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>
62 #include <sys/bitset.h>
63 #include <sys/domainset.h>
64 #include <sys/eventhandler.h>
65 #include <sys/kernel.h>
66 #include <sys/types.h>
67 #include <sys/limits.h>
68 #include <sys/queue.h>
69 #include <sys/malloc.h>
73 #include <sys/mutex.h>
75 #include <sys/random.h>
76 #include <sys/rwlock.h>
78 #include <sys/sched.h>
79 #include <sys/sleepqueue.h>
82 #include <sys/sysctl.h>
83 #include <sys/taskqueue.h>
84 #include <sys/vmmeter.h>
87 #include <vm/vm_param.h>
88 #include <vm/vm_domainset.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_phys.h>
93 #include <vm/vm_pagequeue.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_extern.h>
97 #include <vm/vm_dumpset.h>
99 #include <vm/uma_int.h>
100 #include <vm/uma_dbg.h>
104 #ifdef DEBUG_MEMGUARD
105 #include <vm/memguard.h>
108 #include <machine/md_var.h>
111 #define UMA_ALWAYS_CTORDTOR 1
113 #define UMA_ALWAYS_CTORDTOR 0
117 * This is the zone and keg from which all zones are spawned.
119 static uma_zone_t kegs;
120 static uma_zone_t zones;
123 * On INVARIANTS builds, the slab contains a second bitset of the same size,
124 * "dbg_bits", which is laid out immediately after us_free.
127 #define SLAB_BITSETS 2
129 #define SLAB_BITSETS 1
133 * These are the two zones from which all offpage uma_slab_ts are allocated.
135 * One zone is for slab headers that can represent a larger number of items,
136 * making the slabs themselves more efficient, and the other zone is for
137 * headers that are smaller and represent fewer items, making the headers more
140 #define SLABZONE_SIZE(setsize) \
141 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
142 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16)
143 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE
144 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE)
145 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE)
146 static uma_zone_t slabzones[2];
149 * The initial hash tables come out of this zone so they can be allocated
150 * prior to malloc coming up.
152 static uma_zone_t hashzone;
154 /* The boot-time adjusted value for cache line alignment. */
155 int uma_align_cache = 64 - 1;
157 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
158 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
161 * Are we allowed to allocate buckets?
163 static int bucketdisable = 1;
165 /* Linked list of all kegs in the system */
166 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
168 /* Linked list of all cache-only zones in the system */
169 static LIST_HEAD(,uma_zone) uma_cachezones =
170 LIST_HEAD_INITIALIZER(uma_cachezones);
173 * Mutex for global lists: uma_kegs, uma_cachezones, and the per-keg list of
176 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
178 static struct sx uma_reclaim_lock;
181 * First available virual address for boot time allocations.
183 static vm_offset_t bootstart;
184 static vm_offset_t bootmem;
187 * kmem soft limit, initialized by uma_set_limit(). Ensure that early
188 * allocations don't trigger a wakeup of the reclaim thread.
190 unsigned long uma_kmem_limit = LONG_MAX;
191 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
192 "UMA kernel memory soft limit");
193 unsigned long uma_kmem_total;
194 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
195 "UMA kernel memory usage");
197 /* Is the VM done starting up? */
204 } booted = BOOT_COLD;
207 * This is the handle used to schedule events that need to happen
208 * outside of the allocation fast path.
210 static struct callout uma_callout;
211 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
214 * This structure is passed as the zone ctor arg so that I don't have to create
215 * a special allocation function just for zones.
217 struct uma_zctor_args {
232 struct uma_kctor_args {
241 struct uma_bucket_zone {
243 const char *ubz_name;
244 int ubz_entries; /* Number of items it can hold. */
245 int ubz_maxsize; /* Maximum allocation size per-item. */
249 * Compute the actual number of bucket entries to pack them in power
250 * of two sizes for more efficient space utilization.
252 #define BUCKET_SIZE(n) \
253 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
255 #define BUCKET_MAX BUCKET_SIZE(256)
257 struct uma_bucket_zone bucket_zones[] = {
258 /* Literal bucket sizes. */
259 { NULL, "2 Bucket", 2, 4096 },
260 { NULL, "4 Bucket", 4, 3072 },
261 { NULL, "8 Bucket", 8, 2048 },
262 { NULL, "16 Bucket", 16, 1024 },
263 /* Rounded down power of 2 sizes for efficiency. */
264 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
265 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
266 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
267 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
272 * Flags and enumerations to be passed to internal functions.
276 SKIP_CNT = 0x00000001,
277 SKIP_DTOR = 0x00010000,
278 SKIP_FINI = 0x00020000,
283 void uma_startup1(vm_offset_t);
284 void uma_startup2(void);
286 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
287 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
288 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
289 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
290 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
291 static void page_free(void *, vm_size_t, uint8_t);
292 static void pcpu_page_free(void *, vm_size_t, uint8_t);
293 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
294 static void cache_drain(uma_zone_t);
295 static void bucket_drain(uma_zone_t, uma_bucket_t);
296 static void bucket_cache_reclaim(uma_zone_t zone, bool, int);
297 static bool bucket_cache_reclaim_domain(uma_zone_t, bool, bool, int);
298 static int keg_ctor(void *, int, void *, int);
299 static void keg_dtor(void *, int, void *);
300 static void keg_drain(uma_keg_t keg, int domain);
301 static int zone_ctor(void *, int, void *, int);
302 static void zone_dtor(void *, int, void *);
303 static inline void item_dtor(uma_zone_t zone, void *item, int size,
304 void *udata, enum zfreeskip skip);
305 static int zero_init(void *, int, int);
306 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
307 int itemdomain, bool ws);
308 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
309 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
310 static void zone_timeout(uma_zone_t zone, void *);
311 static int hash_alloc(struct uma_hash *, u_int);
312 static int hash_expand(struct uma_hash *, struct uma_hash *);
313 static void hash_free(struct uma_hash *hash);
314 static void uma_timeout(void *);
315 static void uma_shutdown(void);
316 static void *zone_alloc_item(uma_zone_t, void *, int, int);
317 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
318 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
319 static void zone_free_limit(uma_zone_t zone, int count);
320 static void bucket_enable(void);
321 static void bucket_init(void);
322 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
323 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
324 static void bucket_zone_drain(int domain);
325 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
326 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
327 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
328 static size_t slab_sizeof(int nitems);
329 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
330 uma_fini fini, int align, uint32_t flags);
331 static int zone_import(void *, void **, int, int, int);
332 static void zone_release(void *, void **, int);
333 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
334 static bool cache_free(uma_zone_t, uma_cache_t, void *, int);
336 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
337 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
338 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
339 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
340 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
341 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
342 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
344 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
346 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
347 "Memory allocation debugging");
350 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
351 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
353 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
354 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
355 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
356 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
358 static u_int dbg_divisor = 1;
359 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
360 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
361 "Debug & thrash every this item in memory allocator");
363 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
364 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
365 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
366 &uma_dbg_cnt, "memory items debugged");
367 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
368 &uma_skip_cnt, "memory items skipped, not debugged");
371 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
372 "Universal Memory Allocator");
374 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
375 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
377 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
378 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
380 static int zone_warnings = 1;
381 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
382 "Warn when UMA zones becomes full");
384 static int multipage_slabs = 1;
385 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
386 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
387 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
388 "UMA may choose larger slab sizes for better efficiency");
391 * Select the slab zone for an offpage slab with the given maximum item count.
393 static inline uma_zone_t
397 return (slabzones[ipers > SLABZONE0_SETSIZE]);
401 * This routine checks to see whether or not it's safe to enable buckets.
407 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
408 bucketdisable = vm_page_count_min();
412 * Initialize bucket_zones, the array of zones of buckets of various sizes.
414 * For each zone, calculate the memory required for each bucket, consisting
415 * of the header and an array of pointers.
420 struct uma_bucket_zone *ubz;
423 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
424 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
425 size += sizeof(void *) * ubz->ubz_entries;
426 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
427 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
428 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
429 UMA_ZONE_FIRSTTOUCH);
434 * Given a desired number of entries for a bucket, return the zone from which
435 * to allocate the bucket.
437 static struct uma_bucket_zone *
438 bucket_zone_lookup(int entries)
440 struct uma_bucket_zone *ubz;
442 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
443 if (ubz->ubz_entries >= entries)
450 bucket_select(int size)
452 struct uma_bucket_zone *ubz;
454 ubz = &bucket_zones[0];
455 if (size > ubz->ubz_maxsize)
456 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
458 for (; ubz->ubz_entries != 0; ubz++)
459 if (ubz->ubz_maxsize < size)
462 return (ubz->ubz_entries);
466 bucket_alloc(uma_zone_t zone, void *udata, int flags)
468 struct uma_bucket_zone *ubz;
472 * Don't allocate buckets early in boot.
474 if (__predict_false(booted < BOOT_KVA))
478 * To limit bucket recursion we store the original zone flags
479 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
480 * NOVM flag to persist even through deep recursions. We also
481 * store ZFLAG_BUCKET once we have recursed attempting to allocate
482 * a bucket for a bucket zone so we do not allow infinite bucket
483 * recursion. This cookie will even persist to frees of unused
484 * buckets via the allocation path or bucket allocations in the
487 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
488 udata = (void *)(uintptr_t)zone->uz_flags;
490 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
492 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
494 if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
496 ubz = bucket_zone_lookup(atomic_load_16(&zone->uz_bucket_size));
497 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
499 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
502 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
505 bucket->ub_entries = min(ubz->ubz_entries,
506 zone->uz_bucket_size_max);
507 bucket->ub_seq = SMR_SEQ_INVALID;
508 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
509 zone->uz_name, zone, bucket);
516 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
518 struct uma_bucket_zone *ubz;
520 if (bucket->ub_cnt != 0)
521 bucket_drain(zone, bucket);
523 KASSERT(bucket->ub_cnt == 0,
524 ("bucket_free: Freeing a non free bucket."));
525 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
526 ("bucket_free: Freeing an SMR bucket."));
527 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
528 udata = (void *)(uintptr_t)zone->uz_flags;
529 ubz = bucket_zone_lookup(bucket->ub_entries);
530 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
534 bucket_zone_drain(int domain)
536 struct uma_bucket_zone *ubz;
538 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
539 uma_zone_reclaim_domain(ubz->ubz_zone, UMA_RECLAIM_DRAIN,
544 _Static_assert(UMA_SMALLEST_UNIT % KASAN_SHADOW_SCALE == 0,
545 "Base UMA allocation size not a multiple of the KASAN scale factor");
548 kasan_mark_item_valid(uma_zone_t zone, void *item)
554 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
558 rsz = roundup2(sz, KASAN_SHADOW_SCALE);
559 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
560 kasan_mark(item, sz, rsz, KASAN_GENERIC_REDZONE);
562 pcpu_item = zpcpu_base_to_offset(item);
563 for (i = 0; i <= mp_maxid; i++)
564 kasan_mark(zpcpu_get_cpu(pcpu_item, i), sz, rsz,
565 KASAN_GENERIC_REDZONE);
570 kasan_mark_item_invalid(uma_zone_t zone, void *item)
576 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
579 sz = roundup2(zone->uz_size, KASAN_SHADOW_SCALE);
580 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
581 kasan_mark(item, 0, sz, KASAN_UMA_FREED);
583 pcpu_item = zpcpu_base_to_offset(item);
584 for (i = 0; i <= mp_maxid; i++)
585 kasan_mark(zpcpu_get_cpu(pcpu_item, i), 0, sz,
591 kasan_mark_slab_valid(uma_keg_t keg, void *mem)
595 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
596 sz = keg->uk_ppera * PAGE_SIZE;
597 kasan_mark(mem, sz, sz, 0);
602 kasan_mark_slab_invalid(uma_keg_t keg, void *mem)
606 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
607 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
608 sz = keg->uk_ppera * PAGE_SIZE;
611 kasan_mark(mem, 0, sz, KASAN_UMA_FREED);
616 kasan_mark_item_valid(uma_zone_t zone __unused, void *item __unused)
621 kasan_mark_item_invalid(uma_zone_t zone __unused, void *item __unused)
626 kasan_mark_slab_valid(uma_keg_t keg __unused, void *mem __unused)
631 kasan_mark_slab_invalid(uma_keg_t keg __unused, void *mem __unused)
638 kmsan_mark_item_uninitialized(uma_zone_t zone, void *item)
644 if ((zone->uz_flags &
645 (UMA_ZFLAG_CACHE | UMA_ZONE_SECONDARY | UMA_ZONE_MALLOC)) != 0) {
647 * Cache zones should not be instrumented by default, as UMA
648 * does not have enough information to do so correctly.
649 * Consumers can mark items themselves if it makes sense to do
652 * Items from secondary zones are initialized by the parent
653 * zone and thus cannot safely be marked by UMA.
655 * malloc zones are handled directly by malloc(9) and friends,
656 * since they can provide more precise origin tracking.
660 if (zone->uz_keg->uk_init != NULL) {
662 * By definition, initialized items cannot be marked. The
663 * best we can do is mark items from these zones after they
664 * are freed to the keg.
670 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
671 kmsan_orig(item, sz, KMSAN_TYPE_UMA, KMSAN_RET_ADDR);
672 kmsan_mark(item, sz, KMSAN_STATE_UNINIT);
674 pcpu_item = zpcpu_base_to_offset(item);
675 for (i = 0; i <= mp_maxid; i++) {
676 kmsan_orig(zpcpu_get_cpu(pcpu_item, i), sz,
677 KMSAN_TYPE_UMA, KMSAN_RET_ADDR);
678 kmsan_mark(zpcpu_get_cpu(pcpu_item, i), sz,
685 kmsan_mark_item_uninitialized(uma_zone_t zone __unused, void *item __unused)
691 * Acquire the domain lock and record contention.
693 static uma_zone_domain_t
694 zone_domain_lock(uma_zone_t zone, int domain)
696 uma_zone_domain_t zdom;
699 zdom = ZDOM_GET(zone, domain);
701 if (ZDOM_OWNED(zdom))
704 /* This is unsynchronized. The counter does not need to be precise. */
705 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
706 zone->uz_bucket_size++;
711 * Search for the domain with the least cached items and return it if it
712 * is out of balance with the preferred domain.
714 static __noinline int
715 zone_domain_lowest(uma_zone_t zone, int pref)
717 long least, nitems, prefitems;
721 prefitems = least = LONG_MAX;
723 for (i = 0; i < vm_ndomains; i++) {
724 nitems = ZDOM_GET(zone, i)->uzd_nitems;
725 if (nitems < least) {
732 if (prefitems < least * 2)
739 * Search for the domain with the most cached items and return it or the
740 * preferred domain if it has enough to proceed.
742 static __noinline int
743 zone_domain_highest(uma_zone_t zone, int pref)
749 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
754 for (i = 0; i < vm_ndomains; i++) {
755 nitems = ZDOM_GET(zone, i)->uzd_nitems;
766 * Set the maximum imax value.
769 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
773 old = zdom->uzd_imax;
777 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
780 * We are at new maximum, so do the last WSS update for the old
781 * bimin and prepare to measure next allocation batch.
783 if (zdom->uzd_wss < old - zdom->uzd_bimin)
784 zdom->uzd_wss = old - zdom->uzd_bimin;
785 zdom->uzd_bimin = nitems;
789 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
790 * zone's caches. If a bucket is found the zone is not locked on return.
793 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
800 ZDOM_LOCK_ASSERT(zdom);
802 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
805 /* SMR Buckets can not be re-used until readers expire. */
806 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
807 bucket->ub_seq != SMR_SEQ_INVALID) {
808 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
810 bucket->ub_seq = SMR_SEQ_INVALID;
811 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
812 if (STAILQ_NEXT(bucket, ub_link) != NULL)
813 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
815 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
817 KASSERT(zdom->uzd_nitems >= bucket->ub_cnt,
818 ("%s: item count underflow (%ld, %d)",
819 __func__, zdom->uzd_nitems, bucket->ub_cnt));
820 KASSERT(bucket->ub_cnt > 0,
821 ("%s: empty bucket in bucket cache", __func__));
822 zdom->uzd_nitems -= bucket->ub_cnt;
826 * Shift the bounds of the current WSS interval to avoid
827 * perturbing the estimates.
829 cnt = lmin(zdom->uzd_bimin, bucket->ub_cnt);
830 atomic_subtract_long(&zdom->uzd_imax, cnt);
831 zdom->uzd_bimin -= cnt;
832 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
833 if (zdom->uzd_limin >= bucket->ub_cnt) {
834 zdom->uzd_limin -= bucket->ub_cnt;
839 } else if (zdom->uzd_bimin > zdom->uzd_nitems) {
840 zdom->uzd_bimin = zdom->uzd_nitems;
841 if (zdom->uzd_imin > zdom->uzd_nitems)
842 zdom->uzd_imin = zdom->uzd_nitems;
847 for (i = 0; i < bucket->ub_cnt; i++)
848 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
855 * Insert a full bucket into the specified cache. The "ws" parameter indicates
856 * whether the bucket's contents should be counted as part of the zone's working
857 * set. The bucket may be freed if it exceeds the bucket limit.
860 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
863 uma_zone_domain_t zdom;
865 /* We don't cache empty buckets. This can happen after a reclaim. */
866 if (bucket->ub_cnt == 0)
868 zdom = zone_domain_lock(zone, domain);
871 * Conditionally set the maximum number of items.
873 zdom->uzd_nitems += bucket->ub_cnt;
874 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
876 zone_domain_imax_set(zdom, zdom->uzd_nitems);
879 * Shift the bounds of the current WSS interval to
880 * avoid perturbing the estimates.
882 atomic_add_long(&zdom->uzd_imax, bucket->ub_cnt);
883 zdom->uzd_imin += bucket->ub_cnt;
884 zdom->uzd_bimin += bucket->ub_cnt;
885 zdom->uzd_limin += bucket->ub_cnt;
887 if (STAILQ_EMPTY(&zdom->uzd_buckets))
888 zdom->uzd_seq = bucket->ub_seq;
891 * Try to promote reuse of recently used items. For items
892 * protected by SMR, try to defer reuse to minimize polling.
894 if (bucket->ub_seq == SMR_SEQ_INVALID)
895 STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
897 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
901 zdom->uzd_nitems -= bucket->ub_cnt;
904 bucket_free(zone, bucket, udata);
907 /* Pops an item out of a per-cpu cache bucket. */
909 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
913 CRITICAL_ASSERT(curthread);
916 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
918 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
919 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
926 /* Pushes an item into a per-cpu cache bucket. */
928 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
931 CRITICAL_ASSERT(curthread);
932 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
933 ("uma_zfree: Freeing to non free bucket index."));
935 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
941 * Unload a UMA bucket from a per-cpu cache.
943 static inline uma_bucket_t
944 cache_bucket_unload(uma_cache_bucket_t bucket)
948 b = bucket->ucb_bucket;
950 MPASS(b->ub_entries == bucket->ucb_entries);
951 b->ub_cnt = bucket->ucb_cnt;
952 bucket->ucb_bucket = NULL;
953 bucket->ucb_entries = bucket->ucb_cnt = 0;
959 static inline uma_bucket_t
960 cache_bucket_unload_alloc(uma_cache_t cache)
963 return (cache_bucket_unload(&cache->uc_allocbucket));
966 static inline uma_bucket_t
967 cache_bucket_unload_free(uma_cache_t cache)
970 return (cache_bucket_unload(&cache->uc_freebucket));
973 static inline uma_bucket_t
974 cache_bucket_unload_cross(uma_cache_t cache)
977 return (cache_bucket_unload(&cache->uc_crossbucket));
981 * Load a bucket into a per-cpu cache bucket.
984 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
987 CRITICAL_ASSERT(curthread);
988 MPASS(bucket->ucb_bucket == NULL);
989 MPASS(b->ub_seq == SMR_SEQ_INVALID);
991 bucket->ucb_bucket = b;
992 bucket->ucb_cnt = b->ub_cnt;
993 bucket->ucb_entries = b->ub_entries;
997 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
1000 cache_bucket_load(&cache->uc_allocbucket, b);
1004 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
1007 cache_bucket_load(&cache->uc_freebucket, b);
1012 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
1015 cache_bucket_load(&cache->uc_crossbucket, b);
1020 * Copy and preserve ucb_spare.
1023 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
1026 b1->ucb_bucket = b2->ucb_bucket;
1027 b1->ucb_entries = b2->ucb_entries;
1028 b1->ucb_cnt = b2->ucb_cnt;
1032 * Swap two cache buckets.
1035 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
1037 struct uma_cache_bucket b3;
1039 CRITICAL_ASSERT(curthread);
1041 cache_bucket_copy(&b3, b1);
1042 cache_bucket_copy(b1, b2);
1043 cache_bucket_copy(b2, &b3);
1047 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
1050 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
1052 uma_zone_domain_t zdom;
1053 uma_bucket_t bucket;
1057 * Avoid the lock if possible.
1059 zdom = ZDOM_GET(zone, domain);
1060 if (zdom->uzd_nitems == 0)
1063 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
1064 (seq = atomic_load_32(&zdom->uzd_seq)) != SMR_SEQ_INVALID &&
1065 !smr_poll(zone->uz_smr, seq, false))
1069 * Check the zone's cache of buckets.
1071 zdom = zone_domain_lock(zone, domain);
1072 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL)
1080 zone_log_warning(uma_zone_t zone)
1082 static const struct timeval warninterval = { 300, 0 };
1084 if (!zone_warnings || zone->uz_warning == NULL)
1087 if (ratecheck(&zone->uz_ratecheck, &warninterval))
1088 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
1092 zone_maxaction(uma_zone_t zone)
1095 if (zone->uz_maxaction.ta_func != NULL)
1096 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
1100 * Routine called by timeout which is used to fire off some time interval
1101 * based calculations. (stats, hash size, etc.)
1110 uma_timeout(void *unused)
1113 zone_foreach(zone_timeout, NULL);
1115 /* Reschedule this event */
1116 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1120 * Update the working set size estimates for the zone's bucket cache.
1121 * The constants chosen here are somewhat arbitrary.
1124 zone_domain_update_wss(uma_zone_domain_t zdom)
1128 ZDOM_LOCK_ASSERT(zdom);
1129 MPASS(zdom->uzd_imax >= zdom->uzd_nitems);
1130 MPASS(zdom->uzd_nitems >= zdom->uzd_bimin);
1131 MPASS(zdom->uzd_bimin >= zdom->uzd_imin);
1134 * Estimate WSS as modified moving average of biggest allocation
1135 * batches for each period over few minutes (UMA_TIMEOUT of 20s).
1137 zdom->uzd_wss = lmax(zdom->uzd_wss * 3 / 4,
1138 zdom->uzd_imax - zdom->uzd_bimin);
1141 * Estimate longtime minimum item count as a combination of recent
1142 * minimum item count, adjusted by WSS for safety, and the modified
1143 * moving average over the last several hours (UMA_TIMEOUT of 20s).
1144 * timin measures time since limin tried to go negative, that means
1145 * we were dangerously close to or got out of cache.
1147 m = zdom->uzd_imin - zdom->uzd_wss;
1149 if (zdom->uzd_limin >= m)
1150 zdom->uzd_limin = m;
1152 zdom->uzd_limin = (m + zdom->uzd_limin * 255) / 256;
1155 zdom->uzd_limin = 0;
1156 zdom->uzd_timin = 0;
1159 /* To reduce period edge effects on WSS keep half of the imax. */
1160 atomic_subtract_long(&zdom->uzd_imax,
1161 (zdom->uzd_imax - zdom->uzd_nitems + 1) / 2);
1162 zdom->uzd_imin = zdom->uzd_bimin = zdom->uzd_nitems;
1166 * Routine to perform timeout driven calculations. This expands the
1167 * hashes and does per cpu statistics aggregation.
1172 zone_timeout(uma_zone_t zone, void *unused)
1177 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
1183 * Hash zones are non-numa by definition so the first domain
1184 * is the only one present.
1187 pages = keg->uk_domain[0].ud_pages;
1190 * Expand the keg hash table.
1192 * This is done if the number of slabs is larger than the hash size.
1193 * What I'm trying to do here is completely reduce collisions. This
1194 * may be a little aggressive. Should I allow for two collisions max?
1196 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1197 struct uma_hash newhash;
1198 struct uma_hash oldhash;
1202 * This is so involved because allocating and freeing
1203 * while the keg lock is held will lead to deadlock.
1204 * I have to do everything in stages and check for
1208 ret = hash_alloc(&newhash, 1 << fls(slabs));
1211 if (hash_expand(&keg->uk_hash, &newhash)) {
1212 oldhash = keg->uk_hash;
1213 keg->uk_hash = newhash;
1218 hash_free(&oldhash);
1225 /* Trim caches not used for a long time. */
1226 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0) {
1227 for (int i = 0; i < vm_ndomains; i++) {
1228 if (bucket_cache_reclaim_domain(zone, false, false, i) &&
1229 (zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1230 keg_drain(zone->uz_keg, i);
1236 * Allocate and zero fill the next sized hash table from the appropriate
1240 * hash A new hash structure with the old hash size in uh_hashsize
1243 * 1 on success and 0 on failure.
1246 hash_alloc(struct uma_hash *hash, u_int size)
1250 KASSERT(powerof2(size), ("hash size must be power of 2"));
1251 if (size > UMA_HASH_SIZE_INIT) {
1252 hash->uh_hashsize = size;
1253 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1254 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1256 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1257 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1258 UMA_ANYDOMAIN, M_WAITOK);
1259 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1261 if (hash->uh_slab_hash) {
1262 bzero(hash->uh_slab_hash, alloc);
1263 hash->uh_hashmask = hash->uh_hashsize - 1;
1271 * Expands the hash table for HASH zones. This is done from zone_timeout
1272 * to reduce collisions. This must not be done in the regular allocation
1273 * path, otherwise, we can recurse on the vm while allocating pages.
1276 * oldhash The hash you want to expand
1277 * newhash The hash structure for the new table
1285 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1287 uma_hash_slab_t slab;
1291 if (!newhash->uh_slab_hash)
1294 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1298 * I need to investigate hash algorithms for resizing without a
1302 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1303 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1304 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1305 LIST_REMOVE(slab, uhs_hlink);
1306 hval = UMA_HASH(newhash, slab->uhs_data);
1307 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1315 * Free the hash bucket to the appropriate backing store.
1318 * slab_hash The hash bucket we're freeing
1319 * hashsize The number of entries in that hash bucket
1325 hash_free(struct uma_hash *hash)
1327 if (hash->uh_slab_hash == NULL)
1329 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1330 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1332 free(hash->uh_slab_hash, M_UMAHASH);
1336 * Frees all outstanding items in a bucket
1339 * zone The zone to free to, must be unlocked.
1340 * bucket The free/alloc bucket with items.
1346 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1350 if (bucket->ub_cnt == 0)
1353 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1354 bucket->ub_seq != SMR_SEQ_INVALID) {
1355 smr_wait(zone->uz_smr, bucket->ub_seq);
1356 bucket->ub_seq = SMR_SEQ_INVALID;
1357 for (i = 0; i < bucket->ub_cnt; i++)
1358 item_dtor(zone, bucket->ub_bucket[i],
1359 zone->uz_size, NULL, SKIP_NONE);
1362 for (i = 0; i < bucket->ub_cnt; i++) {
1363 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
1364 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1365 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
1367 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1368 if (zone->uz_max_items > 0)
1369 zone_free_limit(zone, bucket->ub_cnt);
1371 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1377 * Drains the per cpu caches for a zone.
1379 * NOTE: This may only be called while the zone is being torn down, and not
1380 * during normal operation. This is necessary in order that we do not have
1381 * to migrate CPUs to drain the per-CPU caches.
1384 * zone The zone to drain, must be unlocked.
1390 cache_drain(uma_zone_t zone)
1393 uma_bucket_t bucket;
1398 * XXX: It is safe to not lock the per-CPU caches, because we're
1399 * tearing down the zone anyway. I.e., there will be no further use
1400 * of the caches at this point.
1402 * XXX: It would good to be able to assert that the zone is being
1403 * torn down to prevent improper use of cache_drain().
1405 seq = SMR_SEQ_INVALID;
1406 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1407 seq = smr_advance(zone->uz_smr);
1409 cache = &zone->uz_cpu[cpu];
1410 bucket = cache_bucket_unload_alloc(cache);
1412 bucket_free(zone, bucket, NULL);
1413 bucket = cache_bucket_unload_free(cache);
1414 if (bucket != NULL) {
1415 bucket->ub_seq = seq;
1416 bucket_free(zone, bucket, NULL);
1418 bucket = cache_bucket_unload_cross(cache);
1419 if (bucket != NULL) {
1420 bucket->ub_seq = seq;
1421 bucket_free(zone, bucket, NULL);
1424 bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN);
1428 cache_shrink(uma_zone_t zone, void *unused)
1431 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1435 zone->uz_bucket_size =
1436 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1441 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1444 uma_bucket_t b1, b2, b3;
1447 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1450 b1 = b2 = b3 = NULL;
1452 cache = &zone->uz_cpu[curcpu];
1453 domain = PCPU_GET(domain);
1454 b1 = cache_bucket_unload_alloc(cache);
1457 * Don't flush SMR zone buckets. This leaves the zone without a
1458 * bucket and forces every free to synchronize().
1460 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1461 b2 = cache_bucket_unload_free(cache);
1462 b3 = cache_bucket_unload_cross(cache);
1467 zone_free_bucket(zone, b1, NULL, domain, false);
1469 zone_free_bucket(zone, b2, NULL, domain, false);
1471 /* Adjust the domain so it goes to zone_free_cross. */
1472 domain = (domain + 1) % vm_ndomains;
1473 zone_free_bucket(zone, b3, NULL, domain, false);
1478 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1479 * This is an expensive call because it needs to bind to all CPUs
1480 * one by one and enter a critical section on each of them in order
1481 * to safely access their cache buckets.
1482 * Zone lock must not be held on call this function.
1485 pcpu_cache_drain_safe(uma_zone_t zone)
1490 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1493 cache_shrink(zone, NULL);
1495 zone_foreach(cache_shrink, NULL);
1498 thread_lock(curthread);
1499 sched_bind(curthread, cpu);
1500 thread_unlock(curthread);
1503 cache_drain_safe_cpu(zone, NULL);
1505 zone_foreach(cache_drain_safe_cpu, NULL);
1507 thread_lock(curthread);
1508 sched_unbind(curthread);
1509 thread_unlock(curthread);
1513 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1514 * requested a drain, otherwise the per-domain caches are trimmed to either
1515 * estimated working set size.
1518 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain)
1520 uma_zone_domain_t zdom;
1521 uma_bucket_t bucket;
1526 * The cross bucket is partially filled and not part of
1527 * the item count. Reclaim it individually here.
1529 zdom = ZDOM_GET(zone, domain);
1530 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1531 ZONE_CROSS_LOCK(zone);
1532 bucket = zdom->uzd_cross;
1533 zdom->uzd_cross = NULL;
1534 ZONE_CROSS_UNLOCK(zone);
1536 bucket_free(zone, bucket, NULL);
1540 * If we were asked to drain the zone, we are done only once
1541 * this bucket cache is empty. If trim, we reclaim items in
1542 * excess of the zone's estimated working set size. Multiple
1543 * consecutive calls will shrink the WSS and so reclaim more.
1544 * If neither drain nor trim, then voluntarily reclaim 1/4
1545 * (to reduce first spike) of items not used for a long time.
1548 zone_domain_update_wss(zdom);
1552 target = zdom->uzd_wss;
1553 else if (zdom->uzd_timin > 900 / UMA_TIMEOUT)
1554 target = zdom->uzd_nitems - zdom->uzd_limin / 4;
1559 while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL &&
1560 zdom->uzd_nitems >= target + bucket->ub_cnt) {
1561 bucket = zone_fetch_bucket(zone, zdom, true);
1564 bucket_free(zone, bucket, NULL);
1573 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain)
1578 * Shrink the zone bucket size to ensure that the per-CPU caches
1579 * don't grow too large.
1581 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1582 zone->uz_bucket_size--;
1584 if (domain != UMA_ANYDOMAIN &&
1585 (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) {
1586 bucket_cache_reclaim_domain(zone, drain, true, domain);
1588 for (i = 0; i < vm_ndomains; i++)
1589 bucket_cache_reclaim_domain(zone, drain, true, i);
1594 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1601 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1602 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1604 mem = slab_data(slab, keg);
1605 size = PAGE_SIZE * keg->uk_ppera;
1607 kasan_mark_slab_valid(keg, mem);
1608 if (keg->uk_fini != NULL) {
1609 for (i = start - 1; i > -1; i--)
1612 * trash_fini implies that dtor was trash_dtor. trash_fini
1613 * would check that memory hasn't been modified since free,
1614 * which executed trash_dtor.
1615 * That's why we need to run uma_dbg_kskip() check here,
1616 * albeit we don't make skip check for other init/fini
1619 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1620 keg->uk_fini != trash_fini)
1622 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1624 flags = slab->us_flags;
1625 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1626 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1629 keg->uk_freef(mem, size, flags);
1630 uma_total_dec(size);
1634 keg_drain_domain(uma_keg_t keg, int domain)
1636 struct slabhead freeslabs;
1638 uma_slab_t slab, tmp;
1639 uint32_t i, stofree, stokeep, partial;
1641 dom = &keg->uk_domain[domain];
1642 LIST_INIT(&freeslabs);
1644 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1645 keg->uk_name, keg, domain, dom->ud_free_items);
1647 KEG_LOCK(keg, domain);
1650 * Are the free items in partially allocated slabs sufficient to meet
1651 * the reserve? If not, compute the number of fully free slabs that must
1654 partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1655 if (partial < keg->uk_reserve) {
1656 stokeep = min(dom->ud_free_slabs,
1657 howmany(keg->uk_reserve - partial, keg->uk_ipers));
1661 stofree = dom->ud_free_slabs - stokeep;
1664 * Partition the free slabs into two sets: those that must be kept in
1665 * order to maintain the reserve, and those that may be released back to
1666 * the system. Since one set may be much larger than the other,
1667 * populate the smaller of the two sets and swap them if necessary.
1669 for (i = min(stofree, stokeep); i > 0; i--) {
1670 slab = LIST_FIRST(&dom->ud_free_slab);
1671 LIST_REMOVE(slab, us_link);
1672 LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1674 if (stofree > stokeep)
1675 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1677 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1678 LIST_FOREACH(slab, &freeslabs, us_link)
1679 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1681 dom->ud_free_items -= stofree * keg->uk_ipers;
1682 dom->ud_free_slabs -= stofree;
1683 dom->ud_pages -= stofree * keg->uk_ppera;
1684 KEG_UNLOCK(keg, domain);
1686 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1687 keg_free_slab(keg, slab, keg->uk_ipers);
1691 * Frees pages from a keg back to the system. This is done on demand from
1692 * the pageout daemon.
1697 keg_drain(uma_keg_t keg, int domain)
1701 if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1703 if (domain != UMA_ANYDOMAIN) {
1704 keg_drain_domain(keg, domain);
1706 for (i = 0; i < vm_ndomains; i++)
1707 keg_drain_domain(keg, i);
1712 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain)
1715 * Count active reclaim operations in order to interlock with
1716 * zone_dtor(), which removes the zone from global lists before
1717 * attempting to reclaim items itself.
1719 * The zone may be destroyed while sleeping, so only zone_dtor() should
1723 if (waitok == M_WAITOK) {
1724 while (zone->uz_reclaimers > 0)
1725 msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1);
1727 zone->uz_reclaimers++;
1729 bucket_cache_reclaim(zone, drain, domain);
1731 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1732 keg_drain(zone->uz_keg, domain);
1734 zone->uz_reclaimers--;
1735 if (zone->uz_reclaimers == 0)
1741 * Allocate a new slab for a keg and inserts it into the partial slab list.
1742 * The keg should be unlocked on entry. If the allocation succeeds it will
1743 * be locked on return.
1746 * flags Wait flags for the item initialization routine
1747 * aflags Wait flags for the slab allocation
1750 * The slab that was allocated or NULL if there is no memory and the
1751 * caller specified M_NOWAIT.
1754 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1764 KASSERT(domain >= 0 && domain < vm_ndomains,
1765 ("keg_alloc_slab: domain %d out of range", domain));
1769 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1770 uma_hash_slab_t hslab;
1771 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1775 slab = &hslab->uhs_slab;
1779 * This reproduces the old vm_zone behavior of zero filling pages the
1780 * first time they are added to a zone.
1782 * Malloced items are zeroed in uma_zalloc.
1785 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1790 if (keg->uk_flags & UMA_ZONE_NODUMP)
1793 /* zone is passed for legacy reasons. */
1794 size = keg->uk_ppera * PAGE_SIZE;
1795 mem = keg->uk_allocf(zone, size, domain, &sflags, aflags);
1797 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1798 zone_free_item(slabzone(keg->uk_ipers),
1799 slab_tohashslab(slab), NULL, SKIP_NONE);
1802 uma_total_inc(size);
1804 /* For HASH zones all pages go to the same uma_domain. */
1805 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1808 kmsan_mark(mem, size,
1809 (aflags & M_ZERO) != 0 ? KMSAN_STATE_INITED : KMSAN_STATE_UNINIT);
1811 /* Point the slab into the allocated memory */
1812 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1813 slab = (uma_slab_t)(mem + keg->uk_pgoff);
1815 slab_tohashslab(slab)->uhs_data = mem;
1817 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1818 for (i = 0; i < keg->uk_ppera; i++)
1819 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1822 slab->us_freecount = keg->uk_ipers;
1823 slab->us_flags = sflags;
1824 slab->us_domain = domain;
1826 BIT_FILL(keg->uk_ipers, &slab->us_free);
1828 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1831 if (keg->uk_init != NULL) {
1832 for (i = 0; i < keg->uk_ipers; i++)
1833 if (keg->uk_init(slab_item(slab, keg, i),
1834 keg->uk_size, flags) != 0)
1836 if (i != keg->uk_ipers) {
1837 keg_free_slab(keg, slab, i);
1841 kasan_mark_slab_invalid(keg, mem);
1842 KEG_LOCK(keg, domain);
1844 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1845 slab, keg->uk_name, keg);
1847 if (keg->uk_flags & UMA_ZFLAG_HASH)
1848 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1851 * If we got a slab here it's safe to mark it partially used
1852 * and return. We assume that the caller is going to remove
1853 * at least one item.
1855 dom = &keg->uk_domain[domain];
1856 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1857 dom->ud_pages += keg->uk_ppera;
1858 dom->ud_free_items += keg->uk_ipers;
1867 * This function is intended to be used early on in place of page_alloc(). It
1868 * performs contiguous physical memory allocations and uses a bump allocator for
1869 * KVA, so is usable before the kernel map is initialized.
1872 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1879 pages = howmany(bytes, PAGE_SIZE);
1880 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1882 *pflag = UMA_SLAB_BOOT;
1883 m = vm_page_alloc_noobj_contig_domain(domain, malloc2vm_flags(wait) |
1884 VM_ALLOC_WIRED, pages, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
1885 VM_MEMATTR_DEFAULT);
1889 pa = VM_PAGE_TO_PHYS(m);
1890 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1891 #if defined(__aarch64__) || defined(__amd64__) || \
1892 defined(__riscv) || defined(__powerpc64__)
1893 if ((wait & M_NODUMP) == 0)
1898 /* Allocate KVA and indirectly advance bootmem. */
1899 return ((void *)pmap_map(&bootmem, m->phys_addr,
1900 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE));
1904 startup_free(void *mem, vm_size_t bytes)
1909 va = (vm_offset_t)mem;
1910 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1913 * startup_alloc() returns direct-mapped slabs on some platforms. Avoid
1914 * unmapping ranges of the direct map.
1916 if (va >= bootstart && va + bytes <= bootmem)
1917 pmap_remove(kernel_pmap, va, va + bytes);
1918 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1919 #if defined(__aarch64__) || defined(__amd64__) || \
1920 defined(__riscv) || defined(__powerpc64__)
1921 dump_drop_page(VM_PAGE_TO_PHYS(m));
1923 vm_page_unwire_noq(m);
1929 * Allocates a number of pages from the system
1932 * bytes The number of bytes requested
1933 * wait Shall we wait?
1936 * A pointer to the alloced memory or possibly
1937 * NULL if M_NOWAIT is set.
1940 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1943 void *p; /* Returned page */
1945 *pflag = UMA_SLAB_KERNEL;
1946 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1952 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1955 struct pglist alloctail;
1956 vm_offset_t addr, zkva;
1958 vm_page_t p, p_next;
1963 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1965 TAILQ_INIT(&alloctail);
1966 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | malloc2vm_flags(wait);
1967 *pflag = UMA_SLAB_KERNEL;
1968 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1969 if (CPU_ABSENT(cpu)) {
1970 p = vm_page_alloc_noobj(flags);
1973 p = vm_page_alloc_noobj(flags);
1975 pc = pcpu_find(cpu);
1976 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1979 p = vm_page_alloc_noobj_domain(pc->pc_domain,
1981 if (__predict_false(p == NULL))
1982 p = vm_page_alloc_noobj(flags);
1985 if (__predict_false(p == NULL))
1987 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1989 if ((addr = kva_alloc(bytes)) == 0)
1992 TAILQ_FOREACH(p, &alloctail, listq) {
1993 pmap_qenter(zkva, &p, 1);
1996 return ((void*)addr);
1998 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1999 vm_page_unwire_noq(p);
2006 * Allocates a number of pages not belonging to a VM object
2009 * bytes The number of bytes requested
2010 * wait Shall we wait?
2013 * A pointer to the alloced memory or possibly
2014 * NULL if M_NOWAIT is set.
2017 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
2020 TAILQ_HEAD(, vm_page) alloctail;
2022 vm_offset_t retkva, zkva;
2023 vm_page_t p, p_next;
2027 TAILQ_INIT(&alloctail);
2029 req = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
2030 if ((wait & M_WAITOK) != 0)
2031 req |= VM_ALLOC_WAITOK;
2033 npages = howmany(bytes, PAGE_SIZE);
2034 while (npages > 0) {
2035 p = vm_page_alloc_noobj_domain(domain, req);
2038 * Since the page does not belong to an object, its
2041 TAILQ_INSERT_TAIL(&alloctail, p, listq);
2046 * Page allocation failed, free intermediate pages and
2049 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2050 vm_page_unwire_noq(p);
2055 *flags = UMA_SLAB_PRIV;
2056 zkva = keg->uk_kva +
2057 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
2059 TAILQ_FOREACH(p, &alloctail, listq) {
2060 pmap_qenter(zkva, &p, 1);
2064 return ((void *)retkva);
2068 * Allocate physically contiguous pages.
2071 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
2075 *pflag = UMA_SLAB_KERNEL;
2076 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
2077 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
2081 * Frees a number of pages to the system
2084 * mem A pointer to the memory to be freed
2085 * size The size of the memory being freed
2086 * flags The original p->us_flags field
2092 page_free(void *mem, vm_size_t size, uint8_t flags)
2095 if ((flags & UMA_SLAB_BOOT) != 0) {
2096 startup_free(mem, size);
2100 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
2101 ("UMA: page_free used with invalid flags %x", flags));
2103 kmem_free((vm_offset_t)mem, size);
2107 * Frees pcpu zone allocations
2110 * mem A pointer to the memory to be freed
2111 * size The size of the memory being freed
2112 * flags The original p->us_flags field
2118 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
2120 vm_offset_t sva, curva;
2124 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
2126 if ((flags & UMA_SLAB_BOOT) != 0) {
2127 startup_free(mem, size);
2131 sva = (vm_offset_t)mem;
2132 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
2133 paddr = pmap_kextract(curva);
2134 m = PHYS_TO_VM_PAGE(paddr);
2135 vm_page_unwire_noq(m);
2138 pmap_qremove(sva, size >> PAGE_SHIFT);
2139 kva_free(sva, size);
2143 * Zero fill initializer
2145 * Arguments/Returns follow uma_init specifications
2148 zero_init(void *mem, int size, int flags)
2155 static struct noslabbits *
2156 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2159 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2164 * Actual size of embedded struct slab (!OFFPAGE).
2167 slab_sizeof(int nitems)
2171 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2172 return (roundup(s, UMA_ALIGN_PTR + 1));
2175 #define UMA_FIXPT_SHIFT 31
2176 #define UMA_FRAC_FIXPT(n, d) \
2177 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2178 #define UMA_FIXPT_PCT(f) \
2179 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2180 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
2181 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2184 * Compute the number of items that will fit in a slab. If hdr is true, the
2185 * item count may be limited to provide space in the slab for an inline slab
2186 * header. Otherwise, all slab space will be provided for item storage.
2189 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2194 /* The padding between items is not needed after the last item. */
2195 padpi = rsize - size;
2199 * Start with the maximum item count and remove items until
2200 * the slab header first alongside the allocatable memory.
2202 for (ipers = MIN(SLAB_MAX_SETSIZE,
2203 (slabsize + padpi - slab_sizeof(1)) / rsize);
2205 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2209 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2215 struct keg_layout_result {
2223 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2224 struct keg_layout_result *kl)
2229 kl->slabsize = slabsize;
2231 /* Handle INTERNAL as inline with an extra page. */
2232 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2233 kl->format &= ~UMA_ZFLAG_INTERNAL;
2234 kl->slabsize += PAGE_SIZE;
2237 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2238 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2240 /* Account for memory used by an offpage slab header. */
2241 total = kl->slabsize;
2242 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2243 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2245 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2249 * Determine the format of a uma keg. This determines where the slab header
2250 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2253 * keg The zone we should initialize
2259 keg_layout(uma_keg_t keg)
2261 struct keg_layout_result kl = {}, kl_tmp;
2270 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2271 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2272 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2273 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2274 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2276 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2277 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2278 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2281 alignsize = keg->uk_align + 1;
2284 * ASAN requires that each allocation be aligned to the shadow map
2287 if (alignsize < KASAN_SHADOW_SCALE)
2288 alignsize = KASAN_SHADOW_SCALE;
2292 * Calculate the size of each allocation (rsize) according to
2293 * alignment. If the requested size is smaller than we have
2294 * allocation bits for we round it up.
2296 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2297 rsize = roundup2(rsize, alignsize);
2299 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2301 * We want one item to start on every align boundary in a page.
2302 * To do this we will span pages. We will also extend the item
2303 * by the size of align if it is an even multiple of align.
2304 * Otherwise, it would fall on the same boundary every time.
2306 if ((rsize & alignsize) == 0)
2308 slabsize = rsize * (PAGE_SIZE / alignsize);
2309 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2310 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2311 slabsize = round_page(slabsize);
2314 * Start with a slab size of as many pages as it takes to
2315 * represent a single item. We will try to fit as many
2316 * additional items into the slab as possible.
2318 slabsize = round_page(keg->uk_size);
2321 /* Build a list of all of the available formats for this keg. */
2324 /* Evaluate an inline slab layout. */
2325 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2328 /* TODO: vm_page-embedded slab. */
2331 * We can't do OFFPAGE if we're internal or if we've been
2332 * asked to not go to the VM for buckets. If we do this we
2333 * may end up going to the VM for slabs which we do not want
2334 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2335 * In those cases, evaluate a pseudo-format called INTERNAL
2336 * which has an inline slab header and one extra page to
2337 * guarantee that it fits.
2339 * Otherwise, see if using an OFFPAGE slab will improve our
2342 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2343 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2345 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2348 * Choose a slab size and format which satisfy the minimum efficiency.
2349 * Prefer the smallest slab size that meets the constraints.
2351 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2352 * for small items (up to PAGE_SIZE), the iteration increment is one
2353 * page; and for large items, the increment is one item.
2355 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2356 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2357 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2360 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2361 round_page(rsize * (i - 1) + keg->uk_size);
2363 for (j = 0; j < nfmt; j++) {
2364 /* Only if we have no viable format yet. */
2365 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2369 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2370 if (kl_tmp.eff <= kl.eff)
2375 CTR6(KTR_UMA, "keg %s layout: format %#x "
2376 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2377 keg->uk_name, kl.format, kl.ipers, rsize,
2378 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2380 /* Stop when we reach the minimum efficiency. */
2381 if (kl.eff >= UMA_MIN_EFF)
2385 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2386 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2387 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2391 pages = atop(kl.slabsize);
2392 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2393 pages *= mp_maxid + 1;
2395 keg->uk_rsize = rsize;
2396 keg->uk_ipers = kl.ipers;
2397 keg->uk_ppera = pages;
2398 keg->uk_flags |= kl.format;
2401 * How do we find the slab header if it is offpage or if not all item
2402 * start addresses are in the same page? We could solve the latter
2403 * case with vaddr alignment, but we don't.
2405 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2406 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2407 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2408 keg->uk_flags |= UMA_ZFLAG_HASH;
2410 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2413 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2414 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2416 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2417 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2418 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2419 keg->uk_ipers, pages));
2423 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2424 * the keg onto the global keg list.
2426 * Arguments/Returns follow uma_ctor specifications
2427 * udata Actually uma_kctor_args
2430 keg_ctor(void *mem, int size, void *udata, int flags)
2432 struct uma_kctor_args *arg = udata;
2433 uma_keg_t keg = mem;
2438 keg->uk_size = arg->size;
2439 keg->uk_init = arg->uminit;
2440 keg->uk_fini = arg->fini;
2441 keg->uk_align = arg->align;
2442 keg->uk_reserve = 0;
2443 keg->uk_flags = arg->flags;
2446 * We use a global round-robin policy by default. Zones with
2447 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2448 * case the iterator is never run.
2450 keg->uk_dr.dr_policy = DOMAINSET_RR();
2451 keg->uk_dr.dr_iter = 0;
2454 * The primary zone is passed to us at keg-creation time.
2457 keg->uk_name = zone->uz_name;
2459 if (arg->flags & UMA_ZONE_ZINIT)
2460 keg->uk_init = zero_init;
2462 if (arg->flags & UMA_ZONE_MALLOC)
2463 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2466 keg->uk_flags &= ~UMA_ZONE_PCPU;
2472 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2473 * work on. Use round-robin for everything else.
2475 * Zones may override the default by specifying either.
2478 if ((keg->uk_flags &
2479 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2480 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2481 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2482 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2486 * If we haven't booted yet we need allocations to go through the
2487 * startup cache until the vm is ready.
2489 #ifdef UMA_MD_SMALL_ALLOC
2490 if (keg->uk_ppera == 1)
2491 keg->uk_allocf = uma_small_alloc;
2494 if (booted < BOOT_KVA)
2495 keg->uk_allocf = startup_alloc;
2496 else if (keg->uk_flags & UMA_ZONE_PCPU)
2497 keg->uk_allocf = pcpu_page_alloc;
2498 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2499 keg->uk_allocf = contig_alloc;
2501 keg->uk_allocf = page_alloc;
2502 #ifdef UMA_MD_SMALL_ALLOC
2503 if (keg->uk_ppera == 1)
2504 keg->uk_freef = uma_small_free;
2507 if (keg->uk_flags & UMA_ZONE_PCPU)
2508 keg->uk_freef = pcpu_page_free;
2510 keg->uk_freef = page_free;
2513 * Initialize keg's locks.
2515 for (i = 0; i < vm_ndomains; i++)
2516 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2519 * If we're putting the slab header in the actual page we need to
2520 * figure out where in each page it goes. See slab_sizeof
2523 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2526 shsize = slab_sizeof(keg->uk_ipers);
2527 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2529 * The only way the following is possible is if with our
2530 * UMA_ALIGN_PTR adjustments we are now bigger than
2531 * UMA_SLAB_SIZE. I haven't checked whether this is
2532 * mathematically possible for all cases, so we make
2535 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2536 ("zone %s ipers %d rsize %d size %d slab won't fit",
2537 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2540 if (keg->uk_flags & UMA_ZFLAG_HASH)
2541 hash_alloc(&keg->uk_hash, 0);
2543 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2545 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2547 rw_wlock(&uma_rwlock);
2548 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2549 rw_wunlock(&uma_rwlock);
2554 zone_kva_available(uma_zone_t zone, void *unused)
2558 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2562 if (keg->uk_allocf == startup_alloc) {
2563 /* Switch to the real allocator. */
2564 if (keg->uk_flags & UMA_ZONE_PCPU)
2565 keg->uk_allocf = pcpu_page_alloc;
2566 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2568 keg->uk_allocf = contig_alloc;
2570 keg->uk_allocf = page_alloc;
2575 zone_alloc_counters(uma_zone_t zone, void *unused)
2578 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2579 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2580 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2581 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2585 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2587 uma_zone_domain_t zdom;
2590 struct sysctl_oid *oid, *domainoid;
2591 int domains, i, cnt;
2592 static const char *nokeg = "cache zone";
2596 * Make a sysctl safe copy of the zone name by removing
2597 * any special characters and handling dups by appending
2600 if (zone->uz_namecnt != 0) {
2601 /* Count the number of decimal digits and '_' separator. */
2602 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2604 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2606 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2609 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2610 for (c = zone->uz_ctlname; *c != '\0'; c++)
2611 if (strchr("./\\ -", *c) != NULL)
2615 * Basic parameters at the root.
2617 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2618 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2620 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2621 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2622 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2623 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2624 zone, 0, sysctl_handle_uma_zone_flags, "A",
2625 "Allocator configuration flags");
2626 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2627 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2628 "Desired per-cpu cache size");
2629 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2630 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2631 "Maximum allowed per-cpu cache size");
2636 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2637 domains = vm_ndomains;
2640 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2641 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2643 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2644 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2645 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2646 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2647 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2648 "Real object size with alignment");
2649 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2650 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2651 "pages per-slab allocation");
2652 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2653 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2654 "items available per-slab");
2655 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2656 "align", CTLFLAG_RD, &keg->uk_align, 0,
2657 "item alignment mask");
2658 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2659 "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2660 "number of reserved items");
2661 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2662 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2663 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2664 "Slab utilization (100 - internal fragmentation %)");
2665 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2666 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2667 for (i = 0; i < domains; i++) {
2668 dom = &keg->uk_domain[i];
2669 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2670 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2671 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2672 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2673 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2674 "Total pages currently allocated from VM");
2675 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2676 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2677 "Items free in the slab layer");
2678 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2679 "free_slabs", CTLFLAG_RD, &dom->ud_free_slabs, 0,
2683 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2684 "name", CTLFLAG_RD, nokeg, "Keg name");
2687 * Information about zone limits.
2689 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2690 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2691 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2692 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2693 zone, 0, sysctl_handle_uma_zone_items, "QU",
2694 "Current number of allocated items if limit is set");
2695 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2696 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2697 "Maximum number of allocated and cached items");
2698 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2699 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2700 "Number of threads sleeping at limit");
2701 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2702 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2703 "Total zone limit sleeps");
2704 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2705 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2706 "Maximum number of items in each domain's bucket cache");
2709 * Per-domain zone information.
2711 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2712 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2713 for (i = 0; i < domains; i++) {
2714 zdom = ZDOM_GET(zone, i);
2715 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2716 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2717 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2718 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2719 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2720 "number of items in this domain");
2721 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2722 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2723 "maximum item count in this period");
2724 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2725 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2726 "minimum item count in this period");
2727 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2728 "bimin", CTLFLAG_RD, &zdom->uzd_bimin,
2729 "Minimum item count in this batch");
2730 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2731 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2732 "Working set size");
2733 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2734 "limin", CTLFLAG_RD, &zdom->uzd_limin,
2735 "Long time minimum item count");
2736 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2737 "timin", CTLFLAG_RD, &zdom->uzd_timin, 0,
2738 "Time since zero long time minimum item count");
2742 * General statistics.
2744 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2745 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2746 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2747 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2748 zone, 1, sysctl_handle_uma_zone_cur, "I",
2749 "Current number of allocated items");
2750 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2751 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2752 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2753 "Total allocation calls");
2754 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2755 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2756 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2757 "Total free calls");
2758 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2759 "fails", CTLFLAG_RD, &zone->uz_fails,
2760 "Number of allocation failures");
2761 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2762 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2763 "Free calls from the wrong domain");
2766 struct uma_zone_count {
2772 zone_count(uma_zone_t zone, void *arg)
2774 struct uma_zone_count *cnt;
2778 * Some zones are rapidly created with identical names and
2779 * destroyed out of order. This can lead to gaps in the count.
2780 * Use one greater than the maximum observed for this name.
2782 if (strcmp(zone->uz_name, cnt->name) == 0)
2783 cnt->count = MAX(cnt->count,
2784 zone->uz_namecnt + 1);
2788 zone_update_caches(uma_zone_t zone)
2792 for (i = 0; i <= mp_maxid; i++) {
2793 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2794 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2799 * Zone header ctor. This initializes all fields, locks, etc.
2801 * Arguments/Returns follow uma_ctor specifications
2802 * udata Actually uma_zctor_args
2805 zone_ctor(void *mem, int size, void *udata, int flags)
2807 struct uma_zone_count cnt;
2808 struct uma_zctor_args *arg = udata;
2809 uma_zone_domain_t zdom;
2810 uma_zone_t zone = mem;
2816 zone->uz_name = arg->name;
2817 zone->uz_ctor = arg->ctor;
2818 zone->uz_dtor = arg->dtor;
2819 zone->uz_init = NULL;
2820 zone->uz_fini = NULL;
2821 zone->uz_sleeps = 0;
2822 zone->uz_bucket_size = 0;
2823 zone->uz_bucket_size_min = 0;
2824 zone->uz_bucket_size_max = BUCKET_MAX;
2825 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2826 zone->uz_warning = NULL;
2827 /* The domain structures follow the cpu structures. */
2828 zone->uz_bucket_max = ULONG_MAX;
2829 timevalclear(&zone->uz_ratecheck);
2831 /* Count the number of duplicate names. */
2832 cnt.name = arg->name;
2834 zone_foreach(zone_count, &cnt);
2835 zone->uz_namecnt = cnt.count;
2836 ZONE_CROSS_LOCK_INIT(zone);
2838 for (i = 0; i < vm_ndomains; i++) {
2839 zdom = ZDOM_GET(zone, i);
2840 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2841 STAILQ_INIT(&zdom->uzd_buckets);
2844 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
2845 if (arg->uminit == trash_init && arg->fini == trash_fini)
2846 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2847 #elif defined(KASAN)
2848 if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0)
2849 arg->flags |= UMA_ZONE_NOKASAN;
2853 * This is a pure cache zone, no kegs.
2856 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2857 ("zone_ctor: Import specified for non-cache zone."));
2858 zone->uz_flags = arg->flags;
2859 zone->uz_size = arg->size;
2860 zone->uz_import = arg->import;
2861 zone->uz_release = arg->release;
2862 zone->uz_arg = arg->arg;
2865 * Cache zones are round-robin unless a policy is
2866 * specified because they may have incompatible
2869 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2870 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2872 rw_wlock(&uma_rwlock);
2873 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2874 rw_wunlock(&uma_rwlock);
2879 * Use the regular zone/keg/slab allocator.
2881 zone->uz_import = zone_import;
2882 zone->uz_release = zone_release;
2883 zone->uz_arg = zone;
2886 if (arg->flags & UMA_ZONE_SECONDARY) {
2887 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2888 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2889 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2890 zone->uz_init = arg->uminit;
2891 zone->uz_fini = arg->fini;
2892 zone->uz_flags |= UMA_ZONE_SECONDARY;
2893 rw_wlock(&uma_rwlock);
2895 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2896 if (LIST_NEXT(z, uz_link) == NULL) {
2897 LIST_INSERT_AFTER(z, zone, uz_link);
2902 rw_wunlock(&uma_rwlock);
2903 } else if (keg == NULL) {
2904 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2905 arg->align, arg->flags)) == NULL)
2908 struct uma_kctor_args karg;
2911 /* We should only be here from uma_startup() */
2912 karg.size = arg->size;
2913 karg.uminit = arg->uminit;
2914 karg.fini = arg->fini;
2915 karg.align = arg->align;
2916 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2918 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2924 /* Inherit properties from the keg. */
2926 zone->uz_size = keg->uk_size;
2927 zone->uz_flags |= (keg->uk_flags &
2928 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2931 if (booted >= BOOT_PCPU) {
2932 zone_alloc_counters(zone, NULL);
2933 if (booted >= BOOT_RUNNING)
2934 zone_alloc_sysctl(zone, NULL);
2936 zone->uz_allocs = EARLY_COUNTER;
2937 zone->uz_frees = EARLY_COUNTER;
2938 zone->uz_fails = EARLY_COUNTER;
2941 /* Caller requests a private SMR context. */
2942 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2943 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2945 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2946 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2947 ("Invalid zone flag combination"));
2948 if (arg->flags & UMA_ZFLAG_INTERNAL)
2949 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2950 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2951 zone->uz_bucket_size = BUCKET_MAX;
2952 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2953 zone->uz_bucket_size = 0;
2955 zone->uz_bucket_size = bucket_select(zone->uz_size);
2956 zone->uz_bucket_size_min = zone->uz_bucket_size;
2957 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2958 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2959 zone_update_caches(zone);
2965 * Keg header dtor. This frees all data, destroys locks, frees the hash
2966 * table and removes the keg from the global list.
2968 * Arguments/Returns follow uma_dtor specifications
2972 keg_dtor(void *arg, int size, void *udata)
2975 uint32_t free, pages;
2978 keg = (uma_keg_t)arg;
2980 for (i = 0; i < vm_ndomains; i++) {
2981 free += keg->uk_domain[i].ud_free_items;
2982 pages += keg->uk_domain[i].ud_pages;
2983 KEG_LOCK_FINI(keg, i);
2986 printf("Freed UMA keg (%s) was not empty (%u items). "
2987 " Lost %u pages of memory.\n",
2988 keg->uk_name ? keg->uk_name : "",
2989 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2991 hash_free(&keg->uk_hash);
2997 * Arguments/Returns follow uma_dtor specifications
3001 zone_dtor(void *arg, int size, void *udata)
3007 zone = (uma_zone_t)arg;
3009 sysctl_remove_oid(zone->uz_oid, 1, 1);
3011 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
3014 rw_wlock(&uma_rwlock);
3015 LIST_REMOVE(zone, uz_link);
3016 rw_wunlock(&uma_rwlock);
3017 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3019 keg->uk_reserve = 0;
3021 zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true);
3024 * We only destroy kegs from non secondary/non cache zones.
3026 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3028 rw_wlock(&uma_rwlock);
3029 LIST_REMOVE(keg, uk_link);
3030 rw_wunlock(&uma_rwlock);
3031 zone_free_item(kegs, keg, NULL, SKIP_NONE);
3033 counter_u64_free(zone->uz_allocs);
3034 counter_u64_free(zone->uz_frees);
3035 counter_u64_free(zone->uz_fails);
3036 counter_u64_free(zone->uz_xdomain);
3037 free(zone->uz_ctlname, M_UMA);
3038 for (i = 0; i < vm_ndomains; i++)
3039 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
3040 ZONE_CROSS_LOCK_FINI(zone);
3044 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3049 LIST_FOREACH(keg, &uma_kegs, uk_link) {
3050 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
3053 LIST_FOREACH(zone, &uma_cachezones, uz_link)
3058 * Traverses every zone in the system and calls a callback
3061 * zfunc A pointer to a function which accepts a zone
3068 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3071 rw_rlock(&uma_rwlock);
3072 zone_foreach_unlocked(zfunc, arg);
3073 rw_runlock(&uma_rwlock);
3077 * Initialize the kernel memory allocator. This is done after pages can be
3078 * allocated but before general KVA is available.
3081 uma_startup1(vm_offset_t virtual_avail)
3083 struct uma_zctor_args args;
3084 size_t ksize, zsize, size;
3085 uma_keg_t primarykeg;
3090 bootstart = bootmem = virtual_avail;
3092 rw_init(&uma_rwlock, "UMA lock");
3093 sx_init(&uma_reclaim_lock, "umareclaim");
3095 ksize = sizeof(struct uma_keg) +
3096 (sizeof(struct uma_domain) * vm_ndomains);
3097 ksize = roundup(ksize, UMA_SUPER_ALIGN);
3098 zsize = sizeof(struct uma_zone) +
3099 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
3100 (sizeof(struct uma_zone_domain) * vm_ndomains);
3101 zsize = roundup(zsize, UMA_SUPER_ALIGN);
3103 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
3104 size = (zsize * 2) + ksize;
3105 for (domain = 0; domain < vm_ndomains; domain++) {
3106 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
3111 zones = (uma_zone_t)m;
3113 kegs = (uma_zone_t)m;
3115 primarykeg = (uma_keg_t)m;
3117 /* "manually" create the initial zone */
3118 memset(&args, 0, sizeof(args));
3119 args.name = "UMA Kegs";
3121 args.ctor = keg_ctor;
3122 args.dtor = keg_dtor;
3123 args.uminit = zero_init;
3125 args.keg = primarykeg;
3126 args.align = UMA_SUPER_ALIGN - 1;
3127 args.flags = UMA_ZFLAG_INTERNAL;
3128 zone_ctor(kegs, zsize, &args, M_WAITOK);
3130 args.name = "UMA Zones";
3132 args.ctor = zone_ctor;
3133 args.dtor = zone_dtor;
3134 args.uminit = zero_init;
3137 args.align = UMA_SUPER_ALIGN - 1;
3138 args.flags = UMA_ZFLAG_INTERNAL;
3139 zone_ctor(zones, zsize, &args, M_WAITOK);
3141 /* Now make zones for slab headers */
3142 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
3143 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3144 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
3145 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3147 hashzone = uma_zcreate("UMA Hash",
3148 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
3149 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3155 #ifndef UMA_MD_SMALL_ALLOC
3156 extern void vm_radix_reserve_kva(void);
3160 * Advertise the availability of normal kva allocations and switch to
3161 * the default back-end allocator. Marks the KVA we consumed on startup
3162 * as used in the map.
3168 if (bootstart != bootmem) {
3169 vm_map_lock(kernel_map);
3170 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
3171 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
3172 vm_map_unlock(kernel_map);
3175 #ifndef UMA_MD_SMALL_ALLOC
3176 /* Set up radix zone to use noobj_alloc. */
3177 vm_radix_reserve_kva();
3181 zone_foreach_unlocked(zone_kva_available, NULL);
3186 * Allocate counters as early as possible so that boot-time allocations are
3187 * accounted more precisely.
3190 uma_startup_pcpu(void *arg __unused)
3193 zone_foreach_unlocked(zone_alloc_counters, NULL);
3196 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
3199 * Finish our initialization steps.
3202 uma_startup3(void *arg __unused)
3206 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
3207 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
3208 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
3210 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
3211 callout_init(&uma_callout, 1);
3212 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
3213 booted = BOOT_RUNNING;
3215 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
3216 EVENTHANDLER_PRI_FIRST);
3218 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
3224 booted = BOOT_SHUTDOWN;
3228 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3229 int align, uint32_t flags)
3231 struct uma_kctor_args args;
3234 args.uminit = uminit;
3236 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
3239 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3242 /* Public functions */
3245 uma_set_align(int align)
3248 if (align != UMA_ALIGN_CACHE)
3249 uma_align_cache = align;
3254 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3255 uma_init uminit, uma_fini fini, int align, uint32_t flags)
3258 struct uma_zctor_args args;
3261 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
3264 /* This stuff is essential for the zone ctor */
3265 memset(&args, 0, sizeof(args));
3270 args.uminit = uminit;
3272 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
3274 * Inject procedures which check for memory use after free if we are
3275 * allowed to scramble the memory while it is not allocated. This
3276 * requires that: UMA is actually able to access the memory, no init
3277 * or fini procedures, no dependency on the initial value of the
3278 * memory, and no (legitimate) use of the memory after free. Note,
3279 * the ctor and dtor do not need to be empty.
3281 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3282 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3283 args.uminit = trash_init;
3284 args.fini = trash_fini;
3291 sx_xlock(&uma_reclaim_lock);
3292 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3293 sx_xunlock(&uma_reclaim_lock);
3300 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3301 uma_init zinit, uma_fini zfini, uma_zone_t primary)
3303 struct uma_zctor_args args;
3307 keg = primary->uz_keg;
3308 memset(&args, 0, sizeof(args));
3310 args.size = keg->uk_size;
3313 args.uminit = zinit;
3315 args.align = keg->uk_align;
3316 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3319 sx_xlock(&uma_reclaim_lock);
3320 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3321 sx_xunlock(&uma_reclaim_lock);
3328 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3329 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3330 void *arg, int flags)
3332 struct uma_zctor_args args;
3334 memset(&args, 0, sizeof(args));
3339 args.uminit = zinit;
3341 args.import = zimport;
3342 args.release = zrelease;
3345 args.flags = flags | UMA_ZFLAG_CACHE;
3347 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3352 uma_zdestroy(uma_zone_t zone)
3356 * Large slabs are expensive to reclaim, so don't bother doing
3357 * unnecessary work if we're shutting down.
3359 if (booted == BOOT_SHUTDOWN &&
3360 zone->uz_fini == NULL && zone->uz_release == zone_release)
3362 sx_xlock(&uma_reclaim_lock);
3363 zone_free_item(zones, zone, NULL, SKIP_NONE);
3364 sx_xunlock(&uma_reclaim_lock);
3368 uma_zwait(uma_zone_t zone)
3371 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3372 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3373 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3374 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3376 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3380 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3382 void *item, *pcpu_item;
3386 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3388 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3391 pcpu_item = zpcpu_base_to_offset(item);
3392 if (flags & M_ZERO) {
3394 for (i = 0; i <= mp_maxid; i++)
3395 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3397 bzero(item, zone->uz_size);
3404 * A stub while both regular and pcpu cases are identical.
3407 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3412 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3415 /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */
3416 if (pcpu_item == NULL)
3419 item = zpcpu_offset_to_base(pcpu_item);
3420 uma_zfree_arg(zone, item, udata);
3423 static inline void *
3424 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3431 kasan_mark_item_valid(zone, item);
3432 kmsan_mark_item_uninitialized(zone, item);
3435 skipdbg = uma_dbg_zskip(zone, item);
3436 if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3437 zone->uz_ctor != trash_ctor)
3438 trash_ctor(item, size, udata, flags);
3441 /* Check flags before loading ctor pointer. */
3442 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3443 __predict_false(zone->uz_ctor != NULL) &&
3444 zone->uz_ctor(item, size, udata, flags) != 0) {
3445 counter_u64_add(zone->uz_fails, 1);
3446 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3451 uma_dbg_alloc(zone, NULL, item);
3453 if (__predict_false(flags & M_ZERO))
3454 return (memset(item, 0, size));
3460 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3461 enum zfreeskip skip)
3466 skipdbg = uma_dbg_zskip(zone, item);
3467 if (skip == SKIP_NONE && !skipdbg) {
3468 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3469 uma_dbg_free(zone, udata, item);
3471 uma_dbg_free(zone, NULL, item);
3474 if (__predict_true(skip < SKIP_DTOR)) {
3475 if (zone->uz_dtor != NULL)
3476 zone->uz_dtor(item, size, udata);
3478 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3479 zone->uz_dtor != trash_dtor)
3480 trash_dtor(item, size, udata);
3483 kasan_mark_item_invalid(zone, item);
3488 item_domain(void *item)
3492 domain = vm_phys_domain(vtophys(item));
3493 KASSERT(domain >= 0 && domain < vm_ndomains,
3494 ("%s: unknown domain for item %p", __func__, item));
3499 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3500 #if defined(INVARIANTS) && (defined(DDB) || defined(STACK))
3501 #include <sys/stack.h>
3503 #define UMA_ZALLOC_DEBUG
3505 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3511 if (flags & M_WAITOK) {
3512 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3513 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3518 KASSERT((flags & M_EXEC) == 0,
3519 ("uma_zalloc_debug: called with M_EXEC"));
3520 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3521 ("uma_zalloc_debug: called within spinlock or critical section"));
3522 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3523 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3525 _Static_assert(M_NOWAIT != 0 && M_WAITOK != 0,
3526 "M_NOWAIT and M_WAITOK must be non-zero for this assertion:");
3529 * Give the #elif clause time to find problems, then remove it
3530 * and enable this. (Remove <sys/stack.h> above, too.)
3532 KASSERT((flags & (M_NOWAIT|M_WAITOK)) == M_NOWAIT ||
3533 (flags & (M_NOWAIT|M_WAITOK)) == M_WAITOK,
3534 ("uma_zalloc_debug: must pass one of M_NOWAIT or M_WAITOK"));
3535 #elif defined(DDB) || defined(STACK)
3536 if (__predict_false((flags & (M_NOWAIT|M_WAITOK)) != M_NOWAIT &&
3537 (flags & (M_NOWAIT|M_WAITOK)) != M_WAITOK)) {
3538 static int stack_count;
3541 if (stack_count < 10) {
3543 printf("uma_zalloc* called with bad WAIT flags:\n");
3551 #ifdef DEBUG_MEMGUARD
3552 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3554 item = memguard_alloc(zone->uz_size, flags);
3556 error = EJUSTRETURN;
3557 if (zone->uz_init != NULL &&
3558 zone->uz_init(item, zone->uz_size, flags) != 0) {
3562 if (zone->uz_ctor != NULL &&
3563 zone->uz_ctor(item, zone->uz_size, udata,
3565 counter_u64_add(zone->uz_fails, 1);
3566 if (zone->uz_fini != NULL)
3567 zone->uz_fini(item, zone->uz_size);
3574 /* This is unfortunate but should not be fatal. */
3581 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3583 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3584 ("uma_zfree_debug: called with spinlock or critical section held"));
3586 #ifdef DEBUG_MEMGUARD
3587 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3588 if (zone->uz_dtor != NULL)
3589 zone->uz_dtor(item, zone->uz_size, udata);
3590 if (zone->uz_fini != NULL)
3591 zone->uz_fini(item, zone->uz_size);
3592 memguard_free(item);
3593 return (EJUSTRETURN);
3600 static inline void *
3601 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3602 void *udata, int flags)
3607 item = cache_bucket_pop(cache, bucket);
3608 size = cache_uz_size(cache);
3609 uz_flags = cache_uz_flags(cache);
3611 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3614 static __noinline void *
3615 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3617 uma_cache_bucket_t bucket;
3620 while (cache_alloc(zone, cache, udata, flags)) {
3621 cache = &zone->uz_cpu[curcpu];
3622 bucket = &cache->uc_allocbucket;
3623 if (__predict_false(bucket->ucb_cnt == 0))
3625 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3630 * We can not get a bucket so try to return a single item.
3632 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3633 domain = PCPU_GET(domain);
3635 domain = UMA_ANYDOMAIN;
3636 return (zone_alloc_item(zone, udata, domain, flags));
3641 uma_zalloc_smr(uma_zone_t zone, int flags)
3643 uma_cache_bucket_t bucket;
3646 CTR3(KTR_UMA, "uma_zalloc_smr zone %s(%p) flags %d", zone->uz_name,
3649 #ifdef UMA_ZALLOC_DEBUG
3652 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3653 ("uma_zalloc_arg: called with non-SMR zone."));
3654 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3659 cache = &zone->uz_cpu[curcpu];
3660 bucket = &cache->uc_allocbucket;
3661 if (__predict_false(bucket->ucb_cnt == 0))
3662 return (cache_alloc_retry(zone, cache, NULL, flags));
3663 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3668 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3670 uma_cache_bucket_t bucket;
3673 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3674 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3676 /* This is the fast path allocation */
3677 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3680 #ifdef UMA_ZALLOC_DEBUG
3683 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3684 ("uma_zalloc_arg: called with SMR zone."));
3685 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3690 * If possible, allocate from the per-CPU cache. There are two
3691 * requirements for safe access to the per-CPU cache: (1) the thread
3692 * accessing the cache must not be preempted or yield during access,
3693 * and (2) the thread must not migrate CPUs without switching which
3694 * cache it accesses. We rely on a critical section to prevent
3695 * preemption and migration. We release the critical section in
3696 * order to acquire the zone mutex if we are unable to allocate from
3697 * the current cache; when we re-acquire the critical section, we
3698 * must detect and handle migration if it has occurred.
3701 cache = &zone->uz_cpu[curcpu];
3702 bucket = &cache->uc_allocbucket;
3703 if (__predict_false(bucket->ucb_cnt == 0))
3704 return (cache_alloc_retry(zone, cache, udata, flags));
3705 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3709 * Replenish an alloc bucket and possibly restore an old one. Called in
3710 * a critical section. Returns in a critical section.
3712 * A false return value indicates an allocation failure.
3713 * A true return value indicates success and the caller should retry.
3715 static __noinline bool
3716 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3718 uma_bucket_t bucket;
3719 int curdomain, domain;
3722 CRITICAL_ASSERT(curthread);
3725 * If we have run out of items in our alloc bucket see
3726 * if we can switch with the free bucket.
3728 * SMR Zones can't re-use the free bucket until the sequence has
3731 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3732 cache->uc_freebucket.ucb_cnt != 0) {
3733 cache_bucket_swap(&cache->uc_freebucket,
3734 &cache->uc_allocbucket);
3739 * Discard any empty allocation bucket while we hold no locks.
3741 bucket = cache_bucket_unload_alloc(cache);
3744 if (bucket != NULL) {
3745 KASSERT(bucket->ub_cnt == 0,
3746 ("cache_alloc: Entered with non-empty alloc bucket."));
3747 bucket_free(zone, bucket, udata);
3751 * Attempt to retrieve the item from the per-CPU cache has failed, so
3752 * we must go back to the zone. This requires the zdom lock, so we
3753 * must drop the critical section, then re-acquire it when we go back
3754 * to the cache. Since the critical section is released, we may be
3755 * preempted or migrate. As such, make sure not to maintain any
3756 * thread-local state specific to the cache from prior to releasing
3757 * the critical section.
3759 domain = PCPU_GET(domain);
3760 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3761 VM_DOMAIN_EMPTY(domain))
3762 domain = zone_domain_highest(zone, domain);
3763 bucket = cache_fetch_bucket(zone, cache, domain);
3764 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3765 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3771 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3772 zone->uz_name, zone, bucket);
3773 if (bucket == NULL) {
3779 * See if we lost the race or were migrated. Cache the
3780 * initialized bucket to make this less likely or claim
3781 * the memory directly.
3784 cache = &zone->uz_cpu[curcpu];
3785 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3786 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3787 (curdomain = PCPU_GET(domain)) == domain ||
3788 VM_DOMAIN_EMPTY(curdomain))) {
3790 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3792 cache_bucket_load_alloc(cache, bucket);
3797 * We lost the race, release this bucket and start over.
3800 zone_put_bucket(zone, domain, bucket, udata, !new);
3807 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3810 uma_bucket_t bucket;
3811 uma_zone_domain_t zdom;
3815 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3816 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3818 /* This is the fast path allocation */
3819 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3820 zone->uz_name, zone, domain, flags);
3822 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3823 ("uma_zalloc_domain: called with SMR zone."));
3825 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3826 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3828 if (vm_ndomains == 1)
3829 return (uma_zalloc_arg(zone, udata, flags));
3831 #ifdef UMA_ZALLOC_DEBUG
3832 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3837 * Try to allocate from the bucket cache before falling back to the keg.
3838 * We could try harder and attempt to allocate from per-CPU caches or
3839 * the per-domain cross-domain buckets, but the complexity is probably
3840 * not worth it. It is more important that frees of previous
3841 * cross-domain allocations do not blow up the cache.
3843 zdom = zone_domain_lock(zone, domain);
3844 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3845 item = bucket->ub_bucket[bucket->ub_cnt - 1];
3847 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3850 zone_put_bucket(zone, domain, bucket, udata, true);
3851 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3854 KASSERT(item_domain(item) == domain,
3855 ("%s: bucket cache item %p from wrong domain",
3857 counter_u64_add(zone->uz_allocs, 1);
3862 return (zone_alloc_item(zone, udata, domain, flags));
3864 return (uma_zalloc_arg(zone, udata, flags));
3869 * Find a slab with some space. Prefer slabs that are partially used over those
3870 * that are totally full. This helps to reduce fragmentation.
3872 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3876 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3882 KASSERT(domain >= 0 && domain < vm_ndomains,
3883 ("keg_first_slab: domain %d out of range", domain));
3884 KEG_LOCK_ASSERT(keg, domain);
3889 dom = &keg->uk_domain[domain];
3890 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3892 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3893 LIST_REMOVE(slab, us_link);
3894 dom->ud_free_slabs--;
3895 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3899 domain = (domain + 1) % vm_ndomains;
3900 } while (domain != start);
3906 * Fetch an existing slab from a free or partial list. Returns with the
3907 * keg domain lock held if a slab was found or unlocked if not.
3910 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3915 /* HASH has a single free list. */
3916 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3919 KEG_LOCK(keg, domain);
3920 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3921 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3922 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3923 KEG_UNLOCK(keg, domain);
3930 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3932 struct vm_domainset_iter di;
3937 KASSERT((flags & (M_WAITOK | M_NOVM)) != (M_WAITOK | M_NOVM),
3938 ("%s: invalid flags %#x", __func__, flags));
3942 * Use the keg's policy if upper layers haven't already specified a
3943 * domain (as happens with first-touch zones).
3945 * To avoid races we run the iterator with the keg lock held, but that
3946 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3947 * clear M_WAITOK and handle low memory conditions locally.
3949 rr = rdomain == UMA_ANYDOMAIN;
3951 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3952 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3960 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3965 * M_NOVM is used to break the recursion that can otherwise
3966 * occur if low-level memory management routines use UMA.
3968 if ((flags & M_NOVM) == 0) {
3969 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3975 if ((flags & M_USE_RESERVE) != 0) {
3977 * Drain reserves from other domains before
3978 * giving up or sleeping. It may be useful to
3979 * support per-domain reserves eventually.
3981 rdomain = UMA_ANYDOMAIN;
3984 if ((flags & M_WAITOK) == 0)
3986 vm_wait_domain(domain);
3987 } else if (vm_domainset_iter_policy(&di, &domain) != 0) {
3988 if ((flags & M_WAITOK) != 0) {
3989 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
3997 * We might not have been able to get a slab but another cpu
3998 * could have while we were unlocked. Check again before we
4001 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
4008 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
4014 KEG_LOCK_ASSERT(keg, slab->us_domain);
4016 dom = &keg->uk_domain[slab->us_domain];
4017 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
4018 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
4019 item = slab_item(slab, keg, freei);
4020 slab->us_freecount--;
4021 dom->ud_free_items--;
4024 * Move this slab to the full list. It must be on the partial list, so
4025 * we do not need to update the free slab count. In particular,
4026 * keg_fetch_slab() always returns slabs on the partial list.
4028 if (slab->us_freecount == 0) {
4029 LIST_REMOVE(slab, us_link);
4030 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
4037 zone_import(void *arg, void **bucket, int max, int domain, int flags)
4051 /* Try to keep the buckets totally full */
4052 for (i = 0; i < max; ) {
4053 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
4056 stripe = howmany(max, vm_ndomains);
4058 dom = &keg->uk_domain[slab->us_domain];
4060 bucket[i++] = slab_alloc_item(keg, slab);
4061 if (keg->uk_reserve > 0 &&
4062 dom->ud_free_items <= keg->uk_reserve) {
4064 * Avoid depleting the reserve after a
4065 * successful item allocation, even if
4066 * M_USE_RESERVE is specified.
4068 KEG_UNLOCK(keg, slab->us_domain);
4073 * If the zone is striped we pick a new slab for every
4074 * N allocations. Eliminating this conditional will
4075 * instead pick a new domain for each bucket rather
4076 * than stripe within each bucket. The current option
4077 * produces more fragmentation and requires more cpu
4078 * time but yields better distribution.
4080 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
4081 vm_ndomains > 1 && --stripe == 0)
4084 } while (slab->us_freecount != 0 && i < max);
4085 KEG_UNLOCK(keg, slab->us_domain);
4087 /* Don't block if we allocated any successfully. */
4096 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
4098 uint64_t old, new, total, max;
4101 * The hard case. We're going to sleep because there were existing
4102 * sleepers or because we ran out of items. This routine enforces
4103 * fairness by keeping fifo order.
4105 * First release our ill gotten gains and make some noise.
4108 zone_free_limit(zone, count);
4109 zone_log_warning(zone);
4110 zone_maxaction(zone);
4111 if (flags & M_NOWAIT)
4115 * We need to allocate an item or set ourself as a sleeper
4116 * while the sleepq lock is held to avoid wakeup races. This
4117 * is essentially a home rolled semaphore.
4119 sleepq_lock(&zone->uz_max_items);
4120 old = zone->uz_items;
4122 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
4123 /* Cache the max since we will evaluate twice. */
4124 max = zone->uz_max_items;
4125 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
4126 UZ_ITEMS_COUNT(old) >= max)
4127 new = old + UZ_ITEMS_SLEEPER;
4129 new = old + MIN(count, max - old);
4130 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
4132 /* We may have successfully allocated under the sleepq lock. */
4133 if (UZ_ITEMS_SLEEPERS(new) == 0) {
4134 sleepq_release(&zone->uz_max_items);
4139 * This is in a different cacheline from uz_items so that we
4140 * don't constantly invalidate the fastpath cacheline when we
4141 * adjust item counts. This could be limited to toggling on
4144 atomic_add_32(&zone->uz_sleepers, 1);
4145 atomic_add_64(&zone->uz_sleeps, 1);
4148 * We have added ourselves as a sleeper. The sleepq lock
4149 * protects us from wakeup races. Sleep now and then retry.
4151 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
4152 sleepq_wait(&zone->uz_max_items, PVM);
4155 * After wakeup, remove ourselves as a sleeper and try
4156 * again. We no longer have the sleepq lock for protection.
4158 * Subract ourselves as a sleeper while attempting to add
4161 atomic_subtract_32(&zone->uz_sleepers, 1);
4162 old = atomic_fetchadd_64(&zone->uz_items,
4163 -(UZ_ITEMS_SLEEPER - count));
4164 /* We're no longer a sleeper. */
4165 old -= UZ_ITEMS_SLEEPER;
4168 * If we're still at the limit, restart. Notably do not
4169 * block on other sleepers. Cache the max value to protect
4170 * against changes via sysctl.
4172 total = UZ_ITEMS_COUNT(old);
4173 max = zone->uz_max_items;
4176 /* Truncate if necessary, otherwise wake other sleepers. */
4177 if (total + count > max) {
4178 zone_free_limit(zone, total + count - max);
4179 count = max - total;
4180 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
4181 wakeup_one(&zone->uz_max_items);
4188 * Allocate 'count' items from our max_items limit. Returns the number
4189 * available. If M_NOWAIT is not specified it will sleep until at least
4190 * one item can be allocated.
4193 zone_alloc_limit(uma_zone_t zone, int count, int flags)
4198 max = zone->uz_max_items;
4202 * We expect normal allocations to succeed with a simple
4205 old = atomic_fetchadd_64(&zone->uz_items, count);
4206 if (__predict_true(old + count <= max))
4210 * If we had some items and no sleepers just return the
4211 * truncated value. We have to release the excess space
4212 * though because that may wake sleepers who weren't woken
4213 * because we were temporarily over the limit.
4216 zone_free_limit(zone, (old + count) - max);
4219 return (zone_alloc_limit_hard(zone, count, flags));
4223 * Free a number of items back to the limit.
4226 zone_free_limit(uma_zone_t zone, int count)
4233 * In the common case we either have no sleepers or
4234 * are still over the limit and can just return.
4236 old = atomic_fetchadd_64(&zone->uz_items, -count);
4237 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4238 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4242 * Moderate the rate of wakeups. Sleepers will continue
4243 * to generate wakeups if necessary.
4245 wakeup_one(&zone->uz_max_items);
4249 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4251 uma_bucket_t bucket;
4252 int error, maxbucket, cnt;
4254 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4257 /* Avoid allocs targeting empty domains. */
4258 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4259 domain = UMA_ANYDOMAIN;
4260 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4261 domain = UMA_ANYDOMAIN;
4263 if (zone->uz_max_items > 0)
4264 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4267 maxbucket = zone->uz_bucket_size;
4271 /* Don't wait for buckets, preserve caller's NOVM setting. */
4272 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4273 if (bucket == NULL) {
4278 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4279 MIN(maxbucket, bucket->ub_entries), domain, flags);
4282 * Initialize the memory if necessary.
4284 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4287 for (i = 0; i < bucket->ub_cnt; i++) {
4288 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
4289 error = zone->uz_init(bucket->ub_bucket[i],
4290 zone->uz_size, flags);
4291 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
4297 * If we couldn't initialize the whole bucket, put the
4298 * rest back onto the freelist.
4300 if (i != bucket->ub_cnt) {
4301 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4302 bucket->ub_cnt - i);
4304 bzero(&bucket->ub_bucket[i],
4305 sizeof(void *) * (bucket->ub_cnt - i));
4311 cnt = bucket->ub_cnt;
4312 if (bucket->ub_cnt == 0) {
4313 bucket_free(zone, bucket, udata);
4314 counter_u64_add(zone->uz_fails, 1);
4318 if (zone->uz_max_items > 0 && cnt < maxbucket)
4319 zone_free_limit(zone, maxbucket - cnt);
4325 * Allocates a single item from a zone.
4328 * zone The zone to alloc for.
4329 * udata The data to be passed to the constructor.
4330 * domain The domain to allocate from or UMA_ANYDOMAIN.
4331 * flags M_WAITOK, M_NOWAIT, M_ZERO.
4334 * NULL if there is no memory and M_NOWAIT is set
4335 * An item if successful
4339 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4343 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4344 counter_u64_add(zone->uz_fails, 1);
4348 /* Avoid allocs targeting empty domains. */
4349 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4350 domain = UMA_ANYDOMAIN;
4352 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4356 * We have to call both the zone's init (not the keg's init)
4357 * and the zone's ctor. This is because the item is going from
4358 * a keg slab directly to the user, and the user is expecting it
4359 * to be both zone-init'd as well as zone-ctor'd.
4361 if (zone->uz_init != NULL) {
4364 kasan_mark_item_valid(zone, item);
4365 error = zone->uz_init(item, zone->uz_size, flags);
4366 kasan_mark_item_invalid(zone, item);
4368 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4372 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4377 counter_u64_add(zone->uz_allocs, 1);
4378 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4379 zone->uz_name, zone);
4384 counter_u64_add(zone->uz_fails, 1);
4386 if (zone->uz_max_items > 0)
4387 zone_free_limit(zone, 1);
4388 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4389 zone->uz_name, zone);
4396 uma_zfree_smr(uma_zone_t zone, void *item)
4399 uma_cache_bucket_t bucket;
4405 CTR3(KTR_UMA, "uma_zfree_smr zone %s(%p) item %p",
4406 zone->uz_name, zone, item);
4408 #ifdef UMA_ZALLOC_DEBUG
4409 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4410 ("uma_zfree_smr: called with non-SMR zone."));
4411 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4412 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4413 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4416 cache = &zone->uz_cpu[curcpu];
4419 uz_flags = cache_uz_flags(cache);
4420 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4421 itemdomain = item_domain(item);
4425 cache = &zone->uz_cpu[curcpu];
4426 /* SMR Zones must free to the free bucket. */
4427 bucket = &cache->uc_freebucket;
4429 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4430 PCPU_GET(domain) != itemdomain) {
4431 bucket = &cache->uc_crossbucket;
4434 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4435 cache_bucket_push(cache, bucket, item);
4439 } while (cache_free(zone, cache, NULL, itemdomain));
4443 * If nothing else caught this, we'll just do an internal free.
4445 zone_free_item(zone, item, NULL, SKIP_NONE);
4450 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4453 uma_cache_bucket_t bucket;
4454 int itemdomain, uz_flags;
4456 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4457 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4459 CTR3(KTR_UMA, "uma_zfree_arg zone %s(%p) item %p",
4460 zone->uz_name, zone, item);
4462 #ifdef UMA_ZALLOC_DEBUG
4463 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4464 ("uma_zfree_arg: called with SMR zone."));
4465 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4468 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4473 * We are accessing the per-cpu cache without a critical section to
4474 * fetch size and flags. This is acceptable, if we are preempted we
4475 * will simply read another cpu's line.
4477 cache = &zone->uz_cpu[curcpu];
4478 uz_flags = cache_uz_flags(cache);
4479 if (UMA_ALWAYS_CTORDTOR ||
4480 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4481 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4484 * The race here is acceptable. If we miss it we'll just have to wait
4485 * a little longer for the limits to be reset.
4487 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4488 if (atomic_load_32(&zone->uz_sleepers) > 0)
4493 * If possible, free to the per-CPU cache. There are two
4494 * requirements for safe access to the per-CPU cache: (1) the thread
4495 * accessing the cache must not be preempted or yield during access,
4496 * and (2) the thread must not migrate CPUs without switching which
4497 * cache it accesses. We rely on a critical section to prevent
4498 * preemption and migration. We release the critical section in
4499 * order to acquire the zone mutex if we are unable to free to the
4500 * current cache; when we re-acquire the critical section, we must
4501 * detect and handle migration if it has occurred.
4505 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4506 itemdomain = item_domain(item);
4510 cache = &zone->uz_cpu[curcpu];
4512 * Try to free into the allocbucket first to give LIFO
4513 * ordering for cache-hot datastructures. Spill over
4514 * into the freebucket if necessary. Alloc will swap
4515 * them if one runs dry.
4517 bucket = &cache->uc_allocbucket;
4519 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4520 PCPU_GET(domain) != itemdomain) {
4521 bucket = &cache->uc_crossbucket;
4524 if (bucket->ucb_cnt == bucket->ucb_entries &&
4525 cache->uc_freebucket.ucb_cnt <
4526 cache->uc_freebucket.ucb_entries)
4527 cache_bucket_swap(&cache->uc_freebucket,
4528 &cache->uc_allocbucket);
4529 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4530 cache_bucket_push(cache, bucket, item);
4534 } while (cache_free(zone, cache, udata, itemdomain));
4538 * If nothing else caught this, we'll just do an internal free.
4541 zone_free_item(zone, item, udata, SKIP_DTOR);
4546 * sort crossdomain free buckets to domain correct buckets and cache
4550 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4552 struct uma_bucketlist emptybuckets, fullbuckets;
4553 uma_zone_domain_t zdom;
4560 "uma_zfree: zone %s(%p) draining cross bucket %p",
4561 zone->uz_name, zone, bucket);
4564 * It is possible for buckets to arrive here out of order so we fetch
4565 * the current smr seq rather than accepting the bucket's.
4567 seq = SMR_SEQ_INVALID;
4568 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4569 seq = smr_advance(zone->uz_smr);
4572 * To avoid having ndomain * ndomain buckets for sorting we have a
4573 * lock on the current crossfree bucket. A full matrix with
4574 * per-domain locking could be used if necessary.
4576 STAILQ_INIT(&emptybuckets);
4577 STAILQ_INIT(&fullbuckets);
4578 ZONE_CROSS_LOCK(zone);
4579 for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4580 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4581 domain = item_domain(item);
4582 zdom = ZDOM_GET(zone, domain);
4583 if (zdom->uzd_cross == NULL) {
4584 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4585 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4586 zdom->uzd_cross = b;
4589 * Avoid allocating a bucket with the cross lock
4590 * held, since allocation can trigger a
4591 * cross-domain free and bucket zones may
4592 * allocate from each other.
4594 ZONE_CROSS_UNLOCK(zone);
4595 b = bucket_alloc(zone, udata, M_NOWAIT);
4598 ZONE_CROSS_LOCK(zone);
4599 if (zdom->uzd_cross != NULL) {
4600 STAILQ_INSERT_HEAD(&emptybuckets, b,
4603 zdom->uzd_cross = b;
4607 b = zdom->uzd_cross;
4608 b->ub_bucket[b->ub_cnt++] = item;
4610 if (b->ub_cnt == b->ub_entries) {
4611 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4612 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4613 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4614 zdom->uzd_cross = b;
4617 ZONE_CROSS_UNLOCK(zone);
4619 if (bucket->ub_cnt == 0)
4620 bucket->ub_seq = SMR_SEQ_INVALID;
4621 bucket_free(zone, bucket, udata);
4623 while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4624 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4625 bucket_free(zone, b, udata);
4627 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4628 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4629 domain = item_domain(b->ub_bucket[0]);
4630 zone_put_bucket(zone, domain, b, udata, true);
4636 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4637 int itemdomain, bool ws)
4642 * Buckets coming from the wrong domain will be entirely for the
4643 * only other domain on two domain systems. In this case we can
4644 * simply cache them. Otherwise we need to sort them back to
4647 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4648 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4649 zone_free_cross(zone, bucket, udata);
4655 * Attempt to save the bucket in the zone's domain bucket cache.
4658 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4659 zone->uz_name, zone, bucket);
4660 /* ub_cnt is pointing to the last free item */
4661 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4662 itemdomain = zone_domain_lowest(zone, itemdomain);
4663 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4667 * Populate a free or cross bucket for the current cpu cache. Free any
4668 * existing full bucket either to the zone cache or back to the slab layer.
4670 * Enters and returns in a critical section. false return indicates that
4671 * we can not satisfy this free in the cache layer. true indicates that
4672 * the caller should retry.
4674 static __noinline bool
4675 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, int itemdomain)
4677 uma_cache_bucket_t cbucket;
4678 uma_bucket_t newbucket, bucket;
4680 CRITICAL_ASSERT(curthread);
4682 if (zone->uz_bucket_size == 0)
4685 cache = &zone->uz_cpu[curcpu];
4689 * FIRSTTOUCH domains need to free to the correct zdom. When
4690 * enabled this is the zdom of the item. The bucket is the
4691 * cross bucket if the current domain and itemdomain do not match.
4693 cbucket = &cache->uc_freebucket;
4695 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4696 if (PCPU_GET(domain) != itemdomain) {
4697 cbucket = &cache->uc_crossbucket;
4698 if (cbucket->ucb_cnt != 0)
4699 counter_u64_add(zone->uz_xdomain,
4704 bucket = cache_bucket_unload(cbucket);
4705 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4706 ("cache_free: Entered with non-full free bucket."));
4708 /* We are no longer associated with this CPU. */
4712 * Don't let SMR zones operate without a free bucket. Force
4713 * a synchronize and re-use this one. We will only degrade
4714 * to a synchronize every bucket_size items rather than every
4715 * item if we fail to allocate a bucket.
4717 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4719 bucket->ub_seq = smr_advance(zone->uz_smr);
4720 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4721 if (newbucket == NULL && bucket != NULL) {
4722 bucket_drain(zone, bucket);
4726 } else if (!bucketdisable)
4727 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4730 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4733 if ((bucket = newbucket) == NULL)
4735 cache = &zone->uz_cpu[curcpu];
4738 * Check to see if we should be populating the cross bucket. If it
4739 * is already populated we will fall through and attempt to populate
4742 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4743 if (PCPU_GET(domain) != itemdomain &&
4744 cache->uc_crossbucket.ucb_bucket == NULL) {
4745 cache_bucket_load_cross(cache, bucket);
4751 * We may have lost the race to fill the bucket or switched CPUs.
4753 if (cache->uc_freebucket.ucb_bucket != NULL) {
4755 bucket_free(zone, bucket, udata);
4758 cache_bucket_load_free(cache, bucket);
4764 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4771 KEG_LOCK_ASSERT(keg, slab->us_domain);
4773 /* Do we need to remove from any lists? */
4774 dom = &keg->uk_domain[slab->us_domain];
4775 if (slab->us_freecount + 1 == keg->uk_ipers) {
4776 LIST_REMOVE(slab, us_link);
4777 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4778 dom->ud_free_slabs++;
4779 } else if (slab->us_freecount == 0) {
4780 LIST_REMOVE(slab, us_link);
4781 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4784 /* Slab management. */
4785 freei = slab_item_index(slab, keg, item);
4786 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4787 slab->us_freecount++;
4789 /* Keg statistics. */
4790 dom->ud_free_items++;
4794 zone_release(void *arg, void **bucket, int cnt)
4807 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4808 lock = KEG_LOCK(keg, 0);
4809 for (i = 0; i < cnt; i++) {
4811 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4812 slab = vtoslab((vm_offset_t)item);
4814 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4815 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4816 slab = hash_sfind(&keg->uk_hash, mem);
4818 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4820 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4823 lock = KEG_LOCK(keg, slab->us_domain);
4825 slab_free_item(zone, slab, item);
4832 * Frees a single item to any zone.
4835 * zone The zone to free to
4836 * item The item we're freeing
4837 * udata User supplied data for the dtor
4838 * skip Skip dtors and finis
4840 static __noinline void
4841 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4845 * If a free is sent directly to an SMR zone we have to
4846 * synchronize immediately because the item can instantly
4847 * be reallocated. This should only happen in degenerate
4848 * cases when no memory is available for per-cpu caches.
4850 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4851 smr_synchronize(zone->uz_smr);
4853 item_dtor(zone, item, zone->uz_size, udata, skip);
4855 if (skip < SKIP_FINI && zone->uz_fini) {
4856 kasan_mark_item_valid(zone, item);
4857 zone->uz_fini(item, zone->uz_size);
4858 kasan_mark_item_invalid(zone, item);
4861 zone->uz_release(zone->uz_arg, &item, 1);
4863 if (skip & SKIP_CNT)
4866 counter_u64_add(zone->uz_frees, 1);
4868 if (zone->uz_max_items > 0)
4869 zone_free_limit(zone, 1);
4874 uma_zone_set_max(uma_zone_t zone, int nitems)
4878 * If the limit is small, we may need to constrain the maximum per-CPU
4879 * cache size, or disable caching entirely.
4881 uma_zone_set_maxcache(zone, nitems);
4884 * XXX This can misbehave if the zone has any allocations with
4885 * no limit and a limit is imposed. There is currently no
4886 * way to clear a limit.
4889 if (zone->uz_max_items == 0)
4890 ZONE_ASSERT_COLD(zone);
4891 zone->uz_max_items = nitems;
4892 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4893 zone_update_caches(zone);
4894 /* We may need to wake waiters. */
4895 wakeup(&zone->uz_max_items);
4903 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4905 int bpcpu, bpdom, bsize, nb;
4910 * Compute a lower bound on the number of items that may be cached in
4911 * the zone. Each CPU gets at least two buckets, and for cross-domain
4912 * frees we use an additional bucket per CPU and per domain. Select the
4913 * largest bucket size that does not exceed half of the requested limit,
4914 * with the left over space given to the full bucket cache.
4919 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4924 nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4925 bsize = nitems / nb / 2;
4926 if (bsize > BUCKET_MAX)
4928 else if (bsize == 0 && nitems / nb > 0)
4930 zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
4931 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4932 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4933 zone->uz_bucket_max = nitems - nb * bsize;
4939 uma_zone_get_max(uma_zone_t zone)
4943 nitems = atomic_load_64(&zone->uz_max_items);
4950 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4953 ZONE_ASSERT_COLD(zone);
4954 zone->uz_warning = warning;
4959 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4962 ZONE_ASSERT_COLD(zone);
4963 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4968 uma_zone_get_cur(uma_zone_t zone)
4974 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4975 nitems = counter_u64_fetch(zone->uz_allocs) -
4976 counter_u64_fetch(zone->uz_frees);
4978 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4979 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4981 return (nitems < 0 ? 0 : nitems);
4985 uma_zone_get_allocs(uma_zone_t zone)
4991 if (zone->uz_allocs != EARLY_COUNTER)
4992 nitems = counter_u64_fetch(zone->uz_allocs);
4994 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
5000 uma_zone_get_frees(uma_zone_t zone)
5006 if (zone->uz_frees != EARLY_COUNTER)
5007 nitems = counter_u64_fetch(zone->uz_frees);
5009 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
5015 /* Used only for KEG_ASSERT_COLD(). */
5017 uma_keg_get_allocs(uma_keg_t keg)
5023 LIST_FOREACH(z, &keg->uk_zones, uz_link)
5024 nitems += uma_zone_get_allocs(z);
5032 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
5037 KEG_ASSERT_COLD(keg);
5038 keg->uk_init = uminit;
5043 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
5048 KEG_ASSERT_COLD(keg);
5049 keg->uk_fini = fini;
5054 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
5057 ZONE_ASSERT_COLD(zone);
5058 zone->uz_init = zinit;
5063 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
5066 ZONE_ASSERT_COLD(zone);
5067 zone->uz_fini = zfini;
5072 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
5077 KEG_ASSERT_COLD(keg);
5078 keg->uk_freef = freef;
5083 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
5088 KEG_ASSERT_COLD(keg);
5089 keg->uk_allocf = allocf;
5094 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
5097 ZONE_ASSERT_COLD(zone);
5099 KASSERT(smr != NULL, ("Got NULL smr"));
5100 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
5101 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
5102 zone->uz_flags |= UMA_ZONE_SMR;
5104 zone_update_caches(zone);
5108 uma_zone_get_smr(uma_zone_t zone)
5111 return (zone->uz_smr);
5116 uma_zone_reserve(uma_zone_t zone, int items)
5121 KEG_ASSERT_COLD(keg);
5122 keg->uk_reserve = items;
5127 uma_zone_reserve_kva(uma_zone_t zone, int count)
5134 KEG_ASSERT_COLD(keg);
5135 ZONE_ASSERT_COLD(zone);
5137 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
5139 #ifdef UMA_MD_SMALL_ALLOC
5140 if (keg->uk_ppera > 1) {
5144 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
5150 MPASS(keg->uk_kva == 0);
5153 zone->uz_max_items = pages * keg->uk_ipers;
5154 #ifdef UMA_MD_SMALL_ALLOC
5155 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
5157 keg->uk_allocf = noobj_alloc;
5159 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5160 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5161 zone_update_caches(zone);
5168 uma_prealloc(uma_zone_t zone, int items)
5170 struct vm_domainset_iter di;
5174 int aflags, domain, slabs;
5177 slabs = howmany(items, keg->uk_ipers);
5178 while (slabs-- > 0) {
5180 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
5183 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
5186 dom = &keg->uk_domain[slab->us_domain];
5188 * keg_alloc_slab() always returns a slab on the
5191 LIST_REMOVE(slab, us_link);
5192 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
5194 dom->ud_free_slabs++;
5195 KEG_UNLOCK(keg, slab->us_domain);
5198 if (vm_domainset_iter_policy(&di, &domain) != 0)
5199 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
5205 * Returns a snapshot of memory consumption in bytes.
5208 uma_zone_memory(uma_zone_t zone)
5214 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
5215 for (i = 0; i < vm_ndomains; i++)
5216 sz += ZDOM_GET(zone, i)->uzd_nitems;
5217 return (sz * zone->uz_size);
5219 for (i = 0; i < vm_ndomains; i++)
5220 sz += zone->uz_keg->uk_domain[i].ud_pages;
5222 return (sz * PAGE_SIZE);
5225 struct uma_reclaim_args {
5231 uma_reclaim_domain_cb(uma_zone_t zone, void *arg)
5233 struct uma_reclaim_args *args;
5236 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0)
5237 uma_zone_reclaim_domain(zone, args->req, args->domain);
5242 uma_reclaim(int req)
5244 uma_reclaim_domain(req, UMA_ANYDOMAIN);
5248 uma_reclaim_domain(int req, int domain)
5250 struct uma_reclaim_args args;
5254 args.domain = domain;
5257 sx_slock(&uma_reclaim_lock);
5259 case UMA_RECLAIM_TRIM:
5260 case UMA_RECLAIM_DRAIN:
5261 zone_foreach(uma_reclaim_domain_cb, &args);
5263 case UMA_RECLAIM_DRAIN_CPU:
5264 zone_foreach(uma_reclaim_domain_cb, &args);
5265 pcpu_cache_drain_safe(NULL);
5266 zone_foreach(uma_reclaim_domain_cb, &args);
5269 panic("unhandled reclamation request %d", req);
5273 * Some slabs may have been freed but this zone will be visited early
5274 * we visit again so that we can free pages that are empty once other
5275 * zones are drained. We have to do the same for buckets.
5277 uma_zone_reclaim_domain(slabzones[0], UMA_RECLAIM_DRAIN, domain);
5278 uma_zone_reclaim_domain(slabzones[1], UMA_RECLAIM_DRAIN, domain);
5279 bucket_zone_drain(domain);
5280 sx_sunlock(&uma_reclaim_lock);
5283 static volatile int uma_reclaim_needed;
5286 uma_reclaim_wakeup(void)
5289 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5290 wakeup(uma_reclaim);
5294 uma_reclaim_worker(void *arg __unused)
5298 sx_xlock(&uma_reclaim_lock);
5299 while (atomic_load_int(&uma_reclaim_needed) == 0)
5300 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5302 sx_xunlock(&uma_reclaim_lock);
5303 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5304 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5305 atomic_store_int(&uma_reclaim_needed, 0);
5306 /* Don't fire more than once per-second. */
5307 pause("umarclslp", hz);
5313 uma_zone_reclaim(uma_zone_t zone, int req)
5315 uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5319 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5322 case UMA_RECLAIM_TRIM:
5323 zone_reclaim(zone, domain, M_NOWAIT, false);
5325 case UMA_RECLAIM_DRAIN:
5326 zone_reclaim(zone, domain, M_NOWAIT, true);
5328 case UMA_RECLAIM_DRAIN_CPU:
5329 pcpu_cache_drain_safe(zone);
5330 zone_reclaim(zone, domain, M_NOWAIT, true);
5333 panic("unhandled reclamation request %d", req);
5339 uma_zone_exhausted(uma_zone_t zone)
5342 return (atomic_load_32(&zone->uz_sleepers) > 0);
5349 return (uma_kmem_limit);
5353 uma_set_limit(unsigned long limit)
5356 uma_kmem_limit = limit;
5363 return (atomic_load_long(&uma_kmem_total));
5370 return (uma_kmem_limit - uma_size());
5375 * Generate statistics across both the zone and its per-cpu cache's. Return
5376 * desired statistics if the pointer is non-NULL for that statistic.
5378 * Note: does not update the zone statistics, as it can't safely clear the
5379 * per-CPU cache statistic.
5383 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5384 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5387 uint64_t allocs, frees, sleeps, xdomain;
5390 allocs = frees = sleeps = xdomain = 0;
5393 cache = &z->uz_cpu[cpu];
5394 cachefree += cache->uc_allocbucket.ucb_cnt;
5395 cachefree += cache->uc_freebucket.ucb_cnt;
5396 xdomain += cache->uc_crossbucket.ucb_cnt;
5397 cachefree += cache->uc_crossbucket.ucb_cnt;
5398 allocs += cache->uc_allocs;
5399 frees += cache->uc_frees;
5401 allocs += counter_u64_fetch(z->uz_allocs);
5402 frees += counter_u64_fetch(z->uz_frees);
5403 xdomain += counter_u64_fetch(z->uz_xdomain);
5404 sleeps += z->uz_sleeps;
5405 if (cachefreep != NULL)
5406 *cachefreep = cachefree;
5407 if (allocsp != NULL)
5411 if (sleepsp != NULL)
5413 if (xdomainp != NULL)
5414 *xdomainp = xdomain;
5419 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5426 rw_rlock(&uma_rwlock);
5427 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5428 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5431 LIST_FOREACH(z, &uma_cachezones, uz_link)
5434 rw_runlock(&uma_rwlock);
5435 return (sysctl_handle_int(oidp, &count, 0, req));
5439 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5440 struct uma_percpu_stat *ups, bool internal)
5442 uma_zone_domain_t zdom;
5446 for (i = 0; i < vm_ndomains; i++) {
5447 zdom = ZDOM_GET(z, i);
5448 uth->uth_zone_free += zdom->uzd_nitems;
5450 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5451 uth->uth_frees = counter_u64_fetch(z->uz_frees);
5452 uth->uth_fails = counter_u64_fetch(z->uz_fails);
5453 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5454 uth->uth_sleeps = z->uz_sleeps;
5456 for (i = 0; i < mp_maxid + 1; i++) {
5457 bzero(&ups[i], sizeof(*ups));
5458 if (internal || CPU_ABSENT(i))
5460 cache = &z->uz_cpu[i];
5461 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5462 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5463 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5464 ups[i].ups_allocs = cache->uc_allocs;
5465 ups[i].ups_frees = cache->uc_frees;
5470 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5472 struct uma_stream_header ush;
5473 struct uma_type_header uth;
5474 struct uma_percpu_stat *ups;
5479 uint32_t kfree, pages;
5480 int count, error, i;
5482 error = sysctl_wire_old_buffer(req, 0);
5485 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5486 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5487 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5490 rw_rlock(&uma_rwlock);
5491 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5492 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5496 LIST_FOREACH(z, &uma_cachezones, uz_link)
5500 * Insert stream header.
5502 bzero(&ush, sizeof(ush));
5503 ush.ush_version = UMA_STREAM_VERSION;
5504 ush.ush_maxcpus = (mp_maxid + 1);
5505 ush.ush_count = count;
5506 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5508 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5510 for (i = 0; i < vm_ndomains; i++) {
5511 kfree += kz->uk_domain[i].ud_free_items;
5512 pages += kz->uk_domain[i].ud_pages;
5514 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5515 bzero(&uth, sizeof(uth));
5516 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5517 uth.uth_align = kz->uk_align;
5518 uth.uth_size = kz->uk_size;
5519 uth.uth_rsize = kz->uk_rsize;
5520 if (z->uz_max_items > 0) {
5521 items = UZ_ITEMS_COUNT(z->uz_items);
5522 uth.uth_pages = (items / kz->uk_ipers) *
5525 uth.uth_pages = pages;
5526 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5528 uth.uth_limit = z->uz_max_items;
5529 uth.uth_keg_free = kfree;
5532 * A zone is secondary is it is not the first entry
5533 * on the keg's zone list.
5535 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5536 (LIST_FIRST(&kz->uk_zones) != z))
5537 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5538 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5539 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5540 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5541 for (i = 0; i < mp_maxid + 1; i++)
5542 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5545 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5546 bzero(&uth, sizeof(uth));
5547 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5548 uth.uth_size = z->uz_size;
5549 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5550 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5551 for (i = 0; i < mp_maxid + 1; i++)
5552 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5555 rw_runlock(&uma_rwlock);
5556 error = sbuf_finish(&sbuf);
5563 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5565 uma_zone_t zone = *(uma_zone_t *)arg1;
5568 max = uma_zone_get_max(zone);
5569 error = sysctl_handle_int(oidp, &max, 0, req);
5570 if (error || !req->newptr)
5573 uma_zone_set_max(zone, max);
5579 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5585 * Some callers want to add sysctls for global zones that
5586 * may not yet exist so they pass a pointer to a pointer.
5589 zone = *(uma_zone_t *)arg1;
5592 cur = uma_zone_get_cur(zone);
5593 return (sysctl_handle_int(oidp, &cur, 0, req));
5597 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5599 uma_zone_t zone = arg1;
5602 cur = uma_zone_get_allocs(zone);
5603 return (sysctl_handle_64(oidp, &cur, 0, req));
5607 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5609 uma_zone_t zone = arg1;
5612 cur = uma_zone_get_frees(zone);
5613 return (sysctl_handle_64(oidp, &cur, 0, req));
5617 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5620 uma_zone_t zone = arg1;
5623 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5624 if (zone->uz_flags != 0)
5625 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5627 sbuf_printf(&sbuf, "0");
5628 error = sbuf_finish(&sbuf);
5635 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5637 uma_keg_t keg = arg1;
5638 int avail, effpct, total;
5640 total = keg->uk_ppera * PAGE_SIZE;
5641 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5642 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5644 * We consider the client's requested size and alignment here, not the
5645 * real size determination uk_rsize, because we also adjust the real
5646 * size for internal implementation reasons (max bitset size).
5648 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5649 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5650 avail *= mp_maxid + 1;
5651 effpct = 100 * avail / total;
5652 return (sysctl_handle_int(oidp, &effpct, 0, req));
5656 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5658 uma_zone_t zone = arg1;
5661 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5662 return (sysctl_handle_64(oidp, &cur, 0, req));
5667 uma_dbg_getslab(uma_zone_t zone, void *item)
5674 * It is safe to return the slab here even though the
5675 * zone is unlocked because the item's allocation state
5676 * essentially holds a reference.
5678 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5679 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5681 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5682 return (vtoslab((vm_offset_t)mem));
5684 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5685 return ((uma_slab_t)(mem + keg->uk_pgoff));
5687 slab = hash_sfind(&keg->uk_hash, mem);
5694 uma_dbg_zskip(uma_zone_t zone, void *mem)
5697 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5700 return (uma_dbg_kskip(zone->uz_keg, mem));
5704 uma_dbg_kskip(uma_keg_t keg, void *mem)
5708 if (dbg_divisor == 0)
5711 if (dbg_divisor == 1)
5714 idx = (uintptr_t)mem >> PAGE_SHIFT;
5715 if (keg->uk_ipers > 1) {
5716 idx *= keg->uk_ipers;
5717 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5720 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5721 counter_u64_add(uma_skip_cnt, 1);
5724 counter_u64_add(uma_dbg_cnt, 1);
5730 * Set up the slab's freei data such that uma_dbg_free can function.
5734 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5740 slab = uma_dbg_getslab(zone, item);
5742 panic("uma: item %p did not belong to zone %s",
5743 item, zone->uz_name);
5746 freei = slab_item_index(slab, keg, item);
5748 if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5749 slab_dbg_bits(slab, keg)))
5750 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5751 item, zone, zone->uz_name, slab, freei);
5755 * Verifies freed addresses. Checks for alignment, valid slab membership
5756 * and duplicate frees.
5760 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5766 slab = uma_dbg_getslab(zone, item);
5768 panic("uma: Freed item %p did not belong to zone %s",
5769 item, zone->uz_name);
5772 freei = slab_item_index(slab, keg, item);
5774 if (freei >= keg->uk_ipers)
5775 panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5776 item, zone, zone->uz_name, slab, freei);
5778 if (slab_item(slab, keg, freei) != item)
5779 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5780 item, zone, zone->uz_name, slab, freei);
5782 if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5783 slab_dbg_bits(slab, keg)))
5784 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5785 item, zone, zone->uz_name, slab, freei);
5787 #endif /* INVARIANTS */
5791 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5792 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5797 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5798 *allocs = counter_u64_fetch(z->uz_allocs);
5799 frees = counter_u64_fetch(z->uz_frees);
5800 *sleeps = z->uz_sleeps;
5804 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5806 for (i = 0; i < vm_ndomains; i++) {
5807 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5808 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5809 (LIST_FIRST(&kz->uk_zones) != z)))
5810 *cachefree += kz->uk_domain[i].ud_free_items;
5812 *used = *allocs - frees;
5813 return (((int64_t)*used + *cachefree) * kz->uk_size);
5816 DB_SHOW_COMMAND(uma, db_show_uma)
5818 const char *fmt_hdr, *fmt_entry;
5821 uint64_t allocs, used, sleeps, xdomain;
5823 /* variables for sorting */
5825 uma_zone_t cur_zone, last_zone;
5826 int64_t cur_size, last_size, size;
5829 /* /i option produces machine-parseable CSV output */
5830 if (modif[0] == 'i') {
5831 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5832 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5834 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5835 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5838 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5839 "Sleeps", "Bucket", "Total Mem", "XFree");
5841 /* Sort the zones with largest size first. */
5843 last_size = INT64_MAX;
5848 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5849 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5851 * In the case of size ties, print out zones
5852 * in the order they are encountered. That is,
5853 * when we encounter the most recently output
5854 * zone, we have already printed all preceding
5855 * ties, and we must print all following ties.
5857 if (z == last_zone) {
5861 size = get_uma_stats(kz, z, &allocs, &used,
5862 &sleeps, &cachefree, &xdomain);
5863 if (size > cur_size && size < last_size + ties)
5871 if (cur_zone == NULL)
5874 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5875 &sleeps, &cachefree, &xdomain);
5876 db_printf(fmt_entry, cur_zone->uz_name,
5877 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5878 (uintmax_t)allocs, (uintmax_t)sleeps,
5879 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5884 last_zone = cur_zone;
5885 last_size = cur_size;
5889 DB_SHOW_COMMAND(umacache, db_show_umacache)
5892 uint64_t allocs, frees;
5896 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5897 "Requests", "Bucket");
5898 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5899 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5900 for (i = 0; i < vm_ndomains; i++)
5901 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5902 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5903 z->uz_name, (uintmax_t)z->uz_size,
5904 (intmax_t)(allocs - frees), cachefree,
5905 (uintmax_t)allocs, z->uz_bucket_size);