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 /* Point the slab into the allocated memory */
1809 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1810 slab = (uma_slab_t)(mem + keg->uk_pgoff);
1812 slab_tohashslab(slab)->uhs_data = mem;
1814 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1815 for (i = 0; i < keg->uk_ppera; i++)
1816 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1819 slab->us_freecount = keg->uk_ipers;
1820 slab->us_flags = sflags;
1821 slab->us_domain = domain;
1823 BIT_FILL(keg->uk_ipers, &slab->us_free);
1825 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1828 if (keg->uk_init != NULL) {
1829 for (i = 0; i < keg->uk_ipers; i++)
1830 if (keg->uk_init(slab_item(slab, keg, i),
1831 keg->uk_size, flags) != 0)
1833 if (i != keg->uk_ipers) {
1834 keg_free_slab(keg, slab, i);
1838 kasan_mark_slab_invalid(keg, mem);
1839 KEG_LOCK(keg, domain);
1841 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1842 slab, keg->uk_name, keg);
1844 if (keg->uk_flags & UMA_ZFLAG_HASH)
1845 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1848 * If we got a slab here it's safe to mark it partially used
1849 * and return. We assume that the caller is going to remove
1850 * at least one item.
1852 dom = &keg->uk_domain[domain];
1853 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1854 dom->ud_pages += keg->uk_ppera;
1855 dom->ud_free_items += keg->uk_ipers;
1864 * This function is intended to be used early on in place of page_alloc(). It
1865 * performs contiguous physical memory allocations and uses a bump allocator for
1866 * KVA, so is usable before the kernel map is initialized.
1869 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1876 pages = howmany(bytes, PAGE_SIZE);
1877 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1879 *pflag = UMA_SLAB_BOOT;
1880 m = vm_page_alloc_noobj_contig_domain(domain, malloc2vm_flags(wait) |
1881 VM_ALLOC_WIRED, pages, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
1882 VM_MEMATTR_DEFAULT);
1886 pa = VM_PAGE_TO_PHYS(m);
1887 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1888 #if defined(__aarch64__) || defined(__amd64__) || \
1889 defined(__riscv) || defined(__powerpc64__)
1890 if ((wait & M_NODUMP) == 0)
1895 /* Allocate KVA and indirectly advance bootmem. */
1896 return ((void *)pmap_map(&bootmem, m->phys_addr,
1897 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE));
1901 startup_free(void *mem, vm_size_t bytes)
1906 va = (vm_offset_t)mem;
1907 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1910 * startup_alloc() returns direct-mapped slabs on some platforms. Avoid
1911 * unmapping ranges of the direct map.
1913 if (va >= bootstart && va + bytes <= bootmem)
1914 pmap_remove(kernel_pmap, va, va + bytes);
1915 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1916 #if defined(__aarch64__) || defined(__amd64__) || \
1917 defined(__riscv) || defined(__powerpc64__)
1918 dump_drop_page(VM_PAGE_TO_PHYS(m));
1920 vm_page_unwire_noq(m);
1926 * Allocates a number of pages from the system
1929 * bytes The number of bytes requested
1930 * wait Shall we wait?
1933 * A pointer to the alloced memory or possibly
1934 * NULL if M_NOWAIT is set.
1937 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1940 void *p; /* Returned page */
1942 *pflag = UMA_SLAB_KERNEL;
1943 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1949 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1952 struct pglist alloctail;
1953 vm_offset_t addr, zkva;
1955 vm_page_t p, p_next;
1960 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1962 TAILQ_INIT(&alloctail);
1963 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | malloc2vm_flags(wait);
1964 *pflag = UMA_SLAB_KERNEL;
1965 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1966 if (CPU_ABSENT(cpu)) {
1967 p = vm_page_alloc_noobj(flags);
1970 p = vm_page_alloc_noobj(flags);
1972 pc = pcpu_find(cpu);
1973 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1976 p = vm_page_alloc_noobj_domain(pc->pc_domain,
1978 if (__predict_false(p == NULL))
1979 p = vm_page_alloc_noobj(flags);
1982 if (__predict_false(p == NULL))
1984 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1986 if ((addr = kva_alloc(bytes)) == 0)
1989 TAILQ_FOREACH(p, &alloctail, listq) {
1990 pmap_qenter(zkva, &p, 1);
1993 return ((void*)addr);
1995 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1996 vm_page_unwire_noq(p);
2003 * Allocates a number of pages not belonging to a VM object
2006 * bytes The number of bytes requested
2007 * wait Shall we wait?
2010 * A pointer to the alloced memory or possibly
2011 * NULL if M_NOWAIT is set.
2014 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
2017 TAILQ_HEAD(, vm_page) alloctail;
2019 vm_offset_t retkva, zkva;
2020 vm_page_t p, p_next;
2024 TAILQ_INIT(&alloctail);
2026 req = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
2027 if ((wait & M_WAITOK) != 0)
2028 req |= VM_ALLOC_WAITOK;
2030 npages = howmany(bytes, PAGE_SIZE);
2031 while (npages > 0) {
2032 p = vm_page_alloc_noobj_domain(domain, req);
2035 * Since the page does not belong to an object, its
2038 TAILQ_INSERT_TAIL(&alloctail, p, listq);
2043 * Page allocation failed, free intermediate pages and
2046 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2047 vm_page_unwire_noq(p);
2052 *flags = UMA_SLAB_PRIV;
2053 zkva = keg->uk_kva +
2054 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
2056 TAILQ_FOREACH(p, &alloctail, listq) {
2057 pmap_qenter(zkva, &p, 1);
2061 return ((void *)retkva);
2065 * Allocate physically contiguous pages.
2068 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
2072 *pflag = UMA_SLAB_KERNEL;
2073 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
2074 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
2078 * Frees a number of pages to the system
2081 * mem A pointer to the memory to be freed
2082 * size The size of the memory being freed
2083 * flags The original p->us_flags field
2089 page_free(void *mem, vm_size_t size, uint8_t flags)
2092 if ((flags & UMA_SLAB_BOOT) != 0) {
2093 startup_free(mem, size);
2097 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
2098 ("UMA: page_free used with invalid flags %x", flags));
2100 kmem_free((vm_offset_t)mem, size);
2104 * Frees pcpu zone allocations
2107 * mem A pointer to the memory to be freed
2108 * size The size of the memory being freed
2109 * flags The original p->us_flags field
2115 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
2117 vm_offset_t sva, curva;
2121 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
2123 if ((flags & UMA_SLAB_BOOT) != 0) {
2124 startup_free(mem, size);
2128 sva = (vm_offset_t)mem;
2129 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
2130 paddr = pmap_kextract(curva);
2131 m = PHYS_TO_VM_PAGE(paddr);
2132 vm_page_unwire_noq(m);
2135 pmap_qremove(sva, size >> PAGE_SHIFT);
2136 kva_free(sva, size);
2140 * Zero fill initializer
2142 * Arguments/Returns follow uma_init specifications
2145 zero_init(void *mem, int size, int flags)
2152 static struct noslabbits *
2153 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2156 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2161 * Actual size of embedded struct slab (!OFFPAGE).
2164 slab_sizeof(int nitems)
2168 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2169 return (roundup(s, UMA_ALIGN_PTR + 1));
2172 #define UMA_FIXPT_SHIFT 31
2173 #define UMA_FRAC_FIXPT(n, d) \
2174 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2175 #define UMA_FIXPT_PCT(f) \
2176 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2177 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
2178 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2181 * Compute the number of items that will fit in a slab. If hdr is true, the
2182 * item count may be limited to provide space in the slab for an inline slab
2183 * header. Otherwise, all slab space will be provided for item storage.
2186 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2191 /* The padding between items is not needed after the last item. */
2192 padpi = rsize - size;
2196 * Start with the maximum item count and remove items until
2197 * the slab header first alongside the allocatable memory.
2199 for (ipers = MIN(SLAB_MAX_SETSIZE,
2200 (slabsize + padpi - slab_sizeof(1)) / rsize);
2202 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2206 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2212 struct keg_layout_result {
2220 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2221 struct keg_layout_result *kl)
2226 kl->slabsize = slabsize;
2228 /* Handle INTERNAL as inline with an extra page. */
2229 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2230 kl->format &= ~UMA_ZFLAG_INTERNAL;
2231 kl->slabsize += PAGE_SIZE;
2234 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2235 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2237 /* Account for memory used by an offpage slab header. */
2238 total = kl->slabsize;
2239 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2240 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2242 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2246 * Determine the format of a uma keg. This determines where the slab header
2247 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2250 * keg The zone we should initialize
2256 keg_layout(uma_keg_t keg)
2258 struct keg_layout_result kl = {}, kl_tmp;
2267 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2268 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2269 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2270 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2271 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2273 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2274 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2275 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2278 alignsize = keg->uk_align + 1;
2281 * ASAN requires that each allocation be aligned to the shadow map
2284 if (alignsize < KASAN_SHADOW_SCALE)
2285 alignsize = KASAN_SHADOW_SCALE;
2289 * Calculate the size of each allocation (rsize) according to
2290 * alignment. If the requested size is smaller than we have
2291 * allocation bits for we round it up.
2293 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2294 rsize = roundup2(rsize, alignsize);
2296 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2298 * We want one item to start on every align boundary in a page.
2299 * To do this we will span pages. We will also extend the item
2300 * by the size of align if it is an even multiple of align.
2301 * Otherwise, it would fall on the same boundary every time.
2303 if ((rsize & alignsize) == 0)
2305 slabsize = rsize * (PAGE_SIZE / alignsize);
2306 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2307 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2308 slabsize = round_page(slabsize);
2311 * Start with a slab size of as many pages as it takes to
2312 * represent a single item. We will try to fit as many
2313 * additional items into the slab as possible.
2315 slabsize = round_page(keg->uk_size);
2318 /* Build a list of all of the available formats for this keg. */
2321 /* Evaluate an inline slab layout. */
2322 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2325 /* TODO: vm_page-embedded slab. */
2328 * We can't do OFFPAGE if we're internal or if we've been
2329 * asked to not go to the VM for buckets. If we do this we
2330 * may end up going to the VM for slabs which we do not want
2331 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2332 * In those cases, evaluate a pseudo-format called INTERNAL
2333 * which has an inline slab header and one extra page to
2334 * guarantee that it fits.
2336 * Otherwise, see if using an OFFPAGE slab will improve our
2339 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2340 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2342 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2345 * Choose a slab size and format which satisfy the minimum efficiency.
2346 * Prefer the smallest slab size that meets the constraints.
2348 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2349 * for small items (up to PAGE_SIZE), the iteration increment is one
2350 * page; and for large items, the increment is one item.
2352 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2353 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2354 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2357 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2358 round_page(rsize * (i - 1) + keg->uk_size);
2360 for (j = 0; j < nfmt; j++) {
2361 /* Only if we have no viable format yet. */
2362 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2366 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2367 if (kl_tmp.eff <= kl.eff)
2372 CTR6(KTR_UMA, "keg %s layout: format %#x "
2373 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2374 keg->uk_name, kl.format, kl.ipers, rsize,
2375 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2377 /* Stop when we reach the minimum efficiency. */
2378 if (kl.eff >= UMA_MIN_EFF)
2382 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2383 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2384 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2388 pages = atop(kl.slabsize);
2389 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2390 pages *= mp_maxid + 1;
2392 keg->uk_rsize = rsize;
2393 keg->uk_ipers = kl.ipers;
2394 keg->uk_ppera = pages;
2395 keg->uk_flags |= kl.format;
2398 * How do we find the slab header if it is offpage or if not all item
2399 * start addresses are in the same page? We could solve the latter
2400 * case with vaddr alignment, but we don't.
2402 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2403 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2404 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2405 keg->uk_flags |= UMA_ZFLAG_HASH;
2407 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2410 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2411 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2413 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2414 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2415 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2416 keg->uk_ipers, pages));
2420 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2421 * the keg onto the global keg list.
2423 * Arguments/Returns follow uma_ctor specifications
2424 * udata Actually uma_kctor_args
2427 keg_ctor(void *mem, int size, void *udata, int flags)
2429 struct uma_kctor_args *arg = udata;
2430 uma_keg_t keg = mem;
2435 keg->uk_size = arg->size;
2436 keg->uk_init = arg->uminit;
2437 keg->uk_fini = arg->fini;
2438 keg->uk_align = arg->align;
2439 keg->uk_reserve = 0;
2440 keg->uk_flags = arg->flags;
2443 * We use a global round-robin policy by default. Zones with
2444 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2445 * case the iterator is never run.
2447 keg->uk_dr.dr_policy = DOMAINSET_RR();
2448 keg->uk_dr.dr_iter = 0;
2451 * The primary zone is passed to us at keg-creation time.
2454 keg->uk_name = zone->uz_name;
2456 if (arg->flags & UMA_ZONE_ZINIT)
2457 keg->uk_init = zero_init;
2459 if (arg->flags & UMA_ZONE_MALLOC)
2460 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2463 keg->uk_flags &= ~UMA_ZONE_PCPU;
2469 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2470 * work on. Use round-robin for everything else.
2472 * Zones may override the default by specifying either.
2475 if ((keg->uk_flags &
2476 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2477 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2478 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2479 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2483 * If we haven't booted yet we need allocations to go through the
2484 * startup cache until the vm is ready.
2486 #ifdef UMA_MD_SMALL_ALLOC
2487 if (keg->uk_ppera == 1)
2488 keg->uk_allocf = uma_small_alloc;
2491 if (booted < BOOT_KVA)
2492 keg->uk_allocf = startup_alloc;
2493 else if (keg->uk_flags & UMA_ZONE_PCPU)
2494 keg->uk_allocf = pcpu_page_alloc;
2495 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2496 keg->uk_allocf = contig_alloc;
2498 keg->uk_allocf = page_alloc;
2499 #ifdef UMA_MD_SMALL_ALLOC
2500 if (keg->uk_ppera == 1)
2501 keg->uk_freef = uma_small_free;
2504 if (keg->uk_flags & UMA_ZONE_PCPU)
2505 keg->uk_freef = pcpu_page_free;
2507 keg->uk_freef = page_free;
2510 * Initialize keg's locks.
2512 for (i = 0; i < vm_ndomains; i++)
2513 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2516 * If we're putting the slab header in the actual page we need to
2517 * figure out where in each page it goes. See slab_sizeof
2520 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2523 shsize = slab_sizeof(keg->uk_ipers);
2524 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2526 * The only way the following is possible is if with our
2527 * UMA_ALIGN_PTR adjustments we are now bigger than
2528 * UMA_SLAB_SIZE. I haven't checked whether this is
2529 * mathematically possible for all cases, so we make
2532 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2533 ("zone %s ipers %d rsize %d size %d slab won't fit",
2534 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2537 if (keg->uk_flags & UMA_ZFLAG_HASH)
2538 hash_alloc(&keg->uk_hash, 0);
2540 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2542 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2544 rw_wlock(&uma_rwlock);
2545 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2546 rw_wunlock(&uma_rwlock);
2551 zone_kva_available(uma_zone_t zone, void *unused)
2555 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2559 if (keg->uk_allocf == startup_alloc) {
2560 /* Switch to the real allocator. */
2561 if (keg->uk_flags & UMA_ZONE_PCPU)
2562 keg->uk_allocf = pcpu_page_alloc;
2563 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2565 keg->uk_allocf = contig_alloc;
2567 keg->uk_allocf = page_alloc;
2572 zone_alloc_counters(uma_zone_t zone, void *unused)
2575 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2576 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2577 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2578 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2582 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2584 uma_zone_domain_t zdom;
2587 struct sysctl_oid *oid, *domainoid;
2588 int domains, i, cnt;
2589 static const char *nokeg = "cache zone";
2593 * Make a sysctl safe copy of the zone name by removing
2594 * any special characters and handling dups by appending
2597 if (zone->uz_namecnt != 0) {
2598 /* Count the number of decimal digits and '_' separator. */
2599 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2601 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2603 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2606 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2607 for (c = zone->uz_ctlname; *c != '\0'; c++)
2608 if (strchr("./\\ -", *c) != NULL)
2612 * Basic parameters at the root.
2614 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2615 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2617 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2618 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2619 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2620 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2621 zone, 0, sysctl_handle_uma_zone_flags, "A",
2622 "Allocator configuration flags");
2623 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2624 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2625 "Desired per-cpu cache size");
2626 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2627 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2628 "Maximum allowed per-cpu cache size");
2633 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2634 domains = vm_ndomains;
2637 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2638 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2640 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2641 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2642 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2643 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2644 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2645 "Real object size with alignment");
2646 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2647 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2648 "pages per-slab allocation");
2649 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2650 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2651 "items available per-slab");
2652 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2653 "align", CTLFLAG_RD, &keg->uk_align, 0,
2654 "item alignment mask");
2655 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2656 "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2657 "number of reserved items");
2658 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2659 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2660 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2661 "Slab utilization (100 - internal fragmentation %)");
2662 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2663 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2664 for (i = 0; i < domains; i++) {
2665 dom = &keg->uk_domain[i];
2666 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2667 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2668 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2669 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2670 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2671 "Total pages currently allocated from VM");
2672 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2673 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2674 "Items free in the slab layer");
2675 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2676 "free_slabs", CTLFLAG_RD, &dom->ud_free_slabs, 0,
2680 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2681 "name", CTLFLAG_RD, nokeg, "Keg name");
2684 * Information about zone limits.
2686 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2687 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2688 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2689 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2690 zone, 0, sysctl_handle_uma_zone_items, "QU",
2691 "Current number of allocated items if limit is set");
2692 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2693 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2694 "Maximum number of allocated and cached items");
2695 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2696 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2697 "Number of threads sleeping at limit");
2698 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2699 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2700 "Total zone limit sleeps");
2701 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2702 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2703 "Maximum number of items in each domain's bucket cache");
2706 * Per-domain zone information.
2708 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2709 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2710 for (i = 0; i < domains; i++) {
2711 zdom = ZDOM_GET(zone, i);
2712 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2713 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2714 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2715 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2716 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2717 "number of items in this domain");
2718 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2719 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2720 "maximum item count in this period");
2721 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2722 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2723 "minimum item count in this period");
2724 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2725 "bimin", CTLFLAG_RD, &zdom->uzd_bimin,
2726 "Minimum item count in this batch");
2727 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2728 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2729 "Working set size");
2730 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2731 "limin", CTLFLAG_RD, &zdom->uzd_limin,
2732 "Long time minimum item count");
2733 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2734 "timin", CTLFLAG_RD, &zdom->uzd_timin, 0,
2735 "Time since zero long time minimum item count");
2739 * General statistics.
2741 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2742 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2743 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2744 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2745 zone, 1, sysctl_handle_uma_zone_cur, "I",
2746 "Current number of allocated items");
2747 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2748 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2749 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2750 "Total allocation calls");
2751 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2752 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2753 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2754 "Total free calls");
2755 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2756 "fails", CTLFLAG_RD, &zone->uz_fails,
2757 "Number of allocation failures");
2758 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2759 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2760 "Free calls from the wrong domain");
2763 struct uma_zone_count {
2769 zone_count(uma_zone_t zone, void *arg)
2771 struct uma_zone_count *cnt;
2775 * Some zones are rapidly created with identical names and
2776 * destroyed out of order. This can lead to gaps in the count.
2777 * Use one greater than the maximum observed for this name.
2779 if (strcmp(zone->uz_name, cnt->name) == 0)
2780 cnt->count = MAX(cnt->count,
2781 zone->uz_namecnt + 1);
2785 zone_update_caches(uma_zone_t zone)
2789 for (i = 0; i <= mp_maxid; i++) {
2790 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2791 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2796 * Zone header ctor. This initializes all fields, locks, etc.
2798 * Arguments/Returns follow uma_ctor specifications
2799 * udata Actually uma_zctor_args
2802 zone_ctor(void *mem, int size, void *udata, int flags)
2804 struct uma_zone_count cnt;
2805 struct uma_zctor_args *arg = udata;
2806 uma_zone_domain_t zdom;
2807 uma_zone_t zone = mem;
2813 zone->uz_name = arg->name;
2814 zone->uz_ctor = arg->ctor;
2815 zone->uz_dtor = arg->dtor;
2816 zone->uz_init = NULL;
2817 zone->uz_fini = NULL;
2818 zone->uz_sleeps = 0;
2819 zone->uz_bucket_size = 0;
2820 zone->uz_bucket_size_min = 0;
2821 zone->uz_bucket_size_max = BUCKET_MAX;
2822 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2823 zone->uz_warning = NULL;
2824 /* The domain structures follow the cpu structures. */
2825 zone->uz_bucket_max = ULONG_MAX;
2826 timevalclear(&zone->uz_ratecheck);
2828 /* Count the number of duplicate names. */
2829 cnt.name = arg->name;
2831 zone_foreach(zone_count, &cnt);
2832 zone->uz_namecnt = cnt.count;
2833 ZONE_CROSS_LOCK_INIT(zone);
2835 for (i = 0; i < vm_ndomains; i++) {
2836 zdom = ZDOM_GET(zone, i);
2837 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2838 STAILQ_INIT(&zdom->uzd_buckets);
2841 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
2842 if (arg->uminit == trash_init && arg->fini == trash_fini)
2843 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2844 #elif defined(KASAN)
2845 if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0)
2846 arg->flags |= UMA_ZONE_NOKASAN;
2850 * This is a pure cache zone, no kegs.
2853 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2854 ("zone_ctor: Import specified for non-cache zone."));
2855 zone->uz_flags = arg->flags;
2856 zone->uz_size = arg->size;
2857 zone->uz_import = arg->import;
2858 zone->uz_release = arg->release;
2859 zone->uz_arg = arg->arg;
2862 * Cache zones are round-robin unless a policy is
2863 * specified because they may have incompatible
2866 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2867 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2869 rw_wlock(&uma_rwlock);
2870 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2871 rw_wunlock(&uma_rwlock);
2876 * Use the regular zone/keg/slab allocator.
2878 zone->uz_import = zone_import;
2879 zone->uz_release = zone_release;
2880 zone->uz_arg = zone;
2883 if (arg->flags & UMA_ZONE_SECONDARY) {
2884 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2885 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2886 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2887 zone->uz_init = arg->uminit;
2888 zone->uz_fini = arg->fini;
2889 zone->uz_flags |= UMA_ZONE_SECONDARY;
2890 rw_wlock(&uma_rwlock);
2892 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2893 if (LIST_NEXT(z, uz_link) == NULL) {
2894 LIST_INSERT_AFTER(z, zone, uz_link);
2899 rw_wunlock(&uma_rwlock);
2900 } else if (keg == NULL) {
2901 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2902 arg->align, arg->flags)) == NULL)
2905 struct uma_kctor_args karg;
2908 /* We should only be here from uma_startup() */
2909 karg.size = arg->size;
2910 karg.uminit = arg->uminit;
2911 karg.fini = arg->fini;
2912 karg.align = arg->align;
2913 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2915 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2921 /* Inherit properties from the keg. */
2923 zone->uz_size = keg->uk_size;
2924 zone->uz_flags |= (keg->uk_flags &
2925 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2928 if (booted >= BOOT_PCPU) {
2929 zone_alloc_counters(zone, NULL);
2930 if (booted >= BOOT_RUNNING)
2931 zone_alloc_sysctl(zone, NULL);
2933 zone->uz_allocs = EARLY_COUNTER;
2934 zone->uz_frees = EARLY_COUNTER;
2935 zone->uz_fails = EARLY_COUNTER;
2938 /* Caller requests a private SMR context. */
2939 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2940 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2942 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2943 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2944 ("Invalid zone flag combination"));
2945 if (arg->flags & UMA_ZFLAG_INTERNAL)
2946 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2947 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2948 zone->uz_bucket_size = BUCKET_MAX;
2949 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2950 zone->uz_bucket_size = 0;
2952 zone->uz_bucket_size = bucket_select(zone->uz_size);
2953 zone->uz_bucket_size_min = zone->uz_bucket_size;
2954 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2955 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2956 zone_update_caches(zone);
2962 * Keg header dtor. This frees all data, destroys locks, frees the hash
2963 * table and removes the keg from the global list.
2965 * Arguments/Returns follow uma_dtor specifications
2969 keg_dtor(void *arg, int size, void *udata)
2972 uint32_t free, pages;
2975 keg = (uma_keg_t)arg;
2977 for (i = 0; i < vm_ndomains; i++) {
2978 free += keg->uk_domain[i].ud_free_items;
2979 pages += keg->uk_domain[i].ud_pages;
2980 KEG_LOCK_FINI(keg, i);
2983 printf("Freed UMA keg (%s) was not empty (%u items). "
2984 " Lost %u pages of memory.\n",
2985 keg->uk_name ? keg->uk_name : "",
2986 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2988 hash_free(&keg->uk_hash);
2994 * Arguments/Returns follow uma_dtor specifications
2998 zone_dtor(void *arg, int size, void *udata)
3004 zone = (uma_zone_t)arg;
3006 sysctl_remove_oid(zone->uz_oid, 1, 1);
3008 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
3011 rw_wlock(&uma_rwlock);
3012 LIST_REMOVE(zone, uz_link);
3013 rw_wunlock(&uma_rwlock);
3014 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3016 keg->uk_reserve = 0;
3018 zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true);
3021 * We only destroy kegs from non secondary/non cache zones.
3023 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3025 rw_wlock(&uma_rwlock);
3026 LIST_REMOVE(keg, uk_link);
3027 rw_wunlock(&uma_rwlock);
3028 zone_free_item(kegs, keg, NULL, SKIP_NONE);
3030 counter_u64_free(zone->uz_allocs);
3031 counter_u64_free(zone->uz_frees);
3032 counter_u64_free(zone->uz_fails);
3033 counter_u64_free(zone->uz_xdomain);
3034 free(zone->uz_ctlname, M_UMA);
3035 for (i = 0; i < vm_ndomains; i++)
3036 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
3037 ZONE_CROSS_LOCK_FINI(zone);
3041 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3046 LIST_FOREACH(keg, &uma_kegs, uk_link) {
3047 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
3050 LIST_FOREACH(zone, &uma_cachezones, uz_link)
3055 * Traverses every zone in the system and calls a callback
3058 * zfunc A pointer to a function which accepts a zone
3065 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3068 rw_rlock(&uma_rwlock);
3069 zone_foreach_unlocked(zfunc, arg);
3070 rw_runlock(&uma_rwlock);
3074 * Initialize the kernel memory allocator. This is done after pages can be
3075 * allocated but before general KVA is available.
3078 uma_startup1(vm_offset_t virtual_avail)
3080 struct uma_zctor_args args;
3081 size_t ksize, zsize, size;
3082 uma_keg_t primarykeg;
3087 bootstart = bootmem = virtual_avail;
3089 rw_init(&uma_rwlock, "UMA lock");
3090 sx_init(&uma_reclaim_lock, "umareclaim");
3092 ksize = sizeof(struct uma_keg) +
3093 (sizeof(struct uma_domain) * vm_ndomains);
3094 ksize = roundup(ksize, UMA_SUPER_ALIGN);
3095 zsize = sizeof(struct uma_zone) +
3096 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
3097 (sizeof(struct uma_zone_domain) * vm_ndomains);
3098 zsize = roundup(zsize, UMA_SUPER_ALIGN);
3100 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
3101 size = (zsize * 2) + ksize;
3102 for (domain = 0; domain < vm_ndomains; domain++) {
3103 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
3108 zones = (uma_zone_t)m;
3110 kegs = (uma_zone_t)m;
3112 primarykeg = (uma_keg_t)m;
3114 /* "manually" create the initial zone */
3115 memset(&args, 0, sizeof(args));
3116 args.name = "UMA Kegs";
3118 args.ctor = keg_ctor;
3119 args.dtor = keg_dtor;
3120 args.uminit = zero_init;
3122 args.keg = primarykeg;
3123 args.align = UMA_SUPER_ALIGN - 1;
3124 args.flags = UMA_ZFLAG_INTERNAL;
3125 zone_ctor(kegs, zsize, &args, M_WAITOK);
3127 args.name = "UMA Zones";
3129 args.ctor = zone_ctor;
3130 args.dtor = zone_dtor;
3131 args.uminit = zero_init;
3134 args.align = UMA_SUPER_ALIGN - 1;
3135 args.flags = UMA_ZFLAG_INTERNAL;
3136 zone_ctor(zones, zsize, &args, M_WAITOK);
3138 /* Now make zones for slab headers */
3139 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
3140 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3141 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
3142 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3144 hashzone = uma_zcreate("UMA Hash",
3145 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
3146 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3152 #ifndef UMA_MD_SMALL_ALLOC
3153 extern void vm_radix_reserve_kva(void);
3157 * Advertise the availability of normal kva allocations and switch to
3158 * the default back-end allocator. Marks the KVA we consumed on startup
3159 * as used in the map.
3165 if (bootstart != bootmem) {
3166 vm_map_lock(kernel_map);
3167 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
3168 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
3169 vm_map_unlock(kernel_map);
3172 #ifndef UMA_MD_SMALL_ALLOC
3173 /* Set up radix zone to use noobj_alloc. */
3174 vm_radix_reserve_kva();
3178 zone_foreach_unlocked(zone_kva_available, NULL);
3183 * Allocate counters as early as possible so that boot-time allocations are
3184 * accounted more precisely.
3187 uma_startup_pcpu(void *arg __unused)
3190 zone_foreach_unlocked(zone_alloc_counters, NULL);
3193 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
3196 * Finish our initialization steps.
3199 uma_startup3(void *arg __unused)
3203 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
3204 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
3205 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
3207 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
3208 callout_init(&uma_callout, 1);
3209 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
3210 booted = BOOT_RUNNING;
3212 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
3213 EVENTHANDLER_PRI_FIRST);
3215 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
3221 booted = BOOT_SHUTDOWN;
3225 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3226 int align, uint32_t flags)
3228 struct uma_kctor_args args;
3231 args.uminit = uminit;
3233 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
3236 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3239 /* Public functions */
3242 uma_set_align(int align)
3245 if (align != UMA_ALIGN_CACHE)
3246 uma_align_cache = align;
3251 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3252 uma_init uminit, uma_fini fini, int align, uint32_t flags)
3255 struct uma_zctor_args args;
3258 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
3261 /* This stuff is essential for the zone ctor */
3262 memset(&args, 0, sizeof(args));
3267 args.uminit = uminit;
3269 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
3271 * Inject procedures which check for memory use after free if we are
3272 * allowed to scramble the memory while it is not allocated. This
3273 * requires that: UMA is actually able to access the memory, no init
3274 * or fini procedures, no dependency on the initial value of the
3275 * memory, and no (legitimate) use of the memory after free. Note,
3276 * the ctor and dtor do not need to be empty.
3278 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3279 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3280 args.uminit = trash_init;
3281 args.fini = trash_fini;
3288 sx_xlock(&uma_reclaim_lock);
3289 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3290 sx_xunlock(&uma_reclaim_lock);
3297 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3298 uma_init zinit, uma_fini zfini, uma_zone_t primary)
3300 struct uma_zctor_args args;
3304 keg = primary->uz_keg;
3305 memset(&args, 0, sizeof(args));
3307 args.size = keg->uk_size;
3310 args.uminit = zinit;
3312 args.align = keg->uk_align;
3313 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3316 sx_xlock(&uma_reclaim_lock);
3317 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3318 sx_xunlock(&uma_reclaim_lock);
3325 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3326 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3327 void *arg, int flags)
3329 struct uma_zctor_args args;
3331 memset(&args, 0, sizeof(args));
3336 args.uminit = zinit;
3338 args.import = zimport;
3339 args.release = zrelease;
3342 args.flags = flags | UMA_ZFLAG_CACHE;
3344 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3349 uma_zdestroy(uma_zone_t zone)
3353 * Large slabs are expensive to reclaim, so don't bother doing
3354 * unnecessary work if we're shutting down.
3356 if (booted == BOOT_SHUTDOWN &&
3357 zone->uz_fini == NULL && zone->uz_release == zone_release)
3359 sx_xlock(&uma_reclaim_lock);
3360 zone_free_item(zones, zone, NULL, SKIP_NONE);
3361 sx_xunlock(&uma_reclaim_lock);
3365 uma_zwait(uma_zone_t zone)
3368 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3369 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3370 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3371 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3373 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3377 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3379 void *item, *pcpu_item;
3383 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3385 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3388 pcpu_item = zpcpu_base_to_offset(item);
3389 if (flags & M_ZERO) {
3391 for (i = 0; i <= mp_maxid; i++)
3392 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3394 bzero(item, zone->uz_size);
3401 * A stub while both regular and pcpu cases are identical.
3404 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3409 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3412 /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */
3413 if (pcpu_item == NULL)
3416 item = zpcpu_offset_to_base(pcpu_item);
3417 uma_zfree_arg(zone, item, udata);
3420 static inline void *
3421 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3428 kasan_mark_item_valid(zone, item);
3429 kmsan_mark_item_uninitialized(zone, item);
3432 skipdbg = uma_dbg_zskip(zone, item);
3433 if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3434 zone->uz_ctor != trash_ctor)
3435 trash_ctor(item, size, udata, flags);
3438 /* Check flags before loading ctor pointer. */
3439 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3440 __predict_false(zone->uz_ctor != NULL) &&
3441 zone->uz_ctor(item, size, udata, flags) != 0) {
3442 counter_u64_add(zone->uz_fails, 1);
3443 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3448 uma_dbg_alloc(zone, NULL, item);
3450 if (__predict_false(flags & M_ZERO))
3451 return (memset(item, 0, size));
3457 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3458 enum zfreeskip skip)
3463 skipdbg = uma_dbg_zskip(zone, item);
3464 if (skip == SKIP_NONE && !skipdbg) {
3465 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3466 uma_dbg_free(zone, udata, item);
3468 uma_dbg_free(zone, NULL, item);
3471 if (__predict_true(skip < SKIP_DTOR)) {
3472 if (zone->uz_dtor != NULL)
3473 zone->uz_dtor(item, size, udata);
3475 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3476 zone->uz_dtor != trash_dtor)
3477 trash_dtor(item, size, udata);
3480 kasan_mark_item_invalid(zone, item);
3485 item_domain(void *item)
3489 domain = vm_phys_domain(vtophys(item));
3490 KASSERT(domain >= 0 && domain < vm_ndomains,
3491 ("%s: unknown domain for item %p", __func__, item));
3496 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3497 #if defined(INVARIANTS) && (defined(DDB) || defined(STACK))
3498 #include <sys/stack.h>
3500 #define UMA_ZALLOC_DEBUG
3502 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3508 if (flags & M_WAITOK) {
3509 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3510 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3515 KASSERT((flags & M_EXEC) == 0,
3516 ("uma_zalloc_debug: called with M_EXEC"));
3517 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3518 ("uma_zalloc_debug: called within spinlock or critical section"));
3519 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3520 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3522 _Static_assert(M_NOWAIT != 0 && M_WAITOK != 0,
3523 "M_NOWAIT and M_WAITOK must be non-zero for this assertion:");
3526 * Give the #elif clause time to find problems, then remove it
3527 * and enable this. (Remove <sys/stack.h> above, too.)
3529 KASSERT((flags & (M_NOWAIT|M_WAITOK)) == M_NOWAIT ||
3530 (flags & (M_NOWAIT|M_WAITOK)) == M_WAITOK,
3531 ("uma_zalloc_debug: must pass one of M_NOWAIT or M_WAITOK"));
3532 #elif defined(DDB) || defined(STACK)
3533 if (__predict_false((flags & (M_NOWAIT|M_WAITOK)) != M_NOWAIT &&
3534 (flags & (M_NOWAIT|M_WAITOK)) != M_WAITOK)) {
3535 static int stack_count;
3538 if (stack_count < 10) {
3540 printf("uma_zalloc* called with bad WAIT flags:\n");
3548 #ifdef DEBUG_MEMGUARD
3549 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3551 item = memguard_alloc(zone->uz_size, flags);
3553 error = EJUSTRETURN;
3554 if (zone->uz_init != NULL &&
3555 zone->uz_init(item, zone->uz_size, flags) != 0) {
3559 if (zone->uz_ctor != NULL &&
3560 zone->uz_ctor(item, zone->uz_size, udata,
3562 counter_u64_add(zone->uz_fails, 1);
3563 if (zone->uz_fini != NULL)
3564 zone->uz_fini(item, zone->uz_size);
3571 /* This is unfortunate but should not be fatal. */
3578 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3580 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3581 ("uma_zfree_debug: called with spinlock or critical section held"));
3583 #ifdef DEBUG_MEMGUARD
3584 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3585 if (zone->uz_dtor != NULL)
3586 zone->uz_dtor(item, zone->uz_size, udata);
3587 if (zone->uz_fini != NULL)
3588 zone->uz_fini(item, zone->uz_size);
3589 memguard_free(item);
3590 return (EJUSTRETURN);
3597 static inline void *
3598 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3599 void *udata, int flags)
3604 item = cache_bucket_pop(cache, bucket);
3605 size = cache_uz_size(cache);
3606 uz_flags = cache_uz_flags(cache);
3608 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3611 static __noinline void *
3612 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3614 uma_cache_bucket_t bucket;
3617 while (cache_alloc(zone, cache, udata, flags)) {
3618 cache = &zone->uz_cpu[curcpu];
3619 bucket = &cache->uc_allocbucket;
3620 if (__predict_false(bucket->ucb_cnt == 0))
3622 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3627 * We can not get a bucket so try to return a single item.
3629 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3630 domain = PCPU_GET(domain);
3632 domain = UMA_ANYDOMAIN;
3633 return (zone_alloc_item(zone, udata, domain, flags));
3638 uma_zalloc_smr(uma_zone_t zone, int flags)
3640 uma_cache_bucket_t bucket;
3643 CTR3(KTR_UMA, "uma_zalloc_smr zone %s(%p) flags %d", zone->uz_name,
3646 #ifdef UMA_ZALLOC_DEBUG
3649 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3650 ("uma_zalloc_arg: called with non-SMR zone."));
3651 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3656 cache = &zone->uz_cpu[curcpu];
3657 bucket = &cache->uc_allocbucket;
3658 if (__predict_false(bucket->ucb_cnt == 0))
3659 return (cache_alloc_retry(zone, cache, NULL, flags));
3660 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3665 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3667 uma_cache_bucket_t bucket;
3670 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3671 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3673 /* This is the fast path allocation */
3674 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3677 #ifdef UMA_ZALLOC_DEBUG
3680 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3681 ("uma_zalloc_arg: called with SMR zone."));
3682 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3687 * If possible, allocate from the per-CPU cache. There are two
3688 * requirements for safe access to the per-CPU cache: (1) the thread
3689 * accessing the cache must not be preempted or yield during access,
3690 * and (2) the thread must not migrate CPUs without switching which
3691 * cache it accesses. We rely on a critical section to prevent
3692 * preemption and migration. We release the critical section in
3693 * order to acquire the zone mutex if we are unable to allocate from
3694 * the current cache; when we re-acquire the critical section, we
3695 * must detect and handle migration if it has occurred.
3698 cache = &zone->uz_cpu[curcpu];
3699 bucket = &cache->uc_allocbucket;
3700 if (__predict_false(bucket->ucb_cnt == 0))
3701 return (cache_alloc_retry(zone, cache, udata, flags));
3702 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3706 * Replenish an alloc bucket and possibly restore an old one. Called in
3707 * a critical section. Returns in a critical section.
3709 * A false return value indicates an allocation failure.
3710 * A true return value indicates success and the caller should retry.
3712 static __noinline bool
3713 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3715 uma_bucket_t bucket;
3716 int curdomain, domain;
3719 CRITICAL_ASSERT(curthread);
3722 * If we have run out of items in our alloc bucket see
3723 * if we can switch with the free bucket.
3725 * SMR Zones can't re-use the free bucket until the sequence has
3728 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3729 cache->uc_freebucket.ucb_cnt != 0) {
3730 cache_bucket_swap(&cache->uc_freebucket,
3731 &cache->uc_allocbucket);
3736 * Discard any empty allocation bucket while we hold no locks.
3738 bucket = cache_bucket_unload_alloc(cache);
3741 if (bucket != NULL) {
3742 KASSERT(bucket->ub_cnt == 0,
3743 ("cache_alloc: Entered with non-empty alloc bucket."));
3744 bucket_free(zone, bucket, udata);
3748 * Attempt to retrieve the item from the per-CPU cache has failed, so
3749 * we must go back to the zone. This requires the zdom lock, so we
3750 * must drop the critical section, then re-acquire it when we go back
3751 * to the cache. Since the critical section is released, we may be
3752 * preempted or migrate. As such, make sure not to maintain any
3753 * thread-local state specific to the cache from prior to releasing
3754 * the critical section.
3756 domain = PCPU_GET(domain);
3757 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3758 VM_DOMAIN_EMPTY(domain))
3759 domain = zone_domain_highest(zone, domain);
3760 bucket = cache_fetch_bucket(zone, cache, domain);
3761 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3762 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3768 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3769 zone->uz_name, zone, bucket);
3770 if (bucket == NULL) {
3776 * See if we lost the race or were migrated. Cache the
3777 * initialized bucket to make this less likely or claim
3778 * the memory directly.
3781 cache = &zone->uz_cpu[curcpu];
3782 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3783 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3784 (curdomain = PCPU_GET(domain)) == domain ||
3785 VM_DOMAIN_EMPTY(curdomain))) {
3787 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3789 cache_bucket_load_alloc(cache, bucket);
3794 * We lost the race, release this bucket and start over.
3797 zone_put_bucket(zone, domain, bucket, udata, !new);
3804 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3807 uma_bucket_t bucket;
3808 uma_zone_domain_t zdom;
3812 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3813 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3815 /* This is the fast path allocation */
3816 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3817 zone->uz_name, zone, domain, flags);
3819 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3820 ("uma_zalloc_domain: called with SMR zone."));
3822 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3823 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3825 if (vm_ndomains == 1)
3826 return (uma_zalloc_arg(zone, udata, flags));
3828 #ifdef UMA_ZALLOC_DEBUG
3829 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3834 * Try to allocate from the bucket cache before falling back to the keg.
3835 * We could try harder and attempt to allocate from per-CPU caches or
3836 * the per-domain cross-domain buckets, but the complexity is probably
3837 * not worth it. It is more important that frees of previous
3838 * cross-domain allocations do not blow up the cache.
3840 zdom = zone_domain_lock(zone, domain);
3841 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3842 item = bucket->ub_bucket[bucket->ub_cnt - 1];
3844 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3847 zone_put_bucket(zone, domain, bucket, udata, true);
3848 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3851 KASSERT(item_domain(item) == domain,
3852 ("%s: bucket cache item %p from wrong domain",
3854 counter_u64_add(zone->uz_allocs, 1);
3859 return (zone_alloc_item(zone, udata, domain, flags));
3861 return (uma_zalloc_arg(zone, udata, flags));
3866 * Find a slab with some space. Prefer slabs that are partially used over those
3867 * that are totally full. This helps to reduce fragmentation.
3869 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3873 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3879 KASSERT(domain >= 0 && domain < vm_ndomains,
3880 ("keg_first_slab: domain %d out of range", domain));
3881 KEG_LOCK_ASSERT(keg, domain);
3886 dom = &keg->uk_domain[domain];
3887 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3889 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3890 LIST_REMOVE(slab, us_link);
3891 dom->ud_free_slabs--;
3892 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3896 domain = (domain + 1) % vm_ndomains;
3897 } while (domain != start);
3903 * Fetch an existing slab from a free or partial list. Returns with the
3904 * keg domain lock held if a slab was found or unlocked if not.
3907 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3912 /* HASH has a single free list. */
3913 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3916 KEG_LOCK(keg, domain);
3917 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3918 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3919 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3920 KEG_UNLOCK(keg, domain);
3927 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3929 struct vm_domainset_iter di;
3934 KASSERT((flags & (M_WAITOK | M_NOVM)) != (M_WAITOK | M_NOVM),
3935 ("%s: invalid flags %#x", __func__, flags));
3939 * Use the keg's policy if upper layers haven't already specified a
3940 * domain (as happens with first-touch zones).
3942 * To avoid races we run the iterator with the keg lock held, but that
3943 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3944 * clear M_WAITOK and handle low memory conditions locally.
3946 rr = rdomain == UMA_ANYDOMAIN;
3948 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3949 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3957 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3962 * M_NOVM is used to break the recursion that can otherwise
3963 * occur if low-level memory management routines use UMA.
3965 if ((flags & M_NOVM) == 0) {
3966 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3972 if ((flags & M_USE_RESERVE) != 0) {
3974 * Drain reserves from other domains before
3975 * giving up or sleeping. It may be useful to
3976 * support per-domain reserves eventually.
3978 rdomain = UMA_ANYDOMAIN;
3981 if ((flags & M_WAITOK) == 0)
3983 vm_wait_domain(domain);
3984 } else if (vm_domainset_iter_policy(&di, &domain) != 0) {
3985 if ((flags & M_WAITOK) != 0) {
3986 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
3994 * We might not have been able to get a slab but another cpu
3995 * could have while we were unlocked. Check again before we
3998 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
4005 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
4011 KEG_LOCK_ASSERT(keg, slab->us_domain);
4013 dom = &keg->uk_domain[slab->us_domain];
4014 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
4015 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
4016 item = slab_item(slab, keg, freei);
4017 slab->us_freecount--;
4018 dom->ud_free_items--;
4021 * Move this slab to the full list. It must be on the partial list, so
4022 * we do not need to update the free slab count. In particular,
4023 * keg_fetch_slab() always returns slabs on the partial list.
4025 if (slab->us_freecount == 0) {
4026 LIST_REMOVE(slab, us_link);
4027 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
4034 zone_import(void *arg, void **bucket, int max, int domain, int flags)
4048 /* Try to keep the buckets totally full */
4049 for (i = 0; i < max; ) {
4050 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
4053 stripe = howmany(max, vm_ndomains);
4055 dom = &keg->uk_domain[slab->us_domain];
4057 bucket[i++] = slab_alloc_item(keg, slab);
4058 if (keg->uk_reserve > 0 &&
4059 dom->ud_free_items <= keg->uk_reserve) {
4061 * Avoid depleting the reserve after a
4062 * successful item allocation, even if
4063 * M_USE_RESERVE is specified.
4065 KEG_UNLOCK(keg, slab->us_domain);
4070 * If the zone is striped we pick a new slab for every
4071 * N allocations. Eliminating this conditional will
4072 * instead pick a new domain for each bucket rather
4073 * than stripe within each bucket. The current option
4074 * produces more fragmentation and requires more cpu
4075 * time but yields better distribution.
4077 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
4078 vm_ndomains > 1 && --stripe == 0)
4081 } while (slab->us_freecount != 0 && i < max);
4082 KEG_UNLOCK(keg, slab->us_domain);
4084 /* Don't block if we allocated any successfully. */
4093 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
4095 uint64_t old, new, total, max;
4098 * The hard case. We're going to sleep because there were existing
4099 * sleepers or because we ran out of items. This routine enforces
4100 * fairness by keeping fifo order.
4102 * First release our ill gotten gains and make some noise.
4105 zone_free_limit(zone, count);
4106 zone_log_warning(zone);
4107 zone_maxaction(zone);
4108 if (flags & M_NOWAIT)
4112 * We need to allocate an item or set ourself as a sleeper
4113 * while the sleepq lock is held to avoid wakeup races. This
4114 * is essentially a home rolled semaphore.
4116 sleepq_lock(&zone->uz_max_items);
4117 old = zone->uz_items;
4119 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
4120 /* Cache the max since we will evaluate twice. */
4121 max = zone->uz_max_items;
4122 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
4123 UZ_ITEMS_COUNT(old) >= max)
4124 new = old + UZ_ITEMS_SLEEPER;
4126 new = old + MIN(count, max - old);
4127 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
4129 /* We may have successfully allocated under the sleepq lock. */
4130 if (UZ_ITEMS_SLEEPERS(new) == 0) {
4131 sleepq_release(&zone->uz_max_items);
4136 * This is in a different cacheline from uz_items so that we
4137 * don't constantly invalidate the fastpath cacheline when we
4138 * adjust item counts. This could be limited to toggling on
4141 atomic_add_32(&zone->uz_sleepers, 1);
4142 atomic_add_64(&zone->uz_sleeps, 1);
4145 * We have added ourselves as a sleeper. The sleepq lock
4146 * protects us from wakeup races. Sleep now and then retry.
4148 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
4149 sleepq_wait(&zone->uz_max_items, PVM);
4152 * After wakeup, remove ourselves as a sleeper and try
4153 * again. We no longer have the sleepq lock for protection.
4155 * Subract ourselves as a sleeper while attempting to add
4158 atomic_subtract_32(&zone->uz_sleepers, 1);
4159 old = atomic_fetchadd_64(&zone->uz_items,
4160 -(UZ_ITEMS_SLEEPER - count));
4161 /* We're no longer a sleeper. */
4162 old -= UZ_ITEMS_SLEEPER;
4165 * If we're still at the limit, restart. Notably do not
4166 * block on other sleepers. Cache the max value to protect
4167 * against changes via sysctl.
4169 total = UZ_ITEMS_COUNT(old);
4170 max = zone->uz_max_items;
4173 /* Truncate if necessary, otherwise wake other sleepers. */
4174 if (total + count > max) {
4175 zone_free_limit(zone, total + count - max);
4176 count = max - total;
4177 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
4178 wakeup_one(&zone->uz_max_items);
4185 * Allocate 'count' items from our max_items limit. Returns the number
4186 * available. If M_NOWAIT is not specified it will sleep until at least
4187 * one item can be allocated.
4190 zone_alloc_limit(uma_zone_t zone, int count, int flags)
4195 max = zone->uz_max_items;
4199 * We expect normal allocations to succeed with a simple
4202 old = atomic_fetchadd_64(&zone->uz_items, count);
4203 if (__predict_true(old + count <= max))
4207 * If we had some items and no sleepers just return the
4208 * truncated value. We have to release the excess space
4209 * though because that may wake sleepers who weren't woken
4210 * because we were temporarily over the limit.
4213 zone_free_limit(zone, (old + count) - max);
4216 return (zone_alloc_limit_hard(zone, count, flags));
4220 * Free a number of items back to the limit.
4223 zone_free_limit(uma_zone_t zone, int count)
4230 * In the common case we either have no sleepers or
4231 * are still over the limit and can just return.
4233 old = atomic_fetchadd_64(&zone->uz_items, -count);
4234 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4235 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4239 * Moderate the rate of wakeups. Sleepers will continue
4240 * to generate wakeups if necessary.
4242 wakeup_one(&zone->uz_max_items);
4246 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4248 uma_bucket_t bucket;
4249 int error, maxbucket, cnt;
4251 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4254 /* Avoid allocs targeting empty domains. */
4255 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4256 domain = UMA_ANYDOMAIN;
4257 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4258 domain = UMA_ANYDOMAIN;
4260 if (zone->uz_max_items > 0)
4261 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4264 maxbucket = zone->uz_bucket_size;
4268 /* Don't wait for buckets, preserve caller's NOVM setting. */
4269 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4270 if (bucket == NULL) {
4275 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4276 MIN(maxbucket, bucket->ub_entries), domain, flags);
4279 * Initialize the memory if necessary.
4281 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4284 for (i = 0; i < bucket->ub_cnt; i++) {
4285 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
4286 error = zone->uz_init(bucket->ub_bucket[i],
4287 zone->uz_size, flags);
4288 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
4294 * If we couldn't initialize the whole bucket, put the
4295 * rest back onto the freelist.
4297 if (i != bucket->ub_cnt) {
4298 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4299 bucket->ub_cnt - i);
4301 bzero(&bucket->ub_bucket[i],
4302 sizeof(void *) * (bucket->ub_cnt - i));
4308 cnt = bucket->ub_cnt;
4309 if (bucket->ub_cnt == 0) {
4310 bucket_free(zone, bucket, udata);
4311 counter_u64_add(zone->uz_fails, 1);
4315 if (zone->uz_max_items > 0 && cnt < maxbucket)
4316 zone_free_limit(zone, maxbucket - cnt);
4322 * Allocates a single item from a zone.
4325 * zone The zone to alloc for.
4326 * udata The data to be passed to the constructor.
4327 * domain The domain to allocate from or UMA_ANYDOMAIN.
4328 * flags M_WAITOK, M_NOWAIT, M_ZERO.
4331 * NULL if there is no memory and M_NOWAIT is set
4332 * An item if successful
4336 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4340 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4341 counter_u64_add(zone->uz_fails, 1);
4345 /* Avoid allocs targeting empty domains. */
4346 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4347 domain = UMA_ANYDOMAIN;
4349 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4353 * We have to call both the zone's init (not the keg's init)
4354 * and the zone's ctor. This is because the item is going from
4355 * a keg slab directly to the user, and the user is expecting it
4356 * to be both zone-init'd as well as zone-ctor'd.
4358 if (zone->uz_init != NULL) {
4361 kasan_mark_item_valid(zone, item);
4362 error = zone->uz_init(item, zone->uz_size, flags);
4363 kasan_mark_item_invalid(zone, item);
4365 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4369 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4374 counter_u64_add(zone->uz_allocs, 1);
4375 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4376 zone->uz_name, zone);
4381 counter_u64_add(zone->uz_fails, 1);
4383 if (zone->uz_max_items > 0)
4384 zone_free_limit(zone, 1);
4385 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4386 zone->uz_name, zone);
4393 uma_zfree_smr(uma_zone_t zone, void *item)
4396 uma_cache_bucket_t bucket;
4397 int itemdomain, uz_flags;
4399 CTR3(KTR_UMA, "uma_zfree_smr zone %s(%p) item %p",
4400 zone->uz_name, zone, item);
4402 #ifdef UMA_ZALLOC_DEBUG
4403 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4404 ("uma_zfree_smr: called with non-SMR zone."));
4405 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4406 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4407 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4410 cache = &zone->uz_cpu[curcpu];
4411 uz_flags = cache_uz_flags(cache);
4414 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4415 itemdomain = item_domain(item);
4419 cache = &zone->uz_cpu[curcpu];
4420 /* SMR Zones must free to the free bucket. */
4421 bucket = &cache->uc_freebucket;
4423 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4424 PCPU_GET(domain) != itemdomain) {
4425 bucket = &cache->uc_crossbucket;
4428 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4429 cache_bucket_push(cache, bucket, item);
4433 } while (cache_free(zone, cache, NULL, itemdomain));
4437 * If nothing else caught this, we'll just do an internal free.
4439 zone_free_item(zone, item, NULL, SKIP_NONE);
4444 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4447 uma_cache_bucket_t bucket;
4448 int itemdomain, uz_flags;
4450 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4451 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4453 CTR3(KTR_UMA, "uma_zfree_arg zone %s(%p) item %p",
4454 zone->uz_name, zone, item);
4456 #ifdef UMA_ZALLOC_DEBUG
4457 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4458 ("uma_zfree_arg: called with SMR zone."));
4459 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4462 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4467 * We are accessing the per-cpu cache without a critical section to
4468 * fetch size and flags. This is acceptable, if we are preempted we
4469 * will simply read another cpu's line.
4471 cache = &zone->uz_cpu[curcpu];
4472 uz_flags = cache_uz_flags(cache);
4473 if (UMA_ALWAYS_CTORDTOR ||
4474 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4475 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4478 * The race here is acceptable. If we miss it we'll just have to wait
4479 * a little longer for the limits to be reset.
4481 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4482 if (atomic_load_32(&zone->uz_sleepers) > 0)
4487 * If possible, free to the per-CPU cache. There are two
4488 * requirements for safe access to the per-CPU cache: (1) the thread
4489 * accessing the cache must not be preempted or yield during access,
4490 * and (2) the thread must not migrate CPUs without switching which
4491 * cache it accesses. We rely on a critical section to prevent
4492 * preemption and migration. We release the critical section in
4493 * order to acquire the zone mutex if we are unable to free to the
4494 * current cache; when we re-acquire the critical section, we must
4495 * detect and handle migration if it has occurred.
4499 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4500 itemdomain = item_domain(item);
4504 cache = &zone->uz_cpu[curcpu];
4506 * Try to free into the allocbucket first to give LIFO
4507 * ordering for cache-hot datastructures. Spill over
4508 * into the freebucket if necessary. Alloc will swap
4509 * them if one runs dry.
4511 bucket = &cache->uc_allocbucket;
4513 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4514 PCPU_GET(domain) != itemdomain) {
4515 bucket = &cache->uc_crossbucket;
4518 if (bucket->ucb_cnt == bucket->ucb_entries &&
4519 cache->uc_freebucket.ucb_cnt <
4520 cache->uc_freebucket.ucb_entries)
4521 cache_bucket_swap(&cache->uc_freebucket,
4522 &cache->uc_allocbucket);
4523 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4524 cache_bucket_push(cache, bucket, item);
4528 } while (cache_free(zone, cache, udata, itemdomain));
4532 * If nothing else caught this, we'll just do an internal free.
4535 zone_free_item(zone, item, udata, SKIP_DTOR);
4540 * sort crossdomain free buckets to domain correct buckets and cache
4544 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4546 struct uma_bucketlist emptybuckets, fullbuckets;
4547 uma_zone_domain_t zdom;
4554 "uma_zfree: zone %s(%p) draining cross bucket %p",
4555 zone->uz_name, zone, bucket);
4558 * It is possible for buckets to arrive here out of order so we fetch
4559 * the current smr seq rather than accepting the bucket's.
4561 seq = SMR_SEQ_INVALID;
4562 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4563 seq = smr_advance(zone->uz_smr);
4566 * To avoid having ndomain * ndomain buckets for sorting we have a
4567 * lock on the current crossfree bucket. A full matrix with
4568 * per-domain locking could be used if necessary.
4570 STAILQ_INIT(&emptybuckets);
4571 STAILQ_INIT(&fullbuckets);
4572 ZONE_CROSS_LOCK(zone);
4573 for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4574 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4575 domain = item_domain(item);
4576 zdom = ZDOM_GET(zone, domain);
4577 if (zdom->uzd_cross == NULL) {
4578 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4579 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4580 zdom->uzd_cross = b;
4583 * Avoid allocating a bucket with the cross lock
4584 * held, since allocation can trigger a
4585 * cross-domain free and bucket zones may
4586 * allocate from each other.
4588 ZONE_CROSS_UNLOCK(zone);
4589 b = bucket_alloc(zone, udata, M_NOWAIT);
4592 ZONE_CROSS_LOCK(zone);
4593 if (zdom->uzd_cross != NULL) {
4594 STAILQ_INSERT_HEAD(&emptybuckets, b,
4597 zdom->uzd_cross = b;
4601 b = zdom->uzd_cross;
4602 b->ub_bucket[b->ub_cnt++] = item;
4604 if (b->ub_cnt == b->ub_entries) {
4605 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4606 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4607 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4608 zdom->uzd_cross = b;
4611 ZONE_CROSS_UNLOCK(zone);
4613 if (bucket->ub_cnt == 0)
4614 bucket->ub_seq = SMR_SEQ_INVALID;
4615 bucket_free(zone, bucket, udata);
4617 while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4618 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4619 bucket_free(zone, b, udata);
4621 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4622 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4623 domain = item_domain(b->ub_bucket[0]);
4624 zone_put_bucket(zone, domain, b, udata, true);
4630 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4631 int itemdomain, bool ws)
4636 * Buckets coming from the wrong domain will be entirely for the
4637 * only other domain on two domain systems. In this case we can
4638 * simply cache them. Otherwise we need to sort them back to
4641 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4642 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4643 zone_free_cross(zone, bucket, udata);
4649 * Attempt to save the bucket in the zone's domain bucket cache.
4652 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4653 zone->uz_name, zone, bucket);
4654 /* ub_cnt is pointing to the last free item */
4655 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4656 itemdomain = zone_domain_lowest(zone, itemdomain);
4657 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4661 * Populate a free or cross bucket for the current cpu cache. Free any
4662 * existing full bucket either to the zone cache or back to the slab layer.
4664 * Enters and returns in a critical section. false return indicates that
4665 * we can not satisfy this free in the cache layer. true indicates that
4666 * the caller should retry.
4668 static __noinline bool
4669 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, int itemdomain)
4671 uma_cache_bucket_t cbucket;
4672 uma_bucket_t newbucket, bucket;
4674 CRITICAL_ASSERT(curthread);
4676 if (zone->uz_bucket_size == 0)
4679 cache = &zone->uz_cpu[curcpu];
4683 * FIRSTTOUCH domains need to free to the correct zdom. When
4684 * enabled this is the zdom of the item. The bucket is the
4685 * cross bucket if the current domain and itemdomain do not match.
4687 cbucket = &cache->uc_freebucket;
4689 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4690 if (PCPU_GET(domain) != itemdomain) {
4691 cbucket = &cache->uc_crossbucket;
4692 if (cbucket->ucb_cnt != 0)
4693 counter_u64_add(zone->uz_xdomain,
4698 bucket = cache_bucket_unload(cbucket);
4699 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4700 ("cache_free: Entered with non-full free bucket."));
4702 /* We are no longer associated with this CPU. */
4706 * Don't let SMR zones operate without a free bucket. Force
4707 * a synchronize and re-use this one. We will only degrade
4708 * to a synchronize every bucket_size items rather than every
4709 * item if we fail to allocate a bucket.
4711 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4713 bucket->ub_seq = smr_advance(zone->uz_smr);
4714 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4715 if (newbucket == NULL && bucket != NULL) {
4716 bucket_drain(zone, bucket);
4720 } else if (!bucketdisable)
4721 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4724 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4727 if ((bucket = newbucket) == NULL)
4729 cache = &zone->uz_cpu[curcpu];
4732 * Check to see if we should be populating the cross bucket. If it
4733 * is already populated we will fall through and attempt to populate
4736 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4737 if (PCPU_GET(domain) != itemdomain &&
4738 cache->uc_crossbucket.ucb_bucket == NULL) {
4739 cache_bucket_load_cross(cache, bucket);
4745 * We may have lost the race to fill the bucket or switched CPUs.
4747 if (cache->uc_freebucket.ucb_bucket != NULL) {
4749 bucket_free(zone, bucket, udata);
4752 cache_bucket_load_free(cache, bucket);
4758 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4765 KEG_LOCK_ASSERT(keg, slab->us_domain);
4767 /* Do we need to remove from any lists? */
4768 dom = &keg->uk_domain[slab->us_domain];
4769 if (slab->us_freecount + 1 == keg->uk_ipers) {
4770 LIST_REMOVE(slab, us_link);
4771 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4772 dom->ud_free_slabs++;
4773 } else if (slab->us_freecount == 0) {
4774 LIST_REMOVE(slab, us_link);
4775 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4778 /* Slab management. */
4779 freei = slab_item_index(slab, keg, item);
4780 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4781 slab->us_freecount++;
4783 /* Keg statistics. */
4784 dom->ud_free_items++;
4788 zone_release(void *arg, void **bucket, int cnt)
4801 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4802 lock = KEG_LOCK(keg, 0);
4803 for (i = 0; i < cnt; i++) {
4805 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4806 slab = vtoslab((vm_offset_t)item);
4808 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4809 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4810 slab = hash_sfind(&keg->uk_hash, mem);
4812 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4814 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4817 lock = KEG_LOCK(keg, slab->us_domain);
4819 slab_free_item(zone, slab, item);
4826 * Frees a single item to any zone.
4829 * zone The zone to free to
4830 * item The item we're freeing
4831 * udata User supplied data for the dtor
4832 * skip Skip dtors and finis
4834 static __noinline void
4835 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4839 * If a free is sent directly to an SMR zone we have to
4840 * synchronize immediately because the item can instantly
4841 * be reallocated. This should only happen in degenerate
4842 * cases when no memory is available for per-cpu caches.
4844 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4845 smr_synchronize(zone->uz_smr);
4847 item_dtor(zone, item, zone->uz_size, udata, skip);
4849 if (skip < SKIP_FINI && zone->uz_fini) {
4850 kasan_mark_item_valid(zone, item);
4851 zone->uz_fini(item, zone->uz_size);
4852 kasan_mark_item_invalid(zone, item);
4855 zone->uz_release(zone->uz_arg, &item, 1);
4857 if (skip & SKIP_CNT)
4860 counter_u64_add(zone->uz_frees, 1);
4862 if (zone->uz_max_items > 0)
4863 zone_free_limit(zone, 1);
4868 uma_zone_set_max(uma_zone_t zone, int nitems)
4872 * If the limit is small, we may need to constrain the maximum per-CPU
4873 * cache size, or disable caching entirely.
4875 uma_zone_set_maxcache(zone, nitems);
4878 * XXX This can misbehave if the zone has any allocations with
4879 * no limit and a limit is imposed. There is currently no
4880 * way to clear a limit.
4883 zone->uz_max_items = nitems;
4884 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4885 zone_update_caches(zone);
4886 /* We may need to wake waiters. */
4887 wakeup(&zone->uz_max_items);
4895 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4897 int bpcpu, bpdom, bsize, nb;
4902 * Compute a lower bound on the number of items that may be cached in
4903 * the zone. Each CPU gets at least two buckets, and for cross-domain
4904 * frees we use an additional bucket per CPU and per domain. Select the
4905 * largest bucket size that does not exceed half of the requested limit,
4906 * with the left over space given to the full bucket cache.
4911 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4916 nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4917 bsize = nitems / nb / 2;
4918 if (bsize > BUCKET_MAX)
4920 else if (bsize == 0 && nitems / nb > 0)
4922 zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
4923 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4924 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4925 zone->uz_bucket_max = nitems - nb * bsize;
4931 uma_zone_get_max(uma_zone_t zone)
4935 nitems = atomic_load_64(&zone->uz_max_items);
4942 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4945 ZONE_ASSERT_COLD(zone);
4946 zone->uz_warning = warning;
4951 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4954 ZONE_ASSERT_COLD(zone);
4955 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4960 uma_zone_get_cur(uma_zone_t zone)
4966 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4967 nitems = counter_u64_fetch(zone->uz_allocs) -
4968 counter_u64_fetch(zone->uz_frees);
4970 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4971 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4973 return (nitems < 0 ? 0 : nitems);
4977 uma_zone_get_allocs(uma_zone_t zone)
4983 if (zone->uz_allocs != EARLY_COUNTER)
4984 nitems = counter_u64_fetch(zone->uz_allocs);
4986 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4992 uma_zone_get_frees(uma_zone_t zone)
4998 if (zone->uz_frees != EARLY_COUNTER)
4999 nitems = counter_u64_fetch(zone->uz_frees);
5001 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
5007 /* Used only for KEG_ASSERT_COLD(). */
5009 uma_keg_get_allocs(uma_keg_t keg)
5015 LIST_FOREACH(z, &keg->uk_zones, uz_link)
5016 nitems += uma_zone_get_allocs(z);
5024 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
5029 KEG_ASSERT_COLD(keg);
5030 keg->uk_init = uminit;
5035 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
5040 KEG_ASSERT_COLD(keg);
5041 keg->uk_fini = fini;
5046 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
5049 ZONE_ASSERT_COLD(zone);
5050 zone->uz_init = zinit;
5055 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
5058 ZONE_ASSERT_COLD(zone);
5059 zone->uz_fini = zfini;
5064 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
5069 KEG_ASSERT_COLD(keg);
5070 keg->uk_freef = freef;
5075 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
5080 KEG_ASSERT_COLD(keg);
5081 keg->uk_allocf = allocf;
5086 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
5089 ZONE_ASSERT_COLD(zone);
5091 KASSERT(smr != NULL, ("Got NULL smr"));
5092 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
5093 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
5094 zone->uz_flags |= UMA_ZONE_SMR;
5096 zone_update_caches(zone);
5100 uma_zone_get_smr(uma_zone_t zone)
5103 return (zone->uz_smr);
5108 uma_zone_reserve(uma_zone_t zone, int items)
5113 KEG_ASSERT_COLD(keg);
5114 keg->uk_reserve = items;
5119 uma_zone_reserve_kva(uma_zone_t zone, int count)
5126 KEG_ASSERT_COLD(keg);
5127 ZONE_ASSERT_COLD(zone);
5129 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
5131 #ifdef UMA_MD_SMALL_ALLOC
5132 if (keg->uk_ppera > 1) {
5136 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
5142 MPASS(keg->uk_kva == 0);
5145 zone->uz_max_items = pages * keg->uk_ipers;
5146 #ifdef UMA_MD_SMALL_ALLOC
5147 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
5149 keg->uk_allocf = noobj_alloc;
5151 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5152 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5153 zone_update_caches(zone);
5160 uma_prealloc(uma_zone_t zone, int items)
5162 struct vm_domainset_iter di;
5166 int aflags, domain, slabs;
5169 slabs = howmany(items, keg->uk_ipers);
5170 while (slabs-- > 0) {
5172 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
5175 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
5178 dom = &keg->uk_domain[slab->us_domain];
5180 * keg_alloc_slab() always returns a slab on the
5183 LIST_REMOVE(slab, us_link);
5184 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
5186 dom->ud_free_slabs++;
5187 KEG_UNLOCK(keg, slab->us_domain);
5190 if (vm_domainset_iter_policy(&di, &domain) != 0)
5191 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
5197 * Returns a snapshot of memory consumption in bytes.
5200 uma_zone_memory(uma_zone_t zone)
5206 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
5207 for (i = 0; i < vm_ndomains; i++)
5208 sz += ZDOM_GET(zone, i)->uzd_nitems;
5209 return (sz * zone->uz_size);
5211 for (i = 0; i < vm_ndomains; i++)
5212 sz += zone->uz_keg->uk_domain[i].ud_pages;
5214 return (sz * PAGE_SIZE);
5217 struct uma_reclaim_args {
5223 uma_reclaim_domain_cb(uma_zone_t zone, void *arg)
5225 struct uma_reclaim_args *args;
5228 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0)
5229 uma_zone_reclaim_domain(zone, args->req, args->domain);
5234 uma_reclaim(int req)
5236 uma_reclaim_domain(req, UMA_ANYDOMAIN);
5240 uma_reclaim_domain(int req, int domain)
5242 struct uma_reclaim_args args;
5246 args.domain = domain;
5249 sx_slock(&uma_reclaim_lock);
5251 case UMA_RECLAIM_TRIM:
5252 case UMA_RECLAIM_DRAIN:
5253 zone_foreach(uma_reclaim_domain_cb, &args);
5255 case UMA_RECLAIM_DRAIN_CPU:
5256 zone_foreach(uma_reclaim_domain_cb, &args);
5257 pcpu_cache_drain_safe(NULL);
5258 zone_foreach(uma_reclaim_domain_cb, &args);
5261 panic("unhandled reclamation request %d", req);
5265 * Some slabs may have been freed but this zone will be visited early
5266 * we visit again so that we can free pages that are empty once other
5267 * zones are drained. We have to do the same for buckets.
5269 uma_zone_reclaim_domain(slabzones[0], UMA_RECLAIM_DRAIN, domain);
5270 uma_zone_reclaim_domain(slabzones[1], UMA_RECLAIM_DRAIN, domain);
5271 bucket_zone_drain(domain);
5272 sx_sunlock(&uma_reclaim_lock);
5275 static volatile int uma_reclaim_needed;
5278 uma_reclaim_wakeup(void)
5281 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5282 wakeup(uma_reclaim);
5286 uma_reclaim_worker(void *arg __unused)
5290 sx_xlock(&uma_reclaim_lock);
5291 while (atomic_load_int(&uma_reclaim_needed) == 0)
5292 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5294 sx_xunlock(&uma_reclaim_lock);
5295 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5296 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5297 atomic_store_int(&uma_reclaim_needed, 0);
5298 /* Don't fire more than once per-second. */
5299 pause("umarclslp", hz);
5305 uma_zone_reclaim(uma_zone_t zone, int req)
5307 uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5311 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5314 case UMA_RECLAIM_TRIM:
5315 zone_reclaim(zone, domain, M_NOWAIT, false);
5317 case UMA_RECLAIM_DRAIN:
5318 zone_reclaim(zone, domain, M_NOWAIT, true);
5320 case UMA_RECLAIM_DRAIN_CPU:
5321 pcpu_cache_drain_safe(zone);
5322 zone_reclaim(zone, domain, M_NOWAIT, true);
5325 panic("unhandled reclamation request %d", req);
5331 uma_zone_exhausted(uma_zone_t zone)
5334 return (atomic_load_32(&zone->uz_sleepers) > 0);
5341 return (uma_kmem_limit);
5345 uma_set_limit(unsigned long limit)
5348 uma_kmem_limit = limit;
5355 return (atomic_load_long(&uma_kmem_total));
5362 return (uma_kmem_limit - uma_size());
5367 * Generate statistics across both the zone and its per-cpu cache's. Return
5368 * desired statistics if the pointer is non-NULL for that statistic.
5370 * Note: does not update the zone statistics, as it can't safely clear the
5371 * per-CPU cache statistic.
5375 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5376 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5379 uint64_t allocs, frees, sleeps, xdomain;
5382 allocs = frees = sleeps = xdomain = 0;
5385 cache = &z->uz_cpu[cpu];
5386 cachefree += cache->uc_allocbucket.ucb_cnt;
5387 cachefree += cache->uc_freebucket.ucb_cnt;
5388 xdomain += cache->uc_crossbucket.ucb_cnt;
5389 cachefree += cache->uc_crossbucket.ucb_cnt;
5390 allocs += cache->uc_allocs;
5391 frees += cache->uc_frees;
5393 allocs += counter_u64_fetch(z->uz_allocs);
5394 frees += counter_u64_fetch(z->uz_frees);
5395 xdomain += counter_u64_fetch(z->uz_xdomain);
5396 sleeps += z->uz_sleeps;
5397 if (cachefreep != NULL)
5398 *cachefreep = cachefree;
5399 if (allocsp != NULL)
5403 if (sleepsp != NULL)
5405 if (xdomainp != NULL)
5406 *xdomainp = xdomain;
5411 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5418 rw_rlock(&uma_rwlock);
5419 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5420 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5423 LIST_FOREACH(z, &uma_cachezones, uz_link)
5426 rw_runlock(&uma_rwlock);
5427 return (sysctl_handle_int(oidp, &count, 0, req));
5431 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5432 struct uma_percpu_stat *ups, bool internal)
5434 uma_zone_domain_t zdom;
5438 for (i = 0; i < vm_ndomains; i++) {
5439 zdom = ZDOM_GET(z, i);
5440 uth->uth_zone_free += zdom->uzd_nitems;
5442 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5443 uth->uth_frees = counter_u64_fetch(z->uz_frees);
5444 uth->uth_fails = counter_u64_fetch(z->uz_fails);
5445 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5446 uth->uth_sleeps = z->uz_sleeps;
5448 for (i = 0; i < mp_maxid + 1; i++) {
5449 bzero(&ups[i], sizeof(*ups));
5450 if (internal || CPU_ABSENT(i))
5452 cache = &z->uz_cpu[i];
5453 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5454 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5455 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5456 ups[i].ups_allocs = cache->uc_allocs;
5457 ups[i].ups_frees = cache->uc_frees;
5462 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5464 struct uma_stream_header ush;
5465 struct uma_type_header uth;
5466 struct uma_percpu_stat *ups;
5471 uint32_t kfree, pages;
5472 int count, error, i;
5474 error = sysctl_wire_old_buffer(req, 0);
5477 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5478 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5479 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5482 rw_rlock(&uma_rwlock);
5483 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5484 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5488 LIST_FOREACH(z, &uma_cachezones, uz_link)
5492 * Insert stream header.
5494 bzero(&ush, sizeof(ush));
5495 ush.ush_version = UMA_STREAM_VERSION;
5496 ush.ush_maxcpus = (mp_maxid + 1);
5497 ush.ush_count = count;
5498 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5500 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5502 for (i = 0; i < vm_ndomains; i++) {
5503 kfree += kz->uk_domain[i].ud_free_items;
5504 pages += kz->uk_domain[i].ud_pages;
5506 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5507 bzero(&uth, sizeof(uth));
5508 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5509 uth.uth_align = kz->uk_align;
5510 uth.uth_size = kz->uk_size;
5511 uth.uth_rsize = kz->uk_rsize;
5512 if (z->uz_max_items > 0) {
5513 items = UZ_ITEMS_COUNT(z->uz_items);
5514 uth.uth_pages = (items / kz->uk_ipers) *
5517 uth.uth_pages = pages;
5518 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5520 uth.uth_limit = z->uz_max_items;
5521 uth.uth_keg_free = kfree;
5524 * A zone is secondary is it is not the first entry
5525 * on the keg's zone list.
5527 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5528 (LIST_FIRST(&kz->uk_zones) != z))
5529 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5530 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5531 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5532 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5533 for (i = 0; i < mp_maxid + 1; i++)
5534 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5537 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5538 bzero(&uth, sizeof(uth));
5539 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5540 uth.uth_size = z->uz_size;
5541 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5542 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5543 for (i = 0; i < mp_maxid + 1; i++)
5544 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5547 rw_runlock(&uma_rwlock);
5548 error = sbuf_finish(&sbuf);
5555 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5557 uma_zone_t zone = *(uma_zone_t *)arg1;
5560 max = uma_zone_get_max(zone);
5561 error = sysctl_handle_int(oidp, &max, 0, req);
5562 if (error || !req->newptr)
5565 uma_zone_set_max(zone, max);
5571 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5577 * Some callers want to add sysctls for global zones that
5578 * may not yet exist so they pass a pointer to a pointer.
5581 zone = *(uma_zone_t *)arg1;
5584 cur = uma_zone_get_cur(zone);
5585 return (sysctl_handle_int(oidp, &cur, 0, req));
5589 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5591 uma_zone_t zone = arg1;
5594 cur = uma_zone_get_allocs(zone);
5595 return (sysctl_handle_64(oidp, &cur, 0, req));
5599 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5601 uma_zone_t zone = arg1;
5604 cur = uma_zone_get_frees(zone);
5605 return (sysctl_handle_64(oidp, &cur, 0, req));
5609 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5612 uma_zone_t zone = arg1;
5615 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5616 if (zone->uz_flags != 0)
5617 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5619 sbuf_printf(&sbuf, "0");
5620 error = sbuf_finish(&sbuf);
5627 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5629 uma_keg_t keg = arg1;
5630 int avail, effpct, total;
5632 total = keg->uk_ppera * PAGE_SIZE;
5633 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5634 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5636 * We consider the client's requested size and alignment here, not the
5637 * real size determination uk_rsize, because we also adjust the real
5638 * size for internal implementation reasons (max bitset size).
5640 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5641 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5642 avail *= mp_maxid + 1;
5643 effpct = 100 * avail / total;
5644 return (sysctl_handle_int(oidp, &effpct, 0, req));
5648 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5650 uma_zone_t zone = arg1;
5653 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5654 return (sysctl_handle_64(oidp, &cur, 0, req));
5659 uma_dbg_getslab(uma_zone_t zone, void *item)
5666 * It is safe to return the slab here even though the
5667 * zone is unlocked because the item's allocation state
5668 * essentially holds a reference.
5670 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5671 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5673 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5674 return (vtoslab((vm_offset_t)mem));
5676 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5677 return ((uma_slab_t)(mem + keg->uk_pgoff));
5679 slab = hash_sfind(&keg->uk_hash, mem);
5686 uma_dbg_zskip(uma_zone_t zone, void *mem)
5689 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5692 return (uma_dbg_kskip(zone->uz_keg, mem));
5696 uma_dbg_kskip(uma_keg_t keg, void *mem)
5700 if (dbg_divisor == 0)
5703 if (dbg_divisor == 1)
5706 idx = (uintptr_t)mem >> PAGE_SHIFT;
5707 if (keg->uk_ipers > 1) {
5708 idx *= keg->uk_ipers;
5709 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5712 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5713 counter_u64_add(uma_skip_cnt, 1);
5716 counter_u64_add(uma_dbg_cnt, 1);
5722 * Set up the slab's freei data such that uma_dbg_free can function.
5726 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5732 slab = uma_dbg_getslab(zone, item);
5734 panic("uma: item %p did not belong to zone %s",
5735 item, zone->uz_name);
5738 freei = slab_item_index(slab, keg, item);
5740 if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5741 slab_dbg_bits(slab, keg)))
5742 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5743 item, zone, zone->uz_name, slab, freei);
5747 * Verifies freed addresses. Checks for alignment, valid slab membership
5748 * and duplicate frees.
5752 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5758 slab = uma_dbg_getslab(zone, item);
5760 panic("uma: Freed item %p did not belong to zone %s",
5761 item, zone->uz_name);
5764 freei = slab_item_index(slab, keg, item);
5766 if (freei >= keg->uk_ipers)
5767 panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5768 item, zone, zone->uz_name, slab, freei);
5770 if (slab_item(slab, keg, freei) != item)
5771 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5772 item, zone, zone->uz_name, slab, freei);
5774 if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5775 slab_dbg_bits(slab, keg)))
5776 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5777 item, zone, zone->uz_name, slab, freei);
5779 #endif /* INVARIANTS */
5783 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5784 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5789 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5790 *allocs = counter_u64_fetch(z->uz_allocs);
5791 frees = counter_u64_fetch(z->uz_frees);
5792 *sleeps = z->uz_sleeps;
5796 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5798 for (i = 0; i < vm_ndomains; i++) {
5799 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5800 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5801 (LIST_FIRST(&kz->uk_zones) != z)))
5802 *cachefree += kz->uk_domain[i].ud_free_items;
5804 *used = *allocs - frees;
5805 return (((int64_t)*used + *cachefree) * kz->uk_size);
5808 DB_SHOW_COMMAND(uma, db_show_uma)
5810 const char *fmt_hdr, *fmt_entry;
5813 uint64_t allocs, used, sleeps, xdomain;
5815 /* variables for sorting */
5817 uma_zone_t cur_zone, last_zone;
5818 int64_t cur_size, last_size, size;
5821 /* /i option produces machine-parseable CSV output */
5822 if (modif[0] == 'i') {
5823 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5824 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5826 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5827 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5830 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5831 "Sleeps", "Bucket", "Total Mem", "XFree");
5833 /* Sort the zones with largest size first. */
5835 last_size = INT64_MAX;
5840 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5841 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5843 * In the case of size ties, print out zones
5844 * in the order they are encountered. That is,
5845 * when we encounter the most recently output
5846 * zone, we have already printed all preceding
5847 * ties, and we must print all following ties.
5849 if (z == last_zone) {
5853 size = get_uma_stats(kz, z, &allocs, &used,
5854 &sleeps, &cachefree, &xdomain);
5855 if (size > cur_size && size < last_size + ties)
5863 if (cur_zone == NULL)
5866 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5867 &sleeps, &cachefree, &xdomain);
5868 db_printf(fmt_entry, cur_zone->uz_name,
5869 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5870 (uintmax_t)allocs, (uintmax_t)sleeps,
5871 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5876 last_zone = cur_zone;
5877 last_size = cur_size;
5881 DB_SHOW_COMMAND(umacache, db_show_umacache)
5884 uint64_t allocs, frees;
5888 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5889 "Requests", "Bucket");
5890 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5891 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5892 for (i = 0; i < vm_ndomains; i++)
5893 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5894 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5895 z->uz_name, (uintmax_t)z->uz_size,
5896 (intmax_t)(allocs - frees), cachefree,
5897 (uintmax_t)allocs, z->uz_bucket_size);