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 * efficient. 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("$FreeBSD$");
54 #include "opt_param.h"
57 #include <sys/param.h>
58 #include <sys/systm.h>
59 #include <sys/bitset.h>
60 #include <sys/eventhandler.h>
61 #include <sys/kernel.h>
62 #include <sys/types.h>
63 #include <sys/queue.h>
64 #include <sys/malloc.h>
67 #include <sys/sysctl.h>
68 #include <sys/mutex.h>
70 #include <sys/random.h>
71 #include <sys/rwlock.h>
73 #include <sys/sched.h>
75 #include <sys/taskqueue.h>
76 #include <sys/vmmeter.h>
79 #include <vm/vm_object.h>
80 #include <vm/vm_page.h>
81 #include <vm/vm_pageout.h>
82 #include <vm/vm_param.h>
83 #include <vm/vm_map.h>
84 #include <vm/vm_kern.h>
85 #include <vm/vm_extern.h>
87 #include <vm/uma_int.h>
88 #include <vm/uma_dbg.h>
93 #include <vm/memguard.h>
97 * This is the zone and keg from which all zones are spawned. The idea is that
98 * even the zone & keg heads are allocated from the allocator, so we use the
99 * bss section to bootstrap us.
101 static struct uma_keg masterkeg;
102 static struct uma_zone masterzone_k;
103 static struct uma_zone masterzone_z;
104 static uma_zone_t kegs = &masterzone_k;
105 static uma_zone_t zones = &masterzone_z;
107 /* This is the zone from which all of uma_slab_t's are allocated. */
108 static uma_zone_t slabzone;
111 * The initial hash tables come out of this zone so they can be allocated
112 * prior to malloc coming up.
114 static uma_zone_t hashzone;
116 /* The boot-time adjusted value for cache line alignment. */
117 int uma_align_cache = 64 - 1;
119 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
122 * Are we allowed to allocate buckets?
124 static int bucketdisable = 1;
126 /* Linked list of all kegs in the system */
127 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
129 /* Linked list of all cache-only zones in the system */
130 static LIST_HEAD(,uma_zone) uma_cachezones =
131 LIST_HEAD_INITIALIZER(uma_cachezones);
133 /* This RW lock protects the keg list */
134 static struct rwlock_padalign uma_rwlock;
137 * Pointer and counter to pool of pages, that is preallocated at
138 * startup to bootstrap UMA. Early zones continue to use the pool
139 * until it is depleted, so allocations may happen after boot, thus
140 * we need a mutex to protect it.
142 static char *bootmem;
143 static int boot_pages;
144 static struct mtx uma_boot_pages_mtx;
146 static struct sx uma_drain_lock;
148 /* Is the VM done starting up? */
149 static int booted = 0;
150 #define UMA_STARTUP 1
151 #define UMA_STARTUP2 2
154 * This is the handle used to schedule events that need to happen
155 * outside of the allocation fast path.
157 static struct callout uma_callout;
158 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
161 * This structure is passed as the zone ctor arg so that I don't have to create
162 * a special allocation function just for zones.
164 struct uma_zctor_args {
179 struct uma_kctor_args {
188 struct uma_bucket_zone {
191 int ubz_entries; /* Number of items it can hold. */
192 int ubz_maxsize; /* Maximum allocation size per-item. */
196 * Compute the actual number of bucket entries to pack them in power
197 * of two sizes for more efficient space utilization.
199 #define BUCKET_SIZE(n) \
200 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
202 #define BUCKET_MAX BUCKET_SIZE(256)
204 struct uma_bucket_zone bucket_zones[] = {
205 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
206 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
207 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
208 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
209 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
210 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
211 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
212 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
213 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
218 * Flags and enumerations to be passed to internal functions.
220 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
224 static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
225 static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
226 static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
227 static void page_free(void *, vm_size_t, uint8_t);
228 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
229 static void cache_drain(uma_zone_t);
230 static void bucket_drain(uma_zone_t, uma_bucket_t);
231 static void bucket_cache_drain(uma_zone_t zone);
232 static int keg_ctor(void *, int, void *, int);
233 static void keg_dtor(void *, int, void *);
234 static int zone_ctor(void *, int, void *, int);
235 static void zone_dtor(void *, int, void *);
236 static int zero_init(void *, int, int);
237 static void keg_small_init(uma_keg_t keg);
238 static void keg_large_init(uma_keg_t keg);
239 static void zone_foreach(void (*zfunc)(uma_zone_t));
240 static void zone_timeout(uma_zone_t zone);
241 static int hash_alloc(struct uma_hash *);
242 static int hash_expand(struct uma_hash *, struct uma_hash *);
243 static void hash_free(struct uma_hash *hash);
244 static void uma_timeout(void *);
245 static void uma_startup3(void);
246 static void *zone_alloc_item(uma_zone_t, void *, int);
247 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
248 static void bucket_enable(void);
249 static void bucket_init(void);
250 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
251 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
252 static void bucket_zone_drain(void);
253 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
254 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
255 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
256 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
257 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
258 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
259 uma_fini fini, int align, uint32_t flags);
260 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
261 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
262 static void uma_zero_item(void *item, uma_zone_t zone);
264 void uma_print_zone(uma_zone_t);
265 void uma_print_stats(void);
266 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
267 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
270 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
271 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
274 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
276 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
277 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
279 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
280 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
282 static int zone_warnings = 1;
283 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
284 "Warn when UMA zones becomes full");
287 * This routine checks to see whether or not it's safe to enable buckets.
292 bucketdisable = vm_page_count_min();
296 * Initialize bucket_zones, the array of zones of buckets of various sizes.
298 * For each zone, calculate the memory required for each bucket, consisting
299 * of the header and an array of pointers.
304 struct uma_bucket_zone *ubz;
307 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
308 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
309 size += sizeof(void *) * ubz->ubz_entries;
310 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
311 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
312 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
317 * Given a desired number of entries for a bucket, return the zone from which
318 * to allocate the bucket.
320 static struct uma_bucket_zone *
321 bucket_zone_lookup(int entries)
323 struct uma_bucket_zone *ubz;
325 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
326 if (ubz->ubz_entries >= entries)
333 bucket_select(int size)
335 struct uma_bucket_zone *ubz;
337 ubz = &bucket_zones[0];
338 if (size > ubz->ubz_maxsize)
339 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
341 for (; ubz->ubz_entries != 0; ubz++)
342 if (ubz->ubz_maxsize < size)
345 return (ubz->ubz_entries);
349 bucket_alloc(uma_zone_t zone, void *udata, int flags)
351 struct uma_bucket_zone *ubz;
355 * This is to stop us from allocating per cpu buckets while we're
356 * running out of vm.boot_pages. Otherwise, we would exhaust the
357 * boot pages. This also prevents us from allocating buckets in
358 * low memory situations.
363 * To limit bucket recursion we store the original zone flags
364 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
365 * NOVM flag to persist even through deep recursions. We also
366 * store ZFLAG_BUCKET once we have recursed attempting to allocate
367 * a bucket for a bucket zone so we do not allow infinite bucket
368 * recursion. This cookie will even persist to frees of unused
369 * buckets via the allocation path or bucket allocations in the
372 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
373 udata = (void *)(uintptr_t)zone->uz_flags;
375 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
377 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
379 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
381 ubz = bucket_zone_lookup(zone->uz_count);
382 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
384 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
387 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
390 bucket->ub_entries = ubz->ubz_entries;
397 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
399 struct uma_bucket_zone *ubz;
401 KASSERT(bucket->ub_cnt == 0,
402 ("bucket_free: Freeing a non free bucket."));
403 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
404 udata = (void *)(uintptr_t)zone->uz_flags;
405 ubz = bucket_zone_lookup(bucket->ub_entries);
406 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
410 bucket_zone_drain(void)
412 struct uma_bucket_zone *ubz;
414 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
415 zone_drain(ubz->ubz_zone);
419 zone_log_warning(uma_zone_t zone)
421 static const struct timeval warninterval = { 300, 0 };
423 if (!zone_warnings || zone->uz_warning == NULL)
426 if (ratecheck(&zone->uz_ratecheck, &warninterval))
427 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
431 zone_maxaction(uma_zone_t zone)
434 if (zone->uz_maxaction.ta_func != NULL)
435 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
439 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
443 LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
444 kegfn(klink->kl_keg);
448 * Routine called by timeout which is used to fire off some time interval
449 * based calculations. (stats, hash size, etc.)
458 uma_timeout(void *unused)
461 zone_foreach(zone_timeout);
463 /* Reschedule this event */
464 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
468 * Routine to perform timeout driven calculations. This expands the
469 * hashes and does per cpu statistics aggregation.
474 keg_timeout(uma_keg_t keg)
479 * Expand the keg hash table.
481 * This is done if the number of slabs is larger than the hash size.
482 * What I'm trying to do here is completely reduce collisions. This
483 * may be a little aggressive. Should I allow for two collisions max?
485 if (keg->uk_flags & UMA_ZONE_HASH &&
486 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
487 struct uma_hash newhash;
488 struct uma_hash oldhash;
492 * This is so involved because allocating and freeing
493 * while the keg lock is held will lead to deadlock.
494 * I have to do everything in stages and check for
497 newhash = keg->uk_hash;
499 ret = hash_alloc(&newhash);
502 if (hash_expand(&keg->uk_hash, &newhash)) {
503 oldhash = keg->uk_hash;
504 keg->uk_hash = newhash;
517 zone_timeout(uma_zone_t zone)
520 zone_foreach_keg(zone, &keg_timeout);
524 * Allocate and zero fill the next sized hash table from the appropriate
528 * hash A new hash structure with the old hash size in uh_hashsize
531 * 1 on success and 0 on failure.
534 hash_alloc(struct uma_hash *hash)
539 oldsize = hash->uh_hashsize;
541 /* We're just going to go to a power of two greater */
543 hash->uh_hashsize = oldsize * 2;
544 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
545 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
546 M_UMAHASH, M_NOWAIT);
548 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
549 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
551 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
553 if (hash->uh_slab_hash) {
554 bzero(hash->uh_slab_hash, alloc);
555 hash->uh_hashmask = hash->uh_hashsize - 1;
563 * Expands the hash table for HASH zones. This is done from zone_timeout
564 * to reduce collisions. This must not be done in the regular allocation
565 * path, otherwise, we can recurse on the vm while allocating pages.
568 * oldhash The hash you want to expand
569 * newhash The hash structure for the new table
577 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
583 if (!newhash->uh_slab_hash)
586 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
590 * I need to investigate hash algorithms for resizing without a
594 for (i = 0; i < oldhash->uh_hashsize; i++)
595 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
596 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
597 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
598 hval = UMA_HASH(newhash, slab->us_data);
599 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
607 * Free the hash bucket to the appropriate backing store.
610 * slab_hash The hash bucket we're freeing
611 * hashsize The number of entries in that hash bucket
617 hash_free(struct uma_hash *hash)
619 if (hash->uh_slab_hash == NULL)
621 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
622 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
624 free(hash->uh_slab_hash, M_UMAHASH);
628 * Frees all outstanding items in a bucket
631 * zone The zone to free to, must be unlocked.
632 * bucket The free/alloc bucket with items, cpu queue must be locked.
639 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
647 for (i = 0; i < bucket->ub_cnt; i++)
648 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
649 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
654 * Drains the per cpu caches for a zone.
656 * NOTE: This may only be called while the zone is being turn down, and not
657 * during normal operation. This is necessary in order that we do not have
658 * to migrate CPUs to drain the per-CPU caches.
661 * zone The zone to drain, must be unlocked.
667 cache_drain(uma_zone_t zone)
673 * XXX: It is safe to not lock the per-CPU caches, because we're
674 * tearing down the zone anyway. I.e., there will be no further use
675 * of the caches at this point.
677 * XXX: It would good to be able to assert that the zone is being
678 * torn down to prevent improper use of cache_drain().
680 * XXX: We lock the zone before passing into bucket_cache_drain() as
681 * it is used elsewhere. Should the tear-down path be made special
682 * there in some form?
685 cache = &zone->uz_cpu[cpu];
686 bucket_drain(zone, cache->uc_allocbucket);
687 bucket_drain(zone, cache->uc_freebucket);
688 if (cache->uc_allocbucket != NULL)
689 bucket_free(zone, cache->uc_allocbucket, NULL);
690 if (cache->uc_freebucket != NULL)
691 bucket_free(zone, cache->uc_freebucket, NULL);
692 cache->uc_allocbucket = cache->uc_freebucket = NULL;
695 bucket_cache_drain(zone);
700 cache_shrink(uma_zone_t zone)
703 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
707 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
712 cache_drain_safe_cpu(uma_zone_t zone)
717 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
723 cache = &zone->uz_cpu[curcpu];
724 if (cache->uc_allocbucket) {
725 if (cache->uc_allocbucket->ub_cnt != 0)
726 LIST_INSERT_HEAD(&zone->uz_buckets,
727 cache->uc_allocbucket, ub_link);
729 b1 = cache->uc_allocbucket;
730 cache->uc_allocbucket = NULL;
732 if (cache->uc_freebucket) {
733 if (cache->uc_freebucket->ub_cnt != 0)
734 LIST_INSERT_HEAD(&zone->uz_buckets,
735 cache->uc_freebucket, ub_link);
737 b2 = cache->uc_freebucket;
738 cache->uc_freebucket = NULL;
743 bucket_free(zone, b1, NULL);
745 bucket_free(zone, b2, NULL);
749 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
750 * This is an expensive call because it needs to bind to all CPUs
751 * one by one and enter a critical section on each of them in order
752 * to safely access their cache buckets.
753 * Zone lock must not be held on call this function.
756 cache_drain_safe(uma_zone_t zone)
761 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
766 zone_foreach(cache_shrink);
769 thread_lock(curthread);
770 sched_bind(curthread, cpu);
771 thread_unlock(curthread);
774 cache_drain_safe_cpu(zone);
776 zone_foreach(cache_drain_safe_cpu);
778 thread_lock(curthread);
779 sched_unbind(curthread);
780 thread_unlock(curthread);
784 * Drain the cached buckets from a zone. Expects a locked zone on entry.
787 bucket_cache_drain(uma_zone_t zone)
792 * Drain the bucket queues and free the buckets, we just keep two per
795 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
796 LIST_REMOVE(bucket, ub_link);
798 bucket_drain(zone, bucket);
799 bucket_free(zone, bucket, NULL);
804 * Shrink further bucket sizes. Price of single zone lock collision
805 * is probably lower then price of global cache drain.
807 if (zone->uz_count > zone->uz_count_min)
812 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
818 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
819 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
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);
831 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
835 * Frees pages from a keg back to the system. This is done on demand from
836 * the pageout daemon.
841 keg_drain(uma_keg_t keg)
843 struct slabhead freeslabs = { 0 };
844 uma_slab_t slab, tmp;
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)
853 CTR3(KTR_UMA, "keg_drain %s(%p) free items: %u",
854 keg->uk_name, keg, keg->uk_free);
856 if (keg->uk_free == 0)
859 LIST_FOREACH_SAFE(slab, &keg->uk_free_slab, us_link, tmp) {
860 /* We have nowhere to free these to. */
861 if (slab->us_flags & UMA_SLAB_BOOT)
864 LIST_REMOVE(slab, us_link);
865 keg->uk_pages -= keg->uk_ppera;
866 keg->uk_free -= keg->uk_ipers;
868 if (keg->uk_flags & UMA_ZONE_HASH)
869 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
871 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
876 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
877 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
878 keg_free_slab(keg, slab, keg->uk_ipers);
883 zone_drain_wait(uma_zone_t zone, int waitok)
887 * Set draining to interlock with zone_dtor() so we can release our
888 * locks as we go. Only dtor() should do a WAITOK call since it
889 * is the only call that knows the structure will still be available
893 while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
894 if (waitok == M_NOWAIT)
896 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
898 zone->uz_flags |= UMA_ZFLAG_DRAINING;
899 bucket_cache_drain(zone);
902 * The DRAINING flag protects us from being freed while
903 * we're running. Normally the uma_rwlock would protect us but we
904 * must be able to release and acquire the right lock for each keg.
906 zone_foreach_keg(zone, &keg_drain);
908 zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
915 zone_drain(uma_zone_t zone)
918 zone_drain_wait(zone, M_NOWAIT);
922 * Allocate a new slab for a keg. This does not insert the slab onto a list.
925 * wait Shall we wait?
928 * The slab that was allocated or NULL if there is no memory and the
929 * caller specified M_NOWAIT.
932 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
940 mtx_assert(&keg->uk_lock, MA_OWNED);
944 allocf = keg->uk_allocf;
947 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
948 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
954 * This reproduces the old vm_zone behavior of zero filling pages the
955 * first time they are added to a zone.
957 * Malloced items are zeroed in uma_zalloc.
960 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
965 if (keg->uk_flags & UMA_ZONE_NODUMP)
968 /* zone is passed for legacy reasons. */
969 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
971 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
972 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
977 /* Point the slab into the allocated memory */
978 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
979 slab = (uma_slab_t )(mem + keg->uk_pgoff);
981 if (keg->uk_flags & UMA_ZONE_VTOSLAB)
982 for (i = 0; i < keg->uk_ppera; i++)
983 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
987 slab->us_freecount = keg->uk_ipers;
988 slab->us_flags = flags;
989 BIT_FILL(SLAB_SETSIZE, &slab->us_free);
991 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
994 if (keg->uk_init != NULL) {
995 for (i = 0; i < keg->uk_ipers; i++)
996 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
997 keg->uk_size, wait) != 0)
999 if (i != keg->uk_ipers) {
1000 keg_free_slab(keg, slab, i);
1008 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1009 slab, keg->uk_name, keg);
1012 if (keg->uk_flags & UMA_ZONE_HASH)
1013 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1015 keg->uk_pages += keg->uk_ppera;
1016 keg->uk_free += keg->uk_ipers;
1023 * This function is intended to be used early on in place of page_alloc() so
1024 * that we may use the boot time page cache to satisfy allocations before
1028 startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1034 keg = zone_first_keg(zone);
1035 pages = howmany(bytes, PAGE_SIZE);
1036 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1039 * Check our small startup cache to see if it has pages remaining.
1041 mtx_lock(&uma_boot_pages_mtx);
1042 if (pages <= boot_pages) {
1044 boot_pages -= pages;
1045 bootmem += pages * PAGE_SIZE;
1046 mtx_unlock(&uma_boot_pages_mtx);
1047 *pflag = UMA_SLAB_BOOT;
1050 mtx_unlock(&uma_boot_pages_mtx);
1051 if (booted < UMA_STARTUP2)
1052 panic("UMA: Increase vm.boot_pages");
1054 * Now that we've booted reset these users to their real allocator.
1056 #ifdef UMA_MD_SMALL_ALLOC
1057 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1059 keg->uk_allocf = page_alloc;
1061 return keg->uk_allocf(zone, bytes, pflag, wait);
1065 * Allocates a number of pages from the system
1068 * bytes The number of bytes requested
1069 * wait Shall we wait?
1072 * A pointer to the alloced memory or possibly
1073 * NULL if M_NOWAIT is set.
1076 page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1078 void *p; /* Returned page */
1080 *pflag = UMA_SLAB_KMEM;
1081 p = (void *) kmem_malloc(kmem_arena, bytes, wait);
1087 * Allocates a number of pages from within an object
1090 * bytes The number of bytes requested
1091 * wait Shall we wait?
1094 * A pointer to the alloced memory or possibly
1095 * NULL if M_NOWAIT is set.
1098 noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait)
1100 TAILQ_HEAD(, vm_page) alloctail;
1102 vm_offset_t retkva, zkva;
1103 vm_page_t p, p_next;
1106 TAILQ_INIT(&alloctail);
1107 keg = zone_first_keg(zone);
1109 npages = howmany(bytes, PAGE_SIZE);
1110 while (npages > 0) {
1111 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1112 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1115 * Since the page does not belong to an object, its
1118 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1122 if (wait & M_WAITOK) {
1128 * Page allocation failed, free intermediate pages and
1131 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1132 vm_page_unwire(p, PQ_NONE);
1137 *flags = UMA_SLAB_PRIV;
1138 zkva = keg->uk_kva +
1139 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1141 TAILQ_FOREACH(p, &alloctail, listq) {
1142 pmap_qenter(zkva, &p, 1);
1146 return ((void *)retkva);
1150 * Frees a number of pages to the system
1153 * mem A pointer to the memory to be freed
1154 * size The size of the memory being freed
1155 * flags The original p->us_flags field
1161 page_free(void *mem, vm_size_t size, uint8_t flags)
1165 if (flags & UMA_SLAB_KMEM)
1167 else if (flags & UMA_SLAB_KERNEL)
1168 vmem = kernel_arena;
1170 panic("UMA: page_free used with invalid flags %x", flags);
1172 kmem_free(vmem, (vm_offset_t)mem, size);
1176 * Zero fill initializer
1178 * Arguments/Returns follow uma_init specifications
1181 zero_init(void *mem, int size, int flags)
1188 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1191 * keg The zone we should initialize
1197 keg_small_init(uma_keg_t keg)
1205 if (keg->uk_flags & UMA_ZONE_PCPU) {
1206 u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU;
1208 slabsize = sizeof(struct pcpu);
1209 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1212 slabsize = UMA_SLAB_SIZE;
1217 * Calculate the size of each allocation (rsize) according to
1218 * alignment. If the requested size is smaller than we have
1219 * allocation bits for we round it up.
1221 rsize = keg->uk_size;
1222 if (rsize < slabsize / SLAB_SETSIZE)
1223 rsize = slabsize / SLAB_SETSIZE;
1224 if (rsize & keg->uk_align)
1225 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1226 keg->uk_rsize = rsize;
1228 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1229 keg->uk_rsize < sizeof(struct pcpu),
1230 ("%s: size %u too large", __func__, keg->uk_rsize));
1232 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1235 shsize = sizeof(struct uma_slab);
1237 keg->uk_ipers = (slabsize - shsize) / rsize;
1238 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1239 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1241 memused = keg->uk_ipers * rsize + shsize;
1242 wastedspace = slabsize - memused;
1245 * We can't do OFFPAGE if we're internal or if we've been
1246 * asked to not go to the VM for buckets. If we do this we
1247 * may end up going to the VM for slabs which we do not
1248 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1249 * of UMA_ZONE_VM, which clearly forbids it.
1251 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1252 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1256 * See if using an OFFPAGE slab will limit our waste. Only do
1257 * this if it permits more items per-slab.
1259 * XXX We could try growing slabsize to limit max waste as well.
1260 * Historically this was not done because the VM could not
1261 * efficiently handle contiguous allocations.
1263 if ((wastedspace >= slabsize / UMA_MAX_WASTE) &&
1264 (keg->uk_ipers < (slabsize / keg->uk_rsize))) {
1265 keg->uk_ipers = slabsize / keg->uk_rsize;
1266 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1267 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1268 CTR6(KTR_UMA, "UMA decided we need offpage slab headers for "
1269 "keg: %s(%p), calculated wastedspace = %d, "
1270 "maximum wasted space allowed = %d, "
1271 "calculated ipers = %d, "
1272 "new wasted space = %d\n", keg->uk_name, keg, wastedspace,
1273 slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1274 slabsize - keg->uk_ipers * keg->uk_rsize);
1275 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1278 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1279 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1280 keg->uk_flags |= UMA_ZONE_HASH;
1284 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1285 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1289 * keg The keg we should initialize
1295 keg_large_init(uma_keg_t keg)
1299 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1300 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1301 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1302 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1303 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1305 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1307 keg->uk_rsize = keg->uk_size;
1309 /* We can't do OFFPAGE if we're internal, bail out here. */
1310 if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1313 /* Check whether we have enough space to not do OFFPAGE. */
1314 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1315 shsize = sizeof(struct uma_slab);
1316 if (shsize & UMA_ALIGN_PTR)
1317 shsize = (shsize & ~UMA_ALIGN_PTR) +
1318 (UMA_ALIGN_PTR + 1);
1320 if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize)
1321 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1324 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1325 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1326 keg->uk_flags |= UMA_ZONE_HASH;
1330 keg_cachespread_init(uma_keg_t keg)
1337 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1338 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1340 alignsize = keg->uk_align + 1;
1341 rsize = keg->uk_size;
1343 * We want one item to start on every align boundary in a page. To
1344 * do this we will span pages. We will also extend the item by the
1345 * size of align if it is an even multiple of align. Otherwise, it
1346 * would fall on the same boundary every time.
1348 if (rsize & keg->uk_align)
1349 rsize = (rsize & ~keg->uk_align) + alignsize;
1350 if ((rsize & alignsize) == 0)
1352 trailer = rsize - keg->uk_size;
1353 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1354 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1355 keg->uk_rsize = rsize;
1356 keg->uk_ppera = pages;
1357 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1358 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1359 KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1360 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1365 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1366 * the keg onto the global keg list.
1368 * Arguments/Returns follow uma_ctor specifications
1369 * udata Actually uma_kctor_args
1372 keg_ctor(void *mem, int size, void *udata, int flags)
1374 struct uma_kctor_args *arg = udata;
1375 uma_keg_t keg = mem;
1379 keg->uk_size = arg->size;
1380 keg->uk_init = arg->uminit;
1381 keg->uk_fini = arg->fini;
1382 keg->uk_align = arg->align;
1384 keg->uk_reserve = 0;
1386 keg->uk_flags = arg->flags;
1387 keg->uk_slabzone = NULL;
1390 * The master zone is passed to us at keg-creation time.
1393 keg->uk_name = zone->uz_name;
1395 if (arg->flags & UMA_ZONE_VM)
1396 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1398 if (arg->flags & UMA_ZONE_ZINIT)
1399 keg->uk_init = zero_init;
1401 if (arg->flags & UMA_ZONE_MALLOC)
1402 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1404 if (arg->flags & UMA_ZONE_PCPU)
1406 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1408 keg->uk_flags &= ~UMA_ZONE_PCPU;
1411 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1412 keg_cachespread_init(keg);
1414 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1415 keg_large_init(keg);
1417 keg_small_init(keg);
1420 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1421 keg->uk_slabzone = slabzone;
1424 * If we haven't booted yet we need allocations to go through the
1425 * startup cache until the vm is ready.
1427 if (booted < UMA_STARTUP2)
1428 keg->uk_allocf = startup_alloc;
1429 #ifdef UMA_MD_SMALL_ALLOC
1430 else if (keg->uk_ppera == 1)
1431 keg->uk_allocf = uma_small_alloc;
1434 keg->uk_allocf = page_alloc;
1435 #ifdef UMA_MD_SMALL_ALLOC
1436 if (keg->uk_ppera == 1)
1437 keg->uk_freef = uma_small_free;
1440 keg->uk_freef = page_free;
1443 * Initialize keg's lock
1445 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1448 * If we're putting the slab header in the actual page we need to
1449 * figure out where in each page it goes. This calculates a right
1450 * justified offset into the memory on an ALIGN_PTR boundary.
1452 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1455 /* Size of the slab struct and free list */
1456 totsize = sizeof(struct uma_slab);
1458 if (totsize & UMA_ALIGN_PTR)
1459 totsize = (totsize & ~UMA_ALIGN_PTR) +
1460 (UMA_ALIGN_PTR + 1);
1461 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1464 * The only way the following is possible is if with our
1465 * UMA_ALIGN_PTR adjustments we are now bigger than
1466 * UMA_SLAB_SIZE. I haven't checked whether this is
1467 * mathematically possible for all cases, so we make
1470 totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1471 if (totsize > PAGE_SIZE * keg->uk_ppera) {
1472 printf("zone %s ipers %d rsize %d size %d\n",
1473 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1475 panic("UMA slab won't fit.");
1479 if (keg->uk_flags & UMA_ZONE_HASH)
1480 hash_alloc(&keg->uk_hash);
1482 CTR5(KTR_UMA, "keg_ctor %p zone %s(%p) out %d free %d\n",
1483 keg, zone->uz_name, zone,
1484 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free,
1487 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1489 rw_wlock(&uma_rwlock);
1490 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1491 rw_wunlock(&uma_rwlock);
1496 * Zone header ctor. This initializes all fields, locks, etc.
1498 * Arguments/Returns follow uma_ctor specifications
1499 * udata Actually uma_zctor_args
1502 zone_ctor(void *mem, int size, void *udata, int flags)
1504 struct uma_zctor_args *arg = udata;
1505 uma_zone_t zone = mem;
1510 zone->uz_name = arg->name;
1511 zone->uz_ctor = arg->ctor;
1512 zone->uz_dtor = arg->dtor;
1513 zone->uz_slab = zone_fetch_slab;
1514 zone->uz_init = NULL;
1515 zone->uz_fini = NULL;
1516 zone->uz_allocs = 0;
1519 zone->uz_sleeps = 0;
1521 zone->uz_count_min = 0;
1523 zone->uz_warning = NULL;
1524 timevalclear(&zone->uz_ratecheck);
1527 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1530 * This is a pure cache zone, no kegs.
1533 if (arg->flags & UMA_ZONE_VM)
1534 arg->flags |= UMA_ZFLAG_CACHEONLY;
1535 zone->uz_flags = arg->flags;
1536 zone->uz_size = arg->size;
1537 zone->uz_import = arg->import;
1538 zone->uz_release = arg->release;
1539 zone->uz_arg = arg->arg;
1540 zone->uz_lockptr = &zone->uz_lock;
1541 rw_wlock(&uma_rwlock);
1542 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1543 rw_wunlock(&uma_rwlock);
1548 * Use the regular zone/keg/slab allocator.
1550 zone->uz_import = (uma_import)zone_import;
1551 zone->uz_release = (uma_release)zone_release;
1552 zone->uz_arg = zone;
1554 if (arg->flags & UMA_ZONE_SECONDARY) {
1555 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1556 zone->uz_init = arg->uminit;
1557 zone->uz_fini = arg->fini;
1558 zone->uz_lockptr = &keg->uk_lock;
1559 zone->uz_flags |= UMA_ZONE_SECONDARY;
1560 rw_wlock(&uma_rwlock);
1562 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1563 if (LIST_NEXT(z, uz_link) == NULL) {
1564 LIST_INSERT_AFTER(z, zone, uz_link);
1569 rw_wunlock(&uma_rwlock);
1570 } else if (keg == NULL) {
1571 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1572 arg->align, arg->flags)) == NULL)
1575 struct uma_kctor_args karg;
1578 /* We should only be here from uma_startup() */
1579 karg.size = arg->size;
1580 karg.uminit = arg->uminit;
1581 karg.fini = arg->fini;
1582 karg.align = arg->align;
1583 karg.flags = arg->flags;
1585 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1592 * Link in the first keg.
1594 zone->uz_klink.kl_keg = keg;
1595 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1596 zone->uz_lockptr = &keg->uk_lock;
1597 zone->uz_size = keg->uk_size;
1598 zone->uz_flags |= (keg->uk_flags &
1599 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1602 * Some internal zones don't have room allocated for the per cpu
1603 * caches. If we're internal, bail out here.
1605 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1606 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1607 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1612 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1613 zone->uz_count = bucket_select(zone->uz_size);
1615 zone->uz_count = BUCKET_MAX;
1616 zone->uz_count_min = zone->uz_count;
1622 * Keg header dtor. This frees all data, destroys locks, frees the hash
1623 * table and removes the keg from the global list.
1625 * Arguments/Returns follow uma_dtor specifications
1629 keg_dtor(void *arg, int size, void *udata)
1633 keg = (uma_keg_t)arg;
1635 if (keg->uk_free != 0) {
1636 printf("Freed UMA keg (%s) was not empty (%d items). "
1637 " Lost %d pages of memory.\n",
1638 keg->uk_name ? keg->uk_name : "",
1639 keg->uk_free, keg->uk_pages);
1643 hash_free(&keg->uk_hash);
1651 * Arguments/Returns follow uma_dtor specifications
1655 zone_dtor(void *arg, int size, void *udata)
1661 zone = (uma_zone_t)arg;
1662 keg = zone_first_keg(zone);
1664 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1667 rw_wlock(&uma_rwlock);
1668 LIST_REMOVE(zone, uz_link);
1669 rw_wunlock(&uma_rwlock);
1671 * XXX there are some races here where
1672 * the zone can be drained but zone lock
1673 * released and then refilled before we
1674 * remove it... we dont care for now
1676 zone_drain_wait(zone, M_WAITOK);
1678 * Unlink all of our kegs.
1680 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1681 klink->kl_keg = NULL;
1682 LIST_REMOVE(klink, kl_link);
1683 if (klink == &zone->uz_klink)
1685 free(klink, M_TEMP);
1688 * We only destroy kegs from non secondary zones.
1690 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1691 rw_wlock(&uma_rwlock);
1692 LIST_REMOVE(keg, uk_link);
1693 rw_wunlock(&uma_rwlock);
1694 zone_free_item(kegs, keg, NULL, SKIP_NONE);
1696 ZONE_LOCK_FINI(zone);
1700 * Traverses every zone in the system and calls a callback
1703 * zfunc A pointer to a function which accepts a zone
1710 zone_foreach(void (*zfunc)(uma_zone_t))
1715 rw_rlock(&uma_rwlock);
1716 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1717 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1720 rw_runlock(&uma_rwlock);
1723 /* Public functions */
1726 uma_startup(void *mem, int npages)
1728 struct uma_zctor_args args;
1730 rw_init(&uma_rwlock, "UMA lock");
1732 /* "manually" create the initial zone */
1733 memset(&args, 0, sizeof(args));
1734 args.name = "UMA Kegs";
1735 args.size = sizeof(struct uma_keg);
1736 args.ctor = keg_ctor;
1737 args.dtor = keg_dtor;
1738 args.uminit = zero_init;
1740 args.keg = &masterkeg;
1741 args.align = 32 - 1;
1742 args.flags = UMA_ZFLAG_INTERNAL;
1743 /* The initial zone has no Per cpu queues so it's smaller */
1744 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1746 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1748 boot_pages = npages;
1750 args.name = "UMA Zones";
1751 args.size = sizeof(struct uma_zone) +
1752 (sizeof(struct uma_cache) * (mp_maxid + 1));
1753 args.ctor = zone_ctor;
1754 args.dtor = zone_dtor;
1755 args.uminit = zero_init;
1758 args.align = 32 - 1;
1759 args.flags = UMA_ZFLAG_INTERNAL;
1760 /* The initial zone has no Per cpu queues so it's smaller */
1761 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1763 /* Now make a zone for slab headers */
1764 slabzone = uma_zcreate("UMA Slabs",
1765 sizeof(struct uma_slab),
1766 NULL, NULL, NULL, NULL,
1767 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1769 hashzone = uma_zcreate("UMA Hash",
1770 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1771 NULL, NULL, NULL, NULL,
1772 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1776 booted = UMA_STARTUP;
1783 booted = UMA_STARTUP2;
1785 sx_init(&uma_drain_lock, "umadrain");
1789 * Initialize our callout handle
1797 callout_init(&uma_callout, 1);
1798 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1802 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1803 int align, uint32_t flags)
1805 struct uma_kctor_args args;
1808 args.uminit = uminit;
1810 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1813 return (zone_alloc_item(kegs, &args, M_WAITOK));
1818 uma_set_align(int align)
1821 if (align != UMA_ALIGN_CACHE)
1822 uma_align_cache = align;
1827 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1828 uma_init uminit, uma_fini fini, int align, uint32_t flags)
1831 struct uma_zctor_args args;
1835 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
1838 /* This stuff is essential for the zone ctor */
1839 memset(&args, 0, sizeof(args));
1844 args.uminit = uminit;
1848 * If a zone is being created with an empty constructor and
1849 * destructor, pass UMA constructor/destructor which checks for
1850 * memory use after free.
1852 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) &&
1853 ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) {
1854 args.ctor = trash_ctor;
1855 args.dtor = trash_dtor;
1856 args.uminit = trash_init;
1857 args.fini = trash_fini;
1864 if (booted < UMA_STARTUP2) {
1867 sx_slock(&uma_drain_lock);
1870 res = zone_alloc_item(zones, &args, M_WAITOK);
1872 sx_sunlock(&uma_drain_lock);
1878 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1879 uma_init zinit, uma_fini zfini, uma_zone_t master)
1881 struct uma_zctor_args args;
1886 keg = zone_first_keg(master);
1887 memset(&args, 0, sizeof(args));
1889 args.size = keg->uk_size;
1892 args.uminit = zinit;
1894 args.align = keg->uk_align;
1895 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1898 if (booted < UMA_STARTUP2) {
1901 sx_slock(&uma_drain_lock);
1904 /* XXX Attaches only one keg of potentially many. */
1905 res = zone_alloc_item(zones, &args, M_WAITOK);
1907 sx_sunlock(&uma_drain_lock);
1913 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1914 uma_init zinit, uma_fini zfini, uma_import zimport,
1915 uma_release zrelease, void *arg, int flags)
1917 struct uma_zctor_args args;
1919 memset(&args, 0, sizeof(args));
1924 args.uminit = zinit;
1926 args.import = zimport;
1927 args.release = zrelease;
1932 return (zone_alloc_item(zones, &args, M_WAITOK));
1936 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1940 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
1943 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
1948 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1956 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1963 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1965 zone_lock_pair(zone, master);
1967 * zone must use vtoslab() to resolve objects and must already be
1970 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1971 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1976 * The new master must also use vtoslab().
1978 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
1984 * The underlying object must be the same size. rsize
1987 if (master->uz_size != zone->uz_size) {
1992 * Put it at the end of the list.
1994 klink->kl_keg = zone_first_keg(master);
1995 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
1996 if (LIST_NEXT(kl, kl_link) == NULL) {
1997 LIST_INSERT_AFTER(kl, klink, kl_link);
2002 zone->uz_flags |= UMA_ZFLAG_MULTI;
2003 zone->uz_slab = zone_fetch_slab_multi;
2006 zone_unlock_pair(zone, master);
2008 free(klink, M_TEMP);
2016 uma_zdestroy(uma_zone_t zone)
2019 sx_slock(&uma_drain_lock);
2020 zone_free_item(zones, zone, NULL, SKIP_NONE);
2021 sx_sunlock(&uma_drain_lock);
2026 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2030 uma_bucket_t bucket;
2034 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2035 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2037 /* This is the fast path allocation */
2038 CTR4(KTR_UMA, "uma_zalloc_arg thread %x zone %s(%p) flags %d",
2039 curthread, zone->uz_name, zone, flags);
2041 if (flags & M_WAITOK) {
2042 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2043 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2045 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2046 ("uma_zalloc_arg: called with spinlock or critical section held"));
2048 #ifdef DEBUG_MEMGUARD
2049 if (memguard_cmp_zone(zone)) {
2050 item = memguard_alloc(zone->uz_size, flags);
2052 if (zone->uz_init != NULL &&
2053 zone->uz_init(item, zone->uz_size, flags) != 0)
2055 if (zone->uz_ctor != NULL &&
2056 zone->uz_ctor(item, zone->uz_size, udata,
2058 zone->uz_fini(item, zone->uz_size);
2063 /* This is unfortunate but should not be fatal. */
2067 * If possible, allocate from the per-CPU cache. There are two
2068 * requirements for safe access to the per-CPU cache: (1) the thread
2069 * accessing the cache must not be preempted or yield during access,
2070 * and (2) the thread must not migrate CPUs without switching which
2071 * cache it accesses. We rely on a critical section to prevent
2072 * preemption and migration. We release the critical section in
2073 * order to acquire the zone mutex if we are unable to allocate from
2074 * the current cache; when we re-acquire the critical section, we
2075 * must detect and handle migration if it has occurred.
2079 cache = &zone->uz_cpu[cpu];
2082 bucket = cache->uc_allocbucket;
2083 if (bucket != NULL && bucket->ub_cnt > 0) {
2085 item = bucket->ub_bucket[bucket->ub_cnt];
2087 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2089 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2092 if (zone->uz_ctor != NULL &&
2093 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2094 atomic_add_long(&zone->uz_fails, 1);
2095 zone_free_item(zone, item, udata, SKIP_DTOR);
2099 uma_dbg_alloc(zone, NULL, item);
2102 uma_zero_item(item, zone);
2107 * We have run out of items in our alloc bucket.
2108 * See if we can switch with our free bucket.
2110 bucket = cache->uc_freebucket;
2111 if (bucket != NULL && bucket->ub_cnt > 0) {
2113 "uma_zalloc: zone %s(%p) swapping empty with alloc",
2114 zone->uz_name, zone);
2115 cache->uc_freebucket = cache->uc_allocbucket;
2116 cache->uc_allocbucket = bucket;
2121 * Discard any empty allocation bucket while we hold no locks.
2123 bucket = cache->uc_allocbucket;
2124 cache->uc_allocbucket = NULL;
2127 bucket_free(zone, bucket, udata);
2129 /* Short-circuit for zones without buckets and low memory. */
2130 if (zone->uz_count == 0 || bucketdisable)
2134 * Attempt to retrieve the item from the per-CPU cache has failed, so
2135 * we must go back to the zone. This requires the zone lock, so we
2136 * must drop the critical section, then re-acquire it when we go back
2137 * to the cache. Since the critical section is released, we may be
2138 * preempted or migrate. As such, make sure not to maintain any
2139 * thread-local state specific to the cache from prior to releasing
2140 * the critical section.
2143 if (ZONE_TRYLOCK(zone) == 0) {
2144 /* Record contention to size the buckets. */
2150 cache = &zone->uz_cpu[cpu];
2153 * Since we have locked the zone we may as well send back our stats.
2155 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2156 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2157 cache->uc_allocs = 0;
2158 cache->uc_frees = 0;
2160 /* See if we lost the race to fill the cache. */
2161 if (cache->uc_allocbucket != NULL) {
2167 * Check the zone's cache of buckets.
2169 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2170 KASSERT(bucket->ub_cnt != 0,
2171 ("uma_zalloc_arg: Returning an empty bucket."));
2173 LIST_REMOVE(bucket, ub_link);
2174 cache->uc_allocbucket = bucket;
2178 /* We are no longer associated with this CPU. */
2182 * We bump the uz count when the cache size is insufficient to
2183 * handle the working set.
2185 if (lockfail && zone->uz_count < BUCKET_MAX)
2190 * Now lets just fill a bucket and put it on the free list. If that
2191 * works we'll restart the allocation from the beginning and it
2192 * will use the just filled bucket.
2194 bucket = zone_alloc_bucket(zone, udata, flags);
2195 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
2196 zone->uz_name, zone, bucket);
2197 if (bucket != NULL) {
2201 cache = &zone->uz_cpu[cpu];
2203 * See if we lost the race or were migrated. Cache the
2204 * initialized bucket to make this less likely or claim
2205 * the memory directly.
2207 if (cache->uc_allocbucket == NULL)
2208 cache->uc_allocbucket = bucket;
2210 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2216 * We may not be able to get a bucket so return an actual item.
2219 item = zone_alloc_item(zone, udata, flags);
2225 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2230 mtx_assert(&keg->uk_lock, MA_OWNED);
2233 if ((flags & M_USE_RESERVE) == 0)
2234 reserve = keg->uk_reserve;
2238 * Find a slab with some space. Prefer slabs that are partially
2239 * used over those that are totally full. This helps to reduce
2242 if (keg->uk_free > reserve) {
2243 if (!LIST_EMPTY(&keg->uk_part_slab)) {
2244 slab = LIST_FIRST(&keg->uk_part_slab);
2246 slab = LIST_FIRST(&keg->uk_free_slab);
2247 LIST_REMOVE(slab, us_link);
2248 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2251 MPASS(slab->us_keg == keg);
2256 * M_NOVM means don't ask at all!
2261 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2262 keg->uk_flags |= UMA_ZFLAG_FULL;
2264 * If this is not a multi-zone, set the FULL bit.
2265 * Otherwise slab_multi() takes care of it.
2267 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2268 zone->uz_flags |= UMA_ZFLAG_FULL;
2269 zone_log_warning(zone);
2270 zone_maxaction(zone);
2272 if (flags & M_NOWAIT)
2275 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2278 slab = keg_alloc_slab(keg, zone, flags);
2280 * If we got a slab here it's safe to mark it partially used
2281 * and return. We assume that the caller is going to remove
2282 * at least one item.
2285 MPASS(slab->us_keg == keg);
2286 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2290 * We might not have been able to get a slab but another cpu
2291 * could have while we were unlocked. Check again before we
2300 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2305 keg = zone_first_keg(zone);
2310 slab = keg_fetch_slab(keg, zone, flags);
2313 if (flags & (M_NOWAIT | M_NOVM))
2321 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2322 * with the keg locked. On NULL no lock is held.
2324 * The last pointer is used to seed the search. It is not required.
2327 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2337 * Don't wait on the first pass. This will skip limit tests
2338 * as well. We don't want to block if we can find a provider
2341 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2343 * Use the last slab allocated as a hint for where to start
2347 slab = keg_fetch_slab(last, zone, flags);
2353 * Loop until we have a slab incase of transient failures
2354 * while M_WAITOK is specified. I'm not sure this is 100%
2355 * required but we've done it for so long now.
2361 * Search the available kegs for slabs. Be careful to hold the
2362 * correct lock while calling into the keg layer.
2364 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2365 keg = klink->kl_keg;
2367 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2368 slab = keg_fetch_slab(keg, zone, flags);
2372 if (keg->uk_flags & UMA_ZFLAG_FULL)
2378 if (rflags & (M_NOWAIT | M_NOVM))
2382 * All kegs are full. XXX We can't atomically check all kegs
2383 * and sleep so just sleep for a short period and retry.
2385 if (full && !empty) {
2387 zone->uz_flags |= UMA_ZFLAG_FULL;
2389 zone_log_warning(zone);
2390 zone_maxaction(zone);
2391 msleep(zone, zone->uz_lockptr, PVM,
2392 "zonelimit", hz/100);
2393 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2402 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2407 MPASS(keg == slab->us_keg);
2408 mtx_assert(&keg->uk_lock, MA_OWNED);
2410 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2411 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2412 item = slab->us_data + (keg->uk_rsize * freei);
2413 slab->us_freecount--;
2416 /* Move this slab to the full list */
2417 if (slab->us_freecount == 0) {
2418 LIST_REMOVE(slab, us_link);
2419 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2426 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2434 /* Try to keep the buckets totally full */
2435 for (i = 0; i < max; ) {
2436 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2439 while (slab->us_freecount && i < max) {
2440 bucket[i++] = slab_alloc_item(keg, slab);
2441 if (keg->uk_free <= keg->uk_reserve)
2444 /* Don't grab more than one slab at a time. */
2455 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2457 uma_bucket_t bucket;
2460 /* Don't wait for buckets, preserve caller's NOVM setting. */
2461 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2465 max = MIN(bucket->ub_entries, zone->uz_count);
2466 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2470 * Initialize the memory if necessary.
2472 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2475 for (i = 0; i < bucket->ub_cnt; i++)
2476 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2480 * If we couldn't initialize the whole bucket, put the
2481 * rest back onto the freelist.
2483 if (i != bucket->ub_cnt) {
2484 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2485 bucket->ub_cnt - i);
2487 bzero(&bucket->ub_bucket[i],
2488 sizeof(void *) * (bucket->ub_cnt - i));
2494 if (bucket->ub_cnt == 0) {
2495 bucket_free(zone, bucket, udata);
2496 atomic_add_long(&zone->uz_fails, 1);
2504 * Allocates a single item from a zone.
2507 * zone The zone to alloc for.
2508 * udata The data to be passed to the constructor.
2509 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2512 * NULL if there is no memory and M_NOWAIT is set
2513 * An item if successful
2517 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2523 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2525 atomic_add_long(&zone->uz_allocs, 1);
2528 * We have to call both the zone's init (not the keg's init)
2529 * and the zone's ctor. This is because the item is going from
2530 * a keg slab directly to the user, and the user is expecting it
2531 * to be both zone-init'd as well as zone-ctor'd.
2533 if (zone->uz_init != NULL) {
2534 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2535 zone_free_item(zone, item, udata, SKIP_FINI);
2539 if (zone->uz_ctor != NULL) {
2540 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2541 zone_free_item(zone, item, udata, SKIP_DTOR);
2546 uma_dbg_alloc(zone, NULL, item);
2549 uma_zero_item(item, zone);
2551 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
2552 zone->uz_name, zone);
2557 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
2558 zone->uz_name, zone);
2559 atomic_add_long(&zone->uz_fails, 1);
2565 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2568 uma_bucket_t bucket;
2572 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2573 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2575 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2578 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2579 ("uma_zfree_arg: called with spinlock or critical section held"));
2581 /* uma_zfree(..., NULL) does nothing, to match free(9). */
2584 #ifdef DEBUG_MEMGUARD
2585 if (is_memguard_addr(item)) {
2586 if (zone->uz_dtor != NULL)
2587 zone->uz_dtor(item, zone->uz_size, udata);
2588 if (zone->uz_fini != NULL)
2589 zone->uz_fini(item, zone->uz_size);
2590 memguard_free(item);
2595 if (zone->uz_flags & UMA_ZONE_MALLOC)
2596 uma_dbg_free(zone, udata, item);
2598 uma_dbg_free(zone, NULL, item);
2600 if (zone->uz_dtor != NULL)
2601 zone->uz_dtor(item, zone->uz_size, udata);
2604 * The race here is acceptable. If we miss it we'll just have to wait
2605 * a little longer for the limits to be reset.
2607 if (zone->uz_flags & UMA_ZFLAG_FULL)
2611 * If possible, free to the per-CPU cache. There are two
2612 * requirements for safe access to the per-CPU cache: (1) the thread
2613 * accessing the cache must not be preempted or yield during access,
2614 * and (2) the thread must not migrate CPUs without switching which
2615 * cache it accesses. We rely on a critical section to prevent
2616 * preemption and migration. We release the critical section in
2617 * order to acquire the zone mutex if we are unable to free to the
2618 * current cache; when we re-acquire the critical section, we must
2619 * detect and handle migration if it has occurred.
2624 cache = &zone->uz_cpu[cpu];
2628 * Try to free into the allocbucket first to give LIFO ordering
2629 * for cache-hot datastructures. Spill over into the freebucket
2630 * if necessary. Alloc will swap them if one runs dry.
2632 bucket = cache->uc_allocbucket;
2633 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2634 bucket = cache->uc_freebucket;
2635 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2636 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2637 ("uma_zfree: Freeing to non free bucket index."));
2638 bucket->ub_bucket[bucket->ub_cnt] = item;
2646 * We must go back the zone, which requires acquiring the zone lock,
2647 * which in turn means we must release and re-acquire the critical
2648 * section. Since the critical section is released, we may be
2649 * preempted or migrate. As such, make sure not to maintain any
2650 * thread-local state specific to the cache from prior to releasing
2651 * the critical section.
2654 if (zone->uz_count == 0 || bucketdisable)
2658 if (ZONE_TRYLOCK(zone) == 0) {
2659 /* Record contention to size the buckets. */
2665 cache = &zone->uz_cpu[cpu];
2668 * Since we have locked the zone we may as well send back our stats.
2670 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2671 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2672 cache->uc_allocs = 0;
2673 cache->uc_frees = 0;
2675 bucket = cache->uc_freebucket;
2676 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2680 cache->uc_freebucket = NULL;
2681 /* We are no longer associated with this CPU. */
2684 /* Can we throw this on the zone full list? */
2685 if (bucket != NULL) {
2687 "uma_zfree: zone %s(%p) putting bucket %p on free list",
2688 zone->uz_name, zone, bucket);
2689 /* ub_cnt is pointing to the last free item */
2690 KASSERT(bucket->ub_cnt != 0,
2691 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2692 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2696 * We bump the uz count when the cache size is insufficient to
2697 * handle the working set.
2699 if (lockfail && zone->uz_count < BUCKET_MAX)
2703 bucket = bucket_alloc(zone, udata, M_NOWAIT);
2704 CTR3(KTR_UMA, "uma_zfree: zone %s(%p) allocated bucket %p",
2705 zone->uz_name, zone, bucket);
2709 cache = &zone->uz_cpu[cpu];
2710 if (cache->uc_freebucket == NULL) {
2711 cache->uc_freebucket = bucket;
2715 * We lost the race, start over. We have to drop our
2716 * critical section to free the bucket.
2719 bucket_free(zone, bucket, udata);
2724 * If nothing else caught this, we'll just do an internal free.
2727 zone_free_item(zone, item, udata, SKIP_DTOR);
2733 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2737 mtx_assert(&keg->uk_lock, MA_OWNED);
2738 MPASS(keg == slab->us_keg);
2740 /* Do we need to remove from any lists? */
2741 if (slab->us_freecount+1 == keg->uk_ipers) {
2742 LIST_REMOVE(slab, us_link);
2743 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2744 } else if (slab->us_freecount == 0) {
2745 LIST_REMOVE(slab, us_link);
2746 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2749 /* Slab management. */
2750 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2751 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2752 slab->us_freecount++;
2754 /* Keg statistics. */
2759 zone_release(uma_zone_t zone, void **bucket, int cnt)
2769 keg = zone_first_keg(zone);
2771 for (i = 0; i < cnt; i++) {
2773 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2774 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2775 if (zone->uz_flags & UMA_ZONE_HASH) {
2776 slab = hash_sfind(&keg->uk_hash, mem);
2778 mem += keg->uk_pgoff;
2779 slab = (uma_slab_t)mem;
2782 slab = vtoslab((vm_offset_t)item);
2783 if (slab->us_keg != keg) {
2789 slab_free_item(keg, slab, item);
2790 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2791 if (keg->uk_pages < keg->uk_maxpages) {
2792 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2797 * We can handle one more allocation. Since we're
2798 * clearing ZFLAG_FULL, wake up all procs blocked
2799 * on pages. This should be uncommon, so keeping this
2800 * simple for now (rather than adding count of blocked
2809 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2817 * Frees a single item to any zone.
2820 * zone The zone to free to
2821 * item The item we're freeing
2822 * udata User supplied data for the dtor
2823 * skip Skip dtors and finis
2826 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2830 if (skip == SKIP_NONE) {
2831 if (zone->uz_flags & UMA_ZONE_MALLOC)
2832 uma_dbg_free(zone, udata, item);
2834 uma_dbg_free(zone, NULL, item);
2837 if (skip < SKIP_DTOR && zone->uz_dtor)
2838 zone->uz_dtor(item, zone->uz_size, udata);
2840 if (skip < SKIP_FINI && zone->uz_fini)
2841 zone->uz_fini(item, zone->uz_size);
2843 atomic_add_long(&zone->uz_frees, 1);
2844 zone->uz_release(zone->uz_arg, &item, 1);
2849 uma_zone_set_max(uma_zone_t zone, int nitems)
2853 keg = zone_first_keg(zone);
2857 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2858 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2859 keg->uk_maxpages += keg->uk_ppera;
2860 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers;
2868 uma_zone_get_max(uma_zone_t zone)
2873 keg = zone_first_keg(zone);
2877 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers;
2885 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2889 zone->uz_warning = warning;
2895 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
2899 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
2905 uma_zone_get_cur(uma_zone_t zone)
2911 nitems = zone->uz_allocs - zone->uz_frees;
2914 * See the comment in sysctl_vm_zone_stats() regarding the
2915 * safety of accessing the per-cpu caches. With the zone lock
2916 * held, it is safe, but can potentially result in stale data.
2918 nitems += zone->uz_cpu[i].uc_allocs -
2919 zone->uz_cpu[i].uc_frees;
2923 return (nitems < 0 ? 0 : nitems);
2928 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2932 keg = zone_first_keg(zone);
2933 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2935 KASSERT(keg->uk_pages == 0,
2936 ("uma_zone_set_init on non-empty keg"));
2937 keg->uk_init = uminit;
2943 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2947 keg = zone_first_keg(zone);
2948 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
2950 KASSERT(keg->uk_pages == 0,
2951 ("uma_zone_set_fini on non-empty keg"));
2952 keg->uk_fini = fini;
2958 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2962 KASSERT(zone_first_keg(zone)->uk_pages == 0,
2963 ("uma_zone_set_zinit on non-empty keg"));
2964 zone->uz_init = zinit;
2970 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2974 KASSERT(zone_first_keg(zone)->uk_pages == 0,
2975 ("uma_zone_set_zfini on non-empty keg"));
2976 zone->uz_fini = zfini;
2981 /* XXX uk_freef is not actually used with the zone locked */
2983 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2987 keg = zone_first_keg(zone);
2988 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
2990 keg->uk_freef = freef;
2995 /* XXX uk_allocf is not actually used with the zone locked */
2997 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3001 keg = zone_first_keg(zone);
3003 keg->uk_allocf = allocf;
3009 uma_zone_reserve(uma_zone_t zone, int items)
3013 keg = zone_first_keg(zone);
3017 keg->uk_reserve = items;
3025 uma_zone_reserve_kva(uma_zone_t zone, int count)
3031 keg = zone_first_keg(zone);
3034 pages = count / keg->uk_ipers;
3036 if (pages * keg->uk_ipers < count)
3038 pages *= keg->uk_ppera;
3040 #ifdef UMA_MD_SMALL_ALLOC
3041 if (keg->uk_ppera > 1) {
3045 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
3053 keg->uk_maxpages = pages;
3054 #ifdef UMA_MD_SMALL_ALLOC
3055 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3057 keg->uk_allocf = noobj_alloc;
3059 keg->uk_flags |= UMA_ZONE_NOFREE;
3067 uma_prealloc(uma_zone_t zone, int items)
3073 keg = zone_first_keg(zone);
3077 slabs = items / keg->uk_ipers;
3078 if (slabs * keg->uk_ipers < items)
3081 slab = keg_alloc_slab(keg, zone, M_WAITOK);
3084 MPASS(slab->us_keg == keg);
3085 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3093 uma_reclaim_locked(bool kmem_danger)
3096 CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
3097 sx_assert(&uma_drain_lock, SA_XLOCKED);
3099 zone_foreach(zone_drain);
3100 if (vm_page_count_min() || kmem_danger) {
3101 cache_drain_safe(NULL);
3102 zone_foreach(zone_drain);
3105 * Some slabs may have been freed but this zone will be visited early
3106 * we visit again so that we can free pages that are empty once other
3107 * zones are drained. We have to do the same for buckets.
3109 zone_drain(slabzone);
3110 bucket_zone_drain();
3117 sx_xlock(&uma_drain_lock);
3118 uma_reclaim_locked(false);
3119 sx_xunlock(&uma_drain_lock);
3122 static int uma_reclaim_needed;
3125 uma_reclaim_wakeup(void)
3128 uma_reclaim_needed = 1;
3129 wakeup(&uma_reclaim_needed);
3133 uma_reclaim_worker(void *arg __unused)
3136 sx_xlock(&uma_drain_lock);
3138 sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM,
3140 if (uma_reclaim_needed) {
3141 uma_reclaim_needed = 0;
3142 sx_xunlock(&uma_drain_lock);
3143 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
3144 sx_xlock(&uma_drain_lock);
3145 uma_reclaim_locked(true);
3152 uma_zone_exhausted(uma_zone_t zone)
3157 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3163 uma_zone_exhausted_nolock(uma_zone_t zone)
3165 return (zone->uz_flags & UMA_ZFLAG_FULL);
3169 uma_large_malloc(vm_size_t size, int wait)
3175 slab = zone_alloc_item(slabzone, NULL, wait);
3178 mem = page_alloc(NULL, size, &flags, wait);
3180 vsetslab((vm_offset_t)mem, slab);
3181 slab->us_data = mem;
3182 slab->us_flags = flags | UMA_SLAB_MALLOC;
3183 slab->us_size = size;
3185 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3192 uma_large_free(uma_slab_t slab)
3195 page_free(slab->us_data, slab->us_size, slab->us_flags);
3196 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3200 uma_zero_item(void *item, uma_zone_t zone)
3204 if (zone->uz_flags & UMA_ZONE_PCPU) {
3206 bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3208 bzero(item, zone->uz_size);
3212 uma_print_stats(void)
3214 zone_foreach(uma_print_zone);
3218 slab_print(uma_slab_t slab)
3220 printf("slab: keg %p, data %p, freecount %d\n",
3221 slab->us_keg, slab->us_data, slab->us_freecount);
3225 cache_print(uma_cache_t cache)
3227 printf("alloc: %p(%d), free: %p(%d)\n",
3228 cache->uc_allocbucket,
3229 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3230 cache->uc_freebucket,
3231 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3235 uma_print_keg(uma_keg_t keg)
3239 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3240 "out %d free %d limit %d\n",
3241 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3242 keg->uk_ipers, keg->uk_ppera,
3243 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free,
3244 keg->uk_free, (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3245 printf("Part slabs:\n");
3246 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3248 printf("Free slabs:\n");
3249 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3251 printf("Full slabs:\n");
3252 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3257 uma_print_zone(uma_zone_t zone)
3263 printf("zone: %s(%p) size %d flags %#x\n",
3264 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3265 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3266 uma_print_keg(kl->kl_keg);
3268 cache = &zone->uz_cpu[i];
3269 printf("CPU %d Cache:\n", i);
3276 * Generate statistics across both the zone and its per-cpu cache's. Return
3277 * desired statistics if the pointer is non-NULL for that statistic.
3279 * Note: does not update the zone statistics, as it can't safely clear the
3280 * per-CPU cache statistic.
3282 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3283 * safe from off-CPU; we should modify the caches to track this information
3284 * directly so that we don't have to.
3287 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3288 uint64_t *freesp, uint64_t *sleepsp)
3291 uint64_t allocs, frees, sleeps;
3294 allocs = frees = sleeps = 0;
3297 cache = &z->uz_cpu[cpu];
3298 if (cache->uc_allocbucket != NULL)
3299 cachefree += cache->uc_allocbucket->ub_cnt;
3300 if (cache->uc_freebucket != NULL)
3301 cachefree += cache->uc_freebucket->ub_cnt;
3302 allocs += cache->uc_allocs;
3303 frees += cache->uc_frees;
3305 allocs += z->uz_allocs;
3306 frees += z->uz_frees;
3307 sleeps += z->uz_sleeps;
3308 if (cachefreep != NULL)
3309 *cachefreep = cachefree;
3310 if (allocsp != NULL)
3314 if (sleepsp != NULL)
3320 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3327 rw_rlock(&uma_rwlock);
3328 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3329 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3332 rw_runlock(&uma_rwlock);
3333 return (sysctl_handle_int(oidp, &count, 0, req));
3337 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3339 struct uma_stream_header ush;
3340 struct uma_type_header uth;
3341 struct uma_percpu_stat ups;
3342 uma_bucket_t bucket;
3349 int count, error, i;
3351 error = sysctl_wire_old_buffer(req, 0);
3354 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3355 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3358 rw_rlock(&uma_rwlock);
3359 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3360 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3365 * Insert stream header.
3367 bzero(&ush, sizeof(ush));
3368 ush.ush_version = UMA_STREAM_VERSION;
3369 ush.ush_maxcpus = (mp_maxid + 1);
3370 ush.ush_count = count;
3371 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3373 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3374 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3375 bzero(&uth, sizeof(uth));
3377 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3378 uth.uth_align = kz->uk_align;
3379 uth.uth_size = kz->uk_size;
3380 uth.uth_rsize = kz->uk_rsize;
3381 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3383 uth.uth_maxpages += k->uk_maxpages;
3384 uth.uth_pages += k->uk_pages;
3385 uth.uth_keg_free += k->uk_free;
3386 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3391 * A zone is secondary is it is not the first entry
3392 * on the keg's zone list.
3394 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3395 (LIST_FIRST(&kz->uk_zones) != z))
3396 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3398 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3399 uth.uth_zone_free += bucket->ub_cnt;
3400 uth.uth_allocs = z->uz_allocs;
3401 uth.uth_frees = z->uz_frees;
3402 uth.uth_fails = z->uz_fails;
3403 uth.uth_sleeps = z->uz_sleeps;
3404 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3406 * While it is not normally safe to access the cache
3407 * bucket pointers while not on the CPU that owns the
3408 * cache, we only allow the pointers to be exchanged
3409 * without the zone lock held, not invalidated, so
3410 * accept the possible race associated with bucket
3411 * exchange during monitoring.
3413 for (i = 0; i < (mp_maxid + 1); i++) {
3414 bzero(&ups, sizeof(ups));
3415 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3419 cache = &z->uz_cpu[i];
3420 if (cache->uc_allocbucket != NULL)
3421 ups.ups_cache_free +=
3422 cache->uc_allocbucket->ub_cnt;
3423 if (cache->uc_freebucket != NULL)
3424 ups.ups_cache_free +=
3425 cache->uc_freebucket->ub_cnt;
3426 ups.ups_allocs = cache->uc_allocs;
3427 ups.ups_frees = cache->uc_frees;
3429 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3434 rw_runlock(&uma_rwlock);
3435 error = sbuf_finish(&sbuf);
3441 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3443 uma_zone_t zone = *(uma_zone_t *)arg1;
3446 max = uma_zone_get_max(zone);
3447 error = sysctl_handle_int(oidp, &max, 0, req);
3448 if (error || !req->newptr)
3451 uma_zone_set_max(zone, max);
3457 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3459 uma_zone_t zone = *(uma_zone_t *)arg1;
3462 cur = uma_zone_get_cur(zone);
3463 return (sysctl_handle_int(oidp, &cur, 0, req));
3468 uma_dbg_getslab(uma_zone_t zone, void *item)
3474 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
3475 if (zone->uz_flags & UMA_ZONE_VTOSLAB) {
3476 slab = vtoslab((vm_offset_t)mem);
3479 * It is safe to return the slab here even though the
3480 * zone is unlocked because the item's allocation state
3481 * essentially holds a reference.
3484 keg = LIST_FIRST(&zone->uz_kegs)->kl_keg;
3485 if (keg->uk_flags & UMA_ZONE_HASH)
3486 slab = hash_sfind(&keg->uk_hash, mem);
3488 slab = (uma_slab_t)(mem + keg->uk_pgoff);
3496 * Set up the slab's freei data such that uma_dbg_free can function.
3500 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
3505 if (zone_first_keg(zone) == NULL)
3508 slab = uma_dbg_getslab(zone, item);
3510 panic("uma: item %p did not belong to zone %s\n",
3511 item, zone->uz_name);
3514 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3516 if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3517 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
3518 item, zone, zone->uz_name, slab, freei);
3519 BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3525 * Verifies freed addresses. Checks for alignment, valid slab membership
3526 * and duplicate frees.
3530 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
3535 if (zone_first_keg(zone) == NULL)
3538 slab = uma_dbg_getslab(zone, item);
3540 panic("uma: Freed item %p did not belong to zone %s\n",
3541 item, zone->uz_name);
3544 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3546 if (freei >= keg->uk_ipers)
3547 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
3548 item, zone, zone->uz_name, slab, freei);
3550 if (((freei * keg->uk_rsize) + slab->us_data) != item)
3551 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
3552 item, zone, zone->uz_name, slab, freei);
3554 if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3555 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
3556 item, zone, zone->uz_name, slab, freei);
3558 BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3560 #endif /* INVARIANTS */
3563 DB_SHOW_COMMAND(uma, db_show_uma)
3565 uint64_t allocs, frees, sleeps;
3566 uma_bucket_t bucket;
3571 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3572 "Free", "Requests", "Sleeps", "Bucket");
3573 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3574 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3575 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3576 allocs = z->uz_allocs;
3577 frees = z->uz_frees;
3578 sleeps = z->uz_sleeps;
3581 uma_zone_sumstat(z, &cachefree, &allocs,
3583 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3584 (LIST_FIRST(&kz->uk_zones) != z)))
3585 cachefree += kz->uk_free;
3586 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3587 cachefree += bucket->ub_cnt;
3588 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3589 z->uz_name, (uintmax_t)kz->uk_size,
3590 (intmax_t)(allocs - frees), cachefree,
3591 (uintmax_t)allocs, sleeps, z->uz_count);
3598 DB_SHOW_COMMAND(umacache, db_show_umacache)
3600 uint64_t allocs, frees;
3601 uma_bucket_t bucket;
3605 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3606 "Requests", "Bucket");
3607 LIST_FOREACH(z, &uma_cachezones, uz_link) {
3608 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3609 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3610 cachefree += bucket->ub_cnt;
3611 db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3612 z->uz_name, (uintmax_t)z->uz_size,
3613 (intmax_t)(allocs - frees), cachefree,
3614 (uintmax_t)allocs, z->uz_count);