2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org>
5 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
6 * Copyright (c) 2004-2006 Robert N. M. Watson
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice unmodified, this list of conditions, and the following
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
20 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
21 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
22 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
24 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
28 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32 * uma_core.c Implementation of the Universal Memory allocator
34 * This allocator is intended to replace the multitude of similar object caches
35 * in the standard FreeBSD kernel. The intent is to be flexible as well as
36 * efficient. A primary design goal is to return unused memory to the rest of
37 * the system. This will make the system as a whole more flexible due to the
38 * ability to move memory to subsystems which most need it instead of leaving
39 * pools of reserved memory unused.
41 * The basic ideas stem from similar slab/zone based allocators whose algorithms
48 * - Improve memory usage for large allocations
49 * - Investigate cache size adjustments
52 #include <sys/cdefs.h>
53 __FBSDID("$FreeBSD$");
56 #include "opt_param.h"
59 #include <sys/param.h>
60 #include <sys/systm.h>
61 #include <sys/bitset.h>
62 #include <sys/eventhandler.h>
63 #include <sys/kernel.h>
64 #include <sys/types.h>
65 #include <sys/limits.h>
66 #include <sys/queue.h>
67 #include <sys/malloc.h>
70 #include <sys/sysctl.h>
71 #include <sys/mutex.h>
73 #include <sys/random.h>
74 #include <sys/rwlock.h>
76 #include <sys/sched.h>
78 #include <sys/taskqueue.h>
79 #include <sys/vmmeter.h>
82 #include <vm/vm_object.h>
83 #include <vm/vm_page.h>
84 #include <vm/vm_pageout.h>
85 #include <vm/vm_param.h>
86 #include <vm/vm_phys.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_extern.h>
91 #include <vm/uma_int.h>
92 #include <vm/uma_dbg.h>
97 #include <vm/memguard.h>
101 * This is the zone and keg from which all zones are spawned.
103 static uma_zone_t kegs;
104 static uma_zone_t zones;
106 /* This is the zone from which all offpage uma_slab_ts are allocated. */
107 static uma_zone_t slabzone;
110 * The initial hash tables come out of this zone so they can be allocated
111 * prior to malloc coming up.
113 static uma_zone_t hashzone;
115 /* The boot-time adjusted value for cache line alignment. */
116 int uma_align_cache = 64 - 1;
118 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
121 * Are we allowed to allocate buckets?
123 static int bucketdisable = 1;
125 /* Linked list of all kegs in the system */
126 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
128 /* Linked list of all cache-only zones in the system */
129 static LIST_HEAD(,uma_zone) uma_cachezones =
130 LIST_HEAD_INITIALIZER(uma_cachezones);
132 /* This RW lock protects the keg list */
133 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
136 * Pointer and counter to pool of pages, that is preallocated at
137 * startup to bootstrap UMA.
139 static char *bootmem;
140 static int boot_pages;
142 static struct sx uma_drain_lock;
144 /* kmem soft limit. */
145 static unsigned long uma_kmem_limit = LONG_MAX;
146 static volatile unsigned long uma_kmem_total;
148 /* Is the VM done starting up? */
149 static enum { BOOT_COLD = 0, BOOT_STRAPPED, BOOT_PAGEALLOC, BOOT_BUCKETS,
150 BOOT_RUNNING } booted = BOOT_COLD;
153 * This is the handle used to schedule events that need to happen
154 * outside of the allocation fast path.
156 static struct callout uma_callout;
157 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
160 * This structure is passed as the zone ctor arg so that I don't have to create
161 * a special allocation function just for zones.
163 struct uma_zctor_args {
178 struct uma_kctor_args {
187 struct uma_bucket_zone {
190 int ubz_entries; /* Number of items it can hold. */
191 int ubz_maxsize; /* Maximum allocation size per-item. */
195 * Compute the actual number of bucket entries to pack them in power
196 * of two sizes for more efficient space utilization.
198 #define BUCKET_SIZE(n) \
199 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
201 #define BUCKET_MAX BUCKET_SIZE(256)
203 struct uma_bucket_zone bucket_zones[] = {
204 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
205 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
206 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
207 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
208 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
209 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
210 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
211 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
212 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
217 * Flags and enumerations to be passed to internal functions.
219 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
221 #define UMA_ANYDOMAIN -1 /* Special value for domain search. */
225 int uma_startup_count(int);
226 void uma_startup(void *, int);
227 void uma_startup1(void);
228 void uma_startup2(void);
230 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
231 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
232 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
233 static void page_free(void *, vm_size_t, uint8_t);
234 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int);
235 static void cache_drain(uma_zone_t);
236 static void bucket_drain(uma_zone_t, uma_bucket_t);
237 static void bucket_cache_drain(uma_zone_t zone);
238 static int keg_ctor(void *, int, void *, int);
239 static void keg_dtor(void *, int, void *);
240 static int zone_ctor(void *, int, void *, int);
241 static void zone_dtor(void *, int, void *);
242 static int zero_init(void *, int, int);
243 static void keg_small_init(uma_keg_t keg);
244 static void keg_large_init(uma_keg_t keg);
245 static void zone_foreach(void (*zfunc)(uma_zone_t));
246 static void zone_timeout(uma_zone_t zone);
247 static int hash_alloc(struct uma_hash *);
248 static int hash_expand(struct uma_hash *, struct uma_hash *);
249 static void hash_free(struct uma_hash *hash);
250 static void uma_timeout(void *);
251 static void uma_startup3(void);
252 static void *zone_alloc_item(uma_zone_t, void *, int, int);
253 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
254 static void bucket_enable(void);
255 static void bucket_init(void);
256 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
257 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
258 static void bucket_zone_drain(void);
259 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
260 static uma_slab_t zone_fetch_slab(uma_zone_t, uma_keg_t, int, int);
261 static uma_slab_t zone_fetch_slab_multi(uma_zone_t, uma_keg_t, int, int);
262 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
263 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
264 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
265 uma_fini fini, int align, uint32_t flags);
266 static int zone_import(uma_zone_t, void **, int, int, int);
267 static void zone_release(uma_zone_t, void **, int);
268 static void uma_zero_item(void *, uma_zone_t);
270 void uma_print_zone(uma_zone_t);
271 void uma_print_stats(void);
272 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
273 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
276 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
277 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
280 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
282 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
283 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
285 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
286 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
288 static int zone_warnings = 1;
289 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
290 "Warn when UMA zones becomes full");
292 /* Adjust bytes under management by UMA. */
294 uma_total_dec(unsigned long size)
297 atomic_subtract_long(&uma_kmem_total, size);
301 uma_total_inc(unsigned long size)
304 if (atomic_fetchadd_long(&uma_kmem_total, size) > uma_kmem_limit)
305 uma_reclaim_wakeup();
309 * This routine checks to see whether or not it's safe to enable buckets.
314 bucketdisable = vm_page_count_min();
318 * Initialize bucket_zones, the array of zones of buckets of various sizes.
320 * For each zone, calculate the memory required for each bucket, consisting
321 * of the header and an array of pointers.
326 struct uma_bucket_zone *ubz;
329 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
330 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
331 size += sizeof(void *) * ubz->ubz_entries;
332 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
333 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
334 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET | UMA_ZONE_NUMA);
339 * Given a desired number of entries for a bucket, return the zone from which
340 * to allocate the bucket.
342 static struct uma_bucket_zone *
343 bucket_zone_lookup(int entries)
345 struct uma_bucket_zone *ubz;
347 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
348 if (ubz->ubz_entries >= entries)
355 bucket_select(int size)
357 struct uma_bucket_zone *ubz;
359 ubz = &bucket_zones[0];
360 if (size > ubz->ubz_maxsize)
361 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
363 for (; ubz->ubz_entries != 0; ubz++)
364 if (ubz->ubz_maxsize < size)
367 return (ubz->ubz_entries);
371 bucket_alloc(uma_zone_t zone, void *udata, int flags)
373 struct uma_bucket_zone *ubz;
377 * This is to stop us from allocating per cpu buckets while we're
378 * running out of vm.boot_pages. Otherwise, we would exhaust the
379 * boot pages. This also prevents us from allocating buckets in
380 * low memory situations.
385 * To limit bucket recursion we store the original zone flags
386 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
387 * NOVM flag to persist even through deep recursions. We also
388 * store ZFLAG_BUCKET once we have recursed attempting to allocate
389 * a bucket for a bucket zone so we do not allow infinite bucket
390 * recursion. This cookie will even persist to frees of unused
391 * buckets via the allocation path or bucket allocations in the
394 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
395 udata = (void *)(uintptr_t)zone->uz_flags;
397 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
399 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
401 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
403 ubz = bucket_zone_lookup(zone->uz_count);
404 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
406 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
409 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
412 bucket->ub_entries = ubz->ubz_entries;
419 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
421 struct uma_bucket_zone *ubz;
423 KASSERT(bucket->ub_cnt == 0,
424 ("bucket_free: Freeing a non free bucket."));
425 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
426 udata = (void *)(uintptr_t)zone->uz_flags;
427 ubz = bucket_zone_lookup(bucket->ub_entries);
428 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
432 bucket_zone_drain(void)
434 struct uma_bucket_zone *ubz;
436 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
437 zone_drain(ubz->ubz_zone);
441 zone_log_warning(uma_zone_t zone)
443 static const struct timeval warninterval = { 300, 0 };
445 if (!zone_warnings || zone->uz_warning == NULL)
448 if (ratecheck(&zone->uz_ratecheck, &warninterval))
449 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
453 zone_maxaction(uma_zone_t zone)
456 if (zone->uz_maxaction.ta_func != NULL)
457 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
461 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
465 LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
466 kegfn(klink->kl_keg);
470 * Routine called by timeout which is used to fire off some time interval
471 * based calculations. (stats, hash size, etc.)
480 uma_timeout(void *unused)
483 zone_foreach(zone_timeout);
485 /* Reschedule this event */
486 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
490 * Routine to perform timeout driven calculations. This expands the
491 * hashes and does per cpu statistics aggregation.
496 keg_timeout(uma_keg_t keg)
501 * Expand the keg hash table.
503 * This is done if the number of slabs is larger than the hash size.
504 * What I'm trying to do here is completely reduce collisions. This
505 * may be a little aggressive. Should I allow for two collisions max?
507 if (keg->uk_flags & UMA_ZONE_HASH &&
508 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
509 struct uma_hash newhash;
510 struct uma_hash oldhash;
514 * This is so involved because allocating and freeing
515 * while the keg lock is held will lead to deadlock.
516 * I have to do everything in stages and check for
519 newhash = keg->uk_hash;
521 ret = hash_alloc(&newhash);
524 if (hash_expand(&keg->uk_hash, &newhash)) {
525 oldhash = keg->uk_hash;
526 keg->uk_hash = newhash;
539 zone_timeout(uma_zone_t zone)
542 zone_foreach_keg(zone, &keg_timeout);
546 * Allocate and zero fill the next sized hash table from the appropriate
550 * hash A new hash structure with the old hash size in uh_hashsize
553 * 1 on success and 0 on failure.
556 hash_alloc(struct uma_hash *hash)
561 oldsize = hash->uh_hashsize;
563 /* We're just going to go to a power of two greater */
565 hash->uh_hashsize = oldsize * 2;
566 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
567 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
568 M_UMAHASH, M_NOWAIT);
570 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
571 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
572 UMA_ANYDOMAIN, M_WAITOK);
573 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
575 if (hash->uh_slab_hash) {
576 bzero(hash->uh_slab_hash, alloc);
577 hash->uh_hashmask = hash->uh_hashsize - 1;
585 * Expands the hash table for HASH zones. This is done from zone_timeout
586 * to reduce collisions. This must not be done in the regular allocation
587 * path, otherwise, we can recurse on the vm while allocating pages.
590 * oldhash The hash you want to expand
591 * newhash The hash structure for the new table
599 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
605 if (!newhash->uh_slab_hash)
608 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
612 * I need to investigate hash algorithms for resizing without a
616 for (i = 0; i < oldhash->uh_hashsize; i++)
617 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
618 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
619 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
620 hval = UMA_HASH(newhash, slab->us_data);
621 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
629 * Free the hash bucket to the appropriate backing store.
632 * slab_hash The hash bucket we're freeing
633 * hashsize The number of entries in that hash bucket
639 hash_free(struct uma_hash *hash)
641 if (hash->uh_slab_hash == NULL)
643 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
644 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
646 free(hash->uh_slab_hash, M_UMAHASH);
650 * Frees all outstanding items in a bucket
653 * zone The zone to free to, must be unlocked.
654 * bucket The free/alloc bucket with items, cpu queue must be locked.
661 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
669 for (i = 0; i < bucket->ub_cnt; i++)
670 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
671 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
676 * Drains the per cpu caches for a zone.
678 * NOTE: This may only be called while the zone is being turn down, and not
679 * during normal operation. This is necessary in order that we do not have
680 * to migrate CPUs to drain the per-CPU caches.
683 * zone The zone to drain, must be unlocked.
689 cache_drain(uma_zone_t zone)
695 * XXX: It is safe to not lock the per-CPU caches, because we're
696 * tearing down the zone anyway. I.e., there will be no further use
697 * of the caches at this point.
699 * XXX: It would good to be able to assert that the zone is being
700 * torn down to prevent improper use of cache_drain().
702 * XXX: We lock the zone before passing into bucket_cache_drain() as
703 * it is used elsewhere. Should the tear-down path be made special
704 * there in some form?
707 cache = &zone->uz_cpu[cpu];
708 bucket_drain(zone, cache->uc_allocbucket);
709 bucket_drain(zone, cache->uc_freebucket);
710 if (cache->uc_allocbucket != NULL)
711 bucket_free(zone, cache->uc_allocbucket, NULL);
712 if (cache->uc_freebucket != NULL)
713 bucket_free(zone, cache->uc_freebucket, NULL);
714 cache->uc_allocbucket = cache->uc_freebucket = NULL;
717 bucket_cache_drain(zone);
722 cache_shrink(uma_zone_t zone)
725 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
729 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
734 cache_drain_safe_cpu(uma_zone_t zone)
740 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
746 if (zone->uz_flags & UMA_ZONE_NUMA)
747 domain = PCPU_GET(domain);
750 cache = &zone->uz_cpu[curcpu];
751 if (cache->uc_allocbucket) {
752 if (cache->uc_allocbucket->ub_cnt != 0)
753 LIST_INSERT_HEAD(&zone->uz_domain[domain].uzd_buckets,
754 cache->uc_allocbucket, ub_link);
756 b1 = cache->uc_allocbucket;
757 cache->uc_allocbucket = NULL;
759 if (cache->uc_freebucket) {
760 if (cache->uc_freebucket->ub_cnt != 0)
761 LIST_INSERT_HEAD(&zone->uz_domain[domain].uzd_buckets,
762 cache->uc_freebucket, ub_link);
764 b2 = cache->uc_freebucket;
765 cache->uc_freebucket = NULL;
770 bucket_free(zone, b1, NULL);
772 bucket_free(zone, b2, NULL);
776 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
777 * This is an expensive call because it needs to bind to all CPUs
778 * one by one and enter a critical section on each of them in order
779 * to safely access their cache buckets.
780 * Zone lock must not be held on call this function.
783 cache_drain_safe(uma_zone_t zone)
788 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
793 zone_foreach(cache_shrink);
796 thread_lock(curthread);
797 sched_bind(curthread, cpu);
798 thread_unlock(curthread);
801 cache_drain_safe_cpu(zone);
803 zone_foreach(cache_drain_safe_cpu);
805 thread_lock(curthread);
806 sched_unbind(curthread);
807 thread_unlock(curthread);
811 * Drain the cached buckets from a zone. Expects a locked zone on entry.
814 bucket_cache_drain(uma_zone_t zone)
816 uma_zone_domain_t zdom;
821 * Drain the bucket queues and free the buckets.
823 for (i = 0; i < vm_ndomains; i++) {
824 zdom = &zone->uz_domain[i];
825 while ((bucket = LIST_FIRST(&zdom->uzd_buckets)) != NULL) {
826 LIST_REMOVE(bucket, ub_link);
828 bucket_drain(zone, bucket);
829 bucket_free(zone, bucket, NULL);
835 * Shrink further bucket sizes. Price of single zone lock collision
836 * is probably lower then price of global cache drain.
838 if (zone->uz_count > zone->uz_count_min)
843 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
849 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
850 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
853 flags = slab->us_flags;
855 if (keg->uk_fini != NULL) {
856 for (i--; i > -1; i--)
857 keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
860 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
861 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
862 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
863 uma_total_dec(PAGE_SIZE * keg->uk_ppera);
867 * Frees pages from a keg back to the system. This is done on demand from
868 * the pageout daemon.
873 keg_drain(uma_keg_t keg)
875 struct slabhead freeslabs = { 0 };
877 uma_slab_t slab, tmp;
881 * We don't want to take pages from statically allocated kegs at this
884 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
887 CTR3(KTR_UMA, "keg_drain %s(%p) free items: %u",
888 keg->uk_name, keg, keg->uk_free);
890 if (keg->uk_free == 0)
893 for (i = 0; i < vm_ndomains; i++) {
894 dom = &keg->uk_domain[i];
895 LIST_FOREACH_SAFE(slab, &dom->ud_free_slab, us_link, tmp) {
896 /* We have nowhere to free these to. */
897 if (slab->us_flags & UMA_SLAB_BOOT)
900 LIST_REMOVE(slab, us_link);
901 keg->uk_pages -= keg->uk_ppera;
902 keg->uk_free -= keg->uk_ipers;
904 if (keg->uk_flags & UMA_ZONE_HASH)
905 UMA_HASH_REMOVE(&keg->uk_hash, slab,
908 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
915 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
916 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
917 keg_free_slab(keg, slab, keg->uk_ipers);
922 zone_drain_wait(uma_zone_t zone, int waitok)
926 * Set draining to interlock with zone_dtor() so we can release our
927 * locks as we go. Only dtor() should do a WAITOK call since it
928 * is the only call that knows the structure will still be available
932 while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
933 if (waitok == M_NOWAIT)
935 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
937 zone->uz_flags |= UMA_ZFLAG_DRAINING;
938 bucket_cache_drain(zone);
941 * The DRAINING flag protects us from being freed while
942 * we're running. Normally the uma_rwlock would protect us but we
943 * must be able to release and acquire the right lock for each keg.
945 zone_foreach_keg(zone, &keg_drain);
947 zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
954 zone_drain(uma_zone_t zone)
957 zone_drain_wait(zone, M_NOWAIT);
961 * Allocate a new slab for a keg. This does not insert the slab onto a list.
964 * wait Shall we wait?
967 * The slab that was allocated or NULL if there is no memory and the
968 * caller specified M_NOWAIT.
971 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int wait)
980 KASSERT(domain >= 0 && domain < vm_ndomains,
981 ("keg_alloc_slab: domain %d out of range", domain));
982 mtx_assert(&keg->uk_lock, MA_OWNED);
986 allocf = keg->uk_allocf;
988 size = keg->uk_ppera * PAGE_SIZE;
990 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
991 slab = zone_alloc_item(keg->uk_slabzone, NULL, domain, wait);
997 * This reproduces the old vm_zone behavior of zero filling pages the
998 * first time they are added to a zone.
1000 * Malloced items are zeroed in uma_zalloc.
1003 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1008 if (keg->uk_flags & UMA_ZONE_NODUMP)
1011 /* zone is passed for legacy reasons. */
1012 mem = allocf(zone, size, domain, &flags, wait);
1014 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1015 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
1019 uma_total_inc(size);
1021 /* Point the slab into the allocated memory */
1022 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
1023 slab = (uma_slab_t )(mem + keg->uk_pgoff);
1025 if (keg->uk_flags & UMA_ZONE_VTOSLAB)
1026 for (i = 0; i < keg->uk_ppera; i++)
1027 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
1030 slab->us_data = mem;
1031 slab->us_freecount = keg->uk_ipers;
1032 slab->us_flags = flags;
1033 slab->us_domain = domain;
1034 BIT_FILL(SLAB_SETSIZE, &slab->us_free);
1036 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1039 if (keg->uk_init != NULL) {
1040 for (i = 0; i < keg->uk_ipers; i++)
1041 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1042 keg->uk_size, wait) != 0)
1044 if (i != keg->uk_ipers) {
1045 keg_free_slab(keg, slab, i);
1053 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1054 slab, keg->uk_name, keg);
1057 if (keg->uk_flags & UMA_ZONE_HASH)
1058 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1060 keg->uk_pages += keg->uk_ppera;
1061 keg->uk_free += keg->uk_ipers;
1068 * This function is intended to be used early on in place of page_alloc() so
1069 * that we may use the boot time page cache to satisfy allocations before
1073 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1080 keg = zone_first_keg(zone);
1083 * If we are in BOOT_BUCKETS or higher, than switch to real
1084 * allocator. Zones with page sized slabs switch at BOOT_PAGEALLOC.
1090 case BOOT_PAGEALLOC:
1091 if (keg->uk_ppera > 1)
1095 #ifdef UMA_MD_SMALL_ALLOC
1096 keg->uk_allocf = (keg->uk_ppera > 1) ?
1097 page_alloc : uma_small_alloc;
1099 keg->uk_allocf = page_alloc;
1101 return keg->uk_allocf(zone, bytes, domain, pflag, wait);
1105 * Check our small startup cache to see if it has pages remaining.
1107 pages = howmany(bytes, PAGE_SIZE);
1108 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1109 if (pages > boot_pages)
1110 panic("UMA zone \"%s\": Increase vm.boot_pages", zone->uz_name);
1112 printf("%s from \"%s\", %d boot pages left\n", __func__, zone->uz_name,
1116 boot_pages -= pages;
1117 bootmem += pages * PAGE_SIZE;
1118 *pflag = UMA_SLAB_BOOT;
1124 * Allocates a number of pages from the system
1127 * bytes The number of bytes requested
1128 * wait Shall we wait?
1131 * A pointer to the alloced memory or possibly
1132 * NULL if M_NOWAIT is set.
1135 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1138 void *p; /* Returned page */
1140 *pflag = UMA_SLAB_KERNEL;
1141 p = (void *) kmem_malloc_domain(domain, bytes, wait);
1147 * Allocates a number of pages from within an object
1150 * bytes The number of bytes requested
1151 * wait Shall we wait?
1154 * A pointer to the alloced memory or possibly
1155 * NULL if M_NOWAIT is set.
1158 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1161 TAILQ_HEAD(, vm_page) alloctail;
1163 vm_offset_t retkva, zkva;
1164 vm_page_t p, p_next;
1167 TAILQ_INIT(&alloctail);
1168 keg = zone_first_keg(zone);
1170 npages = howmany(bytes, PAGE_SIZE);
1171 while (npages > 0) {
1172 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1173 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1174 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1178 * Since the page does not belong to an object, its
1181 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1186 * Page allocation failed, free intermediate pages and
1189 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1190 vm_page_unwire(p, PQ_NONE);
1195 *flags = UMA_SLAB_PRIV;
1196 zkva = keg->uk_kva +
1197 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1199 TAILQ_FOREACH(p, &alloctail, listq) {
1200 pmap_qenter(zkva, &p, 1);
1204 return ((void *)retkva);
1208 * Frees a number of pages to the system
1211 * mem A pointer to the memory to be freed
1212 * size The size of the memory being freed
1213 * flags The original p->us_flags field
1219 page_free(void *mem, vm_size_t size, uint8_t flags)
1223 if (flags & UMA_SLAB_KERNEL)
1224 vmem = kernel_arena;
1226 panic("UMA: page_free used with invalid flags %x", flags);
1228 kmem_free(vmem, (vm_offset_t)mem, size);
1232 * Zero fill initializer
1234 * Arguments/Returns follow uma_init specifications
1237 zero_init(void *mem, int size, int flags)
1244 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1247 * keg The zone we should initialize
1253 keg_small_init(uma_keg_t keg)
1261 if (keg->uk_flags & UMA_ZONE_PCPU) {
1262 u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU;
1264 slabsize = sizeof(struct pcpu);
1265 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1268 slabsize = UMA_SLAB_SIZE;
1273 * Calculate the size of each allocation (rsize) according to
1274 * alignment. If the requested size is smaller than we have
1275 * allocation bits for we round it up.
1277 rsize = keg->uk_size;
1278 if (rsize < slabsize / SLAB_SETSIZE)
1279 rsize = slabsize / SLAB_SETSIZE;
1280 if (rsize & keg->uk_align)
1281 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1282 keg->uk_rsize = rsize;
1284 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1285 keg->uk_rsize < sizeof(struct pcpu),
1286 ("%s: size %u too large", __func__, keg->uk_rsize));
1288 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1291 shsize = sizeof(struct uma_slab);
1293 keg->uk_ipers = (slabsize - shsize) / rsize;
1294 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1295 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1297 memused = keg->uk_ipers * rsize + shsize;
1298 wastedspace = slabsize - memused;
1301 * We can't do OFFPAGE if we're internal or if we've been
1302 * asked to not go to the VM for buckets. If we do this we
1303 * may end up going to the VM for slabs which we do not
1304 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1305 * of UMA_ZONE_VM, which clearly forbids it.
1307 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1308 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1312 * See if using an OFFPAGE slab will limit our waste. Only do
1313 * this if it permits more items per-slab.
1315 * XXX We could try growing slabsize to limit max waste as well.
1316 * Historically this was not done because the VM could not
1317 * efficiently handle contiguous allocations.
1319 if ((wastedspace >= slabsize / UMA_MAX_WASTE) &&
1320 (keg->uk_ipers < (slabsize / keg->uk_rsize))) {
1321 keg->uk_ipers = slabsize / keg->uk_rsize;
1322 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1323 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1324 CTR6(KTR_UMA, "UMA decided we need offpage slab headers for "
1325 "keg: %s(%p), calculated wastedspace = %d, "
1326 "maximum wasted space allowed = %d, "
1327 "calculated ipers = %d, "
1328 "new wasted space = %d\n", keg->uk_name, keg, wastedspace,
1329 slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1330 slabsize - keg->uk_ipers * keg->uk_rsize);
1331 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1334 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1335 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1336 keg->uk_flags |= UMA_ZONE_HASH;
1340 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1341 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1345 * keg The keg we should initialize
1351 keg_large_init(uma_keg_t keg)
1355 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1356 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1357 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1358 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1359 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1361 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1363 keg->uk_rsize = keg->uk_size;
1365 /* Check whether we have enough space to not do OFFPAGE. */
1366 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1367 shsize = sizeof(struct uma_slab);
1368 if (shsize & UMA_ALIGN_PTR)
1369 shsize = (shsize & ~UMA_ALIGN_PTR) +
1370 (UMA_ALIGN_PTR + 1);
1372 if (PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < shsize) {
1374 * We can't do OFFPAGE if we're internal, in which case
1375 * we need an extra page per allocation to contain the
1378 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) == 0)
1379 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1385 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1386 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1387 keg->uk_flags |= UMA_ZONE_HASH;
1391 keg_cachespread_init(uma_keg_t keg)
1398 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1399 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1401 alignsize = keg->uk_align + 1;
1402 rsize = keg->uk_size;
1404 * We want one item to start on every align boundary in a page. To
1405 * do this we will span pages. We will also extend the item by the
1406 * size of align if it is an even multiple of align. Otherwise, it
1407 * would fall on the same boundary every time.
1409 if (rsize & keg->uk_align)
1410 rsize = (rsize & ~keg->uk_align) + alignsize;
1411 if ((rsize & alignsize) == 0)
1413 trailer = rsize - keg->uk_size;
1414 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1415 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1416 keg->uk_rsize = rsize;
1417 keg->uk_ppera = pages;
1418 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1419 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1420 KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1421 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1426 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1427 * the keg onto the global keg list.
1429 * Arguments/Returns follow uma_ctor specifications
1430 * udata Actually uma_kctor_args
1433 keg_ctor(void *mem, int size, void *udata, int flags)
1435 struct uma_kctor_args *arg = udata;
1436 uma_keg_t keg = mem;
1440 keg->uk_size = arg->size;
1441 keg->uk_init = arg->uminit;
1442 keg->uk_fini = arg->fini;
1443 keg->uk_align = arg->align;
1446 keg->uk_reserve = 0;
1448 keg->uk_flags = arg->flags;
1449 keg->uk_slabzone = NULL;
1452 * The master zone is passed to us at keg-creation time.
1455 keg->uk_name = zone->uz_name;
1457 if (arg->flags & UMA_ZONE_VM)
1458 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1460 if (arg->flags & UMA_ZONE_ZINIT)
1461 keg->uk_init = zero_init;
1463 if (arg->flags & UMA_ZONE_MALLOC)
1464 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1466 if (arg->flags & UMA_ZONE_PCPU)
1468 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1470 keg->uk_flags &= ~UMA_ZONE_PCPU;
1473 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1474 keg_cachespread_init(keg);
1476 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1477 keg_large_init(keg);
1479 keg_small_init(keg);
1482 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1483 keg->uk_slabzone = slabzone;
1486 * If we haven't booted yet we need allocations to go through the
1487 * startup cache until the vm is ready.
1489 if (booted < BOOT_PAGEALLOC)
1490 keg->uk_allocf = startup_alloc;
1491 #ifdef UMA_MD_SMALL_ALLOC
1492 else if (keg->uk_ppera == 1)
1493 keg->uk_allocf = uma_small_alloc;
1496 keg->uk_allocf = page_alloc;
1497 #ifdef UMA_MD_SMALL_ALLOC
1498 if (keg->uk_ppera == 1)
1499 keg->uk_freef = uma_small_free;
1502 keg->uk_freef = page_free;
1505 * Initialize keg's lock
1507 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1510 * If we're putting the slab header in the actual page we need to
1511 * figure out where in each page it goes. This calculates a right
1512 * justified offset into the memory on an ALIGN_PTR boundary.
1514 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1517 /* Size of the slab struct and free list */
1518 totsize = sizeof(struct uma_slab);
1520 if (totsize & UMA_ALIGN_PTR)
1521 totsize = (totsize & ~UMA_ALIGN_PTR) +
1522 (UMA_ALIGN_PTR + 1);
1523 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1526 * The only way the following is possible is if with our
1527 * UMA_ALIGN_PTR adjustments we are now bigger than
1528 * UMA_SLAB_SIZE. I haven't checked whether this is
1529 * mathematically possible for all cases, so we make
1532 totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1533 if (totsize > PAGE_SIZE * keg->uk_ppera) {
1534 printf("zone %s ipers %d rsize %d size %d\n",
1535 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1537 panic("UMA slab won't fit.");
1541 if (keg->uk_flags & UMA_ZONE_HASH)
1542 hash_alloc(&keg->uk_hash);
1544 CTR5(KTR_UMA, "keg_ctor %p zone %s(%p) out %d free %d\n",
1545 keg, zone->uz_name, zone,
1546 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free,
1549 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1551 rw_wlock(&uma_rwlock);
1552 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1553 rw_wunlock(&uma_rwlock);
1558 * Zone header ctor. This initializes all fields, locks, etc.
1560 * Arguments/Returns follow uma_ctor specifications
1561 * udata Actually uma_zctor_args
1564 zone_ctor(void *mem, int size, void *udata, int flags)
1566 struct uma_zctor_args *arg = udata;
1567 uma_zone_t zone = mem;
1572 zone->uz_name = arg->name;
1573 zone->uz_ctor = arg->ctor;
1574 zone->uz_dtor = arg->dtor;
1575 zone->uz_slab = zone_fetch_slab;
1576 zone->uz_init = NULL;
1577 zone->uz_fini = NULL;
1578 zone->uz_allocs = 0;
1581 zone->uz_sleeps = 0;
1583 zone->uz_count_min = 0;
1585 zone->uz_warning = NULL;
1586 /* The domain structures follow the cpu structures. */
1587 zone->uz_domain = (struct uma_zone_domain *)&zone->uz_cpu[mp_ncpus];
1588 timevalclear(&zone->uz_ratecheck);
1591 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1594 * This is a pure cache zone, no kegs.
1597 if (arg->flags & UMA_ZONE_VM)
1598 arg->flags |= UMA_ZFLAG_CACHEONLY;
1599 zone->uz_flags = arg->flags;
1600 zone->uz_size = arg->size;
1601 zone->uz_import = arg->import;
1602 zone->uz_release = arg->release;
1603 zone->uz_arg = arg->arg;
1604 zone->uz_lockptr = &zone->uz_lock;
1605 rw_wlock(&uma_rwlock);
1606 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1607 rw_wunlock(&uma_rwlock);
1612 * Use the regular zone/keg/slab allocator.
1614 zone->uz_import = (uma_import)zone_import;
1615 zone->uz_release = (uma_release)zone_release;
1616 zone->uz_arg = zone;
1618 if (arg->flags & UMA_ZONE_SECONDARY) {
1619 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1620 zone->uz_init = arg->uminit;
1621 zone->uz_fini = arg->fini;
1622 zone->uz_lockptr = &keg->uk_lock;
1623 zone->uz_flags |= UMA_ZONE_SECONDARY;
1624 rw_wlock(&uma_rwlock);
1626 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1627 if (LIST_NEXT(z, uz_link) == NULL) {
1628 LIST_INSERT_AFTER(z, zone, uz_link);
1633 rw_wunlock(&uma_rwlock);
1634 } else if (keg == NULL) {
1635 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1636 arg->align, arg->flags)) == NULL)
1639 struct uma_kctor_args karg;
1642 /* We should only be here from uma_startup() */
1643 karg.size = arg->size;
1644 karg.uminit = arg->uminit;
1645 karg.fini = arg->fini;
1646 karg.align = arg->align;
1647 karg.flags = arg->flags;
1649 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1656 * Link in the first keg.
1658 zone->uz_klink.kl_keg = keg;
1659 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1660 zone->uz_lockptr = &keg->uk_lock;
1661 zone->uz_size = keg->uk_size;
1662 zone->uz_flags |= (keg->uk_flags &
1663 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1666 * Some internal zones don't have room allocated for the per cpu
1667 * caches. If we're internal, bail out here.
1669 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1670 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1671 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1676 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1677 zone->uz_count = bucket_select(zone->uz_size);
1679 zone->uz_count = BUCKET_MAX;
1680 zone->uz_count_min = zone->uz_count;
1686 * Keg header dtor. This frees all data, destroys locks, frees the hash
1687 * table and removes the keg from the global list.
1689 * Arguments/Returns follow uma_dtor specifications
1693 keg_dtor(void *arg, int size, void *udata)
1697 keg = (uma_keg_t)arg;
1699 if (keg->uk_free != 0) {
1700 printf("Freed UMA keg (%s) was not empty (%d items). "
1701 " Lost %d pages of memory.\n",
1702 keg->uk_name ? keg->uk_name : "",
1703 keg->uk_free, keg->uk_pages);
1707 hash_free(&keg->uk_hash);
1715 * Arguments/Returns follow uma_dtor specifications
1719 zone_dtor(void *arg, int size, void *udata)
1725 zone = (uma_zone_t)arg;
1726 keg = zone_first_keg(zone);
1728 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1731 rw_wlock(&uma_rwlock);
1732 LIST_REMOVE(zone, uz_link);
1733 rw_wunlock(&uma_rwlock);
1735 * XXX there are some races here where
1736 * the zone can be drained but zone lock
1737 * released and then refilled before we
1738 * remove it... we dont care for now
1740 zone_drain_wait(zone, M_WAITOK);
1742 * Unlink all of our kegs.
1744 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1745 klink->kl_keg = NULL;
1746 LIST_REMOVE(klink, kl_link);
1747 if (klink == &zone->uz_klink)
1749 free(klink, M_TEMP);
1752 * We only destroy kegs from non secondary zones.
1754 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1755 rw_wlock(&uma_rwlock);
1756 LIST_REMOVE(keg, uk_link);
1757 rw_wunlock(&uma_rwlock);
1758 zone_free_item(kegs, keg, NULL, SKIP_NONE);
1760 ZONE_LOCK_FINI(zone);
1764 * Traverses every zone in the system and calls a callback
1767 * zfunc A pointer to a function which accepts a zone
1774 zone_foreach(void (*zfunc)(uma_zone_t))
1779 rw_rlock(&uma_rwlock);
1780 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1781 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1784 rw_runlock(&uma_rwlock);
1788 * Count how many pages do we need to bootstrap. VM supplies
1789 * its need in early zones in the argument, we add up our zones,
1790 * which consist of: UMA Slabs, UMA Hash and 9 Bucket zones. The
1791 * zone of zones and zone of kegs are accounted separately.
1793 #define UMA_BOOT_ZONES 11
1794 /* Zone of zones and zone of kegs have arbitrary alignment. */
1795 #define UMA_BOOT_ALIGN 32
1796 static int zsize, ksize;
1798 uma_startup_count(int vm_zones)
1802 ksize = sizeof(struct uma_keg) +
1803 (sizeof(struct uma_domain) * vm_ndomains);
1804 zsize = sizeof(struct uma_zone) +
1805 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
1806 (sizeof(struct uma_zone_domain) * vm_ndomains);
1809 * Memory for the zone of kegs and its keg,
1810 * and for zone of zones.
1812 pages = howmany(roundup(zsize, CACHE_LINE_SIZE) * 2 +
1813 roundup(ksize, CACHE_LINE_SIZE), PAGE_SIZE);
1815 #ifdef UMA_MD_SMALL_ALLOC
1816 zones = UMA_BOOT_ZONES;
1818 zones = UMA_BOOT_ZONES + vm_zones;
1822 /* Memory for the rest of startup zones, UMA and VM, ... */
1823 if (zsize > UMA_SLAB_SIZE)
1824 pages += (zones + vm_zones) *
1825 howmany(roundup2(zsize, UMA_BOOT_ALIGN), UMA_SLAB_SIZE);
1827 pages += howmany(zones,
1828 UMA_SLAB_SPACE / roundup2(zsize, UMA_BOOT_ALIGN));
1830 /* ... and their kegs. Note that zone of zones allocates a keg! */
1831 pages += howmany(zones + 1,
1832 UMA_SLAB_SPACE / roundup2(ksize, UMA_BOOT_ALIGN));
1835 * Most of startup zones are not going to be offpages, that's
1836 * why we use UMA_SLAB_SPACE instead of UMA_SLAB_SIZE in all
1837 * calculations. Some large bucket zones will be offpage, and
1838 * thus will allocate hashes. We take conservative approach
1839 * and assume that all zones may allocate hash. This may give
1840 * us some positive inaccuracy, usually an extra single page.
1842 pages += howmany(zones, UMA_SLAB_SPACE /
1843 (sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT));
1849 uma_startup(void *mem, int npages)
1851 struct uma_zctor_args args;
1852 uma_keg_t masterkeg;
1856 printf("Entering %s with %d boot pages configured\n", __func__, npages);
1859 rw_init(&uma_rwlock, "UMA lock");
1861 /* Use bootpages memory for the zone of zones and zone of kegs. */
1863 zones = (uma_zone_t)m;
1864 m += roundup(zsize, CACHE_LINE_SIZE);
1865 kegs = (uma_zone_t)m;
1866 m += roundup(zsize, CACHE_LINE_SIZE);
1867 masterkeg = (uma_keg_t)m;
1868 m += roundup(ksize, CACHE_LINE_SIZE);
1869 m = roundup(m, PAGE_SIZE);
1870 npages -= (m - (uintptr_t)mem) / PAGE_SIZE;
1873 /* "manually" create the initial zone */
1874 memset(&args, 0, sizeof(args));
1875 args.name = "UMA Kegs";
1877 args.ctor = keg_ctor;
1878 args.dtor = keg_dtor;
1879 args.uminit = zero_init;
1881 args.keg = masterkeg;
1882 args.align = UMA_BOOT_ALIGN - 1;
1883 args.flags = UMA_ZFLAG_INTERNAL;
1884 zone_ctor(kegs, zsize, &args, M_WAITOK);
1887 boot_pages = npages;
1889 args.name = "UMA Zones";
1891 args.ctor = zone_ctor;
1892 args.dtor = zone_dtor;
1893 args.uminit = zero_init;
1896 args.align = UMA_BOOT_ALIGN - 1;
1897 args.flags = UMA_ZFLAG_INTERNAL;
1898 zone_ctor(zones, zsize, &args, M_WAITOK);
1900 /* Now make a zone for slab headers */
1901 slabzone = uma_zcreate("UMA Slabs",
1902 sizeof(struct uma_slab),
1903 NULL, NULL, NULL, NULL,
1904 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1906 hashzone = uma_zcreate("UMA Hash",
1907 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1908 NULL, NULL, NULL, NULL,
1909 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1913 booted = BOOT_STRAPPED;
1921 printf("Entering %s with %d boot pages left\n", __func__, boot_pages);
1923 booted = BOOT_PAGEALLOC;
1931 printf("Entering %s with %d boot pages left\n", __func__, boot_pages);
1933 booted = BOOT_BUCKETS;
1934 sx_init(&uma_drain_lock, "umadrain");
1939 * Initialize our callout handle
1946 booted = BOOT_RUNNING;
1947 callout_init(&uma_callout, 1);
1948 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1952 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1953 int align, uint32_t flags)
1955 struct uma_kctor_args args;
1958 args.uminit = uminit;
1960 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1963 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
1966 /* Public functions */
1969 uma_set_align(int align)
1972 if (align != UMA_ALIGN_CACHE)
1973 uma_align_cache = align;
1978 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1979 uma_init uminit, uma_fini fini, int align, uint32_t flags)
1982 struct uma_zctor_args args;
1986 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
1989 /* This stuff is essential for the zone ctor */
1990 memset(&args, 0, sizeof(args));
1995 args.uminit = uminit;
1999 * If a zone is being created with an empty constructor and
2000 * destructor, pass UMA constructor/destructor which checks for
2001 * memory use after free.
2003 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) &&
2004 ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) {
2005 args.ctor = trash_ctor;
2006 args.dtor = trash_dtor;
2007 args.uminit = trash_init;
2008 args.fini = trash_fini;
2015 if (booted < BOOT_BUCKETS) {
2018 sx_slock(&uma_drain_lock);
2021 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
2023 sx_sunlock(&uma_drain_lock);
2029 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
2030 uma_init zinit, uma_fini zfini, uma_zone_t master)
2032 struct uma_zctor_args args;
2037 keg = zone_first_keg(master);
2038 memset(&args, 0, sizeof(args));
2040 args.size = keg->uk_size;
2043 args.uminit = zinit;
2045 args.align = keg->uk_align;
2046 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
2049 if (booted < BOOT_BUCKETS) {
2052 sx_slock(&uma_drain_lock);
2055 /* XXX Attaches only one keg of potentially many. */
2056 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
2058 sx_sunlock(&uma_drain_lock);
2064 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
2065 uma_init zinit, uma_fini zfini, uma_import zimport,
2066 uma_release zrelease, void *arg, int flags)
2068 struct uma_zctor_args args;
2070 memset(&args, 0, sizeof(args));
2075 args.uminit = zinit;
2077 args.import = zimport;
2078 args.release = zrelease;
2083 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
2087 zone_lock_pair(uma_zone_t a, uma_zone_t b)
2091 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
2094 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
2099 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2107 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2114 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2116 zone_lock_pair(zone, master);
2118 * zone must use vtoslab() to resolve objects and must already be
2121 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2122 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2127 * The new master must also use vtoslab().
2129 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2135 * The underlying object must be the same size. rsize
2138 if (master->uz_size != zone->uz_size) {
2143 * Put it at the end of the list.
2145 klink->kl_keg = zone_first_keg(master);
2146 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2147 if (LIST_NEXT(kl, kl_link) == NULL) {
2148 LIST_INSERT_AFTER(kl, klink, kl_link);
2153 zone->uz_flags |= UMA_ZFLAG_MULTI;
2154 zone->uz_slab = zone_fetch_slab_multi;
2157 zone_unlock_pair(zone, master);
2159 free(klink, M_TEMP);
2167 uma_zdestroy(uma_zone_t zone)
2170 sx_slock(&uma_drain_lock);
2171 zone_free_item(zones, zone, NULL, SKIP_NONE);
2172 sx_sunlock(&uma_drain_lock);
2176 uma_zwait(uma_zone_t zone)
2180 item = uma_zalloc_arg(zone, NULL, M_WAITOK);
2181 uma_zfree(zone, item);
2186 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2188 uma_zone_domain_t zdom;
2189 uma_bucket_t bucket;
2192 int cpu, domain, lockfail;
2194 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2195 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2197 /* This is the fast path allocation */
2198 CTR4(KTR_UMA, "uma_zalloc_arg thread %x zone %s(%p) flags %d",
2199 curthread, zone->uz_name, zone, flags);
2201 if (flags & M_WAITOK) {
2202 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2203 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2205 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2206 ("uma_zalloc_arg: called with spinlock or critical section held"));
2208 #ifdef DEBUG_MEMGUARD
2209 if (memguard_cmp_zone(zone)) {
2210 item = memguard_alloc(zone->uz_size, flags);
2212 if (zone->uz_init != NULL &&
2213 zone->uz_init(item, zone->uz_size, flags) != 0)
2215 if (zone->uz_ctor != NULL &&
2216 zone->uz_ctor(item, zone->uz_size, udata,
2218 zone->uz_fini(item, zone->uz_size);
2223 /* This is unfortunate but should not be fatal. */
2227 * If possible, allocate from the per-CPU cache. There are two
2228 * requirements for safe access to the per-CPU cache: (1) the thread
2229 * accessing the cache must not be preempted or yield during access,
2230 * and (2) the thread must not migrate CPUs without switching which
2231 * cache it accesses. We rely on a critical section to prevent
2232 * preemption and migration. We release the critical section in
2233 * order to acquire the zone mutex if we are unable to allocate from
2234 * the current cache; when we re-acquire the critical section, we
2235 * must detect and handle migration if it has occurred.
2239 cache = &zone->uz_cpu[cpu];
2242 bucket = cache->uc_allocbucket;
2243 if (bucket != NULL && bucket->ub_cnt > 0) {
2245 item = bucket->ub_bucket[bucket->ub_cnt];
2247 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2249 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2252 if (zone->uz_ctor != NULL &&
2253 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2254 atomic_add_long(&zone->uz_fails, 1);
2255 zone_free_item(zone, item, udata, SKIP_DTOR);
2259 uma_dbg_alloc(zone, NULL, item);
2262 uma_zero_item(item, zone);
2267 * We have run out of items in our alloc bucket.
2268 * See if we can switch with our free bucket.
2270 bucket = cache->uc_freebucket;
2271 if (bucket != NULL && bucket->ub_cnt > 0) {
2273 "uma_zalloc: zone %s(%p) swapping empty with alloc",
2274 zone->uz_name, zone);
2275 cache->uc_freebucket = cache->uc_allocbucket;
2276 cache->uc_allocbucket = bucket;
2281 * Discard any empty allocation bucket while we hold no locks.
2283 bucket = cache->uc_allocbucket;
2284 cache->uc_allocbucket = NULL;
2287 bucket_free(zone, bucket, udata);
2289 if (zone->uz_flags & UMA_ZONE_NUMA)
2290 domain = PCPU_GET(domain);
2292 domain = UMA_ANYDOMAIN;
2294 /* Short-circuit for zones without buckets and low memory. */
2295 if (zone->uz_count == 0 || bucketdisable)
2299 * Attempt to retrieve the item from the per-CPU cache has failed, so
2300 * we must go back to the zone. This requires the zone lock, so we
2301 * must drop the critical section, then re-acquire it when we go back
2302 * to the cache. Since the critical section is released, we may be
2303 * preempted or migrate. As such, make sure not to maintain any
2304 * thread-local state specific to the cache from prior to releasing
2305 * the critical section.
2308 if (ZONE_TRYLOCK(zone) == 0) {
2309 /* Record contention to size the buckets. */
2315 cache = &zone->uz_cpu[cpu];
2318 * Since we have locked the zone we may as well send back our stats.
2320 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2321 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2322 cache->uc_allocs = 0;
2323 cache->uc_frees = 0;
2325 /* See if we lost the race to fill the cache. */
2326 if (cache->uc_allocbucket != NULL) {
2332 * Check the zone's cache of buckets.
2334 if (domain == UMA_ANYDOMAIN)
2335 zdom = &zone->uz_domain[0];
2337 zdom = &zone->uz_domain[domain];
2338 if ((bucket = LIST_FIRST(&zdom->uzd_buckets)) != NULL) {
2339 KASSERT(bucket->ub_cnt != 0,
2340 ("uma_zalloc_arg: Returning an empty bucket."));
2342 LIST_REMOVE(bucket, ub_link);
2343 cache->uc_allocbucket = bucket;
2347 /* We are no longer associated with this CPU. */
2351 * We bump the uz count when the cache size is insufficient to
2352 * handle the working set.
2354 if (lockfail && zone->uz_count < BUCKET_MAX)
2359 * Now lets just fill a bucket and put it on the free list. If that
2360 * works we'll restart the allocation from the beginning and it
2361 * will use the just filled bucket.
2363 bucket = zone_alloc_bucket(zone, udata, domain, flags);
2364 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
2365 zone->uz_name, zone, bucket);
2366 if (bucket != NULL) {
2370 cache = &zone->uz_cpu[cpu];
2372 * See if we lost the race or were migrated. Cache the
2373 * initialized bucket to make this less likely or claim
2374 * the memory directly.
2376 if (cache->uc_allocbucket != NULL ||
2377 (zone->uz_flags & UMA_ZONE_NUMA &&
2378 domain != PCPU_GET(domain)))
2379 LIST_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
2381 cache->uc_allocbucket = bucket;
2387 * We may not be able to get a bucket so return an actual item.
2390 item = zone_alloc_item(zone, udata, domain, flags);
2396 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
2399 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2400 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2402 /* This is the fast path allocation */
2404 "uma_zalloc_domain thread %x zone %s(%p) domain %d flags %d",
2405 curthread, zone->uz_name, zone, domain, flags);
2407 if (flags & M_WAITOK) {
2408 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2409 "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
2411 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2412 ("uma_zalloc_domain: called with spinlock or critical section held"));
2414 return (zone_alloc_item(zone, udata, domain, flags));
2418 * Find a slab with some space. Prefer slabs that are partially used over those
2419 * that are totally full. This helps to reduce fragmentation.
2421 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
2425 keg_first_slab(uma_keg_t keg, int domain, int rr)
2431 KASSERT(domain >= 0 && domain < vm_ndomains,
2432 ("keg_first_slab: domain %d out of range", domain));
2437 dom = &keg->uk_domain[domain];
2438 if (!LIST_EMPTY(&dom->ud_part_slab))
2439 return (LIST_FIRST(&dom->ud_part_slab));
2440 if (!LIST_EMPTY(&dom->ud_free_slab)) {
2441 slab = LIST_FIRST(&dom->ud_free_slab);
2442 LIST_REMOVE(slab, us_link);
2443 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
2447 domain = (domain + 1) % vm_ndomains;
2448 } while (domain != start);
2454 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, int flags)
2458 int allocflags, domain, reserve, rr, start;
2460 mtx_assert(&keg->uk_lock, MA_OWNED);
2464 if ((flags & M_USE_RESERVE) == 0)
2465 reserve = keg->uk_reserve;
2468 * Round-robin for non first-touch zones when there is more than one
2471 if (vm_ndomains == 1)
2473 rr = rdomain == UMA_ANYDOMAIN;
2475 keg->uk_cursor = (keg->uk_cursor + 1) % vm_ndomains;
2476 domain = start = keg->uk_cursor;
2477 /* Only block on the second pass. */
2478 if ((flags & (M_WAITOK | M_NOVM)) == M_WAITOK)
2479 allocflags = (allocflags & ~M_WAITOK) | M_NOWAIT;
2481 domain = start = rdomain;
2485 if (keg->uk_free > reserve &&
2486 (slab = keg_first_slab(keg, domain, rr)) != NULL) {
2487 MPASS(slab->us_keg == keg);
2492 * M_NOVM means don't ask at all!
2497 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2498 keg->uk_flags |= UMA_ZFLAG_FULL;
2500 * If this is not a multi-zone, set the FULL bit.
2501 * Otherwise slab_multi() takes care of it.
2503 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2504 zone->uz_flags |= UMA_ZFLAG_FULL;
2505 zone_log_warning(zone);
2506 zone_maxaction(zone);
2508 if (flags & M_NOWAIT)
2511 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2514 slab = keg_alloc_slab(keg, zone, domain, allocflags);
2516 * If we got a slab here it's safe to mark it partially used
2517 * and return. We assume that the caller is going to remove
2518 * at least one item.
2521 MPASS(slab->us_keg == keg);
2522 dom = &keg->uk_domain[slab->us_domain];
2523 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
2527 keg->uk_cursor = (keg->uk_cursor + 1) % vm_ndomains;
2528 domain = keg->uk_cursor;
2530 } while (domain != start);
2532 /* Retry domain scan with blocking. */
2533 if (allocflags != flags) {
2539 * We might not have been able to get a slab but another cpu
2540 * could have while we were unlocked. Check again before we
2543 if (keg->uk_free > reserve &&
2544 (slab = keg_first_slab(keg, domain, rr)) != NULL) {
2545 MPASS(slab->us_keg == keg);
2552 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int domain, int flags)
2557 keg = zone_first_keg(zone);
2562 slab = keg_fetch_slab(keg, zone, domain, flags);
2565 if (flags & (M_NOWAIT | M_NOVM))
2573 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2574 * with the keg locked. On NULL no lock is held.
2576 * The last pointer is used to seed the search. It is not required.
2579 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int domain, int rflags)
2589 * Don't wait on the first pass. This will skip limit tests
2590 * as well. We don't want to block if we can find a provider
2593 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2595 * Use the last slab allocated as a hint for where to start
2599 slab = keg_fetch_slab(last, zone, domain, flags);
2605 * Loop until we have a slab incase of transient failures
2606 * while M_WAITOK is specified. I'm not sure this is 100%
2607 * required but we've done it for so long now.
2613 * Search the available kegs for slabs. Be careful to hold the
2614 * correct lock while calling into the keg layer.
2616 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2617 keg = klink->kl_keg;
2619 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2620 slab = keg_fetch_slab(keg, zone, domain, flags);
2624 if (keg->uk_flags & UMA_ZFLAG_FULL)
2630 if (rflags & (M_NOWAIT | M_NOVM))
2634 * All kegs are full. XXX We can't atomically check all kegs
2635 * and sleep so just sleep for a short period and retry.
2637 if (full && !empty) {
2639 zone->uz_flags |= UMA_ZFLAG_FULL;
2641 zone_log_warning(zone);
2642 zone_maxaction(zone);
2643 msleep(zone, zone->uz_lockptr, PVM,
2644 "zonelimit", hz/100);
2645 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2654 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2660 MPASS(keg == slab->us_keg);
2661 mtx_assert(&keg->uk_lock, MA_OWNED);
2663 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2664 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2665 item = slab->us_data + (keg->uk_rsize * freei);
2666 slab->us_freecount--;
2669 /* Move this slab to the full list */
2670 if (slab->us_freecount == 0) {
2671 LIST_REMOVE(slab, us_link);
2672 dom = &keg->uk_domain[slab->us_domain];
2673 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
2680 zone_import(uma_zone_t zone, void **bucket, int max, int domain, int flags)
2689 /* Try to keep the buckets totally full */
2690 for (i = 0; i < max; ) {
2691 if ((slab = zone->uz_slab(zone, keg, domain, flags)) == NULL)
2694 stripe = howmany(max, vm_ndomains);
2695 while (slab->us_freecount && i < max) {
2696 bucket[i++] = slab_alloc_item(keg, slab);
2697 if (keg->uk_free <= keg->uk_reserve)
2701 * If the zone is striped we pick a new slab for every
2702 * N allocations. Eliminating this conditional will
2703 * instead pick a new domain for each bucket rather
2704 * than stripe within each bucket. The current option
2705 * produces more fragmentation and requires more cpu
2706 * time but yields better distribution.
2708 if ((zone->uz_flags & UMA_ZONE_NUMA) == 0 &&
2709 vm_ndomains > 1 && --stripe == 0)
2713 /* Don't block if we allocated any successfully. */
2724 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
2726 uma_bucket_t bucket;
2729 /* Don't wait for buckets, preserve caller's NOVM setting. */
2730 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2734 max = MIN(bucket->ub_entries, zone->uz_count);
2735 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2736 max, domain, flags);
2739 * Initialize the memory if necessary.
2741 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2744 for (i = 0; i < bucket->ub_cnt; i++)
2745 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2749 * If we couldn't initialize the whole bucket, put the
2750 * rest back onto the freelist.
2752 if (i != bucket->ub_cnt) {
2753 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2754 bucket->ub_cnt - i);
2756 bzero(&bucket->ub_bucket[i],
2757 sizeof(void *) * (bucket->ub_cnt - i));
2763 if (bucket->ub_cnt == 0) {
2764 bucket_free(zone, bucket, udata);
2765 atomic_add_long(&zone->uz_fails, 1);
2773 * Allocates a single item from a zone.
2776 * zone The zone to alloc for.
2777 * udata The data to be passed to the constructor.
2778 * domain The domain to allocate from or UMA_ANYDOMAIN.
2779 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2782 * NULL if there is no memory and M_NOWAIT is set
2783 * An item if successful
2787 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
2793 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
2795 atomic_add_long(&zone->uz_allocs, 1);
2798 * We have to call both the zone's init (not the keg's init)
2799 * and the zone's ctor. This is because the item is going from
2800 * a keg slab directly to the user, and the user is expecting it
2801 * to be both zone-init'd as well as zone-ctor'd.
2803 if (zone->uz_init != NULL) {
2804 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2805 zone_free_item(zone, item, udata, SKIP_FINI);
2809 if (zone->uz_ctor != NULL) {
2810 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2811 zone_free_item(zone, item, udata, SKIP_DTOR);
2816 uma_dbg_alloc(zone, NULL, item);
2819 uma_zero_item(item, zone);
2821 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
2822 zone->uz_name, zone);
2827 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
2828 zone->uz_name, zone);
2829 atomic_add_long(&zone->uz_fails, 1);
2835 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2838 uma_bucket_t bucket;
2839 uma_zone_domain_t zdom;
2840 int cpu, domain, lockfail;
2842 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2843 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2845 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2848 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2849 ("uma_zfree_arg: called with spinlock or critical section held"));
2851 /* uma_zfree(..., NULL) does nothing, to match free(9). */
2854 #ifdef DEBUG_MEMGUARD
2855 if (is_memguard_addr(item)) {
2856 if (zone->uz_dtor != NULL)
2857 zone->uz_dtor(item, zone->uz_size, udata);
2858 if (zone->uz_fini != NULL)
2859 zone->uz_fini(item, zone->uz_size);
2860 memguard_free(item);
2865 if (zone->uz_flags & UMA_ZONE_MALLOC)
2866 uma_dbg_free(zone, udata, item);
2868 uma_dbg_free(zone, NULL, item);
2870 if (zone->uz_dtor != NULL)
2871 zone->uz_dtor(item, zone->uz_size, udata);
2874 * The race here is acceptable. If we miss it we'll just have to wait
2875 * a little longer for the limits to be reset.
2877 if (zone->uz_flags & UMA_ZFLAG_FULL)
2881 * If possible, free to the per-CPU cache. There are two
2882 * requirements for safe access to the per-CPU cache: (1) the thread
2883 * accessing the cache must not be preempted or yield during access,
2884 * and (2) the thread must not migrate CPUs without switching which
2885 * cache it accesses. We rely on a critical section to prevent
2886 * preemption and migration. We release the critical section in
2887 * order to acquire the zone mutex if we are unable to free to the
2888 * current cache; when we re-acquire the critical section, we must
2889 * detect and handle migration if it has occurred.
2894 cache = &zone->uz_cpu[cpu];
2898 * Try to free into the allocbucket first to give LIFO ordering
2899 * for cache-hot datastructures. Spill over into the freebucket
2900 * if necessary. Alloc will swap them if one runs dry.
2902 bucket = cache->uc_allocbucket;
2903 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2904 bucket = cache->uc_freebucket;
2905 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2906 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2907 ("uma_zfree: Freeing to non free bucket index."));
2908 bucket->ub_bucket[bucket->ub_cnt] = item;
2916 * We must go back the zone, which requires acquiring the zone lock,
2917 * which in turn means we must release and re-acquire the critical
2918 * section. Since the critical section is released, we may be
2919 * preempted or migrate. As such, make sure not to maintain any
2920 * thread-local state specific to the cache from prior to releasing
2921 * the critical section.
2924 if (zone->uz_count == 0 || bucketdisable)
2928 if (ZONE_TRYLOCK(zone) == 0) {
2929 /* Record contention to size the buckets. */
2935 cache = &zone->uz_cpu[cpu];
2938 * Since we have locked the zone we may as well send back our stats.
2940 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2941 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2942 cache->uc_allocs = 0;
2943 cache->uc_frees = 0;
2945 bucket = cache->uc_freebucket;
2946 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2950 cache->uc_freebucket = NULL;
2951 /* We are no longer associated with this CPU. */
2954 if ((zone->uz_flags & UMA_ZONE_NUMA) != 0)
2955 domain = PCPU_GET(domain);
2958 zdom = &zone->uz_domain[0];
2960 /* Can we throw this on the zone full list? */
2961 if (bucket != NULL) {
2963 "uma_zfree: zone %s(%p) putting bucket %p on free list",
2964 zone->uz_name, zone, bucket);
2965 /* ub_cnt is pointing to the last free item */
2966 KASSERT(bucket->ub_cnt != 0,
2967 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2968 LIST_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
2972 * We bump the uz count when the cache size is insufficient to
2973 * handle the working set.
2975 if (lockfail && zone->uz_count < BUCKET_MAX)
2979 bucket = bucket_alloc(zone, udata, M_NOWAIT);
2980 CTR3(KTR_UMA, "uma_zfree: zone %s(%p) allocated bucket %p",
2981 zone->uz_name, zone, bucket);
2985 cache = &zone->uz_cpu[cpu];
2986 if (cache->uc_freebucket == NULL &&
2987 ((zone->uz_flags & UMA_ZONE_NUMA) == 0 ||
2988 domain == PCPU_GET(domain))) {
2989 cache->uc_freebucket = bucket;
2993 * We lost the race, start over. We have to drop our
2994 * critical section to free the bucket.
2997 bucket_free(zone, bucket, udata);
3002 * If nothing else caught this, we'll just do an internal free.
3005 zone_free_item(zone, item, udata, SKIP_DTOR);
3011 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
3014 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3015 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
3017 CTR2(KTR_UMA, "uma_zfree_domain thread %x zone %s", curthread,
3020 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3021 ("uma_zfree_domain: called with spinlock or critical section held"));
3023 /* uma_zfree(..., NULL) does nothing, to match free(9). */
3026 zone_free_item(zone, item, udata, SKIP_NONE);
3030 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
3035 mtx_assert(&keg->uk_lock, MA_OWNED);
3036 MPASS(keg == slab->us_keg);
3038 dom = &keg->uk_domain[slab->us_domain];
3040 /* Do we need to remove from any lists? */
3041 if (slab->us_freecount+1 == keg->uk_ipers) {
3042 LIST_REMOVE(slab, us_link);
3043 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
3044 } else if (slab->us_freecount == 0) {
3045 LIST_REMOVE(slab, us_link);
3046 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3049 /* Slab management. */
3050 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3051 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
3052 slab->us_freecount++;
3054 /* Keg statistics. */
3059 zone_release(uma_zone_t zone, void **bucket, int cnt)
3069 keg = zone_first_keg(zone);
3071 for (i = 0; i < cnt; i++) {
3073 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
3074 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
3075 if (zone->uz_flags & UMA_ZONE_HASH) {
3076 slab = hash_sfind(&keg->uk_hash, mem);
3078 mem += keg->uk_pgoff;
3079 slab = (uma_slab_t)mem;
3082 slab = vtoslab((vm_offset_t)item);
3083 if (slab->us_keg != keg) {
3089 slab_free_item(keg, slab, item);
3090 if (keg->uk_flags & UMA_ZFLAG_FULL) {
3091 if (keg->uk_pages < keg->uk_maxpages) {
3092 keg->uk_flags &= ~UMA_ZFLAG_FULL;
3097 * We can handle one more allocation. Since we're
3098 * clearing ZFLAG_FULL, wake up all procs blocked
3099 * on pages. This should be uncommon, so keeping this
3100 * simple for now (rather than adding count of blocked
3109 zone->uz_flags &= ~UMA_ZFLAG_FULL;
3117 * Frees a single item to any zone.
3120 * zone The zone to free to
3121 * item The item we're freeing
3122 * udata User supplied data for the dtor
3123 * skip Skip dtors and finis
3126 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
3130 if (skip == SKIP_NONE) {
3131 if (zone->uz_flags & UMA_ZONE_MALLOC)
3132 uma_dbg_free(zone, udata, item);
3134 uma_dbg_free(zone, NULL, item);
3137 if (skip < SKIP_DTOR && zone->uz_dtor)
3138 zone->uz_dtor(item, zone->uz_size, udata);
3140 if (skip < SKIP_FINI && zone->uz_fini)
3141 zone->uz_fini(item, zone->uz_size);
3143 atomic_add_long(&zone->uz_frees, 1);
3144 zone->uz_release(zone->uz_arg, &item, 1);
3149 uma_zone_set_max(uma_zone_t zone, int nitems)
3153 keg = zone_first_keg(zone);
3157 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
3158 if (keg->uk_maxpages * keg->uk_ipers < nitems)
3159 keg->uk_maxpages += keg->uk_ppera;
3160 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers;
3168 uma_zone_get_max(uma_zone_t zone)
3173 keg = zone_first_keg(zone);
3177 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers;
3185 uma_zone_set_warning(uma_zone_t zone, const char *warning)
3189 zone->uz_warning = warning;
3195 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
3199 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
3205 uma_zone_get_cur(uma_zone_t zone)
3211 nitems = zone->uz_allocs - zone->uz_frees;
3214 * See the comment in sysctl_vm_zone_stats() regarding the
3215 * safety of accessing the per-cpu caches. With the zone lock
3216 * held, it is safe, but can potentially result in stale data.
3218 nitems += zone->uz_cpu[i].uc_allocs -
3219 zone->uz_cpu[i].uc_frees;
3223 return (nitems < 0 ? 0 : nitems);
3228 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
3232 keg = zone_first_keg(zone);
3233 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3235 KASSERT(keg->uk_pages == 0,
3236 ("uma_zone_set_init on non-empty keg"));
3237 keg->uk_init = uminit;
3243 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3247 keg = zone_first_keg(zone);
3248 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
3250 KASSERT(keg->uk_pages == 0,
3251 ("uma_zone_set_fini on non-empty keg"));
3252 keg->uk_fini = fini;
3258 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3262 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3263 ("uma_zone_set_zinit on non-empty keg"));
3264 zone->uz_init = zinit;
3270 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3274 KASSERT(zone_first_keg(zone)->uk_pages == 0,
3275 ("uma_zone_set_zfini on non-empty keg"));
3276 zone->uz_fini = zfini;
3281 /* XXX uk_freef is not actually used with the zone locked */
3283 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3287 keg = zone_first_keg(zone);
3288 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
3290 keg->uk_freef = freef;
3295 /* XXX uk_allocf is not actually used with the zone locked */
3297 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3301 keg = zone_first_keg(zone);
3303 keg->uk_allocf = allocf;
3309 uma_zone_reserve(uma_zone_t zone, int items)
3313 keg = zone_first_keg(zone);
3317 keg->uk_reserve = items;
3325 uma_zone_reserve_kva(uma_zone_t zone, int count)
3331 keg = zone_first_keg(zone);
3334 pages = count / keg->uk_ipers;
3336 if (pages * keg->uk_ipers < count)
3338 pages *= keg->uk_ppera;
3340 #ifdef UMA_MD_SMALL_ALLOC
3341 if (keg->uk_ppera > 1) {
3345 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
3353 keg->uk_maxpages = pages;
3354 #ifdef UMA_MD_SMALL_ALLOC
3355 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3357 keg->uk_allocf = noobj_alloc;
3359 keg->uk_flags |= UMA_ZONE_NOFREE;
3367 uma_prealloc(uma_zone_t zone, int items)
3374 keg = zone_first_keg(zone);
3378 slabs = items / keg->uk_ipers;
3380 if (slabs * keg->uk_ipers < items)
3383 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK);
3386 MPASS(slab->us_keg == keg);
3387 dom = &keg->uk_domain[slab->us_domain];
3388 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
3390 domain = (domain + 1) % vm_ndomains;
3397 uma_reclaim_locked(bool kmem_danger)
3400 CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
3401 sx_assert(&uma_drain_lock, SA_XLOCKED);
3403 zone_foreach(zone_drain);
3404 if (vm_page_count_min() || kmem_danger) {
3405 cache_drain_safe(NULL);
3406 zone_foreach(zone_drain);
3409 * Some slabs may have been freed but this zone will be visited early
3410 * we visit again so that we can free pages that are empty once other
3411 * zones are drained. We have to do the same for buckets.
3413 zone_drain(slabzone);
3414 bucket_zone_drain();
3421 sx_xlock(&uma_drain_lock);
3422 uma_reclaim_locked(false);
3423 sx_xunlock(&uma_drain_lock);
3426 static volatile int uma_reclaim_needed;
3429 uma_reclaim_wakeup(void)
3432 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
3433 wakeup(uma_reclaim);
3437 uma_reclaim_worker(void *arg __unused)
3441 sx_xlock(&uma_drain_lock);
3442 while (atomic_load_int(&uma_reclaim_needed) == 0)
3443 sx_sleep(uma_reclaim, &uma_drain_lock, PVM, "umarcl",
3445 sx_xunlock(&uma_drain_lock);
3446 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
3447 sx_xlock(&uma_drain_lock);
3448 uma_reclaim_locked(true);
3449 atomic_store_int(&uma_reclaim_needed, 0);
3450 sx_xunlock(&uma_drain_lock);
3451 /* Don't fire more than once per-second. */
3452 pause("umarclslp", hz);
3458 uma_zone_exhausted(uma_zone_t zone)
3463 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3469 uma_zone_exhausted_nolock(uma_zone_t zone)
3471 return (zone->uz_flags & UMA_ZFLAG_FULL);
3475 uma_large_malloc_domain(vm_size_t size, int domain, int wait)
3480 slab = zone_alloc_item(slabzone, NULL, domain, wait);
3483 if (domain == UMA_ANYDOMAIN)
3484 addr = kmem_malloc(kernel_arena, size, wait);
3486 addr = kmem_malloc_domain(domain, size, wait);
3488 vsetslab(addr, slab);
3489 slab->us_data = (void *)addr;
3490 slab->us_flags = UMA_SLAB_KERNEL | UMA_SLAB_MALLOC;
3491 slab->us_size = size;
3492 slab->us_domain = vm_phys_domain(PHYS_TO_VM_PAGE(
3493 pmap_kextract(addr)));
3494 uma_total_inc(size);
3496 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3499 return ((void *)addr);
3503 uma_large_malloc(vm_size_t size, int wait)
3506 return uma_large_malloc_domain(size, UMA_ANYDOMAIN, wait);
3510 uma_large_free(uma_slab_t slab)
3513 KASSERT((slab->us_flags & UMA_SLAB_KERNEL) != 0,
3514 ("uma_large_free: Memory not allocated with uma_large_malloc."));
3515 kmem_free(kernel_arena, (vm_offset_t)slab->us_data, slab->us_size);
3516 uma_total_dec(slab->us_size);
3517 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3521 uma_zero_item(void *item, uma_zone_t zone)
3525 if (zone->uz_flags & UMA_ZONE_PCPU) {
3527 bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3529 bzero(item, zone->uz_size);
3536 return (uma_kmem_limit);
3540 uma_set_limit(unsigned long limit)
3543 uma_kmem_limit = limit;
3550 return (uma_kmem_total);
3557 return (uma_kmem_limit - uma_kmem_total);
3561 uma_print_stats(void)
3563 zone_foreach(uma_print_zone);
3567 slab_print(uma_slab_t slab)
3569 printf("slab: keg %p, data %p, freecount %d\n",
3570 slab->us_keg, slab->us_data, slab->us_freecount);
3574 cache_print(uma_cache_t cache)
3576 printf("alloc: %p(%d), free: %p(%d)\n",
3577 cache->uc_allocbucket,
3578 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3579 cache->uc_freebucket,
3580 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3584 uma_print_keg(uma_keg_t keg)
3590 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3591 "out %d free %d limit %d\n",
3592 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3593 keg->uk_ipers, keg->uk_ppera,
3594 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free,
3595 keg->uk_free, (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3596 for (i = 0; i < vm_ndomains; i++) {
3597 dom = &keg->uk_domain[i];
3598 printf("Part slabs:\n");
3599 LIST_FOREACH(slab, &dom->ud_part_slab, us_link)
3601 printf("Free slabs:\n");
3602 LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
3604 printf("Full slabs:\n");
3605 LIST_FOREACH(slab, &dom->ud_full_slab, us_link)
3611 uma_print_zone(uma_zone_t zone)
3617 printf("zone: %s(%p) size %d flags %#x\n",
3618 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3619 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3620 uma_print_keg(kl->kl_keg);
3622 cache = &zone->uz_cpu[i];
3623 printf("CPU %d Cache:\n", i);
3630 * Generate statistics across both the zone and its per-cpu cache's. Return
3631 * desired statistics if the pointer is non-NULL for that statistic.
3633 * Note: does not update the zone statistics, as it can't safely clear the
3634 * per-CPU cache statistic.
3636 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3637 * safe from off-CPU; we should modify the caches to track this information
3638 * directly so that we don't have to.
3641 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3642 uint64_t *freesp, uint64_t *sleepsp)
3645 uint64_t allocs, frees, sleeps;
3648 allocs = frees = sleeps = 0;
3651 cache = &z->uz_cpu[cpu];
3652 if (cache->uc_allocbucket != NULL)
3653 cachefree += cache->uc_allocbucket->ub_cnt;
3654 if (cache->uc_freebucket != NULL)
3655 cachefree += cache->uc_freebucket->ub_cnt;
3656 allocs += cache->uc_allocs;
3657 frees += cache->uc_frees;
3659 allocs += z->uz_allocs;
3660 frees += z->uz_frees;
3661 sleeps += z->uz_sleeps;
3662 if (cachefreep != NULL)
3663 *cachefreep = cachefree;
3664 if (allocsp != NULL)
3668 if (sleepsp != NULL)
3674 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3681 rw_rlock(&uma_rwlock);
3682 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3683 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3686 rw_runlock(&uma_rwlock);
3687 return (sysctl_handle_int(oidp, &count, 0, req));
3691 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3693 struct uma_stream_header ush;
3694 struct uma_type_header uth;
3695 struct uma_percpu_stat ups;
3696 uma_bucket_t bucket;
3697 uma_zone_domain_t zdom;
3704 int count, error, i;
3706 error = sysctl_wire_old_buffer(req, 0);
3709 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3710 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3713 rw_rlock(&uma_rwlock);
3714 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3715 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3720 * Insert stream header.
3722 bzero(&ush, sizeof(ush));
3723 ush.ush_version = UMA_STREAM_VERSION;
3724 ush.ush_maxcpus = (mp_maxid + 1);
3725 ush.ush_count = count;
3726 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3728 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3729 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3730 bzero(&uth, sizeof(uth));
3732 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3733 uth.uth_align = kz->uk_align;
3734 uth.uth_size = kz->uk_size;
3735 uth.uth_rsize = kz->uk_rsize;
3736 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3738 uth.uth_maxpages += k->uk_maxpages;
3739 uth.uth_pages += k->uk_pages;
3740 uth.uth_keg_free += k->uk_free;
3741 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3746 * A zone is secondary is it is not the first entry
3747 * on the keg's zone list.
3749 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3750 (LIST_FIRST(&kz->uk_zones) != z))
3751 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3753 for (i = 0; i < vm_ndomains; i++) {
3754 zdom = &z->uz_domain[i];
3755 LIST_FOREACH(bucket, &zdom->uzd_buckets,
3757 uth.uth_zone_free += bucket->ub_cnt;
3759 uth.uth_allocs = z->uz_allocs;
3760 uth.uth_frees = z->uz_frees;
3761 uth.uth_fails = z->uz_fails;
3762 uth.uth_sleeps = z->uz_sleeps;
3763 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3765 * While it is not normally safe to access the cache
3766 * bucket pointers while not on the CPU that owns the
3767 * cache, we only allow the pointers to be exchanged
3768 * without the zone lock held, not invalidated, so
3769 * accept the possible race associated with bucket
3770 * exchange during monitoring.
3772 for (i = 0; i < (mp_maxid + 1); i++) {
3773 bzero(&ups, sizeof(ups));
3774 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3778 cache = &z->uz_cpu[i];
3779 if (cache->uc_allocbucket != NULL)
3780 ups.ups_cache_free +=
3781 cache->uc_allocbucket->ub_cnt;
3782 if (cache->uc_freebucket != NULL)
3783 ups.ups_cache_free +=
3784 cache->uc_freebucket->ub_cnt;
3785 ups.ups_allocs = cache->uc_allocs;
3786 ups.ups_frees = cache->uc_frees;
3788 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3793 rw_runlock(&uma_rwlock);
3794 error = sbuf_finish(&sbuf);
3800 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3802 uma_zone_t zone = *(uma_zone_t *)arg1;
3805 max = uma_zone_get_max(zone);
3806 error = sysctl_handle_int(oidp, &max, 0, req);
3807 if (error || !req->newptr)
3810 uma_zone_set_max(zone, max);
3816 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3818 uma_zone_t zone = *(uma_zone_t *)arg1;
3821 cur = uma_zone_get_cur(zone);
3822 return (sysctl_handle_int(oidp, &cur, 0, req));
3827 uma_dbg_getslab(uma_zone_t zone, void *item)
3833 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
3834 if (zone->uz_flags & UMA_ZONE_VTOSLAB) {
3835 slab = vtoslab((vm_offset_t)mem);
3838 * It is safe to return the slab here even though the
3839 * zone is unlocked because the item's allocation state
3840 * essentially holds a reference.
3843 keg = LIST_FIRST(&zone->uz_kegs)->kl_keg;
3844 if (keg->uk_flags & UMA_ZONE_HASH)
3845 slab = hash_sfind(&keg->uk_hash, mem);
3847 slab = (uma_slab_t)(mem + keg->uk_pgoff);
3855 * Set up the slab's freei data such that uma_dbg_free can function.
3859 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
3864 if (zone_first_keg(zone) == NULL)
3867 slab = uma_dbg_getslab(zone, item);
3869 panic("uma: item %p did not belong to zone %s\n",
3870 item, zone->uz_name);
3873 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3875 if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3876 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
3877 item, zone, zone->uz_name, slab, freei);
3878 BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3884 * Verifies freed addresses. Checks for alignment, valid slab membership
3885 * and duplicate frees.
3889 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
3894 if (zone_first_keg(zone) == NULL)
3897 slab = uma_dbg_getslab(zone, item);
3899 panic("uma: Freed item %p did not belong to zone %s\n",
3900 item, zone->uz_name);
3903 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3905 if (freei >= keg->uk_ipers)
3906 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
3907 item, zone, zone->uz_name, slab, freei);
3909 if (((freei * keg->uk_rsize) + slab->us_data) != item)
3910 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
3911 item, zone, zone->uz_name, slab, freei);
3913 if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3914 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
3915 item, zone, zone->uz_name, slab, freei);
3917 BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3919 #endif /* INVARIANTS */
3922 DB_SHOW_COMMAND(uma, db_show_uma)
3924 uma_bucket_t bucket;
3927 uma_zone_domain_t zdom;
3928 uint64_t allocs, frees, sleeps;
3931 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3932 "Free", "Requests", "Sleeps", "Bucket");
3933 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3934 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3935 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3936 allocs = z->uz_allocs;
3937 frees = z->uz_frees;
3938 sleeps = z->uz_sleeps;
3941 uma_zone_sumstat(z, &cachefree, &allocs,
3943 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3944 (LIST_FIRST(&kz->uk_zones) != z)))
3945 cachefree += kz->uk_free;
3946 for (i = 0; i < vm_ndomains; i++) {
3947 zdom = &z->uz_domain[i];
3948 LIST_FOREACH(bucket, &zdom->uzd_buckets,
3950 cachefree += bucket->ub_cnt;
3952 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3953 z->uz_name, (uintmax_t)kz->uk_size,
3954 (intmax_t)(allocs - frees), cachefree,
3955 (uintmax_t)allocs, sleeps, z->uz_count);
3962 DB_SHOW_COMMAND(umacache, db_show_umacache)
3964 uma_bucket_t bucket;
3966 uma_zone_domain_t zdom;
3967 uint64_t allocs, frees;
3970 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3971 "Requests", "Bucket");
3972 LIST_FOREACH(z, &uma_cachezones, uz_link) {
3973 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3974 for (i = 0; i < vm_ndomains; i++) {
3975 zdom = &z->uz_domain[i];
3976 LIST_FOREACH(bucket, &zdom->uzd_buckets, ub_link)
3977 cachefree += bucket->ub_cnt;
3979 db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3980 z->uz_name, (uintmax_t)z->uz_size,
3981 (intmax_t)(allocs - frees), cachefree,
3982 (uintmax_t)allocs, z->uz_count);