2 * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org>
3 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
4 * Copyright (c) 2004-2006 Robert N. M. Watson
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice unmodified, this list of conditions, and the following
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30 * uma_core.c Implementation of the Universal Memory allocator
32 * This allocator is intended to replace the multitude of similar object caches
33 * in the standard FreeBSD kernel. The intent is to be flexible as well as
34 * effecient. A primary design goal is to return unused memory to the rest of
35 * the system. This will make the system as a whole more flexible due to the
36 * ability to move memory to subsystems which most need it instead of leaving
37 * pools of reserved memory unused.
39 * The basic ideas stem from similar slab/zone based allocators whose algorithms
46 * - Improve memory usage for large allocations
47 * - Investigate cache size adjustments
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
53 /* I should really use ktr.. */
56 #define UMA_DEBUG_ALLOC 1
57 #define UMA_DEBUG_ALLOC_1 1
61 #include "opt_param.h"
64 #include <sys/param.h>
65 #include <sys/systm.h>
66 #include <sys/bitset.h>
67 #include <sys/kernel.h>
68 #include <sys/types.h>
69 #include <sys/queue.h>
70 #include <sys/malloc.h>
73 #include <sys/sysctl.h>
74 #include <sys/mutex.h>
76 #include <sys/random.h>
77 #include <sys/rwlock.h>
79 #include <sys/sched.h>
81 #include <sys/vmmeter.h>
84 #include <vm/vm_object.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_pageout.h>
87 #include <vm/vm_param.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_kern.h>
90 #include <vm/vm_extern.h>
92 #include <vm/uma_int.h>
93 #include <vm/uma_dbg.h>
98 #include <vm/memguard.h>
102 * This is the zone and keg from which all zones are spawned. The idea is that
103 * even the zone & keg heads are allocated from the allocator, so we use the
104 * bss section to bootstrap us.
106 static struct uma_keg masterkeg;
107 static struct uma_zone masterzone_k;
108 static struct uma_zone masterzone_z;
109 static uma_zone_t kegs = &masterzone_k;
110 static uma_zone_t zones = &masterzone_z;
112 /* This is the zone from which all of uma_slab_t's are allocated. */
113 static uma_zone_t slabzone;
114 static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
117 * The initial hash tables come out of this zone so they can be allocated
118 * prior to malloc coming up.
120 static uma_zone_t hashzone;
122 /* The boot-time adjusted value for cache line alignment. */
123 int uma_align_cache = 64 - 1;
125 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
128 * Are we allowed to allocate buckets?
130 static int bucketdisable = 1;
132 /* Linked list of all kegs in the system */
133 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
135 /* Linked list of all cache-only zones in the system */
136 static LIST_HEAD(,uma_zone) uma_cachezones =
137 LIST_HEAD_INITIALIZER(uma_cachezones);
139 /* This RW lock protects the keg list */
140 static struct rwlock_padalign uma_rwlock;
142 /* Linked list of boot time pages */
143 static LIST_HEAD(,uma_slab) uma_boot_pages =
144 LIST_HEAD_INITIALIZER(uma_boot_pages);
146 /* This mutex protects the boot time pages list */
147 static struct mtx_padalign uma_boot_pages_mtx;
149 static struct sx uma_drain_lock;
151 /* Is the VM done starting up? */
152 static int booted = 0;
153 #define UMA_STARTUP 1
154 #define UMA_STARTUP2 2
157 * Only mbuf clusters use ref zones. Just provide enough references
158 * to support the one user. New code should not use the ref facility.
160 static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES;
163 * This is the handle used to schedule events that need to happen
164 * outside of the allocation fast path.
166 static struct callout uma_callout;
167 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
170 * This structure is passed as the zone ctor arg so that I don't have to create
171 * a special allocation function just for zones.
173 struct uma_zctor_args {
188 struct uma_kctor_args {
197 struct uma_bucket_zone {
200 int ubz_entries; /* Number of items it can hold. */
201 int ubz_maxsize; /* Maximum allocation size per-item. */
205 * Compute the actual number of bucket entries to pack them in power
206 * of two sizes for more efficient space utilization.
208 #define BUCKET_SIZE(n) \
209 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
211 #define BUCKET_MAX BUCKET_SIZE(256)
213 struct uma_bucket_zone bucket_zones[] = {
214 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
215 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
216 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
217 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
218 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
219 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
220 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
221 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
222 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
227 * Flags and enumerations to be passed to internal functions.
229 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
233 static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
234 static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
235 static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
236 static void page_free(void *, vm_size_t, uint8_t);
237 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
238 static void cache_drain(uma_zone_t);
239 static void bucket_drain(uma_zone_t, uma_bucket_t);
240 static void bucket_cache_drain(uma_zone_t zone);
241 static int keg_ctor(void *, int, void *, int);
242 static void keg_dtor(void *, int, void *);
243 static int zone_ctor(void *, int, void *, int);
244 static void zone_dtor(void *, int, void *);
245 static int zero_init(void *, int, int);
246 static void keg_small_init(uma_keg_t keg);
247 static void keg_large_init(uma_keg_t keg);
248 static void zone_foreach(void (*zfunc)(uma_zone_t));
249 static void zone_timeout(uma_zone_t zone);
250 static int hash_alloc(struct uma_hash *);
251 static int hash_expand(struct uma_hash *, struct uma_hash *);
252 static void hash_free(struct uma_hash *hash);
253 static void uma_timeout(void *);
254 static void uma_startup3(void);
255 static void *zone_alloc_item(uma_zone_t, void *, int);
256 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
257 static void bucket_enable(void);
258 static void bucket_init(void);
259 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
260 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
261 static void bucket_zone_drain(void);
262 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
263 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
264 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
265 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
266 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
267 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
268 uma_fini fini, int align, uint32_t flags);
269 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
270 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
271 static void uma_zero_item(void *item, uma_zone_t zone);
273 void uma_print_zone(uma_zone_t);
274 void uma_print_stats(void);
275 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
276 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
278 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
280 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
281 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
283 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
284 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
286 static int zone_warnings = 1;
287 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
288 "Warn when UMA zones becomes full");
291 * This routine checks to see whether or not it's safe to enable buckets.
296 bucketdisable = vm_page_count_min();
300 * Initialize bucket_zones, the array of zones of buckets of various sizes.
302 * For each zone, calculate the memory required for each bucket, consisting
303 * of the header and an array of pointers.
308 struct uma_bucket_zone *ubz;
312 for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
313 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
314 size += sizeof(void *) * ubz->ubz_entries;
315 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
316 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
317 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
322 * Given a desired number of entries for a bucket, return the zone from which
323 * to allocate the bucket.
325 static struct uma_bucket_zone *
326 bucket_zone_lookup(int entries)
328 struct uma_bucket_zone *ubz;
330 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
331 if (ubz->ubz_entries >= entries)
338 bucket_select(int size)
340 struct uma_bucket_zone *ubz;
342 ubz = &bucket_zones[0];
343 if (size > ubz->ubz_maxsize)
344 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
346 for (; ubz->ubz_entries != 0; ubz++)
347 if (ubz->ubz_maxsize < size)
350 return (ubz->ubz_entries);
354 bucket_alloc(uma_zone_t zone, void *udata, int flags)
356 struct uma_bucket_zone *ubz;
360 * This is to stop us from allocating per cpu buckets while we're
361 * running out of vm.boot_pages. Otherwise, we would exhaust the
362 * boot pages. This also prevents us from allocating buckets in
363 * low memory situations.
368 * To limit bucket recursion we store the original zone flags
369 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
370 * NOVM flag to persist even through deep recursions. We also
371 * store ZFLAG_BUCKET once we have recursed attempting to allocate
372 * a bucket for a bucket zone so we do not allow infinite bucket
373 * recursion. This cookie will even persist to frees of unused
374 * buckets via the allocation path or bucket allocations in the
377 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
378 udata = (void *)(uintptr_t)zone->uz_flags;
380 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
382 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
384 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
386 ubz = bucket_zone_lookup(zone->uz_count);
387 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
389 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
392 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
395 bucket->ub_entries = ubz->ubz_entries;
402 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
404 struct uma_bucket_zone *ubz;
406 KASSERT(bucket->ub_cnt == 0,
407 ("bucket_free: Freeing a non free bucket."));
408 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
409 udata = (void *)(uintptr_t)zone->uz_flags;
410 ubz = bucket_zone_lookup(bucket->ub_entries);
411 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
415 bucket_zone_drain(void)
417 struct uma_bucket_zone *ubz;
419 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
420 zone_drain(ubz->ubz_zone);
424 zone_log_warning(uma_zone_t zone)
426 static const struct timeval warninterval = { 300, 0 };
428 if (!zone_warnings || zone->uz_warning == NULL)
431 if (ratecheck(&zone->uz_ratecheck, &warninterval))
432 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
436 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
440 LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
441 kegfn(klink->kl_keg);
445 * Routine called by timeout which is used to fire off some time interval
446 * based calculations. (stats, hash size, etc.)
455 uma_timeout(void *unused)
458 zone_foreach(zone_timeout);
460 /* Reschedule this event */
461 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
465 * Routine to perform timeout driven calculations. This expands the
466 * hashes and does per cpu statistics aggregation.
471 keg_timeout(uma_keg_t keg)
476 * Expand the keg hash table.
478 * This is done if the number of slabs is larger than the hash size.
479 * What I'm trying to do here is completely reduce collisions. This
480 * may be a little aggressive. Should I allow for two collisions max?
482 if (keg->uk_flags & UMA_ZONE_HASH &&
483 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
484 struct uma_hash newhash;
485 struct uma_hash oldhash;
489 * This is so involved because allocating and freeing
490 * while the keg lock is held will lead to deadlock.
491 * I have to do everything in stages and check for
494 newhash = keg->uk_hash;
496 ret = hash_alloc(&newhash);
499 if (hash_expand(&keg->uk_hash, &newhash)) {
500 oldhash = keg->uk_hash;
501 keg->uk_hash = newhash;
514 zone_timeout(uma_zone_t zone)
517 zone_foreach_keg(zone, &keg_timeout);
521 * Allocate and zero fill the next sized hash table from the appropriate
525 * hash A new hash structure with the old hash size in uh_hashsize
528 * 1 on sucess and 0 on failure.
531 hash_alloc(struct uma_hash *hash)
536 oldsize = hash->uh_hashsize;
538 /* We're just going to go to a power of two greater */
540 hash->uh_hashsize = oldsize * 2;
541 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
542 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
543 M_UMAHASH, M_NOWAIT);
545 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
546 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
548 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
550 if (hash->uh_slab_hash) {
551 bzero(hash->uh_slab_hash, alloc);
552 hash->uh_hashmask = hash->uh_hashsize - 1;
560 * Expands the hash table for HASH zones. This is done from zone_timeout
561 * to reduce collisions. This must not be done in the regular allocation
562 * path, otherwise, we can recurse on the vm while allocating pages.
565 * oldhash The hash you want to expand
566 * newhash The hash structure for the new table
574 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
580 if (!newhash->uh_slab_hash)
583 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
587 * I need to investigate hash algorithms for resizing without a
591 for (i = 0; i < oldhash->uh_hashsize; i++)
592 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
593 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
594 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
595 hval = UMA_HASH(newhash, slab->us_data);
596 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
604 * Free the hash bucket to the appropriate backing store.
607 * slab_hash The hash bucket we're freeing
608 * hashsize The number of entries in that hash bucket
614 hash_free(struct uma_hash *hash)
616 if (hash->uh_slab_hash == NULL)
618 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
619 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
621 free(hash->uh_slab_hash, M_UMAHASH);
625 * Frees all outstanding items in a bucket
628 * zone The zone to free to, must be unlocked.
629 * bucket The free/alloc bucket with items, cpu queue must be locked.
636 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
644 for (i = 0; i < bucket->ub_cnt; i++)
645 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
646 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
651 * Drains the per cpu caches for a zone.
653 * NOTE: This may only be called while the zone is being turn down, and not
654 * during normal operation. This is necessary in order that we do not have
655 * to migrate CPUs to drain the per-CPU caches.
658 * zone The zone to drain, must be unlocked.
664 cache_drain(uma_zone_t zone)
670 * XXX: It is safe to not lock the per-CPU caches, because we're
671 * tearing down the zone anyway. I.e., there will be no further use
672 * of the caches at this point.
674 * XXX: It would good to be able to assert that the zone is being
675 * torn down to prevent improper use of cache_drain().
677 * XXX: We lock the zone before passing into bucket_cache_drain() as
678 * it is used elsewhere. Should the tear-down path be made special
679 * there in some form?
682 cache = &zone->uz_cpu[cpu];
683 bucket_drain(zone, cache->uc_allocbucket);
684 bucket_drain(zone, cache->uc_freebucket);
685 if (cache->uc_allocbucket != NULL)
686 bucket_free(zone, cache->uc_allocbucket, NULL);
687 if (cache->uc_freebucket != NULL)
688 bucket_free(zone, cache->uc_freebucket, NULL);
689 cache->uc_allocbucket = cache->uc_freebucket = NULL;
692 bucket_cache_drain(zone);
697 cache_shrink(uma_zone_t zone)
700 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
704 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
709 cache_drain_safe_cpu(uma_zone_t zone)
714 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
720 cache = &zone->uz_cpu[curcpu];
721 if (cache->uc_allocbucket) {
722 if (cache->uc_allocbucket->ub_cnt != 0)
723 LIST_INSERT_HEAD(&zone->uz_buckets,
724 cache->uc_allocbucket, ub_link);
726 b1 = cache->uc_allocbucket;
727 cache->uc_allocbucket = NULL;
729 if (cache->uc_freebucket) {
730 if (cache->uc_freebucket->ub_cnt != 0)
731 LIST_INSERT_HEAD(&zone->uz_buckets,
732 cache->uc_freebucket, ub_link);
734 b2 = cache->uc_freebucket;
735 cache->uc_freebucket = NULL;
740 bucket_free(zone, b1, NULL);
742 bucket_free(zone, b2, NULL);
746 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
747 * This is an expensive call because it needs to bind to all CPUs
748 * one by one and enter a critical section on each of them in order
749 * to safely access their cache buckets.
750 * Zone lock must not be held on call this function.
753 cache_drain_safe(uma_zone_t zone)
758 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
763 zone_foreach(cache_shrink);
766 thread_lock(curthread);
767 sched_bind(curthread, cpu);
768 thread_unlock(curthread);
771 cache_drain_safe_cpu(zone);
773 zone_foreach(cache_drain_safe_cpu);
775 thread_lock(curthread);
776 sched_unbind(curthread);
777 thread_unlock(curthread);
781 * Drain the cached buckets from a zone. Expects a locked zone on entry.
784 bucket_cache_drain(uma_zone_t zone)
789 * Drain the bucket queues and free the buckets, we just keep two per
792 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
793 LIST_REMOVE(bucket, ub_link);
795 bucket_drain(zone, bucket);
796 bucket_free(zone, bucket, NULL);
801 * Shrink further bucket sizes. Price of single zone lock collision
802 * is probably lower then price of global cache drain.
804 if (zone->uz_count > zone->uz_count_min)
809 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
816 flags = slab->us_flags;
818 if (keg->uk_fini != NULL) {
819 for (i--; i > -1; i--)
820 keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
823 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
824 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
826 printf("%s: Returning %d bytes.\n", keg->uk_name,
827 PAGE_SIZE * keg->uk_ppera);
829 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
833 * Frees pages from a keg back to the system. This is done on demand from
834 * the pageout daemon.
839 keg_drain(uma_keg_t keg)
841 struct slabhead freeslabs = { 0 };
846 * We don't want to take pages from statically allocated kegs at this
849 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
853 printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
856 if (keg->uk_free == 0)
859 slab = LIST_FIRST(&keg->uk_free_slab);
861 n = LIST_NEXT(slab, us_link);
863 /* We have no where to free these to */
864 if (slab->us_flags & UMA_SLAB_BOOT) {
869 LIST_REMOVE(slab, us_link);
870 keg->uk_pages -= keg->uk_ppera;
871 keg->uk_free -= keg->uk_ipers;
873 if (keg->uk_flags & UMA_ZONE_HASH)
874 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
876 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
883 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
884 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
885 keg_free_slab(keg, slab, keg->uk_ipers);
890 zone_drain_wait(uma_zone_t zone, int waitok)
894 * Set draining to interlock with zone_dtor() so we can release our
895 * locks as we go. Only dtor() should do a WAITOK call since it
896 * is the only call that knows the structure will still be available
900 while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
901 if (waitok == M_NOWAIT)
903 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
905 zone->uz_flags |= UMA_ZFLAG_DRAINING;
906 bucket_cache_drain(zone);
909 * The DRAINING flag protects us from being freed while
910 * we're running. Normally the uma_rwlock would protect us but we
911 * must be able to release and acquire the right lock for each keg.
913 zone_foreach_keg(zone, &keg_drain);
915 zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
922 zone_drain(uma_zone_t zone)
925 zone_drain_wait(zone, M_NOWAIT);
929 * Allocate a new slab for a keg. This does not insert the slab onto a list.
932 * wait Shall we wait?
935 * The slab that was allocated or NULL if there is no memory and the
936 * caller specified M_NOWAIT.
939 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
941 uma_slabrefcnt_t slabref;
948 mtx_assert(&keg->uk_lock, MA_OWNED);
953 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name);
955 allocf = keg->uk_allocf;
958 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
959 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
965 * This reproduces the old vm_zone behavior of zero filling pages the
966 * first time they are added to a zone.
968 * Malloced items are zeroed in uma_zalloc.
971 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
976 if (keg->uk_flags & UMA_ZONE_NODUMP)
979 /* zone is passed for legacy reasons. */
980 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
982 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
983 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
988 /* Point the slab into the allocated memory */
989 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
990 slab = (uma_slab_t )(mem + keg->uk_pgoff);
992 if (keg->uk_flags & UMA_ZONE_VTOSLAB)
993 for (i = 0; i < keg->uk_ppera; i++)
994 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
998 slab->us_freecount = keg->uk_ipers;
999 slab->us_flags = flags;
1000 BIT_FILL(SLAB_SETSIZE, &slab->us_free);
1002 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1004 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1005 slabref = (uma_slabrefcnt_t)slab;
1006 for (i = 0; i < keg->uk_ipers; i++)
1007 slabref->us_refcnt[i] = 0;
1010 if (keg->uk_init != NULL) {
1011 for (i = 0; i < keg->uk_ipers; i++)
1012 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1013 keg->uk_size, wait) != 0)
1015 if (i != keg->uk_ipers) {
1016 keg_free_slab(keg, slab, i);
1025 if (keg->uk_flags & UMA_ZONE_HASH)
1026 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1028 keg->uk_pages += keg->uk_ppera;
1029 keg->uk_free += keg->uk_ipers;
1036 * This function is intended to be used early on in place of page_alloc() so
1037 * that we may use the boot time page cache to satisfy allocations before
1041 startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1045 int pages, check_pages;
1047 keg = zone_first_keg(zone);
1048 pages = howmany(bytes, PAGE_SIZE);
1049 check_pages = pages - 1;
1050 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1053 * Check our small startup cache to see if it has pages remaining.
1055 mtx_lock(&uma_boot_pages_mtx);
1057 /* First check if we have enough room. */
1058 tmps = LIST_FIRST(&uma_boot_pages);
1059 while (tmps != NULL && check_pages-- > 0)
1060 tmps = LIST_NEXT(tmps, us_link);
1063 * It's ok to lose tmps references. The last one will
1064 * have tmps->us_data pointing to the start address of
1065 * "pages" contiguous pages of memory.
1067 while (pages-- > 0) {
1068 tmps = LIST_FIRST(&uma_boot_pages);
1069 LIST_REMOVE(tmps, us_link);
1071 mtx_unlock(&uma_boot_pages_mtx);
1072 *pflag = tmps->us_flags;
1073 return (tmps->us_data);
1075 mtx_unlock(&uma_boot_pages_mtx);
1076 if (booted < UMA_STARTUP2)
1077 panic("UMA: Increase vm.boot_pages");
1079 * Now that we've booted reset these users to their real allocator.
1081 #ifdef UMA_MD_SMALL_ALLOC
1082 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1084 keg->uk_allocf = page_alloc;
1086 return keg->uk_allocf(zone, bytes, pflag, wait);
1090 * Allocates a number of pages from the system
1093 * bytes The number of bytes requested
1094 * wait Shall we wait?
1097 * A pointer to the alloced memory or possibly
1098 * NULL if M_NOWAIT is set.
1101 page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1103 void *p; /* Returned page */
1105 *pflag = UMA_SLAB_KMEM;
1106 p = (void *) kmem_malloc(kmem_arena, bytes, wait);
1112 * Allocates a number of pages from within an object
1115 * bytes The number of bytes requested
1116 * wait Shall we wait?
1119 * A pointer to the alloced memory or possibly
1120 * NULL if M_NOWAIT is set.
1123 noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait)
1125 TAILQ_HEAD(, vm_page) alloctail;
1127 vm_offset_t retkva, zkva;
1128 vm_page_t p, p_next;
1131 TAILQ_INIT(&alloctail);
1132 keg = zone_first_keg(zone);
1134 npages = howmany(bytes, PAGE_SIZE);
1135 while (npages > 0) {
1136 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1137 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1140 * Since the page does not belong to an object, its
1143 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1147 if (wait & M_WAITOK) {
1153 * Page allocation failed, free intermediate pages and
1156 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1157 vm_page_unwire(p, PQ_INACTIVE);
1162 *flags = UMA_SLAB_PRIV;
1163 zkva = keg->uk_kva +
1164 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1166 TAILQ_FOREACH(p, &alloctail, listq) {
1167 pmap_qenter(zkva, &p, 1);
1171 return ((void *)retkva);
1175 * Frees a number of pages to the system
1178 * mem A pointer to the memory to be freed
1179 * size The size of the memory being freed
1180 * flags The original p->us_flags field
1186 page_free(void *mem, vm_size_t size, uint8_t flags)
1190 if (flags & UMA_SLAB_KMEM)
1192 else if (flags & UMA_SLAB_KERNEL)
1193 vmem = kernel_arena;
1195 panic("UMA: page_free used with invalid flags %d", flags);
1197 kmem_free(vmem, (vm_offset_t)mem, size);
1201 * Zero fill initializer
1203 * Arguments/Returns follow uma_init specifications
1206 zero_init(void *mem, int size, int flags)
1213 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1216 * keg The zone we should initialize
1222 keg_small_init(uma_keg_t keg)
1229 if (keg->uk_flags & UMA_ZONE_PCPU) {
1230 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1232 keg->uk_slabsize = sizeof(struct pcpu);
1233 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1236 keg->uk_slabsize = UMA_SLAB_SIZE;
1241 * Calculate the size of each allocation (rsize) according to
1242 * alignment. If the requested size is smaller than we have
1243 * allocation bits for we round it up.
1245 rsize = keg->uk_size;
1246 if (rsize < keg->uk_slabsize / SLAB_SETSIZE)
1247 rsize = keg->uk_slabsize / SLAB_SETSIZE;
1248 if (rsize & keg->uk_align)
1249 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1250 keg->uk_rsize = rsize;
1252 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1253 keg->uk_rsize < sizeof(struct pcpu),
1254 ("%s: size %u too large", __func__, keg->uk_rsize));
1256 if (keg->uk_flags & UMA_ZONE_REFCNT)
1257 rsize += sizeof(uint32_t);
1259 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1262 shsize = sizeof(struct uma_slab);
1264 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1265 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1266 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1268 memused = keg->uk_ipers * rsize + shsize;
1269 wastedspace = keg->uk_slabsize - memused;
1272 * We can't do OFFPAGE if we're internal or if we've been
1273 * asked to not go to the VM for buckets. If we do this we
1274 * may end up going to the VM for slabs which we do not
1275 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1276 * of UMA_ZONE_VM, which clearly forbids it.
1278 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1279 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1283 * See if using an OFFPAGE slab will limit our waste. Only do
1284 * this if it permits more items per-slab.
1286 * XXX We could try growing slabsize to limit max waste as well.
1287 * Historically this was not done because the VM could not
1288 * efficiently handle contiguous allocations.
1290 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1291 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1292 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1293 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1294 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1296 printf("UMA decided we need offpage slab headers for "
1297 "keg: %s, calculated wastedspace = %d, "
1298 "maximum wasted space allowed = %d, "
1299 "calculated ipers = %d, "
1300 "new wasted space = %d\n", keg->uk_name, wastedspace,
1301 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1302 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1304 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1307 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1308 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1309 keg->uk_flags |= UMA_ZONE_HASH;
1313 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1314 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1318 * keg The keg we should initialize
1324 keg_large_init(uma_keg_t keg)
1328 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1329 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1330 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1331 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1332 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1334 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1335 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1337 keg->uk_rsize = keg->uk_size;
1339 /* We can't do OFFPAGE if we're internal, bail out here. */
1340 if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1343 /* Check whether we have enough space to not do OFFPAGE. */
1344 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1345 shsize = sizeof(struct uma_slab);
1346 if (keg->uk_flags & UMA_ZONE_REFCNT)
1347 shsize += keg->uk_ipers * sizeof(uint32_t);
1348 if (shsize & UMA_ALIGN_PTR)
1349 shsize = (shsize & ~UMA_ALIGN_PTR) +
1350 (UMA_ALIGN_PTR + 1);
1352 if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize)
1353 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1356 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1357 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1358 keg->uk_flags |= UMA_ZONE_HASH;
1362 keg_cachespread_init(uma_keg_t keg)
1369 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1370 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1372 alignsize = keg->uk_align + 1;
1373 rsize = keg->uk_size;
1375 * We want one item to start on every align boundary in a page. To
1376 * do this we will span pages. We will also extend the item by the
1377 * size of align if it is an even multiple of align. Otherwise, it
1378 * would fall on the same boundary every time.
1380 if (rsize & keg->uk_align)
1381 rsize = (rsize & ~keg->uk_align) + alignsize;
1382 if ((rsize & alignsize) == 0)
1384 trailer = rsize - keg->uk_size;
1385 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1386 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1387 keg->uk_rsize = rsize;
1388 keg->uk_ppera = pages;
1389 keg->uk_slabsize = UMA_SLAB_SIZE;
1390 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1391 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1392 KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1393 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1398 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1399 * the keg onto the global keg list.
1401 * Arguments/Returns follow uma_ctor specifications
1402 * udata Actually uma_kctor_args
1405 keg_ctor(void *mem, int size, void *udata, int flags)
1407 struct uma_kctor_args *arg = udata;
1408 uma_keg_t keg = mem;
1412 keg->uk_size = arg->size;
1413 keg->uk_init = arg->uminit;
1414 keg->uk_fini = arg->fini;
1415 keg->uk_align = arg->align;
1417 keg->uk_reserve = 0;
1419 keg->uk_flags = arg->flags;
1420 keg->uk_allocf = page_alloc;
1421 keg->uk_freef = page_free;
1422 keg->uk_slabzone = NULL;
1425 * The master zone is passed to us at keg-creation time.
1428 keg->uk_name = zone->uz_name;
1430 if (arg->flags & UMA_ZONE_VM)
1431 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1433 if (arg->flags & UMA_ZONE_ZINIT)
1434 keg->uk_init = zero_init;
1436 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1437 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1439 if (arg->flags & UMA_ZONE_PCPU)
1441 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1443 keg->uk_flags &= ~UMA_ZONE_PCPU;
1446 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1447 keg_cachespread_init(keg);
1448 } else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1450 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) -
1452 keg_large_init(keg);
1454 keg_small_init(keg);
1456 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1457 keg_large_init(keg);
1459 keg_small_init(keg);
1462 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1463 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1464 if (keg->uk_ipers > uma_max_ipers_ref)
1465 panic("Too many ref items per zone: %d > %d\n",
1466 keg->uk_ipers, uma_max_ipers_ref);
1467 keg->uk_slabzone = slabrefzone;
1469 keg->uk_slabzone = slabzone;
1473 * If we haven't booted yet we need allocations to go through the
1474 * startup cache until the vm is ready.
1476 if (keg->uk_ppera == 1) {
1477 #ifdef UMA_MD_SMALL_ALLOC
1478 keg->uk_allocf = uma_small_alloc;
1479 keg->uk_freef = uma_small_free;
1481 if (booted < UMA_STARTUP)
1482 keg->uk_allocf = startup_alloc;
1484 if (booted < UMA_STARTUP2)
1485 keg->uk_allocf = startup_alloc;
1487 } else if (booted < UMA_STARTUP2 &&
1488 (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1489 keg->uk_allocf = startup_alloc;
1492 * Initialize keg's lock
1494 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1497 * If we're putting the slab header in the actual page we need to
1498 * figure out where in each page it goes. This calculates a right
1499 * justified offset into the memory on an ALIGN_PTR boundary.
1501 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1504 /* Size of the slab struct and free list */
1505 totsize = sizeof(struct uma_slab);
1507 /* Size of the reference counts. */
1508 if (keg->uk_flags & UMA_ZONE_REFCNT)
1509 totsize += keg->uk_ipers * sizeof(uint32_t);
1511 if (totsize & UMA_ALIGN_PTR)
1512 totsize = (totsize & ~UMA_ALIGN_PTR) +
1513 (UMA_ALIGN_PTR + 1);
1514 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1517 * The only way the following is possible is if with our
1518 * UMA_ALIGN_PTR adjustments we are now bigger than
1519 * UMA_SLAB_SIZE. I haven't checked whether this is
1520 * mathematically possible for all cases, so we make
1523 totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1524 if (keg->uk_flags & UMA_ZONE_REFCNT)
1525 totsize += keg->uk_ipers * sizeof(uint32_t);
1526 if (totsize > PAGE_SIZE * keg->uk_ppera) {
1527 printf("zone %s ipers %d rsize %d size %d\n",
1528 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1530 panic("UMA slab won't fit.");
1534 if (keg->uk_flags & UMA_ZONE_HASH)
1535 hash_alloc(&keg->uk_hash);
1538 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1539 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1540 keg->uk_ipers, keg->uk_ppera,
1541 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1544 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1546 rw_wlock(&uma_rwlock);
1547 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1548 rw_wunlock(&uma_rwlock);
1553 * Zone header ctor. This initializes all fields, locks, etc.
1555 * Arguments/Returns follow uma_ctor specifications
1556 * udata Actually uma_zctor_args
1559 zone_ctor(void *mem, int size, void *udata, int flags)
1561 struct uma_zctor_args *arg = udata;
1562 uma_zone_t zone = mem;
1567 zone->uz_name = arg->name;
1568 zone->uz_ctor = arg->ctor;
1569 zone->uz_dtor = arg->dtor;
1570 zone->uz_slab = zone_fetch_slab;
1571 zone->uz_init = NULL;
1572 zone->uz_fini = NULL;
1573 zone->uz_allocs = 0;
1576 zone->uz_sleeps = 0;
1578 zone->uz_count_min = 0;
1580 zone->uz_warning = NULL;
1581 timevalclear(&zone->uz_ratecheck);
1584 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1587 * This is a pure cache zone, no kegs.
1590 if (arg->flags & UMA_ZONE_VM)
1591 arg->flags |= UMA_ZFLAG_CACHEONLY;
1592 zone->uz_flags = arg->flags;
1593 zone->uz_size = arg->size;
1594 zone->uz_import = arg->import;
1595 zone->uz_release = arg->release;
1596 zone->uz_arg = arg->arg;
1597 zone->uz_lockptr = &zone->uz_lock;
1598 rw_wlock(&uma_rwlock);
1599 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1600 rw_wunlock(&uma_rwlock);
1605 * Use the regular zone/keg/slab allocator.
1607 zone->uz_import = (uma_import)zone_import;
1608 zone->uz_release = (uma_release)zone_release;
1609 zone->uz_arg = zone;
1611 if (arg->flags & UMA_ZONE_SECONDARY) {
1612 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1613 zone->uz_init = arg->uminit;
1614 zone->uz_fini = arg->fini;
1615 zone->uz_lockptr = &keg->uk_lock;
1616 zone->uz_flags |= UMA_ZONE_SECONDARY;
1617 rw_wlock(&uma_rwlock);
1619 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1620 if (LIST_NEXT(z, uz_link) == NULL) {
1621 LIST_INSERT_AFTER(z, zone, uz_link);
1626 rw_wunlock(&uma_rwlock);
1627 } else if (keg == NULL) {
1628 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1629 arg->align, arg->flags)) == NULL)
1632 struct uma_kctor_args karg;
1635 /* We should only be here from uma_startup() */
1636 karg.size = arg->size;
1637 karg.uminit = arg->uminit;
1638 karg.fini = arg->fini;
1639 karg.align = arg->align;
1640 karg.flags = arg->flags;
1642 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1649 * Link in the first keg.
1651 zone->uz_klink.kl_keg = keg;
1652 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1653 zone->uz_lockptr = &keg->uk_lock;
1654 zone->uz_size = keg->uk_size;
1655 zone->uz_flags |= (keg->uk_flags &
1656 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1659 * Some internal zones don't have room allocated for the per cpu
1660 * caches. If we're internal, bail out here.
1662 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1663 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1664 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1669 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1670 zone->uz_count = bucket_select(zone->uz_size);
1672 zone->uz_count = BUCKET_MAX;
1673 zone->uz_count_min = zone->uz_count;
1679 * Keg header dtor. This frees all data, destroys locks, frees the hash
1680 * table and removes the keg from the global list.
1682 * Arguments/Returns follow uma_dtor specifications
1686 keg_dtor(void *arg, int size, void *udata)
1690 keg = (uma_keg_t)arg;
1692 if (keg->uk_free != 0) {
1693 printf("Freed UMA keg (%s) was not empty (%d items). "
1694 " Lost %d pages of memory.\n",
1695 keg->uk_name ? keg->uk_name : "",
1696 keg->uk_free, keg->uk_pages);
1700 hash_free(&keg->uk_hash);
1708 * Arguments/Returns follow uma_dtor specifications
1712 zone_dtor(void *arg, int size, void *udata)
1718 zone = (uma_zone_t)arg;
1719 keg = zone_first_keg(zone);
1721 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1724 rw_wlock(&uma_rwlock);
1725 LIST_REMOVE(zone, uz_link);
1726 rw_wunlock(&uma_rwlock);
1728 * XXX there are some races here where
1729 * the zone can be drained but zone lock
1730 * released and then refilled before we
1731 * remove it... we dont care for now
1733 zone_drain_wait(zone, M_WAITOK);
1735 * Unlink all of our kegs.
1737 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1738 klink->kl_keg = NULL;
1739 LIST_REMOVE(klink, kl_link);
1740 if (klink == &zone->uz_klink)
1742 free(klink, M_TEMP);
1745 * We only destroy kegs from non secondary zones.
1747 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1748 rw_wlock(&uma_rwlock);
1749 LIST_REMOVE(keg, uk_link);
1750 rw_wunlock(&uma_rwlock);
1751 zone_free_item(kegs, keg, NULL, SKIP_NONE);
1753 ZONE_LOCK_FINI(zone);
1757 * Traverses every zone in the system and calls a callback
1760 * zfunc A pointer to a function which accepts a zone
1767 zone_foreach(void (*zfunc)(uma_zone_t))
1772 rw_rlock(&uma_rwlock);
1773 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1774 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1777 rw_runlock(&uma_rwlock);
1780 /* Public functions */
1783 uma_startup(void *bootmem, int boot_pages)
1785 struct uma_zctor_args args;
1791 printf("Creating uma keg headers zone and keg.\n");
1793 rw_init(&uma_rwlock, "UMA lock");
1795 /* "manually" create the initial zone */
1796 memset(&args, 0, sizeof(args));
1797 args.name = "UMA Kegs";
1798 args.size = sizeof(struct uma_keg);
1799 args.ctor = keg_ctor;
1800 args.dtor = keg_dtor;
1801 args.uminit = zero_init;
1803 args.keg = &masterkeg;
1804 args.align = 32 - 1;
1805 args.flags = UMA_ZFLAG_INTERNAL;
1806 /* The initial zone has no Per cpu queues so it's smaller */
1807 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1810 printf("Filling boot free list.\n");
1812 for (i = 0; i < boot_pages; i++) {
1813 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1814 slab->us_data = (uint8_t *)slab;
1815 slab->us_flags = UMA_SLAB_BOOT;
1816 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1818 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1821 printf("Creating uma zone headers zone and keg.\n");
1823 args.name = "UMA Zones";
1824 args.size = sizeof(struct uma_zone) +
1825 (sizeof(struct uma_cache) * (mp_maxid + 1));
1826 args.ctor = zone_ctor;
1827 args.dtor = zone_dtor;
1828 args.uminit = zero_init;
1831 args.align = 32 - 1;
1832 args.flags = UMA_ZFLAG_INTERNAL;
1833 /* The initial zone has no Per cpu queues so it's smaller */
1834 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1837 printf("Creating slab and hash zones.\n");
1840 /* Now make a zone for slab headers */
1841 slabzone = uma_zcreate("UMA Slabs",
1842 sizeof(struct uma_slab),
1843 NULL, NULL, NULL, NULL,
1844 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1847 * We also create a zone for the bigger slabs with reference
1848 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1850 slabsize = sizeof(struct uma_slab_refcnt);
1851 slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1852 slabrefzone = uma_zcreate("UMA RCntSlabs",
1854 NULL, NULL, NULL, NULL,
1856 UMA_ZFLAG_INTERNAL);
1858 hashzone = uma_zcreate("UMA Hash",
1859 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1860 NULL, NULL, NULL, NULL,
1861 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1865 booted = UMA_STARTUP;
1868 printf("UMA startup complete.\n");
1876 booted = UMA_STARTUP2;
1878 sx_init(&uma_drain_lock, "umadrain");
1880 printf("UMA startup2 complete.\n");
1885 * Initialize our callout handle
1893 printf("Starting callout.\n");
1895 callout_init(&uma_callout, CALLOUT_MPSAFE);
1896 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1898 printf("UMA startup3 complete.\n");
1903 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1904 int align, uint32_t flags)
1906 struct uma_kctor_args args;
1909 args.uminit = uminit;
1911 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1914 return (zone_alloc_item(kegs, &args, M_WAITOK));
1919 uma_set_align(int align)
1922 if (align != UMA_ALIGN_CACHE)
1923 uma_align_cache = align;
1928 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1929 uma_init uminit, uma_fini fini, int align, uint32_t flags)
1932 struct uma_zctor_args args;
1936 /* This stuff is essential for the zone ctor */
1937 memset(&args, 0, sizeof(args));
1942 args.uminit = uminit;
1948 if (booted < UMA_STARTUP2) {
1951 sx_slock(&uma_drain_lock);
1954 res = zone_alloc_item(zones, &args, M_WAITOK);
1956 sx_sunlock(&uma_drain_lock);
1962 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1963 uma_init zinit, uma_fini zfini, uma_zone_t master)
1965 struct uma_zctor_args args;
1970 keg = zone_first_keg(master);
1971 memset(&args, 0, sizeof(args));
1973 args.size = keg->uk_size;
1976 args.uminit = zinit;
1978 args.align = keg->uk_align;
1979 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1982 if (booted < UMA_STARTUP2) {
1985 sx_slock(&uma_drain_lock);
1988 /* XXX Attaches only one keg of potentially many. */
1989 res = zone_alloc_item(zones, &args, M_WAITOK);
1991 sx_sunlock(&uma_drain_lock);
1997 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1998 uma_init zinit, uma_fini zfini, uma_import zimport,
1999 uma_release zrelease, void *arg, int flags)
2001 struct uma_zctor_args args;
2003 memset(&args, 0, sizeof(args));
2008 args.uminit = zinit;
2010 args.import = zimport;
2011 args.release = zrelease;
2016 return (zone_alloc_item(zones, &args, M_WAITOK));
2020 zone_lock_pair(uma_zone_t a, uma_zone_t b)
2024 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
2027 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
2032 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2040 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2047 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2049 zone_lock_pair(zone, master);
2051 * zone must use vtoslab() to resolve objects and must already be
2054 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2055 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2060 * The new master must also use vtoslab().
2062 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2067 * Both must either be refcnt, or not be refcnt.
2069 if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
2070 (master->uz_flags & UMA_ZONE_REFCNT)) {
2075 * The underlying object must be the same size. rsize
2078 if (master->uz_size != zone->uz_size) {
2083 * Put it at the end of the list.
2085 klink->kl_keg = zone_first_keg(master);
2086 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2087 if (LIST_NEXT(kl, kl_link) == NULL) {
2088 LIST_INSERT_AFTER(kl, klink, kl_link);
2093 zone->uz_flags |= UMA_ZFLAG_MULTI;
2094 zone->uz_slab = zone_fetch_slab_multi;
2097 zone_unlock_pair(zone, master);
2099 free(klink, M_TEMP);
2107 uma_zdestroy(uma_zone_t zone)
2110 sx_slock(&uma_drain_lock);
2111 zone_free_item(zones, zone, NULL, SKIP_NONE);
2112 sx_sunlock(&uma_drain_lock);
2117 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2121 uma_bucket_t bucket;
2126 /* XXX: FIX!! Do not enable this in CURRENT!! MarkM */
2127 /* The entropy here is desirable, but the harvesting is expensive */
2128 random_harvest(&(zone->uz_name), sizeof(void *), 1, RANDOM_UMA_ALLOC);
2131 /* This is the fast path allocation */
2132 #ifdef UMA_DEBUG_ALLOC_1
2133 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2135 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2136 zone->uz_name, flags);
2138 if (flags & M_WAITOK) {
2139 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2140 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2142 #ifdef DEBUG_MEMGUARD
2143 if (memguard_cmp_zone(zone)) {
2144 item = memguard_alloc(zone->uz_size, flags);
2147 * Avoid conflict with the use-after-free
2148 * protecting infrastructure from INVARIANTS.
2150 if (zone->uz_init != NULL &&
2151 zone->uz_init != mtrash_init &&
2152 zone->uz_init(item, zone->uz_size, flags) != 0)
2154 if (zone->uz_ctor != NULL &&
2155 zone->uz_ctor != mtrash_ctor &&
2156 zone->uz_ctor(item, zone->uz_size, udata,
2158 zone->uz_fini(item, zone->uz_size);
2162 /* XXX: FIX!! Do not enable this in CURRENT!! MarkM */
2163 /* The entropy here is desirable, but the harvesting is expensive */
2164 random_harvest(&item, sizeof(void *), 1, RANDOM_UMA_ALLOC);
2168 /* This is unfortunate but should not be fatal. */
2172 * If possible, allocate from the per-CPU cache. There are two
2173 * requirements for safe access to the per-CPU cache: (1) the thread
2174 * accessing the cache must not be preempted or yield during access,
2175 * and (2) the thread must not migrate CPUs without switching which
2176 * cache it accesses. We rely on a critical section to prevent
2177 * preemption and migration. We release the critical section in
2178 * order to acquire the zone mutex if we are unable to allocate from
2179 * the current cache; when we re-acquire the critical section, we
2180 * must detect and handle migration if it has occurred.
2184 cache = &zone->uz_cpu[cpu];
2187 bucket = cache->uc_allocbucket;
2188 if (bucket != NULL && bucket->ub_cnt > 0) {
2190 item = bucket->ub_bucket[bucket->ub_cnt];
2192 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2194 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2197 if (zone->uz_ctor != NULL &&
2198 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2199 atomic_add_long(&zone->uz_fails, 1);
2200 zone_free_item(zone, item, udata, SKIP_DTOR);
2204 uma_dbg_alloc(zone, NULL, item);
2207 uma_zero_item(item, zone);
2209 /* XXX: FIX!! Do not enable this in CURRENT!! MarkM */
2210 /* The entropy here is desirable, but the harvesting is expensive */
2211 random_harvest(&item, sizeof(void *), 1, RANDOM_UMA_ALLOC);
2217 * We have run out of items in our alloc bucket.
2218 * See if we can switch with our free bucket.
2220 bucket = cache->uc_freebucket;
2221 if (bucket != NULL && bucket->ub_cnt > 0) {
2222 #ifdef UMA_DEBUG_ALLOC
2223 printf("uma_zalloc: Swapping empty with alloc.\n");
2225 cache->uc_freebucket = cache->uc_allocbucket;
2226 cache->uc_allocbucket = bucket;
2231 * Discard any empty allocation bucket while we hold no locks.
2233 bucket = cache->uc_allocbucket;
2234 cache->uc_allocbucket = NULL;
2237 bucket_free(zone, bucket, udata);
2239 /* Short-circuit for zones without buckets and low memory. */
2240 if (zone->uz_count == 0 || bucketdisable)
2244 * Attempt to retrieve the item from the per-CPU cache has failed, so
2245 * we must go back to the zone. This requires the zone lock, so we
2246 * must drop the critical section, then re-acquire it when we go back
2247 * to the cache. Since the critical section is released, we may be
2248 * preempted or migrate. As such, make sure not to maintain any
2249 * thread-local state specific to the cache from prior to releasing
2250 * the critical section.
2253 if (ZONE_TRYLOCK(zone) == 0) {
2254 /* Record contention to size the buckets. */
2260 cache = &zone->uz_cpu[cpu];
2263 * Since we have locked the zone we may as well send back our stats.
2265 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2266 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2267 cache->uc_allocs = 0;
2268 cache->uc_frees = 0;
2270 /* See if we lost the race to fill the cache. */
2271 if (cache->uc_allocbucket != NULL) {
2277 * Check the zone's cache of buckets.
2279 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2280 KASSERT(bucket->ub_cnt != 0,
2281 ("uma_zalloc_arg: Returning an empty bucket."));
2283 LIST_REMOVE(bucket, ub_link);
2284 cache->uc_allocbucket = bucket;
2288 /* We are no longer associated with this CPU. */
2292 * We bump the uz count when the cache size is insufficient to
2293 * handle the working set.
2295 if (lockfail && zone->uz_count < BUCKET_MAX)
2300 * Now lets just fill a bucket and put it on the free list. If that
2301 * works we'll restart the allocation from the begining and it
2302 * will use the just filled bucket.
2304 bucket = zone_alloc_bucket(zone, udata, flags);
2305 if (bucket != NULL) {
2309 cache = &zone->uz_cpu[cpu];
2311 * See if we lost the race or were migrated. Cache the
2312 * initialized bucket to make this less likely or claim
2313 * the memory directly.
2315 if (cache->uc_allocbucket == NULL)
2316 cache->uc_allocbucket = bucket;
2318 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2324 * We may not be able to get a bucket so return an actual item.
2327 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2331 item = zone_alloc_item(zone, udata, flags);
2334 /* XXX: FIX!! Do not enable this in CURRENT!! MarkM */
2335 /* The entropy here is desirable, but the harvesting is expensive */
2336 random_harvest(&item, sizeof(void *), 1, RANDOM_UMA_ALLOC);
2342 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2347 mtx_assert(&keg->uk_lock, MA_OWNED);
2350 if ((flags & M_USE_RESERVE) == 0)
2351 reserve = keg->uk_reserve;
2355 * Find a slab with some space. Prefer slabs that are partially
2356 * used over those that are totally full. This helps to reduce
2359 if (keg->uk_free > reserve) {
2360 if (!LIST_EMPTY(&keg->uk_part_slab)) {
2361 slab = LIST_FIRST(&keg->uk_part_slab);
2363 slab = LIST_FIRST(&keg->uk_free_slab);
2364 LIST_REMOVE(slab, us_link);
2365 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2368 MPASS(slab->us_keg == keg);
2373 * M_NOVM means don't ask at all!
2378 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2379 keg->uk_flags |= UMA_ZFLAG_FULL;
2381 * If this is not a multi-zone, set the FULL bit.
2382 * Otherwise slab_multi() takes care of it.
2384 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2385 zone->uz_flags |= UMA_ZFLAG_FULL;
2386 zone_log_warning(zone);
2388 if (flags & M_NOWAIT)
2391 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2394 slab = keg_alloc_slab(keg, zone, flags);
2396 * If we got a slab here it's safe to mark it partially used
2397 * and return. We assume that the caller is going to remove
2398 * at least one item.
2401 MPASS(slab->us_keg == keg);
2402 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2406 * We might not have been able to get a slab but another cpu
2407 * could have while we were unlocked. Check again before we
2416 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2421 keg = zone_first_keg(zone);
2426 slab = keg_fetch_slab(keg, zone, flags);
2429 if (flags & (M_NOWAIT | M_NOVM))
2437 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2438 * with the keg locked. On NULL no lock is held.
2440 * The last pointer is used to seed the search. It is not required.
2443 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2453 * Don't wait on the first pass. This will skip limit tests
2454 * as well. We don't want to block if we can find a provider
2457 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2459 * Use the last slab allocated as a hint for where to start
2463 slab = keg_fetch_slab(last, zone, flags);
2469 * Loop until we have a slab incase of transient failures
2470 * while M_WAITOK is specified. I'm not sure this is 100%
2471 * required but we've done it for so long now.
2477 * Search the available kegs for slabs. Be careful to hold the
2478 * correct lock while calling into the keg layer.
2480 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2481 keg = klink->kl_keg;
2483 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2484 slab = keg_fetch_slab(keg, zone, flags);
2488 if (keg->uk_flags & UMA_ZFLAG_FULL)
2494 if (rflags & (M_NOWAIT | M_NOVM))
2498 * All kegs are full. XXX We can't atomically check all kegs
2499 * and sleep so just sleep for a short period and retry.
2501 if (full && !empty) {
2503 zone->uz_flags |= UMA_ZFLAG_FULL;
2505 zone_log_warning(zone);
2506 msleep(zone, zone->uz_lockptr, PVM,
2507 "zonelimit", hz/100);
2508 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2517 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2522 MPASS(keg == slab->us_keg);
2523 mtx_assert(&keg->uk_lock, MA_OWNED);
2525 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2526 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2527 item = slab->us_data + (keg->uk_rsize * freei);
2528 slab->us_freecount--;
2531 /* Move this slab to the full list */
2532 if (slab->us_freecount == 0) {
2533 LIST_REMOVE(slab, us_link);
2534 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2541 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2549 /* Try to keep the buckets totally full */
2550 for (i = 0; i < max; ) {
2551 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2554 while (slab->us_freecount && i < max) {
2555 bucket[i++] = slab_alloc_item(keg, slab);
2556 if (keg->uk_free <= keg->uk_reserve)
2559 /* Don't grab more than one slab at a time. */
2570 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2572 uma_bucket_t bucket;
2575 /* Don't wait for buckets, preserve caller's NOVM setting. */
2576 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2580 max = MIN(bucket->ub_entries, zone->uz_count);
2581 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2585 * Initialize the memory if necessary.
2587 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2590 for (i = 0; i < bucket->ub_cnt; i++)
2591 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2595 * If we couldn't initialize the whole bucket, put the
2596 * rest back onto the freelist.
2598 if (i != bucket->ub_cnt) {
2599 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2600 bucket->ub_cnt - i);
2602 bzero(&bucket->ub_bucket[i],
2603 sizeof(void *) * (bucket->ub_cnt - i));
2609 if (bucket->ub_cnt == 0) {
2610 bucket_free(zone, bucket, udata);
2611 atomic_add_long(&zone->uz_fails, 1);
2619 * Allocates a single item from a zone.
2622 * zone The zone to alloc for.
2623 * udata The data to be passed to the constructor.
2624 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2627 * NULL if there is no memory and M_NOWAIT is set
2628 * An item if successful
2632 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2638 #ifdef UMA_DEBUG_ALLOC
2639 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2641 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2643 atomic_add_long(&zone->uz_allocs, 1);
2646 * We have to call both the zone's init (not the keg's init)
2647 * and the zone's ctor. This is because the item is going from
2648 * a keg slab directly to the user, and the user is expecting it
2649 * to be both zone-init'd as well as zone-ctor'd.
2651 if (zone->uz_init != NULL) {
2652 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2653 zone_free_item(zone, item, udata, SKIP_FINI);
2657 if (zone->uz_ctor != NULL) {
2658 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2659 zone_free_item(zone, item, udata, SKIP_DTOR);
2664 uma_dbg_alloc(zone, NULL, item);
2667 uma_zero_item(item, zone);
2672 atomic_add_long(&zone->uz_fails, 1);
2678 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2681 uma_bucket_t bucket;
2686 /* XXX: FIX!! Do not enable this in CURRENT!! MarkM */
2687 /* The entropy here is desirable, but the harvesting is expensive */
2689 const void *uz_name;
2693 entropy.uz_name = zone->uz_name;
2694 entropy.item = item;
2695 random_harvest(&entropy, sizeof(struct entropy), 2, RANDOM_UMA_ALLOC);
2698 #ifdef UMA_DEBUG_ALLOC_1
2699 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2701 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2704 /* uma_zfree(..., NULL) does nothing, to match free(9). */
2707 #ifdef DEBUG_MEMGUARD
2708 if (is_memguard_addr(item)) {
2709 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2710 zone->uz_dtor(item, zone->uz_size, udata);
2711 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2712 zone->uz_fini(item, zone->uz_size);
2713 memguard_free(item);
2718 if (zone->uz_flags & UMA_ZONE_MALLOC)
2719 uma_dbg_free(zone, udata, item);
2721 uma_dbg_free(zone, NULL, item);
2723 if (zone->uz_dtor != NULL)
2724 zone->uz_dtor(item, zone->uz_size, udata);
2727 * The race here is acceptable. If we miss it we'll just have to wait
2728 * a little longer for the limits to be reset.
2730 if (zone->uz_flags & UMA_ZFLAG_FULL)
2734 * If possible, free to the per-CPU cache. There are two
2735 * requirements for safe access to the per-CPU cache: (1) the thread
2736 * accessing the cache must not be preempted or yield during access,
2737 * and (2) the thread must not migrate CPUs without switching which
2738 * cache it accesses. We rely on a critical section to prevent
2739 * preemption and migration. We release the critical section in
2740 * order to acquire the zone mutex if we are unable to free to the
2741 * current cache; when we re-acquire the critical section, we must
2742 * detect and handle migration if it has occurred.
2747 cache = &zone->uz_cpu[cpu];
2751 * Try to free into the allocbucket first to give LIFO ordering
2752 * for cache-hot datastructures. Spill over into the freebucket
2753 * if necessary. Alloc will swap them if one runs dry.
2755 bucket = cache->uc_allocbucket;
2756 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2757 bucket = cache->uc_freebucket;
2758 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2759 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2760 ("uma_zfree: Freeing to non free bucket index."));
2761 bucket->ub_bucket[bucket->ub_cnt] = item;
2769 * We must go back the zone, which requires acquiring the zone lock,
2770 * which in turn means we must release and re-acquire the critical
2771 * section. Since the critical section is released, we may be
2772 * preempted or migrate. As such, make sure not to maintain any
2773 * thread-local state specific to the cache from prior to releasing
2774 * the critical section.
2777 if (zone->uz_count == 0 || bucketdisable)
2781 if (ZONE_TRYLOCK(zone) == 0) {
2782 /* Record contention to size the buckets. */
2788 cache = &zone->uz_cpu[cpu];
2791 * Since we have locked the zone we may as well send back our stats.
2793 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2794 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2795 cache->uc_allocs = 0;
2796 cache->uc_frees = 0;
2798 bucket = cache->uc_freebucket;
2799 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2803 cache->uc_freebucket = NULL;
2805 /* Can we throw this on the zone full list? */
2806 if (bucket != NULL) {
2807 #ifdef UMA_DEBUG_ALLOC
2808 printf("uma_zfree: Putting old bucket on the free list.\n");
2810 /* ub_cnt is pointing to the last free item */
2811 KASSERT(bucket->ub_cnt != 0,
2812 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2813 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2816 /* We are no longer associated with this CPU. */
2820 * We bump the uz count when the cache size is insufficient to
2821 * handle the working set.
2823 if (lockfail && zone->uz_count < BUCKET_MAX)
2827 #ifdef UMA_DEBUG_ALLOC
2828 printf("uma_zfree: Allocating new free bucket.\n");
2830 bucket = bucket_alloc(zone, udata, M_NOWAIT);
2834 cache = &zone->uz_cpu[cpu];
2835 if (cache->uc_freebucket == NULL) {
2836 cache->uc_freebucket = bucket;
2840 * We lost the race, start over. We have to drop our
2841 * critical section to free the bucket.
2844 bucket_free(zone, bucket, udata);
2849 * If nothing else caught this, we'll just do an internal free.
2852 zone_free_item(zone, item, udata, SKIP_DTOR);
2858 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2862 mtx_assert(&keg->uk_lock, MA_OWNED);
2863 MPASS(keg == slab->us_keg);
2865 /* Do we need to remove from any lists? */
2866 if (slab->us_freecount+1 == keg->uk_ipers) {
2867 LIST_REMOVE(slab, us_link);
2868 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2869 } else if (slab->us_freecount == 0) {
2870 LIST_REMOVE(slab, us_link);
2871 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2874 /* Slab management. */
2875 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2876 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2877 slab->us_freecount++;
2879 /* Keg statistics. */
2884 zone_release(uma_zone_t zone, void **bucket, int cnt)
2894 keg = zone_first_keg(zone);
2896 for (i = 0; i < cnt; i++) {
2898 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2899 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2900 if (zone->uz_flags & UMA_ZONE_HASH) {
2901 slab = hash_sfind(&keg->uk_hash, mem);
2903 mem += keg->uk_pgoff;
2904 slab = (uma_slab_t)mem;
2907 slab = vtoslab((vm_offset_t)item);
2908 if (slab->us_keg != keg) {
2914 slab_free_item(keg, slab, item);
2915 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2916 if (keg->uk_pages < keg->uk_maxpages) {
2917 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2922 * We can handle one more allocation. Since we're
2923 * clearing ZFLAG_FULL, wake up all procs blocked
2924 * on pages. This should be uncommon, so keeping this
2925 * simple for now (rather than adding count of blocked
2934 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2942 * Frees a single item to any zone.
2945 * zone The zone to free to
2946 * item The item we're freeing
2947 * udata User supplied data for the dtor
2948 * skip Skip dtors and finis
2951 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2955 if (skip == SKIP_NONE) {
2956 if (zone->uz_flags & UMA_ZONE_MALLOC)
2957 uma_dbg_free(zone, udata, item);
2959 uma_dbg_free(zone, NULL, item);
2962 if (skip < SKIP_DTOR && zone->uz_dtor)
2963 zone->uz_dtor(item, zone->uz_size, udata);
2965 if (skip < SKIP_FINI && zone->uz_fini)
2966 zone->uz_fini(item, zone->uz_size);
2968 atomic_add_long(&zone->uz_frees, 1);
2969 zone->uz_release(zone->uz_arg, &item, 1);
2974 uma_zone_set_max(uma_zone_t zone, int nitems)
2978 keg = zone_first_keg(zone);
2982 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2983 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2984 keg->uk_maxpages += keg->uk_ppera;
2985 nitems = keg->uk_maxpages * keg->uk_ipers;
2993 uma_zone_get_max(uma_zone_t zone)
2998 keg = zone_first_keg(zone);
3002 nitems = keg->uk_maxpages * keg->uk_ipers;
3010 uma_zone_set_warning(uma_zone_t zone, const char *warning)
3014 zone->uz_warning = warning;
3020 uma_zone_get_cur(uma_zone_t zone)
3026 nitems = zone->uz_allocs - zone->uz_frees;
3029 * See the comment in sysctl_vm_zone_stats() regarding the
3030 * safety of accessing the per-cpu caches. With the zone lock
3031 * held, it is safe, but can potentially result in stale data.
3033 nitems += zone->uz_cpu[i].uc_allocs -
3034 zone->uz_cpu[i].uc_frees;
3038 return (nitems < 0 ? 0 : nitems);
3043 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
3047 keg = zone_first_keg(zone);
3048 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3050 KASSERT(keg->uk_pages == 0,
3051 ("uma_zone_set_init on non-empty keg"));
3052 keg->uk_init = uminit;
3058 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3062 keg = zone_first_keg(zone);
3063 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
3065 KASSERT(keg->uk_pages == 0,
3066 ("uma_zone_set_fini on non-empty keg"));
3067 keg->uk_fini = fini;
3073 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3077 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3078 ("uma_zone_set_zinit on non-empty keg"));
3079 zone->uz_init = zinit;
3085 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3089 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3090 ("uma_zone_set_zfini on non-empty keg"));
3091 zone->uz_fini = zfini;
3096 /* XXX uk_freef is not actually used with the zone locked */
3098 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3102 keg = zone_first_keg(zone);
3103 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
3105 keg->uk_freef = freef;
3110 /* XXX uk_allocf is not actually used with the zone locked */
3112 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3116 keg = zone_first_keg(zone);
3118 keg->uk_allocf = allocf;
3124 uma_zone_reserve(uma_zone_t zone, int items)
3128 keg = zone_first_keg(zone);
3132 keg->uk_reserve = items;
3140 uma_zone_reserve_kva(uma_zone_t zone, int count)
3146 keg = zone_first_keg(zone);
3149 pages = count / keg->uk_ipers;
3151 if (pages * keg->uk_ipers < count)
3154 #ifdef UMA_MD_SMALL_ALLOC
3155 if (keg->uk_ppera > 1) {
3159 kva = kva_alloc(pages * UMA_SLAB_SIZE);
3167 keg->uk_maxpages = pages;
3168 #ifdef UMA_MD_SMALL_ALLOC
3169 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3171 keg->uk_allocf = noobj_alloc;
3173 keg->uk_flags |= UMA_ZONE_NOFREE;
3181 uma_prealloc(uma_zone_t zone, int items)
3187 keg = zone_first_keg(zone);
3191 slabs = items / keg->uk_ipers;
3192 if (slabs * keg->uk_ipers < items)
3195 slab = keg_alloc_slab(keg, zone, M_WAITOK);
3198 MPASS(slab->us_keg == keg);
3199 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3207 uma_find_refcnt(uma_zone_t zone, void *item)
3209 uma_slabrefcnt_t slabref;
3215 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3216 slabref = (uma_slabrefcnt_t)slab;
3218 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3219 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3220 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3221 refcnt = &slabref->us_refcnt[idx];
3230 printf("UMA: vm asked us to release pages!\n");
3232 sx_xlock(&uma_drain_lock);
3234 zone_foreach(zone_drain);
3235 if (vm_page_count_min()) {
3236 cache_drain_safe(NULL);
3237 zone_foreach(zone_drain);
3240 * Some slabs may have been freed but this zone will be visited early
3241 * we visit again so that we can free pages that are empty once other
3242 * zones are drained. We have to do the same for buckets.
3244 zone_drain(slabzone);
3245 zone_drain(slabrefzone);
3246 bucket_zone_drain();
3247 sx_xunlock(&uma_drain_lock);
3252 uma_zone_exhausted(uma_zone_t zone)
3257 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3263 uma_zone_exhausted_nolock(uma_zone_t zone)
3265 return (zone->uz_flags & UMA_ZFLAG_FULL);
3269 uma_large_malloc(vm_size_t size, int wait)
3275 slab = zone_alloc_item(slabzone, NULL, wait);
3278 mem = page_alloc(NULL, size, &flags, wait);
3280 vsetslab((vm_offset_t)mem, slab);
3281 slab->us_data = mem;
3282 slab->us_flags = flags | UMA_SLAB_MALLOC;
3283 slab->us_size = size;
3285 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3292 uma_large_free(uma_slab_t slab)
3295 page_free(slab->us_data, slab->us_size, slab->us_flags);
3296 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3300 uma_zero_item(void *item, uma_zone_t zone)
3303 if (zone->uz_flags & UMA_ZONE_PCPU) {
3304 for (int i = 0; i < mp_ncpus; i++)
3305 bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3307 bzero(item, zone->uz_size);
3311 uma_print_stats(void)
3313 zone_foreach(uma_print_zone);
3317 slab_print(uma_slab_t slab)
3319 printf("slab: keg %p, data %p, freecount %d\n",
3320 slab->us_keg, slab->us_data, slab->us_freecount);
3324 cache_print(uma_cache_t cache)
3326 printf("alloc: %p(%d), free: %p(%d)\n",
3327 cache->uc_allocbucket,
3328 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3329 cache->uc_freebucket,
3330 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3334 uma_print_keg(uma_keg_t keg)
3338 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3339 "out %d free %d limit %d\n",
3340 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3341 keg->uk_ipers, keg->uk_ppera,
3342 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3343 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3344 printf("Part slabs:\n");
3345 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3347 printf("Free slabs:\n");
3348 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3350 printf("Full slabs:\n");
3351 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3356 uma_print_zone(uma_zone_t zone)
3362 printf("zone: %s(%p) size %d flags %#x\n",
3363 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3364 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3365 uma_print_keg(kl->kl_keg);
3367 cache = &zone->uz_cpu[i];
3368 printf("CPU %d Cache:\n", i);
3375 * Generate statistics across both the zone and its per-cpu cache's. Return
3376 * desired statistics if the pointer is non-NULL for that statistic.
3378 * Note: does not update the zone statistics, as it can't safely clear the
3379 * per-CPU cache statistic.
3381 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3382 * safe from off-CPU; we should modify the caches to track this information
3383 * directly so that we don't have to.
3386 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3387 uint64_t *freesp, uint64_t *sleepsp)
3390 uint64_t allocs, frees, sleeps;
3393 allocs = frees = sleeps = 0;
3396 cache = &z->uz_cpu[cpu];
3397 if (cache->uc_allocbucket != NULL)
3398 cachefree += cache->uc_allocbucket->ub_cnt;
3399 if (cache->uc_freebucket != NULL)
3400 cachefree += cache->uc_freebucket->ub_cnt;
3401 allocs += cache->uc_allocs;
3402 frees += cache->uc_frees;
3404 allocs += z->uz_allocs;
3405 frees += z->uz_frees;
3406 sleeps += z->uz_sleeps;
3407 if (cachefreep != NULL)
3408 *cachefreep = cachefree;
3409 if (allocsp != NULL)
3413 if (sleepsp != NULL)
3419 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3426 rw_rlock(&uma_rwlock);
3427 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3428 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3431 rw_runlock(&uma_rwlock);
3432 return (sysctl_handle_int(oidp, &count, 0, req));
3436 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3438 struct uma_stream_header ush;
3439 struct uma_type_header uth;
3440 struct uma_percpu_stat ups;
3441 uma_bucket_t bucket;
3448 int count, error, i;
3450 error = sysctl_wire_old_buffer(req, 0);
3453 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3454 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3457 rw_rlock(&uma_rwlock);
3458 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3459 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3464 * Insert stream header.
3466 bzero(&ush, sizeof(ush));
3467 ush.ush_version = UMA_STREAM_VERSION;
3468 ush.ush_maxcpus = (mp_maxid + 1);
3469 ush.ush_count = count;
3470 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3472 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3473 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3474 bzero(&uth, sizeof(uth));
3476 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3477 uth.uth_align = kz->uk_align;
3478 uth.uth_size = kz->uk_size;
3479 uth.uth_rsize = kz->uk_rsize;
3480 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3482 uth.uth_maxpages += k->uk_maxpages;
3483 uth.uth_pages += k->uk_pages;
3484 uth.uth_keg_free += k->uk_free;
3485 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3490 * A zone is secondary is it is not the first entry
3491 * on the keg's zone list.
3493 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3494 (LIST_FIRST(&kz->uk_zones) != z))
3495 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3497 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3498 uth.uth_zone_free += bucket->ub_cnt;
3499 uth.uth_allocs = z->uz_allocs;
3500 uth.uth_frees = z->uz_frees;
3501 uth.uth_fails = z->uz_fails;
3502 uth.uth_sleeps = z->uz_sleeps;
3503 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3505 * While it is not normally safe to access the cache
3506 * bucket pointers while not on the CPU that owns the
3507 * cache, we only allow the pointers to be exchanged
3508 * without the zone lock held, not invalidated, so
3509 * accept the possible race associated with bucket
3510 * exchange during monitoring.
3512 for (i = 0; i < (mp_maxid + 1); i++) {
3513 bzero(&ups, sizeof(ups));
3514 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3518 cache = &z->uz_cpu[i];
3519 if (cache->uc_allocbucket != NULL)
3520 ups.ups_cache_free +=
3521 cache->uc_allocbucket->ub_cnt;
3522 if (cache->uc_freebucket != NULL)
3523 ups.ups_cache_free +=
3524 cache->uc_freebucket->ub_cnt;
3525 ups.ups_allocs = cache->uc_allocs;
3526 ups.ups_frees = cache->uc_frees;
3528 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3533 rw_runlock(&uma_rwlock);
3534 error = sbuf_finish(&sbuf);
3540 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3542 uma_zone_t zone = *(uma_zone_t *)arg1;
3545 max = uma_zone_get_max(zone);
3546 error = sysctl_handle_int(oidp, &max, 0, req);
3547 if (error || !req->newptr)
3550 uma_zone_set_max(zone, max);
3556 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3558 uma_zone_t zone = *(uma_zone_t *)arg1;
3561 cur = uma_zone_get_cur(zone);
3562 return (sysctl_handle_int(oidp, &cur, 0, req));
3566 DB_SHOW_COMMAND(uma, db_show_uma)
3568 uint64_t allocs, frees, sleeps;
3569 uma_bucket_t bucket;
3574 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3575 "Free", "Requests", "Sleeps", "Bucket");
3576 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3577 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3578 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3579 allocs = z->uz_allocs;
3580 frees = z->uz_frees;
3581 sleeps = z->uz_sleeps;
3584 uma_zone_sumstat(z, &cachefree, &allocs,
3586 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3587 (LIST_FIRST(&kz->uk_zones) != z)))
3588 cachefree += kz->uk_free;
3589 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3590 cachefree += bucket->ub_cnt;
3591 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3592 z->uz_name, (uintmax_t)kz->uk_size,
3593 (intmax_t)(allocs - frees), cachefree,
3594 (uintmax_t)allocs, sleeps, z->uz_count);
3601 DB_SHOW_COMMAND(umacache, db_show_umacache)
3603 uint64_t allocs, frees;
3604 uma_bucket_t bucket;
3608 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3609 "Requests", "Bucket");
3610 LIST_FOREACH(z, &uma_cachezones, uz_link) {
3611 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3612 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3613 cachefree += bucket->ub_cnt;
3614 db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3615 z->uz_name, (uintmax_t)z->uz_size,
3616 (intmax_t)(allocs - frees), cachefree,
3617 (uintmax_t)allocs, z->uz_count);