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,
1765 KASSERT(domain >= 0 && domain < vm_ndomains,
1766 ("keg_alloc_slab: domain %d out of range", domain));
1770 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1771 uma_hash_slab_t hslab;
1772 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1776 slab = &hslab->uhs_slab;
1780 * This reproduces the old vm_zone behavior of zero filling pages the
1781 * first time they are added to a zone.
1783 * Malloced items are zeroed in uma_zalloc.
1786 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1791 if (keg->uk_flags & UMA_ZONE_NODUMP)
1794 /* zone is passed for legacy reasons. */
1795 size = keg->uk_ppera * PAGE_SIZE;
1796 mem = keg->uk_allocf(zone, size, domain, &sflags, aflags);
1798 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1799 zone_free_item(slabzone(keg->uk_ipers),
1800 slab_tohashslab(slab), NULL, SKIP_NONE);
1803 uma_total_inc(size);
1805 /* For HASH zones all pages go to the same uma_domain. */
1806 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1809 kmsan_mark(mem, size,
1810 (aflags & M_ZERO) != 0 ? KMSAN_STATE_INITED : KMSAN_STATE_UNINIT);
1812 /* Point the slab into the allocated memory */
1813 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1814 slab = (uma_slab_t)(mem + keg->uk_pgoff);
1816 slab_tohashslab(slab)->uhs_data = mem;
1818 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1819 for (i = 0; i < keg->uk_ppera; i++)
1820 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1823 slab->us_freecount = keg->uk_ipers;
1824 slab->us_flags = sflags;
1825 slab->us_domain = domain;
1827 BIT_FILL(keg->uk_ipers, &slab->us_free);
1829 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1832 if (keg->uk_init != NULL) {
1833 for (i = 0; i < keg->uk_ipers; i++)
1834 if (keg->uk_init(slab_item(slab, keg, i),
1835 keg->uk_size, flags) != 0)
1837 if (i != keg->uk_ipers) {
1838 keg_free_slab(keg, slab, i);
1842 kasan_mark_slab_invalid(keg, mem);
1843 KEG_LOCK(keg, domain);
1845 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1846 slab, keg->uk_name, keg);
1848 if (keg->uk_flags & UMA_ZFLAG_HASH)
1849 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1852 * If we got a slab here it's safe to mark it partially used
1853 * and return. We assume that the caller is going to remove
1854 * at least one item.
1856 dom = &keg->uk_domain[domain];
1857 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1858 dom->ud_pages += keg->uk_ppera;
1859 dom->ud_free_items += keg->uk_ipers;
1868 * This function is intended to be used early on in place of page_alloc(). It
1869 * performs contiguous physical memory allocations and uses a bump allocator for
1870 * KVA, so is usable before the kernel map is initialized.
1873 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1880 pages = howmany(bytes, PAGE_SIZE);
1881 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1883 *pflag = UMA_SLAB_BOOT;
1884 m = vm_page_alloc_noobj_contig_domain(domain, malloc2vm_flags(wait) |
1885 VM_ALLOC_WIRED, pages, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
1886 VM_MEMATTR_DEFAULT);
1890 pa = VM_PAGE_TO_PHYS(m);
1891 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1892 #if defined(__aarch64__) || defined(__amd64__) || \
1893 defined(__riscv) || defined(__powerpc64__)
1894 if ((wait & M_NODUMP) == 0)
1899 /* Allocate KVA and indirectly advance bootmem. */
1900 return ((void *)pmap_map(&bootmem, m->phys_addr,
1901 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE));
1905 startup_free(void *mem, vm_size_t bytes)
1910 va = (vm_offset_t)mem;
1911 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1914 * startup_alloc() returns direct-mapped slabs on some platforms. Avoid
1915 * unmapping ranges of the direct map.
1917 if (va >= bootstart && va + bytes <= bootmem)
1918 pmap_remove(kernel_pmap, va, va + bytes);
1919 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1920 #if defined(__aarch64__) || defined(__amd64__) || \
1921 defined(__riscv) || defined(__powerpc64__)
1922 dump_drop_page(VM_PAGE_TO_PHYS(m));
1924 vm_page_unwire_noq(m);
1930 * Allocates a number of pages from the system
1933 * bytes The number of bytes requested
1934 * wait Shall we wait?
1937 * A pointer to the alloced memory or possibly
1938 * NULL if M_NOWAIT is set.
1941 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1944 void *p; /* Returned page */
1946 *pflag = UMA_SLAB_KERNEL;
1947 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1953 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1956 struct pglist alloctail;
1957 vm_offset_t addr, zkva;
1959 vm_page_t p, p_next;
1964 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1966 TAILQ_INIT(&alloctail);
1967 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | malloc2vm_flags(wait);
1968 *pflag = UMA_SLAB_KERNEL;
1969 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1970 if (CPU_ABSENT(cpu)) {
1971 p = vm_page_alloc_noobj(flags);
1974 p = vm_page_alloc_noobj(flags);
1976 pc = pcpu_find(cpu);
1977 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1980 p = vm_page_alloc_noobj_domain(pc->pc_domain,
1982 if (__predict_false(p == NULL))
1983 p = vm_page_alloc_noobj(flags);
1986 if (__predict_false(p == NULL))
1988 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1990 if ((addr = kva_alloc(bytes)) == 0)
1993 TAILQ_FOREACH(p, &alloctail, listq) {
1994 pmap_qenter(zkva, &p, 1);
1997 return ((void*)addr);
1999 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2000 vm_page_unwire_noq(p);
2007 * Allocates a number of pages not belonging to a VM object
2010 * bytes The number of bytes requested
2011 * wait Shall we wait?
2014 * A pointer to the alloced memory or possibly
2015 * NULL if M_NOWAIT is set.
2018 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
2021 TAILQ_HEAD(, vm_page) alloctail;
2023 vm_offset_t retkva, zkva;
2024 vm_page_t p, p_next;
2028 TAILQ_INIT(&alloctail);
2030 req = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
2031 if ((wait & M_WAITOK) != 0)
2032 req |= VM_ALLOC_WAITOK;
2034 npages = howmany(bytes, PAGE_SIZE);
2035 while (npages > 0) {
2036 p = vm_page_alloc_noobj_domain(domain, req);
2039 * Since the page does not belong to an object, its
2042 TAILQ_INSERT_TAIL(&alloctail, p, listq);
2047 * Page allocation failed, free intermediate pages and
2050 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2051 vm_page_unwire_noq(p);
2056 *flags = UMA_SLAB_PRIV;
2057 zkva = keg->uk_kva +
2058 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
2060 TAILQ_FOREACH(p, &alloctail, listq) {
2061 pmap_qenter(zkva, &p, 1);
2065 return ((void *)retkva);
2069 * Allocate physically contiguous pages.
2072 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
2076 *pflag = UMA_SLAB_KERNEL;
2077 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
2078 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
2082 * Frees a number of pages to the system
2085 * mem A pointer to the memory to be freed
2086 * size The size of the memory being freed
2087 * flags The original p->us_flags field
2093 page_free(void *mem, vm_size_t size, uint8_t flags)
2096 if ((flags & UMA_SLAB_BOOT) != 0) {
2097 startup_free(mem, size);
2101 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
2102 ("UMA: page_free used with invalid flags %x", flags));
2104 kmem_free((vm_offset_t)mem, size);
2108 * Frees pcpu zone allocations
2111 * mem A pointer to the memory to be freed
2112 * size The size of the memory being freed
2113 * flags The original p->us_flags field
2119 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
2121 vm_offset_t sva, curva;
2125 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
2127 if ((flags & UMA_SLAB_BOOT) != 0) {
2128 startup_free(mem, size);
2132 sva = (vm_offset_t)mem;
2133 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
2134 paddr = pmap_kextract(curva);
2135 m = PHYS_TO_VM_PAGE(paddr);
2136 vm_page_unwire_noq(m);
2139 pmap_qremove(sva, size >> PAGE_SHIFT);
2140 kva_free(sva, size);
2144 * Zero fill initializer
2146 * Arguments/Returns follow uma_init specifications
2149 zero_init(void *mem, int size, int flags)
2156 static struct noslabbits *
2157 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2160 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2165 * Actual size of embedded struct slab (!OFFPAGE).
2168 slab_sizeof(int nitems)
2172 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2173 return (roundup(s, UMA_ALIGN_PTR + 1));
2176 #define UMA_FIXPT_SHIFT 31
2177 #define UMA_FRAC_FIXPT(n, d) \
2178 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2179 #define UMA_FIXPT_PCT(f) \
2180 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2181 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
2182 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2185 * Compute the number of items that will fit in a slab. If hdr is true, the
2186 * item count may be limited to provide space in the slab for an inline slab
2187 * header. Otherwise, all slab space will be provided for item storage.
2190 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2195 /* The padding between items is not needed after the last item. */
2196 padpi = rsize - size;
2200 * Start with the maximum item count and remove items until
2201 * the slab header first alongside the allocatable memory.
2203 for (ipers = MIN(SLAB_MAX_SETSIZE,
2204 (slabsize + padpi - slab_sizeof(1)) / rsize);
2206 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2210 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2216 struct keg_layout_result {
2224 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2225 struct keg_layout_result *kl)
2230 kl->slabsize = slabsize;
2232 /* Handle INTERNAL as inline with an extra page. */
2233 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2234 kl->format &= ~UMA_ZFLAG_INTERNAL;
2235 kl->slabsize += PAGE_SIZE;
2238 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2239 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2241 /* Account for memory used by an offpage slab header. */
2242 total = kl->slabsize;
2243 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2244 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2246 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2250 * Determine the format of a uma keg. This determines where the slab header
2251 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2254 * keg The zone we should initialize
2260 keg_layout(uma_keg_t keg)
2262 struct keg_layout_result kl = {}, kl_tmp;
2271 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2272 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2273 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2274 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2275 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2277 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2278 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2279 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2282 alignsize = keg->uk_align + 1;
2285 * ASAN requires that each allocation be aligned to the shadow map
2288 if (alignsize < KASAN_SHADOW_SCALE)
2289 alignsize = KASAN_SHADOW_SCALE;
2293 * Calculate the size of each allocation (rsize) according to
2294 * alignment. If the requested size is smaller than we have
2295 * allocation bits for we round it up.
2297 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2298 rsize = roundup2(rsize, alignsize);
2300 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2302 * We want one item to start on every align boundary in a page.
2303 * To do this we will span pages. We will also extend the item
2304 * by the size of align if it is an even multiple of align.
2305 * Otherwise, it would fall on the same boundary every time.
2307 if ((rsize & alignsize) == 0)
2309 slabsize = rsize * (PAGE_SIZE / alignsize);
2310 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2311 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2312 slabsize = round_page(slabsize);
2315 * Start with a slab size of as many pages as it takes to
2316 * represent a single item. We will try to fit as many
2317 * additional items into the slab as possible.
2319 slabsize = round_page(keg->uk_size);
2322 /* Build a list of all of the available formats for this keg. */
2325 /* Evaluate an inline slab layout. */
2326 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2329 /* TODO: vm_page-embedded slab. */
2332 * We can't do OFFPAGE if we're internal or if we've been
2333 * asked to not go to the VM for buckets. If we do this we
2334 * may end up going to the VM for slabs which we do not want
2335 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2336 * In those cases, evaluate a pseudo-format called INTERNAL
2337 * which has an inline slab header and one extra page to
2338 * guarantee that it fits.
2340 * Otherwise, see if using an OFFPAGE slab will improve our
2343 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2344 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2346 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2349 * Choose a slab size and format which satisfy the minimum efficiency.
2350 * Prefer the smallest slab size that meets the constraints.
2352 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2353 * for small items (up to PAGE_SIZE), the iteration increment is one
2354 * page; and for large items, the increment is one item.
2356 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2357 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2358 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2361 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2362 round_page(rsize * (i - 1) + keg->uk_size);
2364 for (j = 0; j < nfmt; j++) {
2365 /* Only if we have no viable format yet. */
2366 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2370 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2371 if (kl_tmp.eff <= kl.eff)
2376 CTR6(KTR_UMA, "keg %s layout: format %#x "
2377 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2378 keg->uk_name, kl.format, kl.ipers, rsize,
2379 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2381 /* Stop when we reach the minimum efficiency. */
2382 if (kl.eff >= UMA_MIN_EFF)
2386 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2387 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2388 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2392 pages = atop(kl.slabsize);
2393 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2394 pages *= mp_maxid + 1;
2396 keg->uk_rsize = rsize;
2397 keg->uk_ipers = kl.ipers;
2398 keg->uk_ppera = pages;
2399 keg->uk_flags |= kl.format;
2402 * How do we find the slab header if it is offpage or if not all item
2403 * start addresses are in the same page? We could solve the latter
2404 * case with vaddr alignment, but we don't.
2406 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2407 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2408 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2409 keg->uk_flags |= UMA_ZFLAG_HASH;
2411 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2414 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2415 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2417 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2418 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2419 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2420 keg->uk_ipers, pages));
2424 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2425 * the keg onto the global keg list.
2427 * Arguments/Returns follow uma_ctor specifications
2428 * udata Actually uma_kctor_args
2431 keg_ctor(void *mem, int size, void *udata, int flags)
2433 struct uma_kctor_args *arg = udata;
2434 uma_keg_t keg = mem;
2439 keg->uk_size = arg->size;
2440 keg->uk_init = arg->uminit;
2441 keg->uk_fini = arg->fini;
2442 keg->uk_align = arg->align;
2443 keg->uk_reserve = 0;
2444 keg->uk_flags = arg->flags;
2447 * We use a global round-robin policy by default. Zones with
2448 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2449 * case the iterator is never run.
2451 keg->uk_dr.dr_policy = DOMAINSET_RR();
2452 keg->uk_dr.dr_iter = 0;
2455 * The primary zone is passed to us at keg-creation time.
2458 keg->uk_name = zone->uz_name;
2460 if (arg->flags & UMA_ZONE_ZINIT)
2461 keg->uk_init = zero_init;
2463 if (arg->flags & UMA_ZONE_MALLOC)
2464 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2467 keg->uk_flags &= ~UMA_ZONE_PCPU;
2473 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2474 * work on. Use round-robin for everything else.
2476 * Zones may override the default by specifying either.
2479 if ((keg->uk_flags &
2480 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2481 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2482 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2483 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2487 * If we haven't booted yet we need allocations to go through the
2488 * startup cache until the vm is ready.
2490 #ifdef UMA_MD_SMALL_ALLOC
2491 if (keg->uk_ppera == 1)
2492 keg->uk_allocf = uma_small_alloc;
2495 if (booted < BOOT_KVA)
2496 keg->uk_allocf = startup_alloc;
2497 else if (keg->uk_flags & UMA_ZONE_PCPU)
2498 keg->uk_allocf = pcpu_page_alloc;
2499 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2500 keg->uk_allocf = contig_alloc;
2502 keg->uk_allocf = page_alloc;
2503 #ifdef UMA_MD_SMALL_ALLOC
2504 if (keg->uk_ppera == 1)
2505 keg->uk_freef = uma_small_free;
2508 if (keg->uk_flags & UMA_ZONE_PCPU)
2509 keg->uk_freef = pcpu_page_free;
2511 keg->uk_freef = page_free;
2514 * Initialize keg's locks.
2516 for (i = 0; i < vm_ndomains; i++)
2517 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2520 * If we're putting the slab header in the actual page we need to
2521 * figure out where in each page it goes. See slab_sizeof
2524 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2527 shsize = slab_sizeof(keg->uk_ipers);
2528 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2530 * The only way the following is possible is if with our
2531 * UMA_ALIGN_PTR adjustments we are now bigger than
2532 * UMA_SLAB_SIZE. I haven't checked whether this is
2533 * mathematically possible for all cases, so we make
2536 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2537 ("zone %s ipers %d rsize %d size %d slab won't fit",
2538 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2541 if (keg->uk_flags & UMA_ZFLAG_HASH)
2542 hash_alloc(&keg->uk_hash, 0);
2544 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2546 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2548 rw_wlock(&uma_rwlock);
2549 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2550 rw_wunlock(&uma_rwlock);
2555 zone_kva_available(uma_zone_t zone, void *unused)
2559 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2563 if (keg->uk_allocf == startup_alloc) {
2564 /* Switch to the real allocator. */
2565 if (keg->uk_flags & UMA_ZONE_PCPU)
2566 keg->uk_allocf = pcpu_page_alloc;
2567 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2569 keg->uk_allocf = contig_alloc;
2571 keg->uk_allocf = page_alloc;
2576 zone_alloc_counters(uma_zone_t zone, void *unused)
2579 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2580 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2581 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2582 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2586 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2588 uma_zone_domain_t zdom;
2591 struct sysctl_oid *oid, *domainoid;
2592 int domains, i, cnt;
2593 static const char *nokeg = "cache zone";
2597 * Make a sysctl safe copy of the zone name by removing
2598 * any special characters and handling dups by appending
2601 if (zone->uz_namecnt != 0) {
2602 /* Count the number of decimal digits and '_' separator. */
2603 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2605 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2607 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2610 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2611 for (c = zone->uz_ctlname; *c != '\0'; c++)
2612 if (strchr("./\\ -", *c) != NULL)
2616 * Basic parameters at the root.
2618 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2619 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2621 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2622 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2623 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2624 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2625 zone, 0, sysctl_handle_uma_zone_flags, "A",
2626 "Allocator configuration flags");
2627 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2628 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2629 "Desired per-cpu cache size");
2630 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2631 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2632 "Maximum allowed per-cpu cache size");
2637 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2638 domains = vm_ndomains;
2641 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2642 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2644 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2645 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2646 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2647 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2648 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2649 "Real object size with alignment");
2650 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2651 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2652 "pages per-slab allocation");
2653 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2654 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2655 "items available per-slab");
2656 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2657 "align", CTLFLAG_RD, &keg->uk_align, 0,
2658 "item alignment mask");
2659 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2660 "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2661 "number of reserved items");
2662 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2663 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2664 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2665 "Slab utilization (100 - internal fragmentation %)");
2666 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2667 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2668 for (i = 0; i < domains; i++) {
2669 dom = &keg->uk_domain[i];
2670 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2671 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2672 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2673 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2674 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2675 "Total pages currently allocated from VM");
2676 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2677 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2678 "Items free in the slab layer");
2679 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2680 "free_slabs", CTLFLAG_RD, &dom->ud_free_slabs, 0,
2684 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2685 "name", CTLFLAG_RD, nokeg, "Keg name");
2688 * Information about zone limits.
2690 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2691 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2692 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2693 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2694 zone, 0, sysctl_handle_uma_zone_items, "QU",
2695 "Current number of allocated items if limit is set");
2696 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2697 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2698 "Maximum number of allocated and cached items");
2699 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2700 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2701 "Number of threads sleeping at limit");
2702 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2703 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2704 "Total zone limit sleeps");
2705 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2706 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2707 "Maximum number of items in each domain's bucket cache");
2710 * Per-domain zone information.
2712 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2713 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2714 for (i = 0; i < domains; i++) {
2715 zdom = ZDOM_GET(zone, i);
2716 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2717 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2718 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2719 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2720 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2721 "number of items in this domain");
2722 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2723 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2724 "maximum item count in this period");
2725 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2726 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2727 "minimum item count in this period");
2728 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2729 "bimin", CTLFLAG_RD, &zdom->uzd_bimin,
2730 "Minimum item count in this batch");
2731 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2732 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2733 "Working set size");
2734 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2735 "limin", CTLFLAG_RD, &zdom->uzd_limin,
2736 "Long time minimum item count");
2737 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2738 "timin", CTLFLAG_RD, &zdom->uzd_timin, 0,
2739 "Time since zero long time minimum item count");
2743 * General statistics.
2745 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2746 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2747 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2748 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2749 zone, 1, sysctl_handle_uma_zone_cur, "I",
2750 "Current number of allocated items");
2751 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2752 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2753 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2754 "Total allocation calls");
2755 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2756 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2757 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2758 "Total free calls");
2759 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2760 "fails", CTLFLAG_RD, &zone->uz_fails,
2761 "Number of allocation failures");
2762 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2763 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2764 "Free calls from the wrong domain");
2767 struct uma_zone_count {
2773 zone_count(uma_zone_t zone, void *arg)
2775 struct uma_zone_count *cnt;
2779 * Some zones are rapidly created with identical names and
2780 * destroyed out of order. This can lead to gaps in the count.
2781 * Use one greater than the maximum observed for this name.
2783 if (strcmp(zone->uz_name, cnt->name) == 0)
2784 cnt->count = MAX(cnt->count,
2785 zone->uz_namecnt + 1);
2789 zone_update_caches(uma_zone_t zone)
2793 for (i = 0; i <= mp_maxid; i++) {
2794 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2795 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2800 * Zone header ctor. This initializes all fields, locks, etc.
2802 * Arguments/Returns follow uma_ctor specifications
2803 * udata Actually uma_zctor_args
2806 zone_ctor(void *mem, int size, void *udata, int flags)
2808 struct uma_zone_count cnt;
2809 struct uma_zctor_args *arg = udata;
2810 uma_zone_domain_t zdom;
2811 uma_zone_t zone = mem;
2817 zone->uz_name = arg->name;
2818 zone->uz_ctor = arg->ctor;
2819 zone->uz_dtor = arg->dtor;
2820 zone->uz_init = NULL;
2821 zone->uz_fini = NULL;
2822 zone->uz_sleeps = 0;
2823 zone->uz_bucket_size = 0;
2824 zone->uz_bucket_size_min = 0;
2825 zone->uz_bucket_size_max = BUCKET_MAX;
2826 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2827 zone->uz_warning = NULL;
2828 /* The domain structures follow the cpu structures. */
2829 zone->uz_bucket_max = ULONG_MAX;
2830 timevalclear(&zone->uz_ratecheck);
2832 /* Count the number of duplicate names. */
2833 cnt.name = arg->name;
2835 zone_foreach(zone_count, &cnt);
2836 zone->uz_namecnt = cnt.count;
2837 ZONE_CROSS_LOCK_INIT(zone);
2839 for (i = 0; i < vm_ndomains; i++) {
2840 zdom = ZDOM_GET(zone, i);
2841 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2842 STAILQ_INIT(&zdom->uzd_buckets);
2845 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
2846 if (arg->uminit == trash_init && arg->fini == trash_fini)
2847 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2848 #elif defined(KASAN)
2849 if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0)
2850 arg->flags |= UMA_ZONE_NOKASAN;
2854 * This is a pure cache zone, no kegs.
2857 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2858 ("zone_ctor: Import specified for non-cache zone."));
2859 zone->uz_flags = arg->flags;
2860 zone->uz_size = arg->size;
2861 zone->uz_import = arg->import;
2862 zone->uz_release = arg->release;
2863 zone->uz_arg = arg->arg;
2866 * Cache zones are round-robin unless a policy is
2867 * specified because they may have incompatible
2870 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2871 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2873 rw_wlock(&uma_rwlock);
2874 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2875 rw_wunlock(&uma_rwlock);
2880 * Use the regular zone/keg/slab allocator.
2882 zone->uz_import = zone_import;
2883 zone->uz_release = zone_release;
2884 zone->uz_arg = zone;
2887 if (arg->flags & UMA_ZONE_SECONDARY) {
2888 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2889 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2890 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2891 zone->uz_init = arg->uminit;
2892 zone->uz_fini = arg->fini;
2893 zone->uz_flags |= UMA_ZONE_SECONDARY;
2894 rw_wlock(&uma_rwlock);
2896 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2897 if (LIST_NEXT(z, uz_link) == NULL) {
2898 LIST_INSERT_AFTER(z, zone, uz_link);
2903 rw_wunlock(&uma_rwlock);
2904 } else if (keg == NULL) {
2905 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2906 arg->align, arg->flags)) == NULL)
2909 struct uma_kctor_args karg;
2912 /* We should only be here from uma_startup() */
2913 karg.size = arg->size;
2914 karg.uminit = arg->uminit;
2915 karg.fini = arg->fini;
2916 karg.align = arg->align;
2917 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2919 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2925 /* Inherit properties from the keg. */
2927 zone->uz_size = keg->uk_size;
2928 zone->uz_flags |= (keg->uk_flags &
2929 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2932 if (booted >= BOOT_PCPU) {
2933 zone_alloc_counters(zone, NULL);
2934 if (booted >= BOOT_RUNNING)
2935 zone_alloc_sysctl(zone, NULL);
2937 zone->uz_allocs = EARLY_COUNTER;
2938 zone->uz_frees = EARLY_COUNTER;
2939 zone->uz_fails = EARLY_COUNTER;
2942 /* Caller requests a private SMR context. */
2943 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2944 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2946 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2947 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2948 ("Invalid zone flag combination"));
2949 if (arg->flags & UMA_ZFLAG_INTERNAL)
2950 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2951 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2952 zone->uz_bucket_size = BUCKET_MAX;
2953 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2954 zone->uz_bucket_size = 0;
2956 zone->uz_bucket_size = bucket_select(zone->uz_size);
2957 zone->uz_bucket_size_min = zone->uz_bucket_size;
2958 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2959 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2960 zone_update_caches(zone);
2966 * Keg header dtor. This frees all data, destroys locks, frees the hash
2967 * table and removes the keg from the global list.
2969 * Arguments/Returns follow uma_dtor specifications
2973 keg_dtor(void *arg, int size, void *udata)
2976 uint32_t free, pages;
2979 keg = (uma_keg_t)arg;
2981 for (i = 0; i < vm_ndomains; i++) {
2982 free += keg->uk_domain[i].ud_free_items;
2983 pages += keg->uk_domain[i].ud_pages;
2984 KEG_LOCK_FINI(keg, i);
2987 printf("Freed UMA keg (%s) was not empty (%u items). "
2988 " Lost %u pages of memory.\n",
2989 keg->uk_name ? keg->uk_name : "",
2990 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2992 hash_free(&keg->uk_hash);
2998 * Arguments/Returns follow uma_dtor specifications
3002 zone_dtor(void *arg, int size, void *udata)
3008 zone = (uma_zone_t)arg;
3010 sysctl_remove_oid(zone->uz_oid, 1, 1);
3012 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
3015 rw_wlock(&uma_rwlock);
3016 LIST_REMOVE(zone, uz_link);
3017 rw_wunlock(&uma_rwlock);
3018 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3020 keg->uk_reserve = 0;
3022 zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true);
3025 * We only destroy kegs from non secondary/non cache zones.
3027 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3029 rw_wlock(&uma_rwlock);
3030 LIST_REMOVE(keg, uk_link);
3031 rw_wunlock(&uma_rwlock);
3032 zone_free_item(kegs, keg, NULL, SKIP_NONE);
3034 counter_u64_free(zone->uz_allocs);
3035 counter_u64_free(zone->uz_frees);
3036 counter_u64_free(zone->uz_fails);
3037 counter_u64_free(zone->uz_xdomain);
3038 free(zone->uz_ctlname, M_UMA);
3039 for (i = 0; i < vm_ndomains; i++)
3040 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
3041 ZONE_CROSS_LOCK_FINI(zone);
3045 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3050 LIST_FOREACH(keg, &uma_kegs, uk_link) {
3051 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
3054 LIST_FOREACH(zone, &uma_cachezones, uz_link)
3059 * Traverses every zone in the system and calls a callback
3062 * zfunc A pointer to a function which accepts a zone
3069 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3072 rw_rlock(&uma_rwlock);
3073 zone_foreach_unlocked(zfunc, arg);
3074 rw_runlock(&uma_rwlock);
3078 * Initialize the kernel memory allocator. This is done after pages can be
3079 * allocated but before general KVA is available.
3082 uma_startup1(vm_offset_t virtual_avail)
3084 struct uma_zctor_args args;
3085 size_t ksize, zsize, size;
3086 uma_keg_t primarykeg;
3091 bootstart = bootmem = virtual_avail;
3093 rw_init(&uma_rwlock, "UMA lock");
3094 sx_init(&uma_reclaim_lock, "umareclaim");
3096 ksize = sizeof(struct uma_keg) +
3097 (sizeof(struct uma_domain) * vm_ndomains);
3098 ksize = roundup(ksize, UMA_SUPER_ALIGN);
3099 zsize = sizeof(struct uma_zone) +
3100 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
3101 (sizeof(struct uma_zone_domain) * vm_ndomains);
3102 zsize = roundup(zsize, UMA_SUPER_ALIGN);
3104 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
3105 size = (zsize * 2) + ksize;
3106 for (domain = 0; domain < vm_ndomains; domain++) {
3107 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
3112 zones = (uma_zone_t)m;
3114 kegs = (uma_zone_t)m;
3116 primarykeg = (uma_keg_t)m;
3118 /* "manually" create the initial zone */
3119 memset(&args, 0, sizeof(args));
3120 args.name = "UMA Kegs";
3122 args.ctor = keg_ctor;
3123 args.dtor = keg_dtor;
3124 args.uminit = zero_init;
3126 args.keg = primarykeg;
3127 args.align = UMA_SUPER_ALIGN - 1;
3128 args.flags = UMA_ZFLAG_INTERNAL;
3129 zone_ctor(kegs, zsize, &args, M_WAITOK);
3131 args.name = "UMA Zones";
3133 args.ctor = zone_ctor;
3134 args.dtor = zone_dtor;
3135 args.uminit = zero_init;
3138 args.align = UMA_SUPER_ALIGN - 1;
3139 args.flags = UMA_ZFLAG_INTERNAL;
3140 zone_ctor(zones, zsize, &args, M_WAITOK);
3142 /* Now make zones for slab headers */
3143 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
3144 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3145 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
3146 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3148 hashzone = uma_zcreate("UMA Hash",
3149 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
3150 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3156 #ifndef UMA_MD_SMALL_ALLOC
3157 extern void vm_radix_reserve_kva(void);
3161 * Advertise the availability of normal kva allocations and switch to
3162 * the default back-end allocator. Marks the KVA we consumed on startup
3163 * as used in the map.
3169 if (bootstart != bootmem) {
3170 vm_map_lock(kernel_map);
3171 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
3172 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
3173 vm_map_unlock(kernel_map);
3176 #ifndef UMA_MD_SMALL_ALLOC
3177 /* Set up radix zone to use noobj_alloc. */
3178 vm_radix_reserve_kva();
3182 zone_foreach_unlocked(zone_kva_available, NULL);
3187 * Allocate counters as early as possible so that boot-time allocations are
3188 * accounted more precisely.
3191 uma_startup_pcpu(void *arg __unused)
3194 zone_foreach_unlocked(zone_alloc_counters, NULL);
3197 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
3200 * Finish our initialization steps.
3203 uma_startup3(void *arg __unused)
3207 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
3208 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
3209 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
3211 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
3212 booted = BOOT_RUNNING;
3214 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
3215 EVENTHANDLER_PRI_FIRST);
3217 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
3220 uma_startup4(void *arg __unused)
3222 TIMEOUT_TASK_INIT(taskqueue_thread, &uma_timeout_task, 0, uma_timeout,
3224 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task,
3227 SYSINIT(uma_startup4, SI_SUB_TASKQ, SI_ORDER_ANY, uma_startup4, NULL);
3233 booted = BOOT_SHUTDOWN;
3237 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3238 int align, uint32_t flags)
3240 struct uma_kctor_args args;
3243 args.uminit = uminit;
3245 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
3248 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3251 /* Public functions */
3254 uma_set_align(int align)
3257 if (align != UMA_ALIGN_CACHE)
3258 uma_align_cache = align;
3263 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3264 uma_init uminit, uma_fini fini, int align, uint32_t flags)
3267 struct uma_zctor_args args;
3270 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
3273 /* This stuff is essential for the zone ctor */
3274 memset(&args, 0, sizeof(args));
3279 args.uminit = uminit;
3281 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
3283 * Inject procedures which check for memory use after free if we are
3284 * allowed to scramble the memory while it is not allocated. This
3285 * requires that: UMA is actually able to access the memory, no init
3286 * or fini procedures, no dependency on the initial value of the
3287 * memory, and no (legitimate) use of the memory after free. Note,
3288 * the ctor and dtor do not need to be empty.
3290 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3291 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3292 args.uminit = trash_init;
3293 args.fini = trash_fini;
3300 sx_xlock(&uma_reclaim_lock);
3301 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3302 sx_xunlock(&uma_reclaim_lock);
3309 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3310 uma_init zinit, uma_fini zfini, uma_zone_t primary)
3312 struct uma_zctor_args args;
3316 keg = primary->uz_keg;
3317 memset(&args, 0, sizeof(args));
3319 args.size = keg->uk_size;
3322 args.uminit = zinit;
3324 args.align = keg->uk_align;
3325 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3328 sx_xlock(&uma_reclaim_lock);
3329 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3330 sx_xunlock(&uma_reclaim_lock);
3337 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3338 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3339 void *arg, int flags)
3341 struct uma_zctor_args args;
3343 memset(&args, 0, sizeof(args));
3348 args.uminit = zinit;
3350 args.import = zimport;
3351 args.release = zrelease;
3354 args.flags = flags | UMA_ZFLAG_CACHE;
3356 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3361 uma_zdestroy(uma_zone_t zone)
3365 * Large slabs are expensive to reclaim, so don't bother doing
3366 * unnecessary work if we're shutting down.
3368 if (booted == BOOT_SHUTDOWN &&
3369 zone->uz_fini == NULL && zone->uz_release == zone_release)
3371 sx_xlock(&uma_reclaim_lock);
3372 zone_free_item(zones, zone, NULL, SKIP_NONE);
3373 sx_xunlock(&uma_reclaim_lock);
3377 uma_zwait(uma_zone_t zone)
3380 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3381 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3382 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3383 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3385 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3389 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3391 void *item, *pcpu_item;
3395 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3397 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3400 pcpu_item = zpcpu_base_to_offset(item);
3401 if (flags & M_ZERO) {
3403 for (i = 0; i <= mp_maxid; i++)
3404 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3406 bzero(item, zone->uz_size);
3413 * A stub while both regular and pcpu cases are identical.
3416 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3421 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3424 /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */
3425 if (pcpu_item == NULL)
3428 item = zpcpu_offset_to_base(pcpu_item);
3429 uma_zfree_arg(zone, item, udata);
3432 static inline void *
3433 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3440 kasan_mark_item_valid(zone, item);
3441 kmsan_mark_item_uninitialized(zone, item);
3444 skipdbg = uma_dbg_zskip(zone, item);
3445 if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3446 zone->uz_ctor != trash_ctor)
3447 trash_ctor(item, size, udata, flags);
3450 /* Check flags before loading ctor pointer. */
3451 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3452 __predict_false(zone->uz_ctor != NULL) &&
3453 zone->uz_ctor(item, size, udata, flags) != 0) {
3454 counter_u64_add(zone->uz_fails, 1);
3455 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3460 uma_dbg_alloc(zone, NULL, item);
3462 if (__predict_false(flags & M_ZERO))
3463 return (memset(item, 0, size));
3469 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3470 enum zfreeskip skip)
3475 skipdbg = uma_dbg_zskip(zone, item);
3476 if (skip == SKIP_NONE && !skipdbg) {
3477 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3478 uma_dbg_free(zone, udata, item);
3480 uma_dbg_free(zone, NULL, item);
3483 if (__predict_true(skip < SKIP_DTOR)) {
3484 if (zone->uz_dtor != NULL)
3485 zone->uz_dtor(item, size, udata);
3487 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3488 zone->uz_dtor != trash_dtor)
3489 trash_dtor(item, size, udata);
3492 kasan_mark_item_invalid(zone, item);
3497 item_domain(void *item)
3501 domain = vm_phys_domain(vtophys(item));
3502 KASSERT(domain >= 0 && domain < vm_ndomains,
3503 ("%s: unknown domain for item %p", __func__, item));
3508 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3509 #if defined(INVARIANTS) && (defined(DDB) || defined(STACK))
3510 #include <sys/stack.h>
3512 #define UMA_ZALLOC_DEBUG
3514 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3520 if (flags & M_WAITOK) {
3521 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3522 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3527 KASSERT((flags & M_EXEC) == 0,
3528 ("uma_zalloc_debug: called with M_EXEC"));
3529 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3530 ("uma_zalloc_debug: called within spinlock or critical section"));
3531 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3532 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3534 _Static_assert(M_NOWAIT != 0 && M_WAITOK != 0,
3535 "M_NOWAIT and M_WAITOK must be non-zero for this assertion:");
3538 * Give the #elif clause time to find problems, then remove it
3539 * and enable this. (Remove <sys/stack.h> above, too.)
3541 KASSERT((flags & (M_NOWAIT|M_WAITOK)) == M_NOWAIT ||
3542 (flags & (M_NOWAIT|M_WAITOK)) == M_WAITOK,
3543 ("uma_zalloc_debug: must pass one of M_NOWAIT or M_WAITOK"));
3544 #elif defined(DDB) || defined(STACK)
3545 if (__predict_false((flags & (M_NOWAIT|M_WAITOK)) != M_NOWAIT &&
3546 (flags & (M_NOWAIT|M_WAITOK)) != M_WAITOK)) {
3547 static int stack_count;
3550 if (stack_count < 10) {
3552 printf("uma_zalloc* called with bad WAIT flags:\n");
3560 #ifdef DEBUG_MEMGUARD
3561 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3563 item = memguard_alloc(zone->uz_size, flags);
3565 error = EJUSTRETURN;
3566 if (zone->uz_init != NULL &&
3567 zone->uz_init(item, zone->uz_size, flags) != 0) {
3571 if (zone->uz_ctor != NULL &&
3572 zone->uz_ctor(item, zone->uz_size, udata,
3574 counter_u64_add(zone->uz_fails, 1);
3575 if (zone->uz_fini != NULL)
3576 zone->uz_fini(item, zone->uz_size);
3583 /* This is unfortunate but should not be fatal. */
3590 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3592 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3593 ("uma_zfree_debug: called with spinlock or critical section held"));
3595 #ifdef DEBUG_MEMGUARD
3596 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3597 if (zone->uz_dtor != NULL)
3598 zone->uz_dtor(item, zone->uz_size, udata);
3599 if (zone->uz_fini != NULL)
3600 zone->uz_fini(item, zone->uz_size);
3601 memguard_free(item);
3602 return (EJUSTRETURN);
3609 static inline void *
3610 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3611 void *udata, int flags)
3616 item = cache_bucket_pop(cache, bucket);
3617 size = cache_uz_size(cache);
3618 uz_flags = cache_uz_flags(cache);
3620 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3623 static __noinline void *
3624 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3626 uma_cache_bucket_t bucket;
3629 while (cache_alloc(zone, cache, udata, flags)) {
3630 cache = &zone->uz_cpu[curcpu];
3631 bucket = &cache->uc_allocbucket;
3632 if (__predict_false(bucket->ucb_cnt == 0))
3634 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3639 * We can not get a bucket so try to return a single item.
3641 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3642 domain = PCPU_GET(domain);
3644 domain = UMA_ANYDOMAIN;
3645 return (zone_alloc_item(zone, udata, domain, flags));
3650 uma_zalloc_smr(uma_zone_t zone, int flags)
3652 uma_cache_bucket_t bucket;
3655 CTR3(KTR_UMA, "uma_zalloc_smr zone %s(%p) flags %d", zone->uz_name,
3658 #ifdef UMA_ZALLOC_DEBUG
3661 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3662 ("uma_zalloc_arg: called with non-SMR zone."));
3663 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3668 cache = &zone->uz_cpu[curcpu];
3669 bucket = &cache->uc_allocbucket;
3670 if (__predict_false(bucket->ucb_cnt == 0))
3671 return (cache_alloc_retry(zone, cache, NULL, flags));
3672 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3677 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3679 uma_cache_bucket_t bucket;
3682 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3683 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3685 /* This is the fast path allocation */
3686 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3689 #ifdef UMA_ZALLOC_DEBUG
3692 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3693 ("uma_zalloc_arg: called with SMR zone."));
3694 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3699 * If possible, allocate from the per-CPU cache. There are two
3700 * requirements for safe access to the per-CPU cache: (1) the thread
3701 * accessing the cache must not be preempted or yield during access,
3702 * and (2) the thread must not migrate CPUs without switching which
3703 * cache it accesses. We rely on a critical section to prevent
3704 * preemption and migration. We release the critical section in
3705 * order to acquire the zone mutex if we are unable to allocate from
3706 * the current cache; when we re-acquire the critical section, we
3707 * must detect and handle migration if it has occurred.
3710 cache = &zone->uz_cpu[curcpu];
3711 bucket = &cache->uc_allocbucket;
3712 if (__predict_false(bucket->ucb_cnt == 0))
3713 return (cache_alloc_retry(zone, cache, udata, flags));
3714 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3718 * Replenish an alloc bucket and possibly restore an old one. Called in
3719 * a critical section. Returns in a critical section.
3721 * A false return value indicates an allocation failure.
3722 * A true return value indicates success and the caller should retry.
3724 static __noinline bool
3725 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3727 uma_bucket_t bucket;
3728 int curdomain, domain;
3731 CRITICAL_ASSERT(curthread);
3734 * If we have run out of items in our alloc bucket see
3735 * if we can switch with the free bucket.
3737 * SMR Zones can't re-use the free bucket until the sequence has
3740 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3741 cache->uc_freebucket.ucb_cnt != 0) {
3742 cache_bucket_swap(&cache->uc_freebucket,
3743 &cache->uc_allocbucket);
3748 * Discard any empty allocation bucket while we hold no locks.
3750 bucket = cache_bucket_unload_alloc(cache);
3753 if (bucket != NULL) {
3754 KASSERT(bucket->ub_cnt == 0,
3755 ("cache_alloc: Entered with non-empty alloc bucket."));
3756 bucket_free(zone, bucket, udata);
3760 * Attempt to retrieve the item from the per-CPU cache has failed, so
3761 * we must go back to the zone. This requires the zdom lock, so we
3762 * must drop the critical section, then re-acquire it when we go back
3763 * to the cache. Since the critical section is released, we may be
3764 * preempted or migrate. As such, make sure not to maintain any
3765 * thread-local state specific to the cache from prior to releasing
3766 * the critical section.
3768 domain = PCPU_GET(domain);
3769 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3770 VM_DOMAIN_EMPTY(domain))
3771 domain = zone_domain_highest(zone, domain);
3772 bucket = cache_fetch_bucket(zone, cache, domain);
3773 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3774 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3780 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3781 zone->uz_name, zone, bucket);
3782 if (bucket == NULL) {
3788 * See if we lost the race or were migrated. Cache the
3789 * initialized bucket to make this less likely or claim
3790 * the memory directly.
3793 cache = &zone->uz_cpu[curcpu];
3794 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3795 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3796 (curdomain = PCPU_GET(domain)) == domain ||
3797 VM_DOMAIN_EMPTY(curdomain))) {
3799 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3801 cache_bucket_load_alloc(cache, bucket);
3806 * We lost the race, release this bucket and start over.
3809 zone_put_bucket(zone, domain, bucket, udata, !new);
3816 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3819 uma_bucket_t bucket;
3820 uma_zone_domain_t zdom;
3824 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3825 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3827 /* This is the fast path allocation */
3828 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3829 zone->uz_name, zone, domain, flags);
3831 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3832 ("uma_zalloc_domain: called with SMR zone."));
3834 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3835 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3837 if (vm_ndomains == 1)
3838 return (uma_zalloc_arg(zone, udata, flags));
3840 #ifdef UMA_ZALLOC_DEBUG
3841 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3846 * Try to allocate from the bucket cache before falling back to the keg.
3847 * We could try harder and attempt to allocate from per-CPU caches or
3848 * the per-domain cross-domain buckets, but the complexity is probably
3849 * not worth it. It is more important that frees of previous
3850 * cross-domain allocations do not blow up the cache.
3852 zdom = zone_domain_lock(zone, domain);
3853 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3854 item = bucket->ub_bucket[bucket->ub_cnt - 1];
3856 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3859 zone_put_bucket(zone, domain, bucket, udata, true);
3860 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3863 KASSERT(item_domain(item) == domain,
3864 ("%s: bucket cache item %p from wrong domain",
3866 counter_u64_add(zone->uz_allocs, 1);
3871 return (zone_alloc_item(zone, udata, domain, flags));
3873 return (uma_zalloc_arg(zone, udata, flags));
3878 * Find a slab with some space. Prefer slabs that are partially used over those
3879 * that are totally full. This helps to reduce fragmentation.
3881 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3885 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3891 KASSERT(domain >= 0 && domain < vm_ndomains,
3892 ("keg_first_slab: domain %d out of range", domain));
3893 KEG_LOCK_ASSERT(keg, domain);
3898 dom = &keg->uk_domain[domain];
3899 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3901 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3902 LIST_REMOVE(slab, us_link);
3903 dom->ud_free_slabs--;
3904 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3908 domain = (domain + 1) % vm_ndomains;
3909 } while (domain != start);
3915 * Fetch an existing slab from a free or partial list. Returns with the
3916 * keg domain lock held if a slab was found or unlocked if not.
3919 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3924 /* HASH has a single free list. */
3925 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3928 KEG_LOCK(keg, domain);
3929 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3930 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3931 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3932 KEG_UNLOCK(keg, domain);
3939 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3941 struct vm_domainset_iter di;
3946 KASSERT((flags & (M_WAITOK | M_NOVM)) != (M_WAITOK | M_NOVM),
3947 ("%s: invalid flags %#x", __func__, flags));
3951 * Use the keg's policy if upper layers haven't already specified a
3952 * domain (as happens with first-touch zones).
3954 * To avoid races we run the iterator with the keg lock held, but that
3955 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3956 * clear M_WAITOK and handle low memory conditions locally.
3958 rr = rdomain == UMA_ANYDOMAIN;
3960 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3961 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3969 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3974 * M_NOVM is used to break the recursion that can otherwise
3975 * occur if low-level memory management routines use UMA.
3977 if ((flags & M_NOVM) == 0) {
3978 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3984 if ((flags & M_USE_RESERVE) != 0) {
3986 * Drain reserves from other domains before
3987 * giving up or sleeping. It may be useful to
3988 * support per-domain reserves eventually.
3990 rdomain = UMA_ANYDOMAIN;
3993 if ((flags & M_WAITOK) == 0)
3995 vm_wait_domain(domain);
3996 } else if (vm_domainset_iter_policy(&di, &domain) != 0) {
3997 if ((flags & M_WAITOK) != 0) {
3998 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
4006 * We might not have been able to get a slab but another cpu
4007 * could have while we were unlocked. Check again before we
4010 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
4017 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
4023 KEG_LOCK_ASSERT(keg, slab->us_domain);
4025 dom = &keg->uk_domain[slab->us_domain];
4026 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
4027 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
4028 item = slab_item(slab, keg, freei);
4029 slab->us_freecount--;
4030 dom->ud_free_items--;
4033 * Move this slab to the full list. It must be on the partial list, so
4034 * we do not need to update the free slab count. In particular,
4035 * keg_fetch_slab() always returns slabs on the partial list.
4037 if (slab->us_freecount == 0) {
4038 LIST_REMOVE(slab, us_link);
4039 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
4046 zone_import(void *arg, void **bucket, int max, int domain, int flags)
4060 /* Try to keep the buckets totally full */
4061 for (i = 0; i < max; ) {
4062 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
4065 stripe = howmany(max, vm_ndomains);
4067 dom = &keg->uk_domain[slab->us_domain];
4069 bucket[i++] = slab_alloc_item(keg, slab);
4070 if (keg->uk_reserve > 0 &&
4071 dom->ud_free_items <= keg->uk_reserve) {
4073 * Avoid depleting the reserve after a
4074 * successful item allocation, even if
4075 * M_USE_RESERVE is specified.
4077 KEG_UNLOCK(keg, slab->us_domain);
4082 * If the zone is striped we pick a new slab for every
4083 * N allocations. Eliminating this conditional will
4084 * instead pick a new domain for each bucket rather
4085 * than stripe within each bucket. The current option
4086 * produces more fragmentation and requires more cpu
4087 * time but yields better distribution.
4089 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
4090 vm_ndomains > 1 && --stripe == 0)
4093 } while (slab->us_freecount != 0 && i < max);
4094 KEG_UNLOCK(keg, slab->us_domain);
4096 /* Don't block if we allocated any successfully. */
4105 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
4107 uint64_t old, new, total, max;
4110 * The hard case. We're going to sleep because there were existing
4111 * sleepers or because we ran out of items. This routine enforces
4112 * fairness by keeping fifo order.
4114 * First release our ill gotten gains and make some noise.
4117 zone_free_limit(zone, count);
4118 zone_log_warning(zone);
4119 zone_maxaction(zone);
4120 if (flags & M_NOWAIT)
4124 * We need to allocate an item or set ourself as a sleeper
4125 * while the sleepq lock is held to avoid wakeup races. This
4126 * is essentially a home rolled semaphore.
4128 sleepq_lock(&zone->uz_max_items);
4129 old = zone->uz_items;
4131 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
4132 /* Cache the max since we will evaluate twice. */
4133 max = zone->uz_max_items;
4134 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
4135 UZ_ITEMS_COUNT(old) >= max)
4136 new = old + UZ_ITEMS_SLEEPER;
4138 new = old + MIN(count, max - old);
4139 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
4141 /* We may have successfully allocated under the sleepq lock. */
4142 if (UZ_ITEMS_SLEEPERS(new) == 0) {
4143 sleepq_release(&zone->uz_max_items);
4148 * This is in a different cacheline from uz_items so that we
4149 * don't constantly invalidate the fastpath cacheline when we
4150 * adjust item counts. This could be limited to toggling on
4153 atomic_add_32(&zone->uz_sleepers, 1);
4154 atomic_add_64(&zone->uz_sleeps, 1);
4157 * We have added ourselves as a sleeper. The sleepq lock
4158 * protects us from wakeup races. Sleep now and then retry.
4160 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
4161 sleepq_wait(&zone->uz_max_items, PVM);
4164 * After wakeup, remove ourselves as a sleeper and try
4165 * again. We no longer have the sleepq lock for protection.
4167 * Subract ourselves as a sleeper while attempting to add
4170 atomic_subtract_32(&zone->uz_sleepers, 1);
4171 old = atomic_fetchadd_64(&zone->uz_items,
4172 -(UZ_ITEMS_SLEEPER - count));
4173 /* We're no longer a sleeper. */
4174 old -= UZ_ITEMS_SLEEPER;
4177 * If we're still at the limit, restart. Notably do not
4178 * block on other sleepers. Cache the max value to protect
4179 * against changes via sysctl.
4181 total = UZ_ITEMS_COUNT(old);
4182 max = zone->uz_max_items;
4185 /* Truncate if necessary, otherwise wake other sleepers. */
4186 if (total + count > max) {
4187 zone_free_limit(zone, total + count - max);
4188 count = max - total;
4189 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
4190 wakeup_one(&zone->uz_max_items);
4197 * Allocate 'count' items from our max_items limit. Returns the number
4198 * available. If M_NOWAIT is not specified it will sleep until at least
4199 * one item can be allocated.
4202 zone_alloc_limit(uma_zone_t zone, int count, int flags)
4207 max = zone->uz_max_items;
4211 * We expect normal allocations to succeed with a simple
4214 old = atomic_fetchadd_64(&zone->uz_items, count);
4215 if (__predict_true(old + count <= max))
4219 * If we had some items and no sleepers just return the
4220 * truncated value. We have to release the excess space
4221 * though because that may wake sleepers who weren't woken
4222 * because we were temporarily over the limit.
4225 zone_free_limit(zone, (old + count) - max);
4228 return (zone_alloc_limit_hard(zone, count, flags));
4232 * Free a number of items back to the limit.
4235 zone_free_limit(uma_zone_t zone, int count)
4242 * In the common case we either have no sleepers or
4243 * are still over the limit and can just return.
4245 old = atomic_fetchadd_64(&zone->uz_items, -count);
4246 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4247 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4251 * Moderate the rate of wakeups. Sleepers will continue
4252 * to generate wakeups if necessary.
4254 wakeup_one(&zone->uz_max_items);
4258 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4260 uma_bucket_t bucket;
4261 int error, maxbucket, cnt;
4263 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4266 /* Avoid allocs targeting empty domains. */
4267 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4268 domain = UMA_ANYDOMAIN;
4269 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4270 domain = UMA_ANYDOMAIN;
4272 if (zone->uz_max_items > 0)
4273 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4276 maxbucket = zone->uz_bucket_size;
4280 /* Don't wait for buckets, preserve caller's NOVM setting. */
4281 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4282 if (bucket == NULL) {
4287 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4288 MIN(maxbucket, bucket->ub_entries), domain, flags);
4291 * Initialize the memory if necessary.
4293 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4296 for (i = 0; i < bucket->ub_cnt; i++) {
4297 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
4298 error = zone->uz_init(bucket->ub_bucket[i],
4299 zone->uz_size, flags);
4300 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
4306 * If we couldn't initialize the whole bucket, put the
4307 * rest back onto the freelist.
4309 if (i != bucket->ub_cnt) {
4310 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4311 bucket->ub_cnt - i);
4313 bzero(&bucket->ub_bucket[i],
4314 sizeof(void *) * (bucket->ub_cnt - i));
4320 cnt = bucket->ub_cnt;
4321 if (bucket->ub_cnt == 0) {
4322 bucket_free(zone, bucket, udata);
4323 counter_u64_add(zone->uz_fails, 1);
4327 if (zone->uz_max_items > 0 && cnt < maxbucket)
4328 zone_free_limit(zone, maxbucket - cnt);
4334 * Allocates a single item from a zone.
4337 * zone The zone to alloc for.
4338 * udata The data to be passed to the constructor.
4339 * domain The domain to allocate from or UMA_ANYDOMAIN.
4340 * flags M_WAITOK, M_NOWAIT, M_ZERO.
4343 * NULL if there is no memory and M_NOWAIT is set
4344 * An item if successful
4348 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4352 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4353 counter_u64_add(zone->uz_fails, 1);
4357 /* Avoid allocs targeting empty domains. */
4358 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4359 domain = UMA_ANYDOMAIN;
4361 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4365 * We have to call both the zone's init (not the keg's init)
4366 * and the zone's ctor. This is because the item is going from
4367 * a keg slab directly to the user, and the user is expecting it
4368 * to be both zone-init'd as well as zone-ctor'd.
4370 if (zone->uz_init != NULL) {
4373 kasan_mark_item_valid(zone, item);
4374 error = zone->uz_init(item, zone->uz_size, flags);
4375 kasan_mark_item_invalid(zone, item);
4377 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4381 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4386 counter_u64_add(zone->uz_allocs, 1);
4387 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4388 zone->uz_name, zone);
4393 counter_u64_add(zone->uz_fails, 1);
4395 if (zone->uz_max_items > 0)
4396 zone_free_limit(zone, 1);
4397 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4398 zone->uz_name, zone);
4405 uma_zfree_smr(uma_zone_t zone, void *item)
4408 uma_cache_bucket_t bucket;
4414 CTR3(KTR_UMA, "uma_zfree_smr zone %s(%p) item %p",
4415 zone->uz_name, zone, item);
4417 #ifdef UMA_ZALLOC_DEBUG
4418 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4419 ("uma_zfree_smr: called with non-SMR zone."));
4420 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4421 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4422 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4425 cache = &zone->uz_cpu[curcpu];
4428 uz_flags = cache_uz_flags(cache);
4429 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4430 itemdomain = item_domain(item);
4434 cache = &zone->uz_cpu[curcpu];
4435 /* SMR Zones must free to the free bucket. */
4436 bucket = &cache->uc_freebucket;
4438 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4439 PCPU_GET(domain) != itemdomain) {
4440 bucket = &cache->uc_crossbucket;
4443 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4444 cache_bucket_push(cache, bucket, item);
4448 } while (cache_free(zone, cache, NULL, itemdomain));
4452 * If nothing else caught this, we'll just do an internal free.
4454 zone_free_item(zone, item, NULL, SKIP_NONE);
4459 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4462 uma_cache_bucket_t bucket;
4463 int itemdomain, uz_flags;
4465 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4466 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4468 CTR3(KTR_UMA, "uma_zfree_arg zone %s(%p) item %p",
4469 zone->uz_name, zone, item);
4471 #ifdef UMA_ZALLOC_DEBUG
4472 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4473 ("uma_zfree_arg: called with SMR zone."));
4474 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4477 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4482 * We are accessing the per-cpu cache without a critical section to
4483 * fetch size and flags. This is acceptable, if we are preempted we
4484 * will simply read another cpu's line.
4486 cache = &zone->uz_cpu[curcpu];
4487 uz_flags = cache_uz_flags(cache);
4488 if (UMA_ALWAYS_CTORDTOR ||
4489 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4490 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4493 * The race here is acceptable. If we miss it we'll just have to wait
4494 * a little longer for the limits to be reset.
4496 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4497 if (atomic_load_32(&zone->uz_sleepers) > 0)
4502 * If possible, free to the per-CPU cache. There are two
4503 * requirements for safe access to the per-CPU cache: (1) the thread
4504 * accessing the cache must not be preempted or yield during access,
4505 * and (2) the thread must not migrate CPUs without switching which
4506 * cache it accesses. We rely on a critical section to prevent
4507 * preemption and migration. We release the critical section in
4508 * order to acquire the zone mutex if we are unable to free to the
4509 * current cache; when we re-acquire the critical section, we must
4510 * detect and handle migration if it has occurred.
4514 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4515 itemdomain = item_domain(item);
4519 cache = &zone->uz_cpu[curcpu];
4521 * Try to free into the allocbucket first to give LIFO
4522 * ordering for cache-hot datastructures. Spill over
4523 * into the freebucket if necessary. Alloc will swap
4524 * them if one runs dry.
4526 bucket = &cache->uc_allocbucket;
4528 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4529 PCPU_GET(domain) != itemdomain) {
4530 bucket = &cache->uc_crossbucket;
4533 if (bucket->ucb_cnt == bucket->ucb_entries &&
4534 cache->uc_freebucket.ucb_cnt <
4535 cache->uc_freebucket.ucb_entries)
4536 cache_bucket_swap(&cache->uc_freebucket,
4537 &cache->uc_allocbucket);
4538 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4539 cache_bucket_push(cache, bucket, item);
4543 } while (cache_free(zone, cache, udata, itemdomain));
4547 * If nothing else caught this, we'll just do an internal free.
4550 zone_free_item(zone, item, udata, SKIP_DTOR);
4555 * sort crossdomain free buckets to domain correct buckets and cache
4559 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4561 struct uma_bucketlist emptybuckets, fullbuckets;
4562 uma_zone_domain_t zdom;
4569 "uma_zfree: zone %s(%p) draining cross bucket %p",
4570 zone->uz_name, zone, bucket);
4573 * It is possible for buckets to arrive here out of order so we fetch
4574 * the current smr seq rather than accepting the bucket's.
4576 seq = SMR_SEQ_INVALID;
4577 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4578 seq = smr_advance(zone->uz_smr);
4581 * To avoid having ndomain * ndomain buckets for sorting we have a
4582 * lock on the current crossfree bucket. A full matrix with
4583 * per-domain locking could be used if necessary.
4585 STAILQ_INIT(&emptybuckets);
4586 STAILQ_INIT(&fullbuckets);
4587 ZONE_CROSS_LOCK(zone);
4588 for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4589 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4590 domain = item_domain(item);
4591 zdom = ZDOM_GET(zone, domain);
4592 if (zdom->uzd_cross == NULL) {
4593 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4594 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4595 zdom->uzd_cross = b;
4598 * Avoid allocating a bucket with the cross lock
4599 * held, since allocation can trigger a
4600 * cross-domain free and bucket zones may
4601 * allocate from each other.
4603 ZONE_CROSS_UNLOCK(zone);
4604 b = bucket_alloc(zone, udata, M_NOWAIT);
4607 ZONE_CROSS_LOCK(zone);
4608 if (zdom->uzd_cross != NULL) {
4609 STAILQ_INSERT_HEAD(&emptybuckets, b,
4612 zdom->uzd_cross = b;
4616 b = zdom->uzd_cross;
4617 b->ub_bucket[b->ub_cnt++] = item;
4619 if (b->ub_cnt == b->ub_entries) {
4620 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4621 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4622 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4623 zdom->uzd_cross = b;
4626 ZONE_CROSS_UNLOCK(zone);
4628 if (bucket->ub_cnt == 0)
4629 bucket->ub_seq = SMR_SEQ_INVALID;
4630 bucket_free(zone, bucket, udata);
4632 while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4633 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4634 bucket_free(zone, b, udata);
4636 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4637 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4638 domain = item_domain(b->ub_bucket[0]);
4639 zone_put_bucket(zone, domain, b, udata, true);
4645 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4646 int itemdomain, bool ws)
4651 * Buckets coming from the wrong domain will be entirely for the
4652 * only other domain on two domain systems. In this case we can
4653 * simply cache them. Otherwise we need to sort them back to
4656 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4657 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4658 zone_free_cross(zone, bucket, udata);
4664 * Attempt to save the bucket in the zone's domain bucket cache.
4667 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4668 zone->uz_name, zone, bucket);
4669 /* ub_cnt is pointing to the last free item */
4670 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4671 itemdomain = zone_domain_lowest(zone, itemdomain);
4672 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4676 * Populate a free or cross bucket for the current cpu cache. Free any
4677 * existing full bucket either to the zone cache or back to the slab layer.
4679 * Enters and returns in a critical section. false return indicates that
4680 * we can not satisfy this free in the cache layer. true indicates that
4681 * the caller should retry.
4683 static __noinline bool
4684 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, int itemdomain)
4686 uma_cache_bucket_t cbucket;
4687 uma_bucket_t newbucket, bucket;
4689 CRITICAL_ASSERT(curthread);
4691 if (zone->uz_bucket_size == 0)
4694 cache = &zone->uz_cpu[curcpu];
4698 * FIRSTTOUCH domains need to free to the correct zdom. When
4699 * enabled this is the zdom of the item. The bucket is the
4700 * cross bucket if the current domain and itemdomain do not match.
4702 cbucket = &cache->uc_freebucket;
4704 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4705 if (PCPU_GET(domain) != itemdomain) {
4706 cbucket = &cache->uc_crossbucket;
4707 if (cbucket->ucb_cnt != 0)
4708 counter_u64_add(zone->uz_xdomain,
4713 bucket = cache_bucket_unload(cbucket);
4714 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4715 ("cache_free: Entered with non-full free bucket."));
4717 /* We are no longer associated with this CPU. */
4721 * Don't let SMR zones operate without a free bucket. Force
4722 * a synchronize and re-use this one. We will only degrade
4723 * to a synchronize every bucket_size items rather than every
4724 * item if we fail to allocate a bucket.
4726 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4728 bucket->ub_seq = smr_advance(zone->uz_smr);
4729 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4730 if (newbucket == NULL && bucket != NULL) {
4731 bucket_drain(zone, bucket);
4735 } else if (!bucketdisable)
4736 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4739 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4742 if ((bucket = newbucket) == NULL)
4744 cache = &zone->uz_cpu[curcpu];
4747 * Check to see if we should be populating the cross bucket. If it
4748 * is already populated we will fall through and attempt to populate
4751 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4752 if (PCPU_GET(domain) != itemdomain &&
4753 cache->uc_crossbucket.ucb_bucket == NULL) {
4754 cache_bucket_load_cross(cache, bucket);
4760 * We may have lost the race to fill the bucket or switched CPUs.
4762 if (cache->uc_freebucket.ucb_bucket != NULL) {
4764 bucket_free(zone, bucket, udata);
4767 cache_bucket_load_free(cache, bucket);
4773 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4780 KEG_LOCK_ASSERT(keg, slab->us_domain);
4782 /* Do we need to remove from any lists? */
4783 dom = &keg->uk_domain[slab->us_domain];
4784 if (slab->us_freecount + 1 == keg->uk_ipers) {
4785 LIST_REMOVE(slab, us_link);
4786 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4787 dom->ud_free_slabs++;
4788 } else if (slab->us_freecount == 0) {
4789 LIST_REMOVE(slab, us_link);
4790 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4793 /* Slab management. */
4794 freei = slab_item_index(slab, keg, item);
4795 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4796 slab->us_freecount++;
4798 /* Keg statistics. */
4799 dom->ud_free_items++;
4803 zone_release(void *arg, void **bucket, int cnt)
4816 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4817 lock = KEG_LOCK(keg, 0);
4818 for (i = 0; i < cnt; i++) {
4820 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4821 slab = vtoslab((vm_offset_t)item);
4823 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4824 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4825 slab = hash_sfind(&keg->uk_hash, mem);
4827 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4829 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4832 lock = KEG_LOCK(keg, slab->us_domain);
4834 slab_free_item(zone, slab, item);
4841 * Frees a single item to any zone.
4844 * zone The zone to free to
4845 * item The item we're freeing
4846 * udata User supplied data for the dtor
4847 * skip Skip dtors and finis
4849 static __noinline void
4850 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4854 * If a free is sent directly to an SMR zone we have to
4855 * synchronize immediately because the item can instantly
4856 * be reallocated. This should only happen in degenerate
4857 * cases when no memory is available for per-cpu caches.
4859 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4860 smr_synchronize(zone->uz_smr);
4862 item_dtor(zone, item, zone->uz_size, udata, skip);
4864 if (skip < SKIP_FINI && zone->uz_fini) {
4865 kasan_mark_item_valid(zone, item);
4866 zone->uz_fini(item, zone->uz_size);
4867 kasan_mark_item_invalid(zone, item);
4870 zone->uz_release(zone->uz_arg, &item, 1);
4872 if (skip & SKIP_CNT)
4875 counter_u64_add(zone->uz_frees, 1);
4877 if (zone->uz_max_items > 0)
4878 zone_free_limit(zone, 1);
4883 uma_zone_set_max(uma_zone_t zone, int nitems)
4887 * If the limit is small, we may need to constrain the maximum per-CPU
4888 * cache size, or disable caching entirely.
4890 uma_zone_set_maxcache(zone, nitems);
4893 * XXX This can misbehave if the zone has any allocations with
4894 * no limit and a limit is imposed. There is currently no
4895 * way to clear a limit.
4898 if (zone->uz_max_items == 0)
4899 ZONE_ASSERT_COLD(zone);
4900 zone->uz_max_items = nitems;
4901 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4902 zone_update_caches(zone);
4903 /* We may need to wake waiters. */
4904 wakeup(&zone->uz_max_items);
4912 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4914 int bpcpu, bpdom, bsize, nb;
4919 * Compute a lower bound on the number of items that may be cached in
4920 * the zone. Each CPU gets at least two buckets, and for cross-domain
4921 * frees we use an additional bucket per CPU and per domain. Select the
4922 * largest bucket size that does not exceed half of the requested limit,
4923 * with the left over space given to the full bucket cache.
4928 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4933 nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4934 bsize = nitems / nb / 2;
4935 if (bsize > BUCKET_MAX)
4937 else if (bsize == 0 && nitems / nb > 0)
4939 zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
4940 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4941 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4942 zone->uz_bucket_max = nitems - nb * bsize;
4948 uma_zone_get_max(uma_zone_t zone)
4952 nitems = atomic_load_64(&zone->uz_max_items);
4959 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4962 ZONE_ASSERT_COLD(zone);
4963 zone->uz_warning = warning;
4968 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4971 ZONE_ASSERT_COLD(zone);
4972 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4977 uma_zone_get_cur(uma_zone_t zone)
4983 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4984 nitems = counter_u64_fetch(zone->uz_allocs) -
4985 counter_u64_fetch(zone->uz_frees);
4987 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4988 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4990 return (nitems < 0 ? 0 : nitems);
4994 uma_zone_get_allocs(uma_zone_t zone)
5000 if (zone->uz_allocs != EARLY_COUNTER)
5001 nitems = counter_u64_fetch(zone->uz_allocs);
5003 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
5009 uma_zone_get_frees(uma_zone_t zone)
5015 if (zone->uz_frees != EARLY_COUNTER)
5016 nitems = counter_u64_fetch(zone->uz_frees);
5018 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
5024 /* Used only for KEG_ASSERT_COLD(). */
5026 uma_keg_get_allocs(uma_keg_t keg)
5032 LIST_FOREACH(z, &keg->uk_zones, uz_link)
5033 nitems += uma_zone_get_allocs(z);
5041 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
5046 KEG_ASSERT_COLD(keg);
5047 keg->uk_init = uminit;
5052 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
5057 KEG_ASSERT_COLD(keg);
5058 keg->uk_fini = fini;
5063 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
5066 ZONE_ASSERT_COLD(zone);
5067 zone->uz_init = zinit;
5072 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
5075 ZONE_ASSERT_COLD(zone);
5076 zone->uz_fini = zfini;
5081 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
5086 KEG_ASSERT_COLD(keg);
5087 keg->uk_freef = freef;
5092 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
5097 KEG_ASSERT_COLD(keg);
5098 keg->uk_allocf = allocf;
5103 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
5106 ZONE_ASSERT_COLD(zone);
5108 KASSERT(smr != NULL, ("Got NULL smr"));
5109 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
5110 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
5111 zone->uz_flags |= UMA_ZONE_SMR;
5113 zone_update_caches(zone);
5117 uma_zone_get_smr(uma_zone_t zone)
5120 return (zone->uz_smr);
5125 uma_zone_reserve(uma_zone_t zone, int items)
5130 KEG_ASSERT_COLD(keg);
5131 keg->uk_reserve = items;
5136 uma_zone_reserve_kva(uma_zone_t zone, int count)
5143 KEG_ASSERT_COLD(keg);
5144 ZONE_ASSERT_COLD(zone);
5146 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
5148 #ifdef UMA_MD_SMALL_ALLOC
5149 if (keg->uk_ppera > 1) {
5153 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
5159 MPASS(keg->uk_kva == 0);
5162 zone->uz_max_items = pages * keg->uk_ipers;
5163 #ifdef UMA_MD_SMALL_ALLOC
5164 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
5166 keg->uk_allocf = noobj_alloc;
5168 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5169 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5170 zone_update_caches(zone);
5177 uma_prealloc(uma_zone_t zone, int items)
5179 struct vm_domainset_iter di;
5183 int aflags, domain, slabs;
5186 slabs = howmany(items, keg->uk_ipers);
5187 while (slabs-- > 0) {
5189 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
5192 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
5195 dom = &keg->uk_domain[slab->us_domain];
5197 * keg_alloc_slab() always returns a slab on the
5200 LIST_REMOVE(slab, us_link);
5201 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
5203 dom->ud_free_slabs++;
5204 KEG_UNLOCK(keg, slab->us_domain);
5207 if (vm_domainset_iter_policy(&di, &domain) != 0)
5208 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
5214 * Returns a snapshot of memory consumption in bytes.
5217 uma_zone_memory(uma_zone_t zone)
5223 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
5224 for (i = 0; i < vm_ndomains; i++)
5225 sz += ZDOM_GET(zone, i)->uzd_nitems;
5226 return (sz * zone->uz_size);
5228 for (i = 0; i < vm_ndomains; i++)
5229 sz += zone->uz_keg->uk_domain[i].ud_pages;
5231 return (sz * PAGE_SIZE);
5234 struct uma_reclaim_args {
5240 uma_reclaim_domain_cb(uma_zone_t zone, void *arg)
5242 struct uma_reclaim_args *args;
5245 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0)
5246 uma_zone_reclaim_domain(zone, args->req, args->domain);
5251 uma_reclaim(int req)
5253 uma_reclaim_domain(req, UMA_ANYDOMAIN);
5257 uma_reclaim_domain(int req, int domain)
5259 struct uma_reclaim_args args;
5263 args.domain = domain;
5266 sx_slock(&uma_reclaim_lock);
5268 case UMA_RECLAIM_TRIM:
5269 case UMA_RECLAIM_DRAIN:
5270 zone_foreach(uma_reclaim_domain_cb, &args);
5272 case UMA_RECLAIM_DRAIN_CPU:
5273 zone_foreach(uma_reclaim_domain_cb, &args);
5274 pcpu_cache_drain_safe(NULL);
5275 zone_foreach(uma_reclaim_domain_cb, &args);
5278 panic("unhandled reclamation request %d", req);
5282 * Some slabs may have been freed but this zone will be visited early
5283 * we visit again so that we can free pages that are empty once other
5284 * zones are drained. We have to do the same for buckets.
5286 uma_zone_reclaim_domain(slabzones[0], UMA_RECLAIM_DRAIN, domain);
5287 uma_zone_reclaim_domain(slabzones[1], UMA_RECLAIM_DRAIN, domain);
5288 bucket_zone_drain(domain);
5289 sx_sunlock(&uma_reclaim_lock);
5292 static volatile int uma_reclaim_needed;
5295 uma_reclaim_wakeup(void)
5298 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5299 wakeup(uma_reclaim);
5303 uma_reclaim_worker(void *arg __unused)
5307 sx_xlock(&uma_reclaim_lock);
5308 while (atomic_load_int(&uma_reclaim_needed) == 0)
5309 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5311 sx_xunlock(&uma_reclaim_lock);
5312 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5313 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5314 atomic_store_int(&uma_reclaim_needed, 0);
5315 /* Don't fire more than once per-second. */
5316 pause("umarclslp", hz);
5322 uma_zone_reclaim(uma_zone_t zone, int req)
5324 uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5328 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5331 case UMA_RECLAIM_TRIM:
5332 zone_reclaim(zone, domain, M_NOWAIT, false);
5334 case UMA_RECLAIM_DRAIN:
5335 zone_reclaim(zone, domain, M_NOWAIT, true);
5337 case UMA_RECLAIM_DRAIN_CPU:
5338 pcpu_cache_drain_safe(zone);
5339 zone_reclaim(zone, domain, M_NOWAIT, true);
5342 panic("unhandled reclamation request %d", req);
5348 uma_zone_exhausted(uma_zone_t zone)
5351 return (atomic_load_32(&zone->uz_sleepers) > 0);
5358 return (uma_kmem_limit);
5362 uma_set_limit(unsigned long limit)
5365 uma_kmem_limit = limit;
5372 return (atomic_load_long(&uma_kmem_total));
5379 return (uma_kmem_limit - uma_size());
5384 * Generate statistics across both the zone and its per-cpu cache's. Return
5385 * desired statistics if the pointer is non-NULL for that statistic.
5387 * Note: does not update the zone statistics, as it can't safely clear the
5388 * per-CPU cache statistic.
5392 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5393 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5396 uint64_t allocs, frees, sleeps, xdomain;
5399 allocs = frees = sleeps = xdomain = 0;
5402 cache = &z->uz_cpu[cpu];
5403 cachefree += cache->uc_allocbucket.ucb_cnt;
5404 cachefree += cache->uc_freebucket.ucb_cnt;
5405 xdomain += cache->uc_crossbucket.ucb_cnt;
5406 cachefree += cache->uc_crossbucket.ucb_cnt;
5407 allocs += cache->uc_allocs;
5408 frees += cache->uc_frees;
5410 allocs += counter_u64_fetch(z->uz_allocs);
5411 frees += counter_u64_fetch(z->uz_frees);
5412 xdomain += counter_u64_fetch(z->uz_xdomain);
5413 sleeps += z->uz_sleeps;
5414 if (cachefreep != NULL)
5415 *cachefreep = cachefree;
5416 if (allocsp != NULL)
5420 if (sleepsp != NULL)
5422 if (xdomainp != NULL)
5423 *xdomainp = xdomain;
5428 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5435 rw_rlock(&uma_rwlock);
5436 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5437 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5440 LIST_FOREACH(z, &uma_cachezones, uz_link)
5443 rw_runlock(&uma_rwlock);
5444 return (sysctl_handle_int(oidp, &count, 0, req));
5448 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5449 struct uma_percpu_stat *ups, bool internal)
5451 uma_zone_domain_t zdom;
5455 for (i = 0; i < vm_ndomains; i++) {
5456 zdom = ZDOM_GET(z, i);
5457 uth->uth_zone_free += zdom->uzd_nitems;
5459 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5460 uth->uth_frees = counter_u64_fetch(z->uz_frees);
5461 uth->uth_fails = counter_u64_fetch(z->uz_fails);
5462 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5463 uth->uth_sleeps = z->uz_sleeps;
5465 for (i = 0; i < mp_maxid + 1; i++) {
5466 bzero(&ups[i], sizeof(*ups));
5467 if (internal || CPU_ABSENT(i))
5469 cache = &z->uz_cpu[i];
5470 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5471 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5472 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5473 ups[i].ups_allocs = cache->uc_allocs;
5474 ups[i].ups_frees = cache->uc_frees;
5479 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5481 struct uma_stream_header ush;
5482 struct uma_type_header uth;
5483 struct uma_percpu_stat *ups;
5488 uint32_t kfree, pages;
5489 int count, error, i;
5491 error = sysctl_wire_old_buffer(req, 0);
5494 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5495 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5496 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5499 rw_rlock(&uma_rwlock);
5500 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5501 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5505 LIST_FOREACH(z, &uma_cachezones, uz_link)
5509 * Insert stream header.
5511 bzero(&ush, sizeof(ush));
5512 ush.ush_version = UMA_STREAM_VERSION;
5513 ush.ush_maxcpus = (mp_maxid + 1);
5514 ush.ush_count = count;
5515 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5517 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5519 for (i = 0; i < vm_ndomains; i++) {
5520 kfree += kz->uk_domain[i].ud_free_items;
5521 pages += kz->uk_domain[i].ud_pages;
5523 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5524 bzero(&uth, sizeof(uth));
5525 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5526 uth.uth_align = kz->uk_align;
5527 uth.uth_size = kz->uk_size;
5528 uth.uth_rsize = kz->uk_rsize;
5529 if (z->uz_max_items > 0) {
5530 items = UZ_ITEMS_COUNT(z->uz_items);
5531 uth.uth_pages = (items / kz->uk_ipers) *
5534 uth.uth_pages = pages;
5535 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5537 uth.uth_limit = z->uz_max_items;
5538 uth.uth_keg_free = kfree;
5541 * A zone is secondary is it is not the first entry
5542 * on the keg's zone list.
5544 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5545 (LIST_FIRST(&kz->uk_zones) != z))
5546 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5547 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5548 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5549 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5550 for (i = 0; i < mp_maxid + 1; i++)
5551 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5554 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5555 bzero(&uth, sizeof(uth));
5556 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5557 uth.uth_size = z->uz_size;
5558 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5559 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5560 for (i = 0; i < mp_maxid + 1; i++)
5561 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5564 rw_runlock(&uma_rwlock);
5565 error = sbuf_finish(&sbuf);
5572 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5574 uma_zone_t zone = *(uma_zone_t *)arg1;
5577 max = uma_zone_get_max(zone);
5578 error = sysctl_handle_int(oidp, &max, 0, req);
5579 if (error || !req->newptr)
5582 uma_zone_set_max(zone, max);
5588 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5594 * Some callers want to add sysctls for global zones that
5595 * may not yet exist so they pass a pointer to a pointer.
5598 zone = *(uma_zone_t *)arg1;
5601 cur = uma_zone_get_cur(zone);
5602 return (sysctl_handle_int(oidp, &cur, 0, req));
5606 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5608 uma_zone_t zone = arg1;
5611 cur = uma_zone_get_allocs(zone);
5612 return (sysctl_handle_64(oidp, &cur, 0, req));
5616 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5618 uma_zone_t zone = arg1;
5621 cur = uma_zone_get_frees(zone);
5622 return (sysctl_handle_64(oidp, &cur, 0, req));
5626 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5629 uma_zone_t zone = arg1;
5632 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5633 if (zone->uz_flags != 0)
5634 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5636 sbuf_printf(&sbuf, "0");
5637 error = sbuf_finish(&sbuf);
5644 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5646 uma_keg_t keg = arg1;
5647 int avail, effpct, total;
5649 total = keg->uk_ppera * PAGE_SIZE;
5650 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5651 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5653 * We consider the client's requested size and alignment here, not the
5654 * real size determination uk_rsize, because we also adjust the real
5655 * size for internal implementation reasons (max bitset size).
5657 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5658 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5659 avail *= mp_maxid + 1;
5660 effpct = 100 * avail / total;
5661 return (sysctl_handle_int(oidp, &effpct, 0, req));
5665 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5667 uma_zone_t zone = arg1;
5670 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5671 return (sysctl_handle_64(oidp, &cur, 0, req));
5676 uma_dbg_getslab(uma_zone_t zone, void *item)
5683 * It is safe to return the slab here even though the
5684 * zone is unlocked because the item's allocation state
5685 * essentially holds a reference.
5687 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5688 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5690 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5691 return (vtoslab((vm_offset_t)mem));
5693 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5694 return ((uma_slab_t)(mem + keg->uk_pgoff));
5696 slab = hash_sfind(&keg->uk_hash, mem);
5703 uma_dbg_zskip(uma_zone_t zone, void *mem)
5706 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5709 return (uma_dbg_kskip(zone->uz_keg, mem));
5713 uma_dbg_kskip(uma_keg_t keg, void *mem)
5717 if (dbg_divisor == 0)
5720 if (dbg_divisor == 1)
5723 idx = (uintptr_t)mem >> PAGE_SHIFT;
5724 if (keg->uk_ipers > 1) {
5725 idx *= keg->uk_ipers;
5726 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5729 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5730 counter_u64_add(uma_skip_cnt, 1);
5733 counter_u64_add(uma_dbg_cnt, 1);
5739 * Set up the slab's freei data such that uma_dbg_free can function.
5743 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5749 slab = uma_dbg_getslab(zone, item);
5751 panic("uma: item %p did not belong to zone %s",
5752 item, zone->uz_name);
5755 freei = slab_item_index(slab, keg, item);
5757 if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5758 slab_dbg_bits(slab, keg)))
5759 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5760 item, zone, zone->uz_name, slab, freei);
5764 * Verifies freed addresses. Checks for alignment, valid slab membership
5765 * and duplicate frees.
5769 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5775 slab = uma_dbg_getslab(zone, item);
5777 panic("uma: Freed item %p did not belong to zone %s",
5778 item, zone->uz_name);
5781 freei = slab_item_index(slab, keg, item);
5783 if (freei >= keg->uk_ipers)
5784 panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5785 item, zone, zone->uz_name, slab, freei);
5787 if (slab_item(slab, keg, freei) != item)
5788 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5789 item, zone, zone->uz_name, slab, freei);
5791 if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5792 slab_dbg_bits(slab, keg)))
5793 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5794 item, zone, zone->uz_name, slab, freei);
5796 #endif /* INVARIANTS */
5800 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5801 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5806 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5807 *allocs = counter_u64_fetch(z->uz_allocs);
5808 frees = counter_u64_fetch(z->uz_frees);
5809 *sleeps = z->uz_sleeps;
5813 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5815 for (i = 0; i < vm_ndomains; i++) {
5816 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5817 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5818 (LIST_FIRST(&kz->uk_zones) != z)))
5819 *cachefree += kz->uk_domain[i].ud_free_items;
5821 *used = *allocs - frees;
5822 return (((int64_t)*used + *cachefree) * kz->uk_size);
5825 DB_SHOW_COMMAND(uma, db_show_uma)
5827 const char *fmt_hdr, *fmt_entry;
5830 uint64_t allocs, used, sleeps, xdomain;
5832 /* variables for sorting */
5834 uma_zone_t cur_zone, last_zone;
5835 int64_t cur_size, last_size, size;
5838 /* /i option produces machine-parseable CSV output */
5839 if (modif[0] == 'i') {
5840 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5841 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5843 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5844 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5847 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5848 "Sleeps", "Bucket", "Total Mem", "XFree");
5850 /* Sort the zones with largest size first. */
5852 last_size = INT64_MAX;
5857 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5858 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5860 * In the case of size ties, print out zones
5861 * in the order they are encountered. That is,
5862 * when we encounter the most recently output
5863 * zone, we have already printed all preceding
5864 * ties, and we must print all following ties.
5866 if (z == last_zone) {
5870 size = get_uma_stats(kz, z, &allocs, &used,
5871 &sleeps, &cachefree, &xdomain);
5872 if (size > cur_size && size < last_size + ties)
5880 if (cur_zone == NULL)
5883 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5884 &sleeps, &cachefree, &xdomain);
5885 db_printf(fmt_entry, cur_zone->uz_name,
5886 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5887 (uintmax_t)allocs, (uintmax_t)sleeps,
5888 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5893 last_zone = cur_zone;
5894 last_size = cur_size;
5898 DB_SHOW_COMMAND(umacache, db_show_umacache)
5901 uint64_t allocs, frees;
5905 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5906 "Requests", "Bucket");
5907 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5908 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5909 for (i = 0; i < vm_ndomains; i++)
5910 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5911 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5912 z->uz_name, (uintmax_t)z->uz_size,
5913 (intmax_t)(allocs - frees), cachefree,
5914 (uintmax_t)allocs, z->uz_bucket_size);