2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2002-2019 Jeffrey Roberson <jeff@FreeBSD.org>
5 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
6 * Copyright (c) 2004-2006 Robert N. M. Watson
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice unmodified, this list of conditions, and the following
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
20 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
21 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
22 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
24 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
28 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32 * uma_core.c Implementation of the Universal Memory allocator
34 * This allocator is intended to replace the multitude of similar object caches
35 * in the standard FreeBSD kernel. The intent is to be flexible as well as
36 * efficient. A primary design goal is to return unused memory to the rest of
37 * the system. This will make the system as a whole more flexible due to the
38 * ability to move memory to subsystems which most need it instead of leaving
39 * pools of reserved memory unused.
41 * The basic ideas stem from similar slab/zone based allocators whose algorithms
48 * - Improve memory usage for large allocations
49 * - Investigate cache size adjustments
52 #include <sys/cdefs.h>
53 __FBSDID("$FreeBSD$");
56 #include "opt_param.h"
59 #include <sys/param.h>
60 #include <sys/systm.h>
61 #include <sys/bitset.h>
62 #include <sys/domainset.h>
63 #include <sys/eventhandler.h>
64 #include <sys/kernel.h>
65 #include <sys/types.h>
66 #include <sys/limits.h>
67 #include <sys/queue.h>
68 #include <sys/malloc.h>
71 #include <sys/sysctl.h>
72 #include <sys/mutex.h>
74 #include <sys/random.h>
75 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
81 #include <sys/taskqueue.h>
82 #include <sys/vmmeter.h>
85 #include <vm/vm_param.h>
86 #include <vm/vm_domainset.h>
87 #include <vm/vm_object.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_pageout.h>
90 #include <vm/vm_phys.h>
91 #include <vm/vm_pagequeue.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
95 #include <vm/vm_dumpset.h>
97 #include <vm/uma_int.h>
98 #include <vm/uma_dbg.h>
102 #ifdef DEBUG_MEMGUARD
103 #include <vm/memguard.h>
106 #include <machine/md_var.h>
109 #define UMA_ALWAYS_CTORDTOR 1
111 #define UMA_ALWAYS_CTORDTOR 0
115 * This is the zone and keg from which all zones are spawned.
117 static uma_zone_t kegs;
118 static uma_zone_t zones;
121 * On INVARIANTS builds, the slab contains a second bitset of the same size,
122 * "dbg_bits", which is laid out immediately after us_free.
125 #define SLAB_BITSETS 2
127 #define SLAB_BITSETS 1
131 * These are the two zones from which all offpage uma_slab_ts are allocated.
133 * One zone is for slab headers that can represent a larger number of items,
134 * making the slabs themselves more efficient, and the other zone is for
135 * headers that are smaller and represent fewer items, making the headers more
138 #define SLABZONE_SIZE(setsize) \
139 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
140 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16)
141 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE
142 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE)
143 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE)
144 static uma_zone_t slabzones[2];
147 * The initial hash tables come out of this zone so they can be allocated
148 * prior to malloc coming up.
150 static uma_zone_t hashzone;
152 /* The boot-time adjusted value for cache line alignment. */
153 int uma_align_cache = 64 - 1;
155 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
156 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
159 * Are we allowed to allocate buckets?
161 static int bucketdisable = 1;
163 /* Linked list of all kegs in the system */
164 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
166 /* Linked list of all cache-only zones in the system */
167 static LIST_HEAD(,uma_zone) uma_cachezones =
168 LIST_HEAD_INITIALIZER(uma_cachezones);
170 /* This RW lock protects the keg list */
171 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
174 * First available virual address for boot time allocations.
176 static vm_offset_t bootstart;
177 static vm_offset_t bootmem;
179 static struct sx uma_reclaim_lock;
182 * kmem soft limit, initialized by uma_set_limit(). Ensure that early
183 * allocations don't trigger a wakeup of the reclaim thread.
185 unsigned long uma_kmem_limit = LONG_MAX;
186 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
187 "UMA kernel memory soft limit");
188 unsigned long uma_kmem_total;
189 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
190 "UMA kernel memory usage");
192 /* Is the VM done starting up? */
199 } booted = BOOT_COLD;
202 * This is the handle used to schedule events that need to happen
203 * outside of the allocation fast path.
205 static struct callout uma_callout;
206 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
209 * This structure is passed as the zone ctor arg so that I don't have to create
210 * a special allocation function just for zones.
212 struct uma_zctor_args {
227 struct uma_kctor_args {
236 struct uma_bucket_zone {
238 const char *ubz_name;
239 int ubz_entries; /* Number of items it can hold. */
240 int ubz_maxsize; /* Maximum allocation size per-item. */
244 * Compute the actual number of bucket entries to pack them in power
245 * of two sizes for more efficient space utilization.
247 #define BUCKET_SIZE(n) \
248 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
250 #define BUCKET_MAX BUCKET_SIZE(256)
253 struct uma_bucket_zone bucket_zones[] = {
254 /* Literal bucket sizes. */
255 { NULL, "2 Bucket", 2, 4096 },
256 { NULL, "4 Bucket", 4, 3072 },
257 { NULL, "8 Bucket", 8, 2048 },
258 { NULL, "16 Bucket", 16, 1024 },
259 /* Rounded down power of 2 sizes for efficiency. */
260 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
261 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
262 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
263 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
268 * Flags and enumerations to be passed to internal functions.
272 SKIP_CNT = 0x00000001,
273 SKIP_DTOR = 0x00010000,
274 SKIP_FINI = 0x00020000,
279 void uma_startup1(vm_offset_t);
280 void uma_startup2(void);
282 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
283 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
284 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
285 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
286 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
287 static void page_free(void *, vm_size_t, uint8_t);
288 static void pcpu_page_free(void *, vm_size_t, uint8_t);
289 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
290 static void cache_drain(uma_zone_t);
291 static void bucket_drain(uma_zone_t, uma_bucket_t);
292 static void bucket_cache_reclaim(uma_zone_t zone, bool);
293 static int keg_ctor(void *, int, void *, int);
294 static void keg_dtor(void *, int, void *);
295 static int zone_ctor(void *, int, void *, int);
296 static void zone_dtor(void *, int, void *);
297 static inline void item_dtor(uma_zone_t zone, void *item, int size,
298 void *udata, enum zfreeskip skip);
299 static int zero_init(void *, int, int);
300 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
301 int itemdomain, bool ws);
302 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
303 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
304 static void zone_timeout(uma_zone_t zone, void *);
305 static int hash_alloc(struct uma_hash *, u_int);
306 static int hash_expand(struct uma_hash *, struct uma_hash *);
307 static void hash_free(struct uma_hash *hash);
308 static void uma_timeout(void *);
309 static void uma_shutdown(void);
310 static void *zone_alloc_item(uma_zone_t, void *, int, int);
311 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
312 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
313 static void zone_free_limit(uma_zone_t zone, int count);
314 static void bucket_enable(void);
315 static void bucket_init(void);
316 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
317 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
318 static void bucket_zone_drain(void);
319 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
320 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
321 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
322 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
323 uma_fini fini, int align, uint32_t flags);
324 static int zone_import(void *, void **, int, int, int);
325 static void zone_release(void *, void **, int);
326 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
327 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
329 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
330 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
331 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
332 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
333 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
334 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
335 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
337 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
339 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
340 "Memory allocation debugging");
343 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
344 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
346 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
347 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
348 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
349 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
351 static u_int dbg_divisor = 1;
352 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
353 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
354 "Debug & thrash every this item in memory allocator");
356 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
357 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
358 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
359 &uma_dbg_cnt, "memory items debugged");
360 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
361 &uma_skip_cnt, "memory items skipped, not debugged");
364 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
365 "Universal Memory Allocator");
367 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
368 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
370 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
371 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
373 static int zone_warnings = 1;
374 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
375 "Warn when UMA zones becomes full");
377 static int multipage_slabs = 1;
378 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
379 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
380 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
381 "UMA may choose larger slab sizes for better efficiency");
384 * Select the slab zone for an offpage slab with the given maximum item count.
386 static inline uma_zone_t
390 return (slabzones[ipers > SLABZONE0_SETSIZE]);
394 * This routine checks to see whether or not it's safe to enable buckets.
400 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
401 bucketdisable = vm_page_count_min();
405 * Initialize bucket_zones, the array of zones of buckets of various sizes.
407 * For each zone, calculate the memory required for each bucket, consisting
408 * of the header and an array of pointers.
413 struct uma_bucket_zone *ubz;
416 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
417 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
418 size += sizeof(void *) * ubz->ubz_entries;
419 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
420 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
421 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
422 UMA_ZONE_FIRSTTOUCH);
427 * Given a desired number of entries for a bucket, return the zone from which
428 * to allocate the bucket.
430 static struct uma_bucket_zone *
431 bucket_zone_lookup(int entries)
433 struct uma_bucket_zone *ubz;
435 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
436 if (ubz->ubz_entries >= entries)
442 static struct uma_bucket_zone *
443 bucket_zone_max(uma_zone_t zone, int nitems)
445 struct uma_bucket_zone *ubz;
449 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
450 /* Count the cross-domain bucket. */
453 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
454 if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems)
456 if (ubz == &bucket_zones[0])
464 bucket_select(int size)
466 struct uma_bucket_zone *ubz;
468 ubz = &bucket_zones[0];
469 if (size > ubz->ubz_maxsize)
470 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
472 for (; ubz->ubz_entries != 0; ubz++)
473 if (ubz->ubz_maxsize < size)
476 return (ubz->ubz_entries);
480 bucket_alloc(uma_zone_t zone, void *udata, int flags)
482 struct uma_bucket_zone *ubz;
486 * Don't allocate buckets early in boot.
488 if (__predict_false(booted < BOOT_KVA))
492 * To limit bucket recursion we store the original zone flags
493 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
494 * NOVM flag to persist even through deep recursions. We also
495 * store ZFLAG_BUCKET once we have recursed attempting to allocate
496 * a bucket for a bucket zone so we do not allow infinite bucket
497 * recursion. This cookie will even persist to frees of unused
498 * buckets via the allocation path or bucket allocations in the
501 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
502 udata = (void *)(uintptr_t)zone->uz_flags;
504 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
506 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
508 if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
510 ubz = bucket_zone_lookup(zone->uz_bucket_size);
511 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
513 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
516 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
519 bucket->ub_entries = ubz->ubz_entries;
520 bucket->ub_seq = SMR_SEQ_INVALID;
521 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
522 zone->uz_name, zone, bucket);
529 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
531 struct uma_bucket_zone *ubz;
533 if (bucket->ub_cnt != 0)
534 bucket_drain(zone, bucket);
536 KASSERT(bucket->ub_cnt == 0,
537 ("bucket_free: Freeing a non free bucket."));
538 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
539 ("bucket_free: Freeing an SMR bucket."));
540 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
541 udata = (void *)(uintptr_t)zone->uz_flags;
542 ubz = bucket_zone_lookup(bucket->ub_entries);
543 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
547 bucket_zone_drain(void)
549 struct uma_bucket_zone *ubz;
551 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
552 uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN);
556 * Acquire the domain lock and record contention.
558 static uma_zone_domain_t
559 zone_domain_lock(uma_zone_t zone, int domain)
561 uma_zone_domain_t zdom;
564 zdom = ZDOM_GET(zone, domain);
566 if (ZDOM_OWNED(zdom))
569 /* This is unsynchronized. The counter does not need to be precise. */
570 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
571 zone->uz_bucket_size++;
576 * Search for the domain with the least cached items and return it if it
577 * is out of balance with the preferred domain.
579 static __noinline int
580 zone_domain_lowest(uma_zone_t zone, int pref)
582 long least, nitems, prefitems;
586 prefitems = least = LONG_MAX;
588 for (i = 0; i < vm_ndomains; i++) {
589 nitems = ZDOM_GET(zone, i)->uzd_nitems;
590 if (nitems < least) {
597 if (prefitems < least * 2)
604 * Search for the domain with the most cached items and return it or the
605 * preferred domain if it has enough to proceed.
607 static __noinline int
608 zone_domain_highest(uma_zone_t zone, int pref)
614 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
619 for (i = 0; i < vm_ndomains; i++) {
620 nitems = ZDOM_GET(zone, i)->uzd_nitems;
631 * Safely subtract cnt from imax.
634 zone_domain_imax_sub(uma_zone_domain_t zdom, int cnt)
639 old = zdom->uzd_imax;
645 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, new) == 0);
649 * Set the maximum imax value.
652 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
656 old = zdom->uzd_imax;
660 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
664 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
665 * zone's caches. If a bucket is found the zone is not locked on return.
668 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
674 ZDOM_LOCK_ASSERT(zdom);
676 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
679 /* SMR Buckets can not be re-used until readers expire. */
680 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
681 bucket->ub_seq != SMR_SEQ_INVALID) {
682 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
684 bucket->ub_seq = SMR_SEQ_INVALID;
685 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
686 if (STAILQ_NEXT(bucket, ub_link) != NULL)
687 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
689 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
691 KASSERT(zdom->uzd_nitems >= bucket->ub_cnt,
692 ("%s: item count underflow (%ld, %d)",
693 __func__, zdom->uzd_nitems, bucket->ub_cnt));
694 KASSERT(bucket->ub_cnt > 0,
695 ("%s: empty bucket in bucket cache", __func__));
696 zdom->uzd_nitems -= bucket->ub_cnt;
699 * Shift the bounds of the current WSS interval to avoid
700 * perturbing the estimate.
703 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
704 zone_domain_imax_sub(zdom, bucket->ub_cnt);
705 } else if (zdom->uzd_imin > zdom->uzd_nitems)
706 zdom->uzd_imin = zdom->uzd_nitems;
710 for (i = 0; i < bucket->ub_cnt; i++)
711 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
718 * Insert a full bucket into the specified cache. The "ws" parameter indicates
719 * whether the bucket's contents should be counted as part of the zone's working
720 * set. The bucket may be freed if it exceeds the bucket limit.
723 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
726 uma_zone_domain_t zdom;
728 /* We don't cache empty buckets. This can happen after a reclaim. */
729 if (bucket->ub_cnt == 0)
731 zdom = zone_domain_lock(zone, domain);
734 * Conditionally set the maximum number of items.
736 zdom->uzd_nitems += bucket->ub_cnt;
737 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
739 zone_domain_imax_set(zdom, zdom->uzd_nitems);
740 if (STAILQ_EMPTY(&zdom->uzd_buckets))
741 zdom->uzd_seq = bucket->ub_seq;
744 * Try to promote reuse of recently used items. For items
745 * protected by SMR, try to defer reuse to minimize polling.
747 if (bucket->ub_seq == SMR_SEQ_INVALID)
748 STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
750 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
754 zdom->uzd_nitems -= bucket->ub_cnt;
757 bucket_free(zone, bucket, udata);
760 /* Pops an item out of a per-cpu cache bucket. */
762 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
766 CRITICAL_ASSERT(curthread);
769 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
771 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
772 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
779 /* Pushes an item into a per-cpu cache bucket. */
781 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
784 CRITICAL_ASSERT(curthread);
785 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
786 ("uma_zfree: Freeing to non free bucket index."));
788 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
794 * Unload a UMA bucket from a per-cpu cache.
796 static inline uma_bucket_t
797 cache_bucket_unload(uma_cache_bucket_t bucket)
801 b = bucket->ucb_bucket;
803 MPASS(b->ub_entries == bucket->ucb_entries);
804 b->ub_cnt = bucket->ucb_cnt;
805 bucket->ucb_bucket = NULL;
806 bucket->ucb_entries = bucket->ucb_cnt = 0;
812 static inline uma_bucket_t
813 cache_bucket_unload_alloc(uma_cache_t cache)
816 return (cache_bucket_unload(&cache->uc_allocbucket));
819 static inline uma_bucket_t
820 cache_bucket_unload_free(uma_cache_t cache)
823 return (cache_bucket_unload(&cache->uc_freebucket));
826 static inline uma_bucket_t
827 cache_bucket_unload_cross(uma_cache_t cache)
830 return (cache_bucket_unload(&cache->uc_crossbucket));
834 * Load a bucket into a per-cpu cache bucket.
837 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
840 CRITICAL_ASSERT(curthread);
841 MPASS(bucket->ucb_bucket == NULL);
842 MPASS(b->ub_seq == SMR_SEQ_INVALID);
844 bucket->ucb_bucket = b;
845 bucket->ucb_cnt = b->ub_cnt;
846 bucket->ucb_entries = b->ub_entries;
850 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
853 cache_bucket_load(&cache->uc_allocbucket, b);
857 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
860 cache_bucket_load(&cache->uc_freebucket, b);
865 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
868 cache_bucket_load(&cache->uc_crossbucket, b);
873 * Copy and preserve ucb_spare.
876 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
879 b1->ucb_bucket = b2->ucb_bucket;
880 b1->ucb_entries = b2->ucb_entries;
881 b1->ucb_cnt = b2->ucb_cnt;
885 * Swap two cache buckets.
888 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
890 struct uma_cache_bucket b3;
892 CRITICAL_ASSERT(curthread);
894 cache_bucket_copy(&b3, b1);
895 cache_bucket_copy(b1, b2);
896 cache_bucket_copy(b2, &b3);
900 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
903 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
905 uma_zone_domain_t zdom;
909 * Avoid the lock if possible.
911 zdom = ZDOM_GET(zone, domain);
912 if (zdom->uzd_nitems == 0)
915 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
916 !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
920 * Check the zone's cache of buckets.
922 zdom = zone_domain_lock(zone, domain);
923 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL)
931 zone_log_warning(uma_zone_t zone)
933 static const struct timeval warninterval = { 300, 0 };
935 if (!zone_warnings || zone->uz_warning == NULL)
938 if (ratecheck(&zone->uz_ratecheck, &warninterval))
939 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
943 zone_maxaction(uma_zone_t zone)
946 if (zone->uz_maxaction.ta_func != NULL)
947 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
951 * Routine called by timeout which is used to fire off some time interval
952 * based calculations. (stats, hash size, etc.)
961 uma_timeout(void *unused)
964 zone_foreach(zone_timeout, NULL);
966 /* Reschedule this event */
967 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
971 * Update the working set size estimate for the zone's bucket cache.
972 * The constants chosen here are somewhat arbitrary. With an update period of
973 * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
977 zone_domain_update_wss(uma_zone_domain_t zdom)
982 MPASS(zdom->uzd_imax >= zdom->uzd_imin);
983 wss = zdom->uzd_imax - zdom->uzd_imin;
984 zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
985 zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
990 * Routine to perform timeout driven calculations. This expands the
991 * hashes and does per cpu statistics aggregation.
996 zone_timeout(uma_zone_t zone, void *unused)
1001 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
1007 * Hash zones are non-numa by definition so the first domain
1008 * is the only one present.
1011 pages = keg->uk_domain[0].ud_pages;
1014 * Expand the keg hash table.
1016 * This is done if the number of slabs is larger than the hash size.
1017 * What I'm trying to do here is completely reduce collisions. This
1018 * may be a little aggressive. Should I allow for two collisions max?
1020 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1021 struct uma_hash newhash;
1022 struct uma_hash oldhash;
1026 * This is so involved because allocating and freeing
1027 * while the keg lock is held will lead to deadlock.
1028 * I have to do everything in stages and check for
1032 ret = hash_alloc(&newhash, 1 << fls(slabs));
1035 if (hash_expand(&keg->uk_hash, &newhash)) {
1036 oldhash = keg->uk_hash;
1037 keg->uk_hash = newhash;
1042 hash_free(&oldhash);
1049 for (int i = 0; i < vm_ndomains; i++)
1050 zone_domain_update_wss(ZDOM_GET(zone, i));
1054 * Allocate and zero fill the next sized hash table from the appropriate
1058 * hash A new hash structure with the old hash size in uh_hashsize
1061 * 1 on success and 0 on failure.
1064 hash_alloc(struct uma_hash *hash, u_int size)
1068 KASSERT(powerof2(size), ("hash size must be power of 2"));
1069 if (size > UMA_HASH_SIZE_INIT) {
1070 hash->uh_hashsize = size;
1071 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1072 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1074 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1075 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1076 UMA_ANYDOMAIN, M_WAITOK);
1077 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1079 if (hash->uh_slab_hash) {
1080 bzero(hash->uh_slab_hash, alloc);
1081 hash->uh_hashmask = hash->uh_hashsize - 1;
1089 * Expands the hash table for HASH zones. This is done from zone_timeout
1090 * to reduce collisions. This must not be done in the regular allocation
1091 * path, otherwise, we can recurse on the vm while allocating pages.
1094 * oldhash The hash you want to expand
1095 * newhash The hash structure for the new table
1103 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1105 uma_hash_slab_t slab;
1109 if (!newhash->uh_slab_hash)
1112 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1116 * I need to investigate hash algorithms for resizing without a
1120 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1121 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1122 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1123 LIST_REMOVE(slab, uhs_hlink);
1124 hval = UMA_HASH(newhash, slab->uhs_data);
1125 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1133 * Free the hash bucket to the appropriate backing store.
1136 * slab_hash The hash bucket we're freeing
1137 * hashsize The number of entries in that hash bucket
1143 hash_free(struct uma_hash *hash)
1145 if (hash->uh_slab_hash == NULL)
1147 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1148 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1150 free(hash->uh_slab_hash, M_UMAHASH);
1154 * Frees all outstanding items in a bucket
1157 * zone The zone to free to, must be unlocked.
1158 * bucket The free/alloc bucket with items.
1164 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1168 if (bucket->ub_cnt == 0)
1171 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1172 bucket->ub_seq != SMR_SEQ_INVALID) {
1173 smr_wait(zone->uz_smr, bucket->ub_seq);
1174 bucket->ub_seq = SMR_SEQ_INVALID;
1175 for (i = 0; i < bucket->ub_cnt; i++)
1176 item_dtor(zone, bucket->ub_bucket[i],
1177 zone->uz_size, NULL, SKIP_NONE);
1180 for (i = 0; i < bucket->ub_cnt; i++)
1181 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1182 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1183 if (zone->uz_max_items > 0)
1184 zone_free_limit(zone, bucket->ub_cnt);
1186 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1192 * Drains the per cpu caches for a zone.
1194 * NOTE: This may only be called while the zone is being torn down, and not
1195 * during normal operation. This is necessary in order that we do not have
1196 * to migrate CPUs to drain the per-CPU caches.
1199 * zone The zone to drain, must be unlocked.
1205 cache_drain(uma_zone_t zone)
1208 uma_bucket_t bucket;
1213 * XXX: It is safe to not lock the per-CPU caches, because we're
1214 * tearing down the zone anyway. I.e., there will be no further use
1215 * of the caches at this point.
1217 * XXX: It would good to be able to assert that the zone is being
1218 * torn down to prevent improper use of cache_drain().
1220 seq = SMR_SEQ_INVALID;
1221 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1222 seq = smr_advance(zone->uz_smr);
1224 cache = &zone->uz_cpu[cpu];
1225 bucket = cache_bucket_unload_alloc(cache);
1227 bucket_free(zone, bucket, NULL);
1228 bucket = cache_bucket_unload_free(cache);
1229 if (bucket != NULL) {
1230 bucket->ub_seq = seq;
1231 bucket_free(zone, bucket, NULL);
1233 bucket = cache_bucket_unload_cross(cache);
1234 if (bucket != NULL) {
1235 bucket->ub_seq = seq;
1236 bucket_free(zone, bucket, NULL);
1239 bucket_cache_reclaim(zone, true);
1243 cache_shrink(uma_zone_t zone, void *unused)
1246 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1249 zone->uz_bucket_size =
1250 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1254 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1257 uma_bucket_t b1, b2, b3;
1260 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1263 b1 = b2 = b3 = NULL;
1265 cache = &zone->uz_cpu[curcpu];
1266 domain = PCPU_GET(domain);
1267 b1 = cache_bucket_unload_alloc(cache);
1270 * Don't flush SMR zone buckets. This leaves the zone without a
1271 * bucket and forces every free to synchronize().
1273 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1274 b2 = cache_bucket_unload_free(cache);
1275 b3 = cache_bucket_unload_cross(cache);
1280 zone_free_bucket(zone, b1, NULL, domain, false);
1282 zone_free_bucket(zone, b2, NULL, domain, false);
1284 /* Adjust the domain so it goes to zone_free_cross. */
1285 domain = (domain + 1) % vm_ndomains;
1286 zone_free_bucket(zone, b3, NULL, domain, false);
1291 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1292 * This is an expensive call because it needs to bind to all CPUs
1293 * one by one and enter a critical section on each of them in order
1294 * to safely access their cache buckets.
1295 * Zone lock must not be held on call this function.
1298 pcpu_cache_drain_safe(uma_zone_t zone)
1303 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1306 cache_shrink(zone, NULL);
1308 zone_foreach(cache_shrink, NULL);
1311 thread_lock(curthread);
1312 sched_bind(curthread, cpu);
1313 thread_unlock(curthread);
1316 cache_drain_safe_cpu(zone, NULL);
1318 zone_foreach(cache_drain_safe_cpu, NULL);
1320 thread_lock(curthread);
1321 sched_unbind(curthread);
1322 thread_unlock(curthread);
1326 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1327 * requested a drain, otherwise the per-domain caches are trimmed to either
1328 * estimated working set size.
1331 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1333 uma_zone_domain_t zdom;
1334 uma_bucket_t bucket;
1339 * Shrink the zone bucket size to ensure that the per-CPU caches
1340 * don't grow too large.
1342 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1343 zone->uz_bucket_size--;
1345 for (i = 0; i < vm_ndomains; i++) {
1347 * The cross bucket is partially filled and not part of
1348 * the item count. Reclaim it individually here.
1350 zdom = ZDOM_GET(zone, i);
1351 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1352 ZONE_CROSS_LOCK(zone);
1353 bucket = zdom->uzd_cross;
1354 zdom->uzd_cross = NULL;
1355 ZONE_CROSS_UNLOCK(zone);
1357 bucket_free(zone, bucket, NULL);
1361 * If we were asked to drain the zone, we are done only once
1362 * this bucket cache is empty. Otherwise, we reclaim items in
1363 * excess of the zone's estimated working set size. If the
1364 * difference nitems - imin is larger than the WSS estimate,
1365 * then the estimate will grow at the end of this interval and
1366 * we ignore the historical average.
1369 target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1371 while (zdom->uzd_nitems > target) {
1372 bucket = zone_fetch_bucket(zone, zdom, true);
1375 bucket_free(zone, bucket, NULL);
1383 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1389 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1390 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1392 mem = slab_data(slab, keg);
1393 flags = slab->us_flags;
1395 if (keg->uk_fini != NULL) {
1396 for (i--; i > -1; i--)
1399 * trash_fini implies that dtor was trash_dtor. trash_fini
1400 * would check that memory hasn't been modified since free,
1401 * which executed trash_dtor.
1402 * That's why we need to run uma_dbg_kskip() check here,
1403 * albeit we don't make skip check for other init/fini
1406 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1407 keg->uk_fini != trash_fini)
1409 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1411 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1412 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1414 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1415 uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1419 * Frees pages from a keg back to the system. This is done on demand from
1420 * the pageout daemon.
1425 keg_drain(uma_keg_t keg)
1427 struct slabhead freeslabs;
1429 uma_slab_t slab, tmp;
1432 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1435 for (i = 0; i < vm_ndomains; i++) {
1436 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1437 keg->uk_name, keg, i, dom->ud_free_items);
1438 dom = &keg->uk_domain[i];
1439 LIST_INIT(&freeslabs);
1442 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1443 LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
1444 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1446 n = dom->ud_free_slabs;
1447 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1448 dom->ud_free_slabs = 0;
1449 dom->ud_free_items -= n * keg->uk_ipers;
1450 dom->ud_pages -= n * keg->uk_ppera;
1453 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1454 keg_free_slab(keg, slab, keg->uk_ipers);
1459 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1463 * Set draining to interlock with zone_dtor() so we can release our
1464 * locks as we go. Only dtor() should do a WAITOK call since it
1465 * is the only call that knows the structure will still be available
1469 while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1470 if (waitok == M_NOWAIT)
1472 msleep(zone, &ZDOM_GET(zone, 0)->uzd_lock, PVM, "zonedrain",
1475 zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1477 bucket_cache_reclaim(zone, drain);
1480 * The DRAINING flag protects us from being freed while
1481 * we're running. Normally the uma_rwlock would protect us but we
1482 * must be able to release and acquire the right lock for each keg.
1484 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1485 keg_drain(zone->uz_keg);
1487 zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1494 zone_drain(uma_zone_t zone, void *unused)
1497 zone_reclaim(zone, M_NOWAIT, true);
1501 zone_trim(uma_zone_t zone, void *unused)
1504 zone_reclaim(zone, M_NOWAIT, false);
1508 * Allocate a new slab for a keg and inserts it into the partial slab list.
1509 * The keg should be unlocked on entry. If the allocation succeeds it will
1510 * be locked on return.
1513 * flags Wait flags for the item initialization routine
1514 * aflags Wait flags for the slab allocation
1517 * The slab that was allocated or NULL if there is no memory and the
1518 * caller specified M_NOWAIT.
1521 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1532 KASSERT(domain >= 0 && domain < vm_ndomains,
1533 ("keg_alloc_slab: domain %d out of range", domain));
1535 allocf = keg->uk_allocf;
1538 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1539 uma_hash_slab_t hslab;
1540 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1544 slab = &hslab->uhs_slab;
1548 * This reproduces the old vm_zone behavior of zero filling pages the
1549 * first time they are added to a zone.
1551 * Malloced items are zeroed in uma_zalloc.
1554 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1559 if (keg->uk_flags & UMA_ZONE_NODUMP)
1562 /* zone is passed for legacy reasons. */
1563 size = keg->uk_ppera * PAGE_SIZE;
1564 mem = allocf(zone, size, domain, &sflags, aflags);
1566 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1567 zone_free_item(slabzone(keg->uk_ipers),
1568 slab_tohashslab(slab), NULL, SKIP_NONE);
1571 uma_total_inc(size);
1573 /* For HASH zones all pages go to the same uma_domain. */
1574 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1577 /* Point the slab into the allocated memory */
1578 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1579 slab = (uma_slab_t )(mem + keg->uk_pgoff);
1581 slab_tohashslab(slab)->uhs_data = mem;
1583 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1584 for (i = 0; i < keg->uk_ppera; i++)
1585 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1588 slab->us_freecount = keg->uk_ipers;
1589 slab->us_flags = sflags;
1590 slab->us_domain = domain;
1592 BIT_FILL(keg->uk_ipers, &slab->us_free);
1594 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1597 if (keg->uk_init != NULL) {
1598 for (i = 0; i < keg->uk_ipers; i++)
1599 if (keg->uk_init(slab_item(slab, keg, i),
1600 keg->uk_size, flags) != 0)
1602 if (i != keg->uk_ipers) {
1603 keg_free_slab(keg, slab, i);
1607 KEG_LOCK(keg, domain);
1609 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1610 slab, keg->uk_name, keg);
1612 if (keg->uk_flags & UMA_ZFLAG_HASH)
1613 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1616 * If we got a slab here it's safe to mark it partially used
1617 * and return. We assume that the caller is going to remove
1618 * at least one item.
1620 dom = &keg->uk_domain[domain];
1621 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1622 dom->ud_pages += keg->uk_ppera;
1623 dom->ud_free_items += keg->uk_ipers;
1632 * This function is intended to be used early on in place of page_alloc() so
1633 * that we may use the boot time page cache to satisfy allocations before
1637 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1646 pages = howmany(bytes, PAGE_SIZE);
1647 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1649 *pflag = UMA_SLAB_BOOT;
1650 m = vm_page_alloc_contig_domain(NULL, 0, domain,
1651 malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1652 (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1656 pa = VM_PAGE_TO_PHYS(m);
1657 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1658 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1659 defined(__riscv) || defined(__powerpc64__)
1660 if ((wait & M_NODUMP) == 0)
1664 /* Allocate KVA and indirectly advance bootmem. */
1665 mem = (void *)pmap_map(&bootmem, m->phys_addr,
1666 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1667 if ((wait & M_ZERO) != 0)
1668 bzero(mem, pages * PAGE_SIZE);
1674 startup_free(void *mem, vm_size_t bytes)
1679 va = (vm_offset_t)mem;
1680 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1681 pmap_remove(kernel_pmap, va, va + bytes);
1682 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1683 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1684 defined(__riscv) || defined(__powerpc64__)
1685 dump_drop_page(VM_PAGE_TO_PHYS(m));
1687 vm_page_unwire_noq(m);
1693 * Allocates a number of pages from the system
1696 * bytes The number of bytes requested
1697 * wait Shall we wait?
1700 * A pointer to the alloced memory or possibly
1701 * NULL if M_NOWAIT is set.
1704 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1707 void *p; /* Returned page */
1709 *pflag = UMA_SLAB_KERNEL;
1710 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1716 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1719 struct pglist alloctail;
1720 vm_offset_t addr, zkva;
1722 vm_page_t p, p_next;
1727 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1729 TAILQ_INIT(&alloctail);
1730 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1731 malloc2vm_flags(wait);
1732 *pflag = UMA_SLAB_KERNEL;
1733 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1734 if (CPU_ABSENT(cpu)) {
1735 p = vm_page_alloc(NULL, 0, flags);
1738 p = vm_page_alloc(NULL, 0, flags);
1740 pc = pcpu_find(cpu);
1741 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1744 p = vm_page_alloc_domain(NULL, 0,
1745 pc->pc_domain, flags);
1746 if (__predict_false(p == NULL))
1747 p = vm_page_alloc(NULL, 0, flags);
1750 if (__predict_false(p == NULL))
1752 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1754 if ((addr = kva_alloc(bytes)) == 0)
1757 TAILQ_FOREACH(p, &alloctail, listq) {
1758 pmap_qenter(zkva, &p, 1);
1761 return ((void*)addr);
1763 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1764 vm_page_unwire_noq(p);
1771 * Allocates a number of pages from within an object
1774 * bytes The number of bytes requested
1775 * wait Shall we wait?
1778 * A pointer to the alloced memory or possibly
1779 * NULL if M_NOWAIT is set.
1782 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1785 TAILQ_HEAD(, vm_page) alloctail;
1787 vm_offset_t retkva, zkva;
1788 vm_page_t p, p_next;
1791 TAILQ_INIT(&alloctail);
1794 npages = howmany(bytes, PAGE_SIZE);
1795 while (npages > 0) {
1796 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1797 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1798 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1802 * Since the page does not belong to an object, its
1805 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1810 * Page allocation failed, free intermediate pages and
1813 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1814 vm_page_unwire_noq(p);
1819 *flags = UMA_SLAB_PRIV;
1820 zkva = keg->uk_kva +
1821 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1823 TAILQ_FOREACH(p, &alloctail, listq) {
1824 pmap_qenter(zkva, &p, 1);
1828 return ((void *)retkva);
1832 * Allocate physically contiguous pages.
1835 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1839 *pflag = UMA_SLAB_KERNEL;
1840 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1841 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1845 * Frees a number of pages to the system
1848 * mem A pointer to the memory to be freed
1849 * size The size of the memory being freed
1850 * flags The original p->us_flags field
1856 page_free(void *mem, vm_size_t size, uint8_t flags)
1859 if ((flags & UMA_SLAB_BOOT) != 0) {
1860 startup_free(mem, size);
1864 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1865 ("UMA: page_free used with invalid flags %x", flags));
1867 kmem_free((vm_offset_t)mem, size);
1871 * Frees pcpu zone allocations
1874 * mem A pointer to the memory to be freed
1875 * size The size of the memory being freed
1876 * flags The original p->us_flags field
1882 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1884 vm_offset_t sva, curva;
1888 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1890 if ((flags & UMA_SLAB_BOOT) != 0) {
1891 startup_free(mem, size);
1895 sva = (vm_offset_t)mem;
1896 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1897 paddr = pmap_kextract(curva);
1898 m = PHYS_TO_VM_PAGE(paddr);
1899 vm_page_unwire_noq(m);
1902 pmap_qremove(sva, size >> PAGE_SHIFT);
1903 kva_free(sva, size);
1907 * Zero fill initializer
1909 * Arguments/Returns follow uma_init specifications
1912 zero_init(void *mem, int size, int flags)
1919 static struct noslabbits *
1920 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1923 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1928 * Actual size of embedded struct slab (!OFFPAGE).
1931 slab_sizeof(int nitems)
1935 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1936 return (roundup(s, UMA_ALIGN_PTR + 1));
1939 #define UMA_FIXPT_SHIFT 31
1940 #define UMA_FRAC_FIXPT(n, d) \
1941 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1942 #define UMA_FIXPT_PCT(f) \
1943 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1944 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
1945 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1948 * Compute the number of items that will fit in a slab. If hdr is true, the
1949 * item count may be limited to provide space in the slab for an inline slab
1950 * header. Otherwise, all slab space will be provided for item storage.
1953 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1958 /* The padding between items is not needed after the last item. */
1959 padpi = rsize - size;
1963 * Start with the maximum item count and remove items until
1964 * the slab header first alongside the allocatable memory.
1966 for (ipers = MIN(SLAB_MAX_SETSIZE,
1967 (slabsize + padpi - slab_sizeof(1)) / rsize);
1969 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1973 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1979 struct keg_layout_result {
1987 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
1988 struct keg_layout_result *kl)
1993 kl->slabsize = slabsize;
1995 /* Handle INTERNAL as inline with an extra page. */
1996 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
1997 kl->format &= ~UMA_ZFLAG_INTERNAL;
1998 kl->slabsize += PAGE_SIZE;
2001 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2002 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2004 /* Account for memory used by an offpage slab header. */
2005 total = kl->slabsize;
2006 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2007 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2009 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2013 * Determine the format of a uma keg. This determines where the slab header
2014 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2017 * keg The zone we should initialize
2023 keg_layout(uma_keg_t keg)
2025 struct keg_layout_result kl = {}, kl_tmp;
2034 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2035 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2036 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2037 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2038 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2040 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2041 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2042 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2045 alignsize = keg->uk_align + 1;
2048 * Calculate the size of each allocation (rsize) according to
2049 * alignment. If the requested size is smaller than we have
2050 * allocation bits for we round it up.
2052 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2053 rsize = roundup2(rsize, alignsize);
2055 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2057 * We want one item to start on every align boundary in a page.
2058 * To do this we will span pages. We will also extend the item
2059 * by the size of align if it is an even multiple of align.
2060 * Otherwise, it would fall on the same boundary every time.
2062 if ((rsize & alignsize) == 0)
2064 slabsize = rsize * (PAGE_SIZE / alignsize);
2065 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2066 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2067 slabsize = round_page(slabsize);
2070 * Start with a slab size of as many pages as it takes to
2071 * represent a single item. We will try to fit as many
2072 * additional items into the slab as possible.
2074 slabsize = round_page(keg->uk_size);
2077 /* Build a list of all of the available formats for this keg. */
2080 /* Evaluate an inline slab layout. */
2081 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2084 /* TODO: vm_page-embedded slab. */
2087 * We can't do OFFPAGE if we're internal or if we've been
2088 * asked to not go to the VM for buckets. If we do this we
2089 * may end up going to the VM for slabs which we do not want
2090 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2091 * In those cases, evaluate a pseudo-format called INTERNAL
2092 * which has an inline slab header and one extra page to
2093 * guarantee that it fits.
2095 * Otherwise, see if using an OFFPAGE slab will improve our
2098 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2099 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2101 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2104 * Choose a slab size and format which satisfy the minimum efficiency.
2105 * Prefer the smallest slab size that meets the constraints.
2107 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2108 * for small items (up to PAGE_SIZE), the iteration increment is one
2109 * page; and for large items, the increment is one item.
2111 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2112 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2113 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2116 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2117 round_page(rsize * (i - 1) + keg->uk_size);
2119 for (j = 0; j < nfmt; j++) {
2120 /* Only if we have no viable format yet. */
2121 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2125 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2126 if (kl_tmp.eff <= kl.eff)
2131 CTR6(KTR_UMA, "keg %s layout: format %#x "
2132 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2133 keg->uk_name, kl.format, kl.ipers, rsize,
2134 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2136 /* Stop when we reach the minimum efficiency. */
2137 if (kl.eff >= UMA_MIN_EFF)
2141 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2142 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2143 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2147 pages = atop(kl.slabsize);
2148 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2149 pages *= mp_maxid + 1;
2151 keg->uk_rsize = rsize;
2152 keg->uk_ipers = kl.ipers;
2153 keg->uk_ppera = pages;
2154 keg->uk_flags |= kl.format;
2157 * How do we find the slab header if it is offpage or if not all item
2158 * start addresses are in the same page? We could solve the latter
2159 * case with vaddr alignment, but we don't.
2161 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2162 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2163 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2164 keg->uk_flags |= UMA_ZFLAG_HASH;
2166 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2169 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2170 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2172 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2173 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2174 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2175 keg->uk_ipers, pages));
2179 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2180 * the keg onto the global keg list.
2182 * Arguments/Returns follow uma_ctor specifications
2183 * udata Actually uma_kctor_args
2186 keg_ctor(void *mem, int size, void *udata, int flags)
2188 struct uma_kctor_args *arg = udata;
2189 uma_keg_t keg = mem;
2194 keg->uk_size = arg->size;
2195 keg->uk_init = arg->uminit;
2196 keg->uk_fini = arg->fini;
2197 keg->uk_align = arg->align;
2198 keg->uk_reserve = 0;
2199 keg->uk_flags = arg->flags;
2202 * We use a global round-robin policy by default. Zones with
2203 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2204 * case the iterator is never run.
2206 keg->uk_dr.dr_policy = DOMAINSET_RR();
2207 keg->uk_dr.dr_iter = 0;
2210 * The primary zone is passed to us at keg-creation time.
2213 keg->uk_name = zone->uz_name;
2215 if (arg->flags & UMA_ZONE_ZINIT)
2216 keg->uk_init = zero_init;
2218 if (arg->flags & UMA_ZONE_MALLOC)
2219 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2222 keg->uk_flags &= ~UMA_ZONE_PCPU;
2228 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2229 * work on. Use round-robin for everything else.
2231 * Zones may override the default by specifying either.
2234 if ((keg->uk_flags &
2235 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2236 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2237 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2238 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2242 * If we haven't booted yet we need allocations to go through the
2243 * startup cache until the vm is ready.
2245 #ifdef UMA_MD_SMALL_ALLOC
2246 if (keg->uk_ppera == 1)
2247 keg->uk_allocf = uma_small_alloc;
2250 if (booted < BOOT_KVA)
2251 keg->uk_allocf = startup_alloc;
2252 else if (keg->uk_flags & UMA_ZONE_PCPU)
2253 keg->uk_allocf = pcpu_page_alloc;
2254 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2255 keg->uk_allocf = contig_alloc;
2257 keg->uk_allocf = page_alloc;
2258 #ifdef UMA_MD_SMALL_ALLOC
2259 if (keg->uk_ppera == 1)
2260 keg->uk_freef = uma_small_free;
2263 if (keg->uk_flags & UMA_ZONE_PCPU)
2264 keg->uk_freef = pcpu_page_free;
2266 keg->uk_freef = page_free;
2269 * Initialize keg's locks.
2271 for (i = 0; i < vm_ndomains; i++)
2272 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2275 * If we're putting the slab header in the actual page we need to
2276 * figure out where in each page it goes. See slab_sizeof
2279 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2282 shsize = slab_sizeof(keg->uk_ipers);
2283 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2285 * The only way the following is possible is if with our
2286 * UMA_ALIGN_PTR adjustments we are now bigger than
2287 * UMA_SLAB_SIZE. I haven't checked whether this is
2288 * mathematically possible for all cases, so we make
2291 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2292 ("zone %s ipers %d rsize %d size %d slab won't fit",
2293 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2296 if (keg->uk_flags & UMA_ZFLAG_HASH)
2297 hash_alloc(&keg->uk_hash, 0);
2299 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2301 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2303 rw_wlock(&uma_rwlock);
2304 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2305 rw_wunlock(&uma_rwlock);
2310 zone_kva_available(uma_zone_t zone, void *unused)
2314 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2318 if (keg->uk_allocf == startup_alloc) {
2319 /* Switch to the real allocator. */
2320 if (keg->uk_flags & UMA_ZONE_PCPU)
2321 keg->uk_allocf = pcpu_page_alloc;
2322 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2324 keg->uk_allocf = contig_alloc;
2326 keg->uk_allocf = page_alloc;
2331 zone_alloc_counters(uma_zone_t zone, void *unused)
2334 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2335 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2336 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2337 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2341 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2343 uma_zone_domain_t zdom;
2346 struct sysctl_oid *oid, *domainoid;
2347 int domains, i, cnt;
2348 static const char *nokeg = "cache zone";
2352 * Make a sysctl safe copy of the zone name by removing
2353 * any special characters and handling dups by appending
2356 if (zone->uz_namecnt != 0) {
2357 /* Count the number of decimal digits and '_' separator. */
2358 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2360 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2362 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2365 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2366 for (c = zone->uz_ctlname; *c != '\0'; c++)
2367 if (strchr("./\\ -", *c) != NULL)
2371 * Basic parameters at the root.
2373 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2374 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2376 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2377 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2378 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2379 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2380 zone, 0, sysctl_handle_uma_zone_flags, "A",
2381 "Allocator configuration flags");
2382 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2383 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2384 "Desired per-cpu cache size");
2385 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2386 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2387 "Maximum allowed per-cpu cache size");
2392 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2393 domains = vm_ndomains;
2396 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2397 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2399 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2400 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2401 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2402 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2403 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2404 "Real object size with alignment");
2405 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2406 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2407 "pages per-slab allocation");
2408 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2409 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2410 "items available per-slab");
2411 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2412 "align", CTLFLAG_RD, &keg->uk_align, 0,
2413 "item alignment mask");
2414 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2415 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2416 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2417 "Slab utilization (100 - internal fragmentation %)");
2418 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2419 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2420 for (i = 0; i < domains; i++) {
2421 dom = &keg->uk_domain[i];
2422 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2423 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2424 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2425 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2426 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2427 "Total pages currently allocated from VM");
2428 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2429 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2430 "items free in the slab layer");
2433 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2434 "name", CTLFLAG_RD, nokeg, "Keg name");
2437 * Information about zone limits.
2439 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2440 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2441 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2442 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2443 zone, 0, sysctl_handle_uma_zone_items, "QU",
2444 "current number of allocated items if limit is set");
2445 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2446 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2447 "Maximum number of cached items");
2448 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2449 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2450 "Number of threads sleeping at limit");
2451 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2452 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2453 "Total zone limit sleeps");
2454 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2455 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2456 "Maximum number of items in each domain's bucket cache");
2459 * Per-domain zone information.
2461 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2462 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2463 for (i = 0; i < domains; i++) {
2464 zdom = ZDOM_GET(zone, i);
2465 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2466 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2467 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2468 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2469 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2470 "number of items in this domain");
2471 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2472 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2473 "maximum item count in this period");
2474 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2475 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2476 "minimum item count in this period");
2477 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2478 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2479 "Working set size");
2483 * General statistics.
2485 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2486 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2487 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2488 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2489 zone, 1, sysctl_handle_uma_zone_cur, "I",
2490 "Current number of allocated items");
2491 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2492 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2493 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2494 "Total allocation calls");
2495 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2496 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2497 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2498 "Total free calls");
2499 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2500 "fails", CTLFLAG_RD, &zone->uz_fails,
2501 "Number of allocation failures");
2502 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2503 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2504 "Free calls from the wrong domain");
2507 struct uma_zone_count {
2513 zone_count(uma_zone_t zone, void *arg)
2515 struct uma_zone_count *cnt;
2519 * Some zones are rapidly created with identical names and
2520 * destroyed out of order. This can lead to gaps in the count.
2521 * Use one greater than the maximum observed for this name.
2523 if (strcmp(zone->uz_name, cnt->name) == 0)
2524 cnt->count = MAX(cnt->count,
2525 zone->uz_namecnt + 1);
2529 zone_update_caches(uma_zone_t zone)
2533 for (i = 0; i <= mp_maxid; i++) {
2534 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2535 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2540 * Zone header ctor. This initializes all fields, locks, etc.
2542 * Arguments/Returns follow uma_ctor specifications
2543 * udata Actually uma_zctor_args
2546 zone_ctor(void *mem, int size, void *udata, int flags)
2548 struct uma_zone_count cnt;
2549 struct uma_zctor_args *arg = udata;
2550 uma_zone_domain_t zdom;
2551 uma_zone_t zone = mem;
2557 zone->uz_name = arg->name;
2558 zone->uz_ctor = arg->ctor;
2559 zone->uz_dtor = arg->dtor;
2560 zone->uz_init = NULL;
2561 zone->uz_fini = NULL;
2562 zone->uz_sleeps = 0;
2563 zone->uz_bucket_size = 0;
2564 zone->uz_bucket_size_min = 0;
2565 zone->uz_bucket_size_max = BUCKET_MAX;
2566 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2567 zone->uz_warning = NULL;
2568 /* The domain structures follow the cpu structures. */
2569 zone->uz_bucket_max = ULONG_MAX;
2570 timevalclear(&zone->uz_ratecheck);
2572 /* Count the number of duplicate names. */
2573 cnt.name = arg->name;
2575 zone_foreach(zone_count, &cnt);
2576 zone->uz_namecnt = cnt.count;
2577 ZONE_CROSS_LOCK_INIT(zone);
2579 for (i = 0; i < vm_ndomains; i++) {
2580 zdom = ZDOM_GET(zone, i);
2581 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2582 STAILQ_INIT(&zdom->uzd_buckets);
2586 if (arg->uminit == trash_init && arg->fini == trash_fini)
2587 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2591 * This is a pure cache zone, no kegs.
2594 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2595 ("zone_ctor: Import specified for non-cache zone."));
2596 zone->uz_flags = arg->flags;
2597 zone->uz_size = arg->size;
2598 zone->uz_import = arg->import;
2599 zone->uz_release = arg->release;
2600 zone->uz_arg = arg->arg;
2603 * Cache zones are round-robin unless a policy is
2604 * specified because they may have incompatible
2607 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2608 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2610 rw_wlock(&uma_rwlock);
2611 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2612 rw_wunlock(&uma_rwlock);
2617 * Use the regular zone/keg/slab allocator.
2619 zone->uz_import = zone_import;
2620 zone->uz_release = zone_release;
2621 zone->uz_arg = zone;
2624 if (arg->flags & UMA_ZONE_SECONDARY) {
2625 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2626 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2627 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2628 zone->uz_init = arg->uminit;
2629 zone->uz_fini = arg->fini;
2630 zone->uz_flags |= UMA_ZONE_SECONDARY;
2631 rw_wlock(&uma_rwlock);
2633 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2634 if (LIST_NEXT(z, uz_link) == NULL) {
2635 LIST_INSERT_AFTER(z, zone, uz_link);
2640 rw_wunlock(&uma_rwlock);
2641 } else if (keg == NULL) {
2642 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2643 arg->align, arg->flags)) == NULL)
2646 struct uma_kctor_args karg;
2649 /* We should only be here from uma_startup() */
2650 karg.size = arg->size;
2651 karg.uminit = arg->uminit;
2652 karg.fini = arg->fini;
2653 karg.align = arg->align;
2654 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2656 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2662 /* Inherit properties from the keg. */
2664 zone->uz_size = keg->uk_size;
2665 zone->uz_flags |= (keg->uk_flags &
2666 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2669 if (booted >= BOOT_PCPU) {
2670 zone_alloc_counters(zone, NULL);
2671 if (booted >= BOOT_RUNNING)
2672 zone_alloc_sysctl(zone, NULL);
2674 zone->uz_allocs = EARLY_COUNTER;
2675 zone->uz_frees = EARLY_COUNTER;
2676 zone->uz_fails = EARLY_COUNTER;
2679 /* Caller requests a private SMR context. */
2680 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2681 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2683 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2684 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2685 ("Invalid zone flag combination"));
2686 if (arg->flags & UMA_ZFLAG_INTERNAL)
2687 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2688 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2689 zone->uz_bucket_size = BUCKET_MAX;
2690 else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2691 zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2692 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2693 zone->uz_bucket_size = 0;
2695 zone->uz_bucket_size = bucket_select(zone->uz_size);
2696 zone->uz_bucket_size_min = zone->uz_bucket_size;
2697 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2698 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2699 zone_update_caches(zone);
2705 * Keg header dtor. This frees all data, destroys locks, frees the hash
2706 * table and removes the keg from the global list.
2708 * Arguments/Returns follow uma_dtor specifications
2712 keg_dtor(void *arg, int size, void *udata)
2715 uint32_t free, pages;
2718 keg = (uma_keg_t)arg;
2720 for (i = 0; i < vm_ndomains; i++) {
2721 free += keg->uk_domain[i].ud_free_items;
2722 pages += keg->uk_domain[i].ud_pages;
2723 KEG_LOCK_FINI(keg, i);
2726 printf("Freed UMA keg (%s) was not empty (%u items). "
2727 " Lost %u pages of memory.\n",
2728 keg->uk_name ? keg->uk_name : "",
2729 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2731 hash_free(&keg->uk_hash);
2737 * Arguments/Returns follow uma_dtor specifications
2741 zone_dtor(void *arg, int size, void *udata)
2747 zone = (uma_zone_t)arg;
2749 sysctl_remove_oid(zone->uz_oid, 1, 1);
2751 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2754 rw_wlock(&uma_rwlock);
2755 LIST_REMOVE(zone, uz_link);
2756 rw_wunlock(&uma_rwlock);
2757 zone_reclaim(zone, M_WAITOK, true);
2760 * We only destroy kegs from non secondary/non cache zones.
2762 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2764 rw_wlock(&uma_rwlock);
2765 LIST_REMOVE(keg, uk_link);
2766 rw_wunlock(&uma_rwlock);
2767 zone_free_item(kegs, keg, NULL, SKIP_NONE);
2769 counter_u64_free(zone->uz_allocs);
2770 counter_u64_free(zone->uz_frees);
2771 counter_u64_free(zone->uz_fails);
2772 counter_u64_free(zone->uz_xdomain);
2773 free(zone->uz_ctlname, M_UMA);
2774 for (i = 0; i < vm_ndomains; i++)
2775 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2776 ZONE_CROSS_LOCK_FINI(zone);
2780 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2785 LIST_FOREACH(keg, &uma_kegs, uk_link) {
2786 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2789 LIST_FOREACH(zone, &uma_cachezones, uz_link)
2794 * Traverses every zone in the system and calls a callback
2797 * zfunc A pointer to a function which accepts a zone
2804 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2807 rw_rlock(&uma_rwlock);
2808 zone_foreach_unlocked(zfunc, arg);
2809 rw_runlock(&uma_rwlock);
2813 * Initialize the kernel memory allocator. This is done after pages can be
2814 * allocated but before general KVA is available.
2817 uma_startup1(vm_offset_t virtual_avail)
2819 struct uma_zctor_args args;
2820 size_t ksize, zsize, size;
2821 uma_keg_t primarykeg;
2826 bootstart = bootmem = virtual_avail;
2828 rw_init(&uma_rwlock, "UMA lock");
2829 sx_init(&uma_reclaim_lock, "umareclaim");
2831 ksize = sizeof(struct uma_keg) +
2832 (sizeof(struct uma_domain) * vm_ndomains);
2833 ksize = roundup(ksize, UMA_SUPER_ALIGN);
2834 zsize = sizeof(struct uma_zone) +
2835 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2836 (sizeof(struct uma_zone_domain) * vm_ndomains);
2837 zsize = roundup(zsize, UMA_SUPER_ALIGN);
2839 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2840 size = (zsize * 2) + ksize;
2841 for (domain = 0; domain < vm_ndomains; domain++) {
2842 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
2847 zones = (uma_zone_t)m;
2849 kegs = (uma_zone_t)m;
2851 primarykeg = (uma_keg_t)m;
2853 /* "manually" create the initial zone */
2854 memset(&args, 0, sizeof(args));
2855 args.name = "UMA Kegs";
2857 args.ctor = keg_ctor;
2858 args.dtor = keg_dtor;
2859 args.uminit = zero_init;
2861 args.keg = primarykeg;
2862 args.align = UMA_SUPER_ALIGN - 1;
2863 args.flags = UMA_ZFLAG_INTERNAL;
2864 zone_ctor(kegs, zsize, &args, M_WAITOK);
2866 args.name = "UMA Zones";
2868 args.ctor = zone_ctor;
2869 args.dtor = zone_dtor;
2870 args.uminit = zero_init;
2873 args.align = UMA_SUPER_ALIGN - 1;
2874 args.flags = UMA_ZFLAG_INTERNAL;
2875 zone_ctor(zones, zsize, &args, M_WAITOK);
2877 /* Now make zones for slab headers */
2878 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2879 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2880 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2881 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2883 hashzone = uma_zcreate("UMA Hash",
2884 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2885 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2891 #ifndef UMA_MD_SMALL_ALLOC
2892 extern void vm_radix_reserve_kva(void);
2896 * Advertise the availability of normal kva allocations and switch to
2897 * the default back-end allocator. Marks the KVA we consumed on startup
2898 * as used in the map.
2904 if (bootstart != bootmem) {
2905 vm_map_lock(kernel_map);
2906 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2907 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2908 vm_map_unlock(kernel_map);
2911 #ifndef UMA_MD_SMALL_ALLOC
2912 /* Set up radix zone to use noobj_alloc. */
2913 vm_radix_reserve_kva();
2917 zone_foreach_unlocked(zone_kva_available, NULL);
2922 * Allocate counters as early as possible so that boot-time allocations are
2923 * accounted more precisely.
2926 uma_startup_pcpu(void *arg __unused)
2929 zone_foreach_unlocked(zone_alloc_counters, NULL);
2932 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
2935 * Finish our initialization steps.
2938 uma_startup3(void *arg __unused)
2942 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2943 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2944 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2946 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2947 callout_init(&uma_callout, 1);
2948 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2949 booted = BOOT_RUNNING;
2951 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2952 EVENTHANDLER_PRI_FIRST);
2954 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
2960 booted = BOOT_SHUTDOWN;
2964 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2965 int align, uint32_t flags)
2967 struct uma_kctor_args args;
2970 args.uminit = uminit;
2972 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2975 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2978 /* Public functions */
2981 uma_set_align(int align)
2984 if (align != UMA_ALIGN_CACHE)
2985 uma_align_cache = align;
2990 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2991 uma_init uminit, uma_fini fini, int align, uint32_t flags)
2994 struct uma_zctor_args args;
2997 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
3000 /* This stuff is essential for the zone ctor */
3001 memset(&args, 0, sizeof(args));
3006 args.uminit = uminit;
3010 * Inject procedures which check for memory use after free if we are
3011 * allowed to scramble the memory while it is not allocated. This
3012 * requires that: UMA is actually able to access the memory, no init
3013 * or fini procedures, no dependency on the initial value of the
3014 * memory, and no (legitimate) use of the memory after free. Note,
3015 * the ctor and dtor do not need to be empty.
3017 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3018 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3019 args.uminit = trash_init;
3020 args.fini = trash_fini;
3027 sx_slock(&uma_reclaim_lock);
3028 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3029 sx_sunlock(&uma_reclaim_lock);
3036 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3037 uma_init zinit, uma_fini zfini, uma_zone_t primary)
3039 struct uma_zctor_args args;
3043 keg = primary->uz_keg;
3044 memset(&args, 0, sizeof(args));
3046 args.size = keg->uk_size;
3049 args.uminit = zinit;
3051 args.align = keg->uk_align;
3052 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3055 sx_slock(&uma_reclaim_lock);
3056 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3057 sx_sunlock(&uma_reclaim_lock);
3064 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3065 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3066 void *arg, int flags)
3068 struct uma_zctor_args args;
3070 memset(&args, 0, sizeof(args));
3075 args.uminit = zinit;
3077 args.import = zimport;
3078 args.release = zrelease;
3081 args.flags = flags | UMA_ZFLAG_CACHE;
3083 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3088 uma_zdestroy(uma_zone_t zone)
3092 * Large slabs are expensive to reclaim, so don't bother doing
3093 * unnecessary work if we're shutting down.
3095 if (booted == BOOT_SHUTDOWN &&
3096 zone->uz_fini == NULL && zone->uz_release == zone_release)
3098 sx_slock(&uma_reclaim_lock);
3099 zone_free_item(zones, zone, NULL, SKIP_NONE);
3100 sx_sunlock(&uma_reclaim_lock);
3104 uma_zwait(uma_zone_t zone)
3107 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3108 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3109 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3110 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3112 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3116 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3118 void *item, *pcpu_item;
3122 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3124 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3127 pcpu_item = zpcpu_base_to_offset(item);
3128 if (flags & M_ZERO) {
3130 for (i = 0; i <= mp_maxid; i++)
3131 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3133 bzero(item, zone->uz_size);
3140 * A stub while both regular and pcpu cases are identical.
3143 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3148 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3150 item = zpcpu_offset_to_base(pcpu_item);
3151 uma_zfree_arg(zone, item, udata);
3154 static inline void *
3155 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3161 skipdbg = uma_dbg_zskip(zone, item);
3162 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3163 zone->uz_ctor != trash_ctor)
3164 trash_ctor(item, size, udata, flags);
3166 /* Check flags before loading ctor pointer. */
3167 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3168 __predict_false(zone->uz_ctor != NULL) &&
3169 zone->uz_ctor(item, size, udata, flags) != 0) {
3170 counter_u64_add(zone->uz_fails, 1);
3171 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3176 uma_dbg_alloc(zone, NULL, item);
3178 if (__predict_false(flags & M_ZERO))
3179 return (memset(item, 0, size));
3185 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3186 enum zfreeskip skip)
3191 skipdbg = uma_dbg_zskip(zone, item);
3192 if (skip == SKIP_NONE && !skipdbg) {
3193 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3194 uma_dbg_free(zone, udata, item);
3196 uma_dbg_free(zone, NULL, item);
3199 if (__predict_true(skip < SKIP_DTOR)) {
3200 if (zone->uz_dtor != NULL)
3201 zone->uz_dtor(item, size, udata);
3203 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3204 zone->uz_dtor != trash_dtor)
3205 trash_dtor(item, size, udata);
3212 item_domain(void *item)
3216 domain = _vm_phys_domain(vtophys(item));
3217 KASSERT(domain >= 0 && domain < vm_ndomains,
3218 ("%s: unknown domain for item %p", __func__, item));
3223 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3224 #define UMA_ZALLOC_DEBUG
3226 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3232 if (flags & M_WAITOK) {
3233 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3234 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3239 KASSERT((flags & M_EXEC) == 0,
3240 ("uma_zalloc_debug: called with M_EXEC"));
3241 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3242 ("uma_zalloc_debug: called within spinlock or critical section"));
3243 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3244 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3247 #ifdef DEBUG_MEMGUARD
3248 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3250 item = memguard_alloc(zone->uz_size, flags);
3252 error = EJUSTRETURN;
3253 if (zone->uz_init != NULL &&
3254 zone->uz_init(item, zone->uz_size, flags) != 0) {
3258 if (zone->uz_ctor != NULL &&
3259 zone->uz_ctor(item, zone->uz_size, udata,
3261 counter_u64_add(zone->uz_fails, 1);
3262 zone->uz_fini(item, zone->uz_size);
3269 /* This is unfortunate but should not be fatal. */
3276 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3278 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3279 ("uma_zfree_debug: called with spinlock or critical section held"));
3281 #ifdef DEBUG_MEMGUARD
3282 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3283 if (zone->uz_dtor != NULL)
3284 zone->uz_dtor(item, zone->uz_size, udata);
3285 if (zone->uz_fini != NULL)
3286 zone->uz_fini(item, zone->uz_size);
3287 memguard_free(item);
3288 return (EJUSTRETURN);
3295 static inline void *
3296 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3297 void *udata, int flags)
3302 item = cache_bucket_pop(cache, bucket);
3303 size = cache_uz_size(cache);
3304 uz_flags = cache_uz_flags(cache);
3306 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3309 static __noinline void *
3310 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3312 uma_cache_bucket_t bucket;
3315 while (cache_alloc(zone, cache, udata, flags)) {
3316 cache = &zone->uz_cpu[curcpu];
3317 bucket = &cache->uc_allocbucket;
3318 if (__predict_false(bucket->ucb_cnt == 0))
3320 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3325 * We can not get a bucket so try to return a single item.
3327 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3328 domain = PCPU_GET(domain);
3330 domain = UMA_ANYDOMAIN;
3331 return (zone_alloc_item(zone, udata, domain, flags));
3336 uma_zalloc_smr(uma_zone_t zone, int flags)
3338 uma_cache_bucket_t bucket;
3341 #ifdef UMA_ZALLOC_DEBUG
3344 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3345 ("uma_zalloc_arg: called with non-SMR zone."));
3346 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3351 cache = &zone->uz_cpu[curcpu];
3352 bucket = &cache->uc_allocbucket;
3353 if (__predict_false(bucket->ucb_cnt == 0))
3354 return (cache_alloc_retry(zone, cache, NULL, flags));
3355 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3360 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3362 uma_cache_bucket_t bucket;
3365 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3366 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3368 /* This is the fast path allocation */
3369 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3372 #ifdef UMA_ZALLOC_DEBUG
3375 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3376 ("uma_zalloc_arg: called with SMR zone."));
3377 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3382 * If possible, allocate from the per-CPU cache. There are two
3383 * requirements for safe access to the per-CPU cache: (1) the thread
3384 * accessing the cache must not be preempted or yield during access,
3385 * and (2) the thread must not migrate CPUs without switching which
3386 * cache it accesses. We rely on a critical section to prevent
3387 * preemption and migration. We release the critical section in
3388 * order to acquire the zone mutex if we are unable to allocate from
3389 * the current cache; when we re-acquire the critical section, we
3390 * must detect and handle migration if it has occurred.
3393 cache = &zone->uz_cpu[curcpu];
3394 bucket = &cache->uc_allocbucket;
3395 if (__predict_false(bucket->ucb_cnt == 0))
3396 return (cache_alloc_retry(zone, cache, udata, flags));
3397 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3401 * Replenish an alloc bucket and possibly restore an old one. Called in
3402 * a critical section. Returns in a critical section.
3404 * A false return value indicates an allocation failure.
3405 * A true return value indicates success and the caller should retry.
3407 static __noinline bool
3408 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3410 uma_bucket_t bucket;
3411 int curdomain, domain;
3414 CRITICAL_ASSERT(curthread);
3417 * If we have run out of items in our alloc bucket see
3418 * if we can switch with the free bucket.
3420 * SMR Zones can't re-use the free bucket until the sequence has
3423 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3424 cache->uc_freebucket.ucb_cnt != 0) {
3425 cache_bucket_swap(&cache->uc_freebucket,
3426 &cache->uc_allocbucket);
3431 * Discard any empty allocation bucket while we hold no locks.
3433 bucket = cache_bucket_unload_alloc(cache);
3436 if (bucket != NULL) {
3437 KASSERT(bucket->ub_cnt == 0,
3438 ("cache_alloc: Entered with non-empty alloc bucket."));
3439 bucket_free(zone, bucket, udata);
3443 * Attempt to retrieve the item from the per-CPU cache has failed, so
3444 * we must go back to the zone. This requires the zdom lock, so we
3445 * must drop the critical section, then re-acquire it when we go back
3446 * to the cache. Since the critical section is released, we may be
3447 * preempted or migrate. As such, make sure not to maintain any
3448 * thread-local state specific to the cache from prior to releasing
3449 * the critical section.
3451 domain = PCPU_GET(domain);
3452 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3453 VM_DOMAIN_EMPTY(domain))
3454 domain = zone_domain_highest(zone, domain);
3455 bucket = cache_fetch_bucket(zone, cache, domain);
3456 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3457 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3463 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3464 zone->uz_name, zone, bucket);
3465 if (bucket == NULL) {
3471 * See if we lost the race or were migrated. Cache the
3472 * initialized bucket to make this less likely or claim
3473 * the memory directly.
3476 cache = &zone->uz_cpu[curcpu];
3477 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3478 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3479 (curdomain = PCPU_GET(domain)) == domain ||
3480 VM_DOMAIN_EMPTY(curdomain))) {
3482 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3484 cache_bucket_load_alloc(cache, bucket);
3489 * We lost the race, release this bucket and start over.
3492 zone_put_bucket(zone, domain, bucket, udata, false);
3499 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3502 uma_bucket_t bucket;
3503 uma_zone_domain_t zdom;
3507 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3508 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3510 /* This is the fast path allocation */
3511 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3512 zone->uz_name, zone, domain, flags);
3514 if (flags & M_WAITOK) {
3515 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3516 "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3518 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3519 ("uma_zalloc_domain: called with spinlock or critical section held"));
3520 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3521 ("uma_zalloc_domain: called with SMR zone."));
3523 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3524 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3526 if (vm_ndomains == 1)
3527 return (uma_zalloc_arg(zone, udata, flags));
3530 * Try to allocate from the bucket cache before falling back to the keg.
3531 * We could try harder and attempt to allocate from per-CPU caches or
3532 * the per-domain cross-domain buckets, but the complexity is probably
3533 * not worth it. It is more important that frees of previous
3534 * cross-domain allocations do not blow up the cache.
3536 zdom = zone_domain_lock(zone, domain);
3537 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3538 item = bucket->ub_bucket[bucket->ub_cnt - 1];
3540 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3543 zone_put_bucket(zone, domain, bucket, udata, true);
3544 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3547 KASSERT(item_domain(item) == domain,
3548 ("%s: bucket cache item %p from wrong domain",
3550 counter_u64_add(zone->uz_allocs, 1);
3555 return (zone_alloc_item(zone, udata, domain, flags));
3557 return (uma_zalloc_arg(zone, udata, flags));
3562 * Find a slab with some space. Prefer slabs that are partially used over those
3563 * that are totally full. This helps to reduce fragmentation.
3565 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3569 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3575 KASSERT(domain >= 0 && domain < vm_ndomains,
3576 ("keg_first_slab: domain %d out of range", domain));
3577 KEG_LOCK_ASSERT(keg, domain);
3582 dom = &keg->uk_domain[domain];
3583 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3585 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3586 LIST_REMOVE(slab, us_link);
3587 dom->ud_free_slabs--;
3588 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3592 domain = (domain + 1) % vm_ndomains;
3593 } while (domain != start);
3599 * Fetch an existing slab from a free or partial list. Returns with the
3600 * keg domain lock held if a slab was found or unlocked if not.
3603 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3608 /* HASH has a single free list. */
3609 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3612 KEG_LOCK(keg, domain);
3613 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3614 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3615 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3616 KEG_UNLOCK(keg, domain);
3623 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3625 struct vm_domainset_iter di;
3632 * Use the keg's policy if upper layers haven't already specified a
3633 * domain (as happens with first-touch zones).
3635 * To avoid races we run the iterator with the keg lock held, but that
3636 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3637 * clear M_WAITOK and handle low memory conditions locally.
3639 rr = rdomain == UMA_ANYDOMAIN;
3641 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3642 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3650 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3655 * M_NOVM means don't ask at all!
3660 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3663 if (!rr && (flags & M_WAITOK) == 0)
3665 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3666 if ((flags & M_WAITOK) != 0) {
3667 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
3675 * We might not have been able to get a slab but another cpu
3676 * could have while we were unlocked. Check again before we
3679 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3686 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3692 KEG_LOCK_ASSERT(keg, slab->us_domain);
3694 dom = &keg->uk_domain[slab->us_domain];
3695 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3696 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3697 item = slab_item(slab, keg, freei);
3698 slab->us_freecount--;
3699 dom->ud_free_items--;
3702 * Move this slab to the full list. It must be on the partial list, so
3703 * we do not need to update the free slab count. In particular,
3704 * keg_fetch_slab() always returns slabs on the partial list.
3706 if (slab->us_freecount == 0) {
3707 LIST_REMOVE(slab, us_link);
3708 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3715 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3729 /* Try to keep the buckets totally full */
3730 for (i = 0; i < max; ) {
3731 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3734 stripe = howmany(max, vm_ndomains);
3736 dom = &keg->uk_domain[slab->us_domain];
3738 bucket[i++] = slab_alloc_item(keg, slab);
3739 if (dom->ud_free_items <= keg->uk_reserve) {
3741 * Avoid depleting the reserve after a
3742 * successful item allocation, even if
3743 * M_USE_RESERVE is specified.
3745 KEG_UNLOCK(keg, slab->us_domain);
3750 * If the zone is striped we pick a new slab for every
3751 * N allocations. Eliminating this conditional will
3752 * instead pick a new domain for each bucket rather
3753 * than stripe within each bucket. The current option
3754 * produces more fragmentation and requires more cpu
3755 * time but yields better distribution.
3757 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3758 vm_ndomains > 1 && --stripe == 0)
3761 } while (slab->us_freecount != 0 && i < max);
3762 KEG_UNLOCK(keg, slab->us_domain);
3764 /* Don't block if we allocated any successfully. */
3773 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3775 uint64_t old, new, total, max;
3778 * The hard case. We're going to sleep because there were existing
3779 * sleepers or because we ran out of items. This routine enforces
3780 * fairness by keeping fifo order.
3782 * First release our ill gotten gains and make some noise.
3785 zone_free_limit(zone, count);
3786 zone_log_warning(zone);
3787 zone_maxaction(zone);
3788 if (flags & M_NOWAIT)
3792 * We need to allocate an item or set ourself as a sleeper
3793 * while the sleepq lock is held to avoid wakeup races. This
3794 * is essentially a home rolled semaphore.
3796 sleepq_lock(&zone->uz_max_items);
3797 old = zone->uz_items;
3799 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3800 /* Cache the max since we will evaluate twice. */
3801 max = zone->uz_max_items;
3802 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3803 UZ_ITEMS_COUNT(old) >= max)
3804 new = old + UZ_ITEMS_SLEEPER;
3806 new = old + MIN(count, max - old);
3807 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3809 /* We may have successfully allocated under the sleepq lock. */
3810 if (UZ_ITEMS_SLEEPERS(new) == 0) {
3811 sleepq_release(&zone->uz_max_items);
3816 * This is in a different cacheline from uz_items so that we
3817 * don't constantly invalidate the fastpath cacheline when we
3818 * adjust item counts. This could be limited to toggling on
3821 atomic_add_32(&zone->uz_sleepers, 1);
3822 atomic_add_64(&zone->uz_sleeps, 1);
3825 * We have added ourselves as a sleeper. The sleepq lock
3826 * protects us from wakeup races. Sleep now and then retry.
3828 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3829 sleepq_wait(&zone->uz_max_items, PVM);
3832 * After wakeup, remove ourselves as a sleeper and try
3833 * again. We no longer have the sleepq lock for protection.
3835 * Subract ourselves as a sleeper while attempting to add
3838 atomic_subtract_32(&zone->uz_sleepers, 1);
3839 old = atomic_fetchadd_64(&zone->uz_items,
3840 -(UZ_ITEMS_SLEEPER - count));
3841 /* We're no longer a sleeper. */
3842 old -= UZ_ITEMS_SLEEPER;
3845 * If we're still at the limit, restart. Notably do not
3846 * block on other sleepers. Cache the max value to protect
3847 * against changes via sysctl.
3849 total = UZ_ITEMS_COUNT(old);
3850 max = zone->uz_max_items;
3853 /* Truncate if necessary, otherwise wake other sleepers. */
3854 if (total + count > max) {
3855 zone_free_limit(zone, total + count - max);
3856 count = max - total;
3857 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3858 wakeup_one(&zone->uz_max_items);
3865 * Allocate 'count' items from our max_items limit. Returns the number
3866 * available. If M_NOWAIT is not specified it will sleep until at least
3867 * one item can be allocated.
3870 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3875 max = zone->uz_max_items;
3879 * We expect normal allocations to succeed with a simple
3882 old = atomic_fetchadd_64(&zone->uz_items, count);
3883 if (__predict_true(old + count <= max))
3887 * If we had some items and no sleepers just return the
3888 * truncated value. We have to release the excess space
3889 * though because that may wake sleepers who weren't woken
3890 * because we were temporarily over the limit.
3893 zone_free_limit(zone, (old + count) - max);
3896 return (zone_alloc_limit_hard(zone, count, flags));
3900 * Free a number of items back to the limit.
3903 zone_free_limit(uma_zone_t zone, int count)
3910 * In the common case we either have no sleepers or
3911 * are still over the limit and can just return.
3913 old = atomic_fetchadd_64(&zone->uz_items, -count);
3914 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3915 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3919 * Moderate the rate of wakeups. Sleepers will continue
3920 * to generate wakeups if necessary.
3922 wakeup_one(&zone->uz_max_items);
3926 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3928 uma_bucket_t bucket;
3931 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3934 /* Avoid allocs targeting empty domains. */
3935 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3936 domain = UMA_ANYDOMAIN;
3937 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
3938 domain = UMA_ANYDOMAIN;
3940 if (zone->uz_max_items > 0)
3941 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3944 maxbucket = zone->uz_bucket_size;
3948 /* Don't wait for buckets, preserve caller's NOVM setting. */
3949 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3950 if (bucket == NULL) {
3955 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3956 MIN(maxbucket, bucket->ub_entries), domain, flags);
3959 * Initialize the memory if necessary.
3961 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3964 for (i = 0; i < bucket->ub_cnt; i++)
3965 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3969 * If we couldn't initialize the whole bucket, put the
3970 * rest back onto the freelist.
3972 if (i != bucket->ub_cnt) {
3973 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3974 bucket->ub_cnt - i);
3976 bzero(&bucket->ub_bucket[i],
3977 sizeof(void *) * (bucket->ub_cnt - i));
3983 cnt = bucket->ub_cnt;
3984 if (bucket->ub_cnt == 0) {
3985 bucket_free(zone, bucket, udata);
3986 counter_u64_add(zone->uz_fails, 1);
3990 if (zone->uz_max_items > 0 && cnt < maxbucket)
3991 zone_free_limit(zone, maxbucket - cnt);
3997 * Allocates a single item from a zone.
4000 * zone The zone to alloc for.
4001 * udata The data to be passed to the constructor.
4002 * domain The domain to allocate from or UMA_ANYDOMAIN.
4003 * flags M_WAITOK, M_NOWAIT, M_ZERO.
4006 * NULL if there is no memory and M_NOWAIT is set
4007 * An item if successful
4011 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4015 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4016 counter_u64_add(zone->uz_fails, 1);
4020 /* Avoid allocs targeting empty domains. */
4021 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4022 domain = UMA_ANYDOMAIN;
4024 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4028 * We have to call both the zone's init (not the keg's init)
4029 * and the zone's ctor. This is because the item is going from
4030 * a keg slab directly to the user, and the user is expecting it
4031 * to be both zone-init'd as well as zone-ctor'd.
4033 if (zone->uz_init != NULL) {
4034 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
4035 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4039 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4044 counter_u64_add(zone->uz_allocs, 1);
4045 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4046 zone->uz_name, zone);
4051 counter_u64_add(zone->uz_fails, 1);
4053 if (zone->uz_max_items > 0)
4054 zone_free_limit(zone, 1);
4055 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4056 zone->uz_name, zone);
4063 uma_zfree_smr(uma_zone_t zone, void *item)
4066 uma_cache_bucket_t bucket;
4067 int itemdomain, uz_flags;
4069 #ifdef UMA_ZALLOC_DEBUG
4070 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4071 ("uma_zfree_smr: called with non-SMR zone."));
4072 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4073 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4074 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4077 cache = &zone->uz_cpu[curcpu];
4078 uz_flags = cache_uz_flags(cache);
4081 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4082 itemdomain = item_domain(item);
4086 cache = &zone->uz_cpu[curcpu];
4087 /* SMR Zones must free to the free bucket. */
4088 bucket = &cache->uc_freebucket;
4090 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4091 PCPU_GET(domain) != itemdomain) {
4092 bucket = &cache->uc_crossbucket;
4095 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4096 cache_bucket_push(cache, bucket, item);
4100 } while (cache_free(zone, cache, NULL, item, itemdomain));
4104 * If nothing else caught this, we'll just do an internal free.
4106 zone_free_item(zone, item, NULL, SKIP_NONE);
4111 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4114 uma_cache_bucket_t bucket;
4115 int itemdomain, uz_flags;
4117 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4118 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4120 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4122 #ifdef UMA_ZALLOC_DEBUG
4123 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4124 ("uma_zfree_arg: called with SMR zone."));
4125 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4128 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4133 * We are accessing the per-cpu cache without a critical section to
4134 * fetch size and flags. This is acceptable, if we are preempted we
4135 * will simply read another cpu's line.
4137 cache = &zone->uz_cpu[curcpu];
4138 uz_flags = cache_uz_flags(cache);
4139 if (UMA_ALWAYS_CTORDTOR ||
4140 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4141 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4144 * The race here is acceptable. If we miss it we'll just have to wait
4145 * a little longer for the limits to be reset.
4147 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4148 if (zone->uz_sleepers > 0)
4153 * If possible, free to the per-CPU cache. There are two
4154 * requirements for safe access to the per-CPU cache: (1) the thread
4155 * accessing the cache must not be preempted or yield during access,
4156 * and (2) the thread must not migrate CPUs without switching which
4157 * cache it accesses. We rely on a critical section to prevent
4158 * preemption and migration. We release the critical section in
4159 * order to acquire the zone mutex if we are unable to free to the
4160 * current cache; when we re-acquire the critical section, we must
4161 * detect and handle migration if it has occurred.
4165 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4166 itemdomain = item_domain(item);
4170 cache = &zone->uz_cpu[curcpu];
4172 * Try to free into the allocbucket first to give LIFO
4173 * ordering for cache-hot datastructures. Spill over
4174 * into the freebucket if necessary. Alloc will swap
4175 * them if one runs dry.
4177 bucket = &cache->uc_allocbucket;
4179 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4180 PCPU_GET(domain) != itemdomain) {
4181 bucket = &cache->uc_crossbucket;
4184 if (bucket->ucb_cnt == bucket->ucb_entries &&
4185 cache->uc_freebucket.ucb_cnt <
4186 cache->uc_freebucket.ucb_entries)
4187 cache_bucket_swap(&cache->uc_freebucket,
4188 &cache->uc_allocbucket);
4189 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4190 cache_bucket_push(cache, bucket, item);
4194 } while (cache_free(zone, cache, udata, item, itemdomain));
4198 * If nothing else caught this, we'll just do an internal free.
4201 zone_free_item(zone, item, udata, SKIP_DTOR);
4206 * sort crossdomain free buckets to domain correct buckets and cache
4210 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4212 struct uma_bucketlist fullbuckets;
4213 uma_zone_domain_t zdom;
4220 "uma_zfree: zone %s(%p) draining cross bucket %p",
4221 zone->uz_name, zone, bucket);
4224 * It is possible for buckets to arrive here out of order so we fetch
4225 * the current smr seq rather than accepting the bucket's.
4227 seq = SMR_SEQ_INVALID;
4228 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4229 seq = smr_advance(zone->uz_smr);
4232 * To avoid having ndomain * ndomain buckets for sorting we have a
4233 * lock on the current crossfree bucket. A full matrix with
4234 * per-domain locking could be used if necessary.
4236 STAILQ_INIT(&fullbuckets);
4237 ZONE_CROSS_LOCK(zone);
4238 while (bucket->ub_cnt > 0) {
4239 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4240 domain = item_domain(item);
4241 zdom = ZDOM_GET(zone, domain);
4242 if (zdom->uzd_cross == NULL) {
4243 zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
4244 if (zdom->uzd_cross == NULL)
4247 b = zdom->uzd_cross;
4248 b->ub_bucket[b->ub_cnt++] = item;
4250 if (b->ub_cnt == b->ub_entries) {
4251 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4252 zdom->uzd_cross = NULL;
4256 ZONE_CROSS_UNLOCK(zone);
4257 if (bucket->ub_cnt == 0)
4258 bucket->ub_seq = SMR_SEQ_INVALID;
4259 bucket_free(zone, bucket, udata);
4261 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4262 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4263 domain = item_domain(b->ub_bucket[0]);
4264 zone_put_bucket(zone, domain, b, udata, true);
4270 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4271 int itemdomain, bool ws)
4276 * Buckets coming from the wrong domain will be entirely for the
4277 * only other domain on two domain systems. In this case we can
4278 * simply cache them. Otherwise we need to sort them back to
4281 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4282 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4283 zone_free_cross(zone, bucket, udata);
4289 * Attempt to save the bucket in the zone's domain bucket cache.
4292 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4293 zone->uz_name, zone, bucket);
4294 /* ub_cnt is pointing to the last free item */
4295 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4296 itemdomain = zone_domain_lowest(zone, itemdomain);
4297 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4301 * Populate a free or cross bucket for the current cpu cache. Free any
4302 * existing full bucket either to the zone cache or back to the slab layer.
4304 * Enters and returns in a critical section. false return indicates that
4305 * we can not satisfy this free in the cache layer. true indicates that
4306 * the caller should retry.
4308 static __noinline bool
4309 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4312 uma_cache_bucket_t cbucket;
4313 uma_bucket_t newbucket, bucket;
4315 CRITICAL_ASSERT(curthread);
4317 if (zone->uz_bucket_size == 0)
4320 cache = &zone->uz_cpu[curcpu];
4324 * FIRSTTOUCH domains need to free to the correct zdom. When
4325 * enabled this is the zdom of the item. The bucket is the
4326 * cross bucket if the current domain and itemdomain do not match.
4328 cbucket = &cache->uc_freebucket;
4330 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4331 if (PCPU_GET(domain) != itemdomain) {
4332 cbucket = &cache->uc_crossbucket;
4333 if (cbucket->ucb_cnt != 0)
4334 counter_u64_add(zone->uz_xdomain,
4339 bucket = cache_bucket_unload(cbucket);
4340 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4341 ("cache_free: Entered with non-full free bucket."));
4343 /* We are no longer associated with this CPU. */
4347 * Don't let SMR zones operate without a free bucket. Force
4348 * a synchronize and re-use this one. We will only degrade
4349 * to a synchronize every bucket_size items rather than every
4350 * item if we fail to allocate a bucket.
4352 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4354 bucket->ub_seq = smr_advance(zone->uz_smr);
4355 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4356 if (newbucket == NULL && bucket != NULL) {
4357 bucket_drain(zone, bucket);
4361 } else if (!bucketdisable)
4362 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4365 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4368 if ((bucket = newbucket) == NULL)
4370 cache = &zone->uz_cpu[curcpu];
4373 * Check to see if we should be populating the cross bucket. If it
4374 * is already populated we will fall through and attempt to populate
4377 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4378 if (PCPU_GET(domain) != itemdomain &&
4379 cache->uc_crossbucket.ucb_bucket == NULL) {
4380 cache_bucket_load_cross(cache, bucket);
4386 * We may have lost the race to fill the bucket or switched CPUs.
4388 if (cache->uc_freebucket.ucb_bucket != NULL) {
4390 bucket_free(zone, bucket, udata);
4393 cache_bucket_load_free(cache, bucket);
4399 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4406 KEG_LOCK_ASSERT(keg, slab->us_domain);
4408 /* Do we need to remove from any lists? */
4409 dom = &keg->uk_domain[slab->us_domain];
4410 if (slab->us_freecount + 1 == keg->uk_ipers) {
4411 LIST_REMOVE(slab, us_link);
4412 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4413 dom->ud_free_slabs++;
4414 } else if (slab->us_freecount == 0) {
4415 LIST_REMOVE(slab, us_link);
4416 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4419 /* Slab management. */
4420 freei = slab_item_index(slab, keg, item);
4421 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4422 slab->us_freecount++;
4424 /* Keg statistics. */
4425 dom->ud_free_items++;
4429 zone_release(void *arg, void **bucket, int cnt)
4442 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4443 lock = KEG_LOCK(keg, 0);
4444 for (i = 0; i < cnt; i++) {
4446 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4447 slab = vtoslab((vm_offset_t)item);
4449 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4450 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4451 slab = hash_sfind(&keg->uk_hash, mem);
4453 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4455 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4458 lock = KEG_LOCK(keg, slab->us_domain);
4460 slab_free_item(zone, slab, item);
4467 * Frees a single item to any zone.
4470 * zone The zone to free to
4471 * item The item we're freeing
4472 * udata User supplied data for the dtor
4473 * skip Skip dtors and finis
4475 static __noinline void
4476 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4480 * If a free is sent directly to an SMR zone we have to
4481 * synchronize immediately because the item can instantly
4482 * be reallocated. This should only happen in degenerate
4483 * cases when no memory is available for per-cpu caches.
4485 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4486 smr_synchronize(zone->uz_smr);
4488 item_dtor(zone, item, zone->uz_size, udata, skip);
4490 if (skip < SKIP_FINI && zone->uz_fini)
4491 zone->uz_fini(item, zone->uz_size);
4493 zone->uz_release(zone->uz_arg, &item, 1);
4495 if (skip & SKIP_CNT)
4498 counter_u64_add(zone->uz_frees, 1);
4500 if (zone->uz_max_items > 0)
4501 zone_free_limit(zone, 1);
4506 uma_zone_set_max(uma_zone_t zone, int nitems)
4508 struct uma_bucket_zone *ubz;
4512 * XXX This can misbehave if the zone has any allocations with
4513 * no limit and a limit is imposed. There is currently no
4514 * way to clear a limit.
4517 ubz = bucket_zone_max(zone, nitems);
4518 count = ubz != NULL ? ubz->ubz_entries : 0;
4519 zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4520 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4521 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4522 zone->uz_max_items = nitems;
4523 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4524 zone_update_caches(zone);
4525 /* We may need to wake waiters. */
4526 wakeup(&zone->uz_max_items);
4534 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4536 struct uma_bucket_zone *ubz;
4540 ubz = bucket_zone_max(zone, nitems);
4543 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4544 /* Count the cross-domain bucket. */
4546 nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4547 zone->uz_bucket_size_max = ubz->ubz_entries;
4549 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4551 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4552 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4553 zone->uz_bucket_max = nitems / vm_ndomains;
4559 uma_zone_get_max(uma_zone_t zone)
4563 nitems = atomic_load_64(&zone->uz_max_items);
4570 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4573 ZONE_ASSERT_COLD(zone);
4574 zone->uz_warning = warning;
4579 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4582 ZONE_ASSERT_COLD(zone);
4583 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4588 uma_zone_get_cur(uma_zone_t zone)
4594 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4595 nitems = counter_u64_fetch(zone->uz_allocs) -
4596 counter_u64_fetch(zone->uz_frees);
4598 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4599 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4601 return (nitems < 0 ? 0 : nitems);
4605 uma_zone_get_allocs(uma_zone_t zone)
4611 if (zone->uz_allocs != EARLY_COUNTER)
4612 nitems = counter_u64_fetch(zone->uz_allocs);
4614 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4620 uma_zone_get_frees(uma_zone_t zone)
4626 if (zone->uz_frees != EARLY_COUNTER)
4627 nitems = counter_u64_fetch(zone->uz_frees);
4629 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4635 /* Used only for KEG_ASSERT_COLD(). */
4637 uma_keg_get_allocs(uma_keg_t keg)
4643 LIST_FOREACH(z, &keg->uk_zones, uz_link)
4644 nitems += uma_zone_get_allocs(z);
4652 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4657 KEG_ASSERT_COLD(keg);
4658 keg->uk_init = uminit;
4663 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4668 KEG_ASSERT_COLD(keg);
4669 keg->uk_fini = fini;
4674 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4677 ZONE_ASSERT_COLD(zone);
4678 zone->uz_init = zinit;
4683 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4686 ZONE_ASSERT_COLD(zone);
4687 zone->uz_fini = zfini;
4692 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4697 KEG_ASSERT_COLD(keg);
4698 keg->uk_freef = freef;
4703 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4708 KEG_ASSERT_COLD(keg);
4709 keg->uk_allocf = allocf;
4714 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4717 ZONE_ASSERT_COLD(zone);
4719 KASSERT(smr != NULL, ("Got NULL smr"));
4720 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4721 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
4722 zone->uz_flags |= UMA_ZONE_SMR;
4724 zone_update_caches(zone);
4728 uma_zone_get_smr(uma_zone_t zone)
4731 return (zone->uz_smr);
4736 uma_zone_reserve(uma_zone_t zone, int items)
4741 KEG_ASSERT_COLD(keg);
4742 keg->uk_reserve = items;
4747 uma_zone_reserve_kva(uma_zone_t zone, int count)
4754 KEG_ASSERT_COLD(keg);
4755 ZONE_ASSERT_COLD(zone);
4757 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4759 #ifdef UMA_MD_SMALL_ALLOC
4760 if (keg->uk_ppera > 1) {
4764 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4770 MPASS(keg->uk_kva == 0);
4773 zone->uz_max_items = pages * keg->uk_ipers;
4774 #ifdef UMA_MD_SMALL_ALLOC
4775 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4777 keg->uk_allocf = noobj_alloc;
4779 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4780 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4781 zone_update_caches(zone);
4788 uma_prealloc(uma_zone_t zone, int items)
4790 struct vm_domainset_iter di;
4794 int aflags, domain, slabs;
4797 slabs = howmany(items, keg->uk_ipers);
4798 while (slabs-- > 0) {
4800 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4803 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4806 dom = &keg->uk_domain[slab->us_domain];
4808 * keg_alloc_slab() always returns a slab on the
4811 LIST_REMOVE(slab, us_link);
4812 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4814 dom->ud_free_slabs++;
4815 KEG_UNLOCK(keg, slab->us_domain);
4818 if (vm_domainset_iter_policy(&di, &domain) != 0)
4819 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
4825 * Returns a snapshot of memory consumption in bytes.
4828 uma_zone_memory(uma_zone_t zone)
4834 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
4835 for (i = 0; i < vm_ndomains; i++)
4836 sz += ZDOM_GET(zone, i)->uzd_nitems;
4837 return (sz * zone->uz_size);
4839 for (i = 0; i < vm_ndomains; i++)
4840 sz += zone->uz_keg->uk_domain[i].ud_pages;
4842 return (sz * PAGE_SIZE);
4847 uma_reclaim(int req)
4850 CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4851 sx_xlock(&uma_reclaim_lock);
4855 case UMA_RECLAIM_TRIM:
4856 zone_foreach(zone_trim, NULL);
4858 case UMA_RECLAIM_DRAIN:
4859 case UMA_RECLAIM_DRAIN_CPU:
4860 zone_foreach(zone_drain, NULL);
4861 if (req == UMA_RECLAIM_DRAIN_CPU) {
4862 pcpu_cache_drain_safe(NULL);
4863 zone_foreach(zone_drain, NULL);
4867 panic("unhandled reclamation request %d", req);
4871 * Some slabs may have been freed but this zone will be visited early
4872 * we visit again so that we can free pages that are empty once other
4873 * zones are drained. We have to do the same for buckets.
4875 zone_drain(slabzones[0], NULL);
4876 zone_drain(slabzones[1], NULL);
4877 bucket_zone_drain();
4878 sx_xunlock(&uma_reclaim_lock);
4881 static volatile int uma_reclaim_needed;
4884 uma_reclaim_wakeup(void)
4887 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4888 wakeup(uma_reclaim);
4892 uma_reclaim_worker(void *arg __unused)
4896 sx_xlock(&uma_reclaim_lock);
4897 while (atomic_load_int(&uma_reclaim_needed) == 0)
4898 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4900 sx_xunlock(&uma_reclaim_lock);
4901 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4902 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4903 atomic_store_int(&uma_reclaim_needed, 0);
4904 /* Don't fire more than once per-second. */
4905 pause("umarclslp", hz);
4911 uma_zone_reclaim(uma_zone_t zone, int req)
4915 case UMA_RECLAIM_TRIM:
4916 zone_trim(zone, NULL);
4918 case UMA_RECLAIM_DRAIN:
4919 zone_drain(zone, NULL);
4921 case UMA_RECLAIM_DRAIN_CPU:
4922 pcpu_cache_drain_safe(zone);
4923 zone_drain(zone, NULL);
4926 panic("unhandled reclamation request %d", req);
4932 uma_zone_exhausted(uma_zone_t zone)
4935 return (atomic_load_32(&zone->uz_sleepers) > 0);
4942 return (uma_kmem_limit);
4946 uma_set_limit(unsigned long limit)
4949 uma_kmem_limit = limit;
4956 return (atomic_load_long(&uma_kmem_total));
4963 return (uma_kmem_limit - uma_size());
4968 * Generate statistics across both the zone and its per-cpu cache's. Return
4969 * desired statistics if the pointer is non-NULL for that statistic.
4971 * Note: does not update the zone statistics, as it can't safely clear the
4972 * per-CPU cache statistic.
4976 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4977 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4980 uint64_t allocs, frees, sleeps, xdomain;
4983 allocs = frees = sleeps = xdomain = 0;
4986 cache = &z->uz_cpu[cpu];
4987 cachefree += cache->uc_allocbucket.ucb_cnt;
4988 cachefree += cache->uc_freebucket.ucb_cnt;
4989 xdomain += cache->uc_crossbucket.ucb_cnt;
4990 cachefree += cache->uc_crossbucket.ucb_cnt;
4991 allocs += cache->uc_allocs;
4992 frees += cache->uc_frees;
4994 allocs += counter_u64_fetch(z->uz_allocs);
4995 frees += counter_u64_fetch(z->uz_frees);
4996 xdomain += counter_u64_fetch(z->uz_xdomain);
4997 sleeps += z->uz_sleeps;
4998 if (cachefreep != NULL)
4999 *cachefreep = cachefree;
5000 if (allocsp != NULL)
5004 if (sleepsp != NULL)
5006 if (xdomainp != NULL)
5007 *xdomainp = xdomain;
5012 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5019 rw_rlock(&uma_rwlock);
5020 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5021 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5024 LIST_FOREACH(z, &uma_cachezones, uz_link)
5027 rw_runlock(&uma_rwlock);
5028 return (sysctl_handle_int(oidp, &count, 0, req));
5032 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5033 struct uma_percpu_stat *ups, bool internal)
5035 uma_zone_domain_t zdom;
5039 for (i = 0; i < vm_ndomains; i++) {
5040 zdom = ZDOM_GET(z, i);
5041 uth->uth_zone_free += zdom->uzd_nitems;
5043 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5044 uth->uth_frees = counter_u64_fetch(z->uz_frees);
5045 uth->uth_fails = counter_u64_fetch(z->uz_fails);
5046 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5047 uth->uth_sleeps = z->uz_sleeps;
5049 for (i = 0; i < mp_maxid + 1; i++) {
5050 bzero(&ups[i], sizeof(*ups));
5051 if (internal || CPU_ABSENT(i))
5053 cache = &z->uz_cpu[i];
5054 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5055 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5056 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5057 ups[i].ups_allocs = cache->uc_allocs;
5058 ups[i].ups_frees = cache->uc_frees;
5063 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5065 struct uma_stream_header ush;
5066 struct uma_type_header uth;
5067 struct uma_percpu_stat *ups;
5072 uint32_t kfree, pages;
5073 int count, error, i;
5075 error = sysctl_wire_old_buffer(req, 0);
5078 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5079 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5080 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5083 rw_rlock(&uma_rwlock);
5084 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5085 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5089 LIST_FOREACH(z, &uma_cachezones, uz_link)
5093 * Insert stream header.
5095 bzero(&ush, sizeof(ush));
5096 ush.ush_version = UMA_STREAM_VERSION;
5097 ush.ush_maxcpus = (mp_maxid + 1);
5098 ush.ush_count = count;
5099 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5101 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5103 for (i = 0; i < vm_ndomains; i++) {
5104 kfree += kz->uk_domain[i].ud_free_items;
5105 pages += kz->uk_domain[i].ud_pages;
5107 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5108 bzero(&uth, sizeof(uth));
5109 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5110 uth.uth_align = kz->uk_align;
5111 uth.uth_size = kz->uk_size;
5112 uth.uth_rsize = kz->uk_rsize;
5113 if (z->uz_max_items > 0) {
5114 items = UZ_ITEMS_COUNT(z->uz_items);
5115 uth.uth_pages = (items / kz->uk_ipers) *
5118 uth.uth_pages = pages;
5119 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5121 uth.uth_limit = z->uz_max_items;
5122 uth.uth_keg_free = kfree;
5125 * A zone is secondary is it is not the first entry
5126 * on the keg's zone list.
5128 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5129 (LIST_FIRST(&kz->uk_zones) != z))
5130 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5131 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5132 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5133 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5134 for (i = 0; i < mp_maxid + 1; i++)
5135 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5138 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5139 bzero(&uth, sizeof(uth));
5140 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5141 uth.uth_size = z->uz_size;
5142 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5143 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5144 for (i = 0; i < mp_maxid + 1; i++)
5145 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5148 rw_runlock(&uma_rwlock);
5149 error = sbuf_finish(&sbuf);
5156 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5158 uma_zone_t zone = *(uma_zone_t *)arg1;
5161 max = uma_zone_get_max(zone);
5162 error = sysctl_handle_int(oidp, &max, 0, req);
5163 if (error || !req->newptr)
5166 uma_zone_set_max(zone, max);
5172 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5178 * Some callers want to add sysctls for global zones that
5179 * may not yet exist so they pass a pointer to a pointer.
5182 zone = *(uma_zone_t *)arg1;
5185 cur = uma_zone_get_cur(zone);
5186 return (sysctl_handle_int(oidp, &cur, 0, req));
5190 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5192 uma_zone_t zone = arg1;
5195 cur = uma_zone_get_allocs(zone);
5196 return (sysctl_handle_64(oidp, &cur, 0, req));
5200 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5202 uma_zone_t zone = arg1;
5205 cur = uma_zone_get_frees(zone);
5206 return (sysctl_handle_64(oidp, &cur, 0, req));
5210 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5213 uma_zone_t zone = arg1;
5216 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5217 if (zone->uz_flags != 0)
5218 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5220 sbuf_printf(&sbuf, "0");
5221 error = sbuf_finish(&sbuf);
5228 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5230 uma_keg_t keg = arg1;
5231 int avail, effpct, total;
5233 total = keg->uk_ppera * PAGE_SIZE;
5234 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5235 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5237 * We consider the client's requested size and alignment here, not the
5238 * real size determination uk_rsize, because we also adjust the real
5239 * size for internal implementation reasons (max bitset size).
5241 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5242 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5243 avail *= mp_maxid + 1;
5244 effpct = 100 * avail / total;
5245 return (sysctl_handle_int(oidp, &effpct, 0, req));
5249 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5251 uma_zone_t zone = arg1;
5254 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5255 return (sysctl_handle_64(oidp, &cur, 0, req));
5260 uma_dbg_getslab(uma_zone_t zone, void *item)
5267 * It is safe to return the slab here even though the
5268 * zone is unlocked because the item's allocation state
5269 * essentially holds a reference.
5271 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5272 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5274 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5275 return (vtoslab((vm_offset_t)mem));
5277 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5278 return ((uma_slab_t)(mem + keg->uk_pgoff));
5280 slab = hash_sfind(&keg->uk_hash, mem);
5287 uma_dbg_zskip(uma_zone_t zone, void *mem)
5290 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5293 return (uma_dbg_kskip(zone->uz_keg, mem));
5297 uma_dbg_kskip(uma_keg_t keg, void *mem)
5301 if (dbg_divisor == 0)
5304 if (dbg_divisor == 1)
5307 idx = (uintptr_t)mem >> PAGE_SHIFT;
5308 if (keg->uk_ipers > 1) {
5309 idx *= keg->uk_ipers;
5310 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5313 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5314 counter_u64_add(uma_skip_cnt, 1);
5317 counter_u64_add(uma_dbg_cnt, 1);
5323 * Set up the slab's freei data such that uma_dbg_free can function.
5327 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5333 slab = uma_dbg_getslab(zone, item);
5335 panic("uma: item %p did not belong to zone %s",
5336 item, zone->uz_name);
5339 freei = slab_item_index(slab, keg, item);
5341 if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5342 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5343 item, zone, zone->uz_name, slab, freei);
5344 BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5348 * Verifies freed addresses. Checks for alignment, valid slab membership
5349 * and duplicate frees.
5353 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5359 slab = uma_dbg_getslab(zone, item);
5361 panic("uma: Freed item %p did not belong to zone %s",
5362 item, zone->uz_name);
5365 freei = slab_item_index(slab, keg, item);
5367 if (freei >= keg->uk_ipers)
5368 panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5369 item, zone, zone->uz_name, slab, freei);
5371 if (slab_item(slab, keg, freei) != item)
5372 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5373 item, zone, zone->uz_name, slab, freei);
5375 if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5376 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5377 item, zone, zone->uz_name, slab, freei);
5379 BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5381 #endif /* INVARIANTS */
5385 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5386 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5391 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5392 *allocs = counter_u64_fetch(z->uz_allocs);
5393 frees = counter_u64_fetch(z->uz_frees);
5394 *sleeps = z->uz_sleeps;
5398 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5400 for (i = 0; i < vm_ndomains; i++) {
5401 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5402 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5403 (LIST_FIRST(&kz->uk_zones) != z)))
5404 *cachefree += kz->uk_domain[i].ud_free_items;
5406 *used = *allocs - frees;
5407 return (((int64_t)*used + *cachefree) * kz->uk_size);
5410 DB_SHOW_COMMAND(uma, db_show_uma)
5412 const char *fmt_hdr, *fmt_entry;
5415 uint64_t allocs, used, sleeps, xdomain;
5417 /* variables for sorting */
5419 uma_zone_t cur_zone, last_zone;
5420 int64_t cur_size, last_size, size;
5423 /* /i option produces machine-parseable CSV output */
5424 if (modif[0] == 'i') {
5425 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5426 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5428 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5429 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5432 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5433 "Sleeps", "Bucket", "Total Mem", "XFree");
5435 /* Sort the zones with largest size first. */
5437 last_size = INT64_MAX;
5442 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5443 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5445 * In the case of size ties, print out zones
5446 * in the order they are encountered. That is,
5447 * when we encounter the most recently output
5448 * zone, we have already printed all preceding
5449 * ties, and we must print all following ties.
5451 if (z == last_zone) {
5455 size = get_uma_stats(kz, z, &allocs, &used,
5456 &sleeps, &cachefree, &xdomain);
5457 if (size > cur_size && size < last_size + ties)
5465 if (cur_zone == NULL)
5468 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5469 &sleeps, &cachefree, &xdomain);
5470 db_printf(fmt_entry, cur_zone->uz_name,
5471 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5472 (uintmax_t)allocs, (uintmax_t)sleeps,
5473 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5478 last_zone = cur_zone;
5479 last_size = cur_size;
5483 DB_SHOW_COMMAND(umacache, db_show_umacache)
5486 uint64_t allocs, frees;
5490 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5491 "Requests", "Bucket");
5492 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5493 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5494 for (i = 0; i < vm_ndomains; i++)
5495 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5496 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5497 z->uz_name, (uintmax_t)z->uz_size,
5498 (intmax_t)(allocs - frees), cachefree,
5499 (uintmax_t)allocs, z->uz_bucket_size);