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_domainset.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_param.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>
96 #include <vm/uma_int.h>
97 #include <vm/uma_dbg.h>
101 #ifdef DEBUG_MEMGUARD
102 #include <vm/memguard.h>
105 #include <machine/md_var.h>
108 #define UMA_ALWAYS_CTORDTOR 1
110 #define UMA_ALWAYS_CTORDTOR 0
114 * This is the zone and keg from which all zones are spawned.
116 static uma_zone_t kegs;
117 static uma_zone_t zones;
120 * On INVARIANTS builds, the slab contains a second bitset of the same size,
121 * "dbg_bits", which is laid out immediately after us_free.
124 #define SLAB_BITSETS 2
126 #define SLAB_BITSETS 1
130 * These are the two zones from which all offpage uma_slab_ts are allocated.
132 * One zone is for slab headers that can represent a larger number of items,
133 * making the slabs themselves more efficient, and the other zone is for
134 * headers that are smaller and represent fewer items, making the headers more
137 #define SLABZONE_SIZE(setsize) \
138 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
139 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16)
140 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE
141 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE)
142 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE)
143 static uma_zone_t slabzones[2];
146 * The initial hash tables come out of this zone so they can be allocated
147 * prior to malloc coming up.
149 static uma_zone_t hashzone;
151 /* The boot-time adjusted value for cache line alignment. */
152 int uma_align_cache = 64 - 1;
154 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
155 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
158 * Are we allowed to allocate buckets?
160 static int bucketdisable = 1;
162 /* Linked list of all kegs in the system */
163 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
165 /* Linked list of all cache-only zones in the system */
166 static LIST_HEAD(,uma_zone) uma_cachezones =
167 LIST_HEAD_INITIALIZER(uma_cachezones);
169 /* This RW lock protects the keg list */
170 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
173 * First available virual address for boot time allocations.
175 static vm_offset_t bootstart;
176 static vm_offset_t bootmem;
178 static struct sx uma_reclaim_lock;
181 * kmem soft limit, initialized by uma_set_limit(). Ensure that early
182 * allocations don't trigger a wakeup of the reclaim thread.
184 unsigned long uma_kmem_limit = LONG_MAX;
185 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
186 "UMA kernel memory soft limit");
187 unsigned long uma_kmem_total;
188 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
189 "UMA kernel memory usage");
191 /* Is the VM done starting up? */
198 } booted = BOOT_COLD;
201 * This is the handle used to schedule events that need to happen
202 * outside of the allocation fast path.
204 static struct callout uma_callout;
205 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
208 * This structure is passed as the zone ctor arg so that I don't have to create
209 * a special allocation function just for zones.
211 struct uma_zctor_args {
226 struct uma_kctor_args {
235 struct uma_bucket_zone {
237 const char *ubz_name;
238 int ubz_entries; /* Number of items it can hold. */
239 int ubz_maxsize; /* Maximum allocation size per-item. */
243 * Compute the actual number of bucket entries to pack them in power
244 * of two sizes for more efficient space utilization.
246 #define BUCKET_SIZE(n) \
247 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
249 #define BUCKET_MAX BUCKET_SIZE(256)
252 struct uma_bucket_zone bucket_zones[] = {
253 /* Literal bucket sizes. */
254 { NULL, "2 Bucket", 2, 4096 },
255 { NULL, "4 Bucket", 4, 3072 },
256 { NULL, "8 Bucket", 8, 2048 },
257 { NULL, "16 Bucket", 16, 1024 },
258 /* Rounded down power of 2 sizes for efficiency. */
259 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
260 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
261 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
262 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
267 * Flags and enumerations to be passed to internal functions.
271 SKIP_CNT = 0x00000001,
272 SKIP_DTOR = 0x00010000,
273 SKIP_FINI = 0x00020000,
278 void uma_startup1(vm_offset_t);
279 void uma_startup2(void);
281 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
282 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
283 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
284 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
285 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
286 static void page_free(void *, vm_size_t, uint8_t);
287 static void pcpu_page_free(void *, vm_size_t, uint8_t);
288 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
289 static void cache_drain(uma_zone_t);
290 static void bucket_drain(uma_zone_t, uma_bucket_t);
291 static void bucket_cache_reclaim(uma_zone_t zone, bool);
292 static int keg_ctor(void *, int, void *, int);
293 static void keg_dtor(void *, int, void *);
294 static int zone_ctor(void *, int, void *, int);
295 static void zone_dtor(void *, int, void *);
296 static inline void item_dtor(uma_zone_t zone, void *item, int size,
297 void *udata, enum zfreeskip skip);
298 static int zero_init(void *, int, int);
299 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
300 int itemdomain, bool ws);
301 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
302 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
303 static void zone_timeout(uma_zone_t zone, void *);
304 static int hash_alloc(struct uma_hash *, u_int);
305 static int hash_expand(struct uma_hash *, struct uma_hash *);
306 static void hash_free(struct uma_hash *hash);
307 static void uma_timeout(void *);
308 static void uma_shutdown(void);
309 static void *zone_alloc_item(uma_zone_t, void *, int, int);
310 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
311 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
312 static void zone_free_limit(uma_zone_t zone, int count);
313 static void bucket_enable(void);
314 static void bucket_init(void);
315 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
316 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
317 static void bucket_zone_drain(void);
318 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
319 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
320 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
321 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
322 uma_fini fini, int align, uint32_t flags);
323 static int zone_import(void *, void **, int, int, int);
324 static void zone_release(void *, void **, int);
325 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
326 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
328 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
329 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
330 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
331 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
332 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
333 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
334 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
336 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
338 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
339 "Memory allocation debugging");
342 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
343 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
345 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
346 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
347 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
348 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
350 static u_int dbg_divisor = 1;
351 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
352 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
353 "Debug & thrash every this item in memory allocator");
355 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
356 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
357 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
358 &uma_dbg_cnt, "memory items debugged");
359 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
360 &uma_skip_cnt, "memory items skipped, not debugged");
363 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
364 "Universal Memory Allocator");
366 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
367 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
369 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
370 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
372 static int zone_warnings = 1;
373 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
374 "Warn when UMA zones becomes full");
376 static int multipage_slabs = 1;
377 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
378 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
379 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
380 "UMA may choose larger slab sizes for better efficiency");
383 * Select the slab zone for an offpage slab with the given maximum item count.
385 static inline uma_zone_t
389 return (slabzones[ipers > SLABZONE0_SETSIZE]);
393 * This routine checks to see whether or not it's safe to enable buckets.
399 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
400 bucketdisable = vm_page_count_min();
404 * Initialize bucket_zones, the array of zones of buckets of various sizes.
406 * For each zone, calculate the memory required for each bucket, consisting
407 * of the header and an array of pointers.
412 struct uma_bucket_zone *ubz;
415 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
416 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
417 size += sizeof(void *) * ubz->ubz_entries;
418 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
419 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
420 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
421 UMA_ZONE_FIRSTTOUCH);
426 * Given a desired number of entries for a bucket, return the zone from which
427 * to allocate the bucket.
429 static struct uma_bucket_zone *
430 bucket_zone_lookup(int entries)
432 struct uma_bucket_zone *ubz;
434 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
435 if (ubz->ubz_entries >= entries)
441 static struct uma_bucket_zone *
442 bucket_zone_max(uma_zone_t zone, int nitems)
444 struct uma_bucket_zone *ubz;
448 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
449 /* Count the cross-domain bucket. */
452 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
453 if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems)
455 if (ubz == &bucket_zones[0])
463 bucket_select(int size)
465 struct uma_bucket_zone *ubz;
467 ubz = &bucket_zones[0];
468 if (size > ubz->ubz_maxsize)
469 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
471 for (; ubz->ubz_entries != 0; ubz++)
472 if (ubz->ubz_maxsize < size)
475 return (ubz->ubz_entries);
479 bucket_alloc(uma_zone_t zone, void *udata, int flags)
481 struct uma_bucket_zone *ubz;
485 * Don't allocate buckets early in boot.
487 if (__predict_false(booted < BOOT_KVA))
491 * To limit bucket recursion we store the original zone flags
492 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
493 * NOVM flag to persist even through deep recursions. We also
494 * store ZFLAG_BUCKET once we have recursed attempting to allocate
495 * a bucket for a bucket zone so we do not allow infinite bucket
496 * recursion. This cookie will even persist to frees of unused
497 * buckets via the allocation path or bucket allocations in the
500 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
501 udata = (void *)(uintptr_t)zone->uz_flags;
503 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
505 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
507 if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
509 ubz = bucket_zone_lookup(zone->uz_bucket_size);
510 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
512 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
515 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
518 bucket->ub_entries = ubz->ubz_entries;
519 bucket->ub_seq = SMR_SEQ_INVALID;
520 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
521 zone->uz_name, zone, bucket);
528 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
530 struct uma_bucket_zone *ubz;
532 if (bucket->ub_cnt != 0)
533 bucket_drain(zone, bucket);
535 KASSERT(bucket->ub_cnt == 0,
536 ("bucket_free: Freeing a non free bucket."));
537 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
538 ("bucket_free: Freeing an SMR bucket."));
539 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
540 udata = (void *)(uintptr_t)zone->uz_flags;
541 ubz = bucket_zone_lookup(bucket->ub_entries);
542 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
546 bucket_zone_drain(void)
548 struct uma_bucket_zone *ubz;
550 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
551 uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN);
555 * Acquire the domain lock and record contention.
557 static uma_zone_domain_t
558 zone_domain_lock(uma_zone_t zone, int domain)
560 uma_zone_domain_t zdom;
563 zdom = ZDOM_GET(zone, domain);
565 if (ZDOM_OWNED(zdom))
568 /* This is unsynchronized. The counter does not need to be precise. */
569 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
570 zone->uz_bucket_size++;
575 * Search for the domain with the least cached items and return it if it
576 * is out of balance with the preferred domain.
578 static __noinline int
579 zone_domain_lowest(uma_zone_t zone, int pref)
581 long least, nitems, prefitems;
585 prefitems = least = LONG_MAX;
587 for (i = 0; i < vm_ndomains; i++) {
588 nitems = ZDOM_GET(zone, i)->uzd_nitems;
589 if (nitems < least) {
596 if (prefitems < least * 2)
603 * Search for the domain with the most cached items and return it or the
604 * preferred domain if it has enough to proceed.
606 static __noinline int
607 zone_domain_highest(uma_zone_t zone, int pref)
613 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
618 for (i = 0; i < vm_ndomains; i++) {
619 nitems = ZDOM_GET(zone, i)->uzd_nitems;
630 * Safely subtract cnt from imax.
633 zone_domain_imax_sub(uma_zone_domain_t zdom, int cnt)
638 old = zdom->uzd_imax;
644 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, new) == 0);
648 * Set the maximum imax value.
651 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
655 old = zdom->uzd_imax;
659 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
663 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
664 * zone's caches. If a bucket is found the zone is not locked on return.
667 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
673 ZDOM_LOCK_ASSERT(zdom);
675 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
678 /* SMR Buckets can not be re-used until readers expire. */
679 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
680 bucket->ub_seq != SMR_SEQ_INVALID) {
681 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
683 bucket->ub_seq = SMR_SEQ_INVALID;
684 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
685 if (STAILQ_NEXT(bucket, ub_link) != NULL)
686 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
688 MPASS(zdom->uzd_nitems >= bucket->ub_cnt);
689 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
690 zdom->uzd_nitems -= bucket->ub_cnt;
693 * Shift the bounds of the current WSS interval to avoid
694 * perturbing the estimate.
697 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
698 zone_domain_imax_sub(zdom, bucket->ub_cnt);
699 } else if (zdom->uzd_imin > zdom->uzd_nitems)
700 zdom->uzd_imin = zdom->uzd_nitems;
704 for (i = 0; i < bucket->ub_cnt; i++)
705 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
712 * Insert a full bucket into the specified cache. The "ws" parameter indicates
713 * whether the bucket's contents should be counted as part of the zone's working
714 * set. The bucket may be freed if it exceeds the bucket limit.
717 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
720 uma_zone_domain_t zdom;
722 /* We don't cache empty buckets. This can happen after a reclaim. */
723 if (bucket->ub_cnt == 0)
725 zdom = zone_domain_lock(zone, domain);
728 * Conditionally set the maximum number of items.
730 zdom->uzd_nitems += bucket->ub_cnt;
731 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
733 zone_domain_imax_set(zdom, zdom->uzd_nitems);
734 if (STAILQ_EMPTY(&zdom->uzd_buckets))
735 zdom->uzd_seq = bucket->ub_seq;
736 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
740 zdom->uzd_nitems -= bucket->ub_cnt;
743 bucket_free(zone, bucket, udata);
746 /* Pops an item out of a per-cpu cache bucket. */
748 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
752 CRITICAL_ASSERT(curthread);
755 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
757 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
758 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
765 /* Pushes an item into a per-cpu cache bucket. */
767 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
770 CRITICAL_ASSERT(curthread);
771 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
772 ("uma_zfree: Freeing to non free bucket index."));
774 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
780 * Unload a UMA bucket from a per-cpu cache.
782 static inline uma_bucket_t
783 cache_bucket_unload(uma_cache_bucket_t bucket)
787 b = bucket->ucb_bucket;
789 MPASS(b->ub_entries == bucket->ucb_entries);
790 b->ub_cnt = bucket->ucb_cnt;
791 bucket->ucb_bucket = NULL;
792 bucket->ucb_entries = bucket->ucb_cnt = 0;
798 static inline uma_bucket_t
799 cache_bucket_unload_alloc(uma_cache_t cache)
802 return (cache_bucket_unload(&cache->uc_allocbucket));
805 static inline uma_bucket_t
806 cache_bucket_unload_free(uma_cache_t cache)
809 return (cache_bucket_unload(&cache->uc_freebucket));
812 static inline uma_bucket_t
813 cache_bucket_unload_cross(uma_cache_t cache)
816 return (cache_bucket_unload(&cache->uc_crossbucket));
820 * Load a bucket into a per-cpu cache bucket.
823 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
826 CRITICAL_ASSERT(curthread);
827 MPASS(bucket->ucb_bucket == NULL);
828 MPASS(b->ub_seq == SMR_SEQ_INVALID);
830 bucket->ucb_bucket = b;
831 bucket->ucb_cnt = b->ub_cnt;
832 bucket->ucb_entries = b->ub_entries;
836 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
839 cache_bucket_load(&cache->uc_allocbucket, b);
843 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
846 cache_bucket_load(&cache->uc_freebucket, b);
851 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
854 cache_bucket_load(&cache->uc_crossbucket, b);
859 * Copy and preserve ucb_spare.
862 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
865 b1->ucb_bucket = b2->ucb_bucket;
866 b1->ucb_entries = b2->ucb_entries;
867 b1->ucb_cnt = b2->ucb_cnt;
871 * Swap two cache buckets.
874 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
876 struct uma_cache_bucket b3;
878 CRITICAL_ASSERT(curthread);
880 cache_bucket_copy(&b3, b1);
881 cache_bucket_copy(b1, b2);
882 cache_bucket_copy(b2, &b3);
886 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
889 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
891 uma_zone_domain_t zdom;
895 * Avoid the lock if possible.
897 zdom = ZDOM_GET(zone, domain);
898 if (zdom->uzd_nitems == 0)
901 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
902 !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
906 * Check the zone's cache of buckets.
908 zdom = zone_domain_lock(zone, domain);
909 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
910 KASSERT(bucket->ub_cnt != 0,
911 ("cache_fetch_bucket: Returning an empty bucket."));
920 zone_log_warning(uma_zone_t zone)
922 static const struct timeval warninterval = { 300, 0 };
924 if (!zone_warnings || zone->uz_warning == NULL)
927 if (ratecheck(&zone->uz_ratecheck, &warninterval))
928 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
932 zone_maxaction(uma_zone_t zone)
935 if (zone->uz_maxaction.ta_func != NULL)
936 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
940 * Routine called by timeout which is used to fire off some time interval
941 * based calculations. (stats, hash size, etc.)
950 uma_timeout(void *unused)
953 zone_foreach(zone_timeout, NULL);
955 /* Reschedule this event */
956 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
960 * Update the working set size estimate for the zone's bucket cache.
961 * The constants chosen here are somewhat arbitrary. With an update period of
962 * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
966 zone_domain_update_wss(uma_zone_domain_t zdom)
971 MPASS(zdom->uzd_imax >= zdom->uzd_imin);
972 wss = zdom->uzd_imax - zdom->uzd_imin;
973 zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
974 zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
979 * Routine to perform timeout driven calculations. This expands the
980 * hashes and does per cpu statistics aggregation.
985 zone_timeout(uma_zone_t zone, void *unused)
990 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
996 * Hash zones are non-numa by definition so the first domain
997 * is the only one present.
1000 pages = keg->uk_domain[0].ud_pages;
1003 * Expand the keg hash table.
1005 * This is done if the number of slabs is larger than the hash size.
1006 * What I'm trying to do here is completely reduce collisions. This
1007 * may be a little aggressive. Should I allow for two collisions max?
1009 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1010 struct uma_hash newhash;
1011 struct uma_hash oldhash;
1015 * This is so involved because allocating and freeing
1016 * while the keg lock is held will lead to deadlock.
1017 * I have to do everything in stages and check for
1021 ret = hash_alloc(&newhash, 1 << fls(slabs));
1024 if (hash_expand(&keg->uk_hash, &newhash)) {
1025 oldhash = keg->uk_hash;
1026 keg->uk_hash = newhash;
1031 hash_free(&oldhash);
1038 for (int i = 0; i < vm_ndomains; i++)
1039 zone_domain_update_wss(ZDOM_GET(zone, i));
1043 * Allocate and zero fill the next sized hash table from the appropriate
1047 * hash A new hash structure with the old hash size in uh_hashsize
1050 * 1 on success and 0 on failure.
1053 hash_alloc(struct uma_hash *hash, u_int size)
1057 KASSERT(powerof2(size), ("hash size must be power of 2"));
1058 if (size > UMA_HASH_SIZE_INIT) {
1059 hash->uh_hashsize = size;
1060 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1061 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1063 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1064 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1065 UMA_ANYDOMAIN, M_WAITOK);
1066 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1068 if (hash->uh_slab_hash) {
1069 bzero(hash->uh_slab_hash, alloc);
1070 hash->uh_hashmask = hash->uh_hashsize - 1;
1078 * Expands the hash table for HASH zones. This is done from zone_timeout
1079 * to reduce collisions. This must not be done in the regular allocation
1080 * path, otherwise, we can recurse on the vm while allocating pages.
1083 * oldhash The hash you want to expand
1084 * newhash The hash structure for the new table
1092 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1094 uma_hash_slab_t slab;
1098 if (!newhash->uh_slab_hash)
1101 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1105 * I need to investigate hash algorithms for resizing without a
1109 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1110 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1111 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1112 LIST_REMOVE(slab, uhs_hlink);
1113 hval = UMA_HASH(newhash, slab->uhs_data);
1114 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1122 * Free the hash bucket to the appropriate backing store.
1125 * slab_hash The hash bucket we're freeing
1126 * hashsize The number of entries in that hash bucket
1132 hash_free(struct uma_hash *hash)
1134 if (hash->uh_slab_hash == NULL)
1136 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1137 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1139 free(hash->uh_slab_hash, M_UMAHASH);
1143 * Frees all outstanding items in a bucket
1146 * zone The zone to free to, must be unlocked.
1147 * bucket The free/alloc bucket with items.
1153 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1157 if (bucket->ub_cnt == 0)
1160 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1161 bucket->ub_seq != SMR_SEQ_INVALID) {
1162 smr_wait(zone->uz_smr, bucket->ub_seq);
1163 bucket->ub_seq = SMR_SEQ_INVALID;
1164 for (i = 0; i < bucket->ub_cnt; i++)
1165 item_dtor(zone, bucket->ub_bucket[i],
1166 zone->uz_size, NULL, SKIP_NONE);
1169 for (i = 0; i < bucket->ub_cnt; i++)
1170 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1171 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1172 if (zone->uz_max_items > 0)
1173 zone_free_limit(zone, bucket->ub_cnt);
1175 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1181 * Drains the per cpu caches for a zone.
1183 * NOTE: This may only be called while the zone is being torn down, and not
1184 * during normal operation. This is necessary in order that we do not have
1185 * to migrate CPUs to drain the per-CPU caches.
1188 * zone The zone to drain, must be unlocked.
1194 cache_drain(uma_zone_t zone)
1197 uma_bucket_t bucket;
1202 * XXX: It is safe to not lock the per-CPU caches, because we're
1203 * tearing down the zone anyway. I.e., there will be no further use
1204 * of the caches at this point.
1206 * XXX: It would good to be able to assert that the zone is being
1207 * torn down to prevent improper use of cache_drain().
1209 seq = SMR_SEQ_INVALID;
1210 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1211 seq = smr_advance(zone->uz_smr);
1213 cache = &zone->uz_cpu[cpu];
1214 bucket = cache_bucket_unload_alloc(cache);
1216 bucket_free(zone, bucket, NULL);
1217 bucket = cache_bucket_unload_free(cache);
1218 if (bucket != NULL) {
1219 bucket->ub_seq = seq;
1220 bucket_free(zone, bucket, NULL);
1222 bucket = cache_bucket_unload_cross(cache);
1223 if (bucket != NULL) {
1224 bucket->ub_seq = seq;
1225 bucket_free(zone, bucket, NULL);
1228 bucket_cache_reclaim(zone, true);
1232 cache_shrink(uma_zone_t zone, void *unused)
1235 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1238 zone->uz_bucket_size =
1239 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1243 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1246 uma_bucket_t b1, b2, b3;
1249 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1252 b1 = b2 = b3 = NULL;
1254 cache = &zone->uz_cpu[curcpu];
1255 domain = PCPU_GET(domain);
1256 b1 = cache_bucket_unload_alloc(cache);
1259 * Don't flush SMR zone buckets. This leaves the zone without a
1260 * bucket and forces every free to synchronize().
1262 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1263 b2 = cache_bucket_unload_free(cache);
1264 b3 = cache_bucket_unload_cross(cache);
1269 zone_free_bucket(zone, b1, NULL, domain, false);
1271 zone_free_bucket(zone, b2, NULL, domain, false);
1273 /* Adjust the domain so it goes to zone_free_cross. */
1274 domain = (domain + 1) % vm_ndomains;
1275 zone_free_bucket(zone, b3, NULL, domain, false);
1280 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1281 * This is an expensive call because it needs to bind to all CPUs
1282 * one by one and enter a critical section on each of them in order
1283 * to safely access their cache buckets.
1284 * Zone lock must not be held on call this function.
1287 pcpu_cache_drain_safe(uma_zone_t zone)
1292 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1295 cache_shrink(zone, NULL);
1297 zone_foreach(cache_shrink, NULL);
1300 thread_lock(curthread);
1301 sched_bind(curthread, cpu);
1302 thread_unlock(curthread);
1305 cache_drain_safe_cpu(zone, NULL);
1307 zone_foreach(cache_drain_safe_cpu, NULL);
1309 thread_lock(curthread);
1310 sched_unbind(curthread);
1311 thread_unlock(curthread);
1315 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1316 * requested a drain, otherwise the per-domain caches are trimmed to either
1317 * estimated working set size.
1320 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1322 uma_zone_domain_t zdom;
1323 uma_bucket_t bucket;
1328 * Shrink the zone bucket size to ensure that the per-CPU caches
1329 * don't grow too large.
1331 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1332 zone->uz_bucket_size--;
1334 for (i = 0; i < vm_ndomains; i++) {
1336 * The cross bucket is partially filled and not part of
1337 * the item count. Reclaim it individually here.
1339 zdom = ZDOM_GET(zone, i);
1340 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1341 ZONE_CROSS_LOCK(zone);
1342 bucket = zdom->uzd_cross;
1343 zdom->uzd_cross = NULL;
1344 ZONE_CROSS_UNLOCK(zone);
1346 bucket_free(zone, bucket, NULL);
1350 * If we were asked to drain the zone, we are done only once
1351 * this bucket cache is empty. Otherwise, we reclaim items in
1352 * excess of the zone's estimated working set size. If the
1353 * difference nitems - imin is larger than the WSS estimate,
1354 * then the estimate will grow at the end of this interval and
1355 * we ignore the historical average.
1358 target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1360 while (zdom->uzd_nitems > target) {
1361 bucket = zone_fetch_bucket(zone, zdom, true);
1364 bucket_free(zone, bucket, NULL);
1372 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1378 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1379 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1381 mem = slab_data(slab, keg);
1382 flags = slab->us_flags;
1384 if (keg->uk_fini != NULL) {
1385 for (i--; i > -1; i--)
1388 * trash_fini implies that dtor was trash_dtor. trash_fini
1389 * would check that memory hasn't been modified since free,
1390 * which executed trash_dtor.
1391 * That's why we need to run uma_dbg_kskip() check here,
1392 * albeit we don't make skip check for other init/fini
1395 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1396 keg->uk_fini != trash_fini)
1398 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1400 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1401 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1403 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1404 uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1408 * Frees pages from a keg back to the system. This is done on demand from
1409 * the pageout daemon.
1414 keg_drain(uma_keg_t keg)
1416 struct slabhead freeslabs;
1418 uma_slab_t slab, tmp;
1421 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1424 for (i = 0; i < vm_ndomains; i++) {
1425 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1426 keg->uk_name, keg, i, dom->ud_free_items);
1427 dom = &keg->uk_domain[i];
1428 LIST_INIT(&freeslabs);
1431 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1432 LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
1433 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1435 n = dom->ud_free_slabs;
1436 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1437 dom->ud_free_slabs = 0;
1438 dom->ud_free_items -= n * keg->uk_ipers;
1439 dom->ud_pages -= n * keg->uk_ppera;
1442 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1443 keg_free_slab(keg, slab, keg->uk_ipers);
1448 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1452 * Set draining to interlock with zone_dtor() so we can release our
1453 * locks as we go. Only dtor() should do a WAITOK call since it
1454 * is the only call that knows the structure will still be available
1458 while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1459 if (waitok == M_NOWAIT)
1461 msleep(zone, &ZDOM_GET(zone, 0)->uzd_lock, PVM, "zonedrain",
1464 zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1466 bucket_cache_reclaim(zone, drain);
1469 * The DRAINING flag protects us from being freed while
1470 * we're running. Normally the uma_rwlock would protect us but we
1471 * must be able to release and acquire the right lock for each keg.
1473 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1474 keg_drain(zone->uz_keg);
1476 zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1483 zone_drain(uma_zone_t zone, void *unused)
1486 zone_reclaim(zone, M_NOWAIT, true);
1490 zone_trim(uma_zone_t zone, void *unused)
1493 zone_reclaim(zone, M_NOWAIT, false);
1497 * Allocate a new slab for a keg and inserts it into the partial slab list.
1498 * The keg should be unlocked on entry. If the allocation succeeds it will
1499 * be locked on return.
1502 * flags Wait flags for the item initialization routine
1503 * aflags Wait flags for the slab allocation
1506 * The slab that was allocated or NULL if there is no memory and the
1507 * caller specified M_NOWAIT.
1510 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1521 KASSERT(domain >= 0 && domain < vm_ndomains,
1522 ("keg_alloc_slab: domain %d out of range", domain));
1524 allocf = keg->uk_allocf;
1527 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1528 uma_hash_slab_t hslab;
1529 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1533 slab = &hslab->uhs_slab;
1537 * This reproduces the old vm_zone behavior of zero filling pages the
1538 * first time they are added to a zone.
1540 * Malloced items are zeroed in uma_zalloc.
1543 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1548 if (keg->uk_flags & UMA_ZONE_NODUMP)
1551 /* zone is passed for legacy reasons. */
1552 size = keg->uk_ppera * PAGE_SIZE;
1553 mem = allocf(zone, size, domain, &sflags, aflags);
1555 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1556 zone_free_item(slabzone(keg->uk_ipers),
1557 slab_tohashslab(slab), NULL, SKIP_NONE);
1560 uma_total_inc(size);
1562 /* For HASH zones all pages go to the same uma_domain. */
1563 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1566 /* Point the slab into the allocated memory */
1567 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1568 slab = (uma_slab_t )(mem + keg->uk_pgoff);
1570 slab_tohashslab(slab)->uhs_data = mem;
1572 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1573 for (i = 0; i < keg->uk_ppera; i++)
1574 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1577 slab->us_freecount = keg->uk_ipers;
1578 slab->us_flags = sflags;
1579 slab->us_domain = domain;
1581 BIT_FILL(keg->uk_ipers, &slab->us_free);
1583 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1586 if (keg->uk_init != NULL) {
1587 for (i = 0; i < keg->uk_ipers; i++)
1588 if (keg->uk_init(slab_item(slab, keg, i),
1589 keg->uk_size, flags) != 0)
1591 if (i != keg->uk_ipers) {
1592 keg_free_slab(keg, slab, i);
1596 KEG_LOCK(keg, domain);
1598 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1599 slab, keg->uk_name, keg);
1601 if (keg->uk_flags & UMA_ZFLAG_HASH)
1602 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1605 * If we got a slab here it's safe to mark it partially used
1606 * and return. We assume that the caller is going to remove
1607 * at least one item.
1609 dom = &keg->uk_domain[domain];
1610 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1611 dom->ud_pages += keg->uk_ppera;
1612 dom->ud_free_items += keg->uk_ipers;
1621 * This function is intended to be used early on in place of page_alloc() so
1622 * that we may use the boot time page cache to satisfy allocations before
1626 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1635 pages = howmany(bytes, PAGE_SIZE);
1636 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1638 *pflag = UMA_SLAB_BOOT;
1639 m = vm_page_alloc_contig_domain(NULL, 0, domain,
1640 malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1641 (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1645 pa = VM_PAGE_TO_PHYS(m);
1646 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1647 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1648 defined(__riscv) || defined(__powerpc64__)
1649 if ((wait & M_NODUMP) == 0)
1653 /* Allocate KVA and indirectly advance bootmem. */
1654 mem = (void *)pmap_map(&bootmem, m->phys_addr,
1655 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1656 if ((wait & M_ZERO) != 0)
1657 bzero(mem, pages * PAGE_SIZE);
1663 startup_free(void *mem, vm_size_t bytes)
1668 va = (vm_offset_t)mem;
1669 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1670 pmap_remove(kernel_pmap, va, va + bytes);
1671 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1672 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1673 defined(__riscv) || defined(__powerpc64__)
1674 dump_drop_page(VM_PAGE_TO_PHYS(m));
1676 vm_page_unwire_noq(m);
1682 * Allocates a number of pages from the system
1685 * bytes The number of bytes requested
1686 * wait Shall we wait?
1689 * A pointer to the alloced memory or possibly
1690 * NULL if M_NOWAIT is set.
1693 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1696 void *p; /* Returned page */
1698 *pflag = UMA_SLAB_KERNEL;
1699 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1705 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1708 struct pglist alloctail;
1709 vm_offset_t addr, zkva;
1711 vm_page_t p, p_next;
1716 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1718 TAILQ_INIT(&alloctail);
1719 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1720 malloc2vm_flags(wait);
1721 *pflag = UMA_SLAB_KERNEL;
1722 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1723 if (CPU_ABSENT(cpu)) {
1724 p = vm_page_alloc(NULL, 0, flags);
1727 p = vm_page_alloc(NULL, 0, flags);
1729 pc = pcpu_find(cpu);
1730 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1733 p = vm_page_alloc_domain(NULL, 0,
1734 pc->pc_domain, flags);
1735 if (__predict_false(p == NULL))
1736 p = vm_page_alloc(NULL, 0, flags);
1739 if (__predict_false(p == NULL))
1741 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1743 if ((addr = kva_alloc(bytes)) == 0)
1746 TAILQ_FOREACH(p, &alloctail, listq) {
1747 pmap_qenter(zkva, &p, 1);
1750 return ((void*)addr);
1752 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1753 vm_page_unwire_noq(p);
1760 * Allocates a number of pages from within an object
1763 * bytes The number of bytes requested
1764 * wait Shall we wait?
1767 * A pointer to the alloced memory or possibly
1768 * NULL if M_NOWAIT is set.
1771 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1774 TAILQ_HEAD(, vm_page) alloctail;
1776 vm_offset_t retkva, zkva;
1777 vm_page_t p, p_next;
1780 TAILQ_INIT(&alloctail);
1783 npages = howmany(bytes, PAGE_SIZE);
1784 while (npages > 0) {
1785 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1786 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1787 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1791 * Since the page does not belong to an object, its
1794 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1799 * Page allocation failed, free intermediate pages and
1802 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1803 vm_page_unwire_noq(p);
1808 *flags = UMA_SLAB_PRIV;
1809 zkva = keg->uk_kva +
1810 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1812 TAILQ_FOREACH(p, &alloctail, listq) {
1813 pmap_qenter(zkva, &p, 1);
1817 return ((void *)retkva);
1821 * Allocate physically contiguous pages.
1824 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1828 *pflag = UMA_SLAB_KERNEL;
1829 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1830 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1834 * Frees a number of pages to the system
1837 * mem A pointer to the memory to be freed
1838 * size The size of the memory being freed
1839 * flags The original p->us_flags field
1845 page_free(void *mem, vm_size_t size, uint8_t flags)
1848 if ((flags & UMA_SLAB_BOOT) != 0) {
1849 startup_free(mem, size);
1853 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1854 ("UMA: page_free used with invalid flags %x", flags));
1856 kmem_free((vm_offset_t)mem, size);
1860 * Frees pcpu zone allocations
1863 * mem A pointer to the memory to be freed
1864 * size The size of the memory being freed
1865 * flags The original p->us_flags field
1871 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1873 vm_offset_t sva, curva;
1877 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1879 if ((flags & UMA_SLAB_BOOT) != 0) {
1880 startup_free(mem, size);
1884 sva = (vm_offset_t)mem;
1885 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1886 paddr = pmap_kextract(curva);
1887 m = PHYS_TO_VM_PAGE(paddr);
1888 vm_page_unwire_noq(m);
1891 pmap_qremove(sva, size >> PAGE_SHIFT);
1892 kva_free(sva, size);
1896 * Zero fill initializer
1898 * Arguments/Returns follow uma_init specifications
1901 zero_init(void *mem, int size, int flags)
1908 static struct noslabbits *
1909 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1912 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1917 * Actual size of embedded struct slab (!OFFPAGE).
1920 slab_sizeof(int nitems)
1924 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1925 return (roundup(s, UMA_ALIGN_PTR + 1));
1928 #define UMA_FIXPT_SHIFT 31
1929 #define UMA_FRAC_FIXPT(n, d) \
1930 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1931 #define UMA_FIXPT_PCT(f) \
1932 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1933 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
1934 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1937 * Compute the number of items that will fit in a slab. If hdr is true, the
1938 * item count may be limited to provide space in the slab for an inline slab
1939 * header. Otherwise, all slab space will be provided for item storage.
1942 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1947 /* The padding between items is not needed after the last item. */
1948 padpi = rsize - size;
1952 * Start with the maximum item count and remove items until
1953 * the slab header first alongside the allocatable memory.
1955 for (ipers = MIN(SLAB_MAX_SETSIZE,
1956 (slabsize + padpi - slab_sizeof(1)) / rsize);
1958 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1962 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1968 struct keg_layout_result {
1976 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
1977 struct keg_layout_result *kl)
1982 kl->slabsize = slabsize;
1984 /* Handle INTERNAL as inline with an extra page. */
1985 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
1986 kl->format &= ~UMA_ZFLAG_INTERNAL;
1987 kl->slabsize += PAGE_SIZE;
1990 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
1991 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
1993 /* Account for memory used by an offpage slab header. */
1994 total = kl->slabsize;
1995 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
1996 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
1998 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2002 * Determine the format of a uma keg. This determines where the slab header
2003 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2006 * keg The zone we should initialize
2012 keg_layout(uma_keg_t keg)
2014 struct keg_layout_result kl = {}, kl_tmp;
2023 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2024 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2025 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2026 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2027 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2029 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2030 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2031 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2034 alignsize = keg->uk_align + 1;
2037 * Calculate the size of each allocation (rsize) according to
2038 * alignment. If the requested size is smaller than we have
2039 * allocation bits for we round it up.
2041 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2042 rsize = roundup2(rsize, alignsize);
2044 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2046 * We want one item to start on every align boundary in a page.
2047 * To do this we will span pages. We will also extend the item
2048 * by the size of align if it is an even multiple of align.
2049 * Otherwise, it would fall on the same boundary every time.
2051 if ((rsize & alignsize) == 0)
2053 slabsize = rsize * (PAGE_SIZE / alignsize);
2054 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2055 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2056 slabsize = round_page(slabsize);
2059 * Start with a slab size of as many pages as it takes to
2060 * represent a single item. We will try to fit as many
2061 * additional items into the slab as possible.
2063 slabsize = round_page(keg->uk_size);
2066 /* Build a list of all of the available formats for this keg. */
2069 /* Evaluate an inline slab layout. */
2070 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2073 /* TODO: vm_page-embedded slab. */
2076 * We can't do OFFPAGE if we're internal or if we've been
2077 * asked to not go to the VM for buckets. If we do this we
2078 * may end up going to the VM for slabs which we do not want
2079 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2080 * In those cases, evaluate a pseudo-format called INTERNAL
2081 * which has an inline slab header and one extra page to
2082 * guarantee that it fits.
2084 * Otherwise, see if using an OFFPAGE slab will improve our
2087 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2088 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2090 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2093 * Choose a slab size and format which satisfy the minimum efficiency.
2094 * Prefer the smallest slab size that meets the constraints.
2096 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2097 * for small items (up to PAGE_SIZE), the iteration increment is one
2098 * page; and for large items, the increment is one item.
2100 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2101 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2102 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2105 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2106 round_page(rsize * (i - 1) + keg->uk_size);
2108 for (j = 0; j < nfmt; j++) {
2109 /* Only if we have no viable format yet. */
2110 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2114 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2115 if (kl_tmp.eff <= kl.eff)
2120 CTR6(KTR_UMA, "keg %s layout: format %#x "
2121 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2122 keg->uk_name, kl.format, kl.ipers, rsize,
2123 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2125 /* Stop when we reach the minimum efficiency. */
2126 if (kl.eff >= UMA_MIN_EFF)
2130 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2131 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2132 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2136 pages = atop(kl.slabsize);
2137 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2138 pages *= mp_maxid + 1;
2140 keg->uk_rsize = rsize;
2141 keg->uk_ipers = kl.ipers;
2142 keg->uk_ppera = pages;
2143 keg->uk_flags |= kl.format;
2146 * How do we find the slab header if it is offpage or if not all item
2147 * start addresses are in the same page? We could solve the latter
2148 * case with vaddr alignment, but we don't.
2150 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2151 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2152 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2153 keg->uk_flags |= UMA_ZFLAG_HASH;
2155 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2158 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2159 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2161 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2162 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2163 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2164 keg->uk_ipers, pages));
2168 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2169 * the keg onto the global keg list.
2171 * Arguments/Returns follow uma_ctor specifications
2172 * udata Actually uma_kctor_args
2175 keg_ctor(void *mem, int size, void *udata, int flags)
2177 struct uma_kctor_args *arg = udata;
2178 uma_keg_t keg = mem;
2183 keg->uk_size = arg->size;
2184 keg->uk_init = arg->uminit;
2185 keg->uk_fini = arg->fini;
2186 keg->uk_align = arg->align;
2187 keg->uk_reserve = 0;
2188 keg->uk_flags = arg->flags;
2191 * We use a global round-robin policy by default. Zones with
2192 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2193 * case the iterator is never run.
2195 keg->uk_dr.dr_policy = DOMAINSET_RR();
2196 keg->uk_dr.dr_iter = 0;
2199 * The primary zone is passed to us at keg-creation time.
2202 keg->uk_name = zone->uz_name;
2204 if (arg->flags & UMA_ZONE_ZINIT)
2205 keg->uk_init = zero_init;
2207 if (arg->flags & UMA_ZONE_MALLOC)
2208 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2211 keg->uk_flags &= ~UMA_ZONE_PCPU;
2217 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2218 * work on. Use round-robin for everything else.
2220 * Zones may override the default by specifying either.
2223 if ((keg->uk_flags &
2224 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2225 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2226 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2227 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2231 * If we haven't booted yet we need allocations to go through the
2232 * startup cache until the vm is ready.
2234 #ifdef UMA_MD_SMALL_ALLOC
2235 if (keg->uk_ppera == 1)
2236 keg->uk_allocf = uma_small_alloc;
2239 if (booted < BOOT_KVA)
2240 keg->uk_allocf = startup_alloc;
2241 else if (keg->uk_flags & UMA_ZONE_PCPU)
2242 keg->uk_allocf = pcpu_page_alloc;
2243 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2244 keg->uk_allocf = contig_alloc;
2246 keg->uk_allocf = page_alloc;
2247 #ifdef UMA_MD_SMALL_ALLOC
2248 if (keg->uk_ppera == 1)
2249 keg->uk_freef = uma_small_free;
2252 if (keg->uk_flags & UMA_ZONE_PCPU)
2253 keg->uk_freef = pcpu_page_free;
2255 keg->uk_freef = page_free;
2258 * Initialize keg's locks.
2260 for (i = 0; i < vm_ndomains; i++)
2261 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2264 * If we're putting the slab header in the actual page we need to
2265 * figure out where in each page it goes. See slab_sizeof
2268 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2271 shsize = slab_sizeof(keg->uk_ipers);
2272 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2274 * The only way the following is possible is if with our
2275 * UMA_ALIGN_PTR adjustments we are now bigger than
2276 * UMA_SLAB_SIZE. I haven't checked whether this is
2277 * mathematically possible for all cases, so we make
2280 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2281 ("zone %s ipers %d rsize %d size %d slab won't fit",
2282 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2285 if (keg->uk_flags & UMA_ZFLAG_HASH)
2286 hash_alloc(&keg->uk_hash, 0);
2288 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2290 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2292 rw_wlock(&uma_rwlock);
2293 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2294 rw_wunlock(&uma_rwlock);
2299 zone_kva_available(uma_zone_t zone, void *unused)
2303 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2307 if (keg->uk_allocf == startup_alloc) {
2308 /* Switch to the real allocator. */
2309 if (keg->uk_flags & UMA_ZONE_PCPU)
2310 keg->uk_allocf = pcpu_page_alloc;
2311 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2313 keg->uk_allocf = contig_alloc;
2315 keg->uk_allocf = page_alloc;
2320 zone_alloc_counters(uma_zone_t zone, void *unused)
2323 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2324 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2325 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2326 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2330 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2332 uma_zone_domain_t zdom;
2335 struct sysctl_oid *oid, *domainoid;
2336 int domains, i, cnt;
2337 static const char *nokeg = "cache zone";
2341 * Make a sysctl safe copy of the zone name by removing
2342 * any special characters and handling dups by appending
2345 if (zone->uz_namecnt != 0) {
2346 /* Count the number of decimal digits and '_' separator. */
2347 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2349 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2351 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2354 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2355 for (c = zone->uz_ctlname; *c != '\0'; c++)
2356 if (strchr("./\\ -", *c) != NULL)
2360 * Basic parameters at the root.
2362 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2363 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2365 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2366 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2367 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2368 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2369 zone, 0, sysctl_handle_uma_zone_flags, "A",
2370 "Allocator configuration flags");
2371 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2372 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2373 "Desired per-cpu cache size");
2374 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2375 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2376 "Maximum allowed per-cpu cache size");
2381 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2382 domains = vm_ndomains;
2385 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2386 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2388 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2389 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2390 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2391 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2392 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2393 "Real object size with alignment");
2394 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2395 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2396 "pages per-slab allocation");
2397 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2398 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2399 "items available per-slab");
2400 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2401 "align", CTLFLAG_RD, &keg->uk_align, 0,
2402 "item alignment mask");
2403 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2404 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2405 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2406 "Slab utilization (100 - internal fragmentation %)");
2407 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2408 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2409 for (i = 0; i < domains; i++) {
2410 dom = &keg->uk_domain[i];
2411 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2412 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2413 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2414 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2415 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2416 "Total pages currently allocated from VM");
2417 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2418 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2419 "items free in the slab layer");
2422 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2423 "name", CTLFLAG_RD, nokeg, "Keg name");
2426 * Information about zone limits.
2428 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2429 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2430 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2431 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2432 zone, 0, sysctl_handle_uma_zone_items, "QU",
2433 "current number of allocated items if limit is set");
2434 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2435 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2436 "Maximum number of cached items");
2437 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2438 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2439 "Number of threads sleeping at limit");
2440 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2441 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2442 "Total zone limit sleeps");
2443 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2444 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2445 "Maximum number of items in each domain's bucket cache");
2448 * Per-domain zone information.
2450 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2451 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2452 for (i = 0; i < domains; i++) {
2453 zdom = ZDOM_GET(zone, i);
2454 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2455 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2456 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2457 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2458 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2459 "number of items in this domain");
2460 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2461 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2462 "maximum item count in this period");
2463 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2464 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2465 "minimum item count in this period");
2466 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2467 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2468 "Working set size");
2472 * General statistics.
2474 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2475 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2476 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2477 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2478 zone, 1, sysctl_handle_uma_zone_cur, "I",
2479 "Current number of allocated items");
2480 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2481 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2482 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2483 "Total allocation calls");
2484 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2485 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2486 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2487 "Total free calls");
2488 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2489 "fails", CTLFLAG_RD, &zone->uz_fails,
2490 "Number of allocation failures");
2491 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2492 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2493 "Free calls from the wrong domain");
2496 struct uma_zone_count {
2502 zone_count(uma_zone_t zone, void *arg)
2504 struct uma_zone_count *cnt;
2508 * Some zones are rapidly created with identical names and
2509 * destroyed out of order. This can lead to gaps in the count.
2510 * Use one greater than the maximum observed for this name.
2512 if (strcmp(zone->uz_name, cnt->name) == 0)
2513 cnt->count = MAX(cnt->count,
2514 zone->uz_namecnt + 1);
2518 zone_update_caches(uma_zone_t zone)
2522 for (i = 0; i <= mp_maxid; i++) {
2523 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2524 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2529 * Zone header ctor. This initializes all fields, locks, etc.
2531 * Arguments/Returns follow uma_ctor specifications
2532 * udata Actually uma_zctor_args
2535 zone_ctor(void *mem, int size, void *udata, int flags)
2537 struct uma_zone_count cnt;
2538 struct uma_zctor_args *arg = udata;
2539 uma_zone_domain_t zdom;
2540 uma_zone_t zone = mem;
2546 zone->uz_name = arg->name;
2547 zone->uz_ctor = arg->ctor;
2548 zone->uz_dtor = arg->dtor;
2549 zone->uz_init = NULL;
2550 zone->uz_fini = NULL;
2551 zone->uz_sleeps = 0;
2552 zone->uz_bucket_size = 0;
2553 zone->uz_bucket_size_min = 0;
2554 zone->uz_bucket_size_max = BUCKET_MAX;
2555 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2556 zone->uz_warning = NULL;
2557 /* The domain structures follow the cpu structures. */
2558 zone->uz_bucket_max = ULONG_MAX;
2559 timevalclear(&zone->uz_ratecheck);
2561 /* Count the number of duplicate names. */
2562 cnt.name = arg->name;
2564 zone_foreach(zone_count, &cnt);
2565 zone->uz_namecnt = cnt.count;
2566 ZONE_CROSS_LOCK_INIT(zone);
2568 for (i = 0; i < vm_ndomains; i++) {
2569 zdom = ZDOM_GET(zone, i);
2570 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2571 STAILQ_INIT(&zdom->uzd_buckets);
2575 if (arg->uminit == trash_init && arg->fini == trash_fini)
2576 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2580 * This is a pure cache zone, no kegs.
2583 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2584 ("zone_ctor: Import specified for non-cache zone."));
2585 zone->uz_flags = arg->flags;
2586 zone->uz_size = arg->size;
2587 zone->uz_import = arg->import;
2588 zone->uz_release = arg->release;
2589 zone->uz_arg = arg->arg;
2592 * Cache zones are round-robin unless a policy is
2593 * specified because they may have incompatible
2596 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2597 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2599 rw_wlock(&uma_rwlock);
2600 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2601 rw_wunlock(&uma_rwlock);
2606 * Use the regular zone/keg/slab allocator.
2608 zone->uz_import = zone_import;
2609 zone->uz_release = zone_release;
2610 zone->uz_arg = zone;
2613 if (arg->flags & UMA_ZONE_SECONDARY) {
2614 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2615 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2616 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2617 zone->uz_init = arg->uminit;
2618 zone->uz_fini = arg->fini;
2619 zone->uz_flags |= UMA_ZONE_SECONDARY;
2620 rw_wlock(&uma_rwlock);
2622 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2623 if (LIST_NEXT(z, uz_link) == NULL) {
2624 LIST_INSERT_AFTER(z, zone, uz_link);
2629 rw_wunlock(&uma_rwlock);
2630 } else if (keg == NULL) {
2631 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2632 arg->align, arg->flags)) == NULL)
2635 struct uma_kctor_args karg;
2638 /* We should only be here from uma_startup() */
2639 karg.size = arg->size;
2640 karg.uminit = arg->uminit;
2641 karg.fini = arg->fini;
2642 karg.align = arg->align;
2643 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2645 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2651 /* Inherit properties from the keg. */
2653 zone->uz_size = keg->uk_size;
2654 zone->uz_flags |= (keg->uk_flags &
2655 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2658 if (booted >= BOOT_PCPU) {
2659 zone_alloc_counters(zone, NULL);
2660 if (booted >= BOOT_RUNNING)
2661 zone_alloc_sysctl(zone, NULL);
2663 zone->uz_allocs = EARLY_COUNTER;
2664 zone->uz_frees = EARLY_COUNTER;
2665 zone->uz_fails = EARLY_COUNTER;
2668 /* Caller requests a private SMR context. */
2669 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2670 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2672 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2673 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2674 ("Invalid zone flag combination"));
2675 if (arg->flags & UMA_ZFLAG_INTERNAL)
2676 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2677 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2678 zone->uz_bucket_size = BUCKET_MAX;
2679 else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2680 zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2681 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2682 zone->uz_bucket_size = 0;
2684 zone->uz_bucket_size = bucket_select(zone->uz_size);
2685 zone->uz_bucket_size_min = zone->uz_bucket_size;
2686 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2687 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2688 zone_update_caches(zone);
2694 * Keg header dtor. This frees all data, destroys locks, frees the hash
2695 * table and removes the keg from the global list.
2697 * Arguments/Returns follow uma_dtor specifications
2701 keg_dtor(void *arg, int size, void *udata)
2704 uint32_t free, pages;
2707 keg = (uma_keg_t)arg;
2709 for (i = 0; i < vm_ndomains; i++) {
2710 free += keg->uk_domain[i].ud_free_items;
2711 pages += keg->uk_domain[i].ud_pages;
2712 KEG_LOCK_FINI(keg, i);
2715 printf("Freed UMA keg (%s) was not empty (%u items). "
2716 " Lost %u pages of memory.\n",
2717 keg->uk_name ? keg->uk_name : "",
2718 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2720 hash_free(&keg->uk_hash);
2726 * Arguments/Returns follow uma_dtor specifications
2730 zone_dtor(void *arg, int size, void *udata)
2736 zone = (uma_zone_t)arg;
2738 sysctl_remove_oid(zone->uz_oid, 1, 1);
2740 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2743 rw_wlock(&uma_rwlock);
2744 LIST_REMOVE(zone, uz_link);
2745 rw_wunlock(&uma_rwlock);
2746 zone_reclaim(zone, M_WAITOK, true);
2749 * We only destroy kegs from non secondary/non cache zones.
2751 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2753 rw_wlock(&uma_rwlock);
2754 LIST_REMOVE(keg, uk_link);
2755 rw_wunlock(&uma_rwlock);
2756 zone_free_item(kegs, keg, NULL, SKIP_NONE);
2758 counter_u64_free(zone->uz_allocs);
2759 counter_u64_free(zone->uz_frees);
2760 counter_u64_free(zone->uz_fails);
2761 counter_u64_free(zone->uz_xdomain);
2762 free(zone->uz_ctlname, M_UMA);
2763 for (i = 0; i < vm_ndomains; i++)
2764 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2765 ZONE_CROSS_LOCK_FINI(zone);
2769 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2774 LIST_FOREACH(keg, &uma_kegs, uk_link) {
2775 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2778 LIST_FOREACH(zone, &uma_cachezones, uz_link)
2783 * Traverses every zone in the system and calls a callback
2786 * zfunc A pointer to a function which accepts a zone
2793 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2796 rw_rlock(&uma_rwlock);
2797 zone_foreach_unlocked(zfunc, arg);
2798 rw_runlock(&uma_rwlock);
2802 * Initialize the kernel memory allocator. This is done after pages can be
2803 * allocated but before general KVA is available.
2806 uma_startup1(vm_offset_t virtual_avail)
2808 struct uma_zctor_args args;
2809 size_t ksize, zsize, size;
2810 uma_keg_t primarykeg;
2815 bootstart = bootmem = virtual_avail;
2817 rw_init(&uma_rwlock, "UMA lock");
2818 sx_init(&uma_reclaim_lock, "umareclaim");
2820 ksize = sizeof(struct uma_keg) +
2821 (sizeof(struct uma_domain) * vm_ndomains);
2822 ksize = roundup(ksize, UMA_SUPER_ALIGN);
2823 zsize = sizeof(struct uma_zone) +
2824 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2825 (sizeof(struct uma_zone_domain) * vm_ndomains);
2826 zsize = roundup(zsize, UMA_SUPER_ALIGN);
2828 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2829 size = (zsize * 2) + ksize;
2830 for (domain = 0; domain < vm_ndomains; domain++) {
2831 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
2836 zones = (uma_zone_t)m;
2838 kegs = (uma_zone_t)m;
2840 primarykeg = (uma_keg_t)m;
2842 /* "manually" create the initial zone */
2843 memset(&args, 0, sizeof(args));
2844 args.name = "UMA Kegs";
2846 args.ctor = keg_ctor;
2847 args.dtor = keg_dtor;
2848 args.uminit = zero_init;
2850 args.keg = primarykeg;
2851 args.align = UMA_SUPER_ALIGN - 1;
2852 args.flags = UMA_ZFLAG_INTERNAL;
2853 zone_ctor(kegs, zsize, &args, M_WAITOK);
2855 args.name = "UMA Zones";
2857 args.ctor = zone_ctor;
2858 args.dtor = zone_dtor;
2859 args.uminit = zero_init;
2862 args.align = UMA_SUPER_ALIGN - 1;
2863 args.flags = UMA_ZFLAG_INTERNAL;
2864 zone_ctor(zones, zsize, &args, M_WAITOK);
2866 /* Now make zones for slab headers */
2867 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2868 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2869 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2870 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2872 hashzone = uma_zcreate("UMA Hash",
2873 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2874 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2880 #ifndef UMA_MD_SMALL_ALLOC
2881 extern void vm_radix_reserve_kva(void);
2885 * Advertise the availability of normal kva allocations and switch to
2886 * the default back-end allocator. Marks the KVA we consumed on startup
2887 * as used in the map.
2893 if (bootstart != bootmem) {
2894 vm_map_lock(kernel_map);
2895 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2896 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2897 vm_map_unlock(kernel_map);
2900 #ifndef UMA_MD_SMALL_ALLOC
2901 /* Set up radix zone to use noobj_alloc. */
2902 vm_radix_reserve_kva();
2906 zone_foreach_unlocked(zone_kva_available, NULL);
2911 * Allocate counters as early as possible so that boot-time allocations are
2912 * accounted more precisely.
2915 uma_startup_pcpu(void *arg __unused)
2918 zone_foreach_unlocked(zone_alloc_counters, NULL);
2921 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
2924 * Finish our initialization steps.
2927 uma_startup3(void *arg __unused)
2931 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2932 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2933 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2935 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2936 callout_init(&uma_callout, 1);
2937 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2938 booted = BOOT_RUNNING;
2940 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2941 EVENTHANDLER_PRI_FIRST);
2943 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
2949 booted = BOOT_SHUTDOWN;
2953 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2954 int align, uint32_t flags)
2956 struct uma_kctor_args args;
2959 args.uminit = uminit;
2961 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2964 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2967 /* Public functions */
2970 uma_set_align(int align)
2973 if (align != UMA_ALIGN_CACHE)
2974 uma_align_cache = align;
2979 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2980 uma_init uminit, uma_fini fini, int align, uint32_t flags)
2983 struct uma_zctor_args args;
2986 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2989 /* This stuff is essential for the zone ctor */
2990 memset(&args, 0, sizeof(args));
2995 args.uminit = uminit;
2999 * Inject procedures which check for memory use after free if we are
3000 * allowed to scramble the memory while it is not allocated. This
3001 * requires that: UMA is actually able to access the memory, no init
3002 * or fini procedures, no dependency on the initial value of the
3003 * memory, and no (legitimate) use of the memory after free. Note,
3004 * the ctor and dtor do not need to be empty.
3006 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3007 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3008 args.uminit = trash_init;
3009 args.fini = trash_fini;
3016 sx_slock(&uma_reclaim_lock);
3017 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3018 sx_sunlock(&uma_reclaim_lock);
3025 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3026 uma_init zinit, uma_fini zfini, uma_zone_t primary)
3028 struct uma_zctor_args args;
3032 keg = primary->uz_keg;
3033 memset(&args, 0, sizeof(args));
3035 args.size = keg->uk_size;
3038 args.uminit = zinit;
3040 args.align = keg->uk_align;
3041 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3044 sx_slock(&uma_reclaim_lock);
3045 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3046 sx_sunlock(&uma_reclaim_lock);
3053 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3054 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3055 void *arg, int flags)
3057 struct uma_zctor_args args;
3059 memset(&args, 0, sizeof(args));
3064 args.uminit = zinit;
3066 args.import = zimport;
3067 args.release = zrelease;
3070 args.flags = flags | UMA_ZFLAG_CACHE;
3072 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3077 uma_zdestroy(uma_zone_t zone)
3081 * Large slabs are expensive to reclaim, so don't bother doing
3082 * unnecessary work if we're shutting down.
3084 if (booted == BOOT_SHUTDOWN &&
3085 zone->uz_fini == NULL && zone->uz_release == zone_release)
3087 sx_slock(&uma_reclaim_lock);
3088 zone_free_item(zones, zone, NULL, SKIP_NONE);
3089 sx_sunlock(&uma_reclaim_lock);
3093 uma_zwait(uma_zone_t zone)
3096 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3097 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3098 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3099 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3101 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3105 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3107 void *item, *pcpu_item;
3111 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3113 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3116 pcpu_item = zpcpu_base_to_offset(item);
3117 if (flags & M_ZERO) {
3119 for (i = 0; i <= mp_maxid; i++)
3120 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3122 bzero(item, zone->uz_size);
3129 * A stub while both regular and pcpu cases are identical.
3132 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3137 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3139 item = zpcpu_offset_to_base(pcpu_item);
3140 uma_zfree_arg(zone, item, udata);
3143 static inline void *
3144 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3150 skipdbg = uma_dbg_zskip(zone, item);
3151 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3152 zone->uz_ctor != trash_ctor)
3153 trash_ctor(item, size, udata, flags);
3155 /* Check flags before loading ctor pointer. */
3156 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3157 __predict_false(zone->uz_ctor != NULL) &&
3158 zone->uz_ctor(item, size, udata, flags) != 0) {
3159 counter_u64_add(zone->uz_fails, 1);
3160 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3165 uma_dbg_alloc(zone, NULL, item);
3167 if (__predict_false(flags & M_ZERO))
3168 return (memset(item, 0, size));
3174 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3175 enum zfreeskip skip)
3180 skipdbg = uma_dbg_zskip(zone, item);
3181 if (skip == SKIP_NONE && !skipdbg) {
3182 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3183 uma_dbg_free(zone, udata, item);
3185 uma_dbg_free(zone, NULL, item);
3188 if (__predict_true(skip < SKIP_DTOR)) {
3189 if (zone->uz_dtor != NULL)
3190 zone->uz_dtor(item, size, udata);
3192 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3193 zone->uz_dtor != trash_dtor)
3194 trash_dtor(item, size, udata);
3201 item_domain(void *item)
3205 domain = _vm_phys_domain(vtophys(item));
3206 KASSERT(domain >= 0 && domain < vm_ndomains,
3207 ("%s: unknown domain for item %p", __func__, item));
3212 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3213 #define UMA_ZALLOC_DEBUG
3215 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3221 if (flags & M_WAITOK) {
3222 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3223 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3228 KASSERT((flags & M_EXEC) == 0,
3229 ("uma_zalloc_debug: called with M_EXEC"));
3230 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3231 ("uma_zalloc_debug: called within spinlock or critical section"));
3232 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3233 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3236 #ifdef DEBUG_MEMGUARD
3237 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3239 item = memguard_alloc(zone->uz_size, flags);
3241 error = EJUSTRETURN;
3242 if (zone->uz_init != NULL &&
3243 zone->uz_init(item, zone->uz_size, flags) != 0) {
3247 if (zone->uz_ctor != NULL &&
3248 zone->uz_ctor(item, zone->uz_size, udata,
3250 counter_u64_add(zone->uz_fails, 1);
3251 zone->uz_fini(item, zone->uz_size);
3258 /* This is unfortunate but should not be fatal. */
3265 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3267 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3268 ("uma_zfree_debug: called with spinlock or critical section held"));
3270 #ifdef DEBUG_MEMGUARD
3271 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3272 if (zone->uz_dtor != NULL)
3273 zone->uz_dtor(item, zone->uz_size, udata);
3274 if (zone->uz_fini != NULL)
3275 zone->uz_fini(item, zone->uz_size);
3276 memguard_free(item);
3277 return (EJUSTRETURN);
3284 static inline void *
3285 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3286 void *udata, int flags)
3291 item = cache_bucket_pop(cache, bucket);
3292 size = cache_uz_size(cache);
3293 uz_flags = cache_uz_flags(cache);
3295 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3298 static __noinline void *
3299 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3301 uma_cache_bucket_t bucket;
3304 while (cache_alloc(zone, cache, udata, flags)) {
3305 cache = &zone->uz_cpu[curcpu];
3306 bucket = &cache->uc_allocbucket;
3307 if (__predict_false(bucket->ucb_cnt == 0))
3309 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3314 * We can not get a bucket so try to return a single item.
3316 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3317 domain = PCPU_GET(domain);
3319 domain = UMA_ANYDOMAIN;
3320 return (zone_alloc_item(zone, udata, domain, flags));
3325 uma_zalloc_smr(uma_zone_t zone, int flags)
3327 uma_cache_bucket_t bucket;
3330 #ifdef UMA_ZALLOC_DEBUG
3333 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3334 ("uma_zalloc_arg: called with non-SMR zone.\n"));
3335 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3340 cache = &zone->uz_cpu[curcpu];
3341 bucket = &cache->uc_allocbucket;
3342 if (__predict_false(bucket->ucb_cnt == 0))
3343 return (cache_alloc_retry(zone, cache, NULL, flags));
3344 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3349 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3351 uma_cache_bucket_t bucket;
3354 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3355 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3357 /* This is the fast path allocation */
3358 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3361 #ifdef UMA_ZALLOC_DEBUG
3364 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3365 ("uma_zalloc_arg: called with SMR zone.\n"));
3366 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3371 * If possible, allocate from the per-CPU cache. There are two
3372 * requirements for safe access to the per-CPU cache: (1) the thread
3373 * accessing the cache must not be preempted or yield during access,
3374 * and (2) the thread must not migrate CPUs without switching which
3375 * cache it accesses. We rely on a critical section to prevent
3376 * preemption and migration. We release the critical section in
3377 * order to acquire the zone mutex if we are unable to allocate from
3378 * the current cache; when we re-acquire the critical section, we
3379 * must detect and handle migration if it has occurred.
3382 cache = &zone->uz_cpu[curcpu];
3383 bucket = &cache->uc_allocbucket;
3384 if (__predict_false(bucket->ucb_cnt == 0))
3385 return (cache_alloc_retry(zone, cache, udata, flags));
3386 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3390 * Replenish an alloc bucket and possibly restore an old one. Called in
3391 * a critical section. Returns in a critical section.
3393 * A false return value indicates an allocation failure.
3394 * A true return value indicates success and the caller should retry.
3396 static __noinline bool
3397 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3399 uma_bucket_t bucket;
3400 int curdomain, domain;
3403 CRITICAL_ASSERT(curthread);
3406 * If we have run out of items in our alloc bucket see
3407 * if we can switch with the free bucket.
3409 * SMR Zones can't re-use the free bucket until the sequence has
3412 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3413 cache->uc_freebucket.ucb_cnt != 0) {
3414 cache_bucket_swap(&cache->uc_freebucket,
3415 &cache->uc_allocbucket);
3420 * Discard any empty allocation bucket while we hold no locks.
3422 bucket = cache_bucket_unload_alloc(cache);
3425 if (bucket != NULL) {
3426 KASSERT(bucket->ub_cnt == 0,
3427 ("cache_alloc: Entered with non-empty alloc bucket."));
3428 bucket_free(zone, bucket, udata);
3432 * Attempt to retrieve the item from the per-CPU cache has failed, so
3433 * we must go back to the zone. This requires the zdom lock, so we
3434 * must drop the critical section, then re-acquire it when we go back
3435 * to the cache. Since the critical section is released, we may be
3436 * preempted or migrate. As such, make sure not to maintain any
3437 * thread-local state specific to the cache from prior to releasing
3438 * the critical section.
3440 domain = PCPU_GET(domain);
3441 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3442 VM_DOMAIN_EMPTY(domain))
3443 domain = zone_domain_highest(zone, domain);
3444 bucket = cache_fetch_bucket(zone, cache, domain);
3445 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3446 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3452 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3453 zone->uz_name, zone, bucket);
3454 if (bucket == NULL) {
3460 * See if we lost the race or were migrated. Cache the
3461 * initialized bucket to make this less likely or claim
3462 * the memory directly.
3465 cache = &zone->uz_cpu[curcpu];
3466 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3467 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3468 (curdomain = PCPU_GET(domain)) == domain ||
3469 VM_DOMAIN_EMPTY(curdomain))) {
3471 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3473 cache_bucket_load_alloc(cache, bucket);
3478 * We lost the race, release this bucket and start over.
3481 zone_put_bucket(zone, domain, bucket, udata, false);
3488 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3491 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3492 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3494 /* This is the fast path allocation */
3495 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3496 zone->uz_name, zone, domain, flags);
3498 if (flags & M_WAITOK) {
3499 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3500 "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3502 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3503 ("uma_zalloc_domain: called with spinlock or critical section held"));
3505 return (zone_alloc_item(zone, udata, domain, flags));
3509 * Find a slab with some space. Prefer slabs that are partially used over those
3510 * that are totally full. This helps to reduce fragmentation.
3512 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3516 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3522 KASSERT(domain >= 0 && domain < vm_ndomains,
3523 ("keg_first_slab: domain %d out of range", domain));
3524 KEG_LOCK_ASSERT(keg, domain);
3529 dom = &keg->uk_domain[domain];
3530 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3532 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3533 LIST_REMOVE(slab, us_link);
3534 dom->ud_free_slabs--;
3535 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3539 domain = (domain + 1) % vm_ndomains;
3540 } while (domain != start);
3546 * Fetch an existing slab from a free or partial list. Returns with the
3547 * keg domain lock held if a slab was found or unlocked if not.
3550 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3555 /* HASH has a single free list. */
3556 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3559 KEG_LOCK(keg, domain);
3560 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3561 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3562 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3563 KEG_UNLOCK(keg, domain);
3570 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3572 struct vm_domainset_iter di;
3579 * Use the keg's policy if upper layers haven't already specified a
3580 * domain (as happens with first-touch zones).
3582 * To avoid races we run the iterator with the keg lock held, but that
3583 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3584 * clear M_WAITOK and handle low memory conditions locally.
3586 rr = rdomain == UMA_ANYDOMAIN;
3588 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3589 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3597 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3602 * M_NOVM means don't ask at all!
3607 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3610 if (!rr && (flags & M_WAITOK) == 0)
3612 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3613 if ((flags & M_WAITOK) != 0) {
3614 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
3622 * We might not have been able to get a slab but another cpu
3623 * could have while we were unlocked. Check again before we
3626 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3633 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3639 KEG_LOCK_ASSERT(keg, slab->us_domain);
3641 dom = &keg->uk_domain[slab->us_domain];
3642 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3643 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3644 item = slab_item(slab, keg, freei);
3645 slab->us_freecount--;
3646 dom->ud_free_items--;
3649 * Move this slab to the full list. It must be on the partial list, so
3650 * we do not need to update the free slab count. In particular,
3651 * keg_fetch_slab() always returns slabs on the partial list.
3653 if (slab->us_freecount == 0) {
3654 LIST_REMOVE(slab, us_link);
3655 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3662 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3676 /* Try to keep the buckets totally full */
3677 for (i = 0; i < max; ) {
3678 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3681 stripe = howmany(max, vm_ndomains);
3683 dom = &keg->uk_domain[slab->us_domain];
3684 while (slab->us_freecount && i < max) {
3685 bucket[i++] = slab_alloc_item(keg, slab);
3686 if (dom->ud_free_items <= keg->uk_reserve)
3690 * If the zone is striped we pick a new slab for every
3691 * N allocations. Eliminating this conditional will
3692 * instead pick a new domain for each bucket rather
3693 * than stripe within each bucket. The current option
3694 * produces more fragmentation and requires more cpu
3695 * time but yields better distribution.
3697 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3698 vm_ndomains > 1 && --stripe == 0)
3702 KEG_UNLOCK(keg, slab->us_domain);
3703 /* Don't block if we allocated any successfully. */
3712 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3714 uint64_t old, new, total, max;
3717 * The hard case. We're going to sleep because there were existing
3718 * sleepers or because we ran out of items. This routine enforces
3719 * fairness by keeping fifo order.
3721 * First release our ill gotten gains and make some noise.
3724 zone_free_limit(zone, count);
3725 zone_log_warning(zone);
3726 zone_maxaction(zone);
3727 if (flags & M_NOWAIT)
3731 * We need to allocate an item or set ourself as a sleeper
3732 * while the sleepq lock is held to avoid wakeup races. This
3733 * is essentially a home rolled semaphore.
3735 sleepq_lock(&zone->uz_max_items);
3736 old = zone->uz_items;
3738 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3739 /* Cache the max since we will evaluate twice. */
3740 max = zone->uz_max_items;
3741 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3742 UZ_ITEMS_COUNT(old) >= max)
3743 new = old + UZ_ITEMS_SLEEPER;
3745 new = old + MIN(count, max - old);
3746 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3748 /* We may have successfully allocated under the sleepq lock. */
3749 if (UZ_ITEMS_SLEEPERS(new) == 0) {
3750 sleepq_release(&zone->uz_max_items);
3755 * This is in a different cacheline from uz_items so that we
3756 * don't constantly invalidate the fastpath cacheline when we
3757 * adjust item counts. This could be limited to toggling on
3760 atomic_add_32(&zone->uz_sleepers, 1);
3761 atomic_add_64(&zone->uz_sleeps, 1);
3764 * We have added ourselves as a sleeper. The sleepq lock
3765 * protects us from wakeup races. Sleep now and then retry.
3767 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3768 sleepq_wait(&zone->uz_max_items, PVM);
3771 * After wakeup, remove ourselves as a sleeper and try
3772 * again. We no longer have the sleepq lock for protection.
3774 * Subract ourselves as a sleeper while attempting to add
3777 atomic_subtract_32(&zone->uz_sleepers, 1);
3778 old = atomic_fetchadd_64(&zone->uz_items,
3779 -(UZ_ITEMS_SLEEPER - count));
3780 /* We're no longer a sleeper. */
3781 old -= UZ_ITEMS_SLEEPER;
3784 * If we're still at the limit, restart. Notably do not
3785 * block on other sleepers. Cache the max value to protect
3786 * against changes via sysctl.
3788 total = UZ_ITEMS_COUNT(old);
3789 max = zone->uz_max_items;
3792 /* Truncate if necessary, otherwise wake other sleepers. */
3793 if (total + count > max) {
3794 zone_free_limit(zone, total + count - max);
3795 count = max - total;
3796 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3797 wakeup_one(&zone->uz_max_items);
3804 * Allocate 'count' items from our max_items limit. Returns the number
3805 * available. If M_NOWAIT is not specified it will sleep until at least
3806 * one item can be allocated.
3809 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3814 max = zone->uz_max_items;
3818 * We expect normal allocations to succeed with a simple
3821 old = atomic_fetchadd_64(&zone->uz_items, count);
3822 if (__predict_true(old + count <= max))
3826 * If we had some items and no sleepers just return the
3827 * truncated value. We have to release the excess space
3828 * though because that may wake sleepers who weren't woken
3829 * because we were temporarily over the limit.
3832 zone_free_limit(zone, (old + count) - max);
3835 return (zone_alloc_limit_hard(zone, count, flags));
3839 * Free a number of items back to the limit.
3842 zone_free_limit(uma_zone_t zone, int count)
3849 * In the common case we either have no sleepers or
3850 * are still over the limit and can just return.
3852 old = atomic_fetchadd_64(&zone->uz_items, -count);
3853 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3854 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3858 * Moderate the rate of wakeups. Sleepers will continue
3859 * to generate wakeups if necessary.
3861 wakeup_one(&zone->uz_max_items);
3865 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3867 uma_bucket_t bucket;
3870 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3873 /* Avoid allocs targeting empty domains. */
3874 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3875 domain = UMA_ANYDOMAIN;
3876 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
3877 domain = UMA_ANYDOMAIN;
3879 if (zone->uz_max_items > 0)
3880 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3883 maxbucket = zone->uz_bucket_size;
3887 /* Don't wait for buckets, preserve caller's NOVM setting. */
3888 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3889 if (bucket == NULL) {
3894 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3895 MIN(maxbucket, bucket->ub_entries), domain, flags);
3898 * Initialize the memory if necessary.
3900 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3903 for (i = 0; i < bucket->ub_cnt; i++)
3904 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3908 * If we couldn't initialize the whole bucket, put the
3909 * rest back onto the freelist.
3911 if (i != bucket->ub_cnt) {
3912 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3913 bucket->ub_cnt - i);
3915 bzero(&bucket->ub_bucket[i],
3916 sizeof(void *) * (bucket->ub_cnt - i));
3922 cnt = bucket->ub_cnt;
3923 if (bucket->ub_cnt == 0) {
3924 bucket_free(zone, bucket, udata);
3925 counter_u64_add(zone->uz_fails, 1);
3929 if (zone->uz_max_items > 0 && cnt < maxbucket)
3930 zone_free_limit(zone, maxbucket - cnt);
3936 * Allocates a single item from a zone.
3939 * zone The zone to alloc for.
3940 * udata The data to be passed to the constructor.
3941 * domain The domain to allocate from or UMA_ANYDOMAIN.
3942 * flags M_WAITOK, M_NOWAIT, M_ZERO.
3945 * NULL if there is no memory and M_NOWAIT is set
3946 * An item if successful
3950 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
3954 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
3955 counter_u64_add(zone->uz_fails, 1);
3959 /* Avoid allocs targeting empty domains. */
3960 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3961 domain = UMA_ANYDOMAIN;
3963 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
3967 * We have to call both the zone's init (not the keg's init)
3968 * and the zone's ctor. This is because the item is going from
3969 * a keg slab directly to the user, and the user is expecting it
3970 * to be both zone-init'd as well as zone-ctor'd.
3972 if (zone->uz_init != NULL) {
3973 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
3974 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
3978 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
3983 counter_u64_add(zone->uz_allocs, 1);
3984 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
3985 zone->uz_name, zone);
3990 counter_u64_add(zone->uz_fails, 1);
3992 if (zone->uz_max_items > 0)
3993 zone_free_limit(zone, 1);
3994 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
3995 zone->uz_name, zone);
4002 uma_zfree_smr(uma_zone_t zone, void *item)
4005 uma_cache_bucket_t bucket;
4006 int itemdomain, uz_flags;
4008 #ifdef UMA_ZALLOC_DEBUG
4009 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4010 ("uma_zfree_smr: called with non-SMR zone.\n"));
4011 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4012 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4013 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4016 cache = &zone->uz_cpu[curcpu];
4017 uz_flags = cache_uz_flags(cache);
4020 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4021 itemdomain = item_domain(item);
4025 cache = &zone->uz_cpu[curcpu];
4026 /* SMR Zones must free to the free bucket. */
4027 bucket = &cache->uc_freebucket;
4029 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4030 PCPU_GET(domain) != itemdomain) {
4031 bucket = &cache->uc_crossbucket;
4034 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4035 cache_bucket_push(cache, bucket, item);
4039 } while (cache_free(zone, cache, NULL, item, itemdomain));
4043 * If nothing else caught this, we'll just do an internal free.
4045 zone_free_item(zone, item, NULL, SKIP_NONE);
4050 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4053 uma_cache_bucket_t bucket;
4054 int itemdomain, uz_flags;
4056 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4057 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4059 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4061 #ifdef UMA_ZALLOC_DEBUG
4062 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4063 ("uma_zfree_arg: called with SMR zone.\n"));
4064 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4067 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4072 * We are accessing the per-cpu cache without a critical section to
4073 * fetch size and flags. This is acceptable, if we are preempted we
4074 * will simply read another cpu's line.
4076 cache = &zone->uz_cpu[curcpu];
4077 uz_flags = cache_uz_flags(cache);
4078 if (UMA_ALWAYS_CTORDTOR ||
4079 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4080 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4083 * The race here is acceptable. If we miss it we'll just have to wait
4084 * a little longer for the limits to be reset.
4086 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4087 if (zone->uz_sleepers > 0)
4092 * If possible, free to the per-CPU cache. There are two
4093 * requirements for safe access to the per-CPU cache: (1) the thread
4094 * accessing the cache must not be preempted or yield during access,
4095 * and (2) the thread must not migrate CPUs without switching which
4096 * cache it accesses. We rely on a critical section to prevent
4097 * preemption and migration. We release the critical section in
4098 * order to acquire the zone mutex if we are unable to free to the
4099 * current cache; when we re-acquire the critical section, we must
4100 * detect and handle migration if it has occurred.
4104 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4105 itemdomain = item_domain(item);
4109 cache = &zone->uz_cpu[curcpu];
4111 * Try to free into the allocbucket first to give LIFO
4112 * ordering for cache-hot datastructures. Spill over
4113 * into the freebucket if necessary. Alloc will swap
4114 * them if one runs dry.
4116 bucket = &cache->uc_allocbucket;
4118 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4119 PCPU_GET(domain) != itemdomain) {
4120 bucket = &cache->uc_crossbucket;
4123 if (bucket->ucb_cnt == bucket->ucb_entries &&
4124 cache->uc_freebucket.ucb_cnt <
4125 cache->uc_freebucket.ucb_entries)
4126 cache_bucket_swap(&cache->uc_freebucket,
4127 &cache->uc_allocbucket);
4128 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4129 cache_bucket_push(cache, bucket, item);
4133 } while (cache_free(zone, cache, udata, item, itemdomain));
4137 * If nothing else caught this, we'll just do an internal free.
4140 zone_free_item(zone, item, udata, SKIP_DTOR);
4145 * sort crossdomain free buckets to domain correct buckets and cache
4149 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4151 struct uma_bucketlist fullbuckets;
4152 uma_zone_domain_t zdom;
4159 "uma_zfree: zone %s(%p) draining cross bucket %p",
4160 zone->uz_name, zone, bucket);
4163 * It is possible for buckets to arrive here out of order so we fetch
4164 * the current smr seq rather than accepting the bucket's.
4166 seq = SMR_SEQ_INVALID;
4167 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4168 seq = smr_advance(zone->uz_smr);
4171 * To avoid having ndomain * ndomain buckets for sorting we have a
4172 * lock on the current crossfree bucket. A full matrix with
4173 * per-domain locking could be used if necessary.
4175 STAILQ_INIT(&fullbuckets);
4176 ZONE_CROSS_LOCK(zone);
4177 while (bucket->ub_cnt > 0) {
4178 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4179 domain = item_domain(item);
4180 zdom = ZDOM_GET(zone, domain);
4181 if (zdom->uzd_cross == NULL) {
4182 zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
4183 if (zdom->uzd_cross == NULL)
4186 b = zdom->uzd_cross;
4187 b->ub_bucket[b->ub_cnt++] = item;
4189 if (b->ub_cnt == b->ub_entries) {
4190 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4191 zdom->uzd_cross = NULL;
4195 ZONE_CROSS_UNLOCK(zone);
4196 if (bucket->ub_cnt == 0)
4197 bucket->ub_seq = SMR_SEQ_INVALID;
4198 bucket_free(zone, bucket, udata);
4200 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4201 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4202 domain = item_domain(b->ub_bucket[0]);
4203 zone_put_bucket(zone, domain, b, udata, true);
4209 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4210 int itemdomain, bool ws)
4215 * Buckets coming from the wrong domain will be entirely for the
4216 * only other domain on two domain systems. In this case we can
4217 * simply cache them. Otherwise we need to sort them back to
4220 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4221 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4222 zone_free_cross(zone, bucket, udata);
4228 * Attempt to save the bucket in the zone's domain bucket cache.
4231 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4232 zone->uz_name, zone, bucket);
4233 /* ub_cnt is pointing to the last free item */
4234 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4235 itemdomain = zone_domain_lowest(zone, itemdomain);
4236 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4240 * Populate a free or cross bucket for the current cpu cache. Free any
4241 * existing full bucket either to the zone cache or back to the slab layer.
4243 * Enters and returns in a critical section. false return indicates that
4244 * we can not satisfy this free in the cache layer. true indicates that
4245 * the caller should retry.
4247 static __noinline bool
4248 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4251 uma_cache_bucket_t cbucket;
4252 uma_bucket_t newbucket, bucket;
4254 CRITICAL_ASSERT(curthread);
4256 if (zone->uz_bucket_size == 0)
4259 cache = &zone->uz_cpu[curcpu];
4263 * FIRSTTOUCH domains need to free to the correct zdom. When
4264 * enabled this is the zdom of the item. The bucket is the
4265 * cross bucket if the current domain and itemdomain do not match.
4267 cbucket = &cache->uc_freebucket;
4269 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4270 if (PCPU_GET(domain) != itemdomain) {
4271 cbucket = &cache->uc_crossbucket;
4272 if (cbucket->ucb_cnt != 0)
4273 counter_u64_add(zone->uz_xdomain,
4278 bucket = cache_bucket_unload(cbucket);
4279 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4280 ("cache_free: Entered with non-full free bucket."));
4282 /* We are no longer associated with this CPU. */
4286 * Don't let SMR zones operate without a free bucket. Force
4287 * a synchronize and re-use this one. We will only degrade
4288 * to a synchronize every bucket_size items rather than every
4289 * item if we fail to allocate a bucket.
4291 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4293 bucket->ub_seq = smr_advance(zone->uz_smr);
4294 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4295 if (newbucket == NULL && bucket != NULL) {
4296 bucket_drain(zone, bucket);
4300 } else if (!bucketdisable)
4301 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4304 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4307 if ((bucket = newbucket) == NULL)
4309 cache = &zone->uz_cpu[curcpu];
4312 * Check to see if we should be populating the cross bucket. If it
4313 * is already populated we will fall through and attempt to populate
4316 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4317 if (PCPU_GET(domain) != itemdomain &&
4318 cache->uc_crossbucket.ucb_bucket == NULL) {
4319 cache_bucket_load_cross(cache, bucket);
4325 * We may have lost the race to fill the bucket or switched CPUs.
4327 if (cache->uc_freebucket.ucb_bucket != NULL) {
4329 bucket_free(zone, bucket, udata);
4332 cache_bucket_load_free(cache, bucket);
4338 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4345 KEG_LOCK_ASSERT(keg, slab->us_domain);
4347 /* Do we need to remove from any lists? */
4348 dom = &keg->uk_domain[slab->us_domain];
4349 if (slab->us_freecount + 1 == keg->uk_ipers) {
4350 LIST_REMOVE(slab, us_link);
4351 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4352 dom->ud_free_slabs++;
4353 } else if (slab->us_freecount == 0) {
4354 LIST_REMOVE(slab, us_link);
4355 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4358 /* Slab management. */
4359 freei = slab_item_index(slab, keg, item);
4360 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4361 slab->us_freecount++;
4363 /* Keg statistics. */
4364 dom->ud_free_items++;
4368 zone_release(void *arg, void **bucket, int cnt)
4381 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4382 lock = KEG_LOCK(keg, 0);
4383 for (i = 0; i < cnt; i++) {
4385 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4386 slab = vtoslab((vm_offset_t)item);
4388 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4389 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4390 slab = hash_sfind(&keg->uk_hash, mem);
4392 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4394 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4397 lock = KEG_LOCK(keg, slab->us_domain);
4399 slab_free_item(zone, slab, item);
4406 * Frees a single item to any zone.
4409 * zone The zone to free to
4410 * item The item we're freeing
4411 * udata User supplied data for the dtor
4412 * skip Skip dtors and finis
4414 static __noinline void
4415 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4419 * If a free is sent directly to an SMR zone we have to
4420 * synchronize immediately because the item can instantly
4421 * be reallocated. This should only happen in degenerate
4422 * cases when no memory is available for per-cpu caches.
4424 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4425 smr_synchronize(zone->uz_smr);
4427 item_dtor(zone, item, zone->uz_size, udata, skip);
4429 if (skip < SKIP_FINI && zone->uz_fini)
4430 zone->uz_fini(item, zone->uz_size);
4432 zone->uz_release(zone->uz_arg, &item, 1);
4434 if (skip & SKIP_CNT)
4437 counter_u64_add(zone->uz_frees, 1);
4439 if (zone->uz_max_items > 0)
4440 zone_free_limit(zone, 1);
4445 uma_zone_set_max(uma_zone_t zone, int nitems)
4447 struct uma_bucket_zone *ubz;
4451 * XXX This can misbehave if the zone has any allocations with
4452 * no limit and a limit is imposed. There is currently no
4453 * way to clear a limit.
4456 ubz = bucket_zone_max(zone, nitems);
4457 count = ubz != NULL ? ubz->ubz_entries : 0;
4458 zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4459 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4460 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4461 zone->uz_max_items = nitems;
4462 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4463 zone_update_caches(zone);
4464 /* We may need to wake waiters. */
4465 wakeup(&zone->uz_max_items);
4473 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4475 struct uma_bucket_zone *ubz;
4479 ubz = bucket_zone_max(zone, nitems);
4482 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4483 /* Count the cross-domain bucket. */
4485 nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4486 zone->uz_bucket_size_max = ubz->ubz_entries;
4488 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4490 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4491 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4492 zone->uz_bucket_max = nitems / vm_ndomains;
4498 uma_zone_get_max(uma_zone_t zone)
4502 nitems = atomic_load_64(&zone->uz_max_items);
4509 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4512 ZONE_ASSERT_COLD(zone);
4513 zone->uz_warning = warning;
4518 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4521 ZONE_ASSERT_COLD(zone);
4522 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4527 uma_zone_get_cur(uma_zone_t zone)
4533 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4534 nitems = counter_u64_fetch(zone->uz_allocs) -
4535 counter_u64_fetch(zone->uz_frees);
4537 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4538 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4540 return (nitems < 0 ? 0 : nitems);
4544 uma_zone_get_allocs(uma_zone_t zone)
4550 if (zone->uz_allocs != EARLY_COUNTER)
4551 nitems = counter_u64_fetch(zone->uz_allocs);
4553 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4559 uma_zone_get_frees(uma_zone_t zone)
4565 if (zone->uz_frees != EARLY_COUNTER)
4566 nitems = counter_u64_fetch(zone->uz_frees);
4568 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4574 /* Used only for KEG_ASSERT_COLD(). */
4576 uma_keg_get_allocs(uma_keg_t keg)
4582 LIST_FOREACH(z, &keg->uk_zones, uz_link)
4583 nitems += uma_zone_get_allocs(z);
4591 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4596 KEG_ASSERT_COLD(keg);
4597 keg->uk_init = uminit;
4602 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4607 KEG_ASSERT_COLD(keg);
4608 keg->uk_fini = fini;
4613 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4616 ZONE_ASSERT_COLD(zone);
4617 zone->uz_init = zinit;
4622 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4625 ZONE_ASSERT_COLD(zone);
4626 zone->uz_fini = zfini;
4631 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4636 KEG_ASSERT_COLD(keg);
4637 keg->uk_freef = freef;
4642 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4647 KEG_ASSERT_COLD(keg);
4648 keg->uk_allocf = allocf;
4653 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4656 ZONE_ASSERT_COLD(zone);
4658 KASSERT(smr != NULL, ("Got NULL smr"));
4659 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4660 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
4661 zone->uz_flags |= UMA_ZONE_SMR;
4663 zone_update_caches(zone);
4667 uma_zone_get_smr(uma_zone_t zone)
4670 return (zone->uz_smr);
4675 uma_zone_reserve(uma_zone_t zone, int items)
4680 KEG_ASSERT_COLD(keg);
4681 keg->uk_reserve = items;
4686 uma_zone_reserve_kva(uma_zone_t zone, int count)
4693 KEG_ASSERT_COLD(keg);
4694 ZONE_ASSERT_COLD(zone);
4696 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4698 #ifdef UMA_MD_SMALL_ALLOC
4699 if (keg->uk_ppera > 1) {
4703 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4709 MPASS(keg->uk_kva == 0);
4712 zone->uz_max_items = pages * keg->uk_ipers;
4713 #ifdef UMA_MD_SMALL_ALLOC
4714 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4716 keg->uk_allocf = noobj_alloc;
4718 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4719 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4720 zone_update_caches(zone);
4727 uma_prealloc(uma_zone_t zone, int items)
4729 struct vm_domainset_iter di;
4733 int aflags, domain, slabs;
4736 slabs = howmany(items, keg->uk_ipers);
4737 while (slabs-- > 0) {
4739 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4742 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4745 dom = &keg->uk_domain[slab->us_domain];
4747 * keg_alloc_slab() always returns a slab on the
4750 LIST_REMOVE(slab, us_link);
4751 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4753 dom->ud_free_slabs++;
4754 KEG_UNLOCK(keg, slab->us_domain);
4757 if (vm_domainset_iter_policy(&di, &domain) != 0)
4758 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
4764 * Returns a snapshot of memory consumption in bytes.
4767 uma_zone_memory(uma_zone_t zone)
4773 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
4774 for (i = 0; i < vm_ndomains; i++)
4775 sz += ZDOM_GET(zone, i)->uzd_nitems;
4776 return (sz * zone->uz_size);
4778 for (i = 0; i < vm_ndomains; i++)
4779 sz += zone->uz_keg->uk_domain[i].ud_pages;
4781 return (sz * PAGE_SIZE);
4786 uma_reclaim(int req)
4789 CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4790 sx_xlock(&uma_reclaim_lock);
4794 case UMA_RECLAIM_TRIM:
4795 zone_foreach(zone_trim, NULL);
4797 case UMA_RECLAIM_DRAIN:
4798 case UMA_RECLAIM_DRAIN_CPU:
4799 zone_foreach(zone_drain, NULL);
4800 if (req == UMA_RECLAIM_DRAIN_CPU) {
4801 pcpu_cache_drain_safe(NULL);
4802 zone_foreach(zone_drain, NULL);
4806 panic("unhandled reclamation request %d", req);
4810 * Some slabs may have been freed but this zone will be visited early
4811 * we visit again so that we can free pages that are empty once other
4812 * zones are drained. We have to do the same for buckets.
4814 zone_drain(slabzones[0], NULL);
4815 zone_drain(slabzones[1], NULL);
4816 bucket_zone_drain();
4817 sx_xunlock(&uma_reclaim_lock);
4820 static volatile int uma_reclaim_needed;
4823 uma_reclaim_wakeup(void)
4826 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4827 wakeup(uma_reclaim);
4831 uma_reclaim_worker(void *arg __unused)
4835 sx_xlock(&uma_reclaim_lock);
4836 while (atomic_load_int(&uma_reclaim_needed) == 0)
4837 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4839 sx_xunlock(&uma_reclaim_lock);
4840 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4841 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4842 atomic_store_int(&uma_reclaim_needed, 0);
4843 /* Don't fire more than once per-second. */
4844 pause("umarclslp", hz);
4850 uma_zone_reclaim(uma_zone_t zone, int req)
4854 case UMA_RECLAIM_TRIM:
4855 zone_trim(zone, NULL);
4857 case UMA_RECLAIM_DRAIN:
4858 zone_drain(zone, NULL);
4860 case UMA_RECLAIM_DRAIN_CPU:
4861 pcpu_cache_drain_safe(zone);
4862 zone_drain(zone, NULL);
4865 panic("unhandled reclamation request %d", req);
4871 uma_zone_exhausted(uma_zone_t zone)
4874 return (atomic_load_32(&zone->uz_sleepers) > 0);
4881 return (uma_kmem_limit);
4885 uma_set_limit(unsigned long limit)
4888 uma_kmem_limit = limit;
4895 return (atomic_load_long(&uma_kmem_total));
4902 return (uma_kmem_limit - uma_size());
4907 * Generate statistics across both the zone and its per-cpu cache's. Return
4908 * desired statistics if the pointer is non-NULL for that statistic.
4910 * Note: does not update the zone statistics, as it can't safely clear the
4911 * per-CPU cache statistic.
4915 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4916 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4919 uint64_t allocs, frees, sleeps, xdomain;
4922 allocs = frees = sleeps = xdomain = 0;
4925 cache = &z->uz_cpu[cpu];
4926 cachefree += cache->uc_allocbucket.ucb_cnt;
4927 cachefree += cache->uc_freebucket.ucb_cnt;
4928 xdomain += cache->uc_crossbucket.ucb_cnt;
4929 cachefree += cache->uc_crossbucket.ucb_cnt;
4930 allocs += cache->uc_allocs;
4931 frees += cache->uc_frees;
4933 allocs += counter_u64_fetch(z->uz_allocs);
4934 frees += counter_u64_fetch(z->uz_frees);
4935 xdomain += counter_u64_fetch(z->uz_xdomain);
4936 sleeps += z->uz_sleeps;
4937 if (cachefreep != NULL)
4938 *cachefreep = cachefree;
4939 if (allocsp != NULL)
4943 if (sleepsp != NULL)
4945 if (xdomainp != NULL)
4946 *xdomainp = xdomain;
4951 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
4958 rw_rlock(&uma_rwlock);
4959 LIST_FOREACH(kz, &uma_kegs, uk_link) {
4960 LIST_FOREACH(z, &kz->uk_zones, uz_link)
4963 LIST_FOREACH(z, &uma_cachezones, uz_link)
4966 rw_runlock(&uma_rwlock);
4967 return (sysctl_handle_int(oidp, &count, 0, req));
4971 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
4972 struct uma_percpu_stat *ups, bool internal)
4974 uma_zone_domain_t zdom;
4978 for (i = 0; i < vm_ndomains; i++) {
4979 zdom = ZDOM_GET(z, i);
4980 uth->uth_zone_free += zdom->uzd_nitems;
4982 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
4983 uth->uth_frees = counter_u64_fetch(z->uz_frees);
4984 uth->uth_fails = counter_u64_fetch(z->uz_fails);
4985 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
4986 uth->uth_sleeps = z->uz_sleeps;
4988 for (i = 0; i < mp_maxid + 1; i++) {
4989 bzero(&ups[i], sizeof(*ups));
4990 if (internal || CPU_ABSENT(i))
4992 cache = &z->uz_cpu[i];
4993 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
4994 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
4995 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
4996 ups[i].ups_allocs = cache->uc_allocs;
4997 ups[i].ups_frees = cache->uc_frees;
5002 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5004 struct uma_stream_header ush;
5005 struct uma_type_header uth;
5006 struct uma_percpu_stat *ups;
5011 uint32_t kfree, pages;
5012 int count, error, i;
5014 error = sysctl_wire_old_buffer(req, 0);
5017 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5018 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5019 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5022 rw_rlock(&uma_rwlock);
5023 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5024 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5028 LIST_FOREACH(z, &uma_cachezones, uz_link)
5032 * Insert stream header.
5034 bzero(&ush, sizeof(ush));
5035 ush.ush_version = UMA_STREAM_VERSION;
5036 ush.ush_maxcpus = (mp_maxid + 1);
5037 ush.ush_count = count;
5038 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5040 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5042 for (i = 0; i < vm_ndomains; i++) {
5043 kfree += kz->uk_domain[i].ud_free_items;
5044 pages += kz->uk_domain[i].ud_pages;
5046 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5047 bzero(&uth, sizeof(uth));
5048 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5049 uth.uth_align = kz->uk_align;
5050 uth.uth_size = kz->uk_size;
5051 uth.uth_rsize = kz->uk_rsize;
5052 if (z->uz_max_items > 0) {
5053 items = UZ_ITEMS_COUNT(z->uz_items);
5054 uth.uth_pages = (items / kz->uk_ipers) *
5057 uth.uth_pages = pages;
5058 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5060 uth.uth_limit = z->uz_max_items;
5061 uth.uth_keg_free = kfree;
5064 * A zone is secondary is it is not the first entry
5065 * on the keg's zone list.
5067 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5068 (LIST_FIRST(&kz->uk_zones) != z))
5069 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5070 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5071 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5072 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5073 for (i = 0; i < mp_maxid + 1; i++)
5074 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5077 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5078 bzero(&uth, sizeof(uth));
5079 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5080 uth.uth_size = z->uz_size;
5081 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5082 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5083 for (i = 0; i < mp_maxid + 1; i++)
5084 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5087 rw_runlock(&uma_rwlock);
5088 error = sbuf_finish(&sbuf);
5095 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5097 uma_zone_t zone = *(uma_zone_t *)arg1;
5100 max = uma_zone_get_max(zone);
5101 error = sysctl_handle_int(oidp, &max, 0, req);
5102 if (error || !req->newptr)
5105 uma_zone_set_max(zone, max);
5111 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5117 * Some callers want to add sysctls for global zones that
5118 * may not yet exist so they pass a pointer to a pointer.
5121 zone = *(uma_zone_t *)arg1;
5124 cur = uma_zone_get_cur(zone);
5125 return (sysctl_handle_int(oidp, &cur, 0, req));
5129 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5131 uma_zone_t zone = arg1;
5134 cur = uma_zone_get_allocs(zone);
5135 return (sysctl_handle_64(oidp, &cur, 0, req));
5139 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5141 uma_zone_t zone = arg1;
5144 cur = uma_zone_get_frees(zone);
5145 return (sysctl_handle_64(oidp, &cur, 0, req));
5149 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5152 uma_zone_t zone = arg1;
5155 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5156 if (zone->uz_flags != 0)
5157 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5159 sbuf_printf(&sbuf, "0");
5160 error = sbuf_finish(&sbuf);
5167 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5169 uma_keg_t keg = arg1;
5170 int avail, effpct, total;
5172 total = keg->uk_ppera * PAGE_SIZE;
5173 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5174 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5176 * We consider the client's requested size and alignment here, not the
5177 * real size determination uk_rsize, because we also adjust the real
5178 * size for internal implementation reasons (max bitset size).
5180 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5181 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5182 avail *= mp_maxid + 1;
5183 effpct = 100 * avail / total;
5184 return (sysctl_handle_int(oidp, &effpct, 0, req));
5188 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5190 uma_zone_t zone = arg1;
5193 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5194 return (sysctl_handle_64(oidp, &cur, 0, req));
5199 uma_dbg_getslab(uma_zone_t zone, void *item)
5206 * It is safe to return the slab here even though the
5207 * zone is unlocked because the item's allocation state
5208 * essentially holds a reference.
5210 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5211 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5213 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5214 return (vtoslab((vm_offset_t)mem));
5216 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5217 return ((uma_slab_t)(mem + keg->uk_pgoff));
5219 slab = hash_sfind(&keg->uk_hash, mem);
5226 uma_dbg_zskip(uma_zone_t zone, void *mem)
5229 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5232 return (uma_dbg_kskip(zone->uz_keg, mem));
5236 uma_dbg_kskip(uma_keg_t keg, void *mem)
5240 if (dbg_divisor == 0)
5243 if (dbg_divisor == 1)
5246 idx = (uintptr_t)mem >> PAGE_SHIFT;
5247 if (keg->uk_ipers > 1) {
5248 idx *= keg->uk_ipers;
5249 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5252 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5253 counter_u64_add(uma_skip_cnt, 1);
5256 counter_u64_add(uma_dbg_cnt, 1);
5262 * Set up the slab's freei data such that uma_dbg_free can function.
5266 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5272 slab = uma_dbg_getslab(zone, item);
5274 panic("uma: item %p did not belong to zone %s\n",
5275 item, zone->uz_name);
5278 freei = slab_item_index(slab, keg, item);
5280 if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5281 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
5282 item, zone, zone->uz_name, slab, freei);
5283 BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5287 * Verifies freed addresses. Checks for alignment, valid slab membership
5288 * and duplicate frees.
5292 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5298 slab = uma_dbg_getslab(zone, item);
5300 panic("uma: Freed item %p did not belong to zone %s\n",
5301 item, zone->uz_name);
5304 freei = slab_item_index(slab, keg, item);
5306 if (freei >= keg->uk_ipers)
5307 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
5308 item, zone, zone->uz_name, slab, freei);
5310 if (slab_item(slab, keg, freei) != item)
5311 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
5312 item, zone, zone->uz_name, slab, freei);
5314 if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5315 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
5316 item, zone, zone->uz_name, slab, freei);
5318 BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5320 #endif /* INVARIANTS */
5324 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5325 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5330 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5331 *allocs = counter_u64_fetch(z->uz_allocs);
5332 frees = counter_u64_fetch(z->uz_frees);
5333 *sleeps = z->uz_sleeps;
5337 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5339 for (i = 0; i < vm_ndomains; i++) {
5340 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5341 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5342 (LIST_FIRST(&kz->uk_zones) != z)))
5343 *cachefree += kz->uk_domain[i].ud_free_items;
5345 *used = *allocs - frees;
5346 return (((int64_t)*used + *cachefree) * kz->uk_size);
5349 DB_SHOW_COMMAND(uma, db_show_uma)
5351 const char *fmt_hdr, *fmt_entry;
5354 uint64_t allocs, used, sleeps, xdomain;
5356 /* variables for sorting */
5358 uma_zone_t cur_zone, last_zone;
5359 int64_t cur_size, last_size, size;
5362 /* /i option produces machine-parseable CSV output */
5363 if (modif[0] == 'i') {
5364 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5365 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5367 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5368 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5371 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5372 "Sleeps", "Bucket", "Total Mem", "XFree");
5374 /* Sort the zones with largest size first. */
5376 last_size = INT64_MAX;
5381 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5382 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5384 * In the case of size ties, print out zones
5385 * in the order they are encountered. That is,
5386 * when we encounter the most recently output
5387 * zone, we have already printed all preceding
5388 * ties, and we must print all following ties.
5390 if (z == last_zone) {
5394 size = get_uma_stats(kz, z, &allocs, &used,
5395 &sleeps, &cachefree, &xdomain);
5396 if (size > cur_size && size < last_size + ties)
5404 if (cur_zone == NULL)
5407 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5408 &sleeps, &cachefree, &xdomain);
5409 db_printf(fmt_entry, cur_zone->uz_name,
5410 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5411 (uintmax_t)allocs, (uintmax_t)sleeps,
5412 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5417 last_zone = cur_zone;
5418 last_size = cur_size;
5422 DB_SHOW_COMMAND(umacache, db_show_umacache)
5425 uint64_t allocs, frees;
5429 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5430 "Requests", "Bucket");
5431 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5432 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5433 for (i = 0; i < vm_ndomains; i++)
5434 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5435 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5436 z->uz_name, (uintmax_t)z->uz_size,
5437 (intmax_t)(allocs - frees), cachefree,
5438 (uintmax_t)allocs, z->uz_bucket_size);