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/eventhandler.h>
68 #include <sys/kernel.h>
69 #include <sys/types.h>
70 #include <sys/queue.h>
71 #include <sys/malloc.h>
74 #include <sys/sysctl.h>
75 #include <sys/mutex.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 TUNABLE_INT("vm.zone_warnings", &zone_warnings);
288 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RW, &zone_warnings, 0,
289 "Warn when UMA zones becomes full");
292 * This routine checks to see whether or not it's safe to enable buckets.
297 bucketdisable = vm_page_count_min();
301 * Initialize bucket_zones, the array of zones of buckets of various sizes.
303 * For each zone, calculate the memory required for each bucket, consisting
304 * of the header and an array of pointers.
309 struct uma_bucket_zone *ubz;
313 for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
314 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
315 size += sizeof(void *) * ubz->ubz_entries;
316 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
317 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
318 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
323 * Given a desired number of entries for a bucket, return the zone from which
324 * to allocate the bucket.
326 static struct uma_bucket_zone *
327 bucket_zone_lookup(int entries)
329 struct uma_bucket_zone *ubz;
331 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
332 if (ubz->ubz_entries >= entries)
339 bucket_select(int size)
341 struct uma_bucket_zone *ubz;
343 ubz = &bucket_zones[0];
344 if (size > ubz->ubz_maxsize)
345 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
347 for (; ubz->ubz_entries != 0; ubz++)
348 if (ubz->ubz_maxsize < size)
351 return (ubz->ubz_entries);
355 bucket_alloc(uma_zone_t zone, void *udata, int flags)
357 struct uma_bucket_zone *ubz;
361 * This is to stop us from allocating per cpu buckets while we're
362 * running out of vm.boot_pages. Otherwise, we would exhaust the
363 * boot pages. This also prevents us from allocating buckets in
364 * low memory situations.
369 * To limit bucket recursion we store the original zone flags
370 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
371 * NOVM flag to persist even through deep recursions. We also
372 * store ZFLAG_BUCKET once we have recursed attempting to allocate
373 * a bucket for a bucket zone so we do not allow infinite bucket
374 * recursion. This cookie will even persist to frees of unused
375 * buckets via the allocation path or bucket allocations in the
378 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
379 udata = (void *)(uintptr_t)zone->uz_flags;
381 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
383 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
385 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
387 ubz = bucket_zone_lookup(zone->uz_count);
388 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
390 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
393 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
396 bucket->ub_entries = ubz->ubz_entries;
403 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
405 struct uma_bucket_zone *ubz;
407 KASSERT(bucket->ub_cnt == 0,
408 ("bucket_free: Freeing a non free bucket."));
409 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
410 udata = (void *)(uintptr_t)zone->uz_flags;
411 ubz = bucket_zone_lookup(bucket->ub_entries);
412 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
416 bucket_zone_drain(void)
418 struct uma_bucket_zone *ubz;
420 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
421 zone_drain(ubz->ubz_zone);
425 zone_log_warning(uma_zone_t zone)
427 static const struct timeval warninterval = { 300, 0 };
429 if (!zone_warnings || zone->uz_warning == NULL)
432 if (ratecheck(&zone->uz_ratecheck, &warninterval))
433 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
437 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
441 LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
442 kegfn(klink->kl_keg);
446 * Routine called by timeout which is used to fire off some time interval
447 * based calculations. (stats, hash size, etc.)
456 uma_timeout(void *unused)
459 zone_foreach(zone_timeout);
461 /* Reschedule this event */
462 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
466 * Routine to perform timeout driven calculations. This expands the
467 * hashes and does per cpu statistics aggregation.
472 keg_timeout(uma_keg_t keg)
477 * Expand the keg hash table.
479 * This is done if the number of slabs is larger than the hash size.
480 * What I'm trying to do here is completely reduce collisions. This
481 * may be a little aggressive. Should I allow for two collisions max?
483 if (keg->uk_flags & UMA_ZONE_HASH &&
484 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
485 struct uma_hash newhash;
486 struct uma_hash oldhash;
490 * This is so involved because allocating and freeing
491 * while the keg lock is held will lead to deadlock.
492 * I have to do everything in stages and check for
495 newhash = keg->uk_hash;
497 ret = hash_alloc(&newhash);
500 if (hash_expand(&keg->uk_hash, &newhash)) {
501 oldhash = keg->uk_hash;
502 keg->uk_hash = newhash;
515 zone_timeout(uma_zone_t zone)
518 zone_foreach_keg(zone, &keg_timeout);
522 * Allocate and zero fill the next sized hash table from the appropriate
526 * hash A new hash structure with the old hash size in uh_hashsize
529 * 1 on sucess and 0 on failure.
532 hash_alloc(struct uma_hash *hash)
537 oldsize = hash->uh_hashsize;
539 /* We're just going to go to a power of two greater */
541 hash->uh_hashsize = oldsize * 2;
542 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
543 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
544 M_UMAHASH, M_NOWAIT);
546 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
547 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
549 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
551 if (hash->uh_slab_hash) {
552 bzero(hash->uh_slab_hash, alloc);
553 hash->uh_hashmask = hash->uh_hashsize - 1;
561 * Expands the hash table for HASH zones. This is done from zone_timeout
562 * to reduce collisions. This must not be done in the regular allocation
563 * path, otherwise, we can recurse on the vm while allocating pages.
566 * oldhash The hash you want to expand
567 * newhash The hash structure for the new table
575 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
581 if (!newhash->uh_slab_hash)
584 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
588 * I need to investigate hash algorithms for resizing without a
592 for (i = 0; i < oldhash->uh_hashsize; i++)
593 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
594 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
595 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
596 hval = UMA_HASH(newhash, slab->us_data);
597 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
605 * Free the hash bucket to the appropriate backing store.
608 * slab_hash The hash bucket we're freeing
609 * hashsize The number of entries in that hash bucket
615 hash_free(struct uma_hash *hash)
617 if (hash->uh_slab_hash == NULL)
619 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
620 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
622 free(hash->uh_slab_hash, M_UMAHASH);
626 * Frees all outstanding items in a bucket
629 * zone The zone to free to, must be unlocked.
630 * bucket The free/alloc bucket with items, cpu queue must be locked.
637 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
645 for (i = 0; i < bucket->ub_cnt; i++)
646 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
647 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
652 * Drains the per cpu caches for a zone.
654 * NOTE: This may only be called while the zone is being turn down, and not
655 * during normal operation. This is necessary in order that we do not have
656 * to migrate CPUs to drain the per-CPU caches.
659 * zone The zone to drain, must be unlocked.
665 cache_drain(uma_zone_t zone)
671 * XXX: It is safe to not lock the per-CPU caches, because we're
672 * tearing down the zone anyway. I.e., there will be no further use
673 * of the caches at this point.
675 * XXX: It would good to be able to assert that the zone is being
676 * torn down to prevent improper use of cache_drain().
678 * XXX: We lock the zone before passing into bucket_cache_drain() as
679 * it is used elsewhere. Should the tear-down path be made special
680 * there in some form?
683 cache = &zone->uz_cpu[cpu];
684 bucket_drain(zone, cache->uc_allocbucket);
685 bucket_drain(zone, cache->uc_freebucket);
686 if (cache->uc_allocbucket != NULL)
687 bucket_free(zone, cache->uc_allocbucket, NULL);
688 if (cache->uc_freebucket != NULL)
689 bucket_free(zone, cache->uc_freebucket, NULL);
690 cache->uc_allocbucket = cache->uc_freebucket = NULL;
693 bucket_cache_drain(zone);
698 cache_shrink(uma_zone_t zone)
701 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
705 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
710 cache_drain_safe_cpu(uma_zone_t zone)
715 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
721 cache = &zone->uz_cpu[curcpu];
722 if (cache->uc_allocbucket) {
723 if (cache->uc_allocbucket->ub_cnt != 0)
724 LIST_INSERT_HEAD(&zone->uz_buckets,
725 cache->uc_allocbucket, ub_link);
727 b1 = cache->uc_allocbucket;
728 cache->uc_allocbucket = NULL;
730 if (cache->uc_freebucket) {
731 if (cache->uc_freebucket->ub_cnt != 0)
732 LIST_INSERT_HEAD(&zone->uz_buckets,
733 cache->uc_freebucket, ub_link);
735 b2 = cache->uc_freebucket;
736 cache->uc_freebucket = NULL;
741 bucket_free(zone, b1, NULL);
743 bucket_free(zone, b2, NULL);
747 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
748 * This is an expensive call because it needs to bind to all CPUs
749 * one by one and enter a critical section on each of them in order
750 * to safely access their cache buckets.
751 * Zone lock must not be held on call this function.
754 cache_drain_safe(uma_zone_t zone)
759 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
764 zone_foreach(cache_shrink);
767 thread_lock(curthread);
768 sched_bind(curthread, cpu);
769 thread_unlock(curthread);
772 cache_drain_safe_cpu(zone);
774 zone_foreach(cache_drain_safe_cpu);
776 thread_lock(curthread);
777 sched_unbind(curthread);
778 thread_unlock(curthread);
782 * Drain the cached buckets from a zone. Expects a locked zone on entry.
785 bucket_cache_drain(uma_zone_t zone)
790 * Drain the bucket queues and free the buckets, we just keep two per
793 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
794 LIST_REMOVE(bucket, ub_link);
796 bucket_drain(zone, bucket);
797 bucket_free(zone, bucket, NULL);
802 * Shrink further bucket sizes. Price of single zone lock collision
803 * is probably lower then price of global cache drain.
805 if (zone->uz_count > zone->uz_count_min)
810 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
817 flags = slab->us_flags;
819 if (keg->uk_fini != NULL) {
820 for (i--; i > -1; i--)
821 keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
824 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
825 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
827 printf("%s: Returning %d bytes.\n", keg->uk_name,
828 PAGE_SIZE * keg->uk_ppera);
830 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
834 * Frees pages from a keg back to the system. This is done on demand from
835 * the pageout daemon.
840 keg_drain(uma_keg_t keg)
842 struct slabhead freeslabs = { 0 };
847 * We don't want to take pages from statically allocated kegs at this
850 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
854 printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
857 if (keg->uk_free == 0)
860 slab = LIST_FIRST(&keg->uk_free_slab);
862 n = LIST_NEXT(slab, us_link);
864 /* We have no where to free these to */
865 if (slab->us_flags & UMA_SLAB_BOOT) {
870 LIST_REMOVE(slab, us_link);
871 keg->uk_pages -= keg->uk_ppera;
872 keg->uk_free -= keg->uk_ipers;
874 if (keg->uk_flags & UMA_ZONE_HASH)
875 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
877 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
884 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
885 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
886 keg_free_slab(keg, slab, keg->uk_ipers);
891 zone_drain_wait(uma_zone_t zone, int waitok)
895 * Set draining to interlock with zone_dtor() so we can release our
896 * locks as we go. Only dtor() should do a WAITOK call since it
897 * is the only call that knows the structure will still be available
901 while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
902 if (waitok == M_NOWAIT)
904 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
906 zone->uz_flags |= UMA_ZFLAG_DRAINING;
907 bucket_cache_drain(zone);
910 * The DRAINING flag protects us from being freed while
911 * we're running. Normally the uma_rwlock would protect us but we
912 * must be able to release and acquire the right lock for each keg.
914 zone_foreach_keg(zone, &keg_drain);
916 zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
923 zone_drain(uma_zone_t zone)
926 zone_drain_wait(zone, M_NOWAIT);
930 * Allocate a new slab for a keg. This does not insert the slab onto a list.
933 * wait Shall we wait?
936 * The slab that was allocated or NULL if there is no memory and the
937 * caller specified M_NOWAIT.
940 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
942 uma_slabrefcnt_t slabref;
949 mtx_assert(&keg->uk_lock, MA_OWNED);
954 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name);
956 allocf = keg->uk_allocf;
959 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
960 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
966 * This reproduces the old vm_zone behavior of zero filling pages the
967 * first time they are added to a zone.
969 * Malloced items are zeroed in uma_zalloc.
972 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
977 if (keg->uk_flags & UMA_ZONE_NODUMP)
980 /* zone is passed for legacy reasons. */
981 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
983 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
984 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
989 /* Point the slab into the allocated memory */
990 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
991 slab = (uma_slab_t )(mem + keg->uk_pgoff);
993 if (keg->uk_flags & UMA_ZONE_VTOSLAB)
994 for (i = 0; i < keg->uk_ppera; i++)
995 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
999 slab->us_freecount = keg->uk_ipers;
1000 slab->us_flags = flags;
1001 BIT_FILL(SLAB_SETSIZE, &slab->us_free);
1003 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1005 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1006 slabref = (uma_slabrefcnt_t)slab;
1007 for (i = 0; i < keg->uk_ipers; i++)
1008 slabref->us_refcnt[i] = 0;
1011 if (keg->uk_init != NULL) {
1012 for (i = 0; i < keg->uk_ipers; i++)
1013 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1014 keg->uk_size, wait) != 0)
1016 if (i != keg->uk_ipers) {
1017 keg_free_slab(keg, slab, i);
1026 if (keg->uk_flags & UMA_ZONE_HASH)
1027 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1029 keg->uk_pages += keg->uk_ppera;
1030 keg->uk_free += keg->uk_ipers;
1037 * This function is intended to be used early on in place of page_alloc() so
1038 * that we may use the boot time page cache to satisfy allocations before
1042 startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1046 int pages, check_pages;
1048 keg = zone_first_keg(zone);
1049 pages = howmany(bytes, PAGE_SIZE);
1050 check_pages = pages - 1;
1051 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1054 * Check our small startup cache to see if it has pages remaining.
1056 mtx_lock(&uma_boot_pages_mtx);
1058 /* First check if we have enough room. */
1059 tmps = LIST_FIRST(&uma_boot_pages);
1060 while (tmps != NULL && check_pages-- > 0)
1061 tmps = LIST_NEXT(tmps, us_link);
1064 * It's ok to lose tmps references. The last one will
1065 * have tmps->us_data pointing to the start address of
1066 * "pages" contiguous pages of memory.
1068 while (pages-- > 0) {
1069 tmps = LIST_FIRST(&uma_boot_pages);
1070 LIST_REMOVE(tmps, us_link);
1072 mtx_unlock(&uma_boot_pages_mtx);
1073 *pflag = tmps->us_flags;
1074 return (tmps->us_data);
1076 mtx_unlock(&uma_boot_pages_mtx);
1077 if (booted < UMA_STARTUP2)
1078 panic("UMA: Increase vm.boot_pages");
1080 * Now that we've booted reset these users to their real allocator.
1082 #ifdef UMA_MD_SMALL_ALLOC
1083 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1085 keg->uk_allocf = page_alloc;
1087 return keg->uk_allocf(zone, bytes, pflag, wait);
1091 * Allocates a number of pages from the system
1094 * bytes The number of bytes requested
1095 * wait Shall we wait?
1098 * A pointer to the alloced memory or possibly
1099 * NULL if M_NOWAIT is set.
1102 page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1104 void *p; /* Returned page */
1106 *pflag = UMA_SLAB_KMEM;
1107 p = (void *) kmem_malloc(kmem_arena, bytes, wait);
1113 * Allocates a number of pages from within an object
1116 * bytes The number of bytes requested
1117 * wait Shall we wait?
1120 * A pointer to the alloced memory or possibly
1121 * NULL if M_NOWAIT is set.
1124 noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait)
1126 TAILQ_HEAD(, vm_page) alloctail;
1128 vm_offset_t retkva, zkva;
1129 vm_page_t p, p_next;
1132 TAILQ_INIT(&alloctail);
1133 keg = zone_first_keg(zone);
1135 npages = howmany(bytes, PAGE_SIZE);
1136 while (npages > 0) {
1137 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1138 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1141 * Since the page does not belong to an object, its
1144 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1148 if (wait & M_WAITOK) {
1154 * Page allocation failed, free intermediate pages and
1157 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1158 vm_page_unwire(p, 0);
1163 *flags = UMA_SLAB_PRIV;
1164 zkva = keg->uk_kva +
1165 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1167 TAILQ_FOREACH(p, &alloctail, listq) {
1168 pmap_qenter(zkva, &p, 1);
1172 return ((void *)retkva);
1176 * Frees a number of pages to the system
1179 * mem A pointer to the memory to be freed
1180 * size The size of the memory being freed
1181 * flags The original p->us_flags field
1187 page_free(void *mem, vm_size_t size, uint8_t flags)
1191 if (flags & UMA_SLAB_KMEM)
1193 else if (flags & UMA_SLAB_KERNEL)
1194 vmem = kernel_arena;
1196 panic("UMA: page_free used with invalid flags %d", flags);
1198 kmem_free(vmem, (vm_offset_t)mem, size);
1202 * Zero fill initializer
1204 * Arguments/Returns follow uma_init specifications
1207 zero_init(void *mem, int size, int flags)
1214 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1217 * keg The zone we should initialize
1223 keg_small_init(uma_keg_t keg)
1231 if (keg->uk_flags & UMA_ZONE_PCPU) {
1232 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1234 slabsize = sizeof(struct pcpu);
1235 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1238 slabsize = UMA_SLAB_SIZE;
1243 * Calculate the size of each allocation (rsize) according to
1244 * alignment. If the requested size is smaller than we have
1245 * allocation bits for we round it up.
1247 rsize = keg->uk_size;
1248 if (rsize < slabsize / SLAB_SETSIZE)
1249 rsize = slabsize / SLAB_SETSIZE;
1250 if (rsize & keg->uk_align)
1251 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1252 keg->uk_rsize = rsize;
1254 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1255 keg->uk_rsize < sizeof(struct pcpu),
1256 ("%s: size %u too large", __func__, keg->uk_rsize));
1258 if (keg->uk_flags & UMA_ZONE_REFCNT)
1259 rsize += sizeof(uint32_t);
1261 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1264 shsize = sizeof(struct uma_slab);
1266 keg->uk_ipers = (slabsize - shsize) / rsize;
1267 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1268 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1270 memused = keg->uk_ipers * rsize + shsize;
1271 wastedspace = slabsize - memused;
1274 * We can't do OFFPAGE if we're internal or if we've been
1275 * asked to not go to the VM for buckets. If we do this we
1276 * may end up going to the VM for slabs which we do not
1277 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1278 * of UMA_ZONE_VM, which clearly forbids it.
1280 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1281 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1285 * See if using an OFFPAGE slab will limit our waste. Only do
1286 * this if it permits more items per-slab.
1288 * XXX We could try growing slabsize to limit max waste as well.
1289 * Historically this was not done because the VM could not
1290 * efficiently handle contiguous allocations.
1292 if ((wastedspace >= slabsize / UMA_MAX_WASTE) &&
1293 (keg->uk_ipers < (slabsize / keg->uk_rsize))) {
1294 keg->uk_ipers = slabsize / keg->uk_rsize;
1295 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1296 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1298 printf("UMA decided we need offpage slab headers for "
1299 "keg: %s, calculated wastedspace = %d, "
1300 "maximum wasted space allowed = %d, "
1301 "calculated ipers = %d, "
1302 "new wasted space = %d\n", keg->uk_name, wastedspace,
1303 slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1304 slabsize - keg->uk_ipers * keg->uk_rsize);
1306 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1309 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1310 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1311 keg->uk_flags |= UMA_ZONE_HASH;
1315 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1316 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1320 * keg The keg we should initialize
1326 keg_large_init(uma_keg_t keg)
1330 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1331 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1332 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1333 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1334 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1336 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1338 keg->uk_rsize = keg->uk_size;
1340 /* We can't do OFFPAGE if we're internal, bail out here. */
1341 if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1344 /* Check whether we have enough space to not do OFFPAGE. */
1345 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1346 shsize = sizeof(struct uma_slab);
1347 if (keg->uk_flags & UMA_ZONE_REFCNT)
1348 shsize += keg->uk_ipers * sizeof(uint32_t);
1349 if (shsize & UMA_ALIGN_PTR)
1350 shsize = (shsize & ~UMA_ALIGN_PTR) +
1351 (UMA_ALIGN_PTR + 1);
1353 if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize)
1354 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1357 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1358 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1359 keg->uk_flags |= UMA_ZONE_HASH;
1363 keg_cachespread_init(uma_keg_t keg)
1370 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1371 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1373 alignsize = keg->uk_align + 1;
1374 rsize = keg->uk_size;
1376 * We want one item to start on every align boundary in a page. To
1377 * do this we will span pages. We will also extend the item by the
1378 * size of align if it is an even multiple of align. Otherwise, it
1379 * would fall on the same boundary every time.
1381 if (rsize & keg->uk_align)
1382 rsize = (rsize & ~keg->uk_align) + alignsize;
1383 if ((rsize & alignsize) == 0)
1385 trailer = rsize - keg->uk_size;
1386 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1387 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1388 keg->uk_rsize = rsize;
1389 keg->uk_ppera = pages;
1390 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1391 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1392 KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1393 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1398 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1399 * the keg onto the global keg list.
1401 * Arguments/Returns follow uma_ctor specifications
1402 * udata Actually uma_kctor_args
1405 keg_ctor(void *mem, int size, void *udata, int flags)
1407 struct uma_kctor_args *arg = udata;
1408 uma_keg_t keg = mem;
1412 keg->uk_size = arg->size;
1413 keg->uk_init = arg->uminit;
1414 keg->uk_fini = arg->fini;
1415 keg->uk_align = arg->align;
1417 keg->uk_reserve = 0;
1419 keg->uk_flags = arg->flags;
1420 keg->uk_allocf = page_alloc;
1421 keg->uk_freef = page_free;
1422 keg->uk_slabzone = NULL;
1425 * The master zone is passed to us at keg-creation time.
1428 keg->uk_name = zone->uz_name;
1430 if (arg->flags & UMA_ZONE_VM)
1431 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1433 if (arg->flags & UMA_ZONE_ZINIT)
1434 keg->uk_init = zero_init;
1436 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1437 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1439 if (arg->flags & UMA_ZONE_PCPU)
1441 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1443 keg->uk_flags &= ~UMA_ZONE_PCPU;
1446 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1447 keg_cachespread_init(keg);
1448 } else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1450 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) -
1452 keg_large_init(keg);
1454 keg_small_init(keg);
1456 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1457 keg_large_init(keg);
1459 keg_small_init(keg);
1462 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1463 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1464 if (keg->uk_ipers > uma_max_ipers_ref)
1465 panic("Too many ref items per zone: %d > %d\n",
1466 keg->uk_ipers, uma_max_ipers_ref);
1467 keg->uk_slabzone = slabrefzone;
1469 keg->uk_slabzone = slabzone;
1473 * If we haven't booted yet we need allocations to go through the
1474 * startup cache until the vm is ready.
1476 if (keg->uk_ppera == 1) {
1477 #ifdef UMA_MD_SMALL_ALLOC
1478 keg->uk_allocf = uma_small_alloc;
1479 keg->uk_freef = uma_small_free;
1481 if (booted < UMA_STARTUP)
1482 keg->uk_allocf = startup_alloc;
1484 if (booted < UMA_STARTUP2)
1485 keg->uk_allocf = startup_alloc;
1487 } else if (booted < UMA_STARTUP2 &&
1488 (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1489 keg->uk_allocf = startup_alloc;
1492 * Initialize keg's lock
1494 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1497 * If we're putting the slab header in the actual page we need to
1498 * figure out where in each page it goes. This calculates a right
1499 * justified offset into the memory on an ALIGN_PTR boundary.
1501 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1504 /* Size of the slab struct and free list */
1505 totsize = sizeof(struct uma_slab);
1507 /* Size of the reference counts. */
1508 if (keg->uk_flags & UMA_ZONE_REFCNT)
1509 totsize += keg->uk_ipers * sizeof(uint32_t);
1511 if (totsize & UMA_ALIGN_PTR)
1512 totsize = (totsize & ~UMA_ALIGN_PTR) +
1513 (UMA_ALIGN_PTR + 1);
1514 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1517 * The only way the following is possible is if with our
1518 * UMA_ALIGN_PTR adjustments we are now bigger than
1519 * UMA_SLAB_SIZE. I haven't checked whether this is
1520 * mathematically possible for all cases, so we make
1523 totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1524 if (keg->uk_flags & UMA_ZONE_REFCNT)
1525 totsize += keg->uk_ipers * sizeof(uint32_t);
1526 if (totsize > PAGE_SIZE * keg->uk_ppera) {
1527 printf("zone %s ipers %d rsize %d size %d\n",
1528 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1530 panic("UMA slab won't fit.");
1534 if (keg->uk_flags & UMA_ZONE_HASH)
1535 hash_alloc(&keg->uk_hash);
1538 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1539 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1540 keg->uk_ipers, keg->uk_ppera,
1541 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1544 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1546 rw_wlock(&uma_rwlock);
1547 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1548 rw_wunlock(&uma_rwlock);
1553 * Zone header ctor. This initializes all fields, locks, etc.
1555 * Arguments/Returns follow uma_ctor specifications
1556 * udata Actually uma_zctor_args
1559 zone_ctor(void *mem, int size, void *udata, int flags)
1561 struct uma_zctor_args *arg = udata;
1562 uma_zone_t zone = mem;
1567 zone->uz_name = arg->name;
1568 zone->uz_ctor = arg->ctor;
1569 zone->uz_dtor = arg->dtor;
1570 zone->uz_slab = zone_fetch_slab;
1571 zone->uz_init = NULL;
1572 zone->uz_fini = NULL;
1573 zone->uz_allocs = 0;
1576 zone->uz_sleeps = 0;
1578 zone->uz_count_min = 0;
1580 zone->uz_warning = NULL;
1581 timevalclear(&zone->uz_ratecheck);
1584 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1587 * This is a pure cache zone, no kegs.
1590 if (arg->flags & UMA_ZONE_VM)
1591 arg->flags |= UMA_ZFLAG_CACHEONLY;
1592 zone->uz_flags = arg->flags;
1593 zone->uz_size = arg->size;
1594 zone->uz_import = arg->import;
1595 zone->uz_release = arg->release;
1596 zone->uz_arg = arg->arg;
1597 zone->uz_lockptr = &zone->uz_lock;
1598 rw_wlock(&uma_rwlock);
1599 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1600 rw_wunlock(&uma_rwlock);
1605 * Use the regular zone/keg/slab allocator.
1607 zone->uz_import = (uma_import)zone_import;
1608 zone->uz_release = (uma_release)zone_release;
1609 zone->uz_arg = zone;
1611 if (arg->flags & UMA_ZONE_SECONDARY) {
1612 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1613 zone->uz_init = arg->uminit;
1614 zone->uz_fini = arg->fini;
1615 zone->uz_lockptr = &keg->uk_lock;
1616 zone->uz_flags |= UMA_ZONE_SECONDARY;
1617 rw_wlock(&uma_rwlock);
1619 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1620 if (LIST_NEXT(z, uz_link) == NULL) {
1621 LIST_INSERT_AFTER(z, zone, uz_link);
1626 rw_wunlock(&uma_rwlock);
1627 } else if (keg == NULL) {
1628 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1629 arg->align, arg->flags)) == NULL)
1632 struct uma_kctor_args karg;
1635 /* We should only be here from uma_startup() */
1636 karg.size = arg->size;
1637 karg.uminit = arg->uminit;
1638 karg.fini = arg->fini;
1639 karg.align = arg->align;
1640 karg.flags = arg->flags;
1642 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1649 * Link in the first keg.
1651 zone->uz_klink.kl_keg = keg;
1652 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1653 zone->uz_lockptr = &keg->uk_lock;
1654 zone->uz_size = keg->uk_size;
1655 zone->uz_flags |= (keg->uk_flags &
1656 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1659 * Some internal zones don't have room allocated for the per cpu
1660 * caches. If we're internal, bail out here.
1662 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1663 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1664 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1669 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1670 zone->uz_count = bucket_select(zone->uz_size);
1672 zone->uz_count = BUCKET_MAX;
1673 zone->uz_count_min = zone->uz_count;
1679 * Keg header dtor. This frees all data, destroys locks, frees the hash
1680 * table and removes the keg from the global list.
1682 * Arguments/Returns follow uma_dtor specifications
1686 keg_dtor(void *arg, int size, void *udata)
1690 keg = (uma_keg_t)arg;
1692 if (keg->uk_free != 0) {
1693 printf("Freed UMA keg (%s) was not empty (%d items). "
1694 " Lost %d pages of memory.\n",
1695 keg->uk_name ? keg->uk_name : "",
1696 keg->uk_free, keg->uk_pages);
1700 hash_free(&keg->uk_hash);
1708 * Arguments/Returns follow uma_dtor specifications
1712 zone_dtor(void *arg, int size, void *udata)
1718 zone = (uma_zone_t)arg;
1719 keg = zone_first_keg(zone);
1721 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1724 rw_wlock(&uma_rwlock);
1725 LIST_REMOVE(zone, uz_link);
1726 rw_wunlock(&uma_rwlock);
1728 * XXX there are some races here where
1729 * the zone can be drained but zone lock
1730 * released and then refilled before we
1731 * remove it... we dont care for now
1733 zone_drain_wait(zone, M_WAITOK);
1735 * Unlink all of our kegs.
1737 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1738 klink->kl_keg = NULL;
1739 LIST_REMOVE(klink, kl_link);
1740 if (klink == &zone->uz_klink)
1742 free(klink, M_TEMP);
1745 * We only destroy kegs from non secondary zones.
1747 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1748 rw_wlock(&uma_rwlock);
1749 LIST_REMOVE(keg, uk_link);
1750 rw_wunlock(&uma_rwlock);
1751 zone_free_item(kegs, keg, NULL, SKIP_NONE);
1753 ZONE_LOCK_FINI(zone);
1757 * Traverses every zone in the system and calls a callback
1760 * zfunc A pointer to a function which accepts a zone
1767 zone_foreach(void (*zfunc)(uma_zone_t))
1772 rw_rlock(&uma_rwlock);
1773 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1774 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1777 rw_runlock(&uma_rwlock);
1780 /* Public functions */
1783 uma_startup(void *bootmem, int boot_pages)
1785 struct uma_zctor_args args;
1791 printf("Creating uma keg headers zone and keg.\n");
1793 rw_init(&uma_rwlock, "UMA lock");
1795 /* "manually" create the initial zone */
1796 memset(&args, 0, sizeof(args));
1797 args.name = "UMA Kegs";
1798 args.size = sizeof(struct uma_keg);
1799 args.ctor = keg_ctor;
1800 args.dtor = keg_dtor;
1801 args.uminit = zero_init;
1803 args.keg = &masterkeg;
1804 args.align = 32 - 1;
1805 args.flags = UMA_ZFLAG_INTERNAL;
1806 /* The initial zone has no Per cpu queues so it's smaller */
1807 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1810 printf("Filling boot free list.\n");
1812 for (i = 0; i < boot_pages; i++) {
1813 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1814 slab->us_data = (uint8_t *)slab;
1815 slab->us_flags = UMA_SLAB_BOOT;
1816 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1818 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1821 printf("Creating uma zone headers zone and keg.\n");
1823 args.name = "UMA Zones";
1824 args.size = sizeof(struct uma_zone) +
1825 (sizeof(struct uma_cache) * (mp_maxid + 1));
1826 args.ctor = zone_ctor;
1827 args.dtor = zone_dtor;
1828 args.uminit = zero_init;
1831 args.align = 32 - 1;
1832 args.flags = UMA_ZFLAG_INTERNAL;
1833 /* The initial zone has no Per cpu queues so it's smaller */
1834 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1837 printf("Initializing pcpu cache locks.\n");
1840 printf("Creating slab and hash zones.\n");
1843 /* Now make a zone for slab headers */
1844 slabzone = uma_zcreate("UMA Slabs",
1845 sizeof(struct uma_slab),
1846 NULL, NULL, NULL, NULL,
1847 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1850 * We also create a zone for the bigger slabs with reference
1851 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1853 slabsize = sizeof(struct uma_slab_refcnt);
1854 slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1855 slabrefzone = uma_zcreate("UMA RCntSlabs",
1857 NULL, NULL, NULL, NULL,
1859 UMA_ZFLAG_INTERNAL);
1861 hashzone = uma_zcreate("UMA Hash",
1862 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1863 NULL, NULL, NULL, NULL,
1864 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1868 booted = UMA_STARTUP;
1871 printf("UMA startup complete.\n");
1879 booted = UMA_STARTUP2;
1881 sx_init(&uma_drain_lock, "umadrain");
1883 printf("UMA startup2 complete.\n");
1888 * Initialize our callout handle
1896 printf("Starting callout.\n");
1898 callout_init(&uma_callout, 1);
1899 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1901 printf("UMA startup3 complete.\n");
1906 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1907 int align, uint32_t flags)
1909 struct uma_kctor_args args;
1912 args.uminit = uminit;
1914 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1917 return (zone_alloc_item(kegs, &args, M_WAITOK));
1922 uma_set_align(int align)
1925 if (align != UMA_ALIGN_CACHE)
1926 uma_align_cache = align;
1931 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1932 uma_init uminit, uma_fini fini, int align, uint32_t flags)
1935 struct uma_zctor_args args;
1939 /* This stuff is essential for the zone ctor */
1940 memset(&args, 0, sizeof(args));
1945 args.uminit = uminit;
1951 if (booted < UMA_STARTUP2) {
1954 sx_slock(&uma_drain_lock);
1957 res = zone_alloc_item(zones, &args, M_WAITOK);
1959 sx_sunlock(&uma_drain_lock);
1965 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1966 uma_init zinit, uma_fini zfini, uma_zone_t master)
1968 struct uma_zctor_args args;
1973 keg = zone_first_keg(master);
1974 memset(&args, 0, sizeof(args));
1976 args.size = keg->uk_size;
1979 args.uminit = zinit;
1981 args.align = keg->uk_align;
1982 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1985 if (booted < UMA_STARTUP2) {
1988 sx_slock(&uma_drain_lock);
1991 /* XXX Attaches only one keg of potentially many. */
1992 res = zone_alloc_item(zones, &args, M_WAITOK);
1994 sx_sunlock(&uma_drain_lock);
2000 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
2001 uma_init zinit, uma_fini zfini, uma_import zimport,
2002 uma_release zrelease, void *arg, int flags)
2004 struct uma_zctor_args args;
2006 memset(&args, 0, sizeof(args));
2011 args.uminit = zinit;
2013 args.import = zimport;
2014 args.release = zrelease;
2019 return (zone_alloc_item(zones, &args, M_WAITOK));
2023 zone_lock_pair(uma_zone_t a, uma_zone_t b)
2027 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
2030 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
2035 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2043 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2050 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2052 zone_lock_pair(zone, master);
2054 * zone must use vtoslab() to resolve objects and must already be
2057 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2058 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2063 * The new master must also use vtoslab().
2065 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2070 * Both must either be refcnt, or not be refcnt.
2072 if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
2073 (master->uz_flags & UMA_ZONE_REFCNT)) {
2078 * The underlying object must be the same size. rsize
2081 if (master->uz_size != zone->uz_size) {
2086 * Put it at the end of the list.
2088 klink->kl_keg = zone_first_keg(master);
2089 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2090 if (LIST_NEXT(kl, kl_link) == NULL) {
2091 LIST_INSERT_AFTER(kl, klink, kl_link);
2096 zone->uz_flags |= UMA_ZFLAG_MULTI;
2097 zone->uz_slab = zone_fetch_slab_multi;
2100 zone_unlock_pair(zone, master);
2102 free(klink, M_TEMP);
2110 uma_zdestroy(uma_zone_t zone)
2113 sx_slock(&uma_drain_lock);
2114 zone_free_item(zones, zone, NULL, SKIP_NONE);
2115 sx_sunlock(&uma_drain_lock);
2120 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2124 uma_bucket_t bucket;
2128 /* This is the fast path allocation */
2129 #ifdef UMA_DEBUG_ALLOC_1
2130 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2132 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2133 zone->uz_name, flags);
2135 if (flags & M_WAITOK) {
2136 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2137 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2139 #ifdef DEBUG_MEMGUARD
2140 if (memguard_cmp_zone(zone)) {
2141 item = memguard_alloc(zone->uz_size, flags);
2144 * Avoid conflict with the use-after-free
2145 * protecting infrastructure from INVARIANTS.
2147 if (zone->uz_init != NULL &&
2148 zone->uz_init != mtrash_init &&
2149 zone->uz_init(item, zone->uz_size, flags) != 0)
2151 if (zone->uz_ctor != NULL &&
2152 zone->uz_ctor != mtrash_ctor &&
2153 zone->uz_ctor(item, zone->uz_size, udata,
2155 zone->uz_fini(item, zone->uz_size);
2160 /* This is unfortunate but should not be fatal. */
2164 * If possible, allocate from the per-CPU cache. There are two
2165 * requirements for safe access to the per-CPU cache: (1) the thread
2166 * accessing the cache must not be preempted or yield during access,
2167 * and (2) the thread must not migrate CPUs without switching which
2168 * cache it accesses. We rely on a critical section to prevent
2169 * preemption and migration. We release the critical section in
2170 * order to acquire the zone mutex if we are unable to allocate from
2171 * the current cache; when we re-acquire the critical section, we
2172 * must detect and handle migration if it has occurred.
2176 cache = &zone->uz_cpu[cpu];
2179 bucket = cache->uc_allocbucket;
2180 if (bucket != NULL && bucket->ub_cnt > 0) {
2182 item = bucket->ub_bucket[bucket->ub_cnt];
2184 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2186 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2189 if (zone->uz_ctor != NULL &&
2190 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2191 atomic_add_long(&zone->uz_fails, 1);
2192 zone_free_item(zone, item, udata, SKIP_DTOR);
2196 uma_dbg_alloc(zone, NULL, item);
2199 uma_zero_item(item, zone);
2204 * We have run out of items in our alloc bucket.
2205 * See if we can switch with our free bucket.
2207 bucket = cache->uc_freebucket;
2208 if (bucket != NULL && bucket->ub_cnt > 0) {
2209 #ifdef UMA_DEBUG_ALLOC
2210 printf("uma_zalloc: Swapping empty with alloc.\n");
2212 cache->uc_freebucket = cache->uc_allocbucket;
2213 cache->uc_allocbucket = bucket;
2218 * Discard any empty allocation bucket while we hold no locks.
2220 bucket = cache->uc_allocbucket;
2221 cache->uc_allocbucket = NULL;
2224 bucket_free(zone, bucket, udata);
2226 /* Short-circuit for zones without buckets and low memory. */
2227 if (zone->uz_count == 0 || bucketdisable)
2231 * Attempt to retrieve the item from the per-CPU cache has failed, so
2232 * we must go back to the zone. This requires the zone lock, so we
2233 * must drop the critical section, then re-acquire it when we go back
2234 * to the cache. Since the critical section is released, we may be
2235 * preempted or migrate. As such, make sure not to maintain any
2236 * thread-local state specific to the cache from prior to releasing
2237 * the critical section.
2240 if (ZONE_TRYLOCK(zone) == 0) {
2241 /* Record contention to size the buckets. */
2247 cache = &zone->uz_cpu[cpu];
2250 * Since we have locked the zone we may as well send back our stats.
2252 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2253 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2254 cache->uc_allocs = 0;
2255 cache->uc_frees = 0;
2257 /* See if we lost the race to fill the cache. */
2258 if (cache->uc_allocbucket != NULL) {
2264 * Check the zone's cache of buckets.
2266 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2267 KASSERT(bucket->ub_cnt != 0,
2268 ("uma_zalloc_arg: Returning an empty bucket."));
2270 LIST_REMOVE(bucket, ub_link);
2271 cache->uc_allocbucket = bucket;
2275 /* We are no longer associated with this CPU. */
2279 * We bump the uz count when the cache size is insufficient to
2280 * handle the working set.
2282 if (lockfail && zone->uz_count < BUCKET_MAX)
2287 * Now lets just fill a bucket and put it on the free list. If that
2288 * works we'll restart the allocation from the begining and it
2289 * will use the just filled bucket.
2291 bucket = zone_alloc_bucket(zone, udata, flags);
2292 if (bucket != NULL) {
2296 cache = &zone->uz_cpu[cpu];
2298 * See if we lost the race or were migrated. Cache the
2299 * initialized bucket to make this less likely or claim
2300 * the memory directly.
2302 if (cache->uc_allocbucket == NULL)
2303 cache->uc_allocbucket = bucket;
2305 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2311 * We may not be able to get a bucket so return an actual item.
2314 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2318 item = zone_alloc_item(zone, udata, flags);
2324 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2329 mtx_assert(&keg->uk_lock, MA_OWNED);
2332 if ((flags & M_USE_RESERVE) == 0)
2333 reserve = keg->uk_reserve;
2337 * Find a slab with some space. Prefer slabs that are partially
2338 * used over those that are totally full. This helps to reduce
2341 if (keg->uk_free > reserve) {
2342 if (!LIST_EMPTY(&keg->uk_part_slab)) {
2343 slab = LIST_FIRST(&keg->uk_part_slab);
2345 slab = LIST_FIRST(&keg->uk_free_slab);
2346 LIST_REMOVE(slab, us_link);
2347 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2350 MPASS(slab->us_keg == keg);
2355 * M_NOVM means don't ask at all!
2360 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2361 keg->uk_flags |= UMA_ZFLAG_FULL;
2363 * If this is not a multi-zone, set the FULL bit.
2364 * Otherwise slab_multi() takes care of it.
2366 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2367 zone->uz_flags |= UMA_ZFLAG_FULL;
2368 zone_log_warning(zone);
2370 if (flags & M_NOWAIT)
2373 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2376 slab = keg_alloc_slab(keg, zone, flags);
2378 * If we got a slab here it's safe to mark it partially used
2379 * and return. We assume that the caller is going to remove
2380 * at least one item.
2383 MPASS(slab->us_keg == keg);
2384 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2388 * We might not have been able to get a slab but another cpu
2389 * could have while we were unlocked. Check again before we
2398 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2403 keg = zone_first_keg(zone);
2408 slab = keg_fetch_slab(keg, zone, flags);
2411 if (flags & (M_NOWAIT | M_NOVM))
2419 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2420 * with the keg locked. On NULL no lock is held.
2422 * The last pointer is used to seed the search. It is not required.
2425 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2435 * Don't wait on the first pass. This will skip limit tests
2436 * as well. We don't want to block if we can find a provider
2439 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2441 * Use the last slab allocated as a hint for where to start
2445 slab = keg_fetch_slab(last, zone, flags);
2451 * Loop until we have a slab incase of transient failures
2452 * while M_WAITOK is specified. I'm not sure this is 100%
2453 * required but we've done it for so long now.
2459 * Search the available kegs for slabs. Be careful to hold the
2460 * correct lock while calling into the keg layer.
2462 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2463 keg = klink->kl_keg;
2465 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2466 slab = keg_fetch_slab(keg, zone, flags);
2470 if (keg->uk_flags & UMA_ZFLAG_FULL)
2476 if (rflags & (M_NOWAIT | M_NOVM))
2480 * All kegs are full. XXX We can't atomically check all kegs
2481 * and sleep so just sleep for a short period and retry.
2483 if (full && !empty) {
2485 zone->uz_flags |= UMA_ZFLAG_FULL;
2487 zone_log_warning(zone);
2488 msleep(zone, zone->uz_lockptr, PVM,
2489 "zonelimit", hz/100);
2490 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2499 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2504 MPASS(keg == slab->us_keg);
2505 mtx_assert(&keg->uk_lock, MA_OWNED);
2507 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2508 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2509 item = slab->us_data + (keg->uk_rsize * freei);
2510 slab->us_freecount--;
2513 /* Move this slab to the full list */
2514 if (slab->us_freecount == 0) {
2515 LIST_REMOVE(slab, us_link);
2516 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2523 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2531 /* Try to keep the buckets totally full */
2532 for (i = 0; i < max; ) {
2533 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2536 while (slab->us_freecount && i < max) {
2537 bucket[i++] = slab_alloc_item(keg, slab);
2538 if (keg->uk_free <= keg->uk_reserve)
2541 /* Don't grab more than one slab at a time. */
2552 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2554 uma_bucket_t bucket;
2557 /* Don't wait for buckets, preserve caller's NOVM setting. */
2558 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2562 max = MIN(bucket->ub_entries, zone->uz_count);
2563 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2567 * Initialize the memory if necessary.
2569 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2572 for (i = 0; i < bucket->ub_cnt; i++)
2573 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2577 * If we couldn't initialize the whole bucket, put the
2578 * rest back onto the freelist.
2580 if (i != bucket->ub_cnt) {
2581 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2582 bucket->ub_cnt - i);
2584 bzero(&bucket->ub_bucket[i],
2585 sizeof(void *) * (bucket->ub_cnt - i));
2591 if (bucket->ub_cnt == 0) {
2592 bucket_free(zone, bucket, udata);
2593 atomic_add_long(&zone->uz_fails, 1);
2601 * Allocates a single item from a zone.
2604 * zone The zone to alloc for.
2605 * udata The data to be passed to the constructor.
2606 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2609 * NULL if there is no memory and M_NOWAIT is set
2610 * An item if successful
2614 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2620 #ifdef UMA_DEBUG_ALLOC
2621 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2623 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2625 atomic_add_long(&zone->uz_allocs, 1);
2628 * We have to call both the zone's init (not the keg's init)
2629 * and the zone's ctor. This is because the item is going from
2630 * a keg slab directly to the user, and the user is expecting it
2631 * to be both zone-init'd as well as zone-ctor'd.
2633 if (zone->uz_init != NULL) {
2634 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2635 zone_free_item(zone, item, udata, SKIP_FINI);
2639 if (zone->uz_ctor != NULL) {
2640 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2641 zone_free_item(zone, item, udata, SKIP_DTOR);
2646 uma_dbg_alloc(zone, NULL, item);
2649 uma_zero_item(item, zone);
2654 atomic_add_long(&zone->uz_fails, 1);
2660 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2663 uma_bucket_t bucket;
2667 #ifdef UMA_DEBUG_ALLOC_1
2668 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2670 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2673 /* uma_zfree(..., NULL) does nothing, to match free(9). */
2676 #ifdef DEBUG_MEMGUARD
2677 if (is_memguard_addr(item)) {
2678 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2679 zone->uz_dtor(item, zone->uz_size, udata);
2680 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2681 zone->uz_fini(item, zone->uz_size);
2682 memguard_free(item);
2687 if (zone->uz_flags & UMA_ZONE_MALLOC)
2688 uma_dbg_free(zone, udata, item);
2690 uma_dbg_free(zone, NULL, item);
2692 if (zone->uz_dtor != NULL)
2693 zone->uz_dtor(item, zone->uz_size, udata);
2696 * The race here is acceptable. If we miss it we'll just have to wait
2697 * a little longer for the limits to be reset.
2699 if (zone->uz_flags & UMA_ZFLAG_FULL)
2703 * If possible, free to the per-CPU cache. There are two
2704 * requirements for safe access to the per-CPU cache: (1) the thread
2705 * accessing the cache must not be preempted or yield during access,
2706 * and (2) the thread must not migrate CPUs without switching which
2707 * cache it accesses. We rely on a critical section to prevent
2708 * preemption and migration. We release the critical section in
2709 * order to acquire the zone mutex if we are unable to free to the
2710 * current cache; when we re-acquire the critical section, we must
2711 * detect and handle migration if it has occurred.
2716 cache = &zone->uz_cpu[cpu];
2720 * Try to free into the allocbucket first to give LIFO ordering
2721 * for cache-hot datastructures. Spill over into the freebucket
2722 * if necessary. Alloc will swap them if one runs dry.
2724 bucket = cache->uc_allocbucket;
2725 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2726 bucket = cache->uc_freebucket;
2727 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2728 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2729 ("uma_zfree: Freeing to non free bucket index."));
2730 bucket->ub_bucket[bucket->ub_cnt] = item;
2738 * We must go back the zone, which requires acquiring the zone lock,
2739 * which in turn means we must release and re-acquire the critical
2740 * section. Since the critical section is released, we may be
2741 * preempted or migrate. As such, make sure not to maintain any
2742 * thread-local state specific to the cache from prior to releasing
2743 * the critical section.
2746 if (zone->uz_count == 0 || bucketdisable)
2750 if (ZONE_TRYLOCK(zone) == 0) {
2751 /* Record contention to size the buckets. */
2757 cache = &zone->uz_cpu[cpu];
2760 * Since we have locked the zone we may as well send back our stats.
2762 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2763 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2764 cache->uc_allocs = 0;
2765 cache->uc_frees = 0;
2767 bucket = cache->uc_freebucket;
2768 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2772 cache->uc_freebucket = NULL;
2773 /* We are no longer associated with this CPU. */
2776 /* Can we throw this on the zone full list? */
2777 if (bucket != NULL) {
2778 #ifdef UMA_DEBUG_ALLOC
2779 printf("uma_zfree: Putting old bucket on the free list.\n");
2781 /* ub_cnt is pointing to the last free item */
2782 KASSERT(bucket->ub_cnt != 0,
2783 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2784 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2788 * We bump the uz count when the cache size is insufficient to
2789 * handle the working set.
2791 if (lockfail && zone->uz_count < BUCKET_MAX)
2795 #ifdef UMA_DEBUG_ALLOC
2796 printf("uma_zfree: Allocating new free bucket.\n");
2798 bucket = bucket_alloc(zone, udata, M_NOWAIT);
2802 cache = &zone->uz_cpu[cpu];
2803 if (cache->uc_freebucket == NULL) {
2804 cache->uc_freebucket = bucket;
2808 * We lost the race, start over. We have to drop our
2809 * critical section to free the bucket.
2812 bucket_free(zone, bucket, udata);
2817 * If nothing else caught this, we'll just do an internal free.
2820 zone_free_item(zone, item, udata, SKIP_DTOR);
2826 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2830 mtx_assert(&keg->uk_lock, MA_OWNED);
2831 MPASS(keg == slab->us_keg);
2833 /* Do we need to remove from any lists? */
2834 if (slab->us_freecount+1 == keg->uk_ipers) {
2835 LIST_REMOVE(slab, us_link);
2836 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2837 } else if (slab->us_freecount == 0) {
2838 LIST_REMOVE(slab, us_link);
2839 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2842 /* Slab management. */
2843 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2844 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2845 slab->us_freecount++;
2847 /* Keg statistics. */
2852 zone_release(uma_zone_t zone, void **bucket, int cnt)
2862 keg = zone_first_keg(zone);
2864 for (i = 0; i < cnt; i++) {
2866 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2867 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2868 if (zone->uz_flags & UMA_ZONE_HASH) {
2869 slab = hash_sfind(&keg->uk_hash, mem);
2871 mem += keg->uk_pgoff;
2872 slab = (uma_slab_t)mem;
2875 slab = vtoslab((vm_offset_t)item);
2876 if (slab->us_keg != keg) {
2882 slab_free_item(keg, slab, item);
2883 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2884 if (keg->uk_pages < keg->uk_maxpages) {
2885 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2890 * We can handle one more allocation. Since we're
2891 * clearing ZFLAG_FULL, wake up all procs blocked
2892 * on pages. This should be uncommon, so keeping this
2893 * simple for now (rather than adding count of blocked
2902 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2910 * Frees a single item to any zone.
2913 * zone The zone to free to
2914 * item The item we're freeing
2915 * udata User supplied data for the dtor
2916 * skip Skip dtors and finis
2919 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2923 if (skip == SKIP_NONE) {
2924 if (zone->uz_flags & UMA_ZONE_MALLOC)
2925 uma_dbg_free(zone, udata, item);
2927 uma_dbg_free(zone, NULL, item);
2930 if (skip < SKIP_DTOR && zone->uz_dtor)
2931 zone->uz_dtor(item, zone->uz_size, udata);
2933 if (skip < SKIP_FINI && zone->uz_fini)
2934 zone->uz_fini(item, zone->uz_size);
2936 atomic_add_long(&zone->uz_frees, 1);
2937 zone->uz_release(zone->uz_arg, &item, 1);
2942 uma_zone_set_max(uma_zone_t zone, int nitems)
2946 keg = zone_first_keg(zone);
2950 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2951 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2952 keg->uk_maxpages += keg->uk_ppera;
2953 nitems = keg->uk_maxpages * keg->uk_ipers;
2961 uma_zone_get_max(uma_zone_t zone)
2966 keg = zone_first_keg(zone);
2970 nitems = keg->uk_maxpages * keg->uk_ipers;
2978 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2982 zone->uz_warning = warning;
2988 uma_zone_get_cur(uma_zone_t zone)
2994 nitems = zone->uz_allocs - zone->uz_frees;
2997 * See the comment in sysctl_vm_zone_stats() regarding the
2998 * safety of accessing the per-cpu caches. With the zone lock
2999 * held, it is safe, but can potentially result in stale data.
3001 nitems += zone->uz_cpu[i].uc_allocs -
3002 zone->uz_cpu[i].uc_frees;
3006 return (nitems < 0 ? 0 : nitems);
3011 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
3015 keg = zone_first_keg(zone);
3016 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3018 KASSERT(keg->uk_pages == 0,
3019 ("uma_zone_set_init on non-empty keg"));
3020 keg->uk_init = uminit;
3026 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3030 keg = zone_first_keg(zone);
3031 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
3033 KASSERT(keg->uk_pages == 0,
3034 ("uma_zone_set_fini on non-empty keg"));
3035 keg->uk_fini = fini;
3041 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3045 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3046 ("uma_zone_set_zinit on non-empty keg"));
3047 zone->uz_init = zinit;
3053 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3057 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3058 ("uma_zone_set_zfini on non-empty keg"));
3059 zone->uz_fini = zfini;
3064 /* XXX uk_freef is not actually used with the zone locked */
3066 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3070 keg = zone_first_keg(zone);
3071 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
3073 keg->uk_freef = freef;
3078 /* XXX uk_allocf is not actually used with the zone locked */
3080 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3084 keg = zone_first_keg(zone);
3086 keg->uk_allocf = allocf;
3092 uma_zone_reserve(uma_zone_t zone, int items)
3096 keg = zone_first_keg(zone);
3100 keg->uk_reserve = items;
3108 uma_zone_reserve_kva(uma_zone_t zone, int count)
3114 keg = zone_first_keg(zone);
3117 pages = count / keg->uk_ipers;
3119 if (pages * keg->uk_ipers < count)
3122 #ifdef UMA_MD_SMALL_ALLOC
3123 if (keg->uk_ppera > 1) {
3127 kva = kva_alloc(pages * UMA_SLAB_SIZE);
3135 keg->uk_maxpages = pages;
3136 #ifdef UMA_MD_SMALL_ALLOC
3137 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3139 keg->uk_allocf = noobj_alloc;
3141 keg->uk_flags |= UMA_ZONE_NOFREE;
3149 uma_prealloc(uma_zone_t zone, int items)
3155 keg = zone_first_keg(zone);
3159 slabs = items / keg->uk_ipers;
3160 if (slabs * keg->uk_ipers < items)
3163 slab = keg_alloc_slab(keg, zone, M_WAITOK);
3166 MPASS(slab->us_keg == keg);
3167 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3175 uma_find_refcnt(uma_zone_t zone, void *item)
3177 uma_slabrefcnt_t slabref;
3183 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3184 slabref = (uma_slabrefcnt_t)slab;
3186 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3187 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3188 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3189 refcnt = &slabref->us_refcnt[idx];
3195 uma_reclaim_locked(bool kmem_danger)
3199 printf("UMA: vm asked us to release pages!\n");
3201 sx_assert(&uma_drain_lock, SA_XLOCKED);
3203 zone_foreach(zone_drain);
3204 if (vm_page_count_min() || kmem_danger) {
3205 cache_drain_safe(NULL);
3206 zone_foreach(zone_drain);
3209 * Some slabs may have been freed but this zone will be visited early
3210 * we visit again so that we can free pages that are empty once other
3211 * zones are drained. We have to do the same for buckets.
3213 zone_drain(slabzone);
3214 zone_drain(slabrefzone);
3215 bucket_zone_drain();
3222 sx_xlock(&uma_drain_lock);
3223 uma_reclaim_locked(false);
3224 sx_xunlock(&uma_drain_lock);
3227 static int uma_reclaim_needed;
3230 uma_reclaim_wakeup(void)
3233 uma_reclaim_needed = 1;
3234 wakeup(&uma_reclaim_needed);
3238 uma_reclaim_worker(void *arg __unused)
3241 sx_xlock(&uma_drain_lock);
3243 sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM,
3245 if (uma_reclaim_needed) {
3246 uma_reclaim_needed = 0;
3247 sx_xunlock(&uma_drain_lock);
3248 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
3249 sx_xlock(&uma_drain_lock);
3250 uma_reclaim_locked(true);
3257 uma_zone_exhausted(uma_zone_t zone)
3262 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3268 uma_zone_exhausted_nolock(uma_zone_t zone)
3270 return (zone->uz_flags & UMA_ZFLAG_FULL);
3274 uma_large_malloc(vm_size_t size, int wait)
3280 slab = zone_alloc_item(slabzone, NULL, wait);
3283 mem = page_alloc(NULL, size, &flags, wait);
3285 vsetslab((vm_offset_t)mem, slab);
3286 slab->us_data = mem;
3287 slab->us_flags = flags | UMA_SLAB_MALLOC;
3288 slab->us_size = size;
3290 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3297 uma_large_free(uma_slab_t slab)
3300 page_free(slab->us_data, slab->us_size, slab->us_flags);
3301 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3305 uma_zero_item(void *item, uma_zone_t zone)
3308 if (zone->uz_flags & UMA_ZONE_PCPU) {
3309 for (int i = 0; i < mp_ncpus; i++)
3310 bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3312 bzero(item, zone->uz_size);
3316 uma_print_stats(void)
3318 zone_foreach(uma_print_zone);
3322 slab_print(uma_slab_t slab)
3324 printf("slab: keg %p, data %p, freecount %d\n",
3325 slab->us_keg, slab->us_data, slab->us_freecount);
3329 cache_print(uma_cache_t cache)
3331 printf("alloc: %p(%d), free: %p(%d)\n",
3332 cache->uc_allocbucket,
3333 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3334 cache->uc_freebucket,
3335 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3339 uma_print_keg(uma_keg_t keg)
3343 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3344 "out %d free %d limit %d\n",
3345 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3346 keg->uk_ipers, keg->uk_ppera,
3347 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3348 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3349 printf("Part slabs:\n");
3350 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3352 printf("Free slabs:\n");
3353 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3355 printf("Full slabs:\n");
3356 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3361 uma_print_zone(uma_zone_t zone)
3367 printf("zone: %s(%p) size %d flags %#x\n",
3368 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3369 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3370 uma_print_keg(kl->kl_keg);
3372 cache = &zone->uz_cpu[i];
3373 printf("CPU %d Cache:\n", i);
3380 * Generate statistics across both the zone and its per-cpu cache's. Return
3381 * desired statistics if the pointer is non-NULL for that statistic.
3383 * Note: does not update the zone statistics, as it can't safely clear the
3384 * per-CPU cache statistic.
3386 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3387 * safe from off-CPU; we should modify the caches to track this information
3388 * directly so that we don't have to.
3391 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3392 uint64_t *freesp, uint64_t *sleepsp)
3395 uint64_t allocs, frees, sleeps;
3398 allocs = frees = sleeps = 0;
3401 cache = &z->uz_cpu[cpu];
3402 if (cache->uc_allocbucket != NULL)
3403 cachefree += cache->uc_allocbucket->ub_cnt;
3404 if (cache->uc_freebucket != NULL)
3405 cachefree += cache->uc_freebucket->ub_cnt;
3406 allocs += cache->uc_allocs;
3407 frees += cache->uc_frees;
3409 allocs += z->uz_allocs;
3410 frees += z->uz_frees;
3411 sleeps += z->uz_sleeps;
3412 if (cachefreep != NULL)
3413 *cachefreep = cachefree;
3414 if (allocsp != NULL)
3418 if (sleepsp != NULL)
3424 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3431 rw_rlock(&uma_rwlock);
3432 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3433 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3436 rw_runlock(&uma_rwlock);
3437 return (sysctl_handle_int(oidp, &count, 0, req));
3441 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3443 struct uma_stream_header ush;
3444 struct uma_type_header uth;
3445 struct uma_percpu_stat ups;
3446 uma_bucket_t bucket;
3453 int count, error, i;
3455 error = sysctl_wire_old_buffer(req, 0);
3458 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3461 rw_rlock(&uma_rwlock);
3462 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3463 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3468 * Insert stream header.
3470 bzero(&ush, sizeof(ush));
3471 ush.ush_version = UMA_STREAM_VERSION;
3472 ush.ush_maxcpus = (mp_maxid + 1);
3473 ush.ush_count = count;
3474 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3476 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3477 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3478 bzero(&uth, sizeof(uth));
3480 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3481 uth.uth_align = kz->uk_align;
3482 uth.uth_size = kz->uk_size;
3483 uth.uth_rsize = kz->uk_rsize;
3484 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3486 uth.uth_maxpages += k->uk_maxpages;
3487 uth.uth_pages += k->uk_pages;
3488 uth.uth_keg_free += k->uk_free;
3489 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3494 * A zone is secondary is it is not the first entry
3495 * on the keg's zone list.
3497 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3498 (LIST_FIRST(&kz->uk_zones) != z))
3499 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3501 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3502 uth.uth_zone_free += bucket->ub_cnt;
3503 uth.uth_allocs = z->uz_allocs;
3504 uth.uth_frees = z->uz_frees;
3505 uth.uth_fails = z->uz_fails;
3506 uth.uth_sleeps = z->uz_sleeps;
3507 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3509 * While it is not normally safe to access the cache
3510 * bucket pointers while not on the CPU that owns the
3511 * cache, we only allow the pointers to be exchanged
3512 * without the zone lock held, not invalidated, so
3513 * accept the possible race associated with bucket
3514 * exchange during monitoring.
3516 for (i = 0; i < (mp_maxid + 1); i++) {
3517 bzero(&ups, sizeof(ups));
3518 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3522 cache = &z->uz_cpu[i];
3523 if (cache->uc_allocbucket != NULL)
3524 ups.ups_cache_free +=
3525 cache->uc_allocbucket->ub_cnt;
3526 if (cache->uc_freebucket != NULL)
3527 ups.ups_cache_free +=
3528 cache->uc_freebucket->ub_cnt;
3529 ups.ups_allocs = cache->uc_allocs;
3530 ups.ups_frees = cache->uc_frees;
3532 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3537 rw_runlock(&uma_rwlock);
3538 error = sbuf_finish(&sbuf);
3544 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3546 uma_zone_t zone = *(uma_zone_t *)arg1;
3547 int error, max, old;
3549 old = max = uma_zone_get_max(zone);
3550 error = sysctl_handle_int(oidp, &max, 0, req);
3551 if (error || !req->newptr)
3557 uma_zone_set_max(zone, max);
3563 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3565 uma_zone_t zone = *(uma_zone_t *)arg1;
3568 cur = uma_zone_get_cur(zone);
3569 return (sysctl_handle_int(oidp, &cur, 0, req));
3573 DB_SHOW_COMMAND(uma, db_show_uma)
3575 uint64_t allocs, frees, sleeps;
3576 uma_bucket_t bucket;
3581 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3582 "Free", "Requests", "Sleeps", "Bucket");
3583 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3584 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3585 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3586 allocs = z->uz_allocs;
3587 frees = z->uz_frees;
3588 sleeps = z->uz_sleeps;
3591 uma_zone_sumstat(z, &cachefree, &allocs,
3593 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3594 (LIST_FIRST(&kz->uk_zones) != z)))
3595 cachefree += kz->uk_free;
3596 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3597 cachefree += bucket->ub_cnt;
3598 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3599 z->uz_name, (uintmax_t)kz->uk_size,
3600 (intmax_t)(allocs - frees), cachefree,
3601 (uintmax_t)allocs, sleeps, z->uz_count);
3608 DB_SHOW_COMMAND(umacache, db_show_umacache)
3610 uint64_t allocs, frees;
3611 uma_bucket_t bucket;
3615 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3616 "Requests", "Bucket");
3617 LIST_FOREACH(z, &uma_cachezones, uz_link) {
3618 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3619 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3620 cachefree += bucket->ub_cnt;
3621 db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3622 z->uz_name, (uintmax_t)z->uz_size,
3623 (intmax_t)(allocs - frees), cachefree,
3624 (uintmax_t)allocs, z->uz_count);