so
3554 * accept the possible race associated with bucket
3555 * exchange during monitoring.
3557 for (i = 0; i < (mp_maxid + 1); i++) {
3558 bzero(&ups, sizeof(ups));
3559 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3563 cache = &z->uz_cpu[i];
3564 if (cache->uc_allocbucket != NULL)
3565 ups.ups_cache_free +=
3566 cache->uc_allocbucket->ub_cnt;
3567 if (cache->uc_freebucket != NULL)
3568 ups.ups_cache_free +=
3569 cache->uc_freebucket->ub_cnt;
3570 ups.ups_allocs = cache->uc_allocs;
3571 ups.ups_frees = cache->uc_frees;
3573 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3578 rw_runlock(&uma_rwlock);
3579 error = sbuf_finish(&sbuf);
3585 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3587 uma_zone_t zone = *(uma_zone_t *)arg1;
3590 max = uma_zone_get_max(zone);
3591 error = sysctl_handle_int(oidp, &max, 0, req);
3592 if (error || !req->newptr)
3595 uma_zone_set_max(zone, max);
3601 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3603 uma_zone_t zone = *(uma_zone_t *)arg1;
3606 cur = uma_zone_get_cur(zone);
3607 return (sysctl_handle_int(oidp, &cur, 0, req));
3611 DB_SHOW_COMMAND(uma, db_show_uma)
3613 uint64_t allocs, frees, sleeps;
3614 uma_bucket_t bucket;
3619 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3620 "Free", "Requests", "Sleeps", "Bucket");
3621 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3622 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3623 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3624 allocs = z->uz_allocs;
3625 frees = z->uz_frees;
3626 sleeps = z->uz_sleeps;
3629 uma_zone_sumstat(z, &cachefree, &allocs,
3631 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3632 (LIST_FIRST(&kz->uk_zones) != z)))
3633 cachefree += kz->uk_free;
3634 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3635 cachefree += bucket->ub_cnt;
3636 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3637 z->uz_name, (uintmax_t)kz->uk_size,
3638 (intmax_t)(allocs - frees), cachefree,
3639 (uintmax_t)allocs, sleeps, z->uz_count);
3646 DB_SHOW_COMMAND(umacache, db_show_umacache)
3648 uint64_t allocs, frees;
3649 uma_bucket_t bucket;
3653 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3654 "Requests", "Bucket");
3655 LIST_FOREACH(z, &uma_cachezones, uz_link) {
3656 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3657 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3658 cachefree += bucket->ub_cnt;
3659 db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3660 z->uz_name, (uintmax_t)z->uz_size,
3661 (intmax_t)(allocs - frees), cachefree,
3662 (uintmax_t)allocs, z->uz_count);