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);
279 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
280 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
283 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
285 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
286 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
288 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
289 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
291 static int zone_warnings = 1;
292 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
293 "Warn when UMA zones becomes full");
296 * This routine checks to see whether or not it's safe to enable buckets.
301 bucketdisable = vm_page_count_min();
305 * Initialize bucket_zones, the array of zones of buckets of various sizes.
307 * For each zone, calculate the memory required for each bucket, consisting
308 * of the header and an array of pointers.
313 struct uma_bucket_zone *ubz;
316 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
317 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
318 size += sizeof(void *) * ubz->ubz_entries;
319 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
320 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
321 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
326 * Given a desired number of entries for a bucket, return the zone from which
327 * to allocate the bucket.
329 static struct uma_bucket_zone *
330 bucket_zone_lookup(int entries)
332 struct uma_bucket_zone *ubz;
334 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
335 if (ubz->ubz_entries >= entries)
342 bucket_select(int size)
344 struct uma_bucket_zone *ubz;
346 ubz = &bucket_zones[0];
347 if (size > ubz->ubz_maxsize)
348 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
350 for (; ubz->ubz_entries != 0; ubz++)
351 if (ubz->ubz_maxsize < size)
354 return (ubz->ubz_entries);
358 bucket_alloc(uma_zone_t zone, void *udata, int flags)
360 struct uma_bucket_zone *ubz;
364 * This is to stop us from allocating per cpu buckets while we're
365 * running out of vm.boot_pages. Otherwise, we would exhaust the
366 * boot pages. This also prevents us from allocating buckets in
367 * low memory situations.
372 * To limit bucket recursion we store the original zone flags
373 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
374 * NOVM flag to persist even through deep recursions. We also
375 * store ZFLAG_BUCKET once we have recursed attempting to allocate
376 * a bucket for a bucket zone so we do not allow infinite bucket
377 * recursion. This cookie will even persist to frees of unused
378 * buckets via the allocation path or bucket allocations in the
381 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
382 udata = (void *)(uintptr_t)zone->uz_flags;
384 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
386 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
388 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
390 ubz = bucket_zone_lookup(zone->uz_count);
391 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
393 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
396 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
399 bucket->ub_entries = ubz->ubz_entries;
406 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
408 struct uma_bucket_zone *ubz;
410 KASSERT(bucket->ub_cnt == 0,
411 ("bucket_free: Freeing a non free bucket."));
412 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
413 udata = (void *)(uintptr_t)zone->uz_flags;
414 ubz = bucket_zone_lookup(bucket->ub_entries);
415 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
419 bucket_zone_drain(void)
421 struct uma_bucket_zone *ubz;
423 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
424 zone_drain(ubz->ubz_zone);
428 zone_log_warning(uma_zone_t zone)
430 static const struct timeval warninterval = { 300, 0 };
432 if (!zone_warnings || zone->uz_warning == NULL)
435 if (ratecheck(&zone->uz_ratecheck, &warninterval))
436 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
440 zone_maxaction(uma_zone_t zone)
442 if (zone->uz_maxaction)
443 (*zone->uz_maxaction)(zone);
447 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
451 LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
452 kegfn(klink->kl_keg);
456 * Routine called by timeout which is used to fire off some time interval
457 * based calculations. (stats, hash size, etc.)
466 uma_timeout(void *unused)
469 zone_foreach(zone_timeout);
471 /* Reschedule this event */
472 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
476 * Routine to perform timeout driven calculations. This expands the
477 * hashes and does per cpu statistics aggregation.
482 keg_timeout(uma_keg_t keg)
487 * Expand the keg hash table.
489 * This is done if the number of slabs is larger than the hash size.
490 * What I'm trying to do here is completely reduce collisions. This
491 * may be a little aggressive. Should I allow for two collisions max?
493 if (keg->uk_flags & UMA_ZONE_HASH &&
494 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
495 struct uma_hash newhash;
496 struct uma_hash oldhash;
500 * This is so involved because allocating and freeing
501 * while the keg lock is held will lead to deadlock.
502 * I have to do everything in stages and check for
505 newhash = keg->uk_hash;
507 ret = hash_alloc(&newhash);
510 if (hash_expand(&keg->uk_hash, &newhash)) {
511 oldhash = keg->uk_hash;
512 keg->uk_hash = newhash;
525 zone_timeout(uma_zone_t zone)
528 zone_foreach_keg(zone, &keg_timeout);
532 * Allocate and zero fill the next sized hash table from the appropriate
536 * hash A new hash structure with the old hash size in uh_hashsize
539 * 1 on sucess and 0 on failure.
542 hash_alloc(struct uma_hash *hash)
547 oldsize = hash->uh_hashsize;
549 /* We're just going to go to a power of two greater */
551 hash->uh_hashsize = oldsize * 2;
552 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
553 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
554 M_UMAHASH, M_NOWAIT);
556 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
557 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
559 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
561 if (hash->uh_slab_hash) {
562 bzero(hash->uh_slab_hash, alloc);
563 hash->uh_hashmask = hash->uh_hashsize - 1;
571 * Expands the hash table for HASH zones. This is done from zone_timeout
572 * to reduce collisions. This must not be done in the regular allocation
573 * path, otherwise, we can recurse on the vm while allocating pages.
576 * oldhash The hash you want to expand
577 * newhash The hash structure for the new table
585 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
591 if (!newhash->uh_slab_hash)
594 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
598 * I need to investigate hash algorithms for resizing without a
602 for (i = 0; i < oldhash->uh_hashsize; i++)
603 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
604 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
605 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
606 hval = UMA_HASH(newhash, slab->us_data);
607 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
615 * Free the hash bucket to the appropriate backing store.
618 * slab_hash The hash bucket we're freeing
619 * hashsize The number of entries in that hash bucket
625 hash_free(struct uma_hash *hash)
627 if (hash->uh_slab_hash == NULL)
629 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
630 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
632 free(hash->uh_slab_hash, M_UMAHASH);
636 * Frees all outstanding items in a bucket
639 * zone The zone to free to, must be unlocked.
640 * bucket The free/alloc bucket with items, cpu queue must be locked.
647 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
655 for (i = 0; i < bucket->ub_cnt; i++)
656 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
657 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
662 * Drains the per cpu caches for a zone.
664 * NOTE: This may only be called while the zone is being turn down, and not
665 * during normal operation. This is necessary in order that we do not have
666 * to migrate CPUs to drain the per-CPU caches.
669 * zone The zone to drain, must be unlocked.
675 cache_drain(uma_zone_t zone)
681 * XXX: It is safe to not lock the per-CPU caches, because we're
682 * tearing down the zone anyway. I.e., there will be no further use
683 * of the caches at this point.
685 * XXX: It would good to be able to assert that the zone is being
686 * torn down to prevent improper use of cache_drain().
688 * XXX: We lock the zone before passing into bucket_cache_drain() as
689 * it is used elsewhere. Should the tear-down path be made special
690 * there in some form?
693 cache = &zone->uz_cpu[cpu];
694 bucket_drain(zone, cache->uc_allocbucket);
695 bucket_drain(zone, cache->uc_freebucket);
696 if (cache->uc_allocbucket != NULL)
697 bucket_free(zone, cache->uc_allocbucket, NULL);
698 if (cache->uc_freebucket != NULL)
699 bucket_free(zone, cache->uc_freebucket, NULL);
700 cache->uc_allocbucket = cache->uc_freebucket = NULL;
703 bucket_cache_drain(zone);
708 cache_shrink(uma_zone_t zone)
711 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
715 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
720 cache_drain_safe_cpu(uma_zone_t zone)
725 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
731 cache = &zone->uz_cpu[curcpu];
732 if (cache->uc_allocbucket) {
733 if (cache->uc_allocbucket->ub_cnt != 0)
734 LIST_INSERT_HEAD(&zone->uz_buckets,
735 cache->uc_allocbucket, ub_link);
737 b1 = cache->uc_allocbucket;
738 cache->uc_allocbucket = NULL;
740 if (cache->uc_freebucket) {
741 if (cache->uc_freebucket->ub_cnt != 0)
742 LIST_INSERT_HEAD(&zone->uz_buckets,
743 cache->uc_freebucket, ub_link);
745 b2 = cache->uc_freebucket;
746 cache->uc_freebucket = NULL;
751 bucket_free(zone, b1, NULL);
753 bucket_free(zone, b2, NULL);
757 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
758 * This is an expensive call because it needs to bind to all CPUs
759 * one by one and enter a critical section on each of them in order
760 * to safely access their cache buckets.
761 * Zone lock must not be held on call this function.
764 cache_drain_safe(uma_zone_t zone)
769 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
774 zone_foreach(cache_shrink);
777 thread_lock(curthread);
778 sched_bind(curthread, cpu);
779 thread_unlock(curthread);
782 cache_drain_safe_cpu(zone);
784 zone_foreach(cache_drain_safe_cpu);
786 thread_lock(curthread);
787 sched_unbind(curthread);
788 thread_unlock(curthread);
792 * Drain the cached buckets from a zone. Expects a locked zone on entry.
795 bucket_cache_drain(uma_zone_t zone)
800 * Drain the bucket queues and free the buckets, we just keep two per
803 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
804 LIST_REMOVE(bucket, ub_link);
806 bucket_drain(zone, bucket);
807 bucket_free(zone, bucket, NULL);
812 * Shrink further bucket sizes. Price of single zone lock collision
813 * is probably lower then price of global cache drain.
815 if (zone->uz_count > zone->uz_count_min)
820 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
827 flags = slab->us_flags;
829 if (keg->uk_fini != NULL) {
830 for (i--; i > -1; i--)
831 keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
834 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
835 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
837 printf("%s: Returning %d bytes.\n", keg->uk_name,
838 PAGE_SIZE * keg->uk_ppera);
840 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
844 * Frees pages from a keg back to the system. This is done on demand from
845 * the pageout daemon.
850 keg_drain(uma_keg_t keg)
852 struct slabhead freeslabs = { 0 };
857 * We don't want to take pages from statically allocated kegs at this
860 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
864 printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
867 if (keg->uk_free == 0)
870 slab = LIST_FIRST(&keg->uk_free_slab);
872 n = LIST_NEXT(slab, us_link);
874 /* We have no where to free these to */
875 if (slab->us_flags & UMA_SLAB_BOOT) {
880 LIST_REMOVE(slab, us_link);
881 keg->uk_pages -= keg->uk_ppera;
882 keg->uk_free -= keg->uk_ipers;
884 if (keg->uk_flags & UMA_ZONE_HASH)
885 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
887 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
894 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
895 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
896 keg_free_slab(keg, slab, keg->uk_ipers);
901 zone_drain_wait(uma_zone_t zone, int waitok)
905 * Set draining to interlock with zone_dtor() so we can release our
906 * locks as we go. Only dtor() should do a WAITOK call since it
907 * is the only call that knows the structure will still be available
911 while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
912 if (waitok == M_NOWAIT)
914 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
916 zone->uz_flags |= UMA_ZFLAG_DRAINING;
917 bucket_cache_drain(zone);
920 * The DRAINING flag protects us from being freed while
921 * we're running. Normally the uma_rwlock would protect us but we
922 * must be able to release and acquire the right lock for each keg.
924 zone_foreach_keg(zone, &keg_drain);
926 zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
933 zone_drain(uma_zone_t zone)
936 zone_drain_wait(zone, M_NOWAIT);
940 * Allocate a new slab for a keg. This does not insert the slab onto a list.
943 * wait Shall we wait?
946 * The slab that was allocated or NULL if there is no memory and the
947 * caller specified M_NOWAIT.
950 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
952 uma_slabrefcnt_t slabref;
959 mtx_assert(&keg->uk_lock, MA_OWNED);
964 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name);
966 allocf = keg->uk_allocf;
969 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
970 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
976 * This reproduces the old vm_zone behavior of zero filling pages the
977 * first time they are added to a zone.
979 * Malloced items are zeroed in uma_zalloc.
982 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
987 if (keg->uk_flags & UMA_ZONE_NODUMP)
990 /* zone is passed for legacy reasons. */
991 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
993 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
994 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
999 /* Point the slab into the allocated memory */
1000 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
1001 slab = (uma_slab_t )(mem + keg->uk_pgoff);
1003 if (keg->uk_flags & UMA_ZONE_VTOSLAB)
1004 for (i = 0; i < keg->uk_ppera; i++)
1005 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
1008 slab->us_data = mem;
1009 slab->us_freecount = keg->uk_ipers;
1010 slab->us_flags = flags;
1011 BIT_FILL(SLAB_SETSIZE, &slab->us_free);
1013 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1015 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1016 slabref = (uma_slabrefcnt_t)slab;
1017 for (i = 0; i < keg->uk_ipers; i++)
1018 slabref->us_refcnt[i] = 0;
1021 if (keg->uk_init != NULL) {
1022 for (i = 0; i < keg->uk_ipers; i++)
1023 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1024 keg->uk_size, wait) != 0)
1026 if (i != keg->uk_ipers) {
1027 keg_free_slab(keg, slab, i);
1036 if (keg->uk_flags & UMA_ZONE_HASH)
1037 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1039 keg->uk_pages += keg->uk_ppera;
1040 keg->uk_free += keg->uk_ipers;
1047 * This function is intended to be used early on in place of page_alloc() so
1048 * that we may use the boot time page cache to satisfy allocations before
1052 startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1056 int pages, check_pages;
1058 keg = zone_first_keg(zone);
1059 pages = howmany(bytes, PAGE_SIZE);
1060 check_pages = pages - 1;
1061 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1064 * Check our small startup cache to see if it has pages remaining.
1066 mtx_lock(&uma_boot_pages_mtx);
1068 /* First check if we have enough room. */
1069 tmps = LIST_FIRST(&uma_boot_pages);
1070 while (tmps != NULL && check_pages-- > 0)
1071 tmps = LIST_NEXT(tmps, us_link);
1074 * It's ok to lose tmps references. The last one will
1075 * have tmps->us_data pointing to the start address of
1076 * "pages" contiguous pages of memory.
1078 while (pages-- > 0) {
1079 tmps = LIST_FIRST(&uma_boot_pages);
1080 LIST_REMOVE(tmps, us_link);
1082 mtx_unlock(&uma_boot_pages_mtx);
1083 *pflag = tmps->us_flags;
1084 return (tmps->us_data);
1086 mtx_unlock(&uma_boot_pages_mtx);
1087 if (booted < UMA_STARTUP2)
1088 panic("UMA: Increase vm.boot_pages");
1090 * Now that we've booted reset these users to their real allocator.
1092 #ifdef UMA_MD_SMALL_ALLOC
1093 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1095 keg->uk_allocf = page_alloc;
1097 return keg->uk_allocf(zone, bytes, pflag, wait);
1101 * Allocates a number of pages from the system
1104 * bytes The number of bytes requested
1105 * wait Shall we wait?
1108 * A pointer to the alloced memory or possibly
1109 * NULL if M_NOWAIT is set.
1112 page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1114 void *p; /* Returned page */
1116 *pflag = UMA_SLAB_KMEM;
1117 p = (void *) kmem_malloc(kmem_arena, bytes, wait);
1123 * Allocates a number of pages from within an object
1126 * bytes The number of bytes requested
1127 * wait Shall we wait?
1130 * A pointer to the alloced memory or possibly
1131 * NULL if M_NOWAIT is set.
1134 noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait)
1136 TAILQ_HEAD(, vm_page) alloctail;
1138 vm_offset_t retkva, zkva;
1139 vm_page_t p, p_next;
1142 TAILQ_INIT(&alloctail);
1143 keg = zone_first_keg(zone);
1145 npages = howmany(bytes, PAGE_SIZE);
1146 while (npages > 0) {
1147 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1148 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1151 * Since the page does not belong to an object, its
1154 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1158 if (wait & M_WAITOK) {
1164 * Page allocation failed, free intermediate pages and
1167 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1168 vm_page_unwire(p, PQ_NONE);
1173 *flags = UMA_SLAB_PRIV;
1174 zkva = keg->uk_kva +
1175 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1177 TAILQ_FOREACH(p, &alloctail, listq) {
1178 pmap_qenter(zkva, &p, 1);
1182 return ((void *)retkva);
1186 * Frees a number of pages to the system
1189 * mem A pointer to the memory to be freed
1190 * size The size of the memory being freed
1191 * flags The original p->us_flags field
1197 page_free(void *mem, vm_size_t size, uint8_t flags)
1201 if (flags & UMA_SLAB_KMEM)
1203 else if (flags & UMA_SLAB_KERNEL)
1204 vmem = kernel_arena;
1206 panic("UMA: page_free used with invalid flags %d", flags);
1208 kmem_free(vmem, (vm_offset_t)mem, size);
1212 * Zero fill initializer
1214 * Arguments/Returns follow uma_init specifications
1217 zero_init(void *mem, int size, int flags)
1224 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1227 * keg The zone we should initialize
1233 keg_small_init(uma_keg_t keg)
1240 if (keg->uk_flags & UMA_ZONE_PCPU) {
1241 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1243 keg->uk_slabsize = sizeof(struct pcpu);
1244 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1247 keg->uk_slabsize = UMA_SLAB_SIZE;
1252 * Calculate the size of each allocation (rsize) according to
1253 * alignment. If the requested size is smaller than we have
1254 * allocation bits for we round it up.
1256 rsize = keg->uk_size;
1257 if (rsize < keg->uk_slabsize / SLAB_SETSIZE)
1258 rsize = keg->uk_slabsize / SLAB_SETSIZE;
1259 if (rsize & keg->uk_align)
1260 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1261 keg->uk_rsize = rsize;
1263 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1264 keg->uk_rsize < sizeof(struct pcpu),
1265 ("%s: size %u too large", __func__, keg->uk_rsize));
1267 if (keg->uk_flags & UMA_ZONE_REFCNT)
1268 rsize += sizeof(uint32_t);
1270 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1273 shsize = sizeof(struct uma_slab);
1275 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1276 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1277 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1279 memused = keg->uk_ipers * rsize + shsize;
1280 wastedspace = keg->uk_slabsize - memused;
1283 * We can't do OFFPAGE if we're internal or if we've been
1284 * asked to not go to the VM for buckets. If we do this we
1285 * may end up going to the VM for slabs which we do not
1286 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1287 * of UMA_ZONE_VM, which clearly forbids it.
1289 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1290 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1294 * See if using an OFFPAGE slab will limit our waste. Only do
1295 * this if it permits more items per-slab.
1297 * XXX We could try growing slabsize to limit max waste as well.
1298 * Historically this was not done because the VM could not
1299 * efficiently handle contiguous allocations.
1301 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1302 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1303 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1304 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1305 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1307 printf("UMA decided we need offpage slab headers for "
1308 "keg: %s, calculated wastedspace = %d, "
1309 "maximum wasted space allowed = %d, "
1310 "calculated ipers = %d, "
1311 "new wasted space = %d\n", keg->uk_name, wastedspace,
1312 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1313 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1315 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1318 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1319 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1320 keg->uk_flags |= UMA_ZONE_HASH;
1324 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1325 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1329 * keg The keg we should initialize
1335 keg_large_init(uma_keg_t keg)
1339 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1340 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1341 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1342 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1343 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1345 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1346 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1348 keg->uk_rsize = keg->uk_size;
1350 /* We can't do OFFPAGE if we're internal, bail out here. */
1351 if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1354 /* Check whether we have enough space to not do OFFPAGE. */
1355 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1356 shsize = sizeof(struct uma_slab);
1357 if (keg->uk_flags & UMA_ZONE_REFCNT)
1358 shsize += keg->uk_ipers * sizeof(uint32_t);
1359 if (shsize & UMA_ALIGN_PTR)
1360 shsize = (shsize & ~UMA_ALIGN_PTR) +
1361 (UMA_ALIGN_PTR + 1);
1363 if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize)
1364 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1367 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1368 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1369 keg->uk_flags |= UMA_ZONE_HASH;
1373 keg_cachespread_init(uma_keg_t keg)
1380 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1381 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1383 alignsize = keg->uk_align + 1;
1384 rsize = keg->uk_size;
1386 * We want one item to start on every align boundary in a page. To
1387 * do this we will span pages. We will also extend the item by the
1388 * size of align if it is an even multiple of align. Otherwise, it
1389 * would fall on the same boundary every time.
1391 if (rsize & keg->uk_align)
1392 rsize = (rsize & ~keg->uk_align) + alignsize;
1393 if ((rsize & alignsize) == 0)
1395 trailer = rsize - keg->uk_size;
1396 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1397 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1398 keg->uk_rsize = rsize;
1399 keg->uk_ppera = pages;
1400 keg->uk_slabsize = UMA_SLAB_SIZE;
1401 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1402 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1403 KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1404 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1409 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1410 * the keg onto the global keg list.
1412 * Arguments/Returns follow uma_ctor specifications
1413 * udata Actually uma_kctor_args
1416 keg_ctor(void *mem, int size, void *udata, int flags)
1418 struct uma_kctor_args *arg = udata;
1419 uma_keg_t keg = mem;
1423 keg->uk_size = arg->size;
1424 keg->uk_init = arg->uminit;
1425 keg->uk_fini = arg->fini;
1426 keg->uk_align = arg->align;
1428 keg->uk_reserve = 0;
1430 keg->uk_flags = arg->flags;
1431 keg->uk_allocf = page_alloc;
1432 keg->uk_freef = page_free;
1433 keg->uk_slabzone = NULL;
1436 * The master zone is passed to us at keg-creation time.
1439 keg->uk_name = zone->uz_name;
1441 if (arg->flags & UMA_ZONE_VM)
1442 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1444 if (arg->flags & UMA_ZONE_ZINIT)
1445 keg->uk_init = zero_init;
1447 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1448 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1450 if (arg->flags & UMA_ZONE_PCPU)
1452 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1454 keg->uk_flags &= ~UMA_ZONE_PCPU;
1457 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1458 keg_cachespread_init(keg);
1459 } else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1461 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) -
1463 keg_large_init(keg);
1465 keg_small_init(keg);
1467 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1468 keg_large_init(keg);
1470 keg_small_init(keg);
1473 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1474 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1475 if (keg->uk_ipers > uma_max_ipers_ref)
1476 panic("Too many ref items per zone: %d > %d\n",
1477 keg->uk_ipers, uma_max_ipers_ref);
1478 keg->uk_slabzone = slabrefzone;
1480 keg->uk_slabzone = slabzone;
1484 * If we haven't booted yet we need allocations to go through the
1485 * startup cache until the vm is ready.
1487 if (keg->uk_ppera == 1) {
1488 #ifdef UMA_MD_SMALL_ALLOC
1489 keg->uk_allocf = uma_small_alloc;
1490 keg->uk_freef = uma_small_free;
1492 if (booted < UMA_STARTUP)
1493 keg->uk_allocf = startup_alloc;
1495 if (booted < UMA_STARTUP2)
1496 keg->uk_allocf = startup_alloc;
1498 } else if (booted < UMA_STARTUP2 &&
1499 (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1500 keg->uk_allocf = startup_alloc;
1503 * Initialize keg's lock
1505 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1508 * If we're putting the slab header in the actual page we need to
1509 * figure out where in each page it goes. This calculates a right
1510 * justified offset into the memory on an ALIGN_PTR boundary.
1512 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1515 /* Size of the slab struct and free list */
1516 totsize = sizeof(struct uma_slab);
1518 /* Size of the reference counts. */
1519 if (keg->uk_flags & UMA_ZONE_REFCNT)
1520 totsize += keg->uk_ipers * sizeof(uint32_t);
1522 if (totsize & UMA_ALIGN_PTR)
1523 totsize = (totsize & ~UMA_ALIGN_PTR) +
1524 (UMA_ALIGN_PTR + 1);
1525 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1528 * The only way the following is possible is if with our
1529 * UMA_ALIGN_PTR adjustments we are now bigger than
1530 * UMA_SLAB_SIZE. I haven't checked whether this is
1531 * mathematically possible for all cases, so we make
1534 totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1535 if (keg->uk_flags & UMA_ZONE_REFCNT)
1536 totsize += keg->uk_ipers * sizeof(uint32_t);
1537 if (totsize > PAGE_SIZE * keg->uk_ppera) {
1538 printf("zone %s ipers %d rsize %d size %d\n",
1539 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1541 panic("UMA slab won't fit.");
1545 if (keg->uk_flags & UMA_ZONE_HASH)
1546 hash_alloc(&keg->uk_hash);
1549 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1550 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1551 keg->uk_ipers, keg->uk_ppera,
1552 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1555 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1557 rw_wlock(&uma_rwlock);
1558 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1559 rw_wunlock(&uma_rwlock);
1564 * Zone header ctor. This initializes all fields, locks, etc.
1566 * Arguments/Returns follow uma_ctor specifications
1567 * udata Actually uma_zctor_args
1570 zone_ctor(void *mem, int size, void *udata, int flags)
1572 struct uma_zctor_args *arg = udata;
1573 uma_zone_t zone = mem;
1578 zone->uz_name = arg->name;
1579 zone->uz_ctor = arg->ctor;
1580 zone->uz_dtor = arg->dtor;
1581 zone->uz_slab = zone_fetch_slab;
1582 zone->uz_init = NULL;
1583 zone->uz_fini = NULL;
1584 zone->uz_allocs = 0;
1587 zone->uz_sleeps = 0;
1589 zone->uz_count_min = 0;
1591 zone->uz_warning = NULL;
1592 timevalclear(&zone->uz_ratecheck);
1593 zone->uz_maxaction = NULL;
1596 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1599 * This is a pure cache zone, no kegs.
1602 if (arg->flags & UMA_ZONE_VM)
1603 arg->flags |= UMA_ZFLAG_CACHEONLY;
1604 zone->uz_flags = arg->flags;
1605 zone->uz_size = arg->size;
1606 zone->uz_import = arg->import;
1607 zone->uz_release = arg->release;
1608 zone->uz_arg = arg->arg;
1609 zone->uz_lockptr = &zone->uz_lock;
1610 rw_wlock(&uma_rwlock);
1611 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1612 rw_wunlock(&uma_rwlock);
1617 * Use the regular zone/keg/slab allocator.
1619 zone->uz_import = (uma_import)zone_import;
1620 zone->uz_release = (uma_release)zone_release;
1621 zone->uz_arg = zone;
1623 if (arg->flags & UMA_ZONE_SECONDARY) {
1624 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1625 zone->uz_init = arg->uminit;
1626 zone->uz_fini = arg->fini;
1627 zone->uz_lockptr = &keg->uk_lock;
1628 zone->uz_flags |= UMA_ZONE_SECONDARY;
1629 rw_wlock(&uma_rwlock);
1631 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1632 if (LIST_NEXT(z, uz_link) == NULL) {
1633 LIST_INSERT_AFTER(z, zone, uz_link);
1638 rw_wunlock(&uma_rwlock);
1639 } else if (keg == NULL) {
1640 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1641 arg->align, arg->flags)) == NULL)
1644 struct uma_kctor_args karg;
1647 /* We should only be here from uma_startup() */
1648 karg.size = arg->size;
1649 karg.uminit = arg->uminit;
1650 karg.fini = arg->fini;
1651 karg.align = arg->align;
1652 karg.flags = arg->flags;
1654 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1661 * Link in the first keg.
1663 zone->uz_klink.kl_keg = keg;
1664 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1665 zone->uz_lockptr = &keg->uk_lock;
1666 zone->uz_size = keg->uk_size;
1667 zone->uz_flags |= (keg->uk_flags &
1668 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1671 * Some internal zones don't have room allocated for the per cpu
1672 * caches. If we're internal, bail out here.
1674 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1675 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1676 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1681 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1682 zone->uz_count = bucket_select(zone->uz_size);
1684 zone->uz_count = BUCKET_MAX;
1685 zone->uz_count_min = zone->uz_count;
1691 * Keg header dtor. This frees all data, destroys locks, frees the hash
1692 * table and removes the keg from the global list.
1694 * Arguments/Returns follow uma_dtor specifications
1698 keg_dtor(void *arg, int size, void *udata)
1702 keg = (uma_keg_t)arg;
1704 if (keg->uk_free != 0) {
1705 printf("Freed UMA keg (%s) was not empty (%d items). "
1706 " Lost %d pages of memory.\n",
1707 keg->uk_name ? keg->uk_name : "",
1708 keg->uk_free, keg->uk_pages);
1712 hash_free(&keg->uk_hash);
1720 * Arguments/Returns follow uma_dtor specifications
1724 zone_dtor(void *arg, int size, void *udata)
1730 zone = (uma_zone_t)arg;
1731 keg = zone_first_keg(zone);
1733 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1736 rw_wlock(&uma_rwlock);
1737 LIST_REMOVE(zone, uz_link);
1738 rw_wunlock(&uma_rwlock);
1740 * XXX there are some races here where
1741 * the zone can be drained but zone lock
1742 * released and then refilled before we
1743 * remove it... we dont care for now
1745 zone_drain_wait(zone, M_WAITOK);
1747 * Unlink all of our kegs.
1749 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1750 klink->kl_keg = NULL;
1751 LIST_REMOVE(klink, kl_link);
1752 if (klink == &zone->uz_klink)
1754 free(klink, M_TEMP);
1757 * We only destroy kegs from non secondary zones.
1759 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1760 rw_wlock(&uma_rwlock);
1761 LIST_REMOVE(keg, uk_link);
1762 rw_wunlock(&uma_rwlock);
1763 zone_free_item(kegs, keg, NULL, SKIP_NONE);
1765 ZONE_LOCK_FINI(zone);
1769 * Traverses every zone in the system and calls a callback
1772 * zfunc A pointer to a function which accepts a zone
1779 zone_foreach(void (*zfunc)(uma_zone_t))
1784 rw_rlock(&uma_rwlock);
1785 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1786 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1789 rw_runlock(&uma_rwlock);
1792 /* Public functions */
1795 uma_startup(void *bootmem, int boot_pages)
1797 struct uma_zctor_args args;
1803 printf("Creating uma keg headers zone and keg.\n");
1805 rw_init(&uma_rwlock, "UMA lock");
1807 /* "manually" create the initial zone */
1808 memset(&args, 0, sizeof(args));
1809 args.name = "UMA Kegs";
1810 args.size = sizeof(struct uma_keg);
1811 args.ctor = keg_ctor;
1812 args.dtor = keg_dtor;
1813 args.uminit = zero_init;
1815 args.keg = &masterkeg;
1816 args.align = 32 - 1;
1817 args.flags = UMA_ZFLAG_INTERNAL;
1818 /* The initial zone has no Per cpu queues so it's smaller */
1819 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1822 printf("Filling boot free list.\n");
1824 for (i = 0; i < boot_pages; i++) {
1825 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1826 slab->us_data = (uint8_t *)slab;
1827 slab->us_flags = UMA_SLAB_BOOT;
1828 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1830 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1833 printf("Creating uma zone headers zone and keg.\n");
1835 args.name = "UMA Zones";
1836 args.size = sizeof(struct uma_zone) +
1837 (sizeof(struct uma_cache) * (mp_maxid + 1));
1838 args.ctor = zone_ctor;
1839 args.dtor = zone_dtor;
1840 args.uminit = zero_init;
1843 args.align = 32 - 1;
1844 args.flags = UMA_ZFLAG_INTERNAL;
1845 /* The initial zone has no Per cpu queues so it's smaller */
1846 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1849 printf("Creating slab and hash zones.\n");
1852 /* Now make a zone for slab headers */
1853 slabzone = uma_zcreate("UMA Slabs",
1854 sizeof(struct uma_slab),
1855 NULL, NULL, NULL, NULL,
1856 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1859 * We also create a zone for the bigger slabs with reference
1860 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1862 slabsize = sizeof(struct uma_slab_refcnt);
1863 slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1864 slabrefzone = uma_zcreate("UMA RCntSlabs",
1866 NULL, NULL, NULL, NULL,
1868 UMA_ZFLAG_INTERNAL);
1870 hashzone = uma_zcreate("UMA Hash",
1871 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1872 NULL, NULL, NULL, NULL,
1873 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1877 booted = UMA_STARTUP;
1880 printf("UMA startup complete.\n");
1888 booted = UMA_STARTUP2;
1890 sx_init(&uma_drain_lock, "umadrain");
1892 printf("UMA startup2 complete.\n");
1897 * Initialize our callout handle
1905 printf("Starting callout.\n");
1907 callout_init(&uma_callout, 1);
1908 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1910 printf("UMA startup3 complete.\n");
1915 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1916 int align, uint32_t flags)
1918 struct uma_kctor_args args;
1921 args.uminit = uminit;
1923 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1926 return (zone_alloc_item(kegs, &args, M_WAITOK));
1931 uma_set_align(int align)
1934 if (align != UMA_ALIGN_CACHE)
1935 uma_align_cache = align;
1940 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1941 uma_init uminit, uma_fini fini, int align, uint32_t flags)
1944 struct uma_zctor_args args;
1948 /* This stuff is essential for the zone ctor */
1949 memset(&args, 0, sizeof(args));
1954 args.uminit = uminit;
1958 * If a zone is being created with an empty constructor and
1959 * destructor, pass UMA constructor/destructor which checks for
1960 * memory use after free.
1962 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) &&
1963 ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) {
1964 args.ctor = trash_ctor;
1965 args.dtor = trash_dtor;
1966 args.uminit = trash_init;
1967 args.fini = trash_fini;
1974 if (booted < UMA_STARTUP2) {
1977 sx_slock(&uma_drain_lock);
1980 res = zone_alloc_item(zones, &args, M_WAITOK);
1982 sx_sunlock(&uma_drain_lock);
1988 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1989 uma_init zinit, uma_fini zfini, uma_zone_t master)
1991 struct uma_zctor_args args;
1996 keg = zone_first_keg(master);
1997 memset(&args, 0, sizeof(args));
1999 args.size = keg->uk_size;
2002 args.uminit = zinit;
2004 args.align = keg->uk_align;
2005 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
2008 if (booted < UMA_STARTUP2) {
2011 sx_slock(&uma_drain_lock);
2014 /* XXX Attaches only one keg of potentially many. */
2015 res = zone_alloc_item(zones, &args, M_WAITOK);
2017 sx_sunlock(&uma_drain_lock);
2023 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
2024 uma_init zinit, uma_fini zfini, uma_import zimport,
2025 uma_release zrelease, void *arg, int flags)
2027 struct uma_zctor_args args;
2029 memset(&args, 0, sizeof(args));
2034 args.uminit = zinit;
2036 args.import = zimport;
2037 args.release = zrelease;
2042 return (zone_alloc_item(zones, &args, M_WAITOK));
2046 zone_lock_pair(uma_zone_t a, uma_zone_t b)
2050 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
2053 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
2058 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2066 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2073 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2075 zone_lock_pair(zone, master);
2077 * zone must use vtoslab() to resolve objects and must already be
2080 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2081 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2086 * The new master must also use vtoslab().
2088 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2093 * Both must either be refcnt, or not be refcnt.
2095 if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
2096 (master->uz_flags & UMA_ZONE_REFCNT)) {
2101 * The underlying object must be the same size. rsize
2104 if (master->uz_size != zone->uz_size) {
2109 * Put it at the end of the list.
2111 klink->kl_keg = zone_first_keg(master);
2112 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2113 if (LIST_NEXT(kl, kl_link) == NULL) {
2114 LIST_INSERT_AFTER(kl, klink, kl_link);
2119 zone->uz_flags |= UMA_ZFLAG_MULTI;
2120 zone->uz_slab = zone_fetch_slab_multi;
2123 zone_unlock_pair(zone, master);
2125 free(klink, M_TEMP);
2133 uma_zdestroy(uma_zone_t zone)
2136 sx_slock(&uma_drain_lock);
2137 zone_free_item(zones, zone, NULL, SKIP_NONE);
2138 sx_sunlock(&uma_drain_lock);
2143 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2147 uma_bucket_t bucket;
2151 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2152 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2154 /* This is the fast path allocation */
2155 #ifdef UMA_DEBUG_ALLOC_1
2156 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2158 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2159 zone->uz_name, flags);
2161 if (flags & M_WAITOK) {
2162 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2163 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2165 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2166 ("uma_zalloc_arg: called with spinlock or critical section held"));
2168 #ifdef DEBUG_MEMGUARD
2169 if (memguard_cmp_zone(zone)) {
2170 item = memguard_alloc(zone->uz_size, flags);
2173 * Avoid conflict with the use-after-free
2174 * protecting infrastructure from INVARIANTS.
2176 if (zone->uz_init != NULL &&
2177 zone->uz_init != mtrash_init &&
2178 zone->uz_init(item, zone->uz_size, flags) != 0)
2180 if (zone->uz_ctor != NULL &&
2181 zone->uz_ctor != mtrash_ctor &&
2182 zone->uz_ctor(item, zone->uz_size, udata,
2184 zone->uz_fini(item, zone->uz_size);
2189 /* This is unfortunate but should not be fatal. */
2193 * If possible, allocate from the per-CPU cache. There are two
2194 * requirements for safe access to the per-CPU cache: (1) the thread
2195 * accessing the cache must not be preempted or yield during access,
2196 * and (2) the thread must not migrate CPUs without switching which
2197 * cache it accesses. We rely on a critical section to prevent
2198 * preemption and migration. We release the critical section in
2199 * order to acquire the zone mutex if we are unable to allocate from
2200 * the current cache; when we re-acquire the critical section, we
2201 * must detect and handle migration if it has occurred.
2205 cache = &zone->uz_cpu[cpu];
2208 bucket = cache->uc_allocbucket;
2209 if (bucket != NULL && bucket->ub_cnt > 0) {
2211 item = bucket->ub_bucket[bucket->ub_cnt];
2213 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2215 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2218 if (zone->uz_ctor != NULL &&
2219 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2220 atomic_add_long(&zone->uz_fails, 1);
2221 zone_free_item(zone, item, udata, SKIP_DTOR);
2225 uma_dbg_alloc(zone, NULL, item);
2228 uma_zero_item(item, zone);
2233 * We have run out of items in our alloc bucket.
2234 * See if we can switch with our free bucket.
2236 bucket = cache->uc_freebucket;
2237 if (bucket != NULL && bucket->ub_cnt > 0) {
2238 #ifdef UMA_DEBUG_ALLOC
2239 printf("uma_zalloc: Swapping empty with alloc.\n");
2241 cache->uc_freebucket = cache->uc_allocbucket;
2242 cache->uc_allocbucket = bucket;
2247 * Discard any empty allocation bucket while we hold no locks.
2249 bucket = cache->uc_allocbucket;
2250 cache->uc_allocbucket = NULL;
2253 bucket_free(zone, bucket, udata);
2255 /* Short-circuit for zones without buckets and low memory. */
2256 if (zone->uz_count == 0 || bucketdisable)
2260 * Attempt to retrieve the item from the per-CPU cache has failed, so
2261 * we must go back to the zone. This requires the zone lock, so we
2262 * must drop the critical section, then re-acquire it when we go back
2263 * to the cache. Since the critical section is released, we may be
2264 * preempted or migrate. As such, make sure not to maintain any
2265 * thread-local state specific to the cache from prior to releasing
2266 * the critical section.
2269 if (ZONE_TRYLOCK(zone) == 0) {
2270 /* Record contention to size the buckets. */
2276 cache = &zone->uz_cpu[cpu];
2279 * Since we have locked the zone we may as well send back our stats.
2281 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2282 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2283 cache->uc_allocs = 0;
2284 cache->uc_frees = 0;
2286 /* See if we lost the race to fill the cache. */
2287 if (cache->uc_allocbucket != NULL) {
2293 * Check the zone's cache of buckets.
2295 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2296 KASSERT(bucket->ub_cnt != 0,
2297 ("uma_zalloc_arg: Returning an empty bucket."));
2299 LIST_REMOVE(bucket, ub_link);
2300 cache->uc_allocbucket = bucket;
2304 /* We are no longer associated with this CPU. */
2308 * We bump the uz count when the cache size is insufficient to
2309 * handle the working set.
2311 if (lockfail && zone->uz_count < BUCKET_MAX)
2316 * Now lets just fill a bucket and put it on the free list. If that
2317 * works we'll restart the allocation from the begining and it
2318 * will use the just filled bucket.
2320 bucket = zone_alloc_bucket(zone, udata, flags);
2321 if (bucket != NULL) {
2325 cache = &zone->uz_cpu[cpu];
2327 * See if we lost the race or were migrated. Cache the
2328 * initialized bucket to make this less likely or claim
2329 * the memory directly.
2331 if (cache->uc_allocbucket == NULL)
2332 cache->uc_allocbucket = bucket;
2334 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2340 * We may not be able to get a bucket so return an actual item.
2343 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2347 item = zone_alloc_item(zone, udata, flags);
2353 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2358 mtx_assert(&keg->uk_lock, MA_OWNED);
2361 if ((flags & M_USE_RESERVE) == 0)
2362 reserve = keg->uk_reserve;
2366 * Find a slab with some space. Prefer slabs that are partially
2367 * used over those that are totally full. This helps to reduce
2370 if (keg->uk_free > reserve) {
2371 if (!LIST_EMPTY(&keg->uk_part_slab)) {
2372 slab = LIST_FIRST(&keg->uk_part_slab);
2374 slab = LIST_FIRST(&keg->uk_free_slab);
2375 LIST_REMOVE(slab, us_link);
2376 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2379 MPASS(slab->us_keg == keg);
2384 * M_NOVM means don't ask at all!
2389 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2390 keg->uk_flags |= UMA_ZFLAG_FULL;
2392 * If this is not a multi-zone, set the FULL bit.
2393 * Otherwise slab_multi() takes care of it.
2395 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2396 zone->uz_flags |= UMA_ZFLAG_FULL;
2397 zone_log_warning(zone);
2398 zone_maxaction(zone);
2400 if (flags & M_NOWAIT)
2403 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2406 slab = keg_alloc_slab(keg, zone, flags);
2408 * If we got a slab here it's safe to mark it partially used
2409 * and return. We assume that the caller is going to remove
2410 * at least one item.
2413 MPASS(slab->us_keg == keg);
2414 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2418 * We might not have been able to get a slab but another cpu
2419 * could have while we were unlocked. Check again before we
2428 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2433 keg = zone_first_keg(zone);
2438 slab = keg_fetch_slab(keg, zone, flags);
2441 if (flags & (M_NOWAIT | M_NOVM))
2449 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2450 * with the keg locked. On NULL no lock is held.
2452 * The last pointer is used to seed the search. It is not required.
2455 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2465 * Don't wait on the first pass. This will skip limit tests
2466 * as well. We don't want to block if we can find a provider
2469 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2471 * Use the last slab allocated as a hint for where to start
2475 slab = keg_fetch_slab(last, zone, flags);
2481 * Loop until we have a slab incase of transient failures
2482 * while M_WAITOK is specified. I'm not sure this is 100%
2483 * required but we've done it for so long now.
2489 * Search the available kegs for slabs. Be careful to hold the
2490 * correct lock while calling into the keg layer.
2492 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2493 keg = klink->kl_keg;
2495 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2496 slab = keg_fetch_slab(keg, zone, flags);
2500 if (keg->uk_flags & UMA_ZFLAG_FULL)
2506 if (rflags & (M_NOWAIT | M_NOVM))
2510 * All kegs are full. XXX We can't atomically check all kegs
2511 * and sleep so just sleep for a short period and retry.
2513 if (full && !empty) {
2515 zone->uz_flags |= UMA_ZFLAG_FULL;
2517 zone_log_warning(zone);
2518 zone_maxaction(zone);
2519 msleep(zone, zone->uz_lockptr, PVM,
2520 "zonelimit", hz/100);
2521 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2530 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2535 MPASS(keg == slab->us_keg);
2536 mtx_assert(&keg->uk_lock, MA_OWNED);
2538 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2539 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2540 item = slab->us_data + (keg->uk_rsize * freei);
2541 slab->us_freecount--;
2544 /* Move this slab to the full list */
2545 if (slab->us_freecount == 0) {
2546 LIST_REMOVE(slab, us_link);
2547 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2554 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2562 /* Try to keep the buckets totally full */
2563 for (i = 0; i < max; ) {
2564 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2567 while (slab->us_freecount && i < max) {
2568 bucket[i++] = slab_alloc_item(keg, slab);
2569 if (keg->uk_free <= keg->uk_reserve)
2572 /* Don't grab more than one slab at a time. */
2583 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2585 uma_bucket_t bucket;
2588 /* Don't wait for buckets, preserve caller's NOVM setting. */
2589 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2593 max = MIN(bucket->ub_entries, zone->uz_count);
2594 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2598 * Initialize the memory if necessary.
2600 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2603 for (i = 0; i < bucket->ub_cnt; i++)
2604 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2608 * If we couldn't initialize the whole bucket, put the
2609 * rest back onto the freelist.
2611 if (i != bucket->ub_cnt) {
2612 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2613 bucket->ub_cnt - i);
2615 bzero(&bucket->ub_bucket[i],
2616 sizeof(void *) * (bucket->ub_cnt - i));
2622 if (bucket->ub_cnt == 0) {
2623 bucket_free(zone, bucket, udata);
2624 atomic_add_long(&zone->uz_fails, 1);
2632 * Allocates a single item from a zone.
2635 * zone The zone to alloc for.
2636 * udata The data to be passed to the constructor.
2637 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2640 * NULL if there is no memory and M_NOWAIT is set
2641 * An item if successful
2645 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2651 #ifdef UMA_DEBUG_ALLOC
2652 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2654 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2656 atomic_add_long(&zone->uz_allocs, 1);
2659 * We have to call both the zone's init (not the keg's init)
2660 * and the zone's ctor. This is because the item is going from
2661 * a keg slab directly to the user, and the user is expecting it
2662 * to be both zone-init'd as well as zone-ctor'd.
2664 if (zone->uz_init != NULL) {
2665 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2666 zone_free_item(zone, item, udata, SKIP_FINI);
2670 if (zone->uz_ctor != NULL) {
2671 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2672 zone_free_item(zone, item, udata, SKIP_DTOR);
2677 uma_dbg_alloc(zone, NULL, item);
2680 uma_zero_item(item, zone);
2685 atomic_add_long(&zone->uz_fails, 1);
2691 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2694 uma_bucket_t bucket;
2698 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2699 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2701 #ifdef UMA_DEBUG_ALLOC_1
2702 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2704 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2707 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2708 ("uma_zfree_arg: called with spinlock or critical section held"));
2710 /* uma_zfree(..., NULL) does nothing, to match free(9). */
2713 #ifdef DEBUG_MEMGUARD
2714 if (is_memguard_addr(item)) {
2715 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2716 zone->uz_dtor(item, zone->uz_size, udata);
2717 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2718 zone->uz_fini(item, zone->uz_size);
2719 memguard_free(item);
2724 if (zone->uz_flags & UMA_ZONE_MALLOC)
2725 uma_dbg_free(zone, udata, item);
2727 uma_dbg_free(zone, NULL, item);
2729 if (zone->uz_dtor != NULL)
2730 zone->uz_dtor(item, zone->uz_size, udata);
2733 * The race here is acceptable. If we miss it we'll just have to wait
2734 * a little longer for the limits to be reset.
2736 if (zone->uz_flags & UMA_ZFLAG_FULL)
2740 * If possible, free to the per-CPU cache. There are two
2741 * requirements for safe access to the per-CPU cache: (1) the thread
2742 * accessing the cache must not be preempted or yield during access,
2743 * and (2) the thread must not migrate CPUs without switching which
2744 * cache it accesses. We rely on a critical section to prevent
2745 * preemption and migration. We release the critical section in
2746 * order to acquire the zone mutex if we are unable to free to the
2747 * current cache; when we re-acquire the critical section, we must
2748 * detect and handle migration if it has occurred.
2753 cache = &zone->uz_cpu[cpu];
2757 * Try to free into the allocbucket first to give LIFO ordering
2758 * for cache-hot datastructures. Spill over into the freebucket
2759 * if necessary. Alloc will swap them if one runs dry.
2761 bucket = cache->uc_allocbucket;
2762 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2763 bucket = cache->uc_freebucket;
2764 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2765 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2766 ("uma_zfree: Freeing to non free bucket index."));
2767 bucket->ub_bucket[bucket->ub_cnt] = item;
2775 * We must go back the zone, which requires acquiring the zone lock,
2776 * which in turn means we must release and re-acquire the critical
2777 * section. Since the critical section is released, we may be
2778 * preempted or migrate. As such, make sure not to maintain any
2779 * thread-local state specific to the cache from prior to releasing
2780 * the critical section.
2783 if (zone->uz_count == 0 || bucketdisable)
2787 if (ZONE_TRYLOCK(zone) == 0) {
2788 /* Record contention to size the buckets. */
2794 cache = &zone->uz_cpu[cpu];
2797 * Since we have locked the zone we may as well send back our stats.
2799 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2800 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2801 cache->uc_allocs = 0;
2802 cache->uc_frees = 0;
2804 bucket = cache->uc_freebucket;
2805 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2809 cache->uc_freebucket = NULL;
2811 /* Can we throw this on the zone full list? */
2812 if (bucket != NULL) {
2813 #ifdef UMA_DEBUG_ALLOC
2814 printf("uma_zfree: Putting old bucket on the free list.\n");
2816 /* ub_cnt is pointing to the last free item */
2817 KASSERT(bucket->ub_cnt != 0,
2818 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2819 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2822 /* We are no longer associated with this CPU. */
2826 * We bump the uz count when the cache size is insufficient to
2827 * handle the working set.
2829 if (lockfail && zone->uz_count < BUCKET_MAX)
2833 #ifdef UMA_DEBUG_ALLOC
2834 printf("uma_zfree: Allocating new free bucket.\n");
2836 bucket = bucket_alloc(zone, udata, M_NOWAIT);
2840 cache = &zone->uz_cpu[cpu];
2841 if (cache->uc_freebucket == NULL) {
2842 cache->uc_freebucket = bucket;
2846 * We lost the race, start over. We have to drop our
2847 * critical section to free the bucket.
2850 bucket_free(zone, bucket, udata);
2855 * If nothing else caught this, we'll just do an internal free.
2858 zone_free_item(zone, item, udata, SKIP_DTOR);
2864 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2868 mtx_assert(&keg->uk_lock, MA_OWNED);
2869 MPASS(keg == slab->us_keg);
2871 /* Do we need to remove from any lists? */
2872 if (slab->us_freecount+1 == keg->uk_ipers) {
2873 LIST_REMOVE(slab, us_link);
2874 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2875 } else if (slab->us_freecount == 0) {
2876 LIST_REMOVE(slab, us_link);
2877 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2880 /* Slab management. */
2881 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2882 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2883 slab->us_freecount++;
2885 /* Keg statistics. */
2890 zone_release(uma_zone_t zone, void **bucket, int cnt)
2900 keg = zone_first_keg(zone);
2902 for (i = 0; i < cnt; i++) {
2904 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2905 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2906 if (zone->uz_flags & UMA_ZONE_HASH) {
2907 slab = hash_sfind(&keg->uk_hash, mem);
2909 mem += keg->uk_pgoff;
2910 slab = (uma_slab_t)mem;
2913 slab = vtoslab((vm_offset_t)item);
2914 if (slab->us_keg != keg) {
2920 slab_free_item(keg, slab, item);
2921 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2922 if (keg->uk_pages < keg->uk_maxpages) {
2923 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2928 * We can handle one more allocation. Since we're
2929 * clearing ZFLAG_FULL, wake up all procs blocked
2930 * on pages. This should be uncommon, so keeping this
2931 * simple for now (rather than adding count of blocked
2940 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2948 * Frees a single item to any zone.
2951 * zone The zone to free to
2952 * item The item we're freeing
2953 * udata User supplied data for the dtor
2954 * skip Skip dtors and finis
2957 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2961 if (skip == SKIP_NONE) {
2962 if (zone->uz_flags & UMA_ZONE_MALLOC)
2963 uma_dbg_free(zone, udata, item);
2965 uma_dbg_free(zone, NULL, item);
2968 if (skip < SKIP_DTOR && zone->uz_dtor)
2969 zone->uz_dtor(item, zone->uz_size, udata);
2971 if (skip < SKIP_FINI && zone->uz_fini)
2972 zone->uz_fini(item, zone->uz_size);
2974 atomic_add_long(&zone->uz_frees, 1);
2975 zone->uz_release(zone->uz_arg, &item, 1);
2980 uma_zone_set_max(uma_zone_t zone, int nitems)
2984 keg = zone_first_keg(zone);
2988 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2989 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2990 keg->uk_maxpages += keg->uk_ppera;
2991 nitems = keg->uk_maxpages * keg->uk_ipers;
2999 uma_zone_get_max(uma_zone_t zone)
3004 keg = zone_first_keg(zone);
3008 nitems = keg->uk_maxpages * keg->uk_ipers;
3016 uma_zone_set_warning(uma_zone_t zone, const char *warning)
3020 zone->uz_warning = warning;
3026 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
3030 zone->uz_maxaction = maxaction;
3036 uma_zone_get_cur(uma_zone_t zone)
3042 nitems = zone->uz_allocs - zone->uz_frees;
3045 * See the comment in sysctl_vm_zone_stats() regarding the
3046 * safety of accessing the per-cpu caches. With the zone lock
3047 * held, it is safe, but can potentially result in stale data.
3049 nitems += zone->uz_cpu[i].uc_allocs -
3050 zone->uz_cpu[i].uc_frees;
3054 return (nitems < 0 ? 0 : nitems);
3059 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
3063 keg = zone_first_keg(zone);
3064 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3066 KASSERT(keg->uk_pages == 0,
3067 ("uma_zone_set_init on non-empty keg"));
3068 keg->uk_init = uminit;
3074 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3078 keg = zone_first_keg(zone);
3079 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
3081 KASSERT(keg->uk_pages == 0,
3082 ("uma_zone_set_fini on non-empty keg"));
3083 keg->uk_fini = fini;
3089 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3093 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3094 ("uma_zone_set_zinit on non-empty keg"));
3095 zone->uz_init = zinit;
3101 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3105 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3106 ("uma_zone_set_zfini on non-empty keg"));
3107 zone->uz_fini = zfini;
3112 /* XXX uk_freef is not actually used with the zone locked */
3114 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3118 keg = zone_first_keg(zone);
3119 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
3121 keg->uk_freef = freef;
3126 /* XXX uk_allocf is not actually used with the zone locked */
3128 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3132 keg = zone_first_keg(zone);
3134 keg->uk_allocf = allocf;
3140 uma_zone_reserve(uma_zone_t zone, int items)
3144 keg = zone_first_keg(zone);
3148 keg->uk_reserve = items;
3156 uma_zone_reserve_kva(uma_zone_t zone, int count)
3162 keg = zone_first_keg(zone);
3165 pages = count / keg->uk_ipers;
3167 if (pages * keg->uk_ipers < count)
3170 #ifdef UMA_MD_SMALL_ALLOC
3171 if (keg->uk_ppera > 1) {
3175 kva = kva_alloc((vm_size_t)pages * UMA_SLAB_SIZE);
3183 keg->uk_maxpages = pages;
3184 #ifdef UMA_MD_SMALL_ALLOC
3185 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3187 keg->uk_allocf = noobj_alloc;
3189 keg->uk_flags |= UMA_ZONE_NOFREE;
3197 uma_prealloc(uma_zone_t zone, int items)
3203 keg = zone_first_keg(zone);
3207 slabs = items / keg->uk_ipers;
3208 if (slabs * keg->uk_ipers < items)
3211 slab = keg_alloc_slab(keg, zone, M_WAITOK);
3214 MPASS(slab->us_keg == keg);
3215 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3223 uma_find_refcnt(uma_zone_t zone, void *item)
3225 uma_slabrefcnt_t slabref;
3231 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3232 slabref = (uma_slabrefcnt_t)slab;
3234 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3235 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3236 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3237 refcnt = &slabref->us_refcnt[idx];
3243 uma_reclaim_locked(bool kmem_danger)
3247 printf("UMA: vm asked us to release pages!\n");
3249 sx_assert(&uma_drain_lock, SA_XLOCKED);
3251 zone_foreach(zone_drain);
3252 if (vm_page_count_min() || kmem_danger) {
3253 cache_drain_safe(NULL);
3254 zone_foreach(zone_drain);
3257 * Some slabs may have been freed but this zone will be visited early
3258 * we visit again so that we can free pages that are empty once other
3259 * zones are drained. We have to do the same for buckets.
3261 zone_drain(slabzone);
3262 zone_drain(slabrefzone);
3263 bucket_zone_drain();
3270 sx_xlock(&uma_drain_lock);
3271 uma_reclaim_locked(false);
3272 sx_xunlock(&uma_drain_lock);
3275 static int uma_reclaim_needed;
3278 uma_reclaim_wakeup(void)
3281 uma_reclaim_needed = 1;
3282 wakeup(&uma_reclaim_needed);
3286 uma_reclaim_worker(void *arg __unused)
3289 sx_xlock(&uma_drain_lock);
3291 sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM,
3293 if (uma_reclaim_needed) {
3294 uma_reclaim_needed = 0;
3295 uma_reclaim_locked(true);
3302 uma_zone_exhausted(uma_zone_t zone)
3307 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3313 uma_zone_exhausted_nolock(uma_zone_t zone)
3315 return (zone->uz_flags & UMA_ZFLAG_FULL);
3319 uma_large_malloc(vm_size_t size, int wait)
3325 slab = zone_alloc_item(slabzone, NULL, wait);
3328 mem = page_alloc(NULL, size, &flags, wait);
3330 vsetslab((vm_offset_t)mem, slab);
3331 slab->us_data = mem;
3332 slab->us_flags = flags | UMA_SLAB_MALLOC;
3333 slab->us_size = size;
3335 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3342 uma_large_free(uma_slab_t slab)
3345 page_free(slab->us_data, slab->us_size, slab->us_flags);
3346 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3350 uma_zero_item(void *item, uma_zone_t zone)
3353 if (zone->uz_flags & UMA_ZONE_PCPU) {
3354 for (int i = 0; i < mp_ncpus; i++)
3355 bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3357 bzero(item, zone->uz_size);
3361 uma_print_stats(void)
3363 zone_foreach(uma_print_zone);
3367 slab_print(uma_slab_t slab)
3369 printf("slab: keg %p, data %p, freecount %d\n",
3370 slab->us_keg, slab->us_data, slab->us_freecount);
3374 cache_print(uma_cache_t cache)
3376 printf("alloc: %p(%d), free: %p(%d)\n",
3377 cache->uc_allocbucket,
3378 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3379 cache->uc_freebucket,
3380 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3384 uma_print_keg(uma_keg_t keg)
3388 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3389 "out %d free %d limit %d\n",
3390 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3391 keg->uk_ipers, keg->uk_ppera,
3392 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3393 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3394 printf("Part slabs:\n");
3395 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3397 printf("Free slabs:\n");
3398 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3400 printf("Full slabs:\n");
3401 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3406 uma_print_zone(uma_zone_t zone)
3412 printf("zone: %s(%p) size %d flags %#x\n",
3413 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3414 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3415 uma_print_keg(kl->kl_keg);
3417 cache = &zone->uz_cpu[i];
3418 printf("CPU %d Cache:\n", i);
3425 * Generate statistics across both the zone and its per-cpu cache's. Return
3426 * desired statistics if the pointer is non-NULL for that statistic.
3428 * Note: does not update the zone statistics, as it can't safely clear the
3429 * per-CPU cache statistic.
3431 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3432 * safe from off-CPU; we should modify the caches to track this information
3433 * directly so that we don't have to.
3436 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3437 uint64_t *freesp, uint64_t *sleepsp)
3440 uint64_t allocs, frees, sleeps;
3443 allocs = frees = sleeps = 0;
3446 cache = &z->uz_cpu[cpu];
3447 if (cache->uc_allocbucket != NULL)
3448 cachefree += cache->uc_allocbucket->ub_cnt;
3449 if (cache->uc_freebucket != NULL)
3450 cachefree += cache->uc_freebucket->ub_cnt;
3451 allocs += cache->uc_allocs;
3452 frees += cache->uc_frees;
3454 allocs += z->uz_allocs;
3455 frees += z->uz_frees;
3456 sleeps += z->uz_sleeps;
3457 if (cachefreep != NULL)
3458 *cachefreep = cachefree;
3459 if (allocsp != NULL)
3463 if (sleepsp != NULL)
3469 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3476 rw_rlock(&uma_rwlock);
3477 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3478 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3481 rw_runlock(&uma_rwlock);
3482 return (sysctl_handle_int(oidp, &count, 0, req));
3486 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3488 struct uma_stream_header ush;
3489 struct uma_type_header uth;
3490 struct uma_percpu_stat ups;
3491 uma_bucket_t bucket;
3498 int count, error, i;
3500 error = sysctl_wire_old_buffer(req, 0);
3503 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3504 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3507 rw_rlock(&uma_rwlock);
3508 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3509 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3514 * Insert stream header.
3516 bzero(&ush, sizeof(ush));
3517 ush.ush_version = UMA_STREAM_VERSION;
3518 ush.ush_maxcpus = (mp_maxid + 1);
3519 ush.ush_count = count;
3520 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3522 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3523 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3524 bzero(&uth, sizeof(uth));
3526 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3527 uth.uth_align = kz->uk_align;
3528 uth.uth_size = kz->uk_size;
3529 uth.uth_rsize = kz->uk_rsize;
3530 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3532 uth.uth_maxpages += k->uk_maxpages;
3533 uth.uth_pages += k->uk_pages;
3534 uth.uth_keg_free += k->uk_free;
3535 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3540 * A zone is secondary is it is not the first entry
3541 * on the keg's zone list.
3543 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3544 (LIST_FIRST(&kz->uk_zones) != z))
3545 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3547 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3548 uth.uth_zone_free += bucket->ub_cnt;
3549 uth.uth_allocs = z->uz_allocs;
3550 uth.uth_frees = z->uz_frees;
3551 uth.uth_fails = z->uz_fails;
3552 uth.uth_sleeps = z->uz_sleeps;
3553 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3555 * While it is not normally safe to access the cache
3556 * bucket pointers while not on the CPU that owns the
3557 * cache, we only allow the pointers to be exchanged
3558 * without the zone lock held, not invalidated, so
3559 * accept the possible race associated with bucket
3560 * exchange during monitoring.
3562 for (i = 0; i < (mp_maxid + 1); i++) {
3563 bzero(&ups, sizeof(ups));
3564 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3568 cache = &z->uz_cpu[i];
3569 if (cache->uc_allocbucket != NULL)
3570 ups.ups_cache_free +=
3571 cache->uc_allocbucket->ub_cnt;
3572 if (cache->uc_freebucket != NULL)
3573 ups.ups_cache_free +=
3574 cache->uc_freebucket->ub_cnt;
3575 ups.ups_allocs = cache->uc_allocs;
3576 ups.ups_frees = cache->uc_frees;
3578 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3583 rw_runlock(&uma_rwlock);
3584 error = sbuf_finish(&sbuf);
3590 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3592 uma_zone_t zone = *(uma_zone_t *)arg1;
3595 max = uma_zone_get_max(zone);
3596 error = sysctl_handle_int(oidp, &max, 0, req);
3597 if (error || !req->newptr)
3600 uma_zone_set_max(zone, max);
3606 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3608 uma_zone_t zone = *(uma_zone_t *)arg1;
3611 cur = uma_zone_get_cur(zone);
3612 return (sysctl_handle_int(oidp, &cur, 0, req));
3617 uma_dbg_getslab(uma_zone_t zone, void *item)
3623 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
3624 if (zone->uz_flags & UMA_ZONE_VTOSLAB) {
3625 slab = vtoslab((vm_offset_t)mem);
3628 * It is safe to return the slab here even though the
3629 * zone is unlocked because the item's allocation state
3630 * essentially holds a reference.
3633 keg = LIST_FIRST(&zone->uz_kegs)->kl_keg;
3634 if (keg->uk_flags & UMA_ZONE_HASH)
3635 slab = hash_sfind(&keg->uk_hash, mem);
3637 slab = (uma_slab_t)(mem + keg->uk_pgoff);
3645 * Set up the slab's freei data such that uma_dbg_free can function.
3649 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
3654 if (zone_first_keg(zone) == NULL)
3657 slab = uma_dbg_getslab(zone, item);
3659 panic("uma: item %p did not belong to zone %s\n",
3660 item, zone->uz_name);
3663 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3665 if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3666 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
3667 item, zone, zone->uz_name, slab, freei);
3668 BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3674 * Verifies freed addresses. Checks for alignment, valid slab membership
3675 * and duplicate frees.
3679 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
3684 if (zone_first_keg(zone) == NULL)
3687 slab = uma_dbg_getslab(zone, item);
3689 panic("uma: Freed item %p did not belong to zone %s\n",
3690 item, zone->uz_name);
3693 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3695 if (freei >= keg->uk_ipers)
3696 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
3697 item, zone, zone->uz_name, slab, freei);
3699 if (((freei * keg->uk_rsize) + slab->us_data) != item)
3700 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
3701 item, zone, zone->uz_name, slab, freei);
3703 if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3704 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
3705 item, zone, zone->uz_name, slab, freei);
3707 BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3709 #endif /* INVARIANTS */
3712 DB_SHOW_COMMAND(uma, db_show_uma)
3714 uint64_t allocs, frees, sleeps;
3715 uma_bucket_t bucket;
3720 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3721 "Free", "Requests", "Sleeps", "Bucket");
3722 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3723 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3724 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3725 allocs = z->uz_allocs;
3726 frees = z->uz_frees;
3727 sleeps = z->uz_sleeps;
3730 uma_zone_sumstat(z, &cachefree, &allocs,
3732 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3733 (LIST_FIRST(&kz->uk_zones) != z)))
3734 cachefree += kz->uk_free;
3735 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3736 cachefree += bucket->ub_cnt;
3737 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3738 z->uz_name, (uintmax_t)kz->uk_size,
3739 (intmax_t)(allocs - frees), cachefree,
3740 (uintmax_t)allocs, sleeps, z->uz_count);
3747 DB_SHOW_COMMAND(umacache, db_show_umacache)
3749 uint64_t allocs, frees;
3750 uma_bucket_t bucket;
3754 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3755 "Requests", "Bucket");
3756 LIST_FOREACH(z, &uma_cachezones, uz_link) {
3757 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3758 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3759 cachefree += bucket->ub_cnt;
3760 db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3761 z->uz_name, (uintmax_t)z->uz_size,
3762 (intmax_t)(allocs - frees), cachefree,
3763 (uintmax_t)allocs, z->uz_count);