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 __exclusive_cache_line 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 |
1113 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1117 * Since the page does not belong to an object, its
1120 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1125 * Page allocation failed, free intermediate pages and
1128 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1129 vm_page_unwire(p, PQ_NONE);
1134 *flags = UMA_SLAB_PRIV;
1135 zkva = keg->uk_kva +
1136 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1138 TAILQ_FOREACH(p, &alloctail, listq) {
1139 pmap_qenter(zkva, &p, 1);
1143 return ((void *)retkva);
1147 * Frees a number of pages to the system
1150 * mem A pointer to the memory to be freed
1151 * size The size of the memory being freed
1152 * flags The original p->us_flags field
1158 page_free(void *mem, vm_size_t size, uint8_t flags)
1162 if (flags & UMA_SLAB_KMEM)
1164 else if (flags & UMA_SLAB_KERNEL)
1165 vmem = kernel_arena;
1167 panic("UMA: page_free used with invalid flags %x", flags);
1169 kmem_free(vmem, (vm_offset_t)mem, size);
1173 * Zero fill initializer
1175 * Arguments/Returns follow uma_init specifications
1178 zero_init(void *mem, int size, int flags)
1185 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1188 * keg The zone we should initialize
1194 keg_small_init(uma_keg_t keg)
1202 if (keg->uk_flags & UMA_ZONE_PCPU) {
1203 u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU;
1205 slabsize = sizeof(struct pcpu);
1206 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1209 slabsize = UMA_SLAB_SIZE;
1214 * Calculate the size of each allocation (rsize) according to
1215 * alignment. If the requested size is smaller than we have
1216 * allocation bits for we round it up.
1218 rsize = keg->uk_size;
1219 if (rsize < slabsize / SLAB_SETSIZE)
1220 rsize = slabsize / SLAB_SETSIZE;
1221 if (rsize & keg->uk_align)
1222 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1223 keg->uk_rsize = rsize;
1225 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1226 keg->uk_rsize < sizeof(struct pcpu),
1227 ("%s: size %u too large", __func__, keg->uk_rsize));
1229 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1232 shsize = sizeof(struct uma_slab);
1234 keg->uk_ipers = (slabsize - shsize) / rsize;
1235 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1236 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1238 memused = keg->uk_ipers * rsize + shsize;
1239 wastedspace = slabsize - memused;
1242 * We can't do OFFPAGE if we're internal or if we've been
1243 * asked to not go to the VM for buckets. If we do this we
1244 * may end up going to the VM for slabs which we do not
1245 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1246 * of UMA_ZONE_VM, which clearly forbids it.
1248 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1249 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1253 * See if using an OFFPAGE slab will limit our waste. Only do
1254 * this if it permits more items per-slab.
1256 * XXX We could try growing slabsize to limit max waste as well.
1257 * Historically this was not done because the VM could not
1258 * efficiently handle contiguous allocations.
1260 if ((wastedspace >= slabsize / UMA_MAX_WASTE) &&
1261 (keg->uk_ipers < (slabsize / keg->uk_rsize))) {
1262 keg->uk_ipers = slabsize / keg->uk_rsize;
1263 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1264 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1265 CTR6(KTR_UMA, "UMA decided we need offpage slab headers for "
1266 "keg: %s(%p), calculated wastedspace = %d, "
1267 "maximum wasted space allowed = %d, "
1268 "calculated ipers = %d, "
1269 "new wasted space = %d\n", keg->uk_name, keg, wastedspace,
1270 slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1271 slabsize - keg->uk_ipers * keg->uk_rsize);
1272 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1275 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1276 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1277 keg->uk_flags |= UMA_ZONE_HASH;
1281 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1282 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1286 * keg The keg we should initialize
1292 keg_large_init(uma_keg_t keg)
1296 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1297 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1298 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1299 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1300 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1302 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1304 keg->uk_rsize = keg->uk_size;
1306 /* Check whether we have enough space to not do OFFPAGE. */
1307 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1308 shsize = sizeof(struct uma_slab);
1309 if (shsize & UMA_ALIGN_PTR)
1310 shsize = (shsize & ~UMA_ALIGN_PTR) +
1311 (UMA_ALIGN_PTR + 1);
1313 if (PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < shsize) {
1315 * We can't do OFFPAGE if we're internal, in which case
1316 * we need an extra page per allocation to contain the
1319 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) == 0)
1320 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1326 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1327 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1328 keg->uk_flags |= UMA_ZONE_HASH;
1332 keg_cachespread_init(uma_keg_t keg)
1339 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1340 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1342 alignsize = keg->uk_align + 1;
1343 rsize = keg->uk_size;
1345 * We want one item to start on every align boundary in a page. To
1346 * do this we will span pages. We will also extend the item by the
1347 * size of align if it is an even multiple of align. Otherwise, it
1348 * would fall on the same boundary every time.
1350 if (rsize & keg->uk_align)
1351 rsize = (rsize & ~keg->uk_align) + alignsize;
1352 if ((rsize & alignsize) == 0)
1354 trailer = rsize - keg->uk_size;
1355 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1356 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1357 keg->uk_rsize = rsize;
1358 keg->uk_ppera = pages;
1359 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1360 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1361 KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1362 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1367 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1368 * the keg onto the global keg list.
1370 * Arguments/Returns follow uma_ctor specifications
1371 * udata Actually uma_kctor_args
1374 keg_ctor(void *mem, int size, void *udata, int flags)
1376 struct uma_kctor_args *arg = udata;
1377 uma_keg_t keg = mem;
1381 keg->uk_size = arg->size;
1382 keg->uk_init = arg->uminit;
1383 keg->uk_fini = arg->fini;
1384 keg->uk_align = arg->align;
1386 keg->uk_reserve = 0;
1388 keg->uk_flags = arg->flags;
1389 keg->uk_slabzone = NULL;
1392 * The master zone is passed to us at keg-creation time.
1395 keg->uk_name = zone->uz_name;
1397 if (arg->flags & UMA_ZONE_VM)
1398 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1400 if (arg->flags & UMA_ZONE_ZINIT)
1401 keg->uk_init = zero_init;
1403 if (arg->flags & UMA_ZONE_MALLOC)
1404 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1406 if (arg->flags & UMA_ZONE_PCPU)
1408 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1410 keg->uk_flags &= ~UMA_ZONE_PCPU;
1413 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1414 keg_cachespread_init(keg);
1416 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1417 keg_large_init(keg);
1419 keg_small_init(keg);
1422 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1423 keg->uk_slabzone = slabzone;
1426 * If we haven't booted yet we need allocations to go through the
1427 * startup cache until the vm is ready.
1429 if (booted < UMA_STARTUP2)
1430 keg->uk_allocf = startup_alloc;
1431 #ifdef UMA_MD_SMALL_ALLOC
1432 else if (keg->uk_ppera == 1)
1433 keg->uk_allocf = uma_small_alloc;
1436 keg->uk_allocf = page_alloc;
1437 #ifdef UMA_MD_SMALL_ALLOC
1438 if (keg->uk_ppera == 1)
1439 keg->uk_freef = uma_small_free;
1442 keg->uk_freef = page_free;
1445 * Initialize keg's lock
1447 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1450 * If we're putting the slab header in the actual page we need to
1451 * figure out where in each page it goes. This calculates a right
1452 * justified offset into the memory on an ALIGN_PTR boundary.
1454 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1457 /* Size of the slab struct and free list */
1458 totsize = sizeof(struct uma_slab);
1460 if (totsize & UMA_ALIGN_PTR)
1461 totsize = (totsize & ~UMA_ALIGN_PTR) +
1462 (UMA_ALIGN_PTR + 1);
1463 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1466 * The only way the following is possible is if with our
1467 * UMA_ALIGN_PTR adjustments we are now bigger than
1468 * UMA_SLAB_SIZE. I haven't checked whether this is
1469 * mathematically possible for all cases, so we make
1472 totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1473 if (totsize > PAGE_SIZE * keg->uk_ppera) {
1474 printf("zone %s ipers %d rsize %d size %d\n",
1475 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1477 panic("UMA slab won't fit.");
1481 if (keg->uk_flags & UMA_ZONE_HASH)
1482 hash_alloc(&keg->uk_hash);
1484 CTR5(KTR_UMA, "keg_ctor %p zone %s(%p) out %d free %d\n",
1485 keg, zone->uz_name, zone,
1486 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free,
1489 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1491 rw_wlock(&uma_rwlock);
1492 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1493 rw_wunlock(&uma_rwlock);
1498 * Zone header ctor. This initializes all fields, locks, etc.
1500 * Arguments/Returns follow uma_ctor specifications
1501 * udata Actually uma_zctor_args
1504 zone_ctor(void *mem, int size, void *udata, int flags)
1506 struct uma_zctor_args *arg = udata;
1507 uma_zone_t zone = mem;
1512 zone->uz_name = arg->name;
1513 zone->uz_ctor = arg->ctor;
1514 zone->uz_dtor = arg->dtor;
1515 zone->uz_slab = zone_fetch_slab;
1516 zone->uz_init = NULL;
1517 zone->uz_fini = NULL;
1518 zone->uz_allocs = 0;
1521 zone->uz_sleeps = 0;
1523 zone->uz_count_min = 0;
1525 zone->uz_warning = NULL;
1526 timevalclear(&zone->uz_ratecheck);
1529 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1532 * This is a pure cache zone, no kegs.
1535 if (arg->flags & UMA_ZONE_VM)
1536 arg->flags |= UMA_ZFLAG_CACHEONLY;
1537 zone->uz_flags = arg->flags;
1538 zone->uz_size = arg->size;
1539 zone->uz_import = arg->import;
1540 zone->uz_release = arg->release;
1541 zone->uz_arg = arg->arg;
1542 zone->uz_lockptr = &zone->uz_lock;
1543 rw_wlock(&uma_rwlock);
1544 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1545 rw_wunlock(&uma_rwlock);
1550 * Use the regular zone/keg/slab allocator.
1552 zone->uz_import = (uma_import)zone_import;
1553 zone->uz_release = (uma_release)zone_release;
1554 zone->uz_arg = zone;
1556 if (arg->flags & UMA_ZONE_SECONDARY) {
1557 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1558 zone->uz_init = arg->uminit;
1559 zone->uz_fini = arg->fini;
1560 zone->uz_lockptr = &keg->uk_lock;
1561 zone->uz_flags |= UMA_ZONE_SECONDARY;
1562 rw_wlock(&uma_rwlock);
1564 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1565 if (LIST_NEXT(z, uz_link) == NULL) {
1566 LIST_INSERT_AFTER(z, zone, uz_link);
1571 rw_wunlock(&uma_rwlock);
1572 } else if (keg == NULL) {
1573 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1574 arg->align, arg->flags)) == NULL)
1577 struct uma_kctor_args karg;
1580 /* We should only be here from uma_startup() */
1581 karg.size = arg->size;
1582 karg.uminit = arg->uminit;
1583 karg.fini = arg->fini;
1584 karg.align = arg->align;
1585 karg.flags = arg->flags;
1587 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1594 * Link in the first keg.
1596 zone->uz_klink.kl_keg = keg;
1597 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1598 zone->uz_lockptr = &keg->uk_lock;
1599 zone->uz_size = keg->uk_size;
1600 zone->uz_flags |= (keg->uk_flags &
1601 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1604 * Some internal zones don't have room allocated for the per cpu
1605 * caches. If we're internal, bail out here.
1607 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1608 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1609 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1614 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1615 zone->uz_count = bucket_select(zone->uz_size);
1617 zone->uz_count = BUCKET_MAX;
1618 zone->uz_count_min = zone->uz_count;
1624 * Keg header dtor. This frees all data, destroys locks, frees the hash
1625 * table and removes the keg from the global list.
1627 * Arguments/Returns follow uma_dtor specifications
1631 keg_dtor(void *arg, int size, void *udata)
1635 keg = (uma_keg_t)arg;
1637 if (keg->uk_free != 0) {
1638 printf("Freed UMA keg (%s) was not empty (%d items). "
1639 " Lost %d pages of memory.\n",
1640 keg->uk_name ? keg->uk_name : "",
1641 keg->uk_free, keg->uk_pages);
1645 hash_free(&keg->uk_hash);
1653 * Arguments/Returns follow uma_dtor specifications
1657 zone_dtor(void *arg, int size, void *udata)
1663 zone = (uma_zone_t)arg;
1664 keg = zone_first_keg(zone);
1666 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1669 rw_wlock(&uma_rwlock);
1670 LIST_REMOVE(zone, uz_link);
1671 rw_wunlock(&uma_rwlock);
1673 * XXX there are some races here where
1674 * the zone can be drained but zone lock
1675 * released and then refilled before we
1676 * remove it... we dont care for now
1678 zone_drain_wait(zone, M_WAITOK);
1680 * Unlink all of our kegs.
1682 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1683 klink->kl_keg = NULL;
1684 LIST_REMOVE(klink, kl_link);
1685 if (klink == &zone->uz_klink)
1687 free(klink, M_TEMP);
1690 * We only destroy kegs from non secondary zones.
1692 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1693 rw_wlock(&uma_rwlock);
1694 LIST_REMOVE(keg, uk_link);
1695 rw_wunlock(&uma_rwlock);
1696 zone_free_item(kegs, keg, NULL, SKIP_NONE);
1698 ZONE_LOCK_FINI(zone);
1702 * Traverses every zone in the system and calls a callback
1705 * zfunc A pointer to a function which accepts a zone
1712 zone_foreach(void (*zfunc)(uma_zone_t))
1717 rw_rlock(&uma_rwlock);
1718 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1719 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1722 rw_runlock(&uma_rwlock);
1725 /* Public functions */
1728 uma_startup(void *mem, int npages)
1730 struct uma_zctor_args args;
1732 rw_init(&uma_rwlock, "UMA lock");
1734 /* "manually" create the initial zone */
1735 memset(&args, 0, sizeof(args));
1736 args.name = "UMA Kegs";
1737 args.size = sizeof(struct uma_keg);
1738 args.ctor = keg_ctor;
1739 args.dtor = keg_dtor;
1740 args.uminit = zero_init;
1742 args.keg = &masterkeg;
1743 args.align = 32 - 1;
1744 args.flags = UMA_ZFLAG_INTERNAL;
1745 /* The initial zone has no Per cpu queues so it's smaller */
1746 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1748 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1750 boot_pages = npages;
1752 args.name = "UMA Zones";
1753 args.size = sizeof(struct uma_zone) +
1754 (sizeof(struct uma_cache) * (mp_maxid + 1));
1755 args.ctor = zone_ctor;
1756 args.dtor = zone_dtor;
1757 args.uminit = zero_init;
1760 args.align = 32 - 1;
1761 args.flags = UMA_ZFLAG_INTERNAL;
1762 /* The initial zone has no Per cpu queues so it's smaller */
1763 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1765 /* Now make a zone for slab headers */
1766 slabzone = uma_zcreate("UMA Slabs",
1767 sizeof(struct uma_slab),
1768 NULL, NULL, NULL, NULL,
1769 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1771 hashzone = uma_zcreate("UMA Hash",
1772 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1773 NULL, NULL, NULL, NULL,
1774 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1778 booted = UMA_STARTUP;
1785 booted = UMA_STARTUP2;
1787 sx_init(&uma_drain_lock, "umadrain");
1791 * Initialize our callout handle
1799 callout_init(&uma_callout, 1);
1800 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1804 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1805 int align, uint32_t flags)
1807 struct uma_kctor_args args;
1810 args.uminit = uminit;
1812 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1815 return (zone_alloc_item(kegs, &args, M_WAITOK));
1820 uma_set_align(int align)
1823 if (align != UMA_ALIGN_CACHE)
1824 uma_align_cache = align;
1829 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1830 uma_init uminit, uma_fini fini, int align, uint32_t flags)
1833 struct uma_zctor_args args;
1837 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
1840 /* This stuff is essential for the zone ctor */
1841 memset(&args, 0, sizeof(args));
1846 args.uminit = uminit;
1850 * If a zone is being created with an empty constructor and
1851 * destructor, pass UMA constructor/destructor which checks for
1852 * memory use after free.
1854 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) &&
1855 ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) {
1856 args.ctor = trash_ctor;
1857 args.dtor = trash_dtor;
1858 args.uminit = trash_init;
1859 args.fini = trash_fini;
1866 if (booted < UMA_STARTUP2) {
1869 sx_slock(&uma_drain_lock);
1872 res = zone_alloc_item(zones, &args, M_WAITOK);
1874 sx_sunlock(&uma_drain_lock);
1880 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1881 uma_init zinit, uma_fini zfini, uma_zone_t master)
1883 struct uma_zctor_args args;
1888 keg = zone_first_keg(master);
1889 memset(&args, 0, sizeof(args));
1891 args.size = keg->uk_size;
1894 args.uminit = zinit;
1896 args.align = keg->uk_align;
1897 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1900 if (booted < UMA_STARTUP2) {
1903 sx_slock(&uma_drain_lock);
1906 /* XXX Attaches only one keg of potentially many. */
1907 res = zone_alloc_item(zones, &args, M_WAITOK);
1909 sx_sunlock(&uma_drain_lock);
1915 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1916 uma_init zinit, uma_fini zfini, uma_import zimport,
1917 uma_release zrelease, void *arg, int flags)
1919 struct uma_zctor_args args;
1921 memset(&args, 0, sizeof(args));
1926 args.uminit = zinit;
1928 args.import = zimport;
1929 args.release = zrelease;
1934 return (zone_alloc_item(zones, &args, M_WAITOK));
1938 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1942 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
1945 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
1950 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1958 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1965 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1967 zone_lock_pair(zone, master);
1969 * zone must use vtoslab() to resolve objects and must already be
1972 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1973 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1978 * The new master must also use vtoslab().
1980 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
1986 * The underlying object must be the same size. rsize
1989 if (master->uz_size != zone->uz_size) {
1994 * Put it at the end of the list.
1996 klink->kl_keg = zone_first_keg(master);
1997 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
1998 if (LIST_NEXT(kl, kl_link) == NULL) {
1999 LIST_INSERT_AFTER(kl, klink, kl_link);
2004 zone->uz_flags |= UMA_ZFLAG_MULTI;
2005 zone->uz_slab = zone_fetch_slab_multi;
2008 zone_unlock_pair(zone, master);
2010 free(klink, M_TEMP);
2018 uma_zdestroy(uma_zone_t zone)
2021 sx_slock(&uma_drain_lock);
2022 zone_free_item(zones, zone, NULL, SKIP_NONE);
2023 sx_sunlock(&uma_drain_lock);
2027 uma_zwait(uma_zone_t zone)
2031 item = uma_zalloc_arg(zone, NULL, M_WAITOK);
2032 uma_zfree(zone, item);
2037 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2041 uma_bucket_t bucket;
2045 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2046 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2048 /* This is the fast path allocation */
2049 CTR4(KTR_UMA, "uma_zalloc_arg thread %x zone %s(%p) flags %d",
2050 curthread, zone->uz_name, zone, flags);
2052 if (flags & M_WAITOK) {
2053 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2054 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2056 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2057 ("uma_zalloc_arg: called with spinlock or critical section held"));
2059 #ifdef DEBUG_MEMGUARD
2060 if (memguard_cmp_zone(zone)) {
2061 item = memguard_alloc(zone->uz_size, flags);
2063 if (zone->uz_init != NULL &&
2064 zone->uz_init(item, zone->uz_size, flags) != 0)
2066 if (zone->uz_ctor != NULL &&
2067 zone->uz_ctor(item, zone->uz_size, udata,
2069 zone->uz_fini(item, zone->uz_size);
2074 /* This is unfortunate but should not be fatal. */
2078 * If possible, allocate from the per-CPU cache. There are two
2079 * requirements for safe access to the per-CPU cache: (1) the thread
2080 * accessing the cache must not be preempted or yield during access,
2081 * and (2) the thread must not migrate CPUs without switching which
2082 * cache it accesses. We rely on a critical section to prevent
2083 * preemption and migration. We release the critical section in
2084 * order to acquire the zone mutex if we are unable to allocate from
2085 * the current cache; when we re-acquire the critical section, we
2086 * must detect and handle migration if it has occurred.
2090 cache = &zone->uz_cpu[cpu];
2093 bucket = cache->uc_allocbucket;
2094 if (bucket != NULL && bucket->ub_cnt > 0) {
2096 item = bucket->ub_bucket[bucket->ub_cnt];
2098 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2100 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2103 if (zone->uz_ctor != NULL &&
2104 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2105 atomic_add_long(&zone->uz_fails, 1);
2106 zone_free_item(zone, item, udata, SKIP_DTOR);
2110 uma_dbg_alloc(zone, NULL, item);
2113 uma_zero_item(item, zone);
2118 * We have run out of items in our alloc bucket.
2119 * See if we can switch with our free bucket.
2121 bucket = cache->uc_freebucket;
2122 if (bucket != NULL && bucket->ub_cnt > 0) {
2124 "uma_zalloc: zone %s(%p) swapping empty with alloc",
2125 zone->uz_name, zone);
2126 cache->uc_freebucket = cache->uc_allocbucket;
2127 cache->uc_allocbucket = bucket;
2132 * Discard any empty allocation bucket while we hold no locks.
2134 bucket = cache->uc_allocbucket;
2135 cache->uc_allocbucket = NULL;
2138 bucket_free(zone, bucket, udata);
2140 /* Short-circuit for zones without buckets and low memory. */
2141 if (zone->uz_count == 0 || bucketdisable)
2145 * Attempt to retrieve the item from the per-CPU cache has failed, so
2146 * we must go back to the zone. This requires the zone lock, so we
2147 * must drop the critical section, then re-acquire it when we go back
2148 * to the cache. Since the critical section is released, we may be
2149 * preempted or migrate. As such, make sure not to maintain any
2150 * thread-local state specific to the cache from prior to releasing
2151 * the critical section.
2154 if (ZONE_TRYLOCK(zone) == 0) {
2155 /* Record contention to size the buckets. */
2161 cache = &zone->uz_cpu[cpu];
2164 * Since we have locked the zone we may as well send back our stats.
2166 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2167 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2168 cache->uc_allocs = 0;
2169 cache->uc_frees = 0;
2171 /* See if we lost the race to fill the cache. */
2172 if (cache->uc_allocbucket != NULL) {
2178 * Check the zone's cache of buckets.
2180 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2181 KASSERT(bucket->ub_cnt != 0,
2182 ("uma_zalloc_arg: Returning an empty bucket."));
2184 LIST_REMOVE(bucket, ub_link);
2185 cache->uc_allocbucket = bucket;
2189 /* We are no longer associated with this CPU. */
2193 * We bump the uz count when the cache size is insufficient to
2194 * handle the working set.
2196 if (lockfail && zone->uz_count < BUCKET_MAX)
2201 * Now lets just fill a bucket and put it on the free list. If that
2202 * works we'll restart the allocation from the beginning and it
2203 * will use the just filled bucket.
2205 bucket = zone_alloc_bucket(zone, udata, flags);
2206 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
2207 zone->uz_name, zone, bucket);
2208 if (bucket != NULL) {
2212 cache = &zone->uz_cpu[cpu];
2214 * See if we lost the race or were migrated. Cache the
2215 * initialized bucket to make this less likely or claim
2216 * the memory directly.
2218 if (cache->uc_allocbucket == NULL)
2219 cache->uc_allocbucket = bucket;
2221 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2227 * We may not be able to get a bucket so return an actual item.
2230 item = zone_alloc_item(zone, udata, flags);
2236 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2241 mtx_assert(&keg->uk_lock, MA_OWNED);
2244 if ((flags & M_USE_RESERVE) == 0)
2245 reserve = keg->uk_reserve;
2249 * Find a slab with some space. Prefer slabs that are partially
2250 * used over those that are totally full. This helps to reduce
2253 if (keg->uk_free > reserve) {
2254 if (!LIST_EMPTY(&keg->uk_part_slab)) {
2255 slab = LIST_FIRST(&keg->uk_part_slab);
2257 slab = LIST_FIRST(&keg->uk_free_slab);
2258 LIST_REMOVE(slab, us_link);
2259 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2262 MPASS(slab->us_keg == keg);
2267 * M_NOVM means don't ask at all!
2272 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2273 keg->uk_flags |= UMA_ZFLAG_FULL;
2275 * If this is not a multi-zone, set the FULL bit.
2276 * Otherwise slab_multi() takes care of it.
2278 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2279 zone->uz_flags |= UMA_ZFLAG_FULL;
2280 zone_log_warning(zone);
2281 zone_maxaction(zone);
2283 if (flags & M_NOWAIT)
2286 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2289 slab = keg_alloc_slab(keg, zone, flags);
2291 * If we got a slab here it's safe to mark it partially used
2292 * and return. We assume that the caller is going to remove
2293 * at least one item.
2296 MPASS(slab->us_keg == keg);
2297 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2301 * We might not have been able to get a slab but another cpu
2302 * could have while we were unlocked. Check again before we
2311 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2316 keg = zone_first_keg(zone);
2321 slab = keg_fetch_slab(keg, zone, flags);
2324 if (flags & (M_NOWAIT | M_NOVM))
2332 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2333 * with the keg locked. On NULL no lock is held.
2335 * The last pointer is used to seed the search. It is not required.
2338 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2348 * Don't wait on the first pass. This will skip limit tests
2349 * as well. We don't want to block if we can find a provider
2352 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2354 * Use the last slab allocated as a hint for where to start
2358 slab = keg_fetch_slab(last, zone, flags);
2364 * Loop until we have a slab incase of transient failures
2365 * while M_WAITOK is specified. I'm not sure this is 100%
2366 * required but we've done it for so long now.
2372 * Search the available kegs for slabs. Be careful to hold the
2373 * correct lock while calling into the keg layer.
2375 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2376 keg = klink->kl_keg;
2378 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2379 slab = keg_fetch_slab(keg, zone, flags);
2383 if (keg->uk_flags & UMA_ZFLAG_FULL)
2389 if (rflags & (M_NOWAIT | M_NOVM))
2393 * All kegs are full. XXX We can't atomically check all kegs
2394 * and sleep so just sleep for a short period and retry.
2396 if (full && !empty) {
2398 zone->uz_flags |= UMA_ZFLAG_FULL;
2400 zone_log_warning(zone);
2401 zone_maxaction(zone);
2402 msleep(zone, zone->uz_lockptr, PVM,
2403 "zonelimit", hz/100);
2404 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2413 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2418 MPASS(keg == slab->us_keg);
2419 mtx_assert(&keg->uk_lock, MA_OWNED);
2421 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2422 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2423 item = slab->us_data + (keg->uk_rsize * freei);
2424 slab->us_freecount--;
2427 /* Move this slab to the full list */
2428 if (slab->us_freecount == 0) {
2429 LIST_REMOVE(slab, us_link);
2430 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2437 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2445 /* Try to keep the buckets totally full */
2446 for (i = 0; i < max; ) {
2447 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2450 while (slab->us_freecount && i < max) {
2451 bucket[i++] = slab_alloc_item(keg, slab);
2452 if (keg->uk_free <= keg->uk_reserve)
2455 /* Don't grab more than one slab at a time. */
2466 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2468 uma_bucket_t bucket;
2471 /* Don't wait for buckets, preserve caller's NOVM setting. */
2472 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2476 max = MIN(bucket->ub_entries, zone->uz_count);
2477 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2481 * Initialize the memory if necessary.
2483 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2486 for (i = 0; i < bucket->ub_cnt; i++)
2487 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2491 * If we couldn't initialize the whole bucket, put the
2492 * rest back onto the freelist.
2494 if (i != bucket->ub_cnt) {
2495 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2496 bucket->ub_cnt - i);
2498 bzero(&bucket->ub_bucket[i],
2499 sizeof(void *) * (bucket->ub_cnt - i));
2505 if (bucket->ub_cnt == 0) {
2506 bucket_free(zone, bucket, udata);
2507 atomic_add_long(&zone->uz_fails, 1);
2515 * Allocates a single item from a zone.
2518 * zone The zone to alloc for.
2519 * udata The data to be passed to the constructor.
2520 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2523 * NULL if there is no memory and M_NOWAIT is set
2524 * An item if successful
2528 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2534 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2536 atomic_add_long(&zone->uz_allocs, 1);
2539 * We have to call both the zone's init (not the keg's init)
2540 * and the zone's ctor. This is because the item is going from
2541 * a keg slab directly to the user, and the user is expecting it
2542 * to be both zone-init'd as well as zone-ctor'd.
2544 if (zone->uz_init != NULL) {
2545 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2546 zone_free_item(zone, item, udata, SKIP_FINI);
2550 if (zone->uz_ctor != NULL) {
2551 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2552 zone_free_item(zone, item, udata, SKIP_DTOR);
2557 uma_dbg_alloc(zone, NULL, item);
2560 uma_zero_item(item, zone);
2562 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
2563 zone->uz_name, zone);
2568 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
2569 zone->uz_name, zone);
2570 atomic_add_long(&zone->uz_fails, 1);
2576 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2579 uma_bucket_t bucket;
2583 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2584 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2586 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2589 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2590 ("uma_zfree_arg: called with spinlock or critical section held"));
2592 /* uma_zfree(..., NULL) does nothing, to match free(9). */
2595 #ifdef DEBUG_MEMGUARD
2596 if (is_memguard_addr(item)) {
2597 if (zone->uz_dtor != NULL)
2598 zone->uz_dtor(item, zone->uz_size, udata);
2599 if (zone->uz_fini != NULL)
2600 zone->uz_fini(item, zone->uz_size);
2601 memguard_free(item);
2606 if (zone->uz_flags & UMA_ZONE_MALLOC)
2607 uma_dbg_free(zone, udata, item);
2609 uma_dbg_free(zone, NULL, item);
2611 if (zone->uz_dtor != NULL)
2612 zone->uz_dtor(item, zone->uz_size, udata);
2615 * The race here is acceptable. If we miss it we'll just have to wait
2616 * a little longer for the limits to be reset.
2618 if (zone->uz_flags & UMA_ZFLAG_FULL)
2622 * If possible, free to the per-CPU cache. There are two
2623 * requirements for safe access to the per-CPU cache: (1) the thread
2624 * accessing the cache must not be preempted or yield during access,
2625 * and (2) the thread must not migrate CPUs without switching which
2626 * cache it accesses. We rely on a critical section to prevent
2627 * preemption and migration. We release the critical section in
2628 * order to acquire the zone mutex if we are unable to free to the
2629 * current cache; when we re-acquire the critical section, we must
2630 * detect and handle migration if it has occurred.
2635 cache = &zone->uz_cpu[cpu];
2639 * Try to free into the allocbucket first to give LIFO ordering
2640 * for cache-hot datastructures. Spill over into the freebucket
2641 * if necessary. Alloc will swap them if one runs dry.
2643 bucket = cache->uc_allocbucket;
2644 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2645 bucket = cache->uc_freebucket;
2646 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2647 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2648 ("uma_zfree: Freeing to non free bucket index."));
2649 bucket->ub_bucket[bucket->ub_cnt] = item;
2657 * We must go back the zone, which requires acquiring the zone lock,
2658 * which in turn means we must release and re-acquire the critical
2659 * section. Since the critical section is released, we may be
2660 * preempted or migrate. As such, make sure not to maintain any
2661 * thread-local state specific to the cache from prior to releasing
2662 * the critical section.
2665 if (zone->uz_count == 0 || bucketdisable)
2669 if (ZONE_TRYLOCK(zone) == 0) {
2670 /* Record contention to size the buckets. */
2676 cache = &zone->uz_cpu[cpu];
2679 * Since we have locked the zone we may as well send back our stats.
2681 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2682 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2683 cache->uc_allocs = 0;
2684 cache->uc_frees = 0;
2686 bucket = cache->uc_freebucket;
2687 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2691 cache->uc_freebucket = NULL;
2692 /* We are no longer associated with this CPU. */
2695 /* Can we throw this on the zone full list? */
2696 if (bucket != NULL) {
2698 "uma_zfree: zone %s(%p) putting bucket %p on free list",
2699 zone->uz_name, zone, bucket);
2700 /* ub_cnt is pointing to the last free item */
2701 KASSERT(bucket->ub_cnt != 0,
2702 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2703 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2707 * We bump the uz count when the cache size is insufficient to
2708 * handle the working set.
2710 if (lockfail && zone->uz_count < BUCKET_MAX)
2714 bucket = bucket_alloc(zone, udata, M_NOWAIT);
2715 CTR3(KTR_UMA, "uma_zfree: zone %s(%p) allocated bucket %p",
2716 zone->uz_name, zone, bucket);
2720 cache = &zone->uz_cpu[cpu];
2721 if (cache->uc_freebucket == NULL) {
2722 cache->uc_freebucket = bucket;
2726 * We lost the race, start over. We have to drop our
2727 * critical section to free the bucket.
2730 bucket_free(zone, bucket, udata);
2735 * If nothing else caught this, we'll just do an internal free.
2738 zone_free_item(zone, item, udata, SKIP_DTOR);
2744 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2748 mtx_assert(&keg->uk_lock, MA_OWNED);
2749 MPASS(keg == slab->us_keg);
2751 /* Do we need to remove from any lists? */
2752 if (slab->us_freecount+1 == keg->uk_ipers) {
2753 LIST_REMOVE(slab, us_link);
2754 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2755 } else if (slab->us_freecount == 0) {
2756 LIST_REMOVE(slab, us_link);
2757 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2760 /* Slab management. */
2761 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2762 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2763 slab->us_freecount++;
2765 /* Keg statistics. */
2770 zone_release(uma_zone_t zone, void **bucket, int cnt)
2780 keg = zone_first_keg(zone);
2782 for (i = 0; i < cnt; i++) {
2784 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2785 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2786 if (zone->uz_flags & UMA_ZONE_HASH) {
2787 slab = hash_sfind(&keg->uk_hash, mem);
2789 mem += keg->uk_pgoff;
2790 slab = (uma_slab_t)mem;
2793 slab = vtoslab((vm_offset_t)item);
2794 if (slab->us_keg != keg) {
2800 slab_free_item(keg, slab, item);
2801 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2802 if (keg->uk_pages < keg->uk_maxpages) {
2803 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2808 * We can handle one more allocation. Since we're
2809 * clearing ZFLAG_FULL, wake up all procs blocked
2810 * on pages. This should be uncommon, so keeping this
2811 * simple for now (rather than adding count of blocked
2820 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2828 * Frees a single item to any zone.
2831 * zone The zone to free to
2832 * item The item we're freeing
2833 * udata User supplied data for the dtor
2834 * skip Skip dtors and finis
2837 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2841 if (skip == SKIP_NONE) {
2842 if (zone->uz_flags & UMA_ZONE_MALLOC)
2843 uma_dbg_free(zone, udata, item);
2845 uma_dbg_free(zone, NULL, item);
2848 if (skip < SKIP_DTOR && zone->uz_dtor)
2849 zone->uz_dtor(item, zone->uz_size, udata);
2851 if (skip < SKIP_FINI && zone->uz_fini)
2852 zone->uz_fini(item, zone->uz_size);
2854 atomic_add_long(&zone->uz_frees, 1);
2855 zone->uz_release(zone->uz_arg, &item, 1);
2860 uma_zone_set_max(uma_zone_t zone, int nitems)
2864 keg = zone_first_keg(zone);
2868 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2869 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2870 keg->uk_maxpages += keg->uk_ppera;
2871 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers;
2879 uma_zone_get_max(uma_zone_t zone)
2884 keg = zone_first_keg(zone);
2888 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers;
2896 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2900 zone->uz_warning = warning;
2906 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
2910 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
2916 uma_zone_get_cur(uma_zone_t zone)
2922 nitems = zone->uz_allocs - zone->uz_frees;
2925 * See the comment in sysctl_vm_zone_stats() regarding the
2926 * safety of accessing the per-cpu caches. With the zone lock
2927 * held, it is safe, but can potentially result in stale data.
2929 nitems += zone->uz_cpu[i].uc_allocs -
2930 zone->uz_cpu[i].uc_frees;
2934 return (nitems < 0 ? 0 : nitems);
2939 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2943 keg = zone_first_keg(zone);
2944 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2946 KASSERT(keg->uk_pages == 0,
2947 ("uma_zone_set_init on non-empty keg"));
2948 keg->uk_init = uminit;
2954 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2958 keg = zone_first_keg(zone);
2959 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
2961 KASSERT(keg->uk_pages == 0,
2962 ("uma_zone_set_fini on non-empty keg"));
2963 keg->uk_fini = fini;
2969 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2973 KASSERT(zone_first_keg(zone)->uk_pages == 0,
2974 ("uma_zone_set_zinit on non-empty keg"));
2975 zone->uz_init = zinit;
2981 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2985 KASSERT(zone_first_keg(zone)->uk_pages == 0,
2986 ("uma_zone_set_zfini on non-empty keg"));
2987 zone->uz_fini = zfini;
2992 /* XXX uk_freef is not actually used with the zone locked */
2994 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2998 keg = zone_first_keg(zone);
2999 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
3001 keg->uk_freef = freef;
3006 /* XXX uk_allocf is not actually used with the zone locked */
3008 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3012 keg = zone_first_keg(zone);
3014 keg->uk_allocf = allocf;
3020 uma_zone_reserve(uma_zone_t zone, int items)
3024 keg = zone_first_keg(zone);
3028 keg->uk_reserve = items;
3036 uma_zone_reserve_kva(uma_zone_t zone, int count)
3042 keg = zone_first_keg(zone);
3045 pages = count / keg->uk_ipers;
3047 if (pages * keg->uk_ipers < count)
3049 pages *= keg->uk_ppera;
3051 #ifdef UMA_MD_SMALL_ALLOC
3052 if (keg->uk_ppera > 1) {
3056 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
3064 keg->uk_maxpages = pages;
3065 #ifdef UMA_MD_SMALL_ALLOC
3066 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3068 keg->uk_allocf = noobj_alloc;
3070 keg->uk_flags |= UMA_ZONE_NOFREE;
3078 uma_prealloc(uma_zone_t zone, int items)
3084 keg = zone_first_keg(zone);
3088 slabs = items / keg->uk_ipers;
3089 if (slabs * keg->uk_ipers < items)
3092 slab = keg_alloc_slab(keg, zone, M_WAITOK);
3095 MPASS(slab->us_keg == keg);
3096 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3104 uma_reclaim_locked(bool kmem_danger)
3107 CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
3108 sx_assert(&uma_drain_lock, SA_XLOCKED);
3110 zone_foreach(zone_drain);
3111 if (vm_page_count_min() || kmem_danger) {
3112 cache_drain_safe(NULL);
3113 zone_foreach(zone_drain);
3116 * Some slabs may have been freed but this zone will be visited early
3117 * we visit again so that we can free pages that are empty once other
3118 * zones are drained. We have to do the same for buckets.
3120 zone_drain(slabzone);
3121 bucket_zone_drain();
3128 sx_xlock(&uma_drain_lock);
3129 uma_reclaim_locked(false);
3130 sx_xunlock(&uma_drain_lock);
3133 static int uma_reclaim_needed;
3136 uma_reclaim_wakeup(void)
3139 uma_reclaim_needed = 1;
3140 wakeup(&uma_reclaim_needed);
3144 uma_reclaim_worker(void *arg __unused)
3147 sx_xlock(&uma_drain_lock);
3149 sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM,
3151 if (uma_reclaim_needed) {
3152 uma_reclaim_needed = 0;
3153 sx_xunlock(&uma_drain_lock);
3154 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
3155 sx_xlock(&uma_drain_lock);
3156 uma_reclaim_locked(true);
3163 uma_zone_exhausted(uma_zone_t zone)
3168 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3174 uma_zone_exhausted_nolock(uma_zone_t zone)
3176 return (zone->uz_flags & UMA_ZFLAG_FULL);
3180 uma_large_malloc(vm_size_t size, int wait)
3186 slab = zone_alloc_item(slabzone, NULL, wait);
3189 mem = page_alloc(NULL, size, &flags, wait);
3191 vsetslab((vm_offset_t)mem, slab);
3192 slab->us_data = mem;
3193 slab->us_flags = flags | UMA_SLAB_MALLOC;
3194 slab->us_size = size;
3196 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3203 uma_large_free(uma_slab_t slab)
3206 page_free(slab->us_data, slab->us_size, slab->us_flags);
3207 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3211 uma_zero_item(void *item, uma_zone_t zone)
3215 if (zone->uz_flags & UMA_ZONE_PCPU) {
3217 bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3219 bzero(item, zone->uz_size);
3223 uma_print_stats(void)
3225 zone_foreach(uma_print_zone);
3229 slab_print(uma_slab_t slab)
3231 printf("slab: keg %p, data %p, freecount %d\n",
3232 slab->us_keg, slab->us_data, slab->us_freecount);
3236 cache_print(uma_cache_t cache)
3238 printf("alloc: %p(%d), free: %p(%d)\n",
3239 cache->uc_allocbucket,
3240 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3241 cache->uc_freebucket,
3242 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3246 uma_print_keg(uma_keg_t keg)
3250 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3251 "out %d free %d limit %d\n",
3252 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3253 keg->uk_ipers, keg->uk_ppera,
3254 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free,
3255 keg->uk_free, (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3256 printf("Part slabs:\n");
3257 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3259 printf("Free slabs:\n");
3260 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3262 printf("Full slabs:\n");
3263 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3268 uma_print_zone(uma_zone_t zone)
3274 printf("zone: %s(%p) size %d flags %#x\n",
3275 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3276 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3277 uma_print_keg(kl->kl_keg);
3279 cache = &zone->uz_cpu[i];
3280 printf("CPU %d Cache:\n", i);
3287 * Generate statistics across both the zone and its per-cpu cache's. Return
3288 * desired statistics if the pointer is non-NULL for that statistic.
3290 * Note: does not update the zone statistics, as it can't safely clear the
3291 * per-CPU cache statistic.
3293 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3294 * safe from off-CPU; we should modify the caches to track this information
3295 * directly so that we don't have to.
3298 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3299 uint64_t *freesp, uint64_t *sleepsp)
3302 uint64_t allocs, frees, sleeps;
3305 allocs = frees = sleeps = 0;
3308 cache = &z->uz_cpu[cpu];
3309 if (cache->uc_allocbucket != NULL)
3310 cachefree += cache->uc_allocbucket->ub_cnt;
3311 if (cache->uc_freebucket != NULL)
3312 cachefree += cache->uc_freebucket->ub_cnt;
3313 allocs += cache->uc_allocs;
3314 frees += cache->uc_frees;
3316 allocs += z->uz_allocs;
3317 frees += z->uz_frees;
3318 sleeps += z->uz_sleeps;
3319 if (cachefreep != NULL)
3320 *cachefreep = cachefree;
3321 if (allocsp != NULL)
3325 if (sleepsp != NULL)
3331 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3338 rw_rlock(&uma_rwlock);
3339 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3340 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3343 rw_runlock(&uma_rwlock);
3344 return (sysctl_handle_int(oidp, &count, 0, req));
3348 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3350 struct uma_stream_header ush;
3351 struct uma_type_header uth;
3352 struct uma_percpu_stat ups;
3353 uma_bucket_t bucket;
3360 int count, error, i;
3362 error = sysctl_wire_old_buffer(req, 0);
3365 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3366 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3369 rw_rlock(&uma_rwlock);
3370 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3371 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3376 * Insert stream header.
3378 bzero(&ush, sizeof(ush));
3379 ush.ush_version = UMA_STREAM_VERSION;
3380 ush.ush_maxcpus = (mp_maxid + 1);
3381 ush.ush_count = count;
3382 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3384 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3385 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3386 bzero(&uth, sizeof(uth));
3388 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3389 uth.uth_align = kz->uk_align;
3390 uth.uth_size = kz->uk_size;
3391 uth.uth_rsize = kz->uk_rsize;
3392 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3394 uth.uth_maxpages += k->uk_maxpages;
3395 uth.uth_pages += k->uk_pages;
3396 uth.uth_keg_free += k->uk_free;
3397 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3402 * A zone is secondary is it is not the first entry
3403 * on the keg's zone list.
3405 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3406 (LIST_FIRST(&kz->uk_zones) != z))
3407 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3409 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3410 uth.uth_zone_free += bucket->ub_cnt;
3411 uth.uth_allocs = z->uz_allocs;
3412 uth.uth_frees = z->uz_frees;
3413 uth.uth_fails = z->uz_fails;
3414 uth.uth_sleeps = z->uz_sleeps;
3415 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3417 * While it is not normally safe to access the cache
3418 * bucket pointers while not on the CPU that owns the
3419 * cache, we only allow the pointers to be exchanged
3420 * without the zone lock held, not invalidated, so
3421 * accept the possible race associated with bucket
3422 * exchange during monitoring.
3424 for (i = 0; i < (mp_maxid + 1); i++) {
3425 bzero(&ups, sizeof(ups));
3426 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3430 cache = &z->uz_cpu[i];
3431 if (cache->uc_allocbucket != NULL)
3432 ups.ups_cache_free +=
3433 cache->uc_allocbucket->ub_cnt;
3434 if (cache->uc_freebucket != NULL)
3435 ups.ups_cache_free +=
3436 cache->uc_freebucket->ub_cnt;
3437 ups.ups_allocs = cache->uc_allocs;
3438 ups.ups_frees = cache->uc_frees;
3440 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3445 rw_runlock(&uma_rwlock);
3446 error = sbuf_finish(&sbuf);
3452 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3454 uma_zone_t zone = *(uma_zone_t *)arg1;
3457 max = uma_zone_get_max(zone);
3458 error = sysctl_handle_int(oidp, &max, 0, req);
3459 if (error || !req->newptr)
3462 uma_zone_set_max(zone, max);
3468 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3470 uma_zone_t zone = *(uma_zone_t *)arg1;
3473 cur = uma_zone_get_cur(zone);
3474 return (sysctl_handle_int(oidp, &cur, 0, req));
3479 uma_dbg_getslab(uma_zone_t zone, void *item)
3485 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
3486 if (zone->uz_flags & UMA_ZONE_VTOSLAB) {
3487 slab = vtoslab((vm_offset_t)mem);
3490 * It is safe to return the slab here even though the
3491 * zone is unlocked because the item's allocation state
3492 * essentially holds a reference.
3495 keg = LIST_FIRST(&zone->uz_kegs)->kl_keg;
3496 if (keg->uk_flags & UMA_ZONE_HASH)
3497 slab = hash_sfind(&keg->uk_hash, mem);
3499 slab = (uma_slab_t)(mem + keg->uk_pgoff);
3507 * Set up the slab's freei data such that uma_dbg_free can function.
3511 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
3516 if (zone_first_keg(zone) == NULL)
3519 slab = uma_dbg_getslab(zone, item);
3521 panic("uma: item %p did not belong to zone %s\n",
3522 item, zone->uz_name);
3525 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3527 if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3528 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
3529 item, zone, zone->uz_name, slab, freei);
3530 BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3536 * Verifies freed addresses. Checks for alignment, valid slab membership
3537 * and duplicate frees.
3541 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
3546 if (zone_first_keg(zone) == NULL)
3549 slab = uma_dbg_getslab(zone, item);
3551 panic("uma: Freed item %p did not belong to zone %s\n",
3552 item, zone->uz_name);
3555 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3557 if (freei >= keg->uk_ipers)
3558 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
3559 item, zone, zone->uz_name, slab, freei);
3561 if (((freei * keg->uk_rsize) + slab->us_data) != item)
3562 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
3563 item, zone, zone->uz_name, slab, freei);
3565 if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3566 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
3567 item, zone, zone->uz_name, slab, freei);
3569 BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3571 #endif /* INVARIANTS */
3574 DB_SHOW_COMMAND(uma, db_show_uma)
3576 uint64_t allocs, frees, sleeps;
3577 uma_bucket_t bucket;
3582 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3583 "Free", "Requests", "Sleeps", "Bucket");
3584 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3585 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3586 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3587 allocs = z->uz_allocs;
3588 frees = z->uz_frees;
3589 sleeps = z->uz_sleeps;
3592 uma_zone_sumstat(z, &cachefree, &allocs,
3594 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3595 (LIST_FIRST(&kz->uk_zones) != z)))
3596 cachefree += kz->uk_free;
3597 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3598 cachefree += bucket->ub_cnt;
3599 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3600 z->uz_name, (uintmax_t)kz->uk_size,
3601 (intmax_t)(allocs - frees), cachefree,
3602 (uintmax_t)allocs, sleeps, z->uz_count);
3609 DB_SHOW_COMMAND(umacache, db_show_umacache)
3611 uint64_t allocs, frees;
3612 uma_bucket_t bucket;
3616 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3617 "Requests", "Bucket");
3618 LIST_FOREACH(z, &uma_cachezones, uz_link) {
3619 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3620 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3621 cachefree += bucket->ub_cnt;
3622 db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3623 z->uz_name, (uintmax_t)z->uz_size,
3624 (intmax_t)(allocs - frees), cachefree,
3625 (uintmax_t)allocs, z->uz_count);