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 timeout_task uma_timeout_task;
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 *, int);
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 *context __unused, int pending __unused)
1113 zone_foreach(zone_timeout, NULL);
1115 /* Reschedule this event */
1116 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task,
1121 * Update the working set size estimates for the zone's bucket cache.
1122 * The constants chosen here are somewhat arbitrary.
1125 zone_domain_update_wss(uma_zone_domain_t zdom)
1129 ZDOM_LOCK_ASSERT(zdom);
1130 MPASS(zdom->uzd_imax >= zdom->uzd_nitems);
1131 MPASS(zdom->uzd_nitems >= zdom->uzd_bimin);
1132 MPASS(zdom->uzd_bimin >= zdom->uzd_imin);
1135 * Estimate WSS as modified moving average of biggest allocation
1136 * batches for each period over few minutes (UMA_TIMEOUT of 20s).
1138 zdom->uzd_wss = lmax(zdom->uzd_wss * 3 / 4,
1139 zdom->uzd_imax - zdom->uzd_bimin);
1142 * Estimate longtime minimum item count as a combination of recent
1143 * minimum item count, adjusted by WSS for safety, and the modified
1144 * moving average over the last several hours (UMA_TIMEOUT of 20s).
1145 * timin measures time since limin tried to go negative, that means
1146 * we were dangerously close to or got out of cache.
1148 m = zdom->uzd_imin - zdom->uzd_wss;
1150 if (zdom->uzd_limin >= m)
1151 zdom->uzd_limin = m;
1153 zdom->uzd_limin = (m + zdom->uzd_limin * 255) / 256;
1156 zdom->uzd_limin = 0;
1157 zdom->uzd_timin = 0;
1160 /* To reduce period edge effects on WSS keep half of the imax. */
1161 atomic_subtract_long(&zdom->uzd_imax,
1162 (zdom->uzd_imax - zdom->uzd_nitems + 1) / 2);
1163 zdom->uzd_imin = zdom->uzd_bimin = zdom->uzd_nitems;
1167 * Routine to perform timeout driven calculations. This expands the
1168 * hashes and does per cpu statistics aggregation.
1173 zone_timeout(uma_zone_t zone, void *unused)
1178 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
1184 * Hash zones are non-numa by definition so the first domain
1185 * is the only one present.
1188 pages = keg->uk_domain[0].ud_pages;
1191 * Expand the keg hash table.
1193 * This is done if the number of slabs is larger than the hash size.
1194 * What I'm trying to do here is completely reduce collisions. This
1195 * may be a little aggressive. Should I allow for two collisions max?
1197 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1198 struct uma_hash newhash;
1199 struct uma_hash oldhash;
1203 * This is so involved because allocating and freeing
1204 * while the keg lock is held will lead to deadlock.
1205 * I have to do everything in stages and check for
1209 ret = hash_alloc(&newhash, 1 << fls(slabs));
1212 if (hash_expand(&keg->uk_hash, &newhash)) {
1213 oldhash = keg->uk_hash;
1214 keg->uk_hash = newhash;
1219 hash_free(&oldhash);
1226 /* Trim caches not used for a long time. */
1227 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0) {
1228 for (int i = 0; i < vm_ndomains; i++) {
1229 if (bucket_cache_reclaim_domain(zone, false, false, i) &&
1230 (zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1231 keg_drain(zone->uz_keg, i);
1237 * Allocate and zero fill the next sized hash table from the appropriate
1241 * hash A new hash structure with the old hash size in uh_hashsize
1244 * 1 on success and 0 on failure.
1247 hash_alloc(struct uma_hash *hash, u_int size)
1251 KASSERT(powerof2(size), ("hash size must be power of 2"));
1252 if (size > UMA_HASH_SIZE_INIT) {
1253 hash->uh_hashsize = size;
1254 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1255 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1257 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1258 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1259 UMA_ANYDOMAIN, M_WAITOK);
1260 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1262 if (hash->uh_slab_hash) {
1263 bzero(hash->uh_slab_hash, alloc);
1264 hash->uh_hashmask = hash->uh_hashsize - 1;
1272 * Expands the hash table for HASH zones. This is done from zone_timeout
1273 * to reduce collisions. This must not be done in the regular allocation
1274 * path, otherwise, we can recurse on the vm while allocating pages.
1277 * oldhash The hash you want to expand
1278 * newhash The hash structure for the new table
1286 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1288 uma_hash_slab_t slab;
1292 if (!newhash->uh_slab_hash)
1295 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1299 * I need to investigate hash algorithms for resizing without a
1303 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1304 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1305 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1306 LIST_REMOVE(slab, uhs_hlink);
1307 hval = UMA_HASH(newhash, slab->uhs_data);
1308 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1316 * Free the hash bucket to the appropriate backing store.
1319 * slab_hash The hash bucket we're freeing
1320 * hashsize The number of entries in that hash bucket
1326 hash_free(struct uma_hash *hash)
1328 if (hash->uh_slab_hash == NULL)
1330 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1331 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1333 free(hash->uh_slab_hash, M_UMAHASH);
1337 * Frees all outstanding items in a bucket
1340 * zone The zone to free to, must be unlocked.
1341 * bucket The free/alloc bucket with items.
1347 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1351 if (bucket->ub_cnt == 0)
1354 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1355 bucket->ub_seq != SMR_SEQ_INVALID) {
1356 smr_wait(zone->uz_smr, bucket->ub_seq);
1357 bucket->ub_seq = SMR_SEQ_INVALID;
1358 for (i = 0; i < bucket->ub_cnt; i++)
1359 item_dtor(zone, bucket->ub_bucket[i],
1360 zone->uz_size, NULL, SKIP_NONE);
1363 for (i = 0; i < bucket->ub_cnt; i++) {
1364 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
1365 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1366 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
1368 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1369 if (zone->uz_max_items > 0)
1370 zone_free_limit(zone, bucket->ub_cnt);
1372 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1378 * Drains the per cpu caches for a zone.
1380 * NOTE: This may only be called while the zone is being torn down, and not
1381 * during normal operation. This is necessary in order that we do not have
1382 * to migrate CPUs to drain the per-CPU caches.
1385 * zone The zone to drain, must be unlocked.
1391 cache_drain(uma_zone_t zone)
1394 uma_bucket_t bucket;
1399 * XXX: It is safe to not lock the per-CPU caches, because we're
1400 * tearing down the zone anyway. I.e., there will be no further use
1401 * of the caches at this point.
1403 * XXX: It would good to be able to assert that the zone is being
1404 * torn down to prevent improper use of cache_drain().
1406 seq = SMR_SEQ_INVALID;
1407 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1408 seq = smr_advance(zone->uz_smr);
1410 cache = &zone->uz_cpu[cpu];
1411 bucket = cache_bucket_unload_alloc(cache);
1413 bucket_free(zone, bucket, NULL);
1414 bucket = cache_bucket_unload_free(cache);
1415 if (bucket != NULL) {
1416 bucket->ub_seq = seq;
1417 bucket_free(zone, bucket, NULL);
1419 bucket = cache_bucket_unload_cross(cache);
1420 if (bucket != NULL) {
1421 bucket->ub_seq = seq;
1422 bucket_free(zone, bucket, NULL);
1425 bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN);
1429 cache_shrink(uma_zone_t zone, void *unused)
1432 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1436 zone->uz_bucket_size =
1437 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1442 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1445 uma_bucket_t b1, b2, b3;
1448 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1451 b1 = b2 = b3 = NULL;
1453 cache = &zone->uz_cpu[curcpu];
1454 domain = PCPU_GET(domain);
1455 b1 = cache_bucket_unload_alloc(cache);
1458 * Don't flush SMR zone buckets. This leaves the zone without a
1459 * bucket and forces every free to synchronize().
1461 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1462 b2 = cache_bucket_unload_free(cache);
1463 b3 = cache_bucket_unload_cross(cache);
1468 zone_free_bucket(zone, b1, NULL, domain, false);
1470 zone_free_bucket(zone, b2, NULL, domain, false);
1472 /* Adjust the domain so it goes to zone_free_cross. */
1473 domain = (domain + 1) % vm_ndomains;
1474 zone_free_bucket(zone, b3, NULL, domain, false);
1479 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1480 * This is an expensive call because it needs to bind to all CPUs
1481 * one by one and enter a critical section on each of them in order
1482 * to safely access their cache buckets.
1483 * Zone lock must not be held on call this function.
1486 pcpu_cache_drain_safe(uma_zone_t zone)
1491 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1494 cache_shrink(zone, NULL);
1496 zone_foreach(cache_shrink, NULL);
1499 thread_lock(curthread);
1500 sched_bind(curthread, cpu);
1501 thread_unlock(curthread);
1504 cache_drain_safe_cpu(zone, NULL);
1506 zone_foreach(cache_drain_safe_cpu, NULL);
1508 thread_lock(curthread);
1509 sched_unbind(curthread);
1510 thread_unlock(curthread);
1514 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1515 * requested a drain, otherwise the per-domain caches are trimmed to either
1516 * estimated working set size.
1519 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain)
1521 uma_zone_domain_t zdom;
1522 uma_bucket_t bucket;
1527 * The cross bucket is partially filled and not part of
1528 * the item count. Reclaim it individually here.
1530 zdom = ZDOM_GET(zone, domain);
1531 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1532 ZONE_CROSS_LOCK(zone);
1533 bucket = zdom->uzd_cross;
1534 zdom->uzd_cross = NULL;
1535 ZONE_CROSS_UNLOCK(zone);
1537 bucket_free(zone, bucket, NULL);
1541 * If we were asked to drain the zone, we are done only once
1542 * this bucket cache is empty. If trim, we reclaim items in
1543 * excess of the zone's estimated working set size. Multiple
1544 * consecutive calls will shrink the WSS and so reclaim more.
1545 * If neither drain nor trim, then voluntarily reclaim 1/4
1546 * (to reduce first spike) of items not used for a long time.
1549 zone_domain_update_wss(zdom);
1553 target = zdom->uzd_wss;
1554 else if (zdom->uzd_timin > 900 / UMA_TIMEOUT)
1555 target = zdom->uzd_nitems - zdom->uzd_limin / 4;
1560 while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL &&
1561 zdom->uzd_nitems >= target + bucket->ub_cnt) {
1562 bucket = zone_fetch_bucket(zone, zdom, true);
1565 bucket_free(zone, bucket, NULL);
1574 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain)
1579 * Shrink the zone bucket size to ensure that the per-CPU caches
1580 * don't grow too large.
1582 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1583 zone->uz_bucket_size--;
1585 if (domain != UMA_ANYDOMAIN &&
1586 (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) {
1587 bucket_cache_reclaim_domain(zone, drain, true, domain);
1589 for (i = 0; i < vm_ndomains; i++)
1590 bucket_cache_reclaim_domain(zone, drain, true, i);
1595 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1602 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1603 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1605 mem = slab_data(slab, keg);
1606 size = PAGE_SIZE * keg->uk_ppera;
1608 kasan_mark_slab_valid(keg, mem);
1609 if (keg->uk_fini != NULL) {
1610 for (i = start - 1; i > -1; i--)
1613 * trash_fini implies that dtor was trash_dtor. trash_fini
1614 * would check that memory hasn't been modified since free,
1615 * which executed trash_dtor.
1616 * That's why we need to run uma_dbg_kskip() check here,
1617 * albeit we don't make skip check for other init/fini
1620 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1621 keg->uk_fini != trash_fini)
1623 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1625 flags = slab->us_flags;
1626 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1627 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1630 keg->uk_freef(mem, size, flags);
1631 uma_total_dec(size);
1635 keg_drain_domain(uma_keg_t keg, int domain)
1637 struct slabhead freeslabs;
1639 uma_slab_t slab, tmp;
1640 uint32_t i, stofree, stokeep, partial;
1642 dom = &keg->uk_domain[domain];
1643 LIST_INIT(&freeslabs);
1645 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1646 keg->uk_name, keg, domain, dom->ud_free_items);
1648 KEG_LOCK(keg, domain);
1651 * Are the free items in partially allocated slabs sufficient to meet
1652 * the reserve? If not, compute the number of fully free slabs that must
1655 partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1656 if (partial < keg->uk_reserve) {
1657 stokeep = min(dom->ud_free_slabs,
1658 howmany(keg->uk_reserve - partial, keg->uk_ipers));
1662 stofree = dom->ud_free_slabs - stokeep;
1665 * Partition the free slabs into two sets: those that must be kept in
1666 * order to maintain the reserve, and those that may be released back to
1667 * the system. Since one set may be much larger than the other,
1668 * populate the smaller of the two sets and swap them if necessary.
1670 for (i = min(stofree, stokeep); i > 0; i--) {
1671 slab = LIST_FIRST(&dom->ud_free_slab);
1672 LIST_REMOVE(slab, us_link);
1673 LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1675 if (stofree > stokeep)
1676 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1678 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1679 LIST_FOREACH(slab, &freeslabs, us_link)
1680 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1682 dom->ud_free_items -= stofree * keg->uk_ipers;
1683 dom->ud_free_slabs -= stofree;
1684 dom->ud_pages -= stofree * keg->uk_ppera;
1685 KEG_UNLOCK(keg, domain);
1687 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1688 keg_free_slab(keg, slab, keg->uk_ipers);
1692 * Frees pages from a keg back to the system. This is done on demand from
1693 * the pageout daemon.
1698 keg_drain(uma_keg_t keg, int domain)
1702 if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1704 if (domain != UMA_ANYDOMAIN) {
1705 keg_drain_domain(keg, domain);
1707 for (i = 0; i < vm_ndomains; i++)
1708 keg_drain_domain(keg, i);
1713 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain)
1716 * Count active reclaim operations in order to interlock with
1717 * zone_dtor(), which removes the zone from global lists before
1718 * attempting to reclaim items itself.
1720 * The zone may be destroyed while sleeping, so only zone_dtor() should
1724 if (waitok == M_WAITOK) {
1725 while (zone->uz_reclaimers > 0)
1726 msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1);
1728 zone->uz_reclaimers++;
1730 bucket_cache_reclaim(zone, drain, domain);
1732 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1733 keg_drain(zone->uz_keg, domain);
1735 zone->uz_reclaimers--;
1736 if (zone->uz_reclaimers == 0)
1742 * Allocate a new slab for a keg and inserts it into the partial slab list.
1743 * The keg should be unlocked on entry. If the allocation succeeds it will
1744 * be locked on return.
1747 * flags Wait flags for the item initialization routine
1748 * aflags Wait flags for the slab allocation
1751 * The slab that was allocated or NULL if there is no memory and the
1752 * caller specified M_NOWAIT.
1755 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1767 KASSERT(domain >= 0 && domain < vm_ndomains,
1768 ("keg_alloc_slab: domain %d out of range", domain));
1772 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1773 uma_hash_slab_t hslab;
1774 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1778 slab = &hslab->uhs_slab;
1782 * This reproduces the old vm_zone behavior of zero filling pages the
1783 * first time they are added to a zone.
1785 * Malloced items are zeroed in uma_zalloc.
1788 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1793 if (keg->uk_flags & UMA_ZONE_NODUMP)
1796 /* zone is passed for legacy reasons. */
1797 size = keg->uk_ppera * PAGE_SIZE;
1798 mem = keg->uk_allocf(zone, size, domain, &sflags, aflags);
1800 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1801 zone_free_item(slabzone(keg->uk_ipers),
1802 slab_tohashslab(slab), NULL, SKIP_NONE);
1805 uma_total_inc(size);
1807 /* For HASH zones all pages go to the same uma_domain. */
1808 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1811 kmsan_mark(mem, size,
1812 (aflags & M_ZERO) != 0 ? KMSAN_STATE_INITED : KMSAN_STATE_UNINIT);
1814 /* Point the slab into the allocated memory */
1815 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1816 slab = (uma_slab_t)(mem + keg->uk_pgoff);
1818 slab_tohashslab(slab)->uhs_data = mem;
1820 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1821 for (i = 0; i < keg->uk_ppera; i++)
1822 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1825 slab->us_freecount = keg->uk_ipers;
1826 slab->us_flags = sflags;
1827 slab->us_domain = domain;
1829 BIT_FILL(keg->uk_ipers, &slab->us_free);
1831 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1834 if (keg->uk_init != NULL) {
1835 for (i = 0; i < keg->uk_ipers; i++)
1836 if (keg->uk_init(slab_item(slab, keg, i),
1837 keg->uk_size, flags) != 0)
1839 if (i != keg->uk_ipers) {
1840 keg_free_slab(keg, slab, i);
1844 kasan_mark_slab_invalid(keg, mem);
1845 KEG_LOCK(keg, domain);
1847 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1848 slab, keg->uk_name, keg);
1850 if (keg->uk_flags & UMA_ZFLAG_HASH)
1851 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1854 * If we got a slab here it's safe to mark it partially used
1855 * and return. We assume that the caller is going to remove
1856 * at least one item.
1858 dom = &keg->uk_domain[domain];
1859 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1860 dom->ud_pages += keg->uk_ppera;
1861 dom->ud_free_items += keg->uk_ipers;
1871 * This function is intended to be used early on in place of page_alloc(). It
1872 * performs contiguous physical memory allocations and uses a bump allocator for
1873 * KVA, so is usable before the kernel map is initialized.
1876 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1883 pages = howmany(bytes, PAGE_SIZE);
1884 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1886 *pflag = UMA_SLAB_BOOT;
1887 m = vm_page_alloc_noobj_contig_domain(domain, malloc2vm_flags(wait) |
1888 VM_ALLOC_WIRED, pages, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
1889 VM_MEMATTR_DEFAULT);
1893 pa = VM_PAGE_TO_PHYS(m);
1894 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1895 #if defined(__aarch64__) || defined(__amd64__) || \
1896 defined(__riscv) || defined(__powerpc64__)
1897 if ((wait & M_NODUMP) == 0)
1902 /* Allocate KVA and indirectly advance bootmem. */
1903 return ((void *)pmap_map(&bootmem, m->phys_addr,
1904 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE));
1908 startup_free(void *mem, vm_size_t bytes)
1913 va = (vm_offset_t)mem;
1914 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1917 * startup_alloc() returns direct-mapped slabs on some platforms. Avoid
1918 * unmapping ranges of the direct map.
1920 if (va >= bootstart && va + bytes <= bootmem)
1921 pmap_remove(kernel_pmap, va, va + bytes);
1922 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1923 #if defined(__aarch64__) || defined(__amd64__) || \
1924 defined(__riscv) || defined(__powerpc64__)
1925 dump_drop_page(VM_PAGE_TO_PHYS(m));
1927 vm_page_unwire_noq(m);
1933 * Allocates a number of pages from the system
1936 * bytes The number of bytes requested
1937 * wait Shall we wait?
1940 * A pointer to the alloced memory or possibly
1941 * NULL if M_NOWAIT is set.
1944 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1947 void *p; /* Returned page */
1949 *pflag = UMA_SLAB_KERNEL;
1950 p = kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1956 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1959 struct pglist alloctail;
1960 vm_offset_t addr, zkva;
1962 vm_page_t p, p_next;
1967 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1969 TAILQ_INIT(&alloctail);
1970 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | malloc2vm_flags(wait);
1971 *pflag = UMA_SLAB_KERNEL;
1972 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1973 if (CPU_ABSENT(cpu)) {
1974 p = vm_page_alloc_noobj(flags);
1977 p = vm_page_alloc_noobj(flags);
1979 pc = pcpu_find(cpu);
1980 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1983 p = vm_page_alloc_noobj_domain(pc->pc_domain,
1985 if (__predict_false(p == NULL))
1986 p = vm_page_alloc_noobj(flags);
1989 if (__predict_false(p == NULL))
1991 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1993 if ((addr = kva_alloc(bytes)) == 0)
1996 TAILQ_FOREACH(p, &alloctail, listq) {
1997 pmap_qenter(zkva, &p, 1);
2000 return ((void*)addr);
2002 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2003 vm_page_unwire_noq(p);
2010 * Allocates a number of pages not belonging to a VM object
2013 * bytes The number of bytes requested
2014 * wait Shall we wait?
2017 * A pointer to the alloced memory or possibly
2018 * NULL if M_NOWAIT is set.
2021 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
2024 TAILQ_HEAD(, vm_page) alloctail;
2026 vm_offset_t retkva, zkva;
2027 vm_page_t p, p_next;
2031 TAILQ_INIT(&alloctail);
2033 req = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
2034 if ((wait & M_WAITOK) != 0)
2035 req |= VM_ALLOC_WAITOK;
2037 npages = howmany(bytes, PAGE_SIZE);
2038 while (npages > 0) {
2039 p = vm_page_alloc_noobj_domain(domain, req);
2042 * Since the page does not belong to an object, its
2045 TAILQ_INSERT_TAIL(&alloctail, p, listq);
2050 * Page allocation failed, free intermediate pages and
2053 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2054 vm_page_unwire_noq(p);
2059 *flags = UMA_SLAB_PRIV;
2060 zkva = keg->uk_kva +
2061 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
2063 TAILQ_FOREACH(p, &alloctail, listq) {
2064 pmap_qenter(zkva, &p, 1);
2068 return ((void *)retkva);
2072 * Allocate physically contiguous pages.
2075 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
2079 *pflag = UMA_SLAB_KERNEL;
2080 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
2081 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
2085 * Frees a number of pages to the system
2088 * mem A pointer to the memory to be freed
2089 * size The size of the memory being freed
2090 * flags The original p->us_flags field
2096 page_free(void *mem, vm_size_t size, uint8_t flags)
2099 if ((flags & UMA_SLAB_BOOT) != 0) {
2100 startup_free(mem, size);
2104 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
2105 ("UMA: page_free used with invalid flags %x", flags));
2107 kmem_free(mem, size);
2111 * Frees pcpu zone allocations
2114 * mem A pointer to the memory to be freed
2115 * size The size of the memory being freed
2116 * flags The original p->us_flags field
2122 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
2124 vm_offset_t sva, curva;
2128 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
2130 if ((flags & UMA_SLAB_BOOT) != 0) {
2131 startup_free(mem, size);
2135 sva = (vm_offset_t)mem;
2136 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
2137 paddr = pmap_kextract(curva);
2138 m = PHYS_TO_VM_PAGE(paddr);
2139 vm_page_unwire_noq(m);
2142 pmap_qremove(sva, size >> PAGE_SHIFT);
2143 kva_free(sva, size);
2147 * Zero fill initializer
2149 * Arguments/Returns follow uma_init specifications
2152 zero_init(void *mem, int size, int flags)
2159 static struct noslabbits *
2160 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2163 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2168 * Actual size of embedded struct slab (!OFFPAGE).
2171 slab_sizeof(int nitems)
2175 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2176 return (roundup(s, UMA_ALIGN_PTR + 1));
2179 #define UMA_FIXPT_SHIFT 31
2180 #define UMA_FRAC_FIXPT(n, d) \
2181 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2182 #define UMA_FIXPT_PCT(f) \
2183 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2184 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
2185 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2188 * Compute the number of items that will fit in a slab. If hdr is true, the
2189 * item count may be limited to provide space in the slab for an inline slab
2190 * header. Otherwise, all slab space will be provided for item storage.
2193 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2198 /* The padding between items is not needed after the last item. */
2199 padpi = rsize - size;
2203 * Start with the maximum item count and remove items until
2204 * the slab header first alongside the allocatable memory.
2206 for (ipers = MIN(SLAB_MAX_SETSIZE,
2207 (slabsize + padpi - slab_sizeof(1)) / rsize);
2209 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2213 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2219 struct keg_layout_result {
2227 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2228 struct keg_layout_result *kl)
2233 kl->slabsize = slabsize;
2235 /* Handle INTERNAL as inline with an extra page. */
2236 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2237 kl->format &= ~UMA_ZFLAG_INTERNAL;
2238 kl->slabsize += PAGE_SIZE;
2241 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2242 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2244 /* Account for memory used by an offpage slab header. */
2245 total = kl->slabsize;
2246 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2247 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2249 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2253 * Determine the format of a uma keg. This determines where the slab header
2254 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2257 * keg The zone we should initialize
2263 keg_layout(uma_keg_t keg)
2265 struct keg_layout_result kl = {}, kl_tmp;
2274 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2275 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2276 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2277 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2278 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2280 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2281 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2282 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2285 alignsize = keg->uk_align + 1;
2288 * ASAN requires that each allocation be aligned to the shadow map
2291 if (alignsize < KASAN_SHADOW_SCALE)
2292 alignsize = KASAN_SHADOW_SCALE;
2296 * Calculate the size of each allocation (rsize) according to
2297 * alignment. If the requested size is smaller than we have
2298 * allocation bits for we round it up.
2300 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2301 rsize = roundup2(rsize, alignsize);
2303 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2305 * We want one item to start on every align boundary in a page.
2306 * To do this we will span pages. We will also extend the item
2307 * by the size of align if it is an even multiple of align.
2308 * Otherwise, it would fall on the same boundary every time.
2310 if ((rsize & alignsize) == 0)
2312 slabsize = rsize * (PAGE_SIZE / alignsize);
2313 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2314 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2315 slabsize = round_page(slabsize);
2318 * Start with a slab size of as many pages as it takes to
2319 * represent a single item. We will try to fit as many
2320 * additional items into the slab as possible.
2322 slabsize = round_page(keg->uk_size);
2325 /* Build a list of all of the available formats for this keg. */
2328 /* Evaluate an inline slab layout. */
2329 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2332 /* TODO: vm_page-embedded slab. */
2335 * We can't do OFFPAGE if we're internal or if we've been
2336 * asked to not go to the VM for buckets. If we do this we
2337 * may end up going to the VM for slabs which we do not want
2338 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2339 * In those cases, evaluate a pseudo-format called INTERNAL
2340 * which has an inline slab header and one extra page to
2341 * guarantee that it fits.
2343 * Otherwise, see if using an OFFPAGE slab will improve our
2346 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2347 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2349 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2352 * Choose a slab size and format which satisfy the minimum efficiency.
2353 * Prefer the smallest slab size that meets the constraints.
2355 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2356 * for small items (up to PAGE_SIZE), the iteration increment is one
2357 * page; and for large items, the increment is one item.
2359 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2360 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2361 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2364 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2365 round_page(rsize * (i - 1) + keg->uk_size);
2367 for (j = 0; j < nfmt; j++) {
2368 /* Only if we have no viable format yet. */
2369 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2373 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2374 if (kl_tmp.eff <= kl.eff)
2379 CTR6(KTR_UMA, "keg %s layout: format %#x "
2380 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2381 keg->uk_name, kl.format, kl.ipers, rsize,
2382 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2384 /* Stop when we reach the minimum efficiency. */
2385 if (kl.eff >= UMA_MIN_EFF)
2389 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2390 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2391 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2395 pages = atop(kl.slabsize);
2396 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2397 pages *= mp_maxid + 1;
2399 keg->uk_rsize = rsize;
2400 keg->uk_ipers = kl.ipers;
2401 keg->uk_ppera = pages;
2402 keg->uk_flags |= kl.format;
2405 * How do we find the slab header if it is offpage or if not all item
2406 * start addresses are in the same page? We could solve the latter
2407 * case with vaddr alignment, but we don't.
2409 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2410 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2411 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2412 keg->uk_flags |= UMA_ZFLAG_HASH;
2414 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2417 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2418 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2420 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2421 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2422 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2423 keg->uk_ipers, pages));
2427 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2428 * the keg onto the global keg list.
2430 * Arguments/Returns follow uma_ctor specifications
2431 * udata Actually uma_kctor_args
2434 keg_ctor(void *mem, int size, void *udata, int flags)
2436 struct uma_kctor_args *arg = udata;
2437 uma_keg_t keg = mem;
2442 keg->uk_size = arg->size;
2443 keg->uk_init = arg->uminit;
2444 keg->uk_fini = arg->fini;
2445 keg->uk_align = arg->align;
2446 keg->uk_reserve = 0;
2447 keg->uk_flags = arg->flags;
2450 * We use a global round-robin policy by default. Zones with
2451 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2452 * case the iterator is never run.
2454 keg->uk_dr.dr_policy = DOMAINSET_RR();
2455 keg->uk_dr.dr_iter = 0;
2458 * The primary zone is passed to us at keg-creation time.
2461 keg->uk_name = zone->uz_name;
2463 if (arg->flags & UMA_ZONE_ZINIT)
2464 keg->uk_init = zero_init;
2466 if (arg->flags & UMA_ZONE_MALLOC)
2467 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2470 keg->uk_flags &= ~UMA_ZONE_PCPU;
2476 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2477 * work on. Use round-robin for everything else.
2479 * Zones may override the default by specifying either.
2482 if ((keg->uk_flags &
2483 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2484 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2485 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2486 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2490 * If we haven't booted yet we need allocations to go through the
2491 * startup cache until the vm is ready.
2493 #ifdef UMA_MD_SMALL_ALLOC
2494 if (keg->uk_ppera == 1)
2495 keg->uk_allocf = uma_small_alloc;
2498 if (booted < BOOT_KVA)
2499 keg->uk_allocf = startup_alloc;
2500 else if (keg->uk_flags & UMA_ZONE_PCPU)
2501 keg->uk_allocf = pcpu_page_alloc;
2502 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2503 keg->uk_allocf = contig_alloc;
2505 keg->uk_allocf = page_alloc;
2506 #ifdef UMA_MD_SMALL_ALLOC
2507 if (keg->uk_ppera == 1)
2508 keg->uk_freef = uma_small_free;
2511 if (keg->uk_flags & UMA_ZONE_PCPU)
2512 keg->uk_freef = pcpu_page_free;
2514 keg->uk_freef = page_free;
2517 * Initialize keg's locks.
2519 for (i = 0; i < vm_ndomains; i++)
2520 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2523 * If we're putting the slab header in the actual page we need to
2524 * figure out where in each page it goes. See slab_sizeof
2527 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2530 shsize = slab_sizeof(keg->uk_ipers);
2531 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2533 * The only way the following is possible is if with our
2534 * UMA_ALIGN_PTR adjustments we are now bigger than
2535 * UMA_SLAB_SIZE. I haven't checked whether this is
2536 * mathematically possible for all cases, so we make
2539 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2540 ("zone %s ipers %d rsize %d size %d slab won't fit",
2541 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2544 if (keg->uk_flags & UMA_ZFLAG_HASH)
2545 hash_alloc(&keg->uk_hash, 0);
2547 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2549 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2551 rw_wlock(&uma_rwlock);
2552 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2553 rw_wunlock(&uma_rwlock);
2558 zone_kva_available(uma_zone_t zone, void *unused)
2562 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2566 if (keg->uk_allocf == startup_alloc) {
2567 /* Switch to the real allocator. */
2568 if (keg->uk_flags & UMA_ZONE_PCPU)
2569 keg->uk_allocf = pcpu_page_alloc;
2570 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2572 keg->uk_allocf = contig_alloc;
2574 keg->uk_allocf = page_alloc;
2579 zone_alloc_counters(uma_zone_t zone, void *unused)
2582 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2583 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2584 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2585 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2589 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2591 uma_zone_domain_t zdom;
2594 struct sysctl_oid *oid, *domainoid;
2595 int domains, i, cnt;
2596 static const char *nokeg = "cache zone";
2600 * Make a sysctl safe copy of the zone name by removing
2601 * any special characters and handling dups by appending
2604 if (zone->uz_namecnt != 0) {
2605 /* Count the number of decimal digits and '_' separator. */
2606 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2608 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2610 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2613 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2614 for (c = zone->uz_ctlname; *c != '\0'; c++)
2615 if (strchr("./\\ -", *c) != NULL)
2619 * Basic parameters at the root.
2621 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2622 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2624 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2625 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2626 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2627 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2628 zone, 0, sysctl_handle_uma_zone_flags, "A",
2629 "Allocator configuration flags");
2630 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2631 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2632 "Desired per-cpu cache size");
2633 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2634 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2635 "Maximum allowed per-cpu cache size");
2640 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2641 domains = vm_ndomains;
2644 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2645 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2647 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2648 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2649 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2650 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2651 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2652 "Real object size with alignment");
2653 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2654 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2655 "pages per-slab allocation");
2656 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2657 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2658 "items available per-slab");
2659 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2660 "align", CTLFLAG_RD, &keg->uk_align, 0,
2661 "item alignment mask");
2662 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2663 "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2664 "number of reserved items");
2665 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2666 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2667 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2668 "Slab utilization (100 - internal fragmentation %)");
2669 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2670 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2671 for (i = 0; i < domains; i++) {
2672 dom = &keg->uk_domain[i];
2673 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2674 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2675 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2676 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2677 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2678 "Total pages currently allocated from VM");
2679 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2680 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2681 "Items free in the slab layer");
2682 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2683 "free_slabs", CTLFLAG_RD, &dom->ud_free_slabs, 0,
2687 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2688 "name", CTLFLAG_RD, nokeg, "Keg name");
2691 * Information about zone limits.
2693 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2694 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2695 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2696 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2697 zone, 0, sysctl_handle_uma_zone_items, "QU",
2698 "Current number of allocated items if limit is set");
2699 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2700 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2701 "Maximum number of allocated and cached items");
2702 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2703 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2704 "Number of threads sleeping at limit");
2705 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2706 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2707 "Total zone limit sleeps");
2708 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2709 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2710 "Maximum number of items in each domain's bucket cache");
2713 * Per-domain zone information.
2715 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2716 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2717 for (i = 0; i < domains; i++) {
2718 zdom = ZDOM_GET(zone, i);
2719 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2720 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2721 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2722 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2723 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2724 "number of items in this domain");
2725 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2726 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2727 "maximum item count in this period");
2728 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2729 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2730 "minimum item count in this period");
2731 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2732 "bimin", CTLFLAG_RD, &zdom->uzd_bimin,
2733 "Minimum item count in this batch");
2734 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2735 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2736 "Working set size");
2737 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2738 "limin", CTLFLAG_RD, &zdom->uzd_limin,
2739 "Long time minimum item count");
2740 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2741 "timin", CTLFLAG_RD, &zdom->uzd_timin, 0,
2742 "Time since zero long time minimum item count");
2746 * General statistics.
2748 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2749 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2750 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2751 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2752 zone, 1, sysctl_handle_uma_zone_cur, "I",
2753 "Current number of allocated items");
2754 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2755 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2756 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2757 "Total allocation calls");
2758 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2759 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2760 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2761 "Total free calls");
2762 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2763 "fails", CTLFLAG_RD, &zone->uz_fails,
2764 "Number of allocation failures");
2765 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2766 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2767 "Free calls from the wrong domain");
2770 struct uma_zone_count {
2776 zone_count(uma_zone_t zone, void *arg)
2778 struct uma_zone_count *cnt;
2782 * Some zones are rapidly created with identical names and
2783 * destroyed out of order. This can lead to gaps in the count.
2784 * Use one greater than the maximum observed for this name.
2786 if (strcmp(zone->uz_name, cnt->name) == 0)
2787 cnt->count = MAX(cnt->count,
2788 zone->uz_namecnt + 1);
2792 zone_update_caches(uma_zone_t zone)
2796 for (i = 0; i <= mp_maxid; i++) {
2797 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2798 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2803 * Zone header ctor. This initializes all fields, locks, etc.
2805 * Arguments/Returns follow uma_ctor specifications
2806 * udata Actually uma_zctor_args
2809 zone_ctor(void *mem, int size, void *udata, int flags)
2811 struct uma_zone_count cnt;
2812 struct uma_zctor_args *arg = udata;
2813 uma_zone_domain_t zdom;
2814 uma_zone_t zone = mem;
2820 zone->uz_name = arg->name;
2821 zone->uz_ctor = arg->ctor;
2822 zone->uz_dtor = arg->dtor;
2823 zone->uz_init = NULL;
2824 zone->uz_fini = NULL;
2825 zone->uz_sleeps = 0;
2826 zone->uz_bucket_size = 0;
2827 zone->uz_bucket_size_min = 0;
2828 zone->uz_bucket_size_max = BUCKET_MAX;
2829 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2830 zone->uz_warning = NULL;
2831 /* The domain structures follow the cpu structures. */
2832 zone->uz_bucket_max = ULONG_MAX;
2833 timevalclear(&zone->uz_ratecheck);
2835 /* Count the number of duplicate names. */
2836 cnt.name = arg->name;
2838 zone_foreach(zone_count, &cnt);
2839 zone->uz_namecnt = cnt.count;
2840 ZONE_CROSS_LOCK_INIT(zone);
2842 for (i = 0; i < vm_ndomains; i++) {
2843 zdom = ZDOM_GET(zone, i);
2844 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2845 STAILQ_INIT(&zdom->uzd_buckets);
2848 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
2849 if (arg->uminit == trash_init && arg->fini == trash_fini)
2850 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2851 #elif defined(KASAN)
2852 if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0)
2853 arg->flags |= UMA_ZONE_NOKASAN;
2857 * This is a pure cache zone, no kegs.
2860 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2861 ("zone_ctor: Import specified for non-cache zone."));
2862 zone->uz_flags = arg->flags;
2863 zone->uz_size = arg->size;
2864 zone->uz_import = arg->import;
2865 zone->uz_release = arg->release;
2866 zone->uz_arg = arg->arg;
2869 * Cache zones are round-robin unless a policy is
2870 * specified because they may have incompatible
2873 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2874 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2876 rw_wlock(&uma_rwlock);
2877 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2878 rw_wunlock(&uma_rwlock);
2883 * Use the regular zone/keg/slab allocator.
2885 zone->uz_import = zone_import;
2886 zone->uz_release = zone_release;
2887 zone->uz_arg = zone;
2890 if (arg->flags & UMA_ZONE_SECONDARY) {
2891 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2892 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2893 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2894 zone->uz_init = arg->uminit;
2895 zone->uz_fini = arg->fini;
2896 zone->uz_flags |= UMA_ZONE_SECONDARY;
2897 rw_wlock(&uma_rwlock);
2899 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2900 if (LIST_NEXT(z, uz_link) == NULL) {
2901 LIST_INSERT_AFTER(z, zone, uz_link);
2906 rw_wunlock(&uma_rwlock);
2907 } else if (keg == NULL) {
2908 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2909 arg->align, arg->flags)) == NULL)
2912 struct uma_kctor_args karg;
2915 /* We should only be here from uma_startup() */
2916 karg.size = arg->size;
2917 karg.uminit = arg->uminit;
2918 karg.fini = arg->fini;
2919 karg.align = arg->align;
2920 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2922 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2928 /* Inherit properties from the keg. */
2930 zone->uz_size = keg->uk_size;
2931 zone->uz_flags |= (keg->uk_flags &
2932 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2935 if (booted >= BOOT_PCPU) {
2936 zone_alloc_counters(zone, NULL);
2937 if (booted >= BOOT_RUNNING)
2938 zone_alloc_sysctl(zone, NULL);
2940 zone->uz_allocs = EARLY_COUNTER;
2941 zone->uz_frees = EARLY_COUNTER;
2942 zone->uz_fails = EARLY_COUNTER;
2945 /* Caller requests a private SMR context. */
2946 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2947 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2949 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2950 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2951 ("Invalid zone flag combination"));
2952 if (arg->flags & UMA_ZFLAG_INTERNAL)
2953 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2954 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2955 zone->uz_bucket_size = BUCKET_MAX;
2956 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2957 zone->uz_bucket_size = 0;
2959 zone->uz_bucket_size = bucket_select(zone->uz_size);
2960 zone->uz_bucket_size_min = zone->uz_bucket_size;
2961 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2962 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2963 zone_update_caches(zone);
2969 * Keg header dtor. This frees all data, destroys locks, frees the hash
2970 * table and removes the keg from the global list.
2972 * Arguments/Returns follow uma_dtor specifications
2976 keg_dtor(void *arg, int size, void *udata)
2979 uint32_t free, pages;
2982 keg = (uma_keg_t)arg;
2984 for (i = 0; i < vm_ndomains; i++) {
2985 free += keg->uk_domain[i].ud_free_items;
2986 pages += keg->uk_domain[i].ud_pages;
2987 KEG_LOCK_FINI(keg, i);
2990 printf("Freed UMA keg (%s) was not empty (%u items). "
2991 " Lost %u pages of memory.\n",
2992 keg->uk_name ? keg->uk_name : "",
2993 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2995 hash_free(&keg->uk_hash);
3001 * Arguments/Returns follow uma_dtor specifications
3005 zone_dtor(void *arg, int size, void *udata)
3011 zone = (uma_zone_t)arg;
3013 sysctl_remove_oid(zone->uz_oid, 1, 1);
3015 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
3018 rw_wlock(&uma_rwlock);
3019 LIST_REMOVE(zone, uz_link);
3020 rw_wunlock(&uma_rwlock);
3021 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3023 keg->uk_reserve = 0;
3025 zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true);
3028 * We only destroy kegs from non secondary/non cache zones.
3030 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3032 rw_wlock(&uma_rwlock);
3033 LIST_REMOVE(keg, uk_link);
3034 rw_wunlock(&uma_rwlock);
3035 zone_free_item(kegs, keg, NULL, SKIP_NONE);
3037 counter_u64_free(zone->uz_allocs);
3038 counter_u64_free(zone->uz_frees);
3039 counter_u64_free(zone->uz_fails);
3040 counter_u64_free(zone->uz_xdomain);
3041 free(zone->uz_ctlname, M_UMA);
3042 for (i = 0; i < vm_ndomains; i++)
3043 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
3044 ZONE_CROSS_LOCK_FINI(zone);
3048 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3053 LIST_FOREACH(keg, &uma_kegs, uk_link) {
3054 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
3057 LIST_FOREACH(zone, &uma_cachezones, uz_link)
3062 * Traverses every zone in the system and calls a callback
3065 * zfunc A pointer to a function which accepts a zone
3072 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3075 rw_rlock(&uma_rwlock);
3076 zone_foreach_unlocked(zfunc, arg);
3077 rw_runlock(&uma_rwlock);
3081 * Initialize the kernel memory allocator. This is done after pages can be
3082 * allocated but before general KVA is available.
3085 uma_startup1(vm_offset_t virtual_avail)
3087 struct uma_zctor_args args;
3088 size_t ksize, zsize, size;
3089 uma_keg_t primarykeg;
3094 bootstart = bootmem = virtual_avail;
3096 rw_init(&uma_rwlock, "UMA lock");
3097 sx_init(&uma_reclaim_lock, "umareclaim");
3099 ksize = sizeof(struct uma_keg) +
3100 (sizeof(struct uma_domain) * vm_ndomains);
3101 ksize = roundup(ksize, UMA_SUPER_ALIGN);
3102 zsize = sizeof(struct uma_zone) +
3103 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
3104 (sizeof(struct uma_zone_domain) * vm_ndomains);
3105 zsize = roundup(zsize, UMA_SUPER_ALIGN);
3107 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
3108 size = (zsize * 2) + ksize;
3109 for (domain = 0; domain < vm_ndomains; domain++) {
3110 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
3115 zones = (uma_zone_t)m;
3117 kegs = (uma_zone_t)m;
3119 primarykeg = (uma_keg_t)m;
3121 /* "manually" create the initial zone */
3122 memset(&args, 0, sizeof(args));
3123 args.name = "UMA Kegs";
3125 args.ctor = keg_ctor;
3126 args.dtor = keg_dtor;
3127 args.uminit = zero_init;
3129 args.keg = primarykeg;
3130 args.align = UMA_SUPER_ALIGN - 1;
3131 args.flags = UMA_ZFLAG_INTERNAL;
3132 zone_ctor(kegs, zsize, &args, M_WAITOK);
3134 args.name = "UMA Zones";
3136 args.ctor = zone_ctor;
3137 args.dtor = zone_dtor;
3138 args.uminit = zero_init;
3141 args.align = UMA_SUPER_ALIGN - 1;
3142 args.flags = UMA_ZFLAG_INTERNAL;
3143 zone_ctor(zones, zsize, &args, M_WAITOK);
3145 /* Now make zones for slab headers */
3146 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
3147 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3148 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
3149 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3151 hashzone = uma_zcreate("UMA Hash",
3152 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
3153 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3159 #ifndef UMA_MD_SMALL_ALLOC
3160 extern void vm_radix_reserve_kva(void);
3164 * Advertise the availability of normal kva allocations and switch to
3165 * the default back-end allocator. Marks the KVA we consumed on startup
3166 * as used in the map.
3172 if (bootstart != bootmem) {
3173 vm_map_lock(kernel_map);
3174 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
3175 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
3176 vm_map_unlock(kernel_map);
3179 #ifndef UMA_MD_SMALL_ALLOC
3180 /* Set up radix zone to use noobj_alloc. */
3181 vm_radix_reserve_kva();
3185 zone_foreach_unlocked(zone_kva_available, NULL);
3190 * Allocate counters as early as possible so that boot-time allocations are
3191 * accounted more precisely.
3194 uma_startup_pcpu(void *arg __unused)
3197 zone_foreach_unlocked(zone_alloc_counters, NULL);
3200 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
3203 * Finish our initialization steps.
3206 uma_startup3(void *arg __unused)
3210 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
3211 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
3212 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
3214 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
3215 booted = BOOT_RUNNING;
3217 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
3218 EVENTHANDLER_PRI_FIRST);
3220 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
3223 uma_startup4(void *arg __unused)
3225 TIMEOUT_TASK_INIT(taskqueue_thread, &uma_timeout_task, 0, uma_timeout,
3227 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task,
3230 SYSINIT(uma_startup4, SI_SUB_TASKQ, SI_ORDER_ANY, uma_startup4, NULL);
3236 booted = BOOT_SHUTDOWN;
3240 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3241 int align, uint32_t flags)
3243 struct uma_kctor_args args;
3246 args.uminit = uminit;
3248 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
3251 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3254 /* Public functions */
3257 uma_set_align(int align)
3260 if (align != UMA_ALIGN_CACHE)
3261 uma_align_cache = align;
3266 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3267 uma_init uminit, uma_fini fini, int align, uint32_t flags)
3270 struct uma_zctor_args args;
3273 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
3276 /* This stuff is essential for the zone ctor */
3277 memset(&args, 0, sizeof(args));
3282 args.uminit = uminit;
3284 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
3286 * Inject procedures which check for memory use after free if we are
3287 * allowed to scramble the memory while it is not allocated. This
3288 * requires that: UMA is actually able to access the memory, no init
3289 * or fini procedures, no dependency on the initial value of the
3290 * memory, and no (legitimate) use of the memory after free. Note,
3291 * the ctor and dtor do not need to be empty.
3293 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3294 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3295 args.uminit = trash_init;
3296 args.fini = trash_fini;
3303 sx_xlock(&uma_reclaim_lock);
3304 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3305 sx_xunlock(&uma_reclaim_lock);
3312 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3313 uma_init zinit, uma_fini zfini, uma_zone_t primary)
3315 struct uma_zctor_args args;
3319 keg = primary->uz_keg;
3320 memset(&args, 0, sizeof(args));
3322 args.size = keg->uk_size;
3325 args.uminit = zinit;
3327 args.align = keg->uk_align;
3328 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3331 sx_xlock(&uma_reclaim_lock);
3332 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3333 sx_xunlock(&uma_reclaim_lock);
3340 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3341 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3342 void *arg, int flags)
3344 struct uma_zctor_args args;
3346 memset(&args, 0, sizeof(args));
3351 args.uminit = zinit;
3353 args.import = zimport;
3354 args.release = zrelease;
3357 args.flags = flags | UMA_ZFLAG_CACHE;
3359 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3364 uma_zdestroy(uma_zone_t zone)
3368 * Large slabs are expensive to reclaim, so don't bother doing
3369 * unnecessary work if we're shutting down.
3371 if (booted == BOOT_SHUTDOWN &&
3372 zone->uz_fini == NULL && zone->uz_release == zone_release)
3374 sx_xlock(&uma_reclaim_lock);
3375 zone_free_item(zones, zone, NULL, SKIP_NONE);
3376 sx_xunlock(&uma_reclaim_lock);
3380 uma_zwait(uma_zone_t zone)
3383 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3384 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3385 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3386 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3388 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3392 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3394 void *item, *pcpu_item;
3398 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3400 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3403 pcpu_item = zpcpu_base_to_offset(item);
3404 if (flags & M_ZERO) {
3406 for (i = 0; i <= mp_maxid; i++)
3407 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3409 bzero(item, zone->uz_size);
3416 * A stub while both regular and pcpu cases are identical.
3419 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3424 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3427 /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */
3428 if (pcpu_item == NULL)
3431 item = zpcpu_offset_to_base(pcpu_item);
3432 uma_zfree_arg(zone, item, udata);
3435 static inline void *
3436 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3443 kasan_mark_item_valid(zone, item);
3444 kmsan_mark_item_uninitialized(zone, item);
3447 skipdbg = uma_dbg_zskip(zone, item);
3448 if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3449 zone->uz_ctor != trash_ctor)
3450 trash_ctor(item, size, udata, flags);
3453 /* Check flags before loading ctor pointer. */
3454 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3455 __predict_false(zone->uz_ctor != NULL) &&
3456 zone->uz_ctor(item, size, udata, flags) != 0) {
3457 counter_u64_add(zone->uz_fails, 1);
3458 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3463 uma_dbg_alloc(zone, NULL, item);
3465 if (__predict_false(flags & M_ZERO))
3466 return (memset(item, 0, size));
3472 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3473 enum zfreeskip skip)
3478 skipdbg = uma_dbg_zskip(zone, item);
3479 if (skip == SKIP_NONE && !skipdbg) {
3480 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3481 uma_dbg_free(zone, udata, item);
3483 uma_dbg_free(zone, NULL, item);
3486 if (__predict_true(skip < SKIP_DTOR)) {
3487 if (zone->uz_dtor != NULL)
3488 zone->uz_dtor(item, size, udata);
3490 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3491 zone->uz_dtor != trash_dtor)
3492 trash_dtor(item, size, udata);
3495 kasan_mark_item_invalid(zone, item);
3500 item_domain(void *item)
3504 domain = vm_phys_domain(vtophys(item));
3505 KASSERT(domain >= 0 && domain < vm_ndomains,
3506 ("%s: unknown domain for item %p", __func__, item));
3511 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3512 #if defined(INVARIANTS) && (defined(DDB) || defined(STACK))
3513 #include <sys/stack.h>
3515 #define UMA_ZALLOC_DEBUG
3517 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3523 if (flags & M_WAITOK) {
3524 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3525 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3530 KASSERT((flags & M_EXEC) == 0,
3531 ("uma_zalloc_debug: called with M_EXEC"));
3532 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3533 ("uma_zalloc_debug: called within spinlock or critical section"));
3534 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3535 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3537 _Static_assert(M_NOWAIT != 0 && M_WAITOK != 0,
3538 "M_NOWAIT and M_WAITOK must be non-zero for this assertion:");
3541 * Give the #elif clause time to find problems, then remove it
3542 * and enable this. (Remove <sys/stack.h> above, too.)
3544 KASSERT((flags & (M_NOWAIT|M_WAITOK)) == M_NOWAIT ||
3545 (flags & (M_NOWAIT|M_WAITOK)) == M_WAITOK,
3546 ("uma_zalloc_debug: must pass one of M_NOWAIT or M_WAITOK"));
3547 #elif defined(DDB) || defined(STACK)
3548 if (__predict_false((flags & (M_NOWAIT|M_WAITOK)) != M_NOWAIT &&
3549 (flags & (M_NOWAIT|M_WAITOK)) != M_WAITOK)) {
3550 static int stack_count;
3553 if (stack_count < 10) {
3555 printf("uma_zalloc* called with bad WAIT flags:\n");
3563 #ifdef DEBUG_MEMGUARD
3564 if ((zone->uz_flags & (UMA_ZONE_SMR | UMA_ZFLAG_CACHE)) == 0 &&
3565 memguard_cmp_zone(zone)) {
3567 item = memguard_alloc(zone->uz_size, flags);
3569 error = EJUSTRETURN;
3570 if (zone->uz_init != NULL &&
3571 zone->uz_init(item, zone->uz_size, flags) != 0) {
3575 if (zone->uz_ctor != NULL &&
3576 zone->uz_ctor(item, zone->uz_size, udata,
3578 counter_u64_add(zone->uz_fails, 1);
3579 if (zone->uz_fini != NULL)
3580 zone->uz_fini(item, zone->uz_size);
3587 /* This is unfortunate but should not be fatal. */
3594 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3596 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3597 ("uma_zfree_debug: called with spinlock or critical section held"));
3599 #ifdef DEBUG_MEMGUARD
3600 if ((zone->uz_flags & (UMA_ZONE_SMR | UMA_ZFLAG_CACHE)) == 0 &&
3601 is_memguard_addr(item)) {
3602 if (zone->uz_dtor != NULL)
3603 zone->uz_dtor(item, zone->uz_size, udata);
3604 if (zone->uz_fini != NULL)
3605 zone->uz_fini(item, zone->uz_size);
3606 memguard_free(item);
3607 return (EJUSTRETURN);
3614 static inline void *
3615 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3616 void *udata, int flags)
3621 item = cache_bucket_pop(cache, bucket);
3622 size = cache_uz_size(cache);
3623 uz_flags = cache_uz_flags(cache);
3625 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3628 static __noinline void *
3629 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3631 uma_cache_bucket_t bucket;
3634 while (cache_alloc(zone, cache, udata, flags)) {
3635 cache = &zone->uz_cpu[curcpu];
3636 bucket = &cache->uc_allocbucket;
3637 if (__predict_false(bucket->ucb_cnt == 0))
3639 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3644 * We can not get a bucket so try to return a single item.
3646 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3647 domain = PCPU_GET(domain);
3649 domain = UMA_ANYDOMAIN;
3650 return (zone_alloc_item(zone, udata, domain, flags));
3655 uma_zalloc_smr(uma_zone_t zone, int flags)
3657 uma_cache_bucket_t bucket;
3660 CTR3(KTR_UMA, "uma_zalloc_smr zone %s(%p) flags %d", zone->uz_name,
3663 #ifdef UMA_ZALLOC_DEBUG
3666 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3667 ("uma_zalloc_arg: called with non-SMR zone."));
3668 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3673 cache = &zone->uz_cpu[curcpu];
3674 bucket = &cache->uc_allocbucket;
3675 if (__predict_false(bucket->ucb_cnt == 0))
3676 return (cache_alloc_retry(zone, cache, NULL, flags));
3677 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3682 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3684 uma_cache_bucket_t bucket;
3687 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3688 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3690 /* This is the fast path allocation */
3691 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3694 #ifdef UMA_ZALLOC_DEBUG
3697 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3698 ("uma_zalloc_arg: called with SMR zone."));
3699 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3704 * If possible, allocate from the per-CPU cache. There are two
3705 * requirements for safe access to the per-CPU cache: (1) the thread
3706 * accessing the cache must not be preempted or yield during access,
3707 * and (2) the thread must not migrate CPUs without switching which
3708 * cache it accesses. We rely on a critical section to prevent
3709 * preemption and migration. We release the critical section in
3710 * order to acquire the zone mutex if we are unable to allocate from
3711 * the current cache; when we re-acquire the critical section, we
3712 * must detect and handle migration if it has occurred.
3715 cache = &zone->uz_cpu[curcpu];
3716 bucket = &cache->uc_allocbucket;
3717 if (__predict_false(bucket->ucb_cnt == 0))
3718 return (cache_alloc_retry(zone, cache, udata, flags));
3719 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3723 * Replenish an alloc bucket and possibly restore an old one. Called in
3724 * a critical section. Returns in a critical section.
3726 * A false return value indicates an allocation failure.
3727 * A true return value indicates success and the caller should retry.
3729 static __noinline bool
3730 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3732 uma_bucket_t bucket;
3733 int curdomain, domain;
3736 CRITICAL_ASSERT(curthread);
3739 * If we have run out of items in our alloc bucket see
3740 * if we can switch with the free bucket.
3742 * SMR Zones can't re-use the free bucket until the sequence has
3745 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3746 cache->uc_freebucket.ucb_cnt != 0) {
3747 cache_bucket_swap(&cache->uc_freebucket,
3748 &cache->uc_allocbucket);
3753 * Discard any empty allocation bucket while we hold no locks.
3755 bucket = cache_bucket_unload_alloc(cache);
3758 if (bucket != NULL) {
3759 KASSERT(bucket->ub_cnt == 0,
3760 ("cache_alloc: Entered with non-empty alloc bucket."));
3761 bucket_free(zone, bucket, udata);
3765 * Attempt to retrieve the item from the per-CPU cache has failed, so
3766 * we must go back to the zone. This requires the zdom lock, so we
3767 * must drop the critical section, then re-acquire it when we go back
3768 * to the cache. Since the critical section is released, we may be
3769 * preempted or migrate. As such, make sure not to maintain any
3770 * thread-local state specific to the cache from prior to releasing
3771 * the critical section.
3773 domain = PCPU_GET(domain);
3774 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3775 VM_DOMAIN_EMPTY(domain))
3776 domain = zone_domain_highest(zone, domain);
3777 bucket = cache_fetch_bucket(zone, cache, domain);
3778 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3779 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3785 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3786 zone->uz_name, zone, bucket);
3787 if (bucket == NULL) {
3793 * See if we lost the race or were migrated. Cache the
3794 * initialized bucket to make this less likely or claim
3795 * the memory directly.
3798 cache = &zone->uz_cpu[curcpu];
3799 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3800 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3801 (curdomain = PCPU_GET(domain)) == domain ||
3802 VM_DOMAIN_EMPTY(curdomain))) {
3804 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3806 cache_bucket_load_alloc(cache, bucket);
3811 * We lost the race, release this bucket and start over.
3814 zone_put_bucket(zone, domain, bucket, udata, !new);
3821 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3824 uma_bucket_t bucket;
3825 uma_zone_domain_t zdom;
3829 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3830 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3832 /* This is the fast path allocation */
3833 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3834 zone->uz_name, zone, domain, flags);
3836 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3837 ("uma_zalloc_domain: called with SMR zone."));
3839 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3840 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3842 if (vm_ndomains == 1)
3843 return (uma_zalloc_arg(zone, udata, flags));
3845 #ifdef UMA_ZALLOC_DEBUG
3846 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3851 * Try to allocate from the bucket cache before falling back to the keg.
3852 * We could try harder and attempt to allocate from per-CPU caches or
3853 * the per-domain cross-domain buckets, but the complexity is probably
3854 * not worth it. It is more important that frees of previous
3855 * cross-domain allocations do not blow up the cache.
3857 zdom = zone_domain_lock(zone, domain);
3858 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3859 item = bucket->ub_bucket[bucket->ub_cnt - 1];
3861 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3864 zone_put_bucket(zone, domain, bucket, udata, true);
3865 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3868 KASSERT(item_domain(item) == domain,
3869 ("%s: bucket cache item %p from wrong domain",
3871 counter_u64_add(zone->uz_allocs, 1);
3876 return (zone_alloc_item(zone, udata, domain, flags));
3878 return (uma_zalloc_arg(zone, udata, flags));
3883 * Find a slab with some space. Prefer slabs that are partially used over those
3884 * that are totally full. This helps to reduce fragmentation.
3886 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3890 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3896 KASSERT(domain >= 0 && domain < vm_ndomains,
3897 ("keg_first_slab: domain %d out of range", domain));
3898 KEG_LOCK_ASSERT(keg, domain);
3903 dom = &keg->uk_domain[domain];
3904 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3906 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3907 LIST_REMOVE(slab, us_link);
3908 dom->ud_free_slabs--;
3909 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3913 domain = (domain + 1) % vm_ndomains;
3914 } while (domain != start);
3920 * Fetch an existing slab from a free or partial list. Returns with the
3921 * keg domain lock held if a slab was found or unlocked if not.
3924 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3929 /* HASH has a single free list. */
3930 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3933 KEG_LOCK(keg, domain);
3934 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3935 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3936 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3937 KEG_UNLOCK(keg, domain);
3944 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3946 struct vm_domainset_iter di;
3951 KASSERT((flags & (M_WAITOK | M_NOVM)) != (M_WAITOK | M_NOVM),
3952 ("%s: invalid flags %#x", __func__, flags));
3956 * Use the keg's policy if upper layers haven't already specified a
3957 * domain (as happens with first-touch zones).
3959 * To avoid races we run the iterator with the keg lock held, but that
3960 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3961 * clear M_WAITOK and handle low memory conditions locally.
3963 rr = rdomain == UMA_ANYDOMAIN;
3965 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3966 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3974 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3979 * M_NOVM is used to break the recursion that can otherwise
3980 * occur if low-level memory management routines use UMA.
3982 if ((flags & M_NOVM) == 0) {
3983 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3989 if ((flags & M_USE_RESERVE) != 0) {
3991 * Drain reserves from other domains before
3992 * giving up or sleeping. It may be useful to
3993 * support per-domain reserves eventually.
3995 rdomain = UMA_ANYDOMAIN;
3998 if ((flags & M_WAITOK) == 0)
4000 vm_wait_domain(domain);
4001 } else if (vm_domainset_iter_policy(&di, &domain) != 0) {
4002 if ((flags & M_WAITOK) != 0) {
4003 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
4011 * We might not have been able to get a slab but another cpu
4012 * could have while we were unlocked. Check again before we
4015 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
4022 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
4028 KEG_LOCK_ASSERT(keg, slab->us_domain);
4030 dom = &keg->uk_domain[slab->us_domain];
4031 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
4032 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
4033 item = slab_item(slab, keg, freei);
4034 slab->us_freecount--;
4035 dom->ud_free_items--;
4038 * Move this slab to the full list. It must be on the partial list, so
4039 * we do not need to update the free slab count. In particular,
4040 * keg_fetch_slab() always returns slabs on the partial list.
4042 if (slab->us_freecount == 0) {
4043 LIST_REMOVE(slab, us_link);
4044 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
4051 zone_import(void *arg, void **bucket, int max, int domain, int flags)
4065 /* Try to keep the buckets totally full */
4066 for (i = 0; i < max; ) {
4067 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
4070 stripe = howmany(max, vm_ndomains);
4072 dom = &keg->uk_domain[slab->us_domain];
4074 bucket[i++] = slab_alloc_item(keg, slab);
4075 if (keg->uk_reserve > 0 &&
4076 dom->ud_free_items <= keg->uk_reserve) {
4078 * Avoid depleting the reserve after a
4079 * successful item allocation, even if
4080 * M_USE_RESERVE is specified.
4082 KEG_UNLOCK(keg, slab->us_domain);
4087 * If the zone is striped we pick a new slab for every
4088 * N allocations. Eliminating this conditional will
4089 * instead pick a new domain for each bucket rather
4090 * than stripe within each bucket. The current option
4091 * produces more fragmentation and requires more cpu
4092 * time but yields better distribution.
4094 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
4095 vm_ndomains > 1 && --stripe == 0)
4098 } while (slab->us_freecount != 0 && i < max);
4099 KEG_UNLOCK(keg, slab->us_domain);
4101 /* Don't block if we allocated any successfully. */
4110 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
4112 uint64_t old, new, total, max;
4115 * The hard case. We're going to sleep because there were existing
4116 * sleepers or because we ran out of items. This routine enforces
4117 * fairness by keeping fifo order.
4119 * First release our ill gotten gains and make some noise.
4122 zone_free_limit(zone, count);
4123 zone_log_warning(zone);
4124 zone_maxaction(zone);
4125 if (flags & M_NOWAIT)
4129 * We need to allocate an item or set ourself as a sleeper
4130 * while the sleepq lock is held to avoid wakeup races. This
4131 * is essentially a home rolled semaphore.
4133 sleepq_lock(&zone->uz_max_items);
4134 old = zone->uz_items;
4136 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
4137 /* Cache the max since we will evaluate twice. */
4138 max = zone->uz_max_items;
4139 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
4140 UZ_ITEMS_COUNT(old) >= max)
4141 new = old + UZ_ITEMS_SLEEPER;
4143 new = old + MIN(count, max - old);
4144 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
4146 /* We may have successfully allocated under the sleepq lock. */
4147 if (UZ_ITEMS_SLEEPERS(new) == 0) {
4148 sleepq_release(&zone->uz_max_items);
4153 * This is in a different cacheline from uz_items so that we
4154 * don't constantly invalidate the fastpath cacheline when we
4155 * adjust item counts. This could be limited to toggling on
4158 atomic_add_32(&zone->uz_sleepers, 1);
4159 atomic_add_64(&zone->uz_sleeps, 1);
4162 * We have added ourselves as a sleeper. The sleepq lock
4163 * protects us from wakeup races. Sleep now and then retry.
4165 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
4166 sleepq_wait(&zone->uz_max_items, PVM);
4169 * After wakeup, remove ourselves as a sleeper and try
4170 * again. We no longer have the sleepq lock for protection.
4172 * Subract ourselves as a sleeper while attempting to add
4175 atomic_subtract_32(&zone->uz_sleepers, 1);
4176 old = atomic_fetchadd_64(&zone->uz_items,
4177 -(UZ_ITEMS_SLEEPER - count));
4178 /* We're no longer a sleeper. */
4179 old -= UZ_ITEMS_SLEEPER;
4182 * If we're still at the limit, restart. Notably do not
4183 * block on other sleepers. Cache the max value to protect
4184 * against changes via sysctl.
4186 total = UZ_ITEMS_COUNT(old);
4187 max = zone->uz_max_items;
4190 /* Truncate if necessary, otherwise wake other sleepers. */
4191 if (total + count > max) {
4192 zone_free_limit(zone, total + count - max);
4193 count = max - total;
4194 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
4195 wakeup_one(&zone->uz_max_items);
4202 * Allocate 'count' items from our max_items limit. Returns the number
4203 * available. If M_NOWAIT is not specified it will sleep until at least
4204 * one item can be allocated.
4207 zone_alloc_limit(uma_zone_t zone, int count, int flags)
4212 max = zone->uz_max_items;
4216 * We expect normal allocations to succeed with a simple
4219 old = atomic_fetchadd_64(&zone->uz_items, count);
4220 if (__predict_true(old + count <= max))
4224 * If we had some items and no sleepers just return the
4225 * truncated value. We have to release the excess space
4226 * though because that may wake sleepers who weren't woken
4227 * because we were temporarily over the limit.
4230 zone_free_limit(zone, (old + count) - max);
4233 return (zone_alloc_limit_hard(zone, count, flags));
4237 * Free a number of items back to the limit.
4240 zone_free_limit(uma_zone_t zone, int count)
4247 * In the common case we either have no sleepers or
4248 * are still over the limit and can just return.
4250 old = atomic_fetchadd_64(&zone->uz_items, -count);
4251 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4252 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4256 * Moderate the rate of wakeups. Sleepers will continue
4257 * to generate wakeups if necessary.
4259 wakeup_one(&zone->uz_max_items);
4263 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4265 uma_bucket_t bucket;
4266 int error, maxbucket, cnt;
4268 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4271 /* Avoid allocs targeting empty domains. */
4272 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4273 domain = UMA_ANYDOMAIN;
4274 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4275 domain = UMA_ANYDOMAIN;
4277 if (zone->uz_max_items > 0)
4278 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4281 maxbucket = zone->uz_bucket_size;
4285 /* Don't wait for buckets, preserve caller's NOVM setting. */
4286 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4287 if (bucket == NULL) {
4292 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4293 MIN(maxbucket, bucket->ub_entries), domain, flags);
4296 * Initialize the memory if necessary.
4298 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4301 for (i = 0; i < bucket->ub_cnt; i++) {
4302 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
4303 error = zone->uz_init(bucket->ub_bucket[i],
4304 zone->uz_size, flags);
4305 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
4311 * If we couldn't initialize the whole bucket, put the
4312 * rest back onto the freelist.
4314 if (i != bucket->ub_cnt) {
4315 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4316 bucket->ub_cnt - i);
4318 bzero(&bucket->ub_bucket[i],
4319 sizeof(void *) * (bucket->ub_cnt - i));
4325 cnt = bucket->ub_cnt;
4326 if (bucket->ub_cnt == 0) {
4327 bucket_free(zone, bucket, udata);
4328 counter_u64_add(zone->uz_fails, 1);
4332 if (zone->uz_max_items > 0 && cnt < maxbucket)
4333 zone_free_limit(zone, maxbucket - cnt);
4339 * Allocates a single item from a zone.
4342 * zone The zone to alloc for.
4343 * udata The data to be passed to the constructor.
4344 * domain The domain to allocate from or UMA_ANYDOMAIN.
4345 * flags M_WAITOK, M_NOWAIT, M_ZERO.
4348 * NULL if there is no memory and M_NOWAIT is set
4349 * An item if successful
4353 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4357 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4358 counter_u64_add(zone->uz_fails, 1);
4362 /* Avoid allocs targeting empty domains. */
4363 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4364 domain = UMA_ANYDOMAIN;
4366 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4370 * We have to call both the zone's init (not the keg's init)
4371 * and the zone's ctor. This is because the item is going from
4372 * a keg slab directly to the user, and the user is expecting it
4373 * to be both zone-init'd as well as zone-ctor'd.
4375 if (zone->uz_init != NULL) {
4378 kasan_mark_item_valid(zone, item);
4379 error = zone->uz_init(item, zone->uz_size, flags);
4380 kasan_mark_item_invalid(zone, item);
4382 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4386 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4391 counter_u64_add(zone->uz_allocs, 1);
4392 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4393 zone->uz_name, zone);
4398 counter_u64_add(zone->uz_fails, 1);
4400 if (zone->uz_max_items > 0)
4401 zone_free_limit(zone, 1);
4402 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4403 zone->uz_name, zone);
4410 uma_zfree_smr(uma_zone_t zone, void *item)
4413 uma_cache_bucket_t bucket;
4419 CTR3(KTR_UMA, "uma_zfree_smr zone %s(%p) item %p",
4420 zone->uz_name, zone, item);
4422 #ifdef UMA_ZALLOC_DEBUG
4423 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4424 ("uma_zfree_smr: called with non-SMR zone."));
4425 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4426 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4427 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4430 cache = &zone->uz_cpu[curcpu];
4433 uz_flags = cache_uz_flags(cache);
4434 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4435 itemdomain = item_domain(item);
4439 cache = &zone->uz_cpu[curcpu];
4440 /* SMR Zones must free to the free bucket. */
4441 bucket = &cache->uc_freebucket;
4443 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4444 PCPU_GET(domain) != itemdomain) {
4445 bucket = &cache->uc_crossbucket;
4448 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4449 cache_bucket_push(cache, bucket, item);
4453 } while (cache_free(zone, cache, NULL, itemdomain));
4457 * If nothing else caught this, we'll just do an internal free.
4459 zone_free_item(zone, item, NULL, SKIP_NONE);
4464 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4467 uma_cache_bucket_t bucket;
4468 int itemdomain, uz_flags;
4470 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4471 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4473 CTR3(KTR_UMA, "uma_zfree_arg zone %s(%p) item %p",
4474 zone->uz_name, zone, item);
4476 #ifdef UMA_ZALLOC_DEBUG
4477 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4478 ("uma_zfree_arg: called with SMR zone."));
4479 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4482 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4487 * We are accessing the per-cpu cache without a critical section to
4488 * fetch size and flags. This is acceptable, if we are preempted we
4489 * will simply read another cpu's line.
4491 cache = &zone->uz_cpu[curcpu];
4492 uz_flags = cache_uz_flags(cache);
4493 if (UMA_ALWAYS_CTORDTOR ||
4494 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4495 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4498 * The race here is acceptable. If we miss it we'll just have to wait
4499 * a little longer for the limits to be reset.
4501 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4502 if (atomic_load_32(&zone->uz_sleepers) > 0)
4507 * If possible, free to the per-CPU cache. There are two
4508 * requirements for safe access to the per-CPU cache: (1) the thread
4509 * accessing the cache must not be preempted or yield during access,
4510 * and (2) the thread must not migrate CPUs without switching which
4511 * cache it accesses. We rely on a critical section to prevent
4512 * preemption and migration. We release the critical section in
4513 * order to acquire the zone mutex if we are unable to free to the
4514 * current cache; when we re-acquire the critical section, we must
4515 * detect and handle migration if it has occurred.
4519 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4520 itemdomain = item_domain(item);
4524 cache = &zone->uz_cpu[curcpu];
4526 * Try to free into the allocbucket first to give LIFO
4527 * ordering for cache-hot datastructures. Spill over
4528 * into the freebucket if necessary. Alloc will swap
4529 * them if one runs dry.
4531 bucket = &cache->uc_allocbucket;
4533 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4534 PCPU_GET(domain) != itemdomain) {
4535 bucket = &cache->uc_crossbucket;
4538 if (bucket->ucb_cnt == bucket->ucb_entries &&
4539 cache->uc_freebucket.ucb_cnt <
4540 cache->uc_freebucket.ucb_entries)
4541 cache_bucket_swap(&cache->uc_freebucket,
4542 &cache->uc_allocbucket);
4543 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4544 cache_bucket_push(cache, bucket, item);
4548 } while (cache_free(zone, cache, udata, itemdomain));
4552 * If nothing else caught this, we'll just do an internal free.
4555 zone_free_item(zone, item, udata, SKIP_DTOR);
4560 * sort crossdomain free buckets to domain correct buckets and cache
4564 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4566 struct uma_bucketlist emptybuckets, fullbuckets;
4567 uma_zone_domain_t zdom;
4574 "uma_zfree: zone %s(%p) draining cross bucket %p",
4575 zone->uz_name, zone, bucket);
4578 * It is possible for buckets to arrive here out of order so we fetch
4579 * the current smr seq rather than accepting the bucket's.
4581 seq = SMR_SEQ_INVALID;
4582 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4583 seq = smr_advance(zone->uz_smr);
4586 * To avoid having ndomain * ndomain buckets for sorting we have a
4587 * lock on the current crossfree bucket. A full matrix with
4588 * per-domain locking could be used if necessary.
4590 STAILQ_INIT(&emptybuckets);
4591 STAILQ_INIT(&fullbuckets);
4592 ZONE_CROSS_LOCK(zone);
4593 for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4594 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4595 domain = item_domain(item);
4596 zdom = ZDOM_GET(zone, domain);
4597 if (zdom->uzd_cross == NULL) {
4598 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4599 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4600 zdom->uzd_cross = b;
4603 * Avoid allocating a bucket with the cross lock
4604 * held, since allocation can trigger a
4605 * cross-domain free and bucket zones may
4606 * allocate from each other.
4608 ZONE_CROSS_UNLOCK(zone);
4609 b = bucket_alloc(zone, udata, M_NOWAIT);
4612 ZONE_CROSS_LOCK(zone);
4613 if (zdom->uzd_cross != NULL) {
4614 STAILQ_INSERT_HEAD(&emptybuckets, b,
4617 zdom->uzd_cross = b;
4621 b = zdom->uzd_cross;
4622 b->ub_bucket[b->ub_cnt++] = item;
4624 if (b->ub_cnt == b->ub_entries) {
4625 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4626 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4627 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4628 zdom->uzd_cross = b;
4631 ZONE_CROSS_UNLOCK(zone);
4633 if (bucket->ub_cnt == 0)
4634 bucket->ub_seq = SMR_SEQ_INVALID;
4635 bucket_free(zone, bucket, udata);
4637 while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4638 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4639 bucket_free(zone, b, udata);
4641 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4642 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4643 domain = item_domain(b->ub_bucket[0]);
4644 zone_put_bucket(zone, domain, b, udata, true);
4650 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4651 int itemdomain, bool ws)
4656 * Buckets coming from the wrong domain will be entirely for the
4657 * only other domain on two domain systems. In this case we can
4658 * simply cache them. Otherwise we need to sort them back to
4661 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4662 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4663 zone_free_cross(zone, bucket, udata);
4669 * Attempt to save the bucket in the zone's domain bucket cache.
4672 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4673 zone->uz_name, zone, bucket);
4674 /* ub_cnt is pointing to the last free item */
4675 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4676 itemdomain = zone_domain_lowest(zone, itemdomain);
4677 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4681 * Populate a free or cross bucket for the current cpu cache. Free any
4682 * existing full bucket either to the zone cache or back to the slab layer.
4684 * Enters and returns in a critical section. false return indicates that
4685 * we can not satisfy this free in the cache layer. true indicates that
4686 * the caller should retry.
4688 static __noinline bool
4689 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, int itemdomain)
4691 uma_cache_bucket_t cbucket;
4692 uma_bucket_t newbucket, bucket;
4694 CRITICAL_ASSERT(curthread);
4696 if (zone->uz_bucket_size == 0)
4699 cache = &zone->uz_cpu[curcpu];
4703 * FIRSTTOUCH domains need to free to the correct zdom. When
4704 * enabled this is the zdom of the item. The bucket is the
4705 * cross bucket if the current domain and itemdomain do not match.
4707 cbucket = &cache->uc_freebucket;
4709 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4710 if (PCPU_GET(domain) != itemdomain) {
4711 cbucket = &cache->uc_crossbucket;
4712 if (cbucket->ucb_cnt != 0)
4713 counter_u64_add(zone->uz_xdomain,
4718 bucket = cache_bucket_unload(cbucket);
4719 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4720 ("cache_free: Entered with non-full free bucket."));
4722 /* We are no longer associated with this CPU. */
4726 * Don't let SMR zones operate without a free bucket. Force
4727 * a synchronize and re-use this one. We will only degrade
4728 * to a synchronize every bucket_size items rather than every
4729 * item if we fail to allocate a bucket.
4731 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4733 bucket->ub_seq = smr_advance(zone->uz_smr);
4734 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4735 if (newbucket == NULL && bucket != NULL) {
4736 bucket_drain(zone, bucket);
4740 } else if (!bucketdisable)
4741 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4744 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4747 if ((bucket = newbucket) == NULL)
4749 cache = &zone->uz_cpu[curcpu];
4752 * Check to see if we should be populating the cross bucket. If it
4753 * is already populated we will fall through and attempt to populate
4756 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4757 if (PCPU_GET(domain) != itemdomain &&
4758 cache->uc_crossbucket.ucb_bucket == NULL) {
4759 cache_bucket_load_cross(cache, bucket);
4765 * We may have lost the race to fill the bucket or switched CPUs.
4767 if (cache->uc_freebucket.ucb_bucket != NULL) {
4769 bucket_free(zone, bucket, udata);
4772 cache_bucket_load_free(cache, bucket);
4778 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4785 KEG_LOCK_ASSERT(keg, slab->us_domain);
4787 /* Do we need to remove from any lists? */
4788 dom = &keg->uk_domain[slab->us_domain];
4789 if (slab->us_freecount + 1 == keg->uk_ipers) {
4790 LIST_REMOVE(slab, us_link);
4791 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4792 dom->ud_free_slabs++;
4793 } else if (slab->us_freecount == 0) {
4794 LIST_REMOVE(slab, us_link);
4795 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4798 /* Slab management. */
4799 freei = slab_item_index(slab, keg, item);
4800 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4801 slab->us_freecount++;
4803 /* Keg statistics. */
4804 dom->ud_free_items++;
4808 zone_release(void *arg, void **bucket, int cnt)
4821 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4822 lock = KEG_LOCK(keg, 0);
4823 for (i = 0; i < cnt; i++) {
4825 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4826 slab = vtoslab((vm_offset_t)item);
4828 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4829 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4830 slab = hash_sfind(&keg->uk_hash, mem);
4832 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4834 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4837 lock = KEG_LOCK(keg, slab->us_domain);
4839 slab_free_item(zone, slab, item);
4846 * Frees a single item to any zone.
4849 * zone The zone to free to
4850 * item The item we're freeing
4851 * udata User supplied data for the dtor
4852 * skip Skip dtors and finis
4854 static __noinline void
4855 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4859 * If a free is sent directly to an SMR zone we have to
4860 * synchronize immediately because the item can instantly
4861 * be reallocated. This should only happen in degenerate
4862 * cases when no memory is available for per-cpu caches.
4864 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4865 smr_synchronize(zone->uz_smr);
4867 item_dtor(zone, item, zone->uz_size, udata, skip);
4869 if (skip < SKIP_FINI && zone->uz_fini) {
4870 kasan_mark_item_valid(zone, item);
4871 zone->uz_fini(item, zone->uz_size);
4872 kasan_mark_item_invalid(zone, item);
4875 zone->uz_release(zone->uz_arg, &item, 1);
4877 if (skip & SKIP_CNT)
4880 counter_u64_add(zone->uz_frees, 1);
4882 if (zone->uz_max_items > 0)
4883 zone_free_limit(zone, 1);
4888 uma_zone_set_max(uma_zone_t zone, int nitems)
4892 * If the limit is small, we may need to constrain the maximum per-CPU
4893 * cache size, or disable caching entirely.
4895 uma_zone_set_maxcache(zone, nitems);
4898 * XXX This can misbehave if the zone has any allocations with
4899 * no limit and a limit is imposed. There is currently no
4900 * way to clear a limit.
4903 if (zone->uz_max_items == 0)
4904 ZONE_ASSERT_COLD(zone);
4905 zone->uz_max_items = nitems;
4906 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4907 zone_update_caches(zone);
4908 /* We may need to wake waiters. */
4909 wakeup(&zone->uz_max_items);
4917 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4919 int bpcpu, bpdom, bsize, nb;
4924 * Compute a lower bound on the number of items that may be cached in
4925 * the zone. Each CPU gets at least two buckets, and for cross-domain
4926 * frees we use an additional bucket per CPU and per domain. Select the
4927 * largest bucket size that does not exceed half of the requested limit,
4928 * with the left over space given to the full bucket cache.
4933 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4938 nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4939 bsize = nitems / nb / 2;
4940 if (bsize > BUCKET_MAX)
4942 else if (bsize == 0 && nitems / nb > 0)
4944 zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
4945 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4946 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4947 zone->uz_bucket_max = nitems - nb * bsize;
4953 uma_zone_get_max(uma_zone_t zone)
4957 nitems = atomic_load_64(&zone->uz_max_items);
4964 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4967 ZONE_ASSERT_COLD(zone);
4968 zone->uz_warning = warning;
4973 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4976 ZONE_ASSERT_COLD(zone);
4977 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4982 uma_zone_get_cur(uma_zone_t zone)
4988 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4989 nitems = counter_u64_fetch(zone->uz_allocs) -
4990 counter_u64_fetch(zone->uz_frees);
4992 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4993 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4995 return (nitems < 0 ? 0 : nitems);
4999 uma_zone_get_allocs(uma_zone_t zone)
5005 if (zone->uz_allocs != EARLY_COUNTER)
5006 nitems = counter_u64_fetch(zone->uz_allocs);
5008 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
5014 uma_zone_get_frees(uma_zone_t zone)
5020 if (zone->uz_frees != EARLY_COUNTER)
5021 nitems = counter_u64_fetch(zone->uz_frees);
5023 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
5029 /* Used only for KEG_ASSERT_COLD(). */
5031 uma_keg_get_allocs(uma_keg_t keg)
5037 LIST_FOREACH(z, &keg->uk_zones, uz_link)
5038 nitems += uma_zone_get_allocs(z);
5046 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
5051 KEG_ASSERT_COLD(keg);
5052 keg->uk_init = uminit;
5057 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
5062 KEG_ASSERT_COLD(keg);
5063 keg->uk_fini = fini;
5068 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
5071 ZONE_ASSERT_COLD(zone);
5072 zone->uz_init = zinit;
5077 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
5080 ZONE_ASSERT_COLD(zone);
5081 zone->uz_fini = zfini;
5086 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
5091 KEG_ASSERT_COLD(keg);
5092 keg->uk_freef = freef;
5097 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
5102 KEG_ASSERT_COLD(keg);
5103 keg->uk_allocf = allocf;
5108 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
5111 ZONE_ASSERT_COLD(zone);
5113 KASSERT(smr != NULL, ("Got NULL smr"));
5114 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
5115 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
5116 zone->uz_flags |= UMA_ZONE_SMR;
5118 zone_update_caches(zone);
5122 uma_zone_get_smr(uma_zone_t zone)
5125 return (zone->uz_smr);
5130 uma_zone_reserve(uma_zone_t zone, int items)
5135 KEG_ASSERT_COLD(keg);
5136 keg->uk_reserve = items;
5141 uma_zone_reserve_kva(uma_zone_t zone, int count)
5148 KEG_ASSERT_COLD(keg);
5149 ZONE_ASSERT_COLD(zone);
5151 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
5153 #ifdef UMA_MD_SMALL_ALLOC
5154 if (keg->uk_ppera > 1) {
5158 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
5164 MPASS(keg->uk_kva == 0);
5167 zone->uz_max_items = pages * keg->uk_ipers;
5168 #ifdef UMA_MD_SMALL_ALLOC
5169 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
5171 keg->uk_allocf = noobj_alloc;
5173 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5174 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5175 zone_update_caches(zone);
5182 uma_prealloc(uma_zone_t zone, int items)
5184 struct vm_domainset_iter di;
5188 int aflags, domain, slabs;
5191 slabs = howmany(items, keg->uk_ipers);
5192 while (slabs-- > 0) {
5194 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
5197 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
5200 dom = &keg->uk_domain[slab->us_domain];
5202 * keg_alloc_slab() always returns a slab on the
5205 LIST_REMOVE(slab, us_link);
5206 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
5208 dom->ud_free_slabs++;
5209 KEG_UNLOCK(keg, slab->us_domain);
5212 if (vm_domainset_iter_policy(&di, &domain) != 0)
5213 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
5219 * Returns a snapshot of memory consumption in bytes.
5222 uma_zone_memory(uma_zone_t zone)
5228 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
5229 for (i = 0; i < vm_ndomains; i++)
5230 sz += ZDOM_GET(zone, i)->uzd_nitems;
5231 return (sz * zone->uz_size);
5233 for (i = 0; i < vm_ndomains; i++)
5234 sz += zone->uz_keg->uk_domain[i].ud_pages;
5236 return (sz * PAGE_SIZE);
5239 struct uma_reclaim_args {
5245 uma_reclaim_domain_cb(uma_zone_t zone, void *arg)
5247 struct uma_reclaim_args *args;
5250 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0)
5251 uma_zone_reclaim_domain(zone, args->req, args->domain);
5256 uma_reclaim(int req)
5258 uma_reclaim_domain(req, UMA_ANYDOMAIN);
5262 uma_reclaim_domain(int req, int domain)
5264 struct uma_reclaim_args args;
5268 args.domain = domain;
5271 sx_slock(&uma_reclaim_lock);
5273 case UMA_RECLAIM_TRIM:
5274 case UMA_RECLAIM_DRAIN:
5275 zone_foreach(uma_reclaim_domain_cb, &args);
5277 case UMA_RECLAIM_DRAIN_CPU:
5278 zone_foreach(uma_reclaim_domain_cb, &args);
5279 pcpu_cache_drain_safe(NULL);
5280 zone_foreach(uma_reclaim_domain_cb, &args);
5283 panic("unhandled reclamation request %d", req);
5287 * Some slabs may have been freed but this zone will be visited early
5288 * we visit again so that we can free pages that are empty once other
5289 * zones are drained. We have to do the same for buckets.
5291 uma_zone_reclaim_domain(slabzones[0], UMA_RECLAIM_DRAIN, domain);
5292 uma_zone_reclaim_domain(slabzones[1], UMA_RECLAIM_DRAIN, domain);
5293 bucket_zone_drain(domain);
5294 sx_sunlock(&uma_reclaim_lock);
5297 static volatile int uma_reclaim_needed;
5300 uma_reclaim_wakeup(void)
5303 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5304 wakeup(uma_reclaim);
5308 uma_reclaim_worker(void *arg __unused)
5312 sx_xlock(&uma_reclaim_lock);
5313 while (atomic_load_int(&uma_reclaim_needed) == 0)
5314 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5316 sx_xunlock(&uma_reclaim_lock);
5317 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5318 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5319 atomic_store_int(&uma_reclaim_needed, 0);
5320 /* Don't fire more than once per-second. */
5321 pause("umarclslp", hz);
5327 uma_zone_reclaim(uma_zone_t zone, int req)
5329 uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5333 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5336 case UMA_RECLAIM_TRIM:
5337 zone_reclaim(zone, domain, M_NOWAIT, false);
5339 case UMA_RECLAIM_DRAIN:
5340 zone_reclaim(zone, domain, M_NOWAIT, true);
5342 case UMA_RECLAIM_DRAIN_CPU:
5343 pcpu_cache_drain_safe(zone);
5344 zone_reclaim(zone, domain, M_NOWAIT, true);
5347 panic("unhandled reclamation request %d", req);
5353 uma_zone_exhausted(uma_zone_t zone)
5356 return (atomic_load_32(&zone->uz_sleepers) > 0);
5363 return (uma_kmem_limit);
5367 uma_set_limit(unsigned long limit)
5370 uma_kmem_limit = limit;
5377 return (atomic_load_long(&uma_kmem_total));
5384 return (uma_kmem_limit - uma_size());
5389 * Generate statistics across both the zone and its per-cpu cache's. Return
5390 * desired statistics if the pointer is non-NULL for that statistic.
5392 * Note: does not update the zone statistics, as it can't safely clear the
5393 * per-CPU cache statistic.
5397 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5398 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5401 uint64_t allocs, frees, sleeps, xdomain;
5404 allocs = frees = sleeps = xdomain = 0;
5407 cache = &z->uz_cpu[cpu];
5408 cachefree += cache->uc_allocbucket.ucb_cnt;
5409 cachefree += cache->uc_freebucket.ucb_cnt;
5410 xdomain += cache->uc_crossbucket.ucb_cnt;
5411 cachefree += cache->uc_crossbucket.ucb_cnt;
5412 allocs += cache->uc_allocs;
5413 frees += cache->uc_frees;
5415 allocs += counter_u64_fetch(z->uz_allocs);
5416 frees += counter_u64_fetch(z->uz_frees);
5417 xdomain += counter_u64_fetch(z->uz_xdomain);
5418 sleeps += z->uz_sleeps;
5419 if (cachefreep != NULL)
5420 *cachefreep = cachefree;
5421 if (allocsp != NULL)
5425 if (sleepsp != NULL)
5427 if (xdomainp != NULL)
5428 *xdomainp = xdomain;
5433 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5440 rw_rlock(&uma_rwlock);
5441 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5442 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5445 LIST_FOREACH(z, &uma_cachezones, uz_link)
5448 rw_runlock(&uma_rwlock);
5449 return (sysctl_handle_int(oidp, &count, 0, req));
5453 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5454 struct uma_percpu_stat *ups, bool internal)
5456 uma_zone_domain_t zdom;
5460 for (i = 0; i < vm_ndomains; i++) {
5461 zdom = ZDOM_GET(z, i);
5462 uth->uth_zone_free += zdom->uzd_nitems;
5464 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5465 uth->uth_frees = counter_u64_fetch(z->uz_frees);
5466 uth->uth_fails = counter_u64_fetch(z->uz_fails);
5467 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5468 uth->uth_sleeps = z->uz_sleeps;
5470 for (i = 0; i < mp_maxid + 1; i++) {
5471 bzero(&ups[i], sizeof(*ups));
5472 if (internal || CPU_ABSENT(i))
5474 cache = &z->uz_cpu[i];
5475 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5476 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5477 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5478 ups[i].ups_allocs = cache->uc_allocs;
5479 ups[i].ups_frees = cache->uc_frees;
5484 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5486 struct uma_stream_header ush;
5487 struct uma_type_header uth;
5488 struct uma_percpu_stat *ups;
5493 uint32_t kfree, pages;
5494 int count, error, i;
5496 error = sysctl_wire_old_buffer(req, 0);
5499 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5500 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5501 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5504 rw_rlock(&uma_rwlock);
5505 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5506 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5510 LIST_FOREACH(z, &uma_cachezones, uz_link)
5514 * Insert stream header.
5516 bzero(&ush, sizeof(ush));
5517 ush.ush_version = UMA_STREAM_VERSION;
5518 ush.ush_maxcpus = (mp_maxid + 1);
5519 ush.ush_count = count;
5520 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5522 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5524 for (i = 0; i < vm_ndomains; i++) {
5525 kfree += kz->uk_domain[i].ud_free_items;
5526 pages += kz->uk_domain[i].ud_pages;
5528 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5529 bzero(&uth, sizeof(uth));
5530 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5531 uth.uth_align = kz->uk_align;
5532 uth.uth_size = kz->uk_size;
5533 uth.uth_rsize = kz->uk_rsize;
5534 if (z->uz_max_items > 0) {
5535 items = UZ_ITEMS_COUNT(z->uz_items);
5536 uth.uth_pages = (items / kz->uk_ipers) *
5539 uth.uth_pages = pages;
5540 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5542 uth.uth_limit = z->uz_max_items;
5543 uth.uth_keg_free = kfree;
5546 * A zone is secondary is it is not the first entry
5547 * on the keg's zone list.
5549 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5550 (LIST_FIRST(&kz->uk_zones) != z))
5551 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5552 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5553 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5554 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5555 for (i = 0; i < mp_maxid + 1; i++)
5556 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5559 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5560 bzero(&uth, sizeof(uth));
5561 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5562 uth.uth_size = z->uz_size;
5563 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5564 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5565 for (i = 0; i < mp_maxid + 1; i++)
5566 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5569 rw_runlock(&uma_rwlock);
5570 error = sbuf_finish(&sbuf);
5577 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5579 uma_zone_t zone = *(uma_zone_t *)arg1;
5582 max = uma_zone_get_max(zone);
5583 error = sysctl_handle_int(oidp, &max, 0, req);
5584 if (error || !req->newptr)
5587 uma_zone_set_max(zone, max);
5593 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5599 * Some callers want to add sysctls for global zones that
5600 * may not yet exist so they pass a pointer to a pointer.
5603 zone = *(uma_zone_t *)arg1;
5606 cur = uma_zone_get_cur(zone);
5607 return (sysctl_handle_int(oidp, &cur, 0, req));
5611 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5613 uma_zone_t zone = arg1;
5616 cur = uma_zone_get_allocs(zone);
5617 return (sysctl_handle_64(oidp, &cur, 0, req));
5621 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5623 uma_zone_t zone = arg1;
5626 cur = uma_zone_get_frees(zone);
5627 return (sysctl_handle_64(oidp, &cur, 0, req));
5631 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5634 uma_zone_t zone = arg1;
5637 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5638 if (zone->uz_flags != 0)
5639 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5641 sbuf_printf(&sbuf, "0");
5642 error = sbuf_finish(&sbuf);
5649 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5651 uma_keg_t keg = arg1;
5652 int avail, effpct, total;
5654 total = keg->uk_ppera * PAGE_SIZE;
5655 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5656 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5658 * We consider the client's requested size and alignment here, not the
5659 * real size determination uk_rsize, because we also adjust the real
5660 * size for internal implementation reasons (max bitset size).
5662 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5663 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5664 avail *= mp_maxid + 1;
5665 effpct = 100 * avail / total;
5666 return (sysctl_handle_int(oidp, &effpct, 0, req));
5670 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5672 uma_zone_t zone = arg1;
5675 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5676 return (sysctl_handle_64(oidp, &cur, 0, req));
5681 uma_dbg_getslab(uma_zone_t zone, void *item)
5688 * It is safe to return the slab here even though the
5689 * zone is unlocked because the item's allocation state
5690 * essentially holds a reference.
5692 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5693 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5695 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5696 return (vtoslab((vm_offset_t)mem));
5698 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5699 return ((uma_slab_t)(mem + keg->uk_pgoff));
5701 slab = hash_sfind(&keg->uk_hash, mem);
5708 uma_dbg_zskip(uma_zone_t zone, void *mem)
5711 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5714 return (uma_dbg_kskip(zone->uz_keg, mem));
5718 uma_dbg_kskip(uma_keg_t keg, void *mem)
5722 if (dbg_divisor == 0)
5725 if (dbg_divisor == 1)
5728 idx = (uintptr_t)mem >> PAGE_SHIFT;
5729 if (keg->uk_ipers > 1) {
5730 idx *= keg->uk_ipers;
5731 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5734 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5735 counter_u64_add(uma_skip_cnt, 1);
5738 counter_u64_add(uma_dbg_cnt, 1);
5744 * Set up the slab's freei data such that uma_dbg_free can function.
5748 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5754 slab = uma_dbg_getslab(zone, item);
5756 panic("uma: item %p did not belong to zone %s",
5757 item, zone->uz_name);
5760 freei = slab_item_index(slab, keg, item);
5762 if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5763 slab_dbg_bits(slab, keg)))
5764 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5765 item, zone, zone->uz_name, slab, freei);
5769 * Verifies freed addresses. Checks for alignment, valid slab membership
5770 * and duplicate frees.
5774 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5780 slab = uma_dbg_getslab(zone, item);
5782 panic("uma: Freed item %p did not belong to zone %s",
5783 item, zone->uz_name);
5786 freei = slab_item_index(slab, keg, item);
5788 if (freei >= keg->uk_ipers)
5789 panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5790 item, zone, zone->uz_name, slab, freei);
5792 if (slab_item(slab, keg, freei) != item)
5793 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5794 item, zone, zone->uz_name, slab, freei);
5796 if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5797 slab_dbg_bits(slab, keg)))
5798 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5799 item, zone, zone->uz_name, slab, freei);
5801 #endif /* INVARIANTS */
5805 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5806 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5811 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5812 *allocs = counter_u64_fetch(z->uz_allocs);
5813 frees = counter_u64_fetch(z->uz_frees);
5814 *sleeps = z->uz_sleeps;
5818 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5820 for (i = 0; i < vm_ndomains; i++) {
5821 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5822 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5823 (LIST_FIRST(&kz->uk_zones) != z)))
5824 *cachefree += kz->uk_domain[i].ud_free_items;
5826 *used = *allocs - frees;
5827 return (((int64_t)*used + *cachefree) * kz->uk_size);
5830 DB_SHOW_COMMAND_FLAGS(uma, db_show_uma, DB_CMD_MEMSAFE)
5832 const char *fmt_hdr, *fmt_entry;
5835 uint64_t allocs, used, sleeps, xdomain;
5837 /* variables for sorting */
5839 uma_zone_t cur_zone, last_zone;
5840 int64_t cur_size, last_size, size;
5843 /* /i option produces machine-parseable CSV output */
5844 if (modif[0] == 'i') {
5845 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5846 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5848 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5849 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5852 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5853 "Sleeps", "Bucket", "Total Mem", "XFree");
5855 /* Sort the zones with largest size first. */
5857 last_size = INT64_MAX;
5862 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5863 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5865 * In the case of size ties, print out zones
5866 * in the order they are encountered. That is,
5867 * when we encounter the most recently output
5868 * zone, we have already printed all preceding
5869 * ties, and we must print all following ties.
5871 if (z == last_zone) {
5875 size = get_uma_stats(kz, z, &allocs, &used,
5876 &sleeps, &cachefree, &xdomain);
5877 if (size > cur_size && size < last_size + ties)
5885 if (cur_zone == NULL)
5888 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5889 &sleeps, &cachefree, &xdomain);
5890 db_printf(fmt_entry, cur_zone->uz_name,
5891 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5892 (uintmax_t)allocs, (uintmax_t)sleeps,
5893 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5898 last_zone = cur_zone;
5899 last_size = cur_size;
5903 DB_SHOW_COMMAND_FLAGS(umacache, db_show_umacache, DB_CMD_MEMSAFE)
5906 uint64_t allocs, frees;
5910 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5911 "Requests", "Bucket");
5912 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5913 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5914 for (i = 0; i < vm_ndomains; i++)
5915 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5916 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5917 z->uz_name, (uintmax_t)z->uz_size,
5918 (intmax_t)(allocs - frees), cachefree,
5919 (uintmax_t)allocs, z->uz_bucket_size);