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("$FreeBSD$");
53 /* I should really use ktr.. */
56 #define UMA_DEBUG_ALLOC 1
57 #define UMA_DEBUG_ALLOC_1 1
61 #include "opt_param.h"
64 #include <sys/param.h>
65 #include <sys/systm.h>
66 #include <sys/bitset.h>
67 #include <sys/eventhandler.h>
68 #include <sys/kernel.h>
69 #include <sys/types.h>
70 #include <sys/queue.h>
71 #include <sys/malloc.h>
74 #include <sys/sysctl.h>
75 #include <sys/mutex.h>
77 #include <sys/random.h>
78 #include <sys/rwlock.h>
80 #include <sys/sched.h>
82 #include <sys/taskqueue.h>
83 #include <sys/vmmeter.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_param.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_extern.h>
94 #include <vm/uma_int.h>
95 #include <vm/uma_dbg.h>
100 #include <vm/memguard.h>
104 * This is the zone and keg from which all zones are spawned. The idea is that
105 * even the zone & keg heads are allocated from the allocator, so we use the
106 * bss section to bootstrap us.
108 static struct uma_keg masterkeg;
109 static struct uma_zone masterzone_k;
110 static struct uma_zone masterzone_z;
111 static uma_zone_t kegs = &masterzone_k;
112 static uma_zone_t zones = &masterzone_z;
114 /* This is the zone from which all of uma_slab_t's are allocated. */
115 static uma_zone_t slabzone;
118 * The initial hash tables come out of this zone so they can be allocated
119 * prior to malloc coming up.
121 static uma_zone_t hashzone;
123 /* The boot-time adjusted value for cache line alignment. */
124 int uma_align_cache = 64 - 1;
126 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
129 * Are we allowed to allocate buckets?
131 static int bucketdisable = 1;
133 /* Linked list of all kegs in the system */
134 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
136 /* Linked list of all cache-only zones in the system */
137 static LIST_HEAD(,uma_zone) uma_cachezones =
138 LIST_HEAD_INITIALIZER(uma_cachezones);
140 /* This RW lock protects the keg list */
141 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
143 /* Linked list of boot time pages */
144 static LIST_HEAD(,uma_slab) uma_boot_pages =
145 LIST_HEAD_INITIALIZER(uma_boot_pages);
147 /* This mutex protects the boot time pages list */
148 static struct mtx_padalign uma_boot_pages_mtx;
150 static struct sx uma_drain_lock;
152 /* Is the VM done starting up? */
153 static int booted = 0;
154 #define UMA_STARTUP 1
155 #define UMA_STARTUP2 2
158 * This is the handle used to schedule events that need to happen
159 * outside of the allocation fast path.
161 static struct callout uma_callout;
162 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
165 * This structure is passed as the zone ctor arg so that I don't have to create
166 * a special allocation function just for zones.
168 struct uma_zctor_args {
183 struct uma_kctor_args {
192 struct uma_bucket_zone {
195 int ubz_entries; /* Number of items it can hold. */
196 int ubz_maxsize; /* Maximum allocation size per-item. */
200 * Compute the actual number of bucket entries to pack them in power
201 * of two sizes for more efficient space utilization.
203 #define BUCKET_SIZE(n) \
204 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
206 #define BUCKET_MAX BUCKET_SIZE(256)
208 struct uma_bucket_zone bucket_zones[] = {
209 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
210 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
211 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
212 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
213 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
214 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
215 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
216 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
217 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
222 * Flags and enumerations to be passed to internal functions.
224 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
228 static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
229 static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
230 static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
231 static void page_free(void *, vm_size_t, uint8_t);
232 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
233 static void cache_drain(uma_zone_t);
234 static void bucket_drain(uma_zone_t, uma_bucket_t);
235 static void bucket_cache_drain(uma_zone_t zone);
236 static int keg_ctor(void *, int, void *, int);
237 static void keg_dtor(void *, int, void *);
238 static int zone_ctor(void *, int, void *, int);
239 static void zone_dtor(void *, int, void *);
240 static int zero_init(void *, int, int);
241 static void keg_small_init(uma_keg_t keg);
242 static void keg_large_init(uma_keg_t keg);
243 static void zone_foreach(void (*zfunc)(uma_zone_t));
244 static void zone_timeout(uma_zone_t zone);
245 static int hash_alloc(struct uma_hash *);
246 static int hash_expand(struct uma_hash *, struct uma_hash *);
247 static void hash_free(struct uma_hash *hash);
248 static void uma_timeout(void *);
249 static void uma_startup3(void);
250 static void *zone_alloc_item(uma_zone_t, void *, int);
251 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
252 static void bucket_enable(void);
253 static void bucket_init(void);
254 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
255 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
256 static void bucket_zone_drain(void);
257 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
258 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
259 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
260 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
261 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
262 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
263 uma_fini fini, int align, uint32_t flags);
264 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
265 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
266 static void uma_zero_item(void *item, uma_zone_t zone);
268 void uma_print_zone(uma_zone_t);
269 void uma_print_stats(void);
270 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
271 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
274 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
275 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
278 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
280 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
281 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
283 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
284 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
286 static int zone_warnings = 1;
287 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
288 "Warn when UMA zones becomes full");
291 * This routine checks to see whether or not it's safe to enable buckets.
296 bucketdisable = vm_page_count_min();
300 * Initialize bucket_zones, the array of zones of buckets of various sizes.
302 * For each zone, calculate the memory required for each bucket, consisting
303 * of the header and an array of pointers.
308 struct uma_bucket_zone *ubz;
311 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
312 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
313 size += sizeof(void *) * ubz->ubz_entries;
314 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
315 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
316 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
321 * Given a desired number of entries for a bucket, return the zone from which
322 * to allocate the bucket.
324 static struct uma_bucket_zone *
325 bucket_zone_lookup(int entries)
327 struct uma_bucket_zone *ubz;
329 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
330 if (ubz->ubz_entries >= entries)
337 bucket_select(int size)
339 struct uma_bucket_zone *ubz;
341 ubz = &bucket_zones[0];
342 if (size > ubz->ubz_maxsize)
343 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
345 for (; ubz->ubz_entries != 0; ubz++)
346 if (ubz->ubz_maxsize < size)
349 return (ubz->ubz_entries);
353 bucket_alloc(uma_zone_t zone, void *udata, int flags)
355 struct uma_bucket_zone *ubz;
359 * This is to stop us from allocating per cpu buckets while we're
360 * running out of vm.boot_pages. Otherwise, we would exhaust the
361 * boot pages. This also prevents us from allocating buckets in
362 * low memory situations.
367 * To limit bucket recursion we store the original zone flags
368 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
369 * NOVM flag to persist even through deep recursions. We also
370 * store ZFLAG_BUCKET once we have recursed attempting to allocate
371 * a bucket for a bucket zone so we do not allow infinite bucket
372 * recursion. This cookie will even persist to frees of unused
373 * buckets via the allocation path or bucket allocations in the
376 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
377 udata = (void *)(uintptr_t)zone->uz_flags;
379 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
381 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
383 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
385 ubz = bucket_zone_lookup(zone->uz_count);
386 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
388 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
391 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
394 bucket->ub_entries = ubz->ubz_entries;
401 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
403 struct uma_bucket_zone *ubz;
405 KASSERT(bucket->ub_cnt == 0,
406 ("bucket_free: Freeing a non free bucket."));
407 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
408 udata = (void *)(uintptr_t)zone->uz_flags;
409 ubz = bucket_zone_lookup(bucket->ub_entries);
410 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
414 bucket_zone_drain(void)
416 struct uma_bucket_zone *ubz;
418 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
419 zone_drain(ubz->ubz_zone);
423 zone_log_warning(uma_zone_t zone)
425 static const struct timeval warninterval = { 300, 0 };
427 if (!zone_warnings || zone->uz_warning == NULL)
430 if (ratecheck(&zone->uz_ratecheck, &warninterval))
431 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
435 zone_maxaction(uma_zone_t zone)
438 if (zone->uz_maxaction.ta_func != NULL)
439 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
443 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
447 LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
448 kegfn(klink->kl_keg);
452 * Routine called by timeout which is used to fire off some time interval
453 * based calculations. (stats, hash size, etc.)
462 uma_timeout(void *unused)
465 zone_foreach(zone_timeout);
467 /* Reschedule this event */
468 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
472 * Routine to perform timeout driven calculations. This expands the
473 * hashes and does per cpu statistics aggregation.
478 keg_timeout(uma_keg_t keg)
483 * Expand the keg hash table.
485 * This is done if the number of slabs is larger than the hash size.
486 * What I'm trying to do here is completely reduce collisions. This
487 * may be a little aggressive. Should I allow for two collisions max?
489 if (keg->uk_flags & UMA_ZONE_HASH &&
490 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
491 struct uma_hash newhash;
492 struct uma_hash oldhash;
496 * This is so involved because allocating and freeing
497 * while the keg lock is held will lead to deadlock.
498 * I have to do everything in stages and check for
501 newhash = keg->uk_hash;
503 ret = hash_alloc(&newhash);
506 if (hash_expand(&keg->uk_hash, &newhash)) {
507 oldhash = keg->uk_hash;
508 keg->uk_hash = newhash;
521 zone_timeout(uma_zone_t zone)
524 zone_foreach_keg(zone, &keg_timeout);
528 * Allocate and zero fill the next sized hash table from the appropriate
532 * hash A new hash structure with the old hash size in uh_hashsize
535 * 1 on success and 0 on failure.
538 hash_alloc(struct uma_hash *hash)
543 oldsize = hash->uh_hashsize;
545 /* We're just going to go to a power of two greater */
547 hash->uh_hashsize = oldsize * 2;
548 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
549 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
550 M_UMAHASH, M_NOWAIT);
552 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
553 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
555 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
557 if (hash->uh_slab_hash) {
558 bzero(hash->uh_slab_hash, alloc);
559 hash->uh_hashmask = hash->uh_hashsize - 1;
567 * Expands the hash table for HASH zones. This is done from zone_timeout
568 * to reduce collisions. This must not be done in the regular allocation
569 * path, otherwise, we can recurse on the vm while allocating pages.
572 * oldhash The hash you want to expand
573 * newhash The hash structure for the new table
581 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
587 if (!newhash->uh_slab_hash)
590 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
594 * I need to investigate hash algorithms for resizing without a
598 for (i = 0; i < oldhash->uh_hashsize; i++)
599 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
600 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
601 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
602 hval = UMA_HASH(newhash, slab->us_data);
603 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
611 * Free the hash bucket to the appropriate backing store.
614 * slab_hash The hash bucket we're freeing
615 * hashsize The number of entries in that hash bucket
621 hash_free(struct uma_hash *hash)
623 if (hash->uh_slab_hash == NULL)
625 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
626 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
628 free(hash->uh_slab_hash, M_UMAHASH);
632 * Frees all outstanding items in a bucket
635 * zone The zone to free to, must be unlocked.
636 * bucket The free/alloc bucket with items, cpu queue must be locked.
643 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
651 for (i = 0; i < bucket->ub_cnt; i++)
652 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
653 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
658 * Drains the per cpu caches for a zone.
660 * NOTE: This may only be called while the zone is being turn down, and not
661 * during normal operation. This is necessary in order that we do not have
662 * to migrate CPUs to drain the per-CPU caches.
665 * zone The zone to drain, must be unlocked.
671 cache_drain(uma_zone_t zone)
677 * XXX: It is safe to not lock the per-CPU caches, because we're
678 * tearing down the zone anyway. I.e., there will be no further use
679 * of the caches at this point.
681 * XXX: It would good to be able to assert that the zone is being
682 * torn down to prevent improper use of cache_drain().
684 * XXX: We lock the zone before passing into bucket_cache_drain() as
685 * it is used elsewhere. Should the tear-down path be made special
686 * there in some form?
689 cache = &zone->uz_cpu[cpu];
690 bucket_drain(zone, cache->uc_allocbucket);
691 bucket_drain(zone, cache->uc_freebucket);
692 if (cache->uc_allocbucket != NULL)
693 bucket_free(zone, cache->uc_allocbucket, NULL);
694 if (cache->uc_freebucket != NULL)
695 bucket_free(zone, cache->uc_freebucket, NULL);
696 cache->uc_allocbucket = cache->uc_freebucket = NULL;
699 bucket_cache_drain(zone);
704 cache_shrink(uma_zone_t zone)
707 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
711 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
716 cache_drain_safe_cpu(uma_zone_t zone)
721 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
727 cache = &zone->uz_cpu[curcpu];
728 if (cache->uc_allocbucket) {
729 if (cache->uc_allocbucket->ub_cnt != 0)
730 LIST_INSERT_HEAD(&zone->uz_buckets,
731 cache->uc_allocbucket, ub_link);
733 b1 = cache->uc_allocbucket;
734 cache->uc_allocbucket = NULL;
736 if (cache->uc_freebucket) {
737 if (cache->uc_freebucket->ub_cnt != 0)
738 LIST_INSERT_HEAD(&zone->uz_buckets,
739 cache->uc_freebucket, ub_link);
741 b2 = cache->uc_freebucket;
742 cache->uc_freebucket = NULL;
747 bucket_free(zone, b1, NULL);
749 bucket_free(zone, b2, NULL);
753 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
754 * This is an expensive call because it needs to bind to all CPUs
755 * one by one and enter a critical section on each of them in order
756 * to safely access their cache buckets.
757 * Zone lock must not be held on call this function.
760 cache_drain_safe(uma_zone_t zone)
765 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
770 zone_foreach(cache_shrink);
773 thread_lock(curthread);
774 sched_bind(curthread, cpu);
775 thread_unlock(curthread);
778 cache_drain_safe_cpu(zone);
780 zone_foreach(cache_drain_safe_cpu);
782 thread_lock(curthread);
783 sched_unbind(curthread);
784 thread_unlock(curthread);
788 * Drain the cached buckets from a zone. Expects a locked zone on entry.
791 bucket_cache_drain(uma_zone_t zone)
796 * Drain the bucket queues and free the buckets, we just keep two per
799 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
800 LIST_REMOVE(bucket, ub_link);
802 bucket_drain(zone, bucket);
803 bucket_free(zone, bucket, NULL);
808 * Shrink further bucket sizes. Price of single zone lock collision
809 * is probably lower then price of global cache drain.
811 if (zone->uz_count > zone->uz_count_min)
816 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
823 flags = slab->us_flags;
825 if (keg->uk_fini != NULL) {
826 for (i--; i > -1; i--)
827 keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
830 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
831 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
833 printf("%s: Returning %d bytes.\n", keg->uk_name,
834 PAGE_SIZE * keg->uk_ppera);
836 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
840 * Frees pages from a keg back to the system. This is done on demand from
841 * the pageout daemon.
846 keg_drain(uma_keg_t keg)
848 struct slabhead freeslabs = { 0 };
849 uma_slab_t slab, tmp;
852 * We don't want to take pages from statically allocated kegs at this
855 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
859 printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
862 if (keg->uk_free == 0)
865 LIST_FOREACH_SAFE(slab, &keg->uk_free_slab, us_link, tmp) {
866 /* We have nowhere to free these to. */
867 if (slab->us_flags & UMA_SLAB_BOOT)
870 LIST_REMOVE(slab, us_link);
871 keg->uk_pages -= keg->uk_ppera;
872 keg->uk_free -= keg->uk_ipers;
874 if (keg->uk_flags & UMA_ZONE_HASH)
875 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
877 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
882 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
883 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
884 keg_free_slab(keg, slab, keg->uk_ipers);
889 zone_drain_wait(uma_zone_t zone, int waitok)
893 * Set draining to interlock with zone_dtor() so we can release our
894 * locks as we go. Only dtor() should do a WAITOK call since it
895 * is the only call that knows the structure will still be available
899 while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
900 if (waitok == M_NOWAIT)
902 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
904 zone->uz_flags |= UMA_ZFLAG_DRAINING;
905 bucket_cache_drain(zone);
908 * The DRAINING flag protects us from being freed while
909 * we're running. Normally the uma_rwlock would protect us but we
910 * must be able to release and acquire the right lock for each keg.
912 zone_foreach_keg(zone, &keg_drain);
914 zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
921 zone_drain(uma_zone_t zone)
924 zone_drain_wait(zone, M_NOWAIT);
928 * Allocate a new slab for a keg. This does not insert the slab onto a list.
931 * wait Shall we wait?
934 * The slab that was allocated or NULL if there is no memory and the
935 * caller specified M_NOWAIT.
938 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
946 mtx_assert(&keg->uk_lock, MA_OWNED);
951 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name);
953 allocf = keg->uk_allocf;
956 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
957 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
963 * This reproduces the old vm_zone behavior of zero filling pages the
964 * first time they are added to a zone.
966 * Malloced items are zeroed in uma_zalloc.
969 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
974 if (keg->uk_flags & UMA_ZONE_NODUMP)
977 /* zone is passed for legacy reasons. */
978 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
980 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
981 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
986 /* Point the slab into the allocated memory */
987 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
988 slab = (uma_slab_t )(mem + keg->uk_pgoff);
990 if (keg->uk_flags & UMA_ZONE_VTOSLAB)
991 for (i = 0; i < keg->uk_ppera; i++)
992 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
996 slab->us_freecount = keg->uk_ipers;
997 slab->us_flags = flags;
998 BIT_FILL(SLAB_SETSIZE, &slab->us_free);
1000 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1003 if (keg->uk_init != NULL) {
1004 for (i = 0; i < keg->uk_ipers; i++)
1005 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1006 keg->uk_size, wait) != 0)
1008 if (i != keg->uk_ipers) {
1009 keg_free_slab(keg, slab, i);
1018 if (keg->uk_flags & UMA_ZONE_HASH)
1019 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1021 keg->uk_pages += keg->uk_ppera;
1022 keg->uk_free += keg->uk_ipers;
1029 * This function is intended to be used early on in place of page_alloc() so
1030 * that we may use the boot time page cache to satisfy allocations before
1034 startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1038 int pages, check_pages;
1040 keg = zone_first_keg(zone);
1041 pages = howmany(bytes, PAGE_SIZE);
1042 check_pages = pages - 1;
1043 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1046 * Check our small startup cache to see if it has pages remaining.
1048 mtx_lock(&uma_boot_pages_mtx);
1050 /* First check if we have enough room. */
1051 tmps = LIST_FIRST(&uma_boot_pages);
1052 while (tmps != NULL && check_pages-- > 0)
1053 tmps = LIST_NEXT(tmps, us_link);
1056 * It's ok to lose tmps references. The last one will
1057 * have tmps->us_data pointing to the start address of
1058 * "pages" contiguous pages of memory.
1060 while (pages-- > 0) {
1061 tmps = LIST_FIRST(&uma_boot_pages);
1062 LIST_REMOVE(tmps, us_link);
1064 mtx_unlock(&uma_boot_pages_mtx);
1065 *pflag = tmps->us_flags;
1066 return (tmps->us_data);
1068 mtx_unlock(&uma_boot_pages_mtx);
1069 if (booted < UMA_STARTUP2)
1070 panic("UMA: Increase vm.boot_pages");
1072 * Now that we've booted reset these users to their real allocator.
1074 #ifdef UMA_MD_SMALL_ALLOC
1075 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1077 keg->uk_allocf = page_alloc;
1079 return keg->uk_allocf(zone, bytes, pflag, wait);
1083 * Allocates a number of pages from the system
1086 * bytes The number of bytes requested
1087 * wait Shall we wait?
1090 * A pointer to the alloced memory or possibly
1091 * NULL if M_NOWAIT is set.
1094 page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1096 void *p; /* Returned page */
1098 *pflag = UMA_SLAB_KMEM;
1099 p = (void *) kmem_malloc(kmem_arena, bytes, wait);
1105 * Allocates a number of pages from within an object
1108 * bytes The number of bytes requested
1109 * wait Shall we wait?
1112 * A pointer to the alloced memory or possibly
1113 * NULL if M_NOWAIT is set.
1116 noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait)
1118 TAILQ_HEAD(, vm_page) alloctail;
1120 vm_offset_t retkva, zkva;
1121 vm_page_t p, p_next;
1124 TAILQ_INIT(&alloctail);
1125 keg = zone_first_keg(zone);
1127 npages = howmany(bytes, PAGE_SIZE);
1128 while (npages > 0) {
1129 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1130 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1131 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1135 * Since the page does not belong to an object, its
1138 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1143 * Page allocation failed, free intermediate pages and
1146 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1147 vm_page_unwire(p, PQ_NONE);
1152 *flags = UMA_SLAB_PRIV;
1153 zkva = keg->uk_kva +
1154 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1156 TAILQ_FOREACH(p, &alloctail, listq) {
1157 pmap_qenter(zkva, &p, 1);
1161 return ((void *)retkva);
1165 * Frees a number of pages to the system
1168 * mem A pointer to the memory to be freed
1169 * size The size of the memory being freed
1170 * flags The original p->us_flags field
1176 page_free(void *mem, vm_size_t size, uint8_t flags)
1180 if (flags & UMA_SLAB_KMEM)
1182 else if (flags & UMA_SLAB_KERNEL)
1183 vmem = kernel_arena;
1185 panic("UMA: page_free used with invalid flags %d", flags);
1187 kmem_free(vmem, (vm_offset_t)mem, size);
1191 * Zero fill initializer
1193 * Arguments/Returns follow uma_init specifications
1196 zero_init(void *mem, int size, int flags)
1203 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1206 * keg The zone we should initialize
1212 keg_small_init(uma_keg_t keg)
1220 if (keg->uk_flags & UMA_ZONE_PCPU) {
1221 u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU;
1223 slabsize = sizeof(struct pcpu);
1224 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1227 slabsize = UMA_SLAB_SIZE;
1232 * Calculate the size of each allocation (rsize) according to
1233 * alignment. If the requested size is smaller than we have
1234 * allocation bits for we round it up.
1236 rsize = keg->uk_size;
1237 if (rsize < slabsize / SLAB_SETSIZE)
1238 rsize = slabsize / SLAB_SETSIZE;
1239 if (rsize & keg->uk_align)
1240 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1241 keg->uk_rsize = rsize;
1243 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1244 keg->uk_rsize < sizeof(struct pcpu),
1245 ("%s: size %u too large", __func__, keg->uk_rsize));
1247 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1250 shsize = sizeof(struct uma_slab);
1252 if (rsize <= slabsize - shsize)
1253 keg->uk_ipers = (slabsize - shsize) / rsize;
1255 /* Handle special case when we have 1 item per slab, so
1256 * alignment requirement can be relaxed. */
1257 KASSERT(keg->uk_size <= slabsize - shsize,
1258 ("%s: size %u greater than slab", __func__, keg->uk_size));
1261 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1262 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1264 memused = keg->uk_ipers * rsize + shsize;
1265 wastedspace = slabsize - memused;
1268 * We can't do OFFPAGE if we're internal or if we've been
1269 * asked to not go to the VM for buckets. If we do this we
1270 * may end up going to the VM for slabs which we do not
1271 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1272 * of UMA_ZONE_VM, which clearly forbids it.
1274 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1275 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1279 * See if using an OFFPAGE slab will limit our waste. Only do
1280 * this if it permits more items per-slab.
1282 * XXX We could try growing slabsize to limit max waste as well.
1283 * Historically this was not done because the VM could not
1284 * efficiently handle contiguous allocations.
1286 if ((wastedspace >= slabsize / UMA_MAX_WASTE) &&
1287 (keg->uk_ipers < (slabsize / keg->uk_rsize))) {
1288 keg->uk_ipers = slabsize / keg->uk_rsize;
1289 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1290 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1292 printf("UMA decided we need offpage slab headers for "
1293 "keg: %s, calculated wastedspace = %d, "
1294 "maximum wasted space allowed = %d, "
1295 "calculated ipers = %d, "
1296 "new wasted space = %d\n", keg->uk_name, wastedspace,
1297 slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1298 slabsize - keg->uk_ipers * keg->uk_rsize);
1300 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1303 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1304 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1305 keg->uk_flags |= UMA_ZONE_HASH;
1309 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1310 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1314 * keg The keg we should initialize
1320 keg_large_init(uma_keg_t keg)
1324 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1325 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1326 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1327 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1328 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1330 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1332 keg->uk_rsize = keg->uk_size;
1334 /* Check whether we have enough space to not do OFFPAGE. */
1335 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1336 shsize = sizeof(struct uma_slab);
1337 if (shsize & UMA_ALIGN_PTR)
1338 shsize = (shsize & ~UMA_ALIGN_PTR) +
1339 (UMA_ALIGN_PTR + 1);
1341 if (PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < shsize) {
1343 * We can't do OFFPAGE if we're internal, in which case
1344 * we need an extra page per allocation to contain the
1347 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) == 0)
1348 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1354 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1355 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1356 keg->uk_flags |= UMA_ZONE_HASH;
1360 keg_cachespread_init(uma_keg_t keg)
1367 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1368 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1370 alignsize = keg->uk_align + 1;
1371 rsize = keg->uk_size;
1373 * We want one item to start on every align boundary in a page. To
1374 * do this we will span pages. We will also extend the item by the
1375 * size of align if it is an even multiple of align. Otherwise, it
1376 * would fall on the same boundary every time.
1378 if (rsize & keg->uk_align)
1379 rsize = (rsize & ~keg->uk_align) + alignsize;
1380 if ((rsize & alignsize) == 0)
1382 trailer = rsize - keg->uk_size;
1383 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1384 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1385 keg->uk_rsize = rsize;
1386 keg->uk_ppera = pages;
1387 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1388 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1389 KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1390 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1395 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1396 * the keg onto the global keg list.
1398 * Arguments/Returns follow uma_ctor specifications
1399 * udata Actually uma_kctor_args
1402 keg_ctor(void *mem, int size, void *udata, int flags)
1404 struct uma_kctor_args *arg = udata;
1405 uma_keg_t keg = mem;
1409 keg->uk_size = arg->size;
1410 keg->uk_init = arg->uminit;
1411 keg->uk_fini = arg->fini;
1412 keg->uk_align = arg->align;
1414 keg->uk_reserve = 0;
1416 keg->uk_flags = arg->flags;
1417 keg->uk_allocf = page_alloc;
1418 keg->uk_freef = page_free;
1419 keg->uk_slabzone = NULL;
1422 * The master zone is passed to us at keg-creation time.
1425 keg->uk_name = zone->uz_name;
1427 if (arg->flags & UMA_ZONE_VM)
1428 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1430 if (arg->flags & UMA_ZONE_ZINIT)
1431 keg->uk_init = zero_init;
1433 if (arg->flags & UMA_ZONE_MALLOC)
1434 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1436 if (arg->flags & UMA_ZONE_PCPU)
1438 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1440 keg->uk_flags &= ~UMA_ZONE_PCPU;
1443 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1444 keg_cachespread_init(keg);
1446 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1447 keg_large_init(keg);
1449 keg_small_init(keg);
1452 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1453 keg->uk_slabzone = slabzone;
1456 * If we haven't booted yet we need allocations to go through the
1457 * startup cache until the vm is ready.
1459 if (keg->uk_ppera == 1) {
1460 #ifdef UMA_MD_SMALL_ALLOC
1461 keg->uk_allocf = uma_small_alloc;
1462 keg->uk_freef = uma_small_free;
1464 if (booted < UMA_STARTUP)
1465 keg->uk_allocf = startup_alloc;
1467 if (booted < UMA_STARTUP2)
1468 keg->uk_allocf = startup_alloc;
1470 } else if (booted < UMA_STARTUP2 &&
1471 (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1472 keg->uk_allocf = startup_alloc;
1475 * Initialize keg's lock
1477 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1480 * If we're putting the slab header in the actual page we need to
1481 * figure out where in each page it goes. This calculates a right
1482 * justified offset into the memory on an ALIGN_PTR boundary.
1484 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1487 /* Size of the slab struct and free list */
1488 totsize = sizeof(struct uma_slab);
1490 if (totsize & UMA_ALIGN_PTR)
1491 totsize = (totsize & ~UMA_ALIGN_PTR) +
1492 (UMA_ALIGN_PTR + 1);
1493 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1496 * The only way the following is possible is if with our
1497 * UMA_ALIGN_PTR adjustments we are now bigger than
1498 * UMA_SLAB_SIZE. I haven't checked whether this is
1499 * mathematically possible for all cases, so we make
1502 totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1503 if (totsize > PAGE_SIZE * keg->uk_ppera) {
1504 printf("zone %s ipers %d rsize %d size %d\n",
1505 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1507 panic("UMA slab won't fit.");
1511 if (keg->uk_flags & UMA_ZONE_HASH)
1512 hash_alloc(&keg->uk_hash);
1515 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1516 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1517 keg->uk_ipers, keg->uk_ppera,
1518 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free,
1522 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1524 rw_wlock(&uma_rwlock);
1525 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1526 rw_wunlock(&uma_rwlock);
1531 * Zone header ctor. This initializes all fields, locks, etc.
1533 * Arguments/Returns follow uma_ctor specifications
1534 * udata Actually uma_zctor_args
1537 zone_ctor(void *mem, int size, void *udata, int flags)
1539 struct uma_zctor_args *arg = udata;
1540 uma_zone_t zone = mem;
1545 zone->uz_name = arg->name;
1546 zone->uz_ctor = arg->ctor;
1547 zone->uz_dtor = arg->dtor;
1548 zone->uz_slab = zone_fetch_slab;
1549 zone->uz_init = NULL;
1550 zone->uz_fini = NULL;
1551 zone->uz_allocs = 0;
1554 zone->uz_sleeps = 0;
1556 zone->uz_count_min = 0;
1558 zone->uz_warning = NULL;
1559 timevalclear(&zone->uz_ratecheck);
1562 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1565 * This is a pure cache zone, no kegs.
1568 if (arg->flags & UMA_ZONE_VM)
1569 arg->flags |= UMA_ZFLAG_CACHEONLY;
1570 zone->uz_flags = arg->flags;
1571 zone->uz_size = arg->size;
1572 zone->uz_import = arg->import;
1573 zone->uz_release = arg->release;
1574 zone->uz_arg = arg->arg;
1575 zone->uz_lockptr = &zone->uz_lock;
1576 rw_wlock(&uma_rwlock);
1577 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1578 rw_wunlock(&uma_rwlock);
1583 * Use the regular zone/keg/slab allocator.
1585 zone->uz_import = (uma_import)zone_import;
1586 zone->uz_release = (uma_release)zone_release;
1587 zone->uz_arg = zone;
1589 if (arg->flags & UMA_ZONE_SECONDARY) {
1590 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1591 zone->uz_init = arg->uminit;
1592 zone->uz_fini = arg->fini;
1593 zone->uz_lockptr = &keg->uk_lock;
1594 zone->uz_flags |= UMA_ZONE_SECONDARY;
1595 rw_wlock(&uma_rwlock);
1597 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1598 if (LIST_NEXT(z, uz_link) == NULL) {
1599 LIST_INSERT_AFTER(z, zone, uz_link);
1604 rw_wunlock(&uma_rwlock);
1605 } else if (keg == NULL) {
1606 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1607 arg->align, arg->flags)) == NULL)
1610 struct uma_kctor_args karg;
1613 /* We should only be here from uma_startup() */
1614 karg.size = arg->size;
1615 karg.uminit = arg->uminit;
1616 karg.fini = arg->fini;
1617 karg.align = arg->align;
1618 karg.flags = arg->flags;
1620 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1627 * Link in the first keg.
1629 zone->uz_klink.kl_keg = keg;
1630 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1631 zone->uz_lockptr = &keg->uk_lock;
1632 zone->uz_size = keg->uk_size;
1633 zone->uz_flags |= (keg->uk_flags &
1634 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1637 * Some internal zones don't have room allocated for the per cpu
1638 * caches. If we're internal, bail out here.
1640 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1641 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1642 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1647 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
1648 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
1649 ("Invalid zone flag combination"));
1650 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
1651 zone->uz_count = BUCKET_MAX;
1652 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
1655 zone->uz_count = bucket_select(zone->uz_size);
1656 zone->uz_count_min = zone->uz_count;
1662 * Keg header dtor. This frees all data, destroys locks, frees the hash
1663 * table and removes the keg from the global list.
1665 * Arguments/Returns follow uma_dtor specifications
1669 keg_dtor(void *arg, int size, void *udata)
1673 keg = (uma_keg_t)arg;
1675 if (keg->uk_free != 0) {
1676 printf("Freed UMA keg (%s) was not empty (%d items). "
1677 " Lost %d pages of memory.\n",
1678 keg->uk_name ? keg->uk_name : "",
1679 keg->uk_free, keg->uk_pages);
1683 hash_free(&keg->uk_hash);
1691 * Arguments/Returns follow uma_dtor specifications
1695 zone_dtor(void *arg, int size, void *udata)
1701 zone = (uma_zone_t)arg;
1702 keg = zone_first_keg(zone);
1704 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1707 rw_wlock(&uma_rwlock);
1708 LIST_REMOVE(zone, uz_link);
1709 rw_wunlock(&uma_rwlock);
1711 * XXX there are some races here where
1712 * the zone can be drained but zone lock
1713 * released and then refilled before we
1714 * remove it... we dont care for now
1716 zone_drain_wait(zone, M_WAITOK);
1718 * Unlink all of our kegs.
1720 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1721 klink->kl_keg = NULL;
1722 LIST_REMOVE(klink, kl_link);
1723 if (klink == &zone->uz_klink)
1725 free(klink, M_TEMP);
1728 * We only destroy kegs from non secondary zones.
1730 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1731 rw_wlock(&uma_rwlock);
1732 LIST_REMOVE(keg, uk_link);
1733 rw_wunlock(&uma_rwlock);
1734 zone_free_item(kegs, keg, NULL, SKIP_NONE);
1736 ZONE_LOCK_FINI(zone);
1740 * Traverses every zone in the system and calls a callback
1743 * zfunc A pointer to a function which accepts a zone
1750 zone_foreach(void (*zfunc)(uma_zone_t))
1755 rw_rlock(&uma_rwlock);
1756 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1757 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1760 rw_runlock(&uma_rwlock);
1763 /* Public functions */
1766 uma_startup(void *bootmem, int boot_pages)
1768 struct uma_zctor_args args;
1773 printf("Creating uma keg headers zone and keg.\n");
1775 rw_init(&uma_rwlock, "UMA lock");
1777 /* "manually" create the initial zone */
1778 memset(&args, 0, sizeof(args));
1779 args.name = "UMA Kegs";
1780 args.size = sizeof(struct uma_keg);
1781 args.ctor = keg_ctor;
1782 args.dtor = keg_dtor;
1783 args.uminit = zero_init;
1785 args.keg = &masterkeg;
1786 args.align = 32 - 1;
1787 args.flags = UMA_ZFLAG_INTERNAL;
1788 /* The initial zone has no Per cpu queues so it's smaller */
1789 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1792 printf("Filling boot free list.\n");
1794 for (i = 0; i < boot_pages; i++) {
1795 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1796 slab->us_data = (uint8_t *)slab;
1797 slab->us_flags = UMA_SLAB_BOOT;
1798 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1800 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1803 printf("Creating uma zone headers zone and keg.\n");
1805 args.name = "UMA Zones";
1806 args.size = sizeof(struct uma_zone) +
1807 (sizeof(struct uma_cache) * (mp_maxid + 1));
1808 args.ctor = zone_ctor;
1809 args.dtor = zone_dtor;
1810 args.uminit = zero_init;
1813 args.align = 32 - 1;
1814 args.flags = UMA_ZFLAG_INTERNAL;
1815 /* The initial zone has no Per cpu queues so it's smaller */
1816 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1819 printf("Creating slab and hash zones.\n");
1822 /* Now make a zone for slab headers */
1823 slabzone = uma_zcreate("UMA Slabs",
1824 sizeof(struct uma_slab),
1825 NULL, NULL, NULL, NULL,
1826 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1828 hashzone = uma_zcreate("UMA Hash",
1829 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1830 NULL, NULL, NULL, NULL,
1831 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1835 booted = UMA_STARTUP;
1838 printf("UMA startup complete.\n");
1846 booted = UMA_STARTUP2;
1848 sx_init(&uma_drain_lock, "umadrain");
1850 printf("UMA startup2 complete.\n");
1855 * Initialize our callout handle
1863 printf("Starting callout.\n");
1865 callout_init(&uma_callout, 1);
1866 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1868 printf("UMA startup3 complete.\n");
1873 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1874 int align, uint32_t flags)
1876 struct uma_kctor_args args;
1879 args.uminit = uminit;
1881 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1884 return (zone_alloc_item(kegs, &args, M_WAITOK));
1889 uma_set_align(int align)
1892 if (align != UMA_ALIGN_CACHE)
1893 uma_align_cache = align;
1898 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1899 uma_init uminit, uma_fini fini, int align, uint32_t flags)
1902 struct uma_zctor_args args;
1906 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
1909 /* This stuff is essential for the zone ctor */
1910 memset(&args, 0, sizeof(args));
1915 args.uminit = uminit;
1919 * If a zone is being created with an empty constructor and
1920 * destructor, pass UMA constructor/destructor which checks for
1921 * memory use after free.
1923 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) &&
1924 ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) {
1925 args.ctor = trash_ctor;
1926 args.dtor = trash_dtor;
1927 args.uminit = trash_init;
1928 args.fini = trash_fini;
1935 if (booted < UMA_STARTUP2) {
1938 sx_slock(&uma_drain_lock);
1941 res = zone_alloc_item(zones, &args, M_WAITOK);
1943 sx_sunlock(&uma_drain_lock);
1949 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1950 uma_init zinit, uma_fini zfini, uma_zone_t master)
1952 struct uma_zctor_args args;
1957 keg = zone_first_keg(master);
1958 memset(&args, 0, sizeof(args));
1960 args.size = keg->uk_size;
1963 args.uminit = zinit;
1965 args.align = keg->uk_align;
1966 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1969 if (booted < UMA_STARTUP2) {
1972 sx_slock(&uma_drain_lock);
1975 /* XXX Attaches only one keg of potentially many. */
1976 res = zone_alloc_item(zones, &args, M_WAITOK);
1978 sx_sunlock(&uma_drain_lock);
1984 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1985 uma_init zinit, uma_fini zfini, uma_import zimport,
1986 uma_release zrelease, void *arg, int flags)
1988 struct uma_zctor_args args;
1990 memset(&args, 0, sizeof(args));
1995 args.uminit = zinit;
1997 args.import = zimport;
1998 args.release = zrelease;
2003 return (zone_alloc_item(zones, &args, M_WAITOK));
2007 zone_lock_pair(uma_zone_t a, uma_zone_t b)
2011 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
2014 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
2019 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2027 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2034 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2036 zone_lock_pair(zone, master);
2038 * zone must use vtoslab() to resolve objects and must already be
2041 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2042 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2047 * The new master must also use vtoslab().
2049 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2055 * The underlying object must be the same size. rsize
2058 if (master->uz_size != zone->uz_size) {
2063 * Put it at the end of the list.
2065 klink->kl_keg = zone_first_keg(master);
2066 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2067 if (LIST_NEXT(kl, kl_link) == NULL) {
2068 LIST_INSERT_AFTER(kl, klink, kl_link);
2073 zone->uz_flags |= UMA_ZFLAG_MULTI;
2074 zone->uz_slab = zone_fetch_slab_multi;
2077 zone_unlock_pair(zone, master);
2079 free(klink, M_TEMP);
2087 uma_zdestroy(uma_zone_t zone)
2090 sx_slock(&uma_drain_lock);
2091 zone_free_item(zones, zone, NULL, SKIP_NONE);
2092 sx_sunlock(&uma_drain_lock);
2096 uma_zwait(uma_zone_t zone)
2100 item = uma_zalloc_arg(zone, NULL, M_WAITOK);
2101 uma_zfree(zone, item);
2106 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2110 uma_bucket_t bucket;
2114 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2115 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2117 /* This is the fast path allocation */
2118 #ifdef UMA_DEBUG_ALLOC_1
2119 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2121 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2122 zone->uz_name, flags);
2124 if (flags & M_WAITOK) {
2125 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2126 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2128 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2129 ("uma_zalloc_arg: called with spinlock or critical section held"));
2131 #ifdef DEBUG_MEMGUARD
2132 if (memguard_cmp_zone(zone)) {
2133 item = memguard_alloc(zone->uz_size, flags);
2135 if (zone->uz_init != NULL &&
2136 zone->uz_init(item, zone->uz_size, flags) != 0)
2138 if (zone->uz_ctor != NULL &&
2139 zone->uz_ctor(item, zone->uz_size, udata,
2141 zone->uz_fini(item, zone->uz_size);
2146 /* This is unfortunate but should not be fatal. */
2150 * If possible, allocate from the per-CPU cache. There are two
2151 * requirements for safe access to the per-CPU cache: (1) the thread
2152 * accessing the cache must not be preempted or yield during access,
2153 * and (2) the thread must not migrate CPUs without switching which
2154 * cache it accesses. We rely on a critical section to prevent
2155 * preemption and migration. We release the critical section in
2156 * order to acquire the zone mutex if we are unable to allocate from
2157 * the current cache; when we re-acquire the critical section, we
2158 * must detect and handle migration if it has occurred.
2162 cache = &zone->uz_cpu[cpu];
2165 bucket = cache->uc_allocbucket;
2166 if (bucket != NULL && bucket->ub_cnt > 0) {
2168 item = bucket->ub_bucket[bucket->ub_cnt];
2170 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2172 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2175 if (zone->uz_ctor != NULL &&
2176 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2177 atomic_add_long(&zone->uz_fails, 1);
2178 zone_free_item(zone, item, udata, SKIP_DTOR);
2182 uma_dbg_alloc(zone, NULL, item);
2185 uma_zero_item(item, zone);
2190 * We have run out of items in our alloc bucket.
2191 * See if we can switch with our free bucket.
2193 bucket = cache->uc_freebucket;
2194 if (bucket != NULL && bucket->ub_cnt > 0) {
2195 #ifdef UMA_DEBUG_ALLOC
2196 printf("uma_zalloc: Swapping empty with alloc.\n");
2198 cache->uc_freebucket = cache->uc_allocbucket;
2199 cache->uc_allocbucket = bucket;
2204 * Discard any empty allocation bucket while we hold no locks.
2206 bucket = cache->uc_allocbucket;
2207 cache->uc_allocbucket = NULL;
2210 bucket_free(zone, bucket, udata);
2212 /* Short-circuit for zones without buckets and low memory. */
2213 if (zone->uz_count == 0 || bucketdisable)
2217 * Attempt to retrieve the item from the per-CPU cache has failed, so
2218 * we must go back to the zone. This requires the zone lock, so we
2219 * must drop the critical section, then re-acquire it when we go back
2220 * to the cache. Since the critical section is released, we may be
2221 * preempted or migrate. As such, make sure not to maintain any
2222 * thread-local state specific to the cache from prior to releasing
2223 * the critical section.
2226 if (ZONE_TRYLOCK(zone) == 0) {
2227 /* Record contention to size the buckets. */
2233 cache = &zone->uz_cpu[cpu];
2236 * Since we have locked the zone we may as well send back our stats.
2238 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2239 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2240 cache->uc_allocs = 0;
2241 cache->uc_frees = 0;
2243 /* See if we lost the race to fill the cache. */
2244 if (cache->uc_allocbucket != NULL) {
2250 * Check the zone's cache of buckets.
2252 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2253 KASSERT(bucket->ub_cnt != 0,
2254 ("uma_zalloc_arg: Returning an empty bucket."));
2256 LIST_REMOVE(bucket, ub_link);
2257 cache->uc_allocbucket = bucket;
2261 /* We are no longer associated with this CPU. */
2265 * We bump the uz count when the cache size is insufficient to
2266 * handle the working set.
2268 if (lockfail && zone->uz_count < BUCKET_MAX)
2273 * Now lets just fill a bucket and put it on the free list. If that
2274 * works we'll restart the allocation from the beginning and it
2275 * will use the just filled bucket.
2277 bucket = zone_alloc_bucket(zone, udata, flags);
2278 if (bucket != NULL) {
2282 cache = &zone->uz_cpu[cpu];
2284 * See if we lost the race or were migrated. Cache the
2285 * initialized bucket to make this less likely or claim
2286 * the memory directly.
2288 if (cache->uc_allocbucket == NULL)
2289 cache->uc_allocbucket = bucket;
2291 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2297 * We may not be able to get a bucket so return an actual item.
2300 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2304 item = zone_alloc_item(zone, udata, flags);
2310 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2315 mtx_assert(&keg->uk_lock, MA_OWNED);
2318 if ((flags & M_USE_RESERVE) == 0)
2319 reserve = keg->uk_reserve;
2323 * Find a slab with some space. Prefer slabs that are partially
2324 * used over those that are totally full. This helps to reduce
2327 if (keg->uk_free > reserve) {
2328 if (!LIST_EMPTY(&keg->uk_part_slab)) {
2329 slab = LIST_FIRST(&keg->uk_part_slab);
2331 slab = LIST_FIRST(&keg->uk_free_slab);
2332 LIST_REMOVE(slab, us_link);
2333 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2336 MPASS(slab->us_keg == keg);
2341 * M_NOVM means don't ask at all!
2346 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2347 keg->uk_flags |= UMA_ZFLAG_FULL;
2349 * If this is not a multi-zone, set the FULL bit.
2350 * Otherwise slab_multi() takes care of it.
2352 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2353 zone->uz_flags |= UMA_ZFLAG_FULL;
2354 zone_log_warning(zone);
2355 zone_maxaction(zone);
2357 if (flags & M_NOWAIT)
2360 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2363 slab = keg_alloc_slab(keg, zone, flags);
2365 * If we got a slab here it's safe to mark it partially used
2366 * and return. We assume that the caller is going to remove
2367 * at least one item.
2370 MPASS(slab->us_keg == keg);
2371 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2375 * We might not have been able to get a slab but another cpu
2376 * could have while we were unlocked. Check again before we
2385 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2390 keg = zone_first_keg(zone);
2395 slab = keg_fetch_slab(keg, zone, flags);
2398 if (flags & (M_NOWAIT | M_NOVM))
2406 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2407 * with the keg locked. On NULL no lock is held.
2409 * The last pointer is used to seed the search. It is not required.
2412 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2422 * Don't wait on the first pass. This will skip limit tests
2423 * as well. We don't want to block if we can find a provider
2426 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2428 * Use the last slab allocated as a hint for where to start
2432 slab = keg_fetch_slab(last, zone, flags);
2438 * Loop until we have a slab incase of transient failures
2439 * while M_WAITOK is specified. I'm not sure this is 100%
2440 * required but we've done it for so long now.
2446 * Search the available kegs for slabs. Be careful to hold the
2447 * correct lock while calling into the keg layer.
2449 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2450 keg = klink->kl_keg;
2452 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2453 slab = keg_fetch_slab(keg, zone, flags);
2457 if (keg->uk_flags & UMA_ZFLAG_FULL)
2463 if (rflags & (M_NOWAIT | M_NOVM))
2467 * All kegs are full. XXX We can't atomically check all kegs
2468 * and sleep so just sleep for a short period and retry.
2470 if (full && !empty) {
2472 zone->uz_flags |= UMA_ZFLAG_FULL;
2474 zone_log_warning(zone);
2475 zone_maxaction(zone);
2476 msleep(zone, zone->uz_lockptr, PVM,
2477 "zonelimit", hz/100);
2478 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2487 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2492 MPASS(keg == slab->us_keg);
2493 mtx_assert(&keg->uk_lock, MA_OWNED);
2495 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2496 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2497 item = slab->us_data + (keg->uk_rsize * freei);
2498 slab->us_freecount--;
2501 /* Move this slab to the full list */
2502 if (slab->us_freecount == 0) {
2503 LIST_REMOVE(slab, us_link);
2504 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2511 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2519 /* Try to keep the buckets totally full */
2520 for (i = 0; i < max; ) {
2521 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2524 while (slab->us_freecount && i < max) {
2525 bucket[i++] = slab_alloc_item(keg, slab);
2526 if (keg->uk_free <= keg->uk_reserve)
2529 /* Don't grab more than one slab at a time. */
2540 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2542 uma_bucket_t bucket;
2545 /* Don't wait for buckets, preserve caller's NOVM setting. */
2546 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2550 max = MIN(bucket->ub_entries, zone->uz_count);
2551 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2555 * Initialize the memory if necessary.
2557 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2560 for (i = 0; i < bucket->ub_cnt; i++)
2561 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2565 * If we couldn't initialize the whole bucket, put the
2566 * rest back onto the freelist.
2568 if (i != bucket->ub_cnt) {
2569 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2570 bucket->ub_cnt - i);
2572 bzero(&bucket->ub_bucket[i],
2573 sizeof(void *) * (bucket->ub_cnt - i));
2579 if (bucket->ub_cnt == 0) {
2580 bucket_free(zone, bucket, udata);
2581 atomic_add_long(&zone->uz_fails, 1);
2589 * Allocates a single item from a zone.
2592 * zone The zone to alloc for.
2593 * udata The data to be passed to the constructor.
2594 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2597 * NULL if there is no memory and M_NOWAIT is set
2598 * An item if successful
2602 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2608 #ifdef UMA_DEBUG_ALLOC
2609 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2611 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2613 atomic_add_long(&zone->uz_allocs, 1);
2616 * We have to call both the zone's init (not the keg's init)
2617 * and the zone's ctor. This is because the item is going from
2618 * a keg slab directly to the user, and the user is expecting it
2619 * to be both zone-init'd as well as zone-ctor'd.
2621 if (zone->uz_init != NULL) {
2622 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2623 zone_free_item(zone, item, udata, SKIP_FINI);
2627 if (zone->uz_ctor != NULL) {
2628 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2629 zone_free_item(zone, item, udata, SKIP_DTOR);
2634 uma_dbg_alloc(zone, NULL, item);
2637 uma_zero_item(item, zone);
2642 atomic_add_long(&zone->uz_fails, 1);
2648 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2651 uma_bucket_t bucket;
2655 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2656 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2658 #ifdef UMA_DEBUG_ALLOC_1
2659 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2661 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2664 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2665 ("uma_zfree_arg: called with spinlock or critical section held"));
2667 /* uma_zfree(..., NULL) does nothing, to match free(9). */
2670 #ifdef DEBUG_MEMGUARD
2671 if (is_memguard_addr(item)) {
2672 if (zone->uz_dtor != NULL)
2673 zone->uz_dtor(item, zone->uz_size, udata);
2674 if (zone->uz_fini != NULL)
2675 zone->uz_fini(item, zone->uz_size);
2676 memguard_free(item);
2681 if (zone->uz_flags & UMA_ZONE_MALLOC)
2682 uma_dbg_free(zone, udata, item);
2684 uma_dbg_free(zone, NULL, item);
2686 if (zone->uz_dtor != NULL)
2687 zone->uz_dtor(item, zone->uz_size, udata);
2690 * The race here is acceptable. If we miss it we'll just have to wait
2691 * a little longer for the limits to be reset.
2693 if (zone->uz_flags & UMA_ZFLAG_FULL)
2697 * If possible, free to the per-CPU cache. There are two
2698 * requirements for safe access to the per-CPU cache: (1) the thread
2699 * accessing the cache must not be preempted or yield during access,
2700 * and (2) the thread must not migrate CPUs without switching which
2701 * cache it accesses. We rely on a critical section to prevent
2702 * preemption and migration. We release the critical section in
2703 * order to acquire the zone mutex if we are unable to free to the
2704 * current cache; when we re-acquire the critical section, we must
2705 * detect and handle migration if it has occurred.
2710 cache = &zone->uz_cpu[cpu];
2714 * Try to free into the allocbucket first to give LIFO ordering
2715 * for cache-hot datastructures. Spill over into the freebucket
2716 * if necessary. Alloc will swap them if one runs dry.
2718 bucket = cache->uc_allocbucket;
2719 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2720 bucket = cache->uc_freebucket;
2721 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2722 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2723 ("uma_zfree: Freeing to non free bucket index."));
2724 bucket->ub_bucket[bucket->ub_cnt] = item;
2732 * We must go back the zone, which requires acquiring the zone lock,
2733 * which in turn means we must release and re-acquire the critical
2734 * section. Since the critical section is released, we may be
2735 * preempted or migrate. As such, make sure not to maintain any
2736 * thread-local state specific to the cache from prior to releasing
2737 * the critical section.
2740 if (zone->uz_count == 0 || bucketdisable)
2744 if (ZONE_TRYLOCK(zone) == 0) {
2745 /* Record contention to size the buckets. */
2751 cache = &zone->uz_cpu[cpu];
2754 * Since we have locked the zone we may as well send back our stats.
2756 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2757 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2758 cache->uc_allocs = 0;
2759 cache->uc_frees = 0;
2761 bucket = cache->uc_freebucket;
2762 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2766 cache->uc_freebucket = NULL;
2767 /* We are no longer associated with this CPU. */
2770 /* Can we throw this on the zone full list? */
2771 if (bucket != NULL) {
2772 #ifdef UMA_DEBUG_ALLOC
2773 printf("uma_zfree: Putting old bucket on the free list.\n");
2775 /* ub_cnt is pointing to the last free item */
2776 KASSERT(bucket->ub_cnt != 0,
2777 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2778 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2782 * We bump the uz count when the cache size is insufficient to
2783 * handle the working set.
2785 if (lockfail && zone->uz_count < BUCKET_MAX)
2789 #ifdef UMA_DEBUG_ALLOC
2790 printf("uma_zfree: Allocating new free bucket.\n");
2792 bucket = bucket_alloc(zone, udata, M_NOWAIT);
2796 cache = &zone->uz_cpu[cpu];
2797 if (cache->uc_freebucket == NULL) {
2798 cache->uc_freebucket = bucket;
2802 * We lost the race, start over. We have to drop our
2803 * critical section to free the bucket.
2806 bucket_free(zone, bucket, udata);
2811 * If nothing else caught this, we'll just do an internal free.
2814 zone_free_item(zone, item, udata, SKIP_DTOR);
2820 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2824 mtx_assert(&keg->uk_lock, MA_OWNED);
2825 MPASS(keg == slab->us_keg);
2827 /* Do we need to remove from any lists? */
2828 if (slab->us_freecount+1 == keg->uk_ipers) {
2829 LIST_REMOVE(slab, us_link);
2830 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2831 } else if (slab->us_freecount == 0) {
2832 LIST_REMOVE(slab, us_link);
2833 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2836 /* Slab management. */
2837 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2838 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2839 slab->us_freecount++;
2841 /* Keg statistics. */
2846 zone_release(uma_zone_t zone, void **bucket, int cnt)
2856 keg = zone_first_keg(zone);
2858 for (i = 0; i < cnt; i++) {
2860 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2861 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2862 if (zone->uz_flags & UMA_ZONE_HASH) {
2863 slab = hash_sfind(&keg->uk_hash, mem);
2865 mem += keg->uk_pgoff;
2866 slab = (uma_slab_t)mem;
2869 slab = vtoslab((vm_offset_t)item);
2870 if (slab->us_keg != keg) {
2876 slab_free_item(keg, slab, item);
2877 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2878 if (keg->uk_pages < keg->uk_maxpages) {
2879 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2884 * We can handle one more allocation. Since we're
2885 * clearing ZFLAG_FULL, wake up all procs blocked
2886 * on pages. This should be uncommon, so keeping this
2887 * simple for now (rather than adding count of blocked
2896 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2904 * Frees a single item to any zone.
2907 * zone The zone to free to
2908 * item The item we're freeing
2909 * udata User supplied data for the dtor
2910 * skip Skip dtors and finis
2913 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2917 if (skip == SKIP_NONE) {
2918 if (zone->uz_flags & UMA_ZONE_MALLOC)
2919 uma_dbg_free(zone, udata, item);
2921 uma_dbg_free(zone, NULL, item);
2924 if (skip < SKIP_DTOR && zone->uz_dtor)
2925 zone->uz_dtor(item, zone->uz_size, udata);
2927 if (skip < SKIP_FINI && zone->uz_fini)
2928 zone->uz_fini(item, zone->uz_size);
2930 atomic_add_long(&zone->uz_frees, 1);
2931 zone->uz_release(zone->uz_arg, &item, 1);
2936 uma_zone_set_max(uma_zone_t zone, int nitems)
2940 keg = zone_first_keg(zone);
2944 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2945 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2946 keg->uk_maxpages += keg->uk_ppera;
2947 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers;
2955 uma_zone_get_max(uma_zone_t zone)
2960 keg = zone_first_keg(zone);
2964 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers;
2972 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2976 zone->uz_warning = warning;
2982 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
2986 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
2992 uma_zone_get_cur(uma_zone_t zone)
2998 nitems = zone->uz_allocs - zone->uz_frees;
3001 * See the comment in sysctl_vm_zone_stats() regarding the
3002 * safety of accessing the per-cpu caches. With the zone lock
3003 * held, it is safe, but can potentially result in stale data.
3005 nitems += zone->uz_cpu[i].uc_allocs -
3006 zone->uz_cpu[i].uc_frees;
3010 return (nitems < 0 ? 0 : nitems);
3015 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
3019 keg = zone_first_keg(zone);
3020 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3022 KASSERT(keg->uk_pages == 0,
3023 ("uma_zone_set_init on non-empty keg"));
3024 keg->uk_init = uminit;
3030 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3034 keg = zone_first_keg(zone);
3035 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
3037 KASSERT(keg->uk_pages == 0,
3038 ("uma_zone_set_fini on non-empty keg"));
3039 keg->uk_fini = fini;
3045 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3049 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3050 ("uma_zone_set_zinit on non-empty keg"));
3051 zone->uz_init = zinit;
3057 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3061 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3062 ("uma_zone_set_zfini on non-empty keg"));
3063 zone->uz_fini = zfini;
3068 /* XXX uk_freef is not actually used with the zone locked */
3070 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3074 keg = zone_first_keg(zone);
3075 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
3077 keg->uk_freef = freef;
3082 /* XXX uk_allocf is not actually used with the zone locked */
3084 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3088 keg = zone_first_keg(zone);
3090 keg->uk_allocf = allocf;
3096 uma_zone_reserve(uma_zone_t zone, int items)
3100 keg = zone_first_keg(zone);
3104 keg->uk_reserve = items;
3112 uma_zone_reserve_kva(uma_zone_t zone, int count)
3118 keg = zone_first_keg(zone);
3121 pages = count / keg->uk_ipers;
3123 if (pages * keg->uk_ipers < count)
3125 pages *= keg->uk_ppera;
3127 #ifdef UMA_MD_SMALL_ALLOC
3128 if (keg->uk_ppera > 1) {
3132 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
3140 keg->uk_maxpages = pages;
3141 #ifdef UMA_MD_SMALL_ALLOC
3142 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3144 keg->uk_allocf = noobj_alloc;
3146 keg->uk_flags |= UMA_ZONE_NOFREE;
3154 uma_prealloc(uma_zone_t zone, int items)
3160 keg = zone_first_keg(zone);
3164 slabs = items / keg->uk_ipers;
3165 if (slabs * keg->uk_ipers < items)
3168 slab = keg_alloc_slab(keg, zone, M_WAITOK);
3171 MPASS(slab->us_keg == keg);
3172 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3180 uma_reclaim_locked(bool kmem_danger)
3184 printf("UMA: vm asked us to release pages!\n");
3186 sx_assert(&uma_drain_lock, SA_XLOCKED);
3188 zone_foreach(zone_drain);
3189 if (vm_page_count_min() || kmem_danger) {
3190 cache_drain_safe(NULL);
3191 zone_foreach(zone_drain);
3194 * Some slabs may have been freed but this zone will be visited early
3195 * we visit again so that we can free pages that are empty once other
3196 * zones are drained. We have to do the same for buckets.
3198 zone_drain(slabzone);
3199 bucket_zone_drain();
3206 sx_xlock(&uma_drain_lock);
3207 uma_reclaim_locked(false);
3208 sx_xunlock(&uma_drain_lock);
3211 static int uma_reclaim_needed;
3214 uma_reclaim_wakeup(void)
3217 uma_reclaim_needed = 1;
3218 wakeup(&uma_reclaim_needed);
3222 uma_reclaim_worker(void *arg __unused)
3225 sx_xlock(&uma_drain_lock);
3227 sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM,
3229 if (uma_reclaim_needed) {
3230 uma_reclaim_needed = 0;
3231 sx_xunlock(&uma_drain_lock);
3232 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
3233 sx_xlock(&uma_drain_lock);
3234 uma_reclaim_locked(true);
3241 uma_zone_exhausted(uma_zone_t zone)
3246 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3252 uma_zone_exhausted_nolock(uma_zone_t zone)
3254 return (zone->uz_flags & UMA_ZFLAG_FULL);
3258 uma_large_malloc(vm_size_t size, int wait)
3264 slab = zone_alloc_item(slabzone, NULL, wait);
3267 mem = page_alloc(NULL, size, &flags, wait);
3269 vsetslab((vm_offset_t)mem, slab);
3270 slab->us_data = mem;
3271 slab->us_flags = flags | UMA_SLAB_MALLOC;
3272 slab->us_size = size;
3274 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3281 uma_large_free(uma_slab_t slab)
3284 page_free(slab->us_data, slab->us_size, slab->us_flags);
3285 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3289 uma_zero_item(void *item, uma_zone_t zone)
3293 if (zone->uz_flags & UMA_ZONE_PCPU) {
3295 bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3297 bzero(item, zone->uz_size);
3301 uma_print_stats(void)
3303 zone_foreach(uma_print_zone);
3307 slab_print(uma_slab_t slab)
3309 printf("slab: keg %p, data %p, freecount %d\n",
3310 slab->us_keg, slab->us_data, slab->us_freecount);
3314 cache_print(uma_cache_t cache)
3316 printf("alloc: %p(%d), free: %p(%d)\n",
3317 cache->uc_allocbucket,
3318 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3319 cache->uc_freebucket,
3320 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3324 uma_print_keg(uma_keg_t keg)
3328 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3329 "out %d free %d limit %d\n",
3330 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3331 keg->uk_ipers, keg->uk_ppera,
3332 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free,
3333 keg->uk_free, (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3334 printf("Part slabs:\n");
3335 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3337 printf("Free slabs:\n");
3338 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3340 printf("Full slabs:\n");
3341 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3346 uma_print_zone(uma_zone_t zone)
3352 printf("zone: %s(%p) size %d flags %#x\n",
3353 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3354 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3355 uma_print_keg(kl->kl_keg);
3357 cache = &zone->uz_cpu[i];
3358 printf("CPU %d Cache:\n", i);
3365 * Generate statistics across both the zone and its per-cpu cache's. Return
3366 * desired statistics if the pointer is non-NULL for that statistic.
3368 * Note: does not update the zone statistics, as it can't safely clear the
3369 * per-CPU cache statistic.
3371 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3372 * safe from off-CPU; we should modify the caches to track this information
3373 * directly so that we don't have to.
3376 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3377 uint64_t *freesp, uint64_t *sleepsp)
3380 uint64_t allocs, frees, sleeps;
3383 allocs = frees = sleeps = 0;
3386 cache = &z->uz_cpu[cpu];
3387 if (cache->uc_allocbucket != NULL)
3388 cachefree += cache->uc_allocbucket->ub_cnt;
3389 if (cache->uc_freebucket != NULL)
3390 cachefree += cache->uc_freebucket->ub_cnt;
3391 allocs += cache->uc_allocs;
3392 frees += cache->uc_frees;
3394 allocs += z->uz_allocs;
3395 frees += z->uz_frees;
3396 sleeps += z->uz_sleeps;
3397 if (cachefreep != NULL)
3398 *cachefreep = cachefree;
3399 if (allocsp != NULL)
3403 if (sleepsp != NULL)
3409 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3416 rw_rlock(&uma_rwlock);
3417 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3418 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3421 rw_runlock(&uma_rwlock);
3422 return (sysctl_handle_int(oidp, &count, 0, req));
3426 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3428 struct uma_stream_header ush;
3429 struct uma_type_header uth;
3430 struct uma_percpu_stat *ups;
3431 uma_bucket_t bucket;
3438 int count, error, i;
3440 error = sysctl_wire_old_buffer(req, 0);
3443 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3444 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3445 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
3448 rw_rlock(&uma_rwlock);
3449 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3450 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3455 * Insert stream header.
3457 bzero(&ush, sizeof(ush));
3458 ush.ush_version = UMA_STREAM_VERSION;
3459 ush.ush_maxcpus = (mp_maxid + 1);
3460 ush.ush_count = count;
3461 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3463 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3464 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3465 bzero(&uth, sizeof(uth));
3467 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3468 uth.uth_align = kz->uk_align;
3469 uth.uth_size = kz->uk_size;
3470 uth.uth_rsize = kz->uk_rsize;
3471 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3473 uth.uth_maxpages += k->uk_maxpages;
3474 uth.uth_pages += k->uk_pages;
3475 uth.uth_keg_free += k->uk_free;
3476 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3481 * A zone is secondary is it is not the first entry
3482 * on the keg's zone list.
3484 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3485 (LIST_FIRST(&kz->uk_zones) != z))
3486 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3488 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3489 uth.uth_zone_free += bucket->ub_cnt;
3490 uth.uth_allocs = z->uz_allocs;
3491 uth.uth_frees = z->uz_frees;
3492 uth.uth_fails = z->uz_fails;
3493 uth.uth_sleeps = z->uz_sleeps;
3495 * While it is not normally safe to access the cache
3496 * bucket pointers while not on the CPU that owns the
3497 * cache, we only allow the pointers to be exchanged
3498 * without the zone lock held, not invalidated, so
3499 * accept the possible race associated with bucket
3500 * exchange during monitoring.
3502 for (i = 0; i < mp_maxid + 1; i++) {
3503 bzero(&ups[i], sizeof(*ups));
3504 if (kz->uk_flags & UMA_ZFLAG_INTERNAL ||
3507 cache = &z->uz_cpu[i];
3508 if (cache->uc_allocbucket != NULL)
3509 ups[i].ups_cache_free +=
3510 cache->uc_allocbucket->ub_cnt;
3511 if (cache->uc_freebucket != NULL)
3512 ups[i].ups_cache_free +=
3513 cache->uc_freebucket->ub_cnt;
3514 ups[i].ups_allocs = cache->uc_allocs;
3515 ups[i].ups_frees = cache->uc_frees;
3518 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3519 for (i = 0; i < mp_maxid + 1; i++)
3520 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
3523 rw_runlock(&uma_rwlock);
3524 error = sbuf_finish(&sbuf);
3531 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3533 uma_zone_t zone = *(uma_zone_t *)arg1;
3536 max = uma_zone_get_max(zone);
3537 error = sysctl_handle_int(oidp, &max, 0, req);
3538 if (error || !req->newptr)
3541 uma_zone_set_max(zone, max);
3547 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3549 uma_zone_t zone = *(uma_zone_t *)arg1;
3552 cur = uma_zone_get_cur(zone);
3553 return (sysctl_handle_int(oidp, &cur, 0, req));
3558 uma_dbg_getslab(uma_zone_t zone, void *item)
3564 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
3565 if (zone->uz_flags & UMA_ZONE_VTOSLAB) {
3566 slab = vtoslab((vm_offset_t)mem);
3569 * It is safe to return the slab here even though the
3570 * zone is unlocked because the item's allocation state
3571 * essentially holds a reference.
3574 keg = LIST_FIRST(&zone->uz_kegs)->kl_keg;
3575 if (keg->uk_flags & UMA_ZONE_HASH)
3576 slab = hash_sfind(&keg->uk_hash, mem);
3578 slab = (uma_slab_t)(mem + keg->uk_pgoff);
3586 * Set up the slab's freei data such that uma_dbg_free can function.
3590 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
3595 if (zone_first_keg(zone) == NULL)
3598 slab = uma_dbg_getslab(zone, item);
3600 panic("uma: item %p did not belong to zone %s\n",
3601 item, zone->uz_name);
3604 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3606 if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3607 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
3608 item, zone, zone->uz_name, slab, freei);
3609 BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3615 * Verifies freed addresses. Checks for alignment, valid slab membership
3616 * and duplicate frees.
3620 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
3625 if (zone_first_keg(zone) == NULL)
3628 slab = uma_dbg_getslab(zone, item);
3630 panic("uma: Freed item %p did not belong to zone %s\n",
3631 item, zone->uz_name);
3634 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3636 if (freei >= keg->uk_ipers)
3637 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
3638 item, zone, zone->uz_name, slab, freei);
3640 if (((freei * keg->uk_rsize) + slab->us_data) != item)
3641 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
3642 item, zone, zone->uz_name, slab, freei);
3644 if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3645 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
3646 item, zone, zone->uz_name, slab, freei);
3648 BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3650 #endif /* INVARIANTS */
3653 DB_SHOW_COMMAND(uma, db_show_uma)
3655 uint64_t allocs, frees, sleeps;
3656 uma_bucket_t bucket;
3661 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3662 "Free", "Requests", "Sleeps", "Bucket");
3663 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3664 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3665 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3666 allocs = z->uz_allocs;
3667 frees = z->uz_frees;
3668 sleeps = z->uz_sleeps;
3671 uma_zone_sumstat(z, &cachefree, &allocs,
3673 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3674 (LIST_FIRST(&kz->uk_zones) != z)))
3675 cachefree += kz->uk_free;
3676 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3677 cachefree += bucket->ub_cnt;
3678 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3679 z->uz_name, (uintmax_t)kz->uk_size,
3680 (intmax_t)(allocs - frees), cachefree,
3681 (uintmax_t)allocs, sleeps, z->uz_count);
3688 DB_SHOW_COMMAND(umacache, db_show_umacache)
3690 uint64_t allocs, frees;
3691 uma_bucket_t bucket;
3695 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3696 "Requests", "Bucket");
3697 LIST_FOREACH(z, &uma_cachezones, uz_link) {
3698 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3699 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3700 cachefree += bucket->ub_cnt;
3701 db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3702 z->uz_name, (uintmax_t)z->uz_size,
3703 (intmax_t)(allocs - frees), cachefree,
3704 (uintmax_t)allocs, z->uz_count);