2 * Copyright (c) 2002-2005, 2009 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"
63 #include <sys/param.h>
64 #include <sys/systm.h>
65 #include <sys/kernel.h>
66 #include <sys/types.h>
67 #include <sys/queue.h>
68 #include <sys/malloc.h>
71 #include <sys/sysctl.h>
72 #include <sys/mutex.h>
76 #include <sys/vmmeter.h>
79 #include <vm/vm_object.h>
80 #include <vm/vm_page.h>
81 #include <vm/vm_param.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_kern.h>
84 #include <vm/vm_extern.h>
86 #include <vm/uma_int.h>
87 #include <vm/uma_dbg.h>
89 #include <machine/vmparam.h>
94 * This is the zone and keg from which all zones are spawned. The idea is that
95 * even the zone & keg heads are allocated from the allocator, so we use the
96 * bss section to bootstrap us.
98 static struct uma_keg masterkeg;
99 static struct uma_zone masterzone_k;
100 static struct uma_zone masterzone_z;
101 static uma_zone_t kegs = &masterzone_k;
102 static uma_zone_t zones = &masterzone_z;
104 /* This is the zone from which all of uma_slab_t's are allocated. */
105 static uma_zone_t slabzone;
106 static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
109 * The initial hash tables come out of this zone so they can be allocated
110 * prior to malloc coming up.
112 static uma_zone_t hashzone;
114 /* The boot-time adjusted value for cache line alignment. */
115 static int uma_align_cache = 64 - 1;
117 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
120 * Are we allowed to allocate buckets?
122 static int bucketdisable = 1;
124 /* Linked list of all kegs in the system */
125 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
127 /* This mutex protects the keg list */
128 static struct mtx uma_mtx;
130 /* Linked list of boot time pages */
131 static LIST_HEAD(,uma_slab) uma_boot_pages =
132 LIST_HEAD_INITIALIZER(uma_boot_pages);
134 /* This mutex protects the boot time pages list */
135 static struct mtx uma_boot_pages_mtx;
137 /* Is the VM done starting up? */
138 static int booted = 0;
140 /* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
141 static u_int uma_max_ipers;
142 static u_int uma_max_ipers_ref;
145 * This is the handle used to schedule events that need to happen
146 * outside of the allocation fast path.
148 static struct callout uma_callout;
149 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
152 * This structure is passed as the zone ctor arg so that I don't have to create
153 * a special allocation function just for zones.
155 struct uma_zctor_args {
167 struct uma_kctor_args {
176 struct uma_bucket_zone {
182 #define BUCKET_MAX 128
184 struct uma_bucket_zone bucket_zones[] = {
185 { NULL, "16 Bucket", 16 },
186 { NULL, "32 Bucket", 32 },
187 { NULL, "64 Bucket", 64 },
188 { NULL, "128 Bucket", 128 },
192 #define BUCKET_SHIFT 4
193 #define BUCKET_ZONES ((BUCKET_MAX >> BUCKET_SHIFT) + 1)
196 * bucket_size[] maps requested bucket sizes to zones that allocate a bucket
197 * of approximately the right size.
199 static uint8_t bucket_size[BUCKET_ZONES];
202 * Flags and enumerations to be passed to internal functions.
204 enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI };
206 #define ZFREE_STATFAIL 0x00000001 /* Update zone failure statistic. */
207 #define ZFREE_STATFREE 0x00000002 /* Update zone free statistic. */
211 static void *obj_alloc(uma_zone_t, int, u_int8_t *, int);
212 static void *page_alloc(uma_zone_t, int, u_int8_t *, int);
213 static void *startup_alloc(uma_zone_t, int, u_int8_t *, int);
214 static void page_free(void *, int, u_int8_t);
215 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
216 static void cache_drain(uma_zone_t);
217 static void bucket_drain(uma_zone_t, uma_bucket_t);
218 static void bucket_cache_drain(uma_zone_t zone);
219 static int keg_ctor(void *, int, void *, int);
220 static void keg_dtor(void *, int, void *);
221 static int zone_ctor(void *, int, void *, int);
222 static void zone_dtor(void *, int, void *);
223 static int zero_init(void *, int, int);
224 static void keg_small_init(uma_keg_t keg);
225 static void keg_large_init(uma_keg_t keg);
226 static void zone_foreach(void (*zfunc)(uma_zone_t));
227 static void zone_timeout(uma_zone_t zone);
228 static int hash_alloc(struct uma_hash *);
229 static int hash_expand(struct uma_hash *, struct uma_hash *);
230 static void hash_free(struct uma_hash *hash);
231 static void uma_timeout(void *);
232 static void uma_startup3(void);
233 static void *zone_alloc_item(uma_zone_t, void *, int);
234 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip,
236 static void bucket_enable(void);
237 static void bucket_init(void);
238 static uma_bucket_t bucket_alloc(int, int);
239 static void bucket_free(uma_bucket_t);
240 static void bucket_zone_drain(void);
241 static int zone_alloc_bucket(uma_zone_t zone, int flags);
242 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
243 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
244 static void *slab_alloc_item(uma_zone_t zone, uma_slab_t slab);
245 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
246 uma_fini fini, int align, u_int32_t flags);
247 static inline void zone_relock(uma_zone_t zone, uma_keg_t keg);
248 static inline void keg_relock(uma_keg_t keg, uma_zone_t zone);
250 void uma_print_zone(uma_zone_t);
251 void uma_print_stats(void);
252 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
253 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
255 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
257 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
258 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
260 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
261 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
264 * This routine checks to see whether or not it's safe to enable buckets.
270 if (cnt.v_free_count < cnt.v_free_min)
277 * Initialize bucket_zones, the array of zones of buckets of various sizes.
279 * For each zone, calculate the memory required for each bucket, consisting
280 * of the header and an array of pointers. Initialize bucket_size[] to point
281 * the range of appropriate bucket sizes at the zone.
286 struct uma_bucket_zone *ubz;
290 for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) {
293 ubz = &bucket_zones[j];
294 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
295 size += sizeof(void *) * ubz->ubz_entries;
296 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
297 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
298 UMA_ZFLAG_INTERNAL | UMA_ZFLAG_BUCKET);
299 for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT))
300 bucket_size[i >> BUCKET_SHIFT] = j;
305 * Given a desired number of entries for a bucket, return the zone from which
306 * to allocate the bucket.
308 static struct uma_bucket_zone *
309 bucket_zone_lookup(int entries)
313 idx = howmany(entries, 1 << BUCKET_SHIFT);
314 return (&bucket_zones[bucket_size[idx]]);
318 bucket_alloc(int entries, int bflags)
320 struct uma_bucket_zone *ubz;
324 * This is to stop us from allocating per cpu buckets while we're
325 * running out of vm.boot_pages. Otherwise, we would exhaust the
326 * boot pages. This also prevents us from allocating buckets in
327 * low memory situations.
332 ubz = bucket_zone_lookup(entries);
333 bucket = zone_alloc_item(ubz->ubz_zone, NULL, bflags);
336 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
339 bucket->ub_entries = ubz->ubz_entries;
346 bucket_free(uma_bucket_t bucket)
348 struct uma_bucket_zone *ubz;
350 ubz = bucket_zone_lookup(bucket->ub_entries);
351 zone_free_item(ubz->ubz_zone, bucket, NULL, SKIP_NONE,
356 bucket_zone_drain(void)
358 struct uma_bucket_zone *ubz;
360 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
361 zone_drain(ubz->ubz_zone);
364 static inline uma_keg_t
365 zone_first_keg(uma_zone_t zone)
368 return (LIST_FIRST(&zone->uz_kegs)->kl_keg);
372 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
376 LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
377 kegfn(klink->kl_keg);
381 * Routine called by timeout which is used to fire off some time interval
382 * based calculations. (stats, hash size, etc.)
391 uma_timeout(void *unused)
394 zone_foreach(zone_timeout);
396 /* Reschedule this event */
397 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
401 * Routine to perform timeout driven calculations. This expands the
402 * hashes and does per cpu statistics aggregation.
407 keg_timeout(uma_keg_t keg)
412 * Expand the keg hash table.
414 * This is done if the number of slabs is larger than the hash size.
415 * What I'm trying to do here is completely reduce collisions. This
416 * may be a little aggressive. Should I allow for two collisions max?
418 if (keg->uk_flags & UMA_ZONE_HASH &&
419 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
420 struct uma_hash newhash;
421 struct uma_hash oldhash;
425 * This is so involved because allocating and freeing
426 * while the keg lock is held will lead to deadlock.
427 * I have to do everything in stages and check for
430 newhash = keg->uk_hash;
432 ret = hash_alloc(&newhash);
435 if (hash_expand(&keg->uk_hash, &newhash)) {
436 oldhash = keg->uk_hash;
437 keg->uk_hash = newhash;
450 zone_timeout(uma_zone_t zone)
453 zone_foreach_keg(zone, &keg_timeout);
457 * Allocate and zero fill the next sized hash table from the appropriate
461 * hash A new hash structure with the old hash size in uh_hashsize
464 * 1 on sucess and 0 on failure.
467 hash_alloc(struct uma_hash *hash)
472 oldsize = hash->uh_hashsize;
474 /* We're just going to go to a power of two greater */
476 hash->uh_hashsize = oldsize * 2;
477 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
478 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
479 M_UMAHASH, M_NOWAIT);
481 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
482 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
484 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
486 if (hash->uh_slab_hash) {
487 bzero(hash->uh_slab_hash, alloc);
488 hash->uh_hashmask = hash->uh_hashsize - 1;
496 * Expands the hash table for HASH zones. This is done from zone_timeout
497 * to reduce collisions. This must not be done in the regular allocation
498 * path, otherwise, we can recurse on the vm while allocating pages.
501 * oldhash The hash you want to expand
502 * newhash The hash structure for the new table
510 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
516 if (!newhash->uh_slab_hash)
519 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
523 * I need to investigate hash algorithms for resizing without a
527 for (i = 0; i < oldhash->uh_hashsize; i++)
528 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
529 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
530 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
531 hval = UMA_HASH(newhash, slab->us_data);
532 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
540 * Free the hash bucket to the appropriate backing store.
543 * slab_hash The hash bucket we're freeing
544 * hashsize The number of entries in that hash bucket
550 hash_free(struct uma_hash *hash)
552 if (hash->uh_slab_hash == NULL)
554 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
555 zone_free_item(hashzone,
556 hash->uh_slab_hash, NULL, SKIP_NONE, ZFREE_STATFREE);
558 free(hash->uh_slab_hash, M_UMAHASH);
562 * Frees all outstanding items in a bucket
565 * zone The zone to free to, must be unlocked.
566 * bucket The free/alloc bucket with items, cpu queue must be locked.
573 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
580 while (bucket->ub_cnt > 0) {
582 item = bucket->ub_bucket[bucket->ub_cnt];
584 bucket->ub_bucket[bucket->ub_cnt] = NULL;
585 KASSERT(item != NULL,
586 ("bucket_drain: botched ptr, item is NULL"));
588 zone_free_item(zone, item, NULL, SKIP_DTOR, 0);
593 * Drains the per cpu caches for a zone.
595 * NOTE: This may only be called while the zone is being turn down, and not
596 * during normal operation. This is necessary in order that we do not have
597 * to migrate CPUs to drain the per-CPU caches.
600 * zone The zone to drain, must be unlocked.
606 cache_drain(uma_zone_t zone)
612 * XXX: It is safe to not lock the per-CPU caches, because we're
613 * tearing down the zone anyway. I.e., there will be no further use
614 * of the caches at this point.
616 * XXX: It would good to be able to assert that the zone is being
617 * torn down to prevent improper use of cache_drain().
619 * XXX: We lock the zone before passing into bucket_cache_drain() as
620 * it is used elsewhere. Should the tear-down path be made special
621 * there in some form?
624 cache = &zone->uz_cpu[cpu];
625 bucket_drain(zone, cache->uc_allocbucket);
626 bucket_drain(zone, cache->uc_freebucket);
627 if (cache->uc_allocbucket != NULL)
628 bucket_free(cache->uc_allocbucket);
629 if (cache->uc_freebucket != NULL)
630 bucket_free(cache->uc_freebucket);
631 cache->uc_allocbucket = cache->uc_freebucket = NULL;
634 bucket_cache_drain(zone);
639 * Drain the cached buckets from a zone. Expects a locked zone on entry.
642 bucket_cache_drain(uma_zone_t zone)
647 * Drain the bucket queues and free the buckets, we just keep two per
650 while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
651 LIST_REMOVE(bucket, ub_link);
653 bucket_drain(zone, bucket);
658 /* Now we do the free queue.. */
659 while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
660 LIST_REMOVE(bucket, ub_link);
666 * Frees pages from a keg back to the system. This is done on demand from
667 * the pageout daemon.
672 keg_drain(uma_keg_t keg)
674 struct slabhead freeslabs = { 0 };
682 * We don't want to take pages from statically allocated kegs at this
685 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
689 printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
692 if (keg->uk_free == 0)
695 slab = LIST_FIRST(&keg->uk_free_slab);
697 n = LIST_NEXT(slab, us_link);
699 /* We have no where to free these to */
700 if (slab->us_flags & UMA_SLAB_BOOT) {
705 LIST_REMOVE(slab, us_link);
706 keg->uk_pages -= keg->uk_ppera;
707 keg->uk_free -= keg->uk_ipers;
709 if (keg->uk_flags & UMA_ZONE_HASH)
710 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
712 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
719 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
720 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
722 for (i = 0; i < keg->uk_ipers; i++)
724 slab->us_data + (keg->uk_rsize * i),
726 flags = slab->us_flags;
729 if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
732 if (flags & UMA_SLAB_KMEM)
734 else if (flags & UMA_SLAB_KERNEL)
738 for (i = 0; i < keg->uk_ppera; i++)
739 vsetobj((vm_offset_t)mem + (i * PAGE_SIZE),
742 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
743 zone_free_item(keg->uk_slabzone, slab, NULL,
744 SKIP_NONE, ZFREE_STATFREE);
746 printf("%s: Returning %d bytes.\n",
747 keg->uk_name, UMA_SLAB_SIZE * keg->uk_ppera);
749 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags);
754 zone_drain_wait(uma_zone_t zone, int waitok)
758 * Set draining to interlock with zone_dtor() so we can release our
759 * locks as we go. Only dtor() should do a WAITOK call since it
760 * is the only call that knows the structure will still be available
764 while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
765 if (waitok == M_NOWAIT)
767 mtx_unlock(&uma_mtx);
768 msleep(zone, zone->uz_lock, PVM, "zonedrain", 1);
771 zone->uz_flags |= UMA_ZFLAG_DRAINING;
772 bucket_cache_drain(zone);
775 * The DRAINING flag protects us from being freed while
776 * we're running. Normally the uma_mtx would protect us but we
777 * must be able to release and acquire the right lock for each keg.
779 zone_foreach_keg(zone, &keg_drain);
781 zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
788 zone_drain(uma_zone_t zone)
791 zone_drain_wait(zone, M_NOWAIT);
795 * Allocate a new slab for a keg. This does not insert the slab onto a list.
798 * wait Shall we wait?
801 * The slab that was allocated or NULL if there is no memory and the
802 * caller specified M_NOWAIT.
805 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
807 uma_slabrefcnt_t slabref;
814 mtx_assert(&keg->uk_lock, MA_OWNED);
818 printf("slab_zalloc: Allocating a new slab for %s\n", keg->uk_name);
820 allocf = keg->uk_allocf;
823 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
824 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
832 * This reproduces the old vm_zone behavior of zero filling pages the
833 * first time they are added to a zone.
835 * Malloced items are zeroed in uma_zalloc.
838 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
843 /* zone is passed for legacy reasons. */
844 mem = allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE, &flags, wait);
846 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
847 zone_free_item(keg->uk_slabzone, slab, NULL,
848 SKIP_NONE, ZFREE_STATFREE);
853 /* Point the slab into the allocated memory */
854 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
855 slab = (uma_slab_t )(mem + keg->uk_pgoff);
857 if (keg->uk_flags & UMA_ZONE_VTOSLAB)
858 for (i = 0; i < keg->uk_ppera; i++)
859 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
863 slab->us_freecount = keg->uk_ipers;
864 slab->us_firstfree = 0;
865 slab->us_flags = flags;
867 if (keg->uk_flags & UMA_ZONE_REFCNT) {
868 slabref = (uma_slabrefcnt_t)slab;
869 for (i = 0; i < keg->uk_ipers; i++) {
870 slabref->us_freelist[i].us_refcnt = 0;
871 slabref->us_freelist[i].us_item = i+1;
874 for (i = 0; i < keg->uk_ipers; i++)
875 slab->us_freelist[i].us_item = i+1;
878 if (keg->uk_init != NULL) {
879 for (i = 0; i < keg->uk_ipers; i++)
880 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
881 keg->uk_size, wait) != 0)
883 if (i != keg->uk_ipers) {
884 if (keg->uk_fini != NULL) {
885 for (i--; i > -1; i--)
886 keg->uk_fini(slab->us_data +
890 if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
893 if (flags & UMA_SLAB_KMEM)
895 else if (flags & UMA_SLAB_KERNEL)
899 for (i = 0; i < keg->uk_ppera; i++)
900 vsetobj((vm_offset_t)mem +
901 (i * PAGE_SIZE), obj);
903 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
904 zone_free_item(keg->uk_slabzone, slab,
905 NULL, SKIP_NONE, ZFREE_STATFREE);
906 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera,
914 if (keg->uk_flags & UMA_ZONE_HASH)
915 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
917 keg->uk_pages += keg->uk_ppera;
918 keg->uk_free += keg->uk_ipers;
924 * This function is intended to be used early on in place of page_alloc() so
925 * that we may use the boot time page cache to satisfy allocations before
929 startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
933 int pages, check_pages;
935 keg = zone_first_keg(zone);
936 pages = howmany(bytes, PAGE_SIZE);
937 check_pages = pages - 1;
938 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
941 * Check our small startup cache to see if it has pages remaining.
943 mtx_lock(&uma_boot_pages_mtx);
945 /* First check if we have enough room. */
946 tmps = LIST_FIRST(&uma_boot_pages);
947 while (tmps != NULL && check_pages-- > 0)
948 tmps = LIST_NEXT(tmps, us_link);
951 * It's ok to lose tmps references. The last one will
952 * have tmps->us_data pointing to the start address of
953 * "pages" contiguous pages of memory.
955 while (pages-- > 0) {
956 tmps = LIST_FIRST(&uma_boot_pages);
957 LIST_REMOVE(tmps, us_link);
959 mtx_unlock(&uma_boot_pages_mtx);
960 *pflag = tmps->us_flags;
961 return (tmps->us_data);
963 mtx_unlock(&uma_boot_pages_mtx);
965 panic("UMA: Increase vm.boot_pages");
967 * Now that we've booted reset these users to their real allocator.
969 #ifdef UMA_MD_SMALL_ALLOC
970 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
972 keg->uk_allocf = page_alloc;
974 return keg->uk_allocf(zone, bytes, pflag, wait);
978 * Allocates a number of pages from the system
981 * bytes The number of bytes requested
982 * wait Shall we wait?
985 * A pointer to the alloced memory or possibly
986 * NULL if M_NOWAIT is set.
989 page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
991 void *p; /* Returned page */
993 *pflag = UMA_SLAB_KMEM;
994 p = (void *) kmem_malloc(kmem_map, bytes, wait);
1000 * Allocates a number of pages from within an object
1003 * bytes The number of bytes requested
1004 * wait Shall we wait?
1007 * A pointer to the alloced memory or possibly
1008 * NULL if M_NOWAIT is set.
1011 obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
1014 vm_offset_t retkva, zkva;
1016 int pages, startpages;
1019 keg = zone_first_keg(zone);
1020 object = keg->uk_obj;
1024 * This looks a little weird since we're getting one page at a time.
1026 VM_OBJECT_LOCK(object);
1027 p = TAILQ_LAST(&object->memq, pglist);
1028 pages = p != NULL ? p->pindex + 1 : 0;
1030 zkva = keg->uk_kva + pages * PAGE_SIZE;
1031 for (; bytes > 0; bytes -= PAGE_SIZE) {
1032 p = vm_page_alloc(object, pages,
1033 VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED);
1035 if (pages != startpages)
1036 pmap_qremove(retkva, pages - startpages);
1037 while (pages != startpages) {
1039 p = TAILQ_LAST(&object->memq, pglist);
1040 vm_page_lock_queues();
1041 vm_page_unwire(p, 0);
1043 vm_page_unlock_queues();
1048 pmap_qenter(zkva, &p, 1);
1055 VM_OBJECT_UNLOCK(object);
1056 *flags = UMA_SLAB_PRIV;
1058 return ((void *)retkva);
1062 * Frees a number of pages to the system
1065 * mem A pointer to the memory to be freed
1066 * size The size of the memory being freed
1067 * flags The original p->us_flags field
1073 page_free(void *mem, int size, u_int8_t flags)
1077 if (flags & UMA_SLAB_KMEM)
1079 else if (flags & UMA_SLAB_KERNEL)
1082 panic("UMA: page_free used with invalid flags %d", flags);
1084 kmem_free(map, (vm_offset_t)mem, size);
1088 * Zero fill initializer
1090 * Arguments/Returns follow uma_init specifications
1093 zero_init(void *mem, int size, int flags)
1100 * Finish creating a small uma keg. This calculates ipers, and the keg size.
1103 * keg The zone we should initialize
1109 keg_small_init(uma_keg_t keg)
1116 KASSERT(keg != NULL, ("Keg is null in keg_small_init"));
1117 rsize = keg->uk_size;
1119 if (rsize < UMA_SMALLEST_UNIT)
1120 rsize = UMA_SMALLEST_UNIT;
1121 if (rsize & keg->uk_align)
1122 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1124 keg->uk_rsize = rsize;
1127 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1128 rsize += UMA_FRITMREF_SZ; /* linkage & refcnt */
1129 shsize = sizeof(struct uma_slab_refcnt);
1131 rsize += UMA_FRITM_SZ; /* Account for linkage */
1132 shsize = sizeof(struct uma_slab);
1135 keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize;
1136 KASSERT(keg->uk_ipers != 0, ("keg_small_init: ipers is 0"));
1137 memused = keg->uk_ipers * rsize + shsize;
1138 wastedspace = UMA_SLAB_SIZE - memused;
1141 * We can't do OFFPAGE if we're internal or if we've been
1142 * asked to not go to the VM for buckets. If we do this we
1143 * may end up going to the VM (kmem_map) for slabs which we
1144 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1145 * result of UMA_ZONE_VM, which clearly forbids it.
1147 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1148 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1151 if ((wastedspace >= UMA_MAX_WASTE) &&
1152 (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) {
1153 keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize;
1154 KASSERT(keg->uk_ipers <= 255,
1155 ("keg_small_init: keg->uk_ipers too high!"));
1157 printf("UMA decided we need offpage slab headers for "
1158 "keg: %s, calculated wastedspace = %d, "
1159 "maximum wasted space allowed = %d, "
1160 "calculated ipers = %d, "
1161 "new wasted space = %d\n", keg->uk_name, wastedspace,
1162 UMA_MAX_WASTE, keg->uk_ipers,
1163 UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize);
1165 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1166 if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1167 keg->uk_flags |= UMA_ZONE_HASH;
1172 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do
1173 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1177 * keg The keg we should initialize
1183 keg_large_init(uma_keg_t keg)
1187 KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1188 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1189 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1191 pages = keg->uk_size / UMA_SLAB_SIZE;
1193 /* Account for remainder */
1194 if ((pages * UMA_SLAB_SIZE) < keg->uk_size)
1197 keg->uk_ppera = pages;
1199 keg->uk_rsize = keg->uk_size;
1201 /* We can't do OFFPAGE if we're internal, bail out here. */
1202 if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1205 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1206 if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1207 keg->uk_flags |= UMA_ZONE_HASH;
1211 keg_cachespread_init(uma_keg_t keg)
1218 alignsize = keg->uk_align + 1;
1219 rsize = keg->uk_size;
1221 * We want one item to start on every align boundary in a page. To
1222 * do this we will span pages. We will also extend the item by the
1223 * size of align if it is an even multiple of align. Otherwise, it
1224 * would fall on the same boundary every time.
1226 if (rsize & keg->uk_align)
1227 rsize = (rsize & ~keg->uk_align) + alignsize;
1228 if ((rsize & alignsize) == 0)
1230 trailer = rsize - keg->uk_size;
1231 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1232 pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1233 keg->uk_rsize = rsize;
1234 keg->uk_ppera = pages;
1235 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1236 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1237 KASSERT(keg->uk_ipers <= uma_max_ipers,
1238 ("keg_small_init: keg->uk_ipers too high(%d) increase max_ipers",
1243 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1244 * the keg onto the global keg list.
1246 * Arguments/Returns follow uma_ctor specifications
1247 * udata Actually uma_kctor_args
1250 keg_ctor(void *mem, int size, void *udata, int flags)
1252 struct uma_kctor_args *arg = udata;
1253 uma_keg_t keg = mem;
1257 keg->uk_size = arg->size;
1258 keg->uk_init = arg->uminit;
1259 keg->uk_fini = arg->fini;
1260 keg->uk_align = arg->align;
1263 keg->uk_flags = arg->flags;
1264 keg->uk_allocf = page_alloc;
1265 keg->uk_freef = page_free;
1266 keg->uk_recurse = 0;
1267 keg->uk_slabzone = NULL;
1270 * The master zone is passed to us at keg-creation time.
1273 keg->uk_name = zone->uz_name;
1275 if (arg->flags & UMA_ZONE_VM)
1276 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1278 if (arg->flags & UMA_ZONE_ZINIT)
1279 keg->uk_init = zero_init;
1281 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1282 keg->uk_flags |= UMA_ZONE_VTOSLAB;
1285 * The +UMA_FRITM_SZ added to uk_size is to account for the
1286 * linkage that is added to the size in keg_small_init(). If
1287 * we don't account for this here then we may end up in
1288 * keg_small_init() with a calculated 'ipers' of 0.
1290 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1291 if (keg->uk_flags & UMA_ZONE_CACHESPREAD)
1292 keg_cachespread_init(keg);
1293 else if ((keg->uk_size+UMA_FRITMREF_SZ) >
1294 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)))
1295 keg_large_init(keg);
1297 keg_small_init(keg);
1299 if (keg->uk_flags & UMA_ZONE_CACHESPREAD)
1300 keg_cachespread_init(keg);
1301 else if ((keg->uk_size+UMA_FRITM_SZ) >
1302 (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1303 keg_large_init(keg);
1305 keg_small_init(keg);
1308 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1309 if (keg->uk_flags & UMA_ZONE_REFCNT)
1310 keg->uk_slabzone = slabrefzone;
1312 keg->uk_slabzone = slabzone;
1316 * If we haven't booted yet we need allocations to go through the
1317 * startup cache until the vm is ready.
1319 if (keg->uk_ppera == 1) {
1320 #ifdef UMA_MD_SMALL_ALLOC
1321 keg->uk_allocf = uma_small_alloc;
1322 keg->uk_freef = uma_small_free;
1325 keg->uk_allocf = startup_alloc;
1326 } else if (booted == 0 && (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1327 keg->uk_allocf = startup_alloc;
1330 * Initialize keg's lock (shared among zones).
1332 if (arg->flags & UMA_ZONE_MTXCLASS)
1333 KEG_LOCK_INIT(keg, 1);
1335 KEG_LOCK_INIT(keg, 0);
1338 * If we're putting the slab header in the actual page we need to
1339 * figure out where in each page it goes. This calculates a right
1340 * justified offset into the memory on an ALIGN_PTR boundary.
1342 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1345 /* Size of the slab struct and free list */
1346 if (keg->uk_flags & UMA_ZONE_REFCNT)
1347 totsize = sizeof(struct uma_slab_refcnt) +
1348 keg->uk_ipers * UMA_FRITMREF_SZ;
1350 totsize = sizeof(struct uma_slab) +
1351 keg->uk_ipers * UMA_FRITM_SZ;
1353 if (totsize & UMA_ALIGN_PTR)
1354 totsize = (totsize & ~UMA_ALIGN_PTR) +
1355 (UMA_ALIGN_PTR + 1);
1356 keg->uk_pgoff = (UMA_SLAB_SIZE * keg->uk_ppera) - totsize;
1358 if (keg->uk_flags & UMA_ZONE_REFCNT)
1359 totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt)
1360 + keg->uk_ipers * UMA_FRITMREF_SZ;
1362 totsize = keg->uk_pgoff + sizeof(struct uma_slab)
1363 + keg->uk_ipers * UMA_FRITM_SZ;
1366 * The only way the following is possible is if with our
1367 * UMA_ALIGN_PTR adjustments we are now bigger than
1368 * UMA_SLAB_SIZE. I haven't checked whether this is
1369 * mathematically possible for all cases, so we make
1372 if (totsize > UMA_SLAB_SIZE * keg->uk_ppera) {
1373 printf("zone %s ipers %d rsize %d size %d\n",
1374 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1376 panic("UMA slab won't fit.");
1380 if (keg->uk_flags & UMA_ZONE_HASH)
1381 hash_alloc(&keg->uk_hash);
1384 printf("UMA: %s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n",
1385 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1386 keg->uk_ipers, keg->uk_ppera,
1387 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1390 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1393 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1394 mtx_unlock(&uma_mtx);
1399 * Zone header ctor. This initializes all fields, locks, etc.
1401 * Arguments/Returns follow uma_ctor specifications
1402 * udata Actually uma_zctor_args
1405 zone_ctor(void *mem, int size, void *udata, int flags)
1407 struct uma_zctor_args *arg = udata;
1408 uma_zone_t zone = mem;
1413 zone->uz_name = arg->name;
1414 zone->uz_ctor = arg->ctor;
1415 zone->uz_dtor = arg->dtor;
1416 zone->uz_slab = zone_fetch_slab;
1417 zone->uz_init = NULL;
1418 zone->uz_fini = NULL;
1419 zone->uz_allocs = 0;
1422 zone->uz_fills = zone->uz_count = 0;
1426 if (arg->flags & UMA_ZONE_SECONDARY) {
1427 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1428 zone->uz_init = arg->uminit;
1429 zone->uz_fini = arg->fini;
1430 zone->uz_lock = &keg->uk_lock;
1431 zone->uz_flags |= UMA_ZONE_SECONDARY;
1434 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1435 if (LIST_NEXT(z, uz_link) == NULL) {
1436 LIST_INSERT_AFTER(z, zone, uz_link);
1441 mtx_unlock(&uma_mtx);
1442 } else if (keg == NULL) {
1443 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1444 arg->align, arg->flags)) == NULL)
1447 struct uma_kctor_args karg;
1450 /* We should only be here from uma_startup() */
1451 karg.size = arg->size;
1452 karg.uminit = arg->uminit;
1453 karg.fini = arg->fini;
1454 karg.align = arg->align;
1455 karg.flags = arg->flags;
1457 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1463 * Link in the first keg.
1465 zone->uz_klink.kl_keg = keg;
1466 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1467 zone->uz_lock = &keg->uk_lock;
1468 zone->uz_size = keg->uk_size;
1469 zone->uz_flags |= (keg->uk_flags &
1470 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1473 * Some internal zones don't have room allocated for the per cpu
1474 * caches. If we're internal, bail out here.
1476 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1477 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1478 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1482 if (keg->uk_flags & UMA_ZONE_MAXBUCKET)
1483 zone->uz_count = BUCKET_MAX;
1484 else if (keg->uk_ipers <= BUCKET_MAX)
1485 zone->uz_count = keg->uk_ipers;
1487 zone->uz_count = BUCKET_MAX;
1492 * Keg header dtor. This frees all data, destroys locks, frees the hash
1493 * table and removes the keg from the global list.
1495 * Arguments/Returns follow uma_dtor specifications
1499 keg_dtor(void *arg, int size, void *udata)
1503 keg = (uma_keg_t)arg;
1505 if (keg->uk_free != 0) {
1506 printf("Freed UMA keg was not empty (%d items). "
1507 " Lost %d pages of memory.\n",
1508 keg->uk_free, keg->uk_pages);
1512 hash_free(&keg->uk_hash);
1520 * Arguments/Returns follow uma_dtor specifications
1524 zone_dtor(void *arg, int size, void *udata)
1530 zone = (uma_zone_t)arg;
1531 keg = zone_first_keg(zone);
1533 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1537 LIST_REMOVE(zone, uz_link);
1538 mtx_unlock(&uma_mtx);
1540 * XXX there are some races here where
1541 * the zone can be drained but zone lock
1542 * released and then refilled before we
1543 * remove it... we dont care for now
1545 zone_drain_wait(zone, M_WAITOK);
1547 * Unlink all of our kegs.
1549 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1550 klink->kl_keg = NULL;
1551 LIST_REMOVE(klink, kl_link);
1552 if (klink == &zone->uz_klink)
1554 free(klink, M_TEMP);
1557 * We only destroy kegs from non secondary zones.
1559 if ((zone->uz_flags & UMA_ZONE_SECONDARY) == 0) {
1561 LIST_REMOVE(keg, uk_link);
1562 mtx_unlock(&uma_mtx);
1563 zone_free_item(kegs, keg, NULL, SKIP_NONE,
1569 * Traverses every zone in the system and calls a callback
1572 * zfunc A pointer to a function which accepts a zone
1579 zone_foreach(void (*zfunc)(uma_zone_t))
1585 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1586 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1589 mtx_unlock(&uma_mtx);
1592 /* Public functions */
1595 uma_startup(void *bootmem, int boot_pages)
1597 struct uma_zctor_args args;
1600 u_int objsize, totsize, wsize;
1604 printf("Creating uma keg headers zone and keg.\n");
1606 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1609 * Figure out the maximum number of items-per-slab we'll have if
1610 * we're using the OFFPAGE slab header to track free items, given
1611 * all possible object sizes and the maximum desired wastage
1614 * We iterate until we find an object size for
1615 * which the calculated wastage in keg_small_init() will be
1616 * enough to warrant OFFPAGE. Since wastedspace versus objsize
1617 * is an overall increasing see-saw function, we find the smallest
1618 * objsize such that the wastage is always acceptable for objects
1619 * with that objsize or smaller. Since a smaller objsize always
1620 * generates a larger possible uma_max_ipers, we use this computed
1621 * objsize to calculate the largest ipers possible. Since the
1622 * ipers calculated for OFFPAGE slab headers is always larger than
1623 * the ipers initially calculated in keg_small_init(), we use
1624 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to
1625 * obtain the maximum ipers possible for offpage slab headers.
1627 * It should be noted that ipers versus objsize is an inversly
1628 * proportional function which drops off rather quickly so as
1629 * long as our UMA_MAX_WASTE is such that the objsize we calculate
1630 * falls into the portion of the inverse relation AFTER the steep
1631 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386).
1633 * Note that we have 8-bits (1 byte) to use as a freelist index
1634 * inside the actual slab header itself and this is enough to
1635 * accomodate us. In the worst case, a UMA_SMALLEST_UNIT sized
1636 * object with offpage slab header would have ipers =
1637 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is
1638 * 1 greater than what our byte-integer freelist index can
1639 * accomodate, but we know that this situation never occurs as
1640 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate
1641 * that we need to go to offpage slab headers. Or, if we do,
1642 * then we trap that condition below and panic in the INVARIANTS case.
1644 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE;
1646 objsize = UMA_SMALLEST_UNIT;
1647 while (totsize >= wsize) {
1648 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) /
1649 (objsize + UMA_FRITM_SZ);
1650 totsize *= (UMA_FRITM_SZ + objsize);
1653 if (objsize > UMA_SMALLEST_UNIT)
1655 uma_max_ipers = MAX(UMA_SLAB_SIZE / objsize, 64);
1657 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE;
1659 objsize = UMA_SMALLEST_UNIT;
1660 while (totsize >= wsize) {
1661 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) /
1662 (objsize + UMA_FRITMREF_SZ);
1663 totsize *= (UMA_FRITMREF_SZ + objsize);
1666 if (objsize > UMA_SMALLEST_UNIT)
1668 uma_max_ipers_ref = MAX(UMA_SLAB_SIZE / objsize, 64);
1670 KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255),
1671 ("uma_startup: calculated uma_max_ipers values too large!"));
1674 printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers);
1675 printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n",
1679 /* "manually" create the initial zone */
1680 args.name = "UMA Kegs";
1681 args.size = sizeof(struct uma_keg);
1682 args.ctor = keg_ctor;
1683 args.dtor = keg_dtor;
1684 args.uminit = zero_init;
1686 args.keg = &masterkeg;
1687 args.align = 32 - 1;
1688 args.flags = UMA_ZFLAG_INTERNAL;
1689 /* The initial zone has no Per cpu queues so it's smaller */
1690 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1693 printf("Filling boot free list.\n");
1695 for (i = 0; i < boot_pages; i++) {
1696 slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE));
1697 slab->us_data = (u_int8_t *)slab;
1698 slab->us_flags = UMA_SLAB_BOOT;
1699 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1701 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1704 printf("Creating uma zone headers zone and keg.\n");
1706 args.name = "UMA Zones";
1707 args.size = sizeof(struct uma_zone) +
1708 (sizeof(struct uma_cache) * (mp_maxid + 1));
1709 args.ctor = zone_ctor;
1710 args.dtor = zone_dtor;
1711 args.uminit = zero_init;
1714 args.align = 32 - 1;
1715 args.flags = UMA_ZFLAG_INTERNAL;
1716 /* The initial zone has no Per cpu queues so it's smaller */
1717 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1720 printf("Initializing pcpu cache locks.\n");
1723 printf("Creating slab and hash zones.\n");
1727 * This is the max number of free list items we'll have with
1730 slabsize = uma_max_ipers * UMA_FRITM_SZ;
1731 slabsize += sizeof(struct uma_slab);
1733 /* Now make a zone for slab headers */
1734 slabzone = uma_zcreate("UMA Slabs",
1736 NULL, NULL, NULL, NULL,
1737 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1740 * We also create a zone for the bigger slabs with reference
1741 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1743 slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ;
1744 slabsize += sizeof(struct uma_slab_refcnt);
1745 slabrefzone = uma_zcreate("UMA RCntSlabs",
1747 NULL, NULL, NULL, NULL,
1749 UMA_ZFLAG_INTERNAL);
1751 hashzone = uma_zcreate("UMA Hash",
1752 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1753 NULL, NULL, NULL, NULL,
1754 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1758 #if defined(UMA_MD_SMALL_ALLOC) && !defined(UMA_MD_SMALL_ALLOC_NEEDS_VM)
1763 printf("UMA startup complete.\n");
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, u_int32_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(char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1823 uma_init uminit, uma_fini fini, int align, u_int32_t flags)
1826 struct uma_zctor_args args;
1828 /* This stuff is essential for the zone ctor */
1833 args.uminit = uminit;
1839 return (zone_alloc_item(zones, &args, M_WAITOK));
1844 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1845 uma_init zinit, uma_fini zfini, uma_zone_t master)
1847 struct uma_zctor_args args;
1850 keg = zone_first_keg(master);
1852 args.size = keg->uk_size;
1855 args.uminit = zinit;
1857 args.align = keg->uk_align;
1858 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1861 /* XXX Attaches only one keg of potentially many. */
1862 return (zone_alloc_item(zones, &args, M_WAITOK));
1866 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1870 mtx_lock_flags(b->uz_lock, MTX_DUPOK);
1873 mtx_lock_flags(a->uz_lock, MTX_DUPOK);
1878 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1886 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1893 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1895 zone_lock_pair(zone, master);
1897 * zone must use vtoslab() to resolve objects and must already be
1900 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1901 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1906 * The new master must also use vtoslab().
1908 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
1913 * Both must either be refcnt, or not be refcnt.
1915 if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
1916 (master->uz_flags & UMA_ZONE_REFCNT)) {
1921 * The underlying object must be the same size. rsize
1924 if (master->uz_size != zone->uz_size) {
1929 * Put it at the end of the list.
1931 klink->kl_keg = zone_first_keg(master);
1932 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
1933 if (LIST_NEXT(kl, kl_link) == NULL) {
1934 LIST_INSERT_AFTER(kl, klink, kl_link);
1939 zone->uz_flags |= UMA_ZFLAG_MULTI;
1940 zone->uz_slab = zone_fetch_slab_multi;
1943 zone_unlock_pair(zone, master);
1945 free(klink, M_TEMP);
1953 uma_zdestroy(uma_zone_t zone)
1956 zone_free_item(zones, zone, NULL, SKIP_NONE, ZFREE_STATFREE);
1961 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1965 uma_bucket_t bucket;
1968 /* This is the fast path allocation */
1969 #ifdef UMA_DEBUG_ALLOC_1
1970 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1972 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1973 zone->uz_name, flags);
1975 if (flags & M_WAITOK) {
1976 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1977 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
1981 * If possible, allocate from the per-CPU cache. There are two
1982 * requirements for safe access to the per-CPU cache: (1) the thread
1983 * accessing the cache must not be preempted or yield during access,
1984 * and (2) the thread must not migrate CPUs without switching which
1985 * cache it accesses. We rely on a critical section to prevent
1986 * preemption and migration. We release the critical section in
1987 * order to acquire the zone mutex if we are unable to allocate from
1988 * the current cache; when we re-acquire the critical section, we
1989 * must detect and handle migration if it has occurred.
1994 cache = &zone->uz_cpu[cpu];
1997 bucket = cache->uc_allocbucket;
2000 if (bucket->ub_cnt > 0) {
2002 item = bucket->ub_bucket[bucket->ub_cnt];
2004 bucket->ub_bucket[bucket->ub_cnt] = NULL;
2006 KASSERT(item != NULL,
2007 ("uma_zalloc: Bucket pointer mangled."));
2012 uma_dbg_alloc(zone, NULL, item);
2015 if (zone->uz_ctor != NULL) {
2016 if (zone->uz_ctor(item, zone->uz_size,
2017 udata, flags) != 0) {
2018 zone_free_item(zone, item, udata,
2019 SKIP_DTOR, ZFREE_STATFAIL |
2025 bzero(item, zone->uz_size);
2027 } else if (cache->uc_freebucket) {
2029 * We have run out of items in our allocbucket.
2030 * See if we can switch with our free bucket.
2032 if (cache->uc_freebucket->ub_cnt > 0) {
2033 #ifdef UMA_DEBUG_ALLOC
2034 printf("uma_zalloc: Swapping empty with"
2037 bucket = cache->uc_freebucket;
2038 cache->uc_freebucket = cache->uc_allocbucket;
2039 cache->uc_allocbucket = bucket;
2046 * Attempt to retrieve the item from the per-CPU cache has failed, so
2047 * we must go back to the zone. This requires the zone lock, so we
2048 * must drop the critical section, then re-acquire it when we go back
2049 * to the cache. Since the critical section is released, we may be
2050 * preempted or migrate. As such, make sure not to maintain any
2051 * thread-local state specific to the cache from prior to releasing
2052 * the critical section.
2058 cache = &zone->uz_cpu[cpu];
2059 bucket = cache->uc_allocbucket;
2060 if (bucket != NULL) {
2061 if (bucket->ub_cnt > 0) {
2065 bucket = cache->uc_freebucket;
2066 if (bucket != NULL && bucket->ub_cnt > 0) {
2072 /* Since we have locked the zone we may as well send back our stats */
2073 zone->uz_allocs += cache->uc_allocs;
2074 cache->uc_allocs = 0;
2075 zone->uz_frees += cache->uc_frees;
2076 cache->uc_frees = 0;
2078 /* Our old one is now a free bucket */
2079 if (cache->uc_allocbucket) {
2080 KASSERT(cache->uc_allocbucket->ub_cnt == 0,
2081 ("uma_zalloc_arg: Freeing a non free bucket."));
2082 LIST_INSERT_HEAD(&zone->uz_free_bucket,
2083 cache->uc_allocbucket, ub_link);
2084 cache->uc_allocbucket = NULL;
2087 /* Check the free list for a new alloc bucket */
2088 if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
2089 KASSERT(bucket->ub_cnt != 0,
2090 ("uma_zalloc_arg: Returning an empty bucket."));
2092 LIST_REMOVE(bucket, ub_link);
2093 cache->uc_allocbucket = bucket;
2097 /* We are no longer associated with this CPU. */
2100 /* Bump up our uz_count so we get here less */
2101 if (zone->uz_count < BUCKET_MAX)
2105 * Now lets just fill a bucket and put it on the free list. If that
2106 * works we'll restart the allocation from the begining.
2108 if (zone_alloc_bucket(zone, flags)) {
2110 goto zalloc_restart;
2114 * We may not be able to get a bucket so return an actual item.
2117 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2120 item = zone_alloc_item(zone, udata, flags);
2125 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2129 mtx_assert(&keg->uk_lock, MA_OWNED);
2134 * Find a slab with some space. Prefer slabs that are partially
2135 * used over those that are totally full. This helps to reduce
2138 if (keg->uk_free != 0) {
2139 if (!LIST_EMPTY(&keg->uk_part_slab)) {
2140 slab = LIST_FIRST(&keg->uk_part_slab);
2142 slab = LIST_FIRST(&keg->uk_free_slab);
2143 LIST_REMOVE(slab, us_link);
2144 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2147 MPASS(slab->us_keg == keg);
2152 * M_NOVM means don't ask at all!
2157 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2158 keg->uk_flags |= UMA_ZFLAG_FULL;
2160 * If this is not a multi-zone, set the FULL bit.
2161 * Otherwise slab_multi() takes care of it.
2163 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0)
2164 zone->uz_flags |= UMA_ZFLAG_FULL;
2165 if (flags & M_NOWAIT)
2167 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2171 slab = keg_alloc_slab(keg, zone, flags);
2174 * If we got a slab here it's safe to mark it partially used
2175 * and return. We assume that the caller is going to remove
2176 * at least one item.
2179 MPASS(slab->us_keg == keg);
2180 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2184 * We might not have been able to get a slab but another cpu
2185 * could have while we were unlocked. Check again before we
2194 zone_relock(uma_zone_t zone, uma_keg_t keg)
2196 if (zone->uz_lock != &keg->uk_lock) {
2203 keg_relock(uma_keg_t keg, uma_zone_t zone)
2205 if (zone->uz_lock != &keg->uk_lock) {
2212 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2217 keg = zone_first_keg(zone);
2219 * This is to prevent us from recursively trying to allocate
2220 * buckets. The problem is that if an allocation forces us to
2221 * grab a new bucket we will call page_alloc, which will go off
2222 * and cause the vm to allocate vm_map_entries. If we need new
2223 * buckets there too we will recurse in kmem_alloc and bad
2224 * things happen. So instead we return a NULL bucket, and make
2225 * the code that allocates buckets smart enough to deal with it
2227 if (keg->uk_flags & UMA_ZFLAG_BUCKET && keg->uk_recurse != 0)
2231 slab = keg_fetch_slab(keg, zone, flags);
2234 if (flags & (M_NOWAIT | M_NOVM))
2241 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns
2242 * with the keg locked. Caller must call zone_relock() afterwards if the
2243 * zone lock is required. On NULL the zone lock is held.
2245 * The last pointer is used to seed the search. It is not required.
2248 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2258 * Don't wait on the first pass. This will skip limit tests
2259 * as well. We don't want to block if we can find a provider
2262 flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2264 * Use the last slab allocated as a hint for where to start
2268 slab = keg_fetch_slab(last, zone, flags);
2271 zone_relock(zone, last);
2275 * Loop until we have a slab incase of transient failures
2276 * while M_WAITOK is specified. I'm not sure this is 100%
2277 * required but we've done it for so long now.
2283 * Search the available kegs for slabs. Be careful to hold the
2284 * correct lock while calling into the keg layer.
2286 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2287 keg = klink->kl_keg;
2288 keg_relock(keg, zone);
2289 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2290 slab = keg_fetch_slab(keg, zone, flags);
2294 if (keg->uk_flags & UMA_ZFLAG_FULL)
2298 zone_relock(zone, keg);
2300 if (rflags & (M_NOWAIT | M_NOVM))
2304 * All kegs are full. XXX We can't atomically check all kegs
2305 * and sleep so just sleep for a short period and retry.
2307 if (full && !empty) {
2308 zone->uz_flags |= UMA_ZFLAG_FULL;
2309 msleep(zone, zone->uz_lock, PVM, "zonelimit", hz/100);
2310 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2318 slab_alloc_item(uma_zone_t zone, uma_slab_t slab)
2321 uma_slabrefcnt_t slabref;
2326 mtx_assert(&keg->uk_lock, MA_OWNED);
2328 freei = slab->us_firstfree;
2329 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2330 slabref = (uma_slabrefcnt_t)slab;
2331 slab->us_firstfree = slabref->us_freelist[freei].us_item;
2333 slab->us_firstfree = slab->us_freelist[freei].us_item;
2335 item = slab->us_data + (keg->uk_rsize * freei);
2337 slab->us_freecount--;
2340 uma_dbg_alloc(zone, slab, item);
2342 /* Move this slab to the full list */
2343 if (slab->us_freecount == 0) {
2344 LIST_REMOVE(slab, us_link);
2345 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2352 zone_alloc_bucket(uma_zone_t zone, int flags)
2354 uma_bucket_t bucket;
2358 int max, origflags = flags;
2361 * Try this zone's free list first so we don't allocate extra buckets.
2363 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2364 KASSERT(bucket->ub_cnt == 0,
2365 ("zone_alloc_bucket: Bucket on free list is not empty."));
2366 LIST_REMOVE(bucket, ub_link);
2370 bflags = (flags & ~M_ZERO);
2371 if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2375 bucket = bucket_alloc(zone->uz_count, bflags);
2379 if (bucket == NULL) {
2385 * This code is here to limit the number of simultaneous bucket fills
2386 * for any given zone to the number of per cpu caches in this zone. This
2387 * is done so that we don't allocate more memory than we really need.
2389 if (zone->uz_fills >= mp_ncpus)
2395 max = MIN(bucket->ub_entries, zone->uz_count);
2396 /* Try to keep the buckets totally full */
2397 saved = bucket->ub_cnt;
2400 while (bucket->ub_cnt < max &&
2401 (slab = zone->uz_slab(zone, keg, flags)) != NULL) {
2403 while (slab->us_freecount && bucket->ub_cnt < max) {
2404 bucket->ub_bucket[bucket->ub_cnt++] =
2405 slab_alloc_item(zone, slab);
2408 /* Don't block on the next fill */
2412 zone_relock(zone, keg);
2415 * We unlock here because we need to call the zone's init.
2416 * It should be safe to unlock because the slab dealt with
2417 * above is already on the appropriate list within the keg
2418 * and the bucket we filled is not yet on any list, so we
2421 if (zone->uz_init != NULL) {
2425 for (i = saved; i < bucket->ub_cnt; i++)
2426 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2430 * If we couldn't initialize the whole bucket, put the
2431 * rest back onto the freelist.
2433 if (i != bucket->ub_cnt) {
2436 for (j = i; j < bucket->ub_cnt; j++) {
2437 zone_free_item(zone, bucket->ub_bucket[j],
2438 NULL, SKIP_FINI, 0);
2440 bucket->ub_bucket[j] = NULL;
2449 if (bucket->ub_cnt != 0) {
2450 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2457 bucket_free(bucket);
2462 * Allocates an item for an internal zone
2465 * zone The zone to alloc for.
2466 * udata The data to be passed to the constructor.
2467 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2470 * NULL if there is no memory and M_NOWAIT is set
2471 * An item if successful
2475 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2482 #ifdef UMA_DEBUG_ALLOC
2483 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2487 slab = zone->uz_slab(zone, NULL, flags);
2494 item = slab_alloc_item(zone, slab);
2496 zone_relock(zone, slab->us_keg);
2501 * We have to call both the zone's init (not the keg's init)
2502 * and the zone's ctor. This is because the item is going from
2503 * a keg slab directly to the user, and the user is expecting it
2504 * to be both zone-init'd as well as zone-ctor'd.
2506 if (zone->uz_init != NULL) {
2507 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2508 zone_free_item(zone, item, udata, SKIP_FINI,
2509 ZFREE_STATFAIL | ZFREE_STATFREE);
2513 if (zone->uz_ctor != NULL) {
2514 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2515 zone_free_item(zone, item, udata, SKIP_DTOR,
2516 ZFREE_STATFAIL | ZFREE_STATFREE);
2521 bzero(item, zone->uz_size);
2528 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2531 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). */
2546 zone->uz_dtor(item, zone->uz_size, udata);
2550 if (zone->uz_flags & UMA_ZONE_MALLOC)
2551 uma_dbg_free(zone, udata, item);
2553 uma_dbg_free(zone, NULL, item);
2557 * The race here is acceptable. If we miss it we'll just have to wait
2558 * a little longer for the limits to be reset.
2560 if (zone->uz_flags & UMA_ZFLAG_FULL)
2561 goto zfree_internal;
2564 * If possible, free to the per-CPU cache. There are two
2565 * requirements for safe access to the per-CPU cache: (1) the thread
2566 * accessing the cache must not be preempted or yield during access,
2567 * and (2) the thread must not migrate CPUs without switching which
2568 * cache it accesses. We rely on a critical section to prevent
2569 * preemption and migration. We release the critical section in
2570 * order to acquire the zone mutex if we are unable to free to the
2571 * current cache; when we re-acquire the critical section, we must
2572 * detect and handle migration if it has occurred.
2577 cache = &zone->uz_cpu[cpu];
2580 bucket = cache->uc_freebucket;
2584 * Do we have room in our bucket? It is OK for this uz count
2585 * check to be slightly out of sync.
2588 if (bucket->ub_cnt < bucket->ub_entries) {
2589 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2590 ("uma_zfree: Freeing to non free bucket index."));
2591 bucket->ub_bucket[bucket->ub_cnt] = item;
2596 } else if (cache->uc_allocbucket) {
2597 #ifdef UMA_DEBUG_ALLOC
2598 printf("uma_zfree: Swapping buckets.\n");
2601 * We have run out of space in our freebucket.
2602 * See if we can switch with our alloc bucket.
2604 if (cache->uc_allocbucket->ub_cnt <
2605 cache->uc_freebucket->ub_cnt) {
2606 bucket = cache->uc_freebucket;
2607 cache->uc_freebucket = cache->uc_allocbucket;
2608 cache->uc_allocbucket = bucket;
2614 * We can get here for two reasons:
2616 * 1) The buckets are NULL
2617 * 2) The alloc and free buckets are both somewhat full.
2619 * We must go back the zone, which requires acquiring the zone lock,
2620 * which in turn means we must release and re-acquire the critical
2621 * section. Since the critical section is released, we may be
2622 * preempted or migrate. As such, make sure not to maintain any
2623 * thread-local state specific to the cache from prior to releasing
2624 * the critical section.
2630 cache = &zone->uz_cpu[cpu];
2631 if (cache->uc_freebucket != NULL) {
2632 if (cache->uc_freebucket->ub_cnt <
2633 cache->uc_freebucket->ub_entries) {
2637 if (cache->uc_allocbucket != NULL &&
2638 (cache->uc_allocbucket->ub_cnt <
2639 cache->uc_freebucket->ub_cnt)) {
2645 /* Since we have locked the zone we may as well send back our stats */
2646 zone->uz_allocs += cache->uc_allocs;
2647 cache->uc_allocs = 0;
2648 zone->uz_frees += cache->uc_frees;
2649 cache->uc_frees = 0;
2651 bucket = cache->uc_freebucket;
2652 cache->uc_freebucket = NULL;
2654 /* Can we throw this on the zone full list? */
2655 if (bucket != NULL) {
2656 #ifdef UMA_DEBUG_ALLOC
2657 printf("uma_zfree: Putting old bucket on the free list.\n");
2659 /* ub_cnt is pointing to the last free item */
2660 KASSERT(bucket->ub_cnt != 0,
2661 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2662 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2665 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2666 LIST_REMOVE(bucket, ub_link);
2668 cache->uc_freebucket = bucket;
2671 /* We are no longer associated with this CPU. */
2674 /* And the zone.. */
2677 #ifdef UMA_DEBUG_ALLOC
2678 printf("uma_zfree: Allocating new free bucket.\n");
2682 if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2684 bucket = bucket_alloc(zone->uz_count, bflags);
2687 LIST_INSERT_HEAD(&zone->uz_free_bucket,
2694 * If nothing else caught this, we'll just do an internal free.
2697 zone_free_item(zone, item, udata, SKIP_DTOR, ZFREE_STATFREE);
2703 * Frees an item to an INTERNAL zone or allocates a free bucket
2706 * zone The zone to free to
2707 * item The item we're freeing
2708 * udata User supplied data for the dtor
2709 * skip Skip dtors and finis
2712 zone_free_item(uma_zone_t zone, void *item, void *udata,
2713 enum zfreeskip skip, int flags)
2716 uma_slabrefcnt_t slabref;
2722 if (skip < SKIP_DTOR && zone->uz_dtor)
2723 zone->uz_dtor(item, zone->uz_size, udata);
2725 if (skip < SKIP_FINI && zone->uz_fini)
2726 zone->uz_fini(item, zone->uz_size);
2730 if (flags & ZFREE_STATFAIL)
2732 if (flags & ZFREE_STATFREE)
2735 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2736 mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK));
2737 keg = zone_first_keg(zone); /* Must only be one. */
2738 if (zone->uz_flags & UMA_ZONE_HASH) {
2739 slab = hash_sfind(&keg->uk_hash, mem);
2741 mem += keg->uk_pgoff;
2742 slab = (uma_slab_t)mem;
2745 /* This prevents redundant lookups via free(). */
2746 if ((zone->uz_flags & UMA_ZONE_MALLOC) && udata != NULL)
2747 slab = (uma_slab_t)udata;
2749 slab = vtoslab((vm_offset_t)item);
2751 keg_relock(keg, zone);
2753 MPASS(keg == slab->us_keg);
2755 /* Do we need to remove from any lists? */
2756 if (slab->us_freecount+1 == keg->uk_ipers) {
2757 LIST_REMOVE(slab, us_link);
2758 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2759 } else if (slab->us_freecount == 0) {
2760 LIST_REMOVE(slab, us_link);
2761 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2764 /* Slab management stuff */
2765 freei = ((unsigned long)item - (unsigned long)slab->us_data)
2770 uma_dbg_free(zone, slab, item);
2773 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2774 slabref = (uma_slabrefcnt_t)slab;
2775 slabref->us_freelist[freei].us_item = slab->us_firstfree;
2777 slab->us_freelist[freei].us_item = slab->us_firstfree;
2779 slab->us_firstfree = freei;
2780 slab->us_freecount++;
2782 /* Zone statistics */
2786 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2787 if (keg->uk_pages < keg->uk_maxpages) {
2788 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2793 * We can handle one more allocation. Since we're clearing ZFLAG_FULL,
2794 * wake up all procs blocked on pages. This should be uncommon, so
2795 * keeping this simple for now (rather than adding count of blocked
2801 zone_relock(zone, keg);
2802 zone->uz_flags &= ~UMA_ZFLAG_FULL;
2811 uma_zone_set_max(uma_zone_t zone, int nitems)
2816 keg = zone_first_keg(zone);
2817 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2818 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2819 keg->uk_maxpages += keg->uk_ppera;
2826 uma_zone_get_max(uma_zone_t zone)
2832 keg = zone_first_keg(zone);
2833 nitems = keg->uk_maxpages * keg->uk_ipers;
2841 uma_zone_get_cur(uma_zone_t zone)
2847 nitems = zone->uz_allocs - zone->uz_frees;
2850 * See the comment in sysctl_vm_zone_stats() regarding the
2851 * safety of accessing the per-cpu caches. With the zone lock
2852 * held, it is safe, but can potentially result in stale data.
2854 nitems += zone->uz_cpu[i].uc_allocs -
2855 zone->uz_cpu[i].uc_frees;
2859 return (nitems < 0 ? 0 : nitems);
2864 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2869 keg = zone_first_keg(zone);
2870 KASSERT(keg->uk_pages == 0,
2871 ("uma_zone_set_init on non-empty keg"));
2872 keg->uk_init = uminit;
2878 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2883 keg = zone_first_keg(zone);
2884 KASSERT(keg->uk_pages == 0,
2885 ("uma_zone_set_fini on non-empty keg"));
2886 keg->uk_fini = fini;
2892 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2895 KASSERT(zone_first_keg(zone)->uk_pages == 0,
2896 ("uma_zone_set_zinit on non-empty keg"));
2897 zone->uz_init = zinit;
2903 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2906 KASSERT(zone_first_keg(zone)->uk_pages == 0,
2907 ("uma_zone_set_zfini on non-empty keg"));
2908 zone->uz_fini = zfini;
2913 /* XXX uk_freef is not actually used with the zone locked */
2915 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2919 zone_first_keg(zone)->uk_freef = freef;
2924 /* XXX uk_allocf is not actually used with the zone locked */
2926 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2931 keg = zone_first_keg(zone);
2932 keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2933 keg->uk_allocf = allocf;
2939 uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count)
2945 keg = zone_first_keg(zone);
2946 pages = count / keg->uk_ipers;
2948 if (pages * keg->uk_ipers < count)
2951 kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2956 obj = vm_object_allocate(OBJT_DEFAULT,
2959 VM_OBJECT_LOCK_INIT(obj, "uma object");
2960 _vm_object_allocate(OBJT_DEFAULT,
2966 keg->uk_maxpages = pages;
2967 keg->uk_allocf = obj_alloc;
2968 keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
2975 uma_prealloc(uma_zone_t zone, int items)
2981 keg = zone_first_keg(zone);
2983 slabs = items / keg->uk_ipers;
2984 if (slabs * keg->uk_ipers < items)
2987 slab = keg_alloc_slab(keg, zone, M_WAITOK);
2990 MPASS(slab->us_keg == keg);
2991 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2999 uma_find_refcnt(uma_zone_t zone, void *item)
3001 uma_slabrefcnt_t slabref;
3006 slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item &
3008 keg = slabref->us_keg;
3009 KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT,
3010 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3011 idx = ((unsigned long)item - (unsigned long)slabref->us_data)
3013 refcnt = &slabref->us_freelist[idx].us_refcnt;
3022 printf("UMA: vm asked us to release pages!\n");
3025 zone_foreach(zone_drain);
3027 * Some slabs may have been freed but this zone will be visited early
3028 * we visit again so that we can free pages that are empty once other
3029 * zones are drained. We have to do the same for buckets.
3031 zone_drain(slabzone);
3032 zone_drain(slabrefzone);
3033 bucket_zone_drain();
3038 uma_zone_exhausted(uma_zone_t zone)
3043 full = (zone->uz_flags & UMA_ZFLAG_FULL);
3049 uma_zone_exhausted_nolock(uma_zone_t zone)
3051 return (zone->uz_flags & UMA_ZFLAG_FULL);
3055 uma_large_malloc(int size, int wait)
3061 slab = zone_alloc_item(slabzone, NULL, wait);
3064 mem = page_alloc(NULL, size, &flags, wait);
3066 vsetslab((vm_offset_t)mem, slab);
3067 slab->us_data = mem;
3068 slab->us_flags = flags | UMA_SLAB_MALLOC;
3069 slab->us_size = size;
3071 zone_free_item(slabzone, slab, NULL, SKIP_NONE,
3072 ZFREE_STATFAIL | ZFREE_STATFREE);
3079 uma_large_free(uma_slab_t slab)
3081 vsetobj((vm_offset_t)slab->us_data, kmem_object);
3082 page_free(slab->us_data, slab->us_size, slab->us_flags);
3083 zone_free_item(slabzone, slab, NULL, SKIP_NONE, ZFREE_STATFREE);
3087 uma_print_stats(void)
3089 zone_foreach(uma_print_zone);
3093 slab_print(uma_slab_t slab)
3095 printf("slab: keg %p, data %p, freecount %d, firstfree %d\n",
3096 slab->us_keg, slab->us_data, slab->us_freecount,
3097 slab->us_firstfree);
3101 cache_print(uma_cache_t cache)
3103 printf("alloc: %p(%d), free: %p(%d)\n",
3104 cache->uc_allocbucket,
3105 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3106 cache->uc_freebucket,
3107 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3111 uma_print_keg(uma_keg_t keg)
3115 printf("keg: %s(%p) size %d(%d) flags %d ipers %d ppera %d "
3116 "out %d free %d limit %d\n",
3117 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3118 keg->uk_ipers, keg->uk_ppera,
3119 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3120 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3121 printf("Part slabs:\n");
3122 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3124 printf("Free slabs:\n");
3125 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3127 printf("Full slabs:\n");
3128 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3133 uma_print_zone(uma_zone_t zone)
3139 printf("zone: %s(%p) size %d flags %d\n",
3140 zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3141 LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3142 uma_print_keg(kl->kl_keg);
3144 cache = &zone->uz_cpu[i];
3145 printf("CPU %d Cache:\n", i);
3152 * Generate statistics across both the zone and its per-cpu cache's. Return
3153 * desired statistics if the pointer is non-NULL for that statistic.
3155 * Note: does not update the zone statistics, as it can't safely clear the
3156 * per-CPU cache statistic.
3158 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3159 * safe from off-CPU; we should modify the caches to track this information
3160 * directly so that we don't have to.
3163 uma_zone_sumstat(uma_zone_t z, int *cachefreep, u_int64_t *allocsp,
3167 u_int64_t allocs, frees;
3173 cache = &z->uz_cpu[cpu];
3174 if (cache->uc_allocbucket != NULL)
3175 cachefree += cache->uc_allocbucket->ub_cnt;
3176 if (cache->uc_freebucket != NULL)
3177 cachefree += cache->uc_freebucket->ub_cnt;
3178 allocs += cache->uc_allocs;
3179 frees += cache->uc_frees;
3181 allocs += z->uz_allocs;
3182 frees += z->uz_frees;
3183 if (cachefreep != NULL)
3184 *cachefreep = cachefree;
3185 if (allocsp != NULL)
3193 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3201 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3202 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3205 mtx_unlock(&uma_mtx);
3206 return (sysctl_handle_int(oidp, &count, 0, req));
3210 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3212 struct uma_stream_header ush;
3213 struct uma_type_header uth;
3214 struct uma_percpu_stat ups;
3215 uma_bucket_t bucket;
3223 int buflen, count, error, i;
3227 mtx_assert(&uma_mtx, MA_OWNED);
3229 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3230 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3233 mtx_unlock(&uma_mtx);
3235 buflen = sizeof(ush) + count * (sizeof(uth) + sizeof(ups) *
3236 (mp_maxid + 1)) + 1;
3237 buffer = malloc(buflen, M_TEMP, M_WAITOK | M_ZERO);
3241 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3242 LIST_FOREACH(z, &kz->uk_zones, uz_link)
3246 free(buffer, M_TEMP);
3251 sbuf_new(&sbuf, buffer, buflen, SBUF_FIXEDLEN);
3254 * Insert stream header.
3256 bzero(&ush, sizeof(ush));
3257 ush.ush_version = UMA_STREAM_VERSION;
3258 ush.ush_maxcpus = (mp_maxid + 1);
3259 ush.ush_count = count;
3260 if (sbuf_bcat(&sbuf, &ush, sizeof(ush)) < 0) {
3261 mtx_unlock(&uma_mtx);
3266 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3267 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3268 bzero(&uth, sizeof(uth));
3270 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3271 uth.uth_align = kz->uk_align;
3272 uth.uth_size = kz->uk_size;
3273 uth.uth_rsize = kz->uk_rsize;
3274 LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3276 uth.uth_maxpages += k->uk_maxpages;
3277 uth.uth_pages += k->uk_pages;
3278 uth.uth_keg_free += k->uk_free;
3279 uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3284 * A zone is secondary is it is not the first entry
3285 * on the keg's zone list.
3287 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3288 (LIST_FIRST(&kz->uk_zones) != z))
3289 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3291 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3292 uth.uth_zone_free += bucket->ub_cnt;
3293 uth.uth_allocs = z->uz_allocs;
3294 uth.uth_frees = z->uz_frees;
3295 uth.uth_fails = z->uz_fails;
3296 if (sbuf_bcat(&sbuf, &uth, sizeof(uth)) < 0) {
3298 mtx_unlock(&uma_mtx);
3303 * While it is not normally safe to access the cache
3304 * bucket pointers while not on the CPU that owns the
3305 * cache, we only allow the pointers to be exchanged
3306 * without the zone lock held, not invalidated, so
3307 * accept the possible race associated with bucket
3308 * exchange during monitoring.
3310 for (i = 0; i < (mp_maxid + 1); i++) {
3311 bzero(&ups, sizeof(ups));
3312 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3316 cache = &z->uz_cpu[i];
3317 if (cache->uc_allocbucket != NULL)
3318 ups.ups_cache_free +=
3319 cache->uc_allocbucket->ub_cnt;
3320 if (cache->uc_freebucket != NULL)
3321 ups.ups_cache_free +=
3322 cache->uc_freebucket->ub_cnt;
3323 ups.ups_allocs = cache->uc_allocs;
3324 ups.ups_frees = cache->uc_frees;
3326 if (sbuf_bcat(&sbuf, &ups, sizeof(ups)) < 0) {
3328 mtx_unlock(&uma_mtx);
3336 mtx_unlock(&uma_mtx);
3338 error = SYSCTL_OUT(req, sbuf_data(&sbuf), sbuf_len(&sbuf));
3340 free(buffer, M_TEMP);
3345 DB_SHOW_COMMAND(uma, db_show_uma)
3347 u_int64_t allocs, frees;
3348 uma_bucket_t bucket;
3353 db_printf("%18s %8s %8s %8s %12s\n", "Zone", "Size", "Used", "Free",
3355 LIST_FOREACH(kz, &uma_kegs, uk_link) {
3356 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3357 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3358 allocs = z->uz_allocs;
3359 frees = z->uz_frees;
3362 uma_zone_sumstat(z, &cachefree, &allocs,
3364 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3365 (LIST_FIRST(&kz->uk_zones) != z)))
3366 cachefree += kz->uk_free;
3367 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3368 cachefree += bucket->ub_cnt;
3369 db_printf("%18s %8ju %8jd %8d %12ju\n", z->uz_name,
3370 (uintmax_t)kz->uk_size,
3371 (intmax_t)(allocs - frees), cachefree,