2 * Copyright (c) 2004-2005 Robert N. M. Watson
3 * Copyright (c) 2004, 2005,
4 * Bosko Milekic <bmilekic@FreeBSD.org>. All rights reserved.
5 * Copyright (c) 2002, 2003, 2004, 2005,
6 * Jeffrey Roberson <jeff@FreeBSD.org>. All rights reserved.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice unmodified, this list of conditions, and the following
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
23 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
24 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
27 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
31 * uma_core.c Implementation of the Universal Memory allocator
33 * This allocator is intended to replace the multitude of similar object caches
34 * in the standard FreeBSD kernel. The intent is to be flexible as well as
35 * effecient. A primary design goal is to return unused memory to the rest of
36 * the system. This will make the system as a whole more flexible due to the
37 * ability to move memory to subsystems which most need it instead of leaving
38 * pools of reserved memory unused.
40 * The basic ideas stem from similar slab/zone based allocators whose algorithms
47 * - Improve memory usage for large allocations
48 * - Investigate cache size adjustments
51 #include <sys/cdefs.h>
52 __FBSDID("$FreeBSD$");
54 /* I should really use ktr.. */
57 #define UMA_DEBUG_ALLOC 1
58 #define UMA_DEBUG_ALLOC_1 1
61 #include "opt_param.h"
62 #include <sys/param.h>
63 #include <sys/systm.h>
64 #include <sys/kernel.h>
65 #include <sys/types.h>
66 #include <sys/queue.h>
67 #include <sys/malloc.h>
70 #include <sys/sysctl.h>
71 #include <sys/mutex.h>
74 #include <sys/vmmeter.h>
77 #include <vm/vm_object.h>
78 #include <vm/vm_page.h>
79 #include <vm/vm_param.h>
80 #include <vm/vm_map.h>
81 #include <vm/vm_kern.h>
82 #include <vm/vm_extern.h>
84 #include <vm/uma_int.h>
85 #include <vm/uma_dbg.h>
87 #include <machine/vmparam.h>
90 * This is the zone and keg from which all zones are spawned. The idea is that
91 * even the zone & keg heads are allocated from the allocator, so we use the
92 * bss section to bootstrap us.
94 static struct uma_keg masterkeg;
95 static struct uma_zone masterzone_k;
96 static struct uma_zone masterzone_z;
97 static uma_zone_t kegs = &masterzone_k;
98 static uma_zone_t zones = &masterzone_z;
100 /* This is the zone from which all of uma_slab_t's are allocated. */
101 static uma_zone_t slabzone;
102 static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
105 * The initial hash tables come out of this zone so they can be allocated
106 * prior to malloc coming up.
108 static uma_zone_t hashzone;
110 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
113 * Are we allowed to allocate buckets?
115 static int bucketdisable = 1;
117 /* Linked list of all kegs in the system */
118 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(&uma_kegs);
120 /* This mutex protects the keg list */
121 static struct mtx uma_mtx;
123 /* Linked list of boot time pages */
124 static LIST_HEAD(,uma_slab) uma_boot_pages =
125 LIST_HEAD_INITIALIZER(&uma_boot_pages);
127 /* Count of free boottime pages */
128 static int uma_boot_free = 0;
130 /* Is the VM done starting up? */
131 static int booted = 0;
133 /* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
134 static u_int uma_max_ipers;
135 static u_int uma_max_ipers_ref;
138 * This is the handle used to schedule events that need to happen
139 * outside of the allocation fast path.
141 static struct callout uma_callout;
142 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
145 * This structure is passed as the zone ctor arg so that I don't have to create
146 * a special allocation function just for zones.
148 struct uma_zctor_args {
160 struct uma_kctor_args {
169 struct uma_bucket_zone {
175 #define BUCKET_MAX 128
177 struct uma_bucket_zone bucket_zones[] = {
178 { NULL, "16 Bucket", 16 },
179 { NULL, "32 Bucket", 32 },
180 { NULL, "64 Bucket", 64 },
181 { NULL, "128 Bucket", 128 },
185 #define BUCKET_SHIFT 4
186 #define BUCKET_ZONES ((BUCKET_MAX >> BUCKET_SHIFT) + 1)
189 * bucket_size[] maps requested bucket sizes to zones that allocate a bucket
190 * of approximately the right size.
192 static uint8_t bucket_size[BUCKET_ZONES];
194 enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI };
198 static void *obj_alloc(uma_zone_t, int, u_int8_t *, int);
199 static void *page_alloc(uma_zone_t, int, u_int8_t *, int);
200 static void *startup_alloc(uma_zone_t, int, u_int8_t *, int);
201 static void page_free(void *, int, u_int8_t);
202 static uma_slab_t slab_zalloc(uma_zone_t, int);
203 static void cache_drain(uma_zone_t);
204 static void bucket_drain(uma_zone_t, uma_bucket_t);
205 static void bucket_cache_drain(uma_zone_t zone);
206 static int keg_ctor(void *, int, void *, int);
207 static void keg_dtor(void *, int, void *);
208 static int zone_ctor(void *, int, void *, int);
209 static void zone_dtor(void *, int, void *);
210 static int zero_init(void *, int, int);
211 static void zone_small_init(uma_zone_t zone);
212 static void zone_large_init(uma_zone_t zone);
213 static void zone_foreach(void (*zfunc)(uma_zone_t));
214 static void zone_timeout(uma_zone_t zone);
215 static int hash_alloc(struct uma_hash *);
216 static int hash_expand(struct uma_hash *, struct uma_hash *);
217 static void hash_free(struct uma_hash *hash);
218 static void uma_timeout(void *);
219 static void uma_startup3(void);
220 static void *uma_zalloc_internal(uma_zone_t, void *, int);
221 static void uma_zfree_internal(uma_zone_t, void *, void *, enum zfreeskip);
222 static void bucket_enable(void);
223 static void bucket_init(void);
224 static uma_bucket_t bucket_alloc(int, int);
225 static void bucket_free(uma_bucket_t);
226 static void bucket_zone_drain(void);
227 static int uma_zalloc_bucket(uma_zone_t zone, int flags);
228 static uma_slab_t uma_zone_slab(uma_zone_t zone, int flags);
229 static void *uma_slab_alloc(uma_zone_t zone, uma_slab_t slab);
230 static void zone_drain(uma_zone_t);
231 static uma_zone_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
232 uma_fini fini, int align, u_int16_t flags);
234 void uma_print_zone(uma_zone_t);
235 void uma_print_stats(void);
236 static int sysctl_vm_zone(SYSCTL_HANDLER_ARGS);
239 static int nosleepwithlocks = 1;
240 SYSCTL_INT(_debug, OID_AUTO, nosleepwithlocks, CTLFLAG_RW, &nosleepwithlocks,
241 0, "Convert M_WAITOK to M_NOWAIT to avoid lock-held-across-sleep paths");
243 static int nosleepwithlocks = 0;
244 SYSCTL_INT(_debug, OID_AUTO, nosleepwithlocks, CTLFLAG_RW, &nosleepwithlocks,
245 0, "Convert M_WAITOK to M_NOWAIT to avoid lock-held-across-sleep paths");
247 SYSCTL_OID(_vm, OID_AUTO, zone, CTLTYPE_STRING|CTLFLAG_RD,
248 NULL, 0, sysctl_vm_zone, "A", "Zone Info");
249 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
252 * This routine checks to see whether or not it's safe to enable buckets.
258 if (cnt.v_free_count < cnt.v_free_min)
265 * Initialize bucket_zones, the array of zones of buckets of various sizes.
267 * For each zone, calculate the memory required for each bucket, consisting
268 * of the header and an array of pointers. Initialize bucket_size[] to point
269 * the range of appropriate bucket sizes at the zone.
274 struct uma_bucket_zone *ubz;
278 for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) {
281 ubz = &bucket_zones[j];
282 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
283 size += sizeof(void *) * ubz->ubz_entries;
284 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
285 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
286 for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT))
287 bucket_size[i >> BUCKET_SHIFT] = j;
292 * Given a desired number of entries for a bucket, return the zone from which
293 * to allocate the bucket.
295 static struct uma_bucket_zone *
296 bucket_zone_lookup(int entries)
300 idx = howmany(entries, 1 << BUCKET_SHIFT);
301 return (&bucket_zones[bucket_size[idx]]);
305 bucket_alloc(int entries, int bflags)
307 struct uma_bucket_zone *ubz;
311 * This is to stop us from allocating per cpu buckets while we're
312 * running out of UMA_BOOT_PAGES. Otherwise, we would exhaust the
313 * boot pages. This also prevents us from allocating buckets in
314 * low memory situations.
319 ubz = bucket_zone_lookup(entries);
320 bucket = uma_zalloc_internal(ubz->ubz_zone, NULL, bflags);
323 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
326 bucket->ub_entries = ubz->ubz_entries;
333 bucket_free(uma_bucket_t bucket)
335 struct uma_bucket_zone *ubz;
337 ubz = bucket_zone_lookup(bucket->ub_entries);
338 uma_zfree_internal(ubz->ubz_zone, bucket, NULL, SKIP_NONE);
342 bucket_zone_drain(void)
344 struct uma_bucket_zone *ubz;
346 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
347 zone_drain(ubz->ubz_zone);
352 * Routine called by timeout which is used to fire off some time interval
353 * based calculations. (stats, hash size, etc.)
362 uma_timeout(void *unused)
365 zone_foreach(zone_timeout);
367 /* Reschedule this event */
368 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
372 * Routine to perform timeout driven calculations. This expands the
373 * hashes and does per cpu statistics aggregation.
376 * zone The zone to operate on
382 zone_timeout(uma_zone_t zone)
391 * Expand the zone hash table.
393 * This is done if the number of slabs is larger than the hash size.
394 * What I'm trying to do here is completely reduce collisions. This
395 * may be a little aggressive. Should I allow for two collisions max?
398 if (keg->uk_flags & UMA_ZONE_HASH &&
399 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
400 struct uma_hash newhash;
401 struct uma_hash oldhash;
405 * This is so involved because allocating and freeing
406 * while the zone lock is held will lead to deadlock.
407 * I have to do everything in stages and check for
410 newhash = keg->uk_hash;
412 ret = hash_alloc(&newhash);
415 if (hash_expand(&keg->uk_hash, &newhash)) {
416 oldhash = keg->uk_hash;
417 keg->uk_hash = newhash;
430 * Allocate and zero fill the next sized hash table from the appropriate
434 * hash A new hash structure with the old hash size in uh_hashsize
437 * 1 on sucess and 0 on failure.
440 hash_alloc(struct uma_hash *hash)
445 oldsize = hash->uh_hashsize;
447 /* We're just going to go to a power of two greater */
449 hash->uh_hashsize = oldsize * 2;
450 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
451 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
452 M_UMAHASH, M_NOWAIT);
454 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
455 hash->uh_slab_hash = uma_zalloc_internal(hashzone, NULL,
457 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
459 if (hash->uh_slab_hash) {
460 bzero(hash->uh_slab_hash, alloc);
461 hash->uh_hashmask = hash->uh_hashsize - 1;
469 * Expands the hash table for HASH zones. This is done from zone_timeout
470 * to reduce collisions. This must not be done in the regular allocation
471 * path, otherwise, we can recurse on the vm while allocating pages.
474 * oldhash The hash you want to expand
475 * newhash The hash structure for the new table
483 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
489 if (!newhash->uh_slab_hash)
492 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
496 * I need to investigate hash algorithms for resizing without a
500 for (i = 0; i < oldhash->uh_hashsize; i++)
501 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
502 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
503 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
504 hval = UMA_HASH(newhash, slab->us_data);
505 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
513 * Free the hash bucket to the appropriate backing store.
516 * slab_hash The hash bucket we're freeing
517 * hashsize The number of entries in that hash bucket
523 hash_free(struct uma_hash *hash)
525 if (hash->uh_slab_hash == NULL)
527 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
528 uma_zfree_internal(hashzone,
529 hash->uh_slab_hash, NULL, SKIP_NONE);
531 free(hash->uh_slab_hash, M_UMAHASH);
535 * Frees all outstanding items in a bucket
538 * zone The zone to free to, must be unlocked.
539 * bucket The free/alloc bucket with items, cpu queue must be locked.
546 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
558 /* We have to lookup the slab again for malloc.. */
559 if (zone->uz_keg->uk_flags & UMA_ZONE_MALLOC)
562 while (bucket->ub_cnt > 0) {
564 item = bucket->ub_bucket[bucket->ub_cnt];
566 bucket->ub_bucket[bucket->ub_cnt] = NULL;
567 KASSERT(item != NULL,
568 ("bucket_drain: botched ptr, item is NULL"));
571 * This is extremely inefficient. The slab pointer was passed
572 * to uma_zfree_arg, but we lost it because the buckets don't
573 * hold them. This will go away when free() gets a size passed
577 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
578 uma_zfree_internal(zone, item, slab, SKIP_DTOR);
583 * Drains the per cpu caches for a zone.
585 * NOTE: This may only be called while the zone is being turn down, and not
586 * during normal operation. This is necessary in order that we do not have
587 * to migrate CPUs to drain the per-CPU caches.
590 * zone The zone to drain, must be unlocked.
596 cache_drain(uma_zone_t zone)
602 * XXX: It is safe to not lock the per-CPU caches, because we're
603 * tearing down the zone anyway. I.e., there will be no further use
604 * of the caches at this point.
606 * XXX: It would good to be able to assert that the zone is being
607 * torn down to prevent improper use of cache_drain().
609 * XXX: We lock the zone before passing into bucket_cache_drain() as
610 * it is used elsewhere. Should the tear-down path be made special
611 * there in some form?
613 for (cpu = 0; cpu <= mp_maxid; cpu++) {
616 cache = &zone->uz_cpu[cpu];
617 bucket_drain(zone, cache->uc_allocbucket);
618 bucket_drain(zone, cache->uc_freebucket);
619 if (cache->uc_allocbucket != NULL)
620 bucket_free(cache->uc_allocbucket);
621 if (cache->uc_freebucket != NULL)
622 bucket_free(cache->uc_freebucket);
623 cache->uc_allocbucket = cache->uc_freebucket = NULL;
626 bucket_cache_drain(zone);
631 * Drain the cached buckets from a zone. Expects a locked zone on entry.
634 bucket_cache_drain(uma_zone_t zone)
639 * Drain the bucket queues and free the buckets, we just keep two per
642 while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
643 LIST_REMOVE(bucket, ub_link);
645 bucket_drain(zone, bucket);
650 /* Now we do the free queue.. */
651 while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
652 LIST_REMOVE(bucket, ub_link);
658 * Frees pages from a zone back to the system. This is done on demand from
659 * the pageout daemon.
662 * zone The zone to free pages from
663 * all Should we drain all items?
669 zone_drain(uma_zone_t zone)
671 struct slabhead freeslabs = { 0 };
682 * We don't want to take pages from statically allocated zones at this
685 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
691 printf("%s free items: %u\n", zone->uz_name, keg->uk_free);
693 bucket_cache_drain(zone);
694 if (keg->uk_free == 0)
697 slab = LIST_FIRST(&keg->uk_free_slab);
699 n = LIST_NEXT(slab, us_link);
701 /* We have no where to free these to */
702 if (slab->us_flags & UMA_SLAB_BOOT) {
707 LIST_REMOVE(slab, us_link);
708 keg->uk_pages -= keg->uk_ppera;
709 keg->uk_free -= keg->uk_ipers;
711 if (keg->uk_flags & UMA_ZONE_HASH)
712 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
714 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
721 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
722 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
724 for (i = 0; i < keg->uk_ipers; i++)
726 slab->us_data + (keg->uk_rsize * i),
728 flags = slab->us_flags;
731 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
732 (keg->uk_flags & UMA_ZONE_REFCNT)) {
735 if (flags & UMA_SLAB_KMEM)
739 for (i = 0; i < keg->uk_ppera; i++)
740 vsetobj((vm_offset_t)mem + (i * PAGE_SIZE),
743 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
744 uma_zfree_internal(keg->uk_slabzone, slab, NULL,
747 printf("%s: Returning %d bytes.\n",
748 zone->uz_name, UMA_SLAB_SIZE * keg->uk_ppera);
750 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags);
755 * Allocate a new slab for a zone. This does not insert the slab onto a list.
758 * zone The zone to allocate slabs for
759 * wait Shall we wait?
762 * The slab that was allocated or NULL if there is no memory and the
763 * caller specified M_NOWAIT.
766 slab_zalloc(uma_zone_t zone, int wait)
768 uma_slabrefcnt_t slabref;
779 printf("slab_zalloc: Allocating a new slab for %s\n", zone->uz_name);
783 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
784 slab = uma_zalloc_internal(keg->uk_slabzone, NULL, wait);
792 * This reproduces the old vm_zone behavior of zero filling pages the
793 * first time they are added to a zone.
795 * Malloced items are zeroed in uma_zalloc.
798 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
803 mem = keg->uk_allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE,
806 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
807 uma_zfree_internal(keg->uk_slabzone, slab, NULL,
813 /* Point the slab into the allocated memory */
814 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
815 slab = (uma_slab_t )(mem + keg->uk_pgoff);
817 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
818 (keg->uk_flags & UMA_ZONE_REFCNT))
819 for (i = 0; i < keg->uk_ppera; i++)
820 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
824 slab->us_freecount = keg->uk_ipers;
825 slab->us_firstfree = 0;
826 slab->us_flags = flags;
828 if (keg->uk_flags & UMA_ZONE_REFCNT) {
829 slabref = (uma_slabrefcnt_t)slab;
830 for (i = 0; i < keg->uk_ipers; i++) {
831 slabref->us_freelist[i].us_refcnt = 0;
832 slabref->us_freelist[i].us_item = i+1;
835 for (i = 0; i < keg->uk_ipers; i++)
836 slab->us_freelist[i].us_item = i+1;
839 if (keg->uk_init != NULL) {
840 for (i = 0; i < keg->uk_ipers; i++)
841 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
842 keg->uk_size, wait) != 0)
844 if (i != keg->uk_ipers) {
845 if (keg->uk_fini != NULL) {
846 for (i--; i > -1; i--)
847 keg->uk_fini(slab->us_data +
851 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
852 (keg->uk_flags & UMA_ZONE_REFCNT)) {
855 if (flags & UMA_SLAB_KMEM)
859 for (i = 0; i < keg->uk_ppera; i++)
860 vsetobj((vm_offset_t)mem +
861 (i * PAGE_SIZE), obj);
863 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
864 uma_zfree_internal(keg->uk_slabzone, slab,
866 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera,
874 if (keg->uk_flags & UMA_ZONE_HASH)
875 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
877 keg->uk_pages += keg->uk_ppera;
878 keg->uk_free += keg->uk_ipers;
884 * This function is intended to be used early on in place of page_alloc() so
885 * that we may use the boot time page cache to satisfy allocations before
889 startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
896 * Check our small startup cache to see if it has pages remaining.
899 if (uma_boot_free != 0) {
902 tmps = LIST_FIRST(&uma_boot_pages);
903 LIST_REMOVE(tmps, us_link);
905 mtx_unlock(&uma_mtx);
906 *pflag = tmps->us_flags;
907 return (tmps->us_data);
909 mtx_unlock(&uma_mtx);
911 panic("UMA: Increase UMA_BOOT_PAGES");
913 * Now that we've booted reset these users to their real allocator.
915 #ifdef UMA_MD_SMALL_ALLOC
916 keg->uk_allocf = uma_small_alloc;
918 keg->uk_allocf = page_alloc;
920 return keg->uk_allocf(zone, bytes, pflag, wait);
924 * Allocates a number of pages from the system
928 * bytes The number of bytes requested
929 * wait Shall we wait?
932 * A pointer to the alloced memory or possibly
933 * NULL if M_NOWAIT is set.
936 page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
938 void *p; /* Returned page */
940 *pflag = UMA_SLAB_KMEM;
941 p = (void *) kmem_malloc(kmem_map, bytes, wait);
947 * Allocates a number of pages from within an object
951 * bytes The number of bytes requested
952 * wait Shall we wait?
955 * A pointer to the alloced memory or possibly
956 * NULL if M_NOWAIT is set.
959 obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
962 vm_offset_t retkva, zkva;
964 int pages, startpages;
966 object = zone->uz_keg->uk_obj;
970 * This looks a little weird since we're getting one page at a time.
972 VM_OBJECT_LOCK(object);
973 p = TAILQ_LAST(&object->memq, pglist);
974 pages = p != NULL ? p->pindex + 1 : 0;
976 zkva = zone->uz_keg->uk_kva + pages * PAGE_SIZE;
977 for (; bytes > 0; bytes -= PAGE_SIZE) {
978 p = vm_page_alloc(object, pages,
979 VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED);
981 if (pages != startpages)
982 pmap_qremove(retkva, pages - startpages);
983 while (pages != startpages) {
985 p = TAILQ_LAST(&object->memq, pglist);
986 vm_page_lock_queues();
987 vm_page_unwire(p, 0);
989 vm_page_unlock_queues();
994 pmap_qenter(zkva, &p, 1);
1001 VM_OBJECT_UNLOCK(object);
1002 *flags = UMA_SLAB_PRIV;
1004 return ((void *)retkva);
1008 * Frees a number of pages to the system
1011 * mem A pointer to the memory to be freed
1012 * size The size of the memory being freed
1013 * flags The original p->us_flags field
1019 page_free(void *mem, int size, u_int8_t flags)
1023 if (flags & UMA_SLAB_KMEM)
1026 panic("UMA: page_free used with invalid flags %d\n", flags);
1028 kmem_free(map, (vm_offset_t)mem, size);
1032 * Zero fill initializer
1034 * Arguments/Returns follow uma_init specifications
1037 zero_init(void *mem, int size, int flags)
1044 * Finish creating a small uma zone. This calculates ipers, and the zone size.
1047 * zone The zone we should initialize
1053 zone_small_init(uma_zone_t zone)
1062 KASSERT(keg != NULL, ("Keg is null in zone_small_init"));
1063 rsize = keg->uk_size;
1065 if (rsize < UMA_SMALLEST_UNIT)
1066 rsize = UMA_SMALLEST_UNIT;
1067 if (rsize & keg->uk_align)
1068 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1070 keg->uk_rsize = rsize;
1073 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1074 rsize += UMA_FRITMREF_SZ; /* linkage & refcnt */
1075 shsize = sizeof(struct uma_slab_refcnt);
1077 rsize += UMA_FRITM_SZ; /* Account for linkage */
1078 shsize = sizeof(struct uma_slab);
1081 keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize;
1082 KASSERT(keg->uk_ipers != 0, ("zone_small_init: ipers is 0"));
1083 memused = keg->uk_ipers * rsize + shsize;
1084 wastedspace = UMA_SLAB_SIZE - memused;
1087 * We can't do OFFPAGE if we're internal or if we've been
1088 * asked to not go to the VM for buckets. If we do this we
1089 * may end up going to the VM (kmem_map) for slabs which we
1090 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1091 * result of UMA_ZONE_VM, which clearly forbids it.
1093 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1094 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1097 if ((wastedspace >= UMA_MAX_WASTE) &&
1098 (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) {
1099 keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize;
1100 KASSERT(keg->uk_ipers <= 255,
1101 ("zone_small_init: keg->uk_ipers too high!"));
1103 printf("UMA decided we need offpage slab headers for "
1104 "zone: %s, calculated wastedspace = %d, "
1105 "maximum wasted space allowed = %d, "
1106 "calculated ipers = %d, "
1107 "new wasted space = %d\n", zone->uz_name, wastedspace,
1108 UMA_MAX_WASTE, keg->uk_ipers,
1109 UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize);
1111 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1112 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1113 keg->uk_flags |= UMA_ZONE_HASH;
1118 * Finish creating a large (> UMA_SLAB_SIZE) uma zone. Just give in and do
1119 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1123 * zone The zone we should initialize
1129 zone_large_init(uma_zone_t zone)
1136 KASSERT(keg != NULL, ("Keg is null in zone_large_init"));
1137 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1138 ("zone_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY zone"));
1140 pages = keg->uk_size / UMA_SLAB_SIZE;
1142 /* Account for remainder */
1143 if ((pages * UMA_SLAB_SIZE) < keg->uk_size)
1146 keg->uk_ppera = pages;
1149 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1150 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1151 keg->uk_flags |= UMA_ZONE_HASH;
1153 keg->uk_rsize = keg->uk_size;
1157 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1158 * the keg onto the global keg list.
1160 * Arguments/Returns follow uma_ctor specifications
1161 * udata Actually uma_kctor_args
1164 keg_ctor(void *mem, int size, void *udata, int flags)
1166 struct uma_kctor_args *arg = udata;
1167 uma_keg_t keg = mem;
1171 keg->uk_size = arg->size;
1172 keg->uk_init = arg->uminit;
1173 keg->uk_fini = arg->fini;
1174 keg->uk_align = arg->align;
1177 keg->uk_flags = arg->flags;
1178 keg->uk_allocf = page_alloc;
1179 keg->uk_freef = page_free;
1180 keg->uk_recurse = 0;
1181 keg->uk_slabzone = NULL;
1184 * The master zone is passed to us at keg-creation time.
1189 if (arg->flags & UMA_ZONE_VM)
1190 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1192 if (arg->flags & UMA_ZONE_ZINIT)
1193 keg->uk_init = zero_init;
1196 * The +UMA_FRITM_SZ added to uk_size is to account for the
1197 * linkage that is added to the size in zone_small_init(). If
1198 * we don't account for this here then we may end up in
1199 * zone_small_init() with a calculated 'ipers' of 0.
1201 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1202 if ((keg->uk_size+UMA_FRITMREF_SZ) >
1203 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)))
1204 zone_large_init(zone);
1206 zone_small_init(zone);
1208 if ((keg->uk_size+UMA_FRITM_SZ) >
1209 (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1210 zone_large_init(zone);
1212 zone_small_init(zone);
1215 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1216 if (keg->uk_flags & UMA_ZONE_REFCNT)
1217 keg->uk_slabzone = slabrefzone;
1219 keg->uk_slabzone = slabzone;
1223 * If we haven't booted yet we need allocations to go through the
1224 * startup cache until the vm is ready.
1226 if (keg->uk_ppera == 1) {
1227 #ifdef UMA_MD_SMALL_ALLOC
1228 keg->uk_allocf = uma_small_alloc;
1229 keg->uk_freef = uma_small_free;
1232 keg->uk_allocf = startup_alloc;
1236 * Initialize keg's lock (shared among zones) through
1239 zone->uz_lock = &keg->uk_lock;
1240 if (arg->flags & UMA_ZONE_MTXCLASS)
1241 ZONE_LOCK_INIT(zone, 1);
1243 ZONE_LOCK_INIT(zone, 0);
1246 * If we're putting the slab header in the actual page we need to
1247 * figure out where in each page it goes. This calculates a right
1248 * justified offset into the memory on an ALIGN_PTR boundary.
1250 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1253 /* Size of the slab struct and free list */
1254 if (keg->uk_flags & UMA_ZONE_REFCNT)
1255 totsize = sizeof(struct uma_slab_refcnt) +
1256 keg->uk_ipers * UMA_FRITMREF_SZ;
1258 totsize = sizeof(struct uma_slab) +
1259 keg->uk_ipers * UMA_FRITM_SZ;
1261 if (totsize & UMA_ALIGN_PTR)
1262 totsize = (totsize & ~UMA_ALIGN_PTR) +
1263 (UMA_ALIGN_PTR + 1);
1264 keg->uk_pgoff = UMA_SLAB_SIZE - totsize;
1266 if (keg->uk_flags & UMA_ZONE_REFCNT)
1267 totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt)
1268 + keg->uk_ipers * UMA_FRITMREF_SZ;
1270 totsize = keg->uk_pgoff + sizeof(struct uma_slab)
1271 + keg->uk_ipers * UMA_FRITM_SZ;
1274 * The only way the following is possible is if with our
1275 * UMA_ALIGN_PTR adjustments we are now bigger than
1276 * UMA_SLAB_SIZE. I haven't checked whether this is
1277 * mathematically possible for all cases, so we make
1280 if (totsize > UMA_SLAB_SIZE) {
1281 printf("zone %s ipers %d rsize %d size %d\n",
1282 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1284 panic("UMA slab won't fit.\n");
1288 if (keg->uk_flags & UMA_ZONE_HASH)
1289 hash_alloc(&keg->uk_hash);
1292 printf("%s(%p) size = %d ipers = %d ppera = %d pgoff = %d\n",
1293 zone->uz_name, zone,
1294 keg->uk_size, keg->uk_ipers,
1295 keg->uk_ppera, keg->uk_pgoff);
1298 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1301 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1302 mtx_unlock(&uma_mtx);
1307 * Zone header ctor. This initializes all fields, locks, etc.
1309 * Arguments/Returns follow uma_ctor specifications
1310 * udata Actually uma_zctor_args
1314 zone_ctor(void *mem, int size, void *udata, int flags)
1316 struct uma_zctor_args *arg = udata;
1317 uma_zone_t zone = mem;
1322 zone->uz_name = arg->name;
1323 zone->uz_ctor = arg->ctor;
1324 zone->uz_dtor = arg->dtor;
1325 zone->uz_init = NULL;
1326 zone->uz_fini = NULL;
1327 zone->uz_allocs = 0;
1328 zone->uz_fills = zone->uz_count = 0;
1330 if (arg->flags & UMA_ZONE_SECONDARY) {
1331 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1334 zone->uz_init = arg->uminit;
1335 zone->uz_fini = arg->fini;
1336 zone->uz_lock = &keg->uk_lock;
1339 keg->uk_flags |= UMA_ZONE_SECONDARY;
1340 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1341 if (LIST_NEXT(z, uz_link) == NULL) {
1342 LIST_INSERT_AFTER(z, zone, uz_link);
1347 mtx_unlock(&uma_mtx);
1348 } else if (arg->keg == NULL) {
1349 if (uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1350 arg->align, arg->flags) == NULL)
1353 struct uma_kctor_args karg;
1356 /* We should only be here from uma_startup() */
1357 karg.size = arg->size;
1358 karg.uminit = arg->uminit;
1359 karg.fini = arg->fini;
1360 karg.align = arg->align;
1361 karg.flags = arg->flags;
1363 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1369 zone->uz_lock = &keg->uk_lock;
1372 * Some internal zones don't have room allocated for the per cpu
1373 * caches. If we're internal, bail out here.
1375 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1376 KASSERT((keg->uk_flags & UMA_ZONE_SECONDARY) == 0,
1377 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1381 if (keg->uk_flags & UMA_ZONE_MAXBUCKET)
1382 zone->uz_count = BUCKET_MAX;
1383 else if (keg->uk_ipers <= BUCKET_MAX)
1384 zone->uz_count = keg->uk_ipers;
1386 zone->uz_count = BUCKET_MAX;
1391 * Keg header dtor. This frees all data, destroys locks, frees the hash
1392 * table and removes the keg from the global list.
1394 * Arguments/Returns follow uma_dtor specifications
1398 keg_dtor(void *arg, int size, void *udata)
1402 keg = (uma_keg_t)arg;
1403 mtx_lock(&keg->uk_lock);
1404 if (keg->uk_free != 0) {
1405 printf("Freed UMA keg was not empty (%d items). "
1406 " Lost %d pages of memory.\n",
1407 keg->uk_free, keg->uk_pages);
1409 mtx_unlock(&keg->uk_lock);
1411 if (keg->uk_flags & UMA_ZONE_HASH)
1412 hash_free(&keg->uk_hash);
1414 mtx_destroy(&keg->uk_lock);
1420 * Arguments/Returns follow uma_dtor specifications
1424 zone_dtor(void *arg, int size, void *udata)
1429 zone = (uma_zone_t)arg;
1432 if (!(keg->uk_flags & UMA_ZFLAG_INTERNAL))
1437 if (keg->uk_flags & UMA_ZONE_SECONDARY) {
1438 LIST_REMOVE(zone, uz_link);
1440 * XXX there are some races here where
1441 * the zone can be drained but zone lock
1442 * released and then refilled before we
1443 * remove it... we dont care for now
1446 if (LIST_EMPTY(&keg->uk_zones))
1447 keg->uk_flags &= ~UMA_ZONE_SECONDARY;
1449 mtx_unlock(&uma_mtx);
1451 LIST_REMOVE(keg, uk_link);
1452 LIST_REMOVE(zone, uz_link);
1453 mtx_unlock(&uma_mtx);
1454 uma_zfree_internal(kegs, keg, NULL, SKIP_NONE);
1456 zone->uz_keg = NULL;
1460 * Traverses every zone in the system and calls a callback
1463 * zfunc A pointer to a function which accepts a zone
1470 zone_foreach(void (*zfunc)(uma_zone_t))
1476 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1477 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1480 mtx_unlock(&uma_mtx);
1483 /* Public functions */
1486 uma_startup(void *bootmem)
1488 struct uma_zctor_args args;
1491 u_int objsize, totsize, wsize;
1495 printf("Creating uma keg headers zone and keg.\n");
1498 * The general UMA lock is a recursion-allowed lock because
1499 * there is a code path where, while we're still configured
1500 * to use startup_alloc() for backend page allocations, we
1501 * may end up in uma_reclaim() which calls zone_foreach(zone_drain),
1502 * which grabs uma_mtx, only to later call into startup_alloc()
1503 * because while freeing we needed to allocate a bucket. Since
1504 * startup_alloc() also takes uma_mtx, we need to be able to
1507 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF | MTX_RECURSE);
1510 * Figure out the maximum number of items-per-slab we'll have if
1511 * we're using the OFFPAGE slab header to track free items, given
1512 * all possible object sizes and the maximum desired wastage
1515 * We iterate until we find an object size for
1516 * which the calculated wastage in zone_small_init() will be
1517 * enough to warrant OFFPAGE. Since wastedspace versus objsize
1518 * is an overall increasing see-saw function, we find the smallest
1519 * objsize such that the wastage is always acceptable for objects
1520 * with that objsize or smaller. Since a smaller objsize always
1521 * generates a larger possible uma_max_ipers, we use this computed
1522 * objsize to calculate the largest ipers possible. Since the
1523 * ipers calculated for OFFPAGE slab headers is always larger than
1524 * the ipers initially calculated in zone_small_init(), we use
1525 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to
1526 * obtain the maximum ipers possible for offpage slab headers.
1528 * It should be noted that ipers versus objsize is an inversly
1529 * proportional function which drops off rather quickly so as
1530 * long as our UMA_MAX_WASTE is such that the objsize we calculate
1531 * falls into the portion of the inverse relation AFTER the steep
1532 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386).
1534 * Note that we have 8-bits (1 byte) to use as a freelist index
1535 * inside the actual slab header itself and this is enough to
1536 * accomodate us. In the worst case, a UMA_SMALLEST_UNIT sized
1537 * object with offpage slab header would have ipers =
1538 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is
1539 * 1 greater than what our byte-integer freelist index can
1540 * accomodate, but we know that this situation never occurs as
1541 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate
1542 * that we need to go to offpage slab headers. Or, if we do,
1543 * then we trap that condition below and panic in the INVARIANTS case.
1545 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE;
1547 objsize = UMA_SMALLEST_UNIT;
1548 while (totsize >= wsize) {
1549 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) /
1550 (objsize + UMA_FRITM_SZ);
1551 totsize *= (UMA_FRITM_SZ + objsize);
1554 if (objsize > UMA_SMALLEST_UNIT)
1556 uma_max_ipers = UMA_SLAB_SIZE / objsize;
1558 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE;
1560 objsize = UMA_SMALLEST_UNIT;
1561 while (totsize >= wsize) {
1562 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) /
1563 (objsize + UMA_FRITMREF_SZ);
1564 totsize *= (UMA_FRITMREF_SZ + objsize);
1567 if (objsize > UMA_SMALLEST_UNIT)
1569 uma_max_ipers_ref = UMA_SLAB_SIZE / objsize;
1571 KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255),
1572 ("uma_startup: calculated uma_max_ipers values too large!"));
1575 printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers);
1576 printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n",
1580 /* "manually" create the initial zone */
1581 args.name = "UMA Kegs";
1582 args.size = sizeof(struct uma_keg);
1583 args.ctor = keg_ctor;
1584 args.dtor = keg_dtor;
1585 args.uminit = zero_init;
1587 args.keg = &masterkeg;
1588 args.align = 32 - 1;
1589 args.flags = UMA_ZFLAG_INTERNAL;
1590 /* The initial zone has no Per cpu queues so it's smaller */
1591 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1594 printf("Filling boot free list.\n");
1596 for (i = 0; i < UMA_BOOT_PAGES; i++) {
1597 slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE));
1598 slab->us_data = (u_int8_t *)slab;
1599 slab->us_flags = UMA_SLAB_BOOT;
1600 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1605 printf("Creating uma zone headers zone and keg.\n");
1607 args.name = "UMA Zones";
1608 args.size = sizeof(struct uma_zone) +
1609 (sizeof(struct uma_cache) * (mp_maxid + 1));
1610 args.ctor = zone_ctor;
1611 args.dtor = zone_dtor;
1612 args.uminit = zero_init;
1615 args.align = 32 - 1;
1616 args.flags = UMA_ZFLAG_INTERNAL;
1617 /* The initial zone has no Per cpu queues so it's smaller */
1618 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1621 printf("Initializing pcpu cache locks.\n");
1624 printf("Creating slab and hash zones.\n");
1628 * This is the max number of free list items we'll have with
1631 slabsize = uma_max_ipers * UMA_FRITM_SZ;
1632 slabsize += sizeof(struct uma_slab);
1634 /* Now make a zone for slab headers */
1635 slabzone = uma_zcreate("UMA Slabs",
1637 NULL, NULL, NULL, NULL,
1638 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1641 * We also create a zone for the bigger slabs with reference
1642 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1644 slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ;
1645 slabsize += sizeof(struct uma_slab_refcnt);
1646 slabrefzone = uma_zcreate("UMA RCntSlabs",
1648 NULL, NULL, NULL, NULL,
1650 UMA_ZFLAG_INTERNAL);
1652 hashzone = uma_zcreate("UMA Hash",
1653 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1654 NULL, NULL, NULL, NULL,
1655 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1659 #ifdef UMA_MD_SMALL_ALLOC
1664 printf("UMA startup complete.\n");
1675 printf("UMA startup2 complete.\n");
1680 * Initialize our callout handle
1688 printf("Starting callout.\n");
1690 callout_init(&uma_callout, CALLOUT_MPSAFE);
1691 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1693 printf("UMA startup3 complete.\n");
1698 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1699 int align, u_int16_t flags)
1701 struct uma_kctor_args args;
1704 args.uminit = uminit;
1709 return (uma_zalloc_internal(kegs, &args, M_WAITOK));
1714 uma_zcreate(char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1715 uma_init uminit, uma_fini fini, int align, u_int16_t flags)
1718 struct uma_zctor_args args;
1720 /* This stuff is essential for the zone ctor */
1725 args.uminit = uminit;
1731 return (uma_zalloc_internal(zones, &args, M_WAITOK));
1736 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1737 uma_init zinit, uma_fini zfini, uma_zone_t master)
1739 struct uma_zctor_args args;
1742 args.size = master->uz_keg->uk_size;
1745 args.uminit = zinit;
1747 args.align = master->uz_keg->uk_align;
1748 args.flags = master->uz_keg->uk_flags | UMA_ZONE_SECONDARY;
1749 args.keg = master->uz_keg;
1751 return (uma_zalloc_internal(zones, &args, M_WAITOK));
1756 uma_zdestroy(uma_zone_t zone)
1758 uma_zfree_internal(zones, zone, NULL, SKIP_NONE);
1763 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1767 uma_bucket_t bucket;
1771 /* This is the fast path allocation */
1772 #ifdef UMA_DEBUG_ALLOC_1
1773 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1775 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1776 zone->uz_name, flags);
1778 if (!(flags & M_NOWAIT)) {
1779 KASSERT(curthread->td_intr_nesting_level == 0,
1780 ("malloc(M_WAITOK) in interrupt context"));
1781 if (nosleepwithlocks) {
1783 badness = WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK,
1785 "malloc(M_WAITOK) of \"%s\", forcing M_NOWAIT",
1793 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1794 "malloc(M_WAITOK) of \"%s\"", zone->uz_name);
1804 * If possible, allocate from the per-CPU cache. There are two
1805 * requirements for safe access to the per-CPU cache: (1) the thread
1806 * accessing the cache must not be preempted or yield during access,
1807 * and (2) the thread must not migrate CPUs without switching which
1808 * cache it accesses. We rely on a critical section to prevent
1809 * preemption and migration. We release the critical section in
1810 * order to acquire the zone mutex if we are unable to allocate from
1811 * the current cache; when we re-acquire the critical section, we
1812 * must detect and handle migration if it has occurred.
1817 cache = &zone->uz_cpu[cpu];
1820 bucket = cache->uc_allocbucket;
1823 if (bucket->ub_cnt > 0) {
1825 item = bucket->ub_bucket[bucket->ub_cnt];
1827 bucket->ub_bucket[bucket->ub_cnt] = NULL;
1829 KASSERT(item != NULL,
1830 ("uma_zalloc: Bucket pointer mangled."));
1835 uma_dbg_alloc(zone, NULL, item);
1838 if (zone->uz_ctor != NULL) {
1839 if (zone->uz_ctor(item, zone->uz_keg->uk_size,
1840 udata, flags) != 0) {
1841 uma_zfree_internal(zone, item, udata,
1847 bzero(item, zone->uz_keg->uk_size);
1849 } else if (cache->uc_freebucket) {
1851 * We have run out of items in our allocbucket.
1852 * See if we can switch with our free bucket.
1854 if (cache->uc_freebucket->ub_cnt > 0) {
1855 #ifdef UMA_DEBUG_ALLOC
1856 printf("uma_zalloc: Swapping empty with"
1859 bucket = cache->uc_freebucket;
1860 cache->uc_freebucket = cache->uc_allocbucket;
1861 cache->uc_allocbucket = bucket;
1868 * Attempt to retrieve the item from the per-CPU cache has failed, so
1869 * we must go back to the zone. This requires the zone lock, so we
1870 * must drop the critical section, then re-acquire it when we go back
1871 * to the cache. Since the critical section is released, we may be
1872 * preempted or migrate. As such, make sure not to maintain any
1873 * thread-local state specific to the cache from prior to releasing
1874 * the critical section.
1880 cache = &zone->uz_cpu[cpu];
1881 bucket = cache->uc_allocbucket;
1882 if (bucket != NULL) {
1883 if (bucket->ub_cnt > 0) {
1887 bucket = cache->uc_freebucket;
1888 if (bucket != NULL && bucket->ub_cnt > 0) {
1894 /* Since we have locked the zone we may as well send back our stats */
1895 zone->uz_allocs += cache->uc_allocs;
1896 cache->uc_allocs = 0;
1898 /* Our old one is now a free bucket */
1899 if (cache->uc_allocbucket) {
1900 KASSERT(cache->uc_allocbucket->ub_cnt == 0,
1901 ("uma_zalloc_arg: Freeing a non free bucket."));
1902 LIST_INSERT_HEAD(&zone->uz_free_bucket,
1903 cache->uc_allocbucket, ub_link);
1904 cache->uc_allocbucket = NULL;
1907 /* Check the free list for a new alloc bucket */
1908 if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
1909 KASSERT(bucket->ub_cnt != 0,
1910 ("uma_zalloc_arg: Returning an empty bucket."));
1912 LIST_REMOVE(bucket, ub_link);
1913 cache->uc_allocbucket = bucket;
1917 /* We are no longer associated with this CPU. */
1920 /* Bump up our uz_count so we get here less */
1921 if (zone->uz_count < BUCKET_MAX)
1925 * Now lets just fill a bucket and put it on the free list. If that
1926 * works we'll restart the allocation from the begining.
1928 if (uma_zalloc_bucket(zone, flags)) {
1930 goto zalloc_restart;
1934 * We may not be able to get a bucket so return an actual item.
1937 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
1940 return (uma_zalloc_internal(zone, udata, flags));
1944 uma_zone_slab(uma_zone_t zone, int flags)
1952 * This is to prevent us from recursively trying to allocate
1953 * buckets. The problem is that if an allocation forces us to
1954 * grab a new bucket we will call page_alloc, which will go off
1955 * and cause the vm to allocate vm_map_entries. If we need new
1956 * buckets there too we will recurse in kmem_alloc and bad
1957 * things happen. So instead we return a NULL bucket, and make
1958 * the code that allocates buckets smart enough to deal with it
1960 * XXX: While we want this protection for the bucket zones so that
1961 * recursion from the VM is handled (and the calling code that
1962 * allocates buckets knows how to deal with it), we do not want
1963 * to prevent allocation from the slab header zones (slabzone
1964 * and slabrefzone) if uk_recurse is not zero for them. The
1965 * reason is that it could lead to NULL being returned for
1966 * slab header allocations even in the M_WAITOK case, and the
1967 * caller can't handle that.
1969 if (keg->uk_flags & UMA_ZFLAG_INTERNAL && keg->uk_recurse != 0)
1970 if ((zone != slabzone) && (zone != slabrefzone))
1977 * Find a slab with some space. Prefer slabs that are partially
1978 * used over those that are totally full. This helps to reduce
1981 if (keg->uk_free != 0) {
1982 if (!LIST_EMPTY(&keg->uk_part_slab)) {
1983 slab = LIST_FIRST(&keg->uk_part_slab);
1985 slab = LIST_FIRST(&keg->uk_free_slab);
1986 LIST_REMOVE(slab, us_link);
1987 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
1994 * M_NOVM means don't ask at all!
1999 if (keg->uk_maxpages &&
2000 keg->uk_pages >= keg->uk_maxpages) {
2001 keg->uk_flags |= UMA_ZFLAG_FULL;
2003 if (flags & M_NOWAIT)
2006 msleep(keg, &keg->uk_lock, PVM,
2011 slab = slab_zalloc(zone, flags);
2015 * If we got a slab here it's safe to mark it partially used
2016 * and return. We assume that the caller is going to remove
2017 * at least one item.
2020 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2024 * We might not have been able to get a slab but another cpu
2025 * could have while we were unlocked. Check again before we
2028 if (flags & M_NOWAIT)
2035 uma_slab_alloc(uma_zone_t zone, uma_slab_t slab)
2038 uma_slabrefcnt_t slabref;
2044 freei = slab->us_firstfree;
2045 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2046 slabref = (uma_slabrefcnt_t)slab;
2047 slab->us_firstfree = slabref->us_freelist[freei].us_item;
2049 slab->us_firstfree = slab->us_freelist[freei].us_item;
2051 item = slab->us_data + (keg->uk_rsize * freei);
2053 slab->us_freecount--;
2056 uma_dbg_alloc(zone, slab, item);
2058 /* Move this slab to the full list */
2059 if (slab->us_freecount == 0) {
2060 LIST_REMOVE(slab, us_link);
2061 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2068 uma_zalloc_bucket(uma_zone_t zone, int flags)
2070 uma_bucket_t bucket;
2073 int max, origflags = flags;
2076 * Try this zone's free list first so we don't allocate extra buckets.
2078 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2079 KASSERT(bucket->ub_cnt == 0,
2080 ("uma_zalloc_bucket: Bucket on free list is not empty."));
2081 LIST_REMOVE(bucket, ub_link);
2085 bflags = (flags & ~M_ZERO);
2086 if (zone->uz_keg->uk_flags & UMA_ZFLAG_CACHEONLY)
2090 bucket = bucket_alloc(zone->uz_count, bflags);
2099 * This code is here to limit the number of simultaneous bucket fills
2100 * for any given zone to the number of per cpu caches in this zone. This
2101 * is done so that we don't allocate more memory than we really need.
2103 if (zone->uz_fills >= mp_ncpus)
2109 max = MIN(bucket->ub_entries, zone->uz_count);
2110 /* Try to keep the buckets totally full */
2111 saved = bucket->ub_cnt;
2112 while (bucket->ub_cnt < max &&
2113 (slab = uma_zone_slab(zone, flags)) != NULL) {
2114 while (slab->us_freecount && bucket->ub_cnt < max) {
2115 bucket->ub_bucket[bucket->ub_cnt++] =
2116 uma_slab_alloc(zone, slab);
2119 /* Don't block on the next fill */
2124 * We unlock here because we need to call the zone's init.
2125 * It should be safe to unlock because the slab dealt with
2126 * above is already on the appropriate list within the keg
2127 * and the bucket we filled is not yet on any list, so we
2130 if (zone->uz_init != NULL) {
2134 for (i = saved; i < bucket->ub_cnt; i++)
2135 if (zone->uz_init(bucket->ub_bucket[i],
2136 zone->uz_keg->uk_size, origflags) != 0)
2139 * If we couldn't initialize the whole bucket, put the
2140 * rest back onto the freelist.
2142 if (i != bucket->ub_cnt) {
2145 for (j = i; j < bucket->ub_cnt; j++) {
2146 uma_zfree_internal(zone, bucket->ub_bucket[j],
2149 bucket->ub_bucket[j] = NULL;
2158 if (bucket->ub_cnt != 0) {
2159 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2166 bucket_free(bucket);
2171 * Allocates an item for an internal zone
2174 * zone The zone to alloc for.
2175 * udata The data to be passed to the constructor.
2176 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2179 * NULL if there is no memory and M_NOWAIT is set
2180 * An item if successful
2184 uma_zalloc_internal(uma_zone_t zone, void *udata, int flags)
2193 #ifdef UMA_DEBUG_ALLOC
2194 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2198 slab = uma_zone_slab(zone, flags);
2204 item = uma_slab_alloc(zone, slab);
2209 * We have to call both the zone's init (not the keg's init)
2210 * and the zone's ctor. This is because the item is going from
2211 * a keg slab directly to the user, and the user is expecting it
2212 * to be both zone-init'd as well as zone-ctor'd.
2214 if (zone->uz_init != NULL) {
2215 if (zone->uz_init(item, keg->uk_size, flags) != 0) {
2216 uma_zfree_internal(zone, item, udata, SKIP_FINI);
2220 if (zone->uz_ctor != NULL) {
2221 if (zone->uz_ctor(item, keg->uk_size, udata, flags) != 0) {
2222 uma_zfree_internal(zone, item, udata, SKIP_DTOR);
2227 bzero(item, keg->uk_size);
2234 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2238 uma_bucket_t bucket;
2244 #ifdef UMA_DEBUG_ALLOC_1
2245 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2247 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2251 zone->uz_dtor(item, keg->uk_size, udata);
2254 if (keg->uk_flags & UMA_ZONE_MALLOC)
2255 uma_dbg_free(zone, udata, item);
2257 uma_dbg_free(zone, NULL, item);
2261 * The race here is acceptable. If we miss it we'll just have to wait
2262 * a little longer for the limits to be reset.
2264 if (keg->uk_flags & UMA_ZFLAG_FULL)
2265 goto zfree_internal;
2268 * If possible, free to the per-CPU cache. There are two
2269 * requirements for safe access to the per-CPU cache: (1) the thread
2270 * accessing the cache must not be preempted or yield during access,
2271 * and (2) the thread must not migrate CPUs without switching which
2272 * cache it accesses. We rely on a critical section to prevent
2273 * preemption and migration. We release the critical section in
2274 * order to acquire the zone mutex if we are unable to free to the
2275 * current cache; when we re-acquire the critical section, we must
2276 * detect and handle migration if it has occurred.
2281 cache = &zone->uz_cpu[cpu];
2284 bucket = cache->uc_freebucket;
2288 * Do we have room in our bucket? It is OK for this uz count
2289 * check to be slightly out of sync.
2292 if (bucket->ub_cnt < bucket->ub_entries) {
2293 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2294 ("uma_zfree: Freeing to non free bucket index."));
2295 bucket->ub_bucket[bucket->ub_cnt] = item;
2299 } else if (cache->uc_allocbucket) {
2300 #ifdef UMA_DEBUG_ALLOC
2301 printf("uma_zfree: Swapping buckets.\n");
2304 * We have run out of space in our freebucket.
2305 * See if we can switch with our alloc bucket.
2307 if (cache->uc_allocbucket->ub_cnt <
2308 cache->uc_freebucket->ub_cnt) {
2309 bucket = cache->uc_freebucket;
2310 cache->uc_freebucket = cache->uc_allocbucket;
2311 cache->uc_allocbucket = bucket;
2317 * We can get here for two reasons:
2319 * 1) The buckets are NULL
2320 * 2) The alloc and free buckets are both somewhat full.
2322 * We must go back the zone, which requires acquiring the zone lock,
2323 * which in turn means we must release and re-acquire the critical
2324 * section. Since the critical section is released, we may be
2325 * preempted or migrate. As such, make sure not to maintain any
2326 * thread-local state specific to the cache from prior to releasing
2327 * the critical section.
2333 cache = &zone->uz_cpu[cpu];
2334 if (cache->uc_freebucket != NULL) {
2335 if (cache->uc_freebucket->ub_cnt <
2336 cache->uc_freebucket->ub_entries) {
2340 if (cache->uc_allocbucket != NULL &&
2341 (cache->uc_allocbucket->ub_cnt <
2342 cache->uc_freebucket->ub_cnt)) {
2348 bucket = cache->uc_freebucket;
2349 cache->uc_freebucket = NULL;
2351 /* Can we throw this on the zone full list? */
2352 if (bucket != NULL) {
2353 #ifdef UMA_DEBUG_ALLOC
2354 printf("uma_zfree: Putting old bucket on the free list.\n");
2356 /* ub_cnt is pointing to the last free item */
2357 KASSERT(bucket->ub_cnt != 0,
2358 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2359 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2362 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2363 LIST_REMOVE(bucket, ub_link);
2365 cache->uc_freebucket = bucket;
2368 /* We are no longer associated with this CPU. */
2371 /* And the zone.. */
2374 #ifdef UMA_DEBUG_ALLOC
2375 printf("uma_zfree: Allocating new free bucket.\n");
2379 if (keg->uk_flags & UMA_ZFLAG_CACHEONLY)
2381 bucket = bucket_alloc(zone->uz_count, bflags);
2384 LIST_INSERT_HEAD(&zone->uz_free_bucket,
2391 * If nothing else caught this, we'll just do an internal free.
2394 uma_zfree_internal(zone, item, udata, SKIP_DTOR);
2400 * Frees an item to an INTERNAL zone or allocates a free bucket
2403 * zone The zone to free to
2404 * item The item we're freeing
2405 * udata User supplied data for the dtor
2406 * skip Skip dtors and finis
2409 uma_zfree_internal(uma_zone_t zone, void *item, void *udata,
2410 enum zfreeskip skip)
2413 uma_slabrefcnt_t slabref;
2420 if (skip < SKIP_DTOR && zone->uz_dtor)
2421 zone->uz_dtor(item, keg->uk_size, udata);
2422 if (skip < SKIP_FINI && zone->uz_fini)
2423 zone->uz_fini(item, keg->uk_size);
2427 if (!(keg->uk_flags & UMA_ZONE_MALLOC)) {
2428 mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK));
2429 if (keg->uk_flags & UMA_ZONE_HASH)
2430 slab = hash_sfind(&keg->uk_hash, mem);
2432 mem += keg->uk_pgoff;
2433 slab = (uma_slab_t)mem;
2436 slab = (uma_slab_t)udata;
2439 /* Do we need to remove from any lists? */
2440 if (slab->us_freecount+1 == keg->uk_ipers) {
2441 LIST_REMOVE(slab, us_link);
2442 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2443 } else if (slab->us_freecount == 0) {
2444 LIST_REMOVE(slab, us_link);
2445 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2448 /* Slab management stuff */
2449 freei = ((unsigned long)item - (unsigned long)slab->us_data)
2454 uma_dbg_free(zone, slab, item);
2457 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2458 slabref = (uma_slabrefcnt_t)slab;
2459 slabref->us_freelist[freei].us_item = slab->us_firstfree;
2461 slab->us_freelist[freei].us_item = slab->us_firstfree;
2463 slab->us_firstfree = freei;
2464 slab->us_freecount++;
2466 /* Zone statistics */
2469 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2470 if (keg->uk_pages < keg->uk_maxpages)
2471 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2473 /* We can handle one more allocation */
2482 uma_zone_set_max(uma_zone_t zone, int nitems)
2488 if (keg->uk_ppera > 1)
2489 keg->uk_maxpages = nitems * keg->uk_ppera;
2491 keg->uk_maxpages = nitems / keg->uk_ipers;
2493 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2501 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2504 KASSERT(zone->uz_keg->uk_pages == 0,
2505 ("uma_zone_set_init on non-empty keg"));
2506 zone->uz_keg->uk_init = uminit;
2512 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2515 KASSERT(zone->uz_keg->uk_pages == 0,
2516 ("uma_zone_set_fini on non-empty keg"));
2517 zone->uz_keg->uk_fini = fini;
2523 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2526 KASSERT(zone->uz_keg->uk_pages == 0,
2527 ("uma_zone_set_zinit on non-empty keg"));
2528 zone->uz_init = zinit;
2534 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2537 KASSERT(zone->uz_keg->uk_pages == 0,
2538 ("uma_zone_set_zfini on non-empty keg"));
2539 zone->uz_fini = zfini;
2544 /* XXX uk_freef is not actually used with the zone locked */
2546 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2549 zone->uz_keg->uk_freef = freef;
2554 /* XXX uk_allocf is not actually used with the zone locked */
2556 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2559 zone->uz_keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2560 zone->uz_keg->uk_allocf = allocf;
2566 uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count)
2573 pages = count / keg->uk_ipers;
2575 if (pages * keg->uk_ipers < count)
2578 kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2583 obj = vm_object_allocate(OBJT_DEFAULT,
2586 VM_OBJECT_LOCK_INIT(obj, "uma object");
2587 _vm_object_allocate(OBJT_DEFAULT,
2593 keg->uk_maxpages = pages;
2594 keg->uk_allocf = obj_alloc;
2595 keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
2602 uma_prealloc(uma_zone_t zone, int items)
2610 slabs = items / keg->uk_ipers;
2611 if (slabs * keg->uk_ipers < items)
2614 slab = slab_zalloc(zone, M_WAITOK);
2615 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2623 uma_find_refcnt(uma_zone_t zone, void *item)
2625 uma_slabrefcnt_t slabref;
2631 slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item &
2633 KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT,
2634 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
2635 idx = ((unsigned long)item - (unsigned long)slabref->us_data)
2637 refcnt = &slabref->us_freelist[idx].us_refcnt;
2646 printf("UMA: vm asked us to release pages!\n");
2649 zone_foreach(zone_drain);
2651 * Some slabs may have been freed but this zone will be visited early
2652 * we visit again so that we can free pages that are empty once other
2653 * zones are drained. We have to do the same for buckets.
2655 zone_drain(slabzone);
2656 zone_drain(slabrefzone);
2657 bucket_zone_drain();
2661 uma_large_malloc(int size, int wait)
2667 slab = uma_zalloc_internal(slabzone, NULL, wait);
2670 mem = page_alloc(NULL, size, &flags, wait);
2672 vsetslab((vm_offset_t)mem, slab);
2673 slab->us_data = mem;
2674 slab->us_flags = flags | UMA_SLAB_MALLOC;
2675 slab->us_size = size;
2677 uma_zfree_internal(slabzone, slab, NULL, SKIP_NONE);
2684 uma_large_free(uma_slab_t slab)
2686 vsetobj((vm_offset_t)slab->us_data, kmem_object);
2687 page_free(slab->us_data, slab->us_size, slab->us_flags);
2688 uma_zfree_internal(slabzone, slab, NULL, SKIP_NONE);
2692 uma_print_stats(void)
2694 zone_foreach(uma_print_zone);
2698 slab_print(uma_slab_t slab)
2700 printf("slab: keg %p, data %p, freecount %d, firstfree %d\n",
2701 slab->us_keg, slab->us_data, slab->us_freecount,
2702 slab->us_firstfree);
2706 cache_print(uma_cache_t cache)
2708 printf("alloc: %p(%d), free: %p(%d)\n",
2709 cache->uc_allocbucket,
2710 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
2711 cache->uc_freebucket,
2712 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
2716 uma_print_zone(uma_zone_t zone)
2724 printf("%s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n",
2725 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
2726 keg->uk_ipers, keg->uk_ppera,
2727 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
2728 printf("Part slabs:\n");
2729 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
2731 printf("Free slabs:\n");
2732 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
2734 printf("Full slabs:\n");
2735 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
2737 for (i = 0; i <= mp_maxid; i++) {
2740 cache = &zone->uz_cpu[i];
2741 printf("CPU %d Cache:\n", i);
2747 * Sysctl handler for vm.zone
2749 * stolen from vm_zone.c
2752 sysctl_vm_zone(SYSCTL_HANDLER_ARGS)
2754 int error, len, cnt;
2755 const int linesize = 128; /* conservative */
2757 char *tmpbuf, *offset;
2763 uma_bucket_t bucket;
2769 LIST_FOREACH(zk, &uma_kegs, uk_link) {
2770 LIST_FOREACH(z, &zk->uk_zones, uz_link)
2773 mtx_unlock(&uma_mtx);
2774 MALLOC(tmpbuf, char *, (cnt == 0 ? 1 : cnt) * linesize,
2776 len = snprintf(tmpbuf, linesize,
2777 "\nITEM SIZE LIMIT USED FREE REQUESTS\n\n");
2779 tmpbuf[len - 1] = '\0';
2780 error = SYSCTL_OUT(req, tmpbuf, cnt == 0 ? len-1 : len);
2781 if (error || cnt == 0)
2785 LIST_FOREACH(zk, &uma_kegs, uk_link) {
2786 LIST_FOREACH(z, &zk->uk_zones, uz_link) {
2787 if (cnt == 0) /* list may have changed size */
2792 if (!(zk->uk_flags & UMA_ZFLAG_INTERNAL)) {
2793 for (cpu = 0; cpu <= mp_maxid; cpu++) {
2794 if (CPU_ABSENT(cpu))
2796 cache = &z->uz_cpu[cpu];
2797 if (cache->uc_allocbucket != NULL)
2798 cachefree += cache->uc_allocbucket->ub_cnt;
2799 if (cache->uc_freebucket != NULL)
2800 cachefree += cache->uc_freebucket->ub_cnt;
2801 alloc += cache->uc_allocs;
2802 cache->uc_allocs = 0;
2805 alloc += z->uz_allocs;
2807 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link) {
2808 cachefree += bucket->ub_cnt;
2810 totalfree = zk->uk_free + cachefree;
2811 len = snprintf(offset, linesize,
2812 "%-12.12s %6.6u, %8.8u, %6.6u, %6.6u, %8.8llu\n",
2813 z->uz_name, zk->uk_size,
2814 zk->uk_maxpages * zk->uk_ipers,
2815 (zk->uk_ipers * (zk->uk_pages / zk->uk_ppera)) - totalfree,
2817 (unsigned long long)alloc);
2819 for (p = offset + 12; p > offset && *p == ' '; --p)
2826 mtx_unlock(&uma_mtx);
2828 error = SYSCTL_OUT(req, tmpbuf, offset - tmpbuf);
2830 FREE(tmpbuf, M_TEMP);