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/rwlock.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_map.h>
87 #include <vm/vm_kern.h>
88 #include <vm/vm_extern.h>
90 #include <vm/uma_int.h>
91 #include <vm/uma_dbg.h>
96 #include <vm/memguard.h>
100 * This is the zone and keg from which all zones are spawned. The idea is that
101 * even the zone & keg heads are allocated from the allocator, so we use the
102 * bss section to bootstrap us.
104 static struct uma_keg masterkeg;
105 static struct uma_zone masterzone_k;
106 static struct uma_zone masterzone_z;
107 static uma_zone_t kegs = &masterzone_k;
108 static uma_zone_t zones = &masterzone_z;
110 /* This is the zone from which all of uma_slab_t's are allocated. */
111 static uma_zone_t slabzone;
112 static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
115 * The initial hash tables come out of this zone so they can be allocated
116 * prior to malloc coming up.
118 static uma_zone_t hashzone;
120 /* The boot-time adjusted value for cache line alignment. */
121 int uma_align_cache = 64 - 1;
123 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
126 * Are we allowed to allocate buckets?
128 static int bucketdisable = 1;
130 /* Linked list of all kegs in the system */
131 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
133 /* This mutex protects the keg list */
134 static struct mtx_padalign uma_mtx;
136 /* Linked list of boot time pages */
137 static LIST_HEAD(,uma_slab) uma_boot_pages =
138 LIST_HEAD_INITIALIZER(uma_boot_pages);
140 /* This mutex protects the boot time pages list */
141 static struct mtx_padalign uma_boot_pages_mtx;
143 /* Is the VM done starting up? */
144 static int booted = 0;
145 #define UMA_STARTUP 1
146 #define UMA_STARTUP2 2
148 /* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
149 static const u_int uma_max_ipers = SLAB_SETSIZE;
152 * Only mbuf clusters use ref zones. Just provide enough references
153 * to support the one user. New code should not use the ref facility.
155 static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES;
158 * This is the handle used to schedule events that need to happen
159 * outside of the allocation fast path.
161 static struct callout uma_callout;
162 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
165 * This structure is passed as the zone ctor arg so that I don't have to create
166 * a special allocation function just for zones.
168 struct uma_zctor_args {
183 struct uma_kctor_args {
192 struct uma_bucket_zone {
195 int ubz_entries; /* Number of items it can hold. */
196 int ubz_maxsize; /* Maximum allocation size per-item. */
200 * Compute the actual number of bucket entries to pack them in power
201 * of two sizes for more efficient space utilization.
203 #define BUCKET_SIZE(n) \
204 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
206 #define BUCKET_MAX BUCKET_SIZE(128)
208 struct uma_bucket_zone bucket_zones[] = {
209 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
210 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
211 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
212 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
213 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
214 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
219 * Flags and enumerations to be passed to internal functions.
221 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
225 static void *noobj_alloc(uma_zone_t, int, uint8_t *, int);
226 static void *page_alloc(uma_zone_t, int, uint8_t *, int);
227 static void *startup_alloc(uma_zone_t, int, uint8_t *, int);
228 static void page_free(void *, int, uint8_t);
229 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
230 static void cache_drain(uma_zone_t);
231 static void bucket_drain(uma_zone_t, uma_bucket_t);
232 static void bucket_cache_drain(uma_zone_t zone);
233 static int keg_ctor(void *, int, void *, int);
234 static void keg_dtor(void *, int, void *);
235 static int zone_ctor(void *, int, void *, int);
236 static void zone_dtor(void *, int, void *);
237 static int zero_init(void *, int, int);
238 static void keg_small_init(uma_keg_t keg);
239 static void keg_large_init(uma_keg_t keg);
240 static void zone_foreach(void (*zfunc)(uma_zone_t));
241 static void zone_timeout(uma_zone_t zone);
242 static int hash_alloc(struct uma_hash *);
243 static int hash_expand(struct uma_hash *, struct uma_hash *);
244 static void hash_free(struct uma_hash *hash);
245 static void uma_timeout(void *);
246 static void uma_startup3(void);
247 static void *zone_alloc_item(uma_zone_t, void *, int);
248 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
249 static void bucket_enable(void);
250 static void bucket_init(void);
251 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
252 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
253 static void bucket_zone_drain(void);
254 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
255 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
256 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
257 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
258 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
259 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
260 uma_fini fini, int align, uint32_t flags);
261 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
262 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
264 void uma_print_zone(uma_zone_t);
265 void uma_print_stats(void);
266 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
267 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
269 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
271 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
272 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
274 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
275 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
277 static int zone_warnings = 1;
278 TUNABLE_INT("vm.zone_warnings", &zone_warnings);
279 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RW, &zone_warnings, 0,
280 "Warn when UMA zones becomes full");
283 * This routine checks to see whether or not it's safe to enable buckets.
288 bucketdisable = vm_page_count_min();
292 * Initialize bucket_zones, the array of zones of buckets of various sizes.
294 * For each zone, calculate the memory required for each bucket, consisting
295 * of the header and an array of pointers.
300 struct uma_bucket_zone *ubz;
304 for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
305 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
306 size += sizeof(void *) * ubz->ubz_entries;
307 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
308 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
309 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
314 * Given a desired number of entries for a bucket, return the zone from which
315 * to allocate the bucket.
317 static struct uma_bucket_zone *
318 bucket_zone_lookup(int entries)
320 struct uma_bucket_zone *ubz;
322 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
323 if (ubz->ubz_entries >= entries)
330 bucket_select(int size)
332 struct uma_bucket_zone *ubz;
334 ubz = &bucket_zones[0];
335 if (size > ubz->ubz_maxsize)
336 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
338 for (; ubz->ubz_entries != 0; ubz++)
339 if (ubz->ubz_maxsize < size)
342 return (ubz->ubz_entries);
346 bucket_alloc(uma_zone_t zone, void *udata, int flags)
348 struct uma_bucket_zone *ubz;
352 * This is to stop us from allocating per cpu buckets while we're
353 * running out of vm.boot_pages. Otherwise, we would exhaust the
354 * boot pages. This also prevents us from allocating buckets in
355 * low memory situations.
360 * To limit bucket recursion we store the original zone flags
361 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
362 * NOVM flag to persist even through deep recursions. We also
363 * store ZFLAG_BUCKET once we have recursed attempting to allocate
364 * a bucket for a bucket zone so we do not allow infinite bucket
365 * recursion. This cookie will even persist to frees of unused
366 * buckets via the allocation path or bucket allocations in the
369 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
371 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
372 udata = (void *)(uintptr_t)zone->uz_flags;
374 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
375 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
377 ubz = bucket_zone_lookup(zone->uz_count);
378 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
381 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
384 bucket->ub_entries = ubz->ubz_entries;
391 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
393 struct uma_bucket_zone *ubz;
395 KASSERT(bucket->ub_cnt == 0,
396 ("bucket_free: Freeing a non free bucket."));
397 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
398 udata = (void *)(uintptr_t)zone->uz_flags;
399 ubz = bucket_zone_lookup(bucket->ub_entries);
400 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
404 bucket_zone_drain(void)
406 struct uma_bucket_zone *ubz;
408 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
409 zone_drain(ubz->ubz_zone);
413 zone_log_warning(uma_zone_t zone)
415 static const struct timeval warninterval = { 300, 0 };
417 if (!zone_warnings || zone->uz_warning == NULL)
420 if (ratecheck(&zone->uz_ratecheck, &warninterval))
421 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
425 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
429 LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
430 kegfn(klink->kl_keg);
434 * Routine called by timeout which is used to fire off some time interval
435 * based calculations. (stats, hash size, etc.)
444 uma_timeout(void *unused)
447 zone_foreach(zone_timeout);
449 /* Reschedule this event */
450 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
454 * Routine to perform timeout driven calculations. This expands the
455 * hashes and does per cpu statistics aggregation.
460 keg_timeout(uma_keg_t keg)
465 * Expand the keg hash table.
467 * This is done if the number of slabs is larger than the hash size.
468 * What I'm trying to do here is completely reduce collisions. This
469 * may be a little aggressive. Should I allow for two collisions max?
471 if (keg->uk_flags & UMA_ZONE_HASH &&
472 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
473 struct uma_hash newhash;
474 struct uma_hash oldhash;
478 * This is so involved because allocating and freeing
479 * while the keg lock is held will lead to deadlock.
480 * I have to do everything in stages and check for
483 newhash = keg->uk_hash;
485 ret = hash_alloc(&newhash);
488 if (hash_expand(&keg->uk_hash, &newhash)) {
489 oldhash = keg->uk_hash;
490 keg->uk_hash = newhash;
503 zone_timeout(uma_zone_t zone)
506 zone_foreach_keg(zone, &keg_timeout);
510 * Allocate and zero fill the next sized hash table from the appropriate
514 * hash A new hash structure with the old hash size in uh_hashsize
517 * 1 on sucess and 0 on failure.
520 hash_alloc(struct uma_hash *hash)
525 oldsize = hash->uh_hashsize;
527 /* We're just going to go to a power of two greater */
529 hash->uh_hashsize = oldsize * 2;
530 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
531 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
532 M_UMAHASH, M_NOWAIT);
534 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
535 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
537 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
539 if (hash->uh_slab_hash) {
540 bzero(hash->uh_slab_hash, alloc);
541 hash->uh_hashmask = hash->uh_hashsize - 1;
549 * Expands the hash table for HASH zones. This is done from zone_timeout
550 * to reduce collisions. This must not be done in the regular allocation
551 * path, otherwise, we can recurse on the vm while allocating pages.
554 * oldhash The hash you want to expand
555 * newhash The hash structure for the new table
563 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
569 if (!newhash->uh_slab_hash)
572 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
576 * I need to investigate hash algorithms for resizing without a
580 for (i = 0; i < oldhash->uh_hashsize; i++)
581 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
582 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
583 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
584 hval = UMA_HASH(newhash, slab->us_data);
585 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
593 * Free the hash bucket to the appropriate backing store.
596 * slab_hash The hash bucket we're freeing
597 * hashsize The number of entries in that hash bucket
603 hash_free(struct uma_hash *hash)
605 if (hash->uh_slab_hash == NULL)
607 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
608 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
610 free(hash->uh_slab_hash, M_UMAHASH);
614 * Frees all outstanding items in a bucket
617 * zone The zone to free to, must be unlocked.
618 * bucket The free/alloc bucket with items, cpu queue must be locked.
625 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
633 for (i = 0; i < bucket->ub_cnt; i++)
634 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
635 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
640 * Drains the per cpu caches for a zone.
642 * NOTE: This may only be called while the zone is being turn down, and not
643 * during normal operation. This is necessary in order that we do not have
644 * to migrate CPUs to drain the per-CPU caches.
647 * zone The zone to drain, must be unlocked.
653 cache_drain(uma_zone_t zone)
659 * XXX: It is safe to not lock the per-CPU caches, because we're
660 * tearing down the zone anyway. I.e., there will be no further use
661 * of the caches at this point.
663 * XXX: It would good to be able to assert that the zone is being
664 * torn down to prevent improper use of cache_drain().
666 * XXX: We lock the zone before passing into bucket_cache_drain() as
667 * it is used elsewhere. Should the tear-down path be made special
668 * there in some form?
671 cache = &zone->uz_cpu[cpu];
672 bucket_drain(zone, cache->uc_allocbucket);
673 bucket_drain(zone, cache->uc_freebucket);
674 if (cache->uc_allocbucket != NULL)
675 bucket_free(zone, cache->uc_allocbucket, NULL);
676 if (cache->uc_freebucket != NULL)
677 bucket_free(zone, cache->uc_freebucket, NULL);
678 cache->uc_allocbucket = cache->uc_freebucket = NULL;
681 bucket_cache_drain(zone);
686 * Drain the cached buckets from a zone. Expects a locked zone on entry.
689 bucket_cache_drain(uma_zone_t zone)
694 * Drain the bucket queues and free the buckets, we just keep two per
697 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
698 LIST_REMOVE(bucket, ub_link);
700 bucket_drain(zone, bucket);
701 bucket_free(zone, bucket, NULL);
707 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
714 flags = slab->us_flags;
716 if (keg->uk_fini != NULL) {
717 for (i--; i > -1; i--)
718 keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
721 if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
724 if (flags & UMA_SLAB_KMEM)
726 else if (flags & UMA_SLAB_KERNEL)
730 for (i = 0; i < keg->uk_ppera; i++)
731 vsetobj((vm_offset_t)mem + (i * PAGE_SIZE), obj);
733 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
734 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
736 printf("%s: Returning %d bytes.\n", keg->uk_name,
737 PAGE_SIZE * keg->uk_ppera);
739 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
743 * Frees pages from a keg back to the system. This is done on demand from
744 * the pageout daemon.
749 keg_drain(uma_keg_t keg)
751 struct slabhead freeslabs = { 0 };
756 * We don't want to take pages from statically allocated kegs at this
759 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
763 printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
766 if (keg->uk_free == 0)
769 slab = LIST_FIRST(&keg->uk_free_slab);
771 n = LIST_NEXT(slab, us_link);
773 /* We have no where to free these to */
774 if (slab->us_flags & UMA_SLAB_BOOT) {
779 LIST_REMOVE(slab, us_link);
780 keg->uk_pages -= keg->uk_ppera;
781 keg->uk_free -= keg->uk_ipers;
783 if (keg->uk_flags & UMA_ZONE_HASH)
784 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
786 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
793 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
794 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
795 keg_free_slab(keg, slab, 0);
800 zone_drain_wait(uma_zone_t zone, int waitok)
804 * Set draining to interlock with zone_dtor() so we can release our
805 * locks as we go. Only dtor() should do a WAITOK call since it
806 * is the only call that knows the structure will still be available
810 while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
811 if (waitok == M_NOWAIT)
813 mtx_unlock(&uma_mtx);
814 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
817 zone->uz_flags |= UMA_ZFLAG_DRAINING;
818 bucket_cache_drain(zone);
821 * The DRAINING flag protects us from being freed while
822 * we're running. Normally the uma_mtx would protect us but we
823 * must be able to release and acquire the right lock for each keg.
825 zone_foreach_keg(zone, &keg_drain);
827 zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
834 zone_drain(uma_zone_t zone)
837 zone_drain_wait(zone, M_NOWAIT);
841 * Allocate a new slab for a keg. This does not insert the slab onto a list.
844 * wait Shall we wait?
847 * The slab that was allocated or NULL if there is no memory and the
848 * caller specified M_NOWAIT.
851 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
853 uma_slabrefcnt_t slabref;
860 mtx_assert(&keg->uk_lock, MA_OWNED);
865 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name);
867 allocf = keg->uk_allocf;
870 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
871 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
877 * This reproduces the old vm_zone behavior of zero filling pages the
878 * first time they are added to a zone.
880 * Malloced items are zeroed in uma_zalloc.
883 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
888 if (keg->uk_flags & UMA_ZONE_NODUMP)
891 /* zone is passed for legacy reasons. */
892 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
894 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
895 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
900 /* Point the slab into the allocated memory */
901 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
902 slab = (uma_slab_t )(mem + keg->uk_pgoff);
904 if (keg->uk_flags & UMA_ZONE_VTOSLAB)
905 for (i = 0; i < keg->uk_ppera; i++)
906 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
910 slab->us_freecount = keg->uk_ipers;
911 slab->us_flags = flags;
912 BIT_FILL(SLAB_SETSIZE, &slab->us_free);
914 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
916 if (keg->uk_flags & UMA_ZONE_REFCNT) {
917 slabref = (uma_slabrefcnt_t)slab;
918 for (i = 0; i < keg->uk_ipers; i++)
919 slabref->us_refcnt[i] = 0;
922 if (keg->uk_init != NULL) {
923 for (i = 0; i < keg->uk_ipers; i++)
924 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
925 keg->uk_size, wait) != 0)
927 if (i != keg->uk_ipers) {
928 keg_free_slab(keg, slab, i);
937 if (keg->uk_flags & UMA_ZONE_HASH)
938 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
940 keg->uk_pages += keg->uk_ppera;
941 keg->uk_free += keg->uk_ipers;
948 * This function is intended to be used early on in place of page_alloc() so
949 * that we may use the boot time page cache to satisfy allocations before
953 startup_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
957 int pages, check_pages;
959 keg = zone_first_keg(zone);
960 pages = howmany(bytes, PAGE_SIZE);
961 check_pages = pages - 1;
962 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
965 * Check our small startup cache to see if it has pages remaining.
967 mtx_lock(&uma_boot_pages_mtx);
969 /* First check if we have enough room. */
970 tmps = LIST_FIRST(&uma_boot_pages);
971 while (tmps != NULL && check_pages-- > 0)
972 tmps = LIST_NEXT(tmps, us_link);
975 * It's ok to lose tmps references. The last one will
976 * have tmps->us_data pointing to the start address of
977 * "pages" contiguous pages of memory.
979 while (pages-- > 0) {
980 tmps = LIST_FIRST(&uma_boot_pages);
981 LIST_REMOVE(tmps, us_link);
983 mtx_unlock(&uma_boot_pages_mtx);
984 *pflag = tmps->us_flags;
985 return (tmps->us_data);
987 mtx_unlock(&uma_boot_pages_mtx);
988 if (booted < UMA_STARTUP2)
989 panic("UMA: Increase vm.boot_pages");
991 * Now that we've booted reset these users to their real allocator.
993 #ifdef UMA_MD_SMALL_ALLOC
994 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
996 keg->uk_allocf = page_alloc;
998 return keg->uk_allocf(zone, bytes, pflag, wait);
1002 * Allocates a number of pages from the system
1005 * bytes The number of bytes requested
1006 * wait Shall we wait?
1009 * A pointer to the alloced memory or possibly
1010 * NULL if M_NOWAIT is set.
1013 page_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
1015 void *p; /* Returned page */
1017 *pflag = UMA_SLAB_KMEM;
1018 p = (void *) kmem_malloc(kmem_map, bytes, wait);
1024 * Allocates a number of pages from within an object
1027 * bytes The number of bytes requested
1028 * wait Shall we wait?
1031 * A pointer to the alloced memory or possibly
1032 * NULL if M_NOWAIT is set.
1035 noobj_alloc(uma_zone_t zone, int bytes, uint8_t *flags, int wait)
1037 TAILQ_HEAD(, vm_page) alloctail;
1039 vm_offset_t retkva, zkva;
1040 vm_page_t p, p_next;
1043 TAILQ_INIT(&alloctail);
1044 keg = zone_first_keg(zone);
1046 npages = howmany(bytes, PAGE_SIZE);
1047 while (npages > 0) {
1048 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1049 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1052 * Since the page does not belong to an object, its
1055 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1059 if (wait & M_WAITOK) {
1065 * Page allocation failed, free intermediate pages and
1068 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1069 vm_page_unwire(p, 0);
1074 *flags = UMA_SLAB_PRIV;
1075 zkva = keg->uk_kva +
1076 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1078 TAILQ_FOREACH(p, &alloctail, listq) {
1079 pmap_qenter(zkva, &p, 1);
1083 return ((void *)retkva);
1087 * Frees a number of pages to the system
1090 * mem A pointer to the memory to be freed
1091 * size The size of the memory being freed
1092 * flags The original p->us_flags field
1098 page_free(void *mem, int size, uint8_t flags)
1102 if (flags & UMA_SLAB_KMEM)
1104 else if (flags & UMA_SLAB_KERNEL)
1107 panic("UMA: page_free used with invalid flags %d", flags);
1109 kmem_free(map, (vm_offset_t)mem, size);
1113 * Zero fill initializer
1115 * Arguments/Returns follow uma_init specifications
1118 zero_init(void *mem, int size, int flags)
1125 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1128 * keg The zone we should initialize
1134 keg_small_init(uma_keg_t keg)
1141 if (keg->uk_flags & UMA_ZONE_PCPU) {
1142 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1144 keg->uk_slabsize = sizeof(struct pcpu);
1145 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1148 keg->uk_slabsize = UMA_SLAB_SIZE;
1153 * Calculate the size of each allocation (rsize) according to
1154 * alignment. If the requested size is smaller than we have
1155 * allocation bits for we round it up.
1157 rsize = keg->uk_size;
1158 if (rsize < keg->uk_slabsize / SLAB_SETSIZE)
1159 rsize = keg->uk_slabsize / SLAB_SETSIZE;
1160 if (rsize & keg->uk_align)
1161 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1162 keg->uk_rsize = rsize;
1164 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1165 keg->uk_rsize < sizeof(struct pcpu),
1166 ("%s: size %u too large", __func__, keg->uk_rsize));
1168 if (keg->uk_flags & UMA_ZONE_REFCNT)
1169 rsize += sizeof(uint32_t);
1171 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1174 shsize = sizeof(struct uma_slab);
1176 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1177 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1178 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1180 memused = keg->uk_ipers * rsize + shsize;
1181 wastedspace = keg->uk_slabsize - memused;
1184 * We can't do OFFPAGE if we're internal or if we've been
1185 * asked to not go to the VM for buckets. If we do this we
1186 * may end up going to the VM for slabs which we do not
1187 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1188 * of UMA_ZONE_VM, which clearly forbids it.
1190 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1191 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1195 * See if using an OFFPAGE slab will limit our waste. Only do
1196 * this if it permits more items per-slab.
1198 * XXX We could try growing slabsize to limit max waste as well.
1199 * Historically this was not done because the VM could not
1200 * efficiently handle contiguous allocations.
1202 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1203 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1204 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1205 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1206 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1208 printf("UMA decided we need offpage slab headers for "
1209 "keg: %s, calculated wastedspace = %d, "
1210 "maximum wasted space allowed = %d, "
1211 "calculated ipers = %d, "
1212 "new wasted space = %d\n", keg->uk_name, wastedspace,
1213 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1214 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1216 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1219 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1220 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1221 keg->uk_flags |= UMA_ZONE_HASH;
1225 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1226 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1230 * keg The keg we should initialize
1236 keg_large_init(uma_keg_t keg)
1239 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1240 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1241 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1242 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1243 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1245 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1246 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1248 keg->uk_rsize = keg->uk_size;
1250 /* We can't do OFFPAGE if we're internal, bail out here. */
1251 if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1254 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1255 if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1256 keg->uk_flags |= UMA_ZONE_HASH;
1260 keg_cachespread_init(uma_keg_t keg)
1267 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1268 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1270 alignsize = keg->uk_align + 1;
1271 rsize = keg->uk_size;
1273 * We want one item to start on every align boundary in a page. To
1274 * do this we will span pages. We will also extend the item by the
1275 * size of align if it is an even multiple of align. Otherwise, it
1276 * would fall on the same boundary every time.
1278 if (rsize & keg->uk_align)
1279 rsize = (rsize & ~keg->uk_align) + alignsize;
1280 if ((rsize & alignsize) == 0)
1282 trailer = rsize - keg->uk_size;
1283 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1284 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1285 keg->uk_rsize = rsize;
1286 keg->uk_ppera = pages;
1287 keg->uk_slabsize = UMA_SLAB_SIZE;
1288 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1289 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1290 KASSERT(keg->uk_ipers <= uma_max_ipers,
1291 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1296 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1297 * the keg onto the global keg list.
1299 * Arguments/Returns follow uma_ctor specifications
1300 * udata Actually uma_kctor_args
1303 keg_ctor(void *mem, int size, void *udata, int flags)
1305 struct uma_kctor_args *arg = udata;
1306 uma_keg_t keg = mem;
1310 keg->uk_size = arg->size;
1311 keg->uk_init = arg->uminit;
1312 keg->uk_fini = arg->fini;
1313 keg->uk_align = arg->align;
1315 keg->uk_reserve = 0;
1317 keg->uk_flags = arg->flags;
1318 keg->uk_allocf = page_alloc;
1319 keg->uk_freef = page_free;
1320 keg->uk_slabzone = NULL;
1323 * The master zone is passed to us at keg-creation time.
1326 keg->uk_name = zone->uz_name;
1328 if (arg->flags & UMA_ZONE_VM)
1329 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1331 if (arg->flags & UMA_ZONE_ZINIT)
1332 keg->uk_init = zero_init;
1334 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1335 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1337 if (arg->flags & UMA_ZONE_PCPU)
1339 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1341 keg->uk_flags &= ~UMA_ZONE_PCPU;
1344 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1345 keg_cachespread_init(keg);
1346 } else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1348 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) -
1350 keg_large_init(keg);
1352 keg_small_init(keg);
1354 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1355 keg_large_init(keg);
1357 keg_small_init(keg);
1360 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1361 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1362 if (keg->uk_ipers > uma_max_ipers_ref)
1363 panic("Too many ref items per zone: %d > %d\n",
1364 keg->uk_ipers, uma_max_ipers_ref);
1365 keg->uk_slabzone = slabrefzone;
1367 keg->uk_slabzone = slabzone;
1371 * If we haven't booted yet we need allocations to go through the
1372 * startup cache until the vm is ready.
1374 if (keg->uk_ppera == 1) {
1375 #ifdef UMA_MD_SMALL_ALLOC
1376 keg->uk_allocf = uma_small_alloc;
1377 keg->uk_freef = uma_small_free;
1379 if (booted < UMA_STARTUP)
1380 keg->uk_allocf = startup_alloc;
1382 if (booted < UMA_STARTUP2)
1383 keg->uk_allocf = startup_alloc;
1385 } else if (booted < UMA_STARTUP2 &&
1386 (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1387 keg->uk_allocf = startup_alloc;
1390 * Initialize keg's lock
1392 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1395 * If we're putting the slab header in the actual page we need to
1396 * figure out where in each page it goes. This calculates a right
1397 * justified offset into the memory on an ALIGN_PTR boundary.
1399 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1402 /* Size of the slab struct and free list */
1403 totsize = sizeof(struct uma_slab);
1405 /* Size of the reference counts. */
1406 if (keg->uk_flags & UMA_ZONE_REFCNT)
1407 totsize += keg->uk_ipers * sizeof(uint32_t);
1409 if (totsize & UMA_ALIGN_PTR)
1410 totsize = (totsize & ~UMA_ALIGN_PTR) +
1411 (UMA_ALIGN_PTR + 1);
1412 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1415 * The only way the following is possible is if with our
1416 * UMA_ALIGN_PTR adjustments we are now bigger than
1417 * UMA_SLAB_SIZE. I haven't checked whether this is
1418 * mathematically possible for all cases, so we make
1421 totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1422 if (keg->uk_flags & UMA_ZONE_REFCNT)
1423 totsize += keg->uk_ipers * sizeof(uint32_t);
1424 if (totsize > PAGE_SIZE * keg->uk_ppera) {
1425 printf("zone %s ipers %d rsize %d size %d\n",
1426 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1428 panic("UMA slab won't fit.");
1432 if (keg->uk_flags & UMA_ZONE_HASH)
1433 hash_alloc(&keg->uk_hash);
1436 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1437 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1438 keg->uk_ipers, keg->uk_ppera,
1439 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1442 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1445 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1446 mtx_unlock(&uma_mtx);
1451 * Zone header ctor. This initializes all fields, locks, etc.
1453 * Arguments/Returns follow uma_ctor specifications
1454 * udata Actually uma_zctor_args
1457 zone_ctor(void *mem, int size, void *udata, int flags)
1459 struct uma_zctor_args *arg = udata;
1460 uma_zone_t zone = mem;
1465 zone->uz_name = arg->name;
1466 zone->uz_ctor = arg->ctor;
1467 zone->uz_dtor = arg->dtor;
1468 zone->uz_slab = zone_fetch_slab;
1469 zone->uz_init = NULL;
1470 zone->uz_fini = NULL;
1471 zone->uz_allocs = 0;
1474 zone->uz_sleeps = 0;
1477 zone->uz_warning = NULL;
1478 timevalclear(&zone->uz_ratecheck);
1481 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1484 * This is a pure cache zone, no kegs.
1487 if (arg->flags & UMA_ZONE_VM)
1488 arg->flags |= UMA_ZFLAG_CACHEONLY;
1489 zone->uz_flags = arg->flags;
1490 zone->uz_size = arg->size;
1491 zone->uz_import = arg->import;
1492 zone->uz_release = arg->release;
1493 zone->uz_arg = arg->arg;
1494 zone->uz_lockptr = &zone->uz_lock;
1499 * Use the regular zone/keg/slab allocator.
1501 zone->uz_import = (uma_import)zone_import;
1502 zone->uz_release = (uma_release)zone_release;
1503 zone->uz_arg = zone;
1505 if (arg->flags & UMA_ZONE_SECONDARY) {
1506 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1507 zone->uz_init = arg->uminit;
1508 zone->uz_fini = arg->fini;
1509 zone->uz_lockptr = &keg->uk_lock;
1510 zone->uz_flags |= UMA_ZONE_SECONDARY;
1513 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1514 if (LIST_NEXT(z, uz_link) == NULL) {
1515 LIST_INSERT_AFTER(z, zone, uz_link);
1520 mtx_unlock(&uma_mtx);
1521 } else if (keg == NULL) {
1522 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1523 arg->align, arg->flags)) == NULL)
1526 struct uma_kctor_args karg;
1529 /* We should only be here from uma_startup() */
1530 karg.size = arg->size;
1531 karg.uminit = arg->uminit;
1532 karg.fini = arg->fini;
1533 karg.align = arg->align;
1534 karg.flags = arg->flags;
1536 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1543 * Link in the first keg.
1545 zone->uz_klink.kl_keg = keg;
1546 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1547 zone->uz_lockptr = &keg->uk_lock;
1548 zone->uz_size = keg->uk_size;
1549 zone->uz_flags |= (keg->uk_flags &
1550 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1553 * Some internal zones don't have room allocated for the per cpu
1554 * caches. If we're internal, bail out here.
1556 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1557 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1558 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1563 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1564 zone->uz_count = bucket_select(zone->uz_size);
1566 zone->uz_count = BUCKET_MAX;
1572 * Keg header dtor. This frees all data, destroys locks, frees the hash
1573 * table and removes the keg from the global list.
1575 * Arguments/Returns follow uma_dtor specifications
1579 keg_dtor(void *arg, int size, void *udata)
1583 keg = (uma_keg_t)arg;
1585 if (keg->uk_free != 0) {
1586 printf("Freed UMA keg was not empty (%d items). "
1587 " Lost %d pages of memory.\n",
1588 keg->uk_free, keg->uk_pages);
1592 hash_free(&keg->uk_hash);
1600 * Arguments/Returns follow uma_dtor specifications
1604 zone_dtor(void *arg, int size, void *udata)
1610 zone = (uma_zone_t)arg;
1611 keg = zone_first_keg(zone);
1613 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1617 LIST_REMOVE(zone, uz_link);
1618 mtx_unlock(&uma_mtx);
1620 * XXX there are some races here where
1621 * the zone can be drained but zone lock
1622 * released and then refilled before we
1623 * remove it... we dont care for now
1625 zone_drain_wait(zone, M_WAITOK);
1627 * Unlink all of our kegs.
1629 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1630 klink->kl_keg = NULL;
1631 LIST_REMOVE(klink, kl_link);
1632 if (klink == &zone->uz_klink)
1634 free(klink, M_TEMP);
1637 * We only destroy kegs from non secondary zones.
1639 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1641 LIST_REMOVE(keg, uk_link);
1642 mtx_unlock(&uma_mtx);
1643 zone_free_item(kegs, keg, NULL, SKIP_NONE);
1645 ZONE_LOCK_FINI(zone);
1649 * Traverses every zone in the system and calls a callback
1652 * zfunc A pointer to a function which accepts a zone
1659 zone_foreach(void (*zfunc)(uma_zone_t))
1665 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1666 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1669 mtx_unlock(&uma_mtx);
1672 /* Public functions */
1675 uma_startup(void *bootmem, int boot_pages)
1677 struct uma_zctor_args args;
1683 printf("Creating uma keg headers zone and keg.\n");
1685 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1687 /* "manually" create the initial zone */
1688 memset(&args, 0, sizeof(args));
1689 args.name = "UMA Kegs";
1690 args.size = sizeof(struct uma_keg);
1691 args.ctor = keg_ctor;
1692 args.dtor = keg_dtor;
1693 args.uminit = zero_init;
1695 args.keg = &masterkeg;
1696 args.align = 32 - 1;
1697 args.flags = UMA_ZFLAG_INTERNAL;
1698 /* The initial zone has no Per cpu queues so it's smaller */
1699 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1702 printf("Filling boot free list.\n");
1704 for (i = 0; i < boot_pages; i++) {
1705 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1706 slab->us_data = (uint8_t *)slab;
1707 slab->us_flags = UMA_SLAB_BOOT;
1708 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1710 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1713 printf("Creating uma zone headers zone and keg.\n");
1715 args.name = "UMA Zones";
1716 args.size = sizeof(struct uma_zone) +
1717 (sizeof(struct uma_cache) * (mp_maxid + 1));
1718 args.ctor = zone_ctor;
1719 args.dtor = zone_dtor;
1720 args.uminit = zero_init;
1723 args.align = 32 - 1;
1724 args.flags = UMA_ZFLAG_INTERNAL;
1725 /* The initial zone has no Per cpu queues so it's smaller */
1726 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1729 printf("Initializing pcpu cache locks.\n");
1732 printf("Creating slab and hash zones.\n");
1735 /* Now make a zone for slab headers */
1736 slabzone = uma_zcreate("UMA Slabs",
1737 sizeof(struct uma_slab),
1738 NULL, NULL, NULL, NULL,
1739 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1742 * We also create a zone for the bigger slabs with reference
1743 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1745 slabsize = sizeof(struct uma_slab_refcnt);
1746 slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1747 slabrefzone = uma_zcreate("UMA RCntSlabs",
1749 NULL, NULL, NULL, NULL,
1751 UMA_ZFLAG_INTERNAL);
1753 hashzone = uma_zcreate("UMA Hash",
1754 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1755 NULL, NULL, NULL, NULL,
1756 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1760 booted = UMA_STARTUP;
1763 printf("UMA startup complete.\n");
1771 booted = UMA_STARTUP2;
1774 printf("UMA startup2 complete.\n");
1779 * Initialize our callout handle
1787 printf("Starting callout.\n");
1789 callout_init(&uma_callout, CALLOUT_MPSAFE);
1790 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1792 printf("UMA startup3 complete.\n");
1797 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1798 int align, uint32_t flags)
1800 struct uma_kctor_args args;
1803 args.uminit = uminit;
1805 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1808 return (zone_alloc_item(kegs, &args, M_WAITOK));
1813 uma_set_align(int align)
1816 if (align != UMA_ALIGN_CACHE)
1817 uma_align_cache = align;
1822 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1823 uma_init uminit, uma_fini fini, int align, uint32_t flags)
1826 struct uma_zctor_args args;
1828 /* This stuff is essential for the zone ctor */
1829 memset(&args, 0, sizeof(args));
1834 args.uminit = uminit;
1840 return (zone_alloc_item(zones, &args, M_WAITOK));
1845 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1846 uma_init zinit, uma_fini zfini, uma_zone_t master)
1848 struct uma_zctor_args args;
1851 keg = zone_first_keg(master);
1852 memset(&args, 0, sizeof(args));
1854 args.size = keg->uk_size;
1857 args.uminit = zinit;
1859 args.align = keg->uk_align;
1860 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1863 /* XXX Attaches only one keg of potentially many. */
1864 return (zone_alloc_item(zones, &args, M_WAITOK));
1869 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1870 uma_init zinit, uma_fini zfini, uma_import zimport,
1871 uma_release zrelease, void *arg, int flags)
1873 struct uma_zctor_args args;
1875 memset(&args, 0, sizeof(args));
1880 args.uminit = zinit;
1882 args.import = zimport;
1883 args.release = zrelease;
1888 return (zone_alloc_item(zones, &args, M_WAITOK));
1892 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1896 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
1899 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
1904 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1912 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1919 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1921 zone_lock_pair(zone, master);
1923 * zone must use vtoslab() to resolve objects and must already be
1926 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1927 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1932 * The new master must also use vtoslab().
1934 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
1939 * Both must either be refcnt, or not be refcnt.
1941 if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
1942 (master->uz_flags & UMA_ZONE_REFCNT)) {
1947 * The underlying object must be the same size. rsize
1950 if (master->uz_size != zone->uz_size) {
1955 * Put it at the end of the list.
1957 klink->kl_keg = zone_first_keg(master);
1958 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
1959 if (LIST_NEXT(kl, kl_link) == NULL) {
1960 LIST_INSERT_AFTER(kl, klink, kl_link);
1965 zone->uz_flags |= UMA_ZFLAG_MULTI;
1966 zone->uz_slab = zone_fetch_slab_multi;
1969 zone_unlock_pair(zone, master);
1971 free(klink, M_TEMP);
1979 uma_zdestroy(uma_zone_t zone)
1982 zone_free_item(zones, zone, NULL, SKIP_NONE);
1987 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1991 uma_bucket_t bucket;
1995 /* This is the fast path allocation */
1996 #ifdef UMA_DEBUG_ALLOC_1
1997 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1999 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2000 zone->uz_name, flags);
2002 if (flags & M_WAITOK) {
2003 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2004 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2006 #ifdef DEBUG_MEMGUARD
2007 if (memguard_cmp_zone(zone)) {
2008 item = memguard_alloc(zone->uz_size, flags);
2011 * Avoid conflict with the use-after-free
2012 * protecting infrastructure from INVARIANTS.
2014 if (zone->uz_init != NULL &&
2015 zone->uz_init != mtrash_init &&
2016 zone->uz_init(item, zone->uz_size, flags) != 0)
2018 if (zone->uz_ctor != NULL &&
2019 zone->uz_ctor != mtrash_ctor &&
2020 zone->uz_ctor(item, zone->uz_size, udata,
2022 zone->uz_fini(item, zone->uz_size);
2027 /* This is unfortunate but should not be fatal. */
2031 * If possible, allocate from the per-CPU cache. There are two
2032 * requirements for safe access to the per-CPU cache: (1) the thread
2033 * accessing the cache must not be preempted or yield during access,
2034 * and (2) the thread must not migrate CPUs without switching which
2035 * cache it accesses. We rely on a critical section to prevent
2036 * preemption and migration. We release the critical section in
2037 * order to acquire the zone mutex if we are unable to allocate from
2038 * the current cache; when we re-acquire the critical section, we
2039 * must detect and handle migration if it has occurred.
2043 cache = &zone->uz_cpu[cpu];
2046 bucket = cache->uc_allocbucket;
2047 if (bucket != NULL && bucket->ub_cnt > 0) {
2049 item = bucket->ub_bucket[bucket->ub_cnt];
2051 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2053 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2056 if (zone->uz_ctor != NULL &&
2057 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2058 atomic_add_long(&zone->uz_fails, 1);
2059 zone_free_item(zone, item, udata, SKIP_DTOR);
2063 uma_dbg_alloc(zone, NULL, item);
2066 bzero(item, zone->uz_size);
2071 * We have run out of items in our alloc bucket.
2072 * See if we can switch with our free bucket.
2074 bucket = cache->uc_freebucket;
2075 if (bucket != NULL && bucket->ub_cnt > 0) {
2076 #ifdef UMA_DEBUG_ALLOC
2077 printf("uma_zalloc: Swapping empty with alloc.\n");
2079 cache->uc_freebucket = cache->uc_allocbucket;
2080 cache->uc_allocbucket = bucket;
2085 * Discard any empty allocation bucket while we hold no locks.
2087 bucket = cache->uc_allocbucket;
2088 cache->uc_allocbucket = NULL;
2091 bucket_free(zone, bucket, udata);
2093 /* Short-circuit for zones without buckets and low memory. */
2094 if (zone->uz_count == 0 || bucketdisable)
2098 * Attempt to retrieve the item from the per-CPU cache has failed, so
2099 * we must go back to the zone. This requires the zone lock, so we
2100 * must drop the critical section, then re-acquire it when we go back
2101 * to the cache. Since the critical section is released, we may be
2102 * preempted or migrate. As such, make sure not to maintain any
2103 * thread-local state specific to the cache from prior to releasing
2104 * the critical section.
2107 if (ZONE_TRYLOCK(zone) == 0) {
2108 /* Record contention to size the buckets. */
2114 cache = &zone->uz_cpu[cpu];
2117 * Since we have locked the zone we may as well send back our stats.
2119 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2120 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2121 cache->uc_allocs = 0;
2122 cache->uc_frees = 0;
2124 /* See if we lost the race to fill the cache. */
2125 if (cache->uc_allocbucket != NULL) {
2131 * Check the zone's cache of buckets.
2133 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2134 KASSERT(bucket->ub_cnt != 0,
2135 ("uma_zalloc_arg: Returning an empty bucket."));
2137 LIST_REMOVE(bucket, ub_link);
2138 cache->uc_allocbucket = bucket;
2142 /* We are no longer associated with this CPU. */
2146 * We bump the uz count when the cache size is insufficient to
2147 * handle the working set.
2149 if (lockfail && zone->uz_count < BUCKET_MAX)
2154 * Now lets just fill a bucket and put it on the free list. If that
2155 * works we'll restart the allocation from the begining and it
2156 * will use the just filled bucket.
2158 bucket = zone_alloc_bucket(zone, udata, flags);
2159 if (bucket != NULL) {
2163 cache = &zone->uz_cpu[cpu];
2165 * See if we lost the race or were migrated. Cache the
2166 * initialized bucket to make this less likely or claim
2167 * the memory directly.
2169 if (cache->uc_allocbucket == NULL)
2170 cache->uc_allocbucket = bucket;
2172 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2178 * We may not be able to get a bucket so return an actual item.
2181 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2185 item = zone_alloc_item(zone, udata, flags);
2191 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2196 mtx_assert(&keg->uk_lock, MA_OWNED);
2199 if ((flags & M_USE_RESERVE) == 0)
2200 reserve = keg->uk_reserve;
2204 * Find a slab with some space. Prefer slabs that are partially
2205 * used over those that are totally full. This helps to reduce
2208 if (keg->uk_free > reserve) {
2209 if (!LIST_EMPTY(&keg->uk_part_slab)) {
2210 slab = LIST_FIRST(&keg->uk_part_slab);
2212 slab = LIST_FIRST(&keg->uk_free_slab);
2213 LIST_REMOVE(slab, us_link);
2214 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2217 MPASS(slab->us_keg == keg);
2222 * M_NOVM means don't ask at all!
2227 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2228 keg->uk_flags |= UMA_ZFLAG_FULL;
2230 * If this is not a multi-zone, set the FULL bit.
2231 * Otherwise slab_multi() takes care of it.
2233 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2234 zone->uz_flags |= UMA_ZFLAG_FULL;
2235 zone_log_warning(zone);
2237 if (flags & M_NOWAIT)
2240 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2243 slab = keg_alloc_slab(keg, zone, flags);
2245 * If we got a slab here it's safe to mark it partially used
2246 * and return. We assume that the caller is going to remove
2247 * at least one item.
2250 MPASS(slab->us_keg == keg);
2251 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2255 * We might not have been able to get a slab but another cpu
2256 * could have while we were unlocked. Check again before we
2265 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2270 keg = zone_first_keg(zone);
2275 slab = keg_fetch_slab(keg, zone, flags);
2278 if (flags & (M_NOWAIT | M_NOVM))
2286 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2287 * with the keg locked. On NULL no lock is held.
2289 * The last pointer is used to seed the search. It is not required.
2292 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2302 * Don't wait on the first pass. This will skip limit tests
2303 * as well. We don't want to block if we can find a provider
2306 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2308 * Use the last slab allocated as a hint for where to start
2312 slab = keg_fetch_slab(last, zone, flags);
2318 * Loop until we have a slab incase of transient failures
2319 * while M_WAITOK is specified. I'm not sure this is 100%
2320 * required but we've done it for so long now.
2326 * Search the available kegs for slabs. Be careful to hold the
2327 * correct lock while calling into the keg layer.
2329 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2330 keg = klink->kl_keg;
2332 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2333 slab = keg_fetch_slab(keg, zone, flags);
2337 if (keg->uk_flags & UMA_ZFLAG_FULL)
2343 if (rflags & (M_NOWAIT | M_NOVM))
2347 * All kegs are full. XXX We can't atomically check all kegs
2348 * and sleep so just sleep for a short period and retry.
2350 if (full && !empty) {
2352 zone->uz_flags |= UMA_ZFLAG_FULL;
2354 zone_log_warning(zone);
2355 msleep(zone, zone->uz_lockptr, PVM,
2356 "zonelimit", hz/100);
2357 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2366 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2371 MPASS(keg == slab->us_keg);
2372 mtx_assert(&keg->uk_lock, MA_OWNED);
2374 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2375 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2376 item = slab->us_data + (keg->uk_rsize * freei);
2377 slab->us_freecount--;
2380 /* Move this slab to the full list */
2381 if (slab->us_freecount == 0) {
2382 LIST_REMOVE(slab, us_link);
2383 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2390 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2398 /* Try to keep the buckets totally full */
2399 for (i = 0; i < max; ) {
2400 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2403 while (slab->us_freecount && i < max) {
2404 bucket[i++] = slab_alloc_item(keg, slab);
2405 if (keg->uk_free <= keg->uk_reserve)
2408 /* Don't grab more than one slab at a time. */
2419 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2421 uma_bucket_t bucket;
2424 /* Don't wait for buckets, preserve caller's NOVM setting. */
2425 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2429 max = MIN(bucket->ub_entries, zone->uz_count);
2430 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2434 * Initialize the memory if necessary.
2436 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2439 for (i = 0; i < bucket->ub_cnt; i++)
2440 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2444 * If we couldn't initialize the whole bucket, put the
2445 * rest back onto the freelist.
2447 if (i != bucket->ub_cnt) {
2448 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2449 bucket->ub_cnt - i);
2451 bzero(&bucket->ub_bucket[i],
2452 sizeof(void *) * (bucket->ub_cnt - i));
2459 if (bucket == NULL || bucket->ub_cnt == 0) {
2461 bucket_free(zone, bucket, udata);
2462 atomic_add_long(&zone->uz_fails, 1);
2470 * Allocates a single item from a zone.
2473 * zone The zone to alloc for.
2474 * udata The data to be passed to the constructor.
2475 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2478 * NULL if there is no memory and M_NOWAIT is set
2479 * An item if successful
2483 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2489 #ifdef UMA_DEBUG_ALLOC
2490 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2492 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2494 atomic_add_long(&zone->uz_allocs, 1);
2497 * We have to call both the zone's init (not the keg's init)
2498 * and the zone's ctor. This is because the item is going from
2499 * a keg slab directly to the user, and the user is expecting it
2500 * to be both zone-init'd as well as zone-ctor'd.
2502 if (zone->uz_init != NULL) {
2503 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2504 zone_free_item(zone, item, udata, SKIP_FINI);
2508 if (zone->uz_ctor != NULL) {
2509 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2510 zone_free_item(zone, item, udata, SKIP_DTOR);
2515 uma_dbg_alloc(zone, NULL, item);
2518 bzero(item, zone->uz_size);
2523 atomic_add_long(&zone->uz_fails, 1);
2529 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2532 uma_bucket_t bucket;
2535 #ifdef UMA_DEBUG_ALLOC_1
2536 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2538 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2541 /* uma_zfree(..., NULL) does nothing, to match free(9). */
2544 #ifdef DEBUG_MEMGUARD
2545 if (is_memguard_addr(item)) {
2546 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2547 zone->uz_dtor(item, zone->uz_size, udata);
2548 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2549 zone->uz_fini(item, zone->uz_size);
2550 memguard_free(item);
2555 if (zone->uz_flags & UMA_ZONE_MALLOC)
2556 uma_dbg_free(zone, udata, item);
2558 uma_dbg_free(zone, NULL, item);
2560 if (zone->uz_dtor != NULL)
2561 zone->uz_dtor(item, zone->uz_size, udata);
2564 * The race here is acceptable. If we miss it we'll just have to wait
2565 * a little longer for the limits to be reset.
2567 if (zone->uz_flags & UMA_ZFLAG_FULL)
2571 * If possible, free to the per-CPU cache. There are two
2572 * requirements for safe access to the per-CPU cache: (1) the thread
2573 * accessing the cache must not be preempted or yield during access,
2574 * and (2) the thread must not migrate CPUs without switching which
2575 * cache it accesses. We rely on a critical section to prevent
2576 * preemption and migration. We release the critical section in
2577 * order to acquire the zone mutex if we are unable to free to the
2578 * current cache; when we re-acquire the critical section, we must
2579 * detect and handle migration if it has occurred.
2584 cache = &zone->uz_cpu[cpu];
2588 * Try to free into the allocbucket first to give LIFO ordering
2589 * for cache-hot datastructures. Spill over into the freebucket
2590 * if necessary. Alloc will swap them if one runs dry.
2592 bucket = cache->uc_allocbucket;
2593 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2594 bucket = cache->uc_freebucket;
2595 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2596 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2597 ("uma_zfree: Freeing to non free bucket index."));
2598 bucket->ub_bucket[bucket->ub_cnt] = item;
2606 * We must go back the zone, which requires acquiring the zone lock,
2607 * which in turn means we must release and re-acquire the critical
2608 * section. Since the critical section is released, we may be
2609 * preempted or migrate. As such, make sure not to maintain any
2610 * thread-local state specific to the cache from prior to releasing
2611 * the critical section.
2614 if (zone->uz_count == 0 || bucketdisable)
2620 cache = &zone->uz_cpu[cpu];
2623 * Since we have locked the zone we may as well send back our stats.
2625 atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2626 atomic_add_long(&zone->uz_frees, cache->uc_frees);
2627 cache->uc_allocs = 0;
2628 cache->uc_frees = 0;
2630 bucket = cache->uc_freebucket;
2631 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2635 cache->uc_freebucket = NULL;
2637 /* Can we throw this on the zone full list? */
2638 if (bucket != NULL) {
2639 #ifdef UMA_DEBUG_ALLOC
2640 printf("uma_zfree: Putting old bucket on the free list.\n");
2642 /* ub_cnt is pointing to the last free item */
2643 KASSERT(bucket->ub_cnt != 0,
2644 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2645 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2648 /* We are no longer associated with this CPU. */
2651 /* And the zone.. */
2654 #ifdef UMA_DEBUG_ALLOC
2655 printf("uma_zfree: Allocating new free bucket.\n");
2657 bucket = bucket_alloc(zone, udata, M_NOWAIT);
2661 cache = &zone->uz_cpu[cpu];
2662 if (cache->uc_freebucket == NULL) {
2663 cache->uc_freebucket = bucket;
2667 * We lost the race, start over. We have to drop our
2668 * critical section to free the bucket.
2671 bucket_free(zone, bucket, udata);
2676 * If nothing else caught this, we'll just do an internal free.
2679 zone_free_item(zone, item, udata, SKIP_DTOR);
2685 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2689 mtx_assert(&keg->uk_lock, MA_OWNED);
2690 MPASS(keg == slab->us_keg);
2692 /* Do we need to remove from any lists? */
2693 if (slab->us_freecount+1 == keg->uk_ipers) {
2694 LIST_REMOVE(slab, us_link);
2695 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2696 } else if (slab->us_freecount == 0) {
2697 LIST_REMOVE(slab, us_link);
2698 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2701 /* Slab management. */
2702 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2703 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2704 slab->us_freecount++;
2706 /* Keg statistics. */
2711 zone_release(uma_zone_t zone, void **bucket, int cnt)
2721 keg = zone_first_keg(zone);
2723 for (i = 0; i < cnt; i++) {
2725 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2726 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2727 if (zone->uz_flags & UMA_ZONE_HASH) {
2728 slab = hash_sfind(&keg->uk_hash, mem);
2730 mem += keg->uk_pgoff;
2731 slab = (uma_slab_t)mem;
2734 slab = vtoslab((vm_offset_t)item);
2735 if (slab->us_keg != keg) {
2741 slab_free_item(keg, slab, item);
2742 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2743 if (keg->uk_pages < keg->uk_maxpages) {
2744 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2749 * We can handle one more allocation. Since we're
2750 * clearing ZFLAG_FULL, wake up all procs blocked
2751 * on pages. This should be uncommon, so keeping this
2752 * simple for now (rather than adding count of blocked
2761 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2769 * Frees a single item to any zone.
2772 * zone The zone to free to
2773 * item The item we're freeing
2774 * udata User supplied data for the dtor
2775 * skip Skip dtors and finis
2778 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2782 if (skip == SKIP_NONE) {
2783 if (zone->uz_flags & UMA_ZONE_MALLOC)
2784 uma_dbg_free(zone, udata, item);
2786 uma_dbg_free(zone, NULL, item);
2789 if (skip < SKIP_DTOR && zone->uz_dtor)
2790 zone->uz_dtor(item, zone->uz_size, udata);
2792 if (skip < SKIP_FINI && zone->uz_fini)
2793 zone->uz_fini(item, zone->uz_size);
2795 atomic_add_long(&zone->uz_frees, 1);
2796 zone->uz_release(zone->uz_arg, &item, 1);
2801 uma_zone_set_max(uma_zone_t zone, int nitems)
2805 keg = zone_first_keg(zone);
2809 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2810 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2811 keg->uk_maxpages += keg->uk_ppera;
2812 nitems = keg->uk_maxpages * keg->uk_ipers;
2820 uma_zone_get_max(uma_zone_t zone)
2825 keg = zone_first_keg(zone);
2829 nitems = keg->uk_maxpages * keg->uk_ipers;
2837 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2841 zone->uz_warning = warning;
2847 uma_zone_get_cur(uma_zone_t zone)
2853 nitems = zone->uz_allocs - zone->uz_frees;
2856 * See the comment in sysctl_vm_zone_stats() regarding the
2857 * safety of accessing the per-cpu caches. With the zone lock
2858 * held, it is safe, but can potentially result in stale data.
2860 nitems += zone->uz_cpu[i].uc_allocs -
2861 zone->uz_cpu[i].uc_frees;
2865 return (nitems < 0 ? 0 : nitems);
2870 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2874 keg = zone_first_keg(zone);
2875 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2877 KASSERT(keg->uk_pages == 0,
2878 ("uma_zone_set_init on non-empty keg"));
2879 keg->uk_init = uminit;
2885 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2889 keg = zone_first_keg(zone);
2890 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2892 KASSERT(keg->uk_pages == 0,
2893 ("uma_zone_set_fini on non-empty keg"));
2894 keg->uk_fini = fini;
2900 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2904 KASSERT(zone_first_keg(zone)->uk_pages == 0,
2905 ("uma_zone_set_zinit on non-empty keg"));
2906 zone->uz_init = zinit;
2912 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2916 KASSERT(zone_first_keg(zone)->uk_pages == 0,
2917 ("uma_zone_set_zfini on non-empty keg"));
2918 zone->uz_fini = zfini;
2923 /* XXX uk_freef is not actually used with the zone locked */
2925 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2929 keg = zone_first_keg(zone);
2930 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2932 keg->uk_freef = freef;
2937 /* XXX uk_allocf is not actually used with the zone locked */
2939 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2943 keg = zone_first_keg(zone);
2945 keg->uk_allocf = allocf;
2951 uma_zone_reserve(uma_zone_t zone, int items)
2955 keg = zone_first_keg(zone);
2959 keg->uk_reserve = items;
2967 uma_zone_reserve_kva(uma_zone_t zone, int count)
2973 keg = zone_first_keg(zone);
2976 pages = count / keg->uk_ipers;
2978 if (pages * keg->uk_ipers < count)
2981 #ifdef UMA_MD_SMALL_ALLOC
2982 if (keg->uk_ppera > 1) {
2986 kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2994 keg->uk_maxpages = pages;
2995 #ifdef UMA_MD_SMALL_ALLOC
2996 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
2998 keg->uk_allocf = noobj_alloc;
3000 keg->uk_flags |= UMA_ZONE_NOFREE;
3008 uma_prealloc(uma_zone_t zone, int items)
3014 keg = zone_first_keg(zone);
3018 slabs = items / keg->uk_ipers;
3019 if (slabs * keg->uk_ipers < items)
3022 slab = keg_alloc_slab(keg, zone, M_WAITOK);
3025 MPASS(slab->us_keg == keg);
3026 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3034 uma_find_refcnt(uma_zone_t zone, void *item)
3036 uma_slabrefcnt_t slabref;
3042 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3043 slabref = (uma_slabrefcnt_t)slab;
3045 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3046 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3047 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3048 refcnt = &slabref->us_refcnt[idx];
3057 printf("UMA: vm asked us to release pages!\n");
3060 zone_foreach(zone_drain);
3062 * Some slabs may have been freed but this zone will be visited early
3063 * we visit again so that we can free pages that are empty once other
3064 * zones are drained. We have to do the same for buckets.
3066 zone_drain(slabzone);
3067 zone_drain(slabrefzone);
3068 bucket_zone_drain();
3073 uma_zone_exhausted(uma_zone_t zone)
3078 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3084 uma_zone_exhausted_nolock(uma_zone_t zone)
3086 return (zone->uz_flags & UMA_ZFLAG_FULL);
3090 uma_large_malloc(int size, int wait)
3096 slab = zone_alloc_item(slabzone, NULL, wait);
3099 mem = page_alloc(NULL, size, &flags, wait);
3101 vsetslab((vm_offset_t)mem, slab);
3102 slab->us_data = mem;
3103 slab->us_flags = flags | UMA_SLAB_MALLOC;
3104 slab->us_size = size;
3106 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3113 uma_large_free(uma_slab_t slab)
3115 vsetobj((vm_offset_t)slab->us_data, kmem_object);
3116 page_free(slab->us_data, slab->us_size, slab->us_flags);
3117 zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3121 uma_print_stats(void)
3123 zone_foreach(uma_print_zone);
3127 slab_print(uma_slab_t slab)
3129 printf("slab: keg %p, data %p, freecount %d\n",
3130 slab->us_keg, slab->us_data, slab->us_freecount);
3134 cache_print(uma_cache_t cache)
3136 printf("alloc: %p(%d), free: %p(%d)\n",
3137 cache->uc_allocbucket,
3138 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3139 cache->uc_freebucket,
3140 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3144 uma_print_keg(uma_keg_t keg)
3148 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3149 "out %d free %d limit %d\n",
3150 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3151 keg->uk_ipers, keg->uk_ppera,
3152 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3153 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3154 printf("Part slabs:\n");
3155 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3157 printf("Free slabs:\n");
3158 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3160 printf("Full slabs:\n");
3161 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3166 uma_print_zone(uma_zone_t zone)
3172 printf("zone: %s(%p) size %d flags %#x\n",
3173 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3174 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3175 uma_print_keg(kl->kl_keg);
3177 cache = &zone->uz_cpu[i];
3178 printf("CPU %d Cache:\n", i);
3185 * Generate statistics across both the zone and its per-cpu cache's. Return
3186 * desired statistics if the pointer is non-NULL for that statistic.
3188 * Note: does not update the zone statistics, as it can't safely clear the
3189 * per-CPU cache statistic.
3191 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3192 * safe from off-CPU; we should modify the caches to track this information
3193 * directly so that we don't have to.
3196 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3197 uint64_t *freesp, uint64_t *sleepsp)
3200 uint64_t allocs, frees, sleeps;
3203 allocs = frees = sleeps = 0;
3206 cache = &z->uz_cpu[cpu];
3207 if (cache->uc_allocbucket != NULL)
3208 cachefree += cache->uc_allocbucket->ub_cnt;
3209 if (cache->uc_freebucket != NULL)
3210 cachefree += cache->uc_freebucket->ub_cnt;
3211 allocs += cache->uc_allocs;
3212 frees += cache->uc_frees;
3214 allocs += z->uz_allocs;
3215 frees += z->uz_frees;
3216 sleeps += z->uz_sleeps;
3217 if (cachefreep != NULL)
3218 *cachefreep = cachefree;
3219 if (allocsp != NULL)
3223 if (sleepsp != NULL)
3229 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3237 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3238 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3241 mtx_unlock(&uma_mtx);
3242 return (sysctl_handle_int(oidp, &count, 0, req));
3246 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3248 struct uma_stream_header ush;
3249 struct uma_type_header uth;
3250 struct uma_percpu_stat ups;
3251 uma_bucket_t bucket;
3258 int count, error, i;
3260 error = sysctl_wire_old_buffer(req, 0);
3263 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3267 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3268 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3273 * Insert stream header.
3275 bzero(&ush, sizeof(ush));
3276 ush.ush_version = UMA_STREAM_VERSION;
3277 ush.ush_maxcpus = (mp_maxid + 1);
3278 ush.ush_count = count;
3279 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3281 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3282 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3283 bzero(&uth, sizeof(uth));
3285 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3286 uth.uth_align = kz->uk_align;
3287 uth.uth_size = kz->uk_size;
3288 uth.uth_rsize = kz->uk_rsize;
3289 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3291 uth.uth_maxpages += k->uk_maxpages;
3292 uth.uth_pages += k->uk_pages;
3293 uth.uth_keg_free += k->uk_free;
3294 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3299 * A zone is secondary is it is not the first entry
3300 * on the keg's zone list.
3302 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3303 (LIST_FIRST(&kz->uk_zones) != z))
3304 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3306 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3307 uth.uth_zone_free += bucket->ub_cnt;
3308 uth.uth_allocs = z->uz_allocs;
3309 uth.uth_frees = z->uz_frees;
3310 uth.uth_fails = z->uz_fails;
3311 uth.uth_sleeps = z->uz_sleeps;
3312 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3314 * While it is not normally safe to access the cache
3315 * bucket pointers while not on the CPU that owns the
3316 * cache, we only allow the pointers to be exchanged
3317 * without the zone lock held, not invalidated, so
3318 * accept the possible race associated with bucket
3319 * exchange during monitoring.
3321 for (i = 0; i < (mp_maxid + 1); i++) {
3322 bzero(&ups, sizeof(ups));
3323 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3327 cache = &z->uz_cpu[i];
3328 if (cache->uc_allocbucket != NULL)
3329 ups.ups_cache_free +=
3330 cache->uc_allocbucket->ub_cnt;
3331 if (cache->uc_freebucket != NULL)
3332 ups.ups_cache_free +=
3333 cache->uc_freebucket->ub_cnt;
3334 ups.ups_allocs = cache->uc_allocs;
3335 ups.ups_frees = cache->uc_frees;
3337 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3342 mtx_unlock(&uma_mtx);
3343 error = sbuf_finish(&sbuf);
3349 DB_SHOW_COMMAND(uma, db_show_uma)
3351 uint64_t allocs, frees, sleeps;
3352 uma_bucket_t bucket;
3357 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3358 "Requests", "Sleeps");
3359 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3360 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3361 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3362 allocs = z->uz_allocs;
3363 frees = z->uz_frees;
3364 sleeps = z->uz_sleeps;
3367 uma_zone_sumstat(z, &cachefree, &allocs,
3369 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3370 (LIST_FIRST(&kz->uk_zones) != z)))
3371 cachefree += kz->uk_free;
3372 LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3373 cachefree += bucket->ub_cnt;
3374 db_printf("%18s %8ju %8jd %8d %12ju %8ju\n", z->uz_name,
3375 (uintmax_t)kz->uk_size,
3376 (intmax_t)(allocs - frees), cachefree,
3377 (uintmax_t)allocs, sleeps);