2 * SPDX-License-Identifier: BSD-2-Clause
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>
54 #include "opt_param.h"
57 #include <sys/param.h>
58 #include <sys/systm.h>
60 #include <sys/bitset.h>
61 #include <sys/domainset.h>
62 #include <sys/eventhandler.h>
63 #include <sys/kernel.h>
64 #include <sys/types.h>
65 #include <sys/limits.h>
66 #include <sys/queue.h>
67 #include <sys/malloc.h>
71 #include <sys/mutex.h>
73 #include <sys/random.h>
74 #include <sys/rwlock.h>
76 #include <sys/sched.h>
77 #include <sys/sleepqueue.h>
80 #include <sys/sysctl.h>
81 #include <sys/taskqueue.h>
82 #include <sys/vmmeter.h>
85 #include <vm/vm_param.h>
86 #include <vm/vm_domainset.h>
87 #include <vm/vm_object.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_pageout.h>
90 #include <vm/vm_phys.h>
91 #include <vm/vm_pagequeue.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
95 #include <vm/vm_dumpset.h>
97 #include <vm/uma_int.h>
98 #include <vm/uma_dbg.h>
102 #ifdef DEBUG_MEMGUARD
103 #include <vm/memguard.h>
106 #include <machine/md_var.h>
109 #define UMA_ALWAYS_CTORDTOR 1
111 #define UMA_ALWAYS_CTORDTOR 0
115 * This is the zone and keg from which all zones are spawned.
117 static uma_zone_t kegs;
118 static uma_zone_t zones;
121 * On INVARIANTS builds, the slab contains a second bitset of the same size,
122 * "dbg_bits", which is laid out immediately after us_free.
125 #define SLAB_BITSETS 2
127 #define SLAB_BITSETS 1
131 * These are the two zones from which all offpage uma_slab_ts are allocated.
133 * One zone is for slab headers that can represent a larger number of items,
134 * making the slabs themselves more efficient, and the other zone is for
135 * headers that are smaller and represent fewer items, making the headers more
138 #define SLABZONE_SIZE(setsize) \
139 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
140 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16)
141 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE
142 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE)
143 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE)
144 static uma_zone_t slabzones[2];
147 * The initial hash tables come out of this zone so they can be allocated
148 * prior to malloc coming up.
150 static uma_zone_t hashzone;
152 /* The boot-time adjusted value for cache line alignment. */
153 int uma_align_cache = 64 - 1;
155 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
156 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
159 * Are we allowed to allocate buckets?
161 static int bucketdisable = 1;
163 /* Linked list of all kegs in the system */
164 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
166 /* Linked list of all cache-only zones in the system */
167 static LIST_HEAD(,uma_zone) uma_cachezones =
168 LIST_HEAD_INITIALIZER(uma_cachezones);
171 * Mutex for global lists: uma_kegs, uma_cachezones, and the per-keg list of
174 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
176 static struct sx uma_reclaim_lock;
179 * First available virual address for boot time allocations.
181 static vm_offset_t bootstart;
182 static vm_offset_t bootmem;
185 * kmem soft limit, initialized by uma_set_limit(). Ensure that early
186 * allocations don't trigger a wakeup of the reclaim thread.
188 unsigned long uma_kmem_limit = LONG_MAX;
189 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
190 "UMA kernel memory soft limit");
191 unsigned long uma_kmem_total;
192 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
193 "UMA kernel memory usage");
195 /* Is the VM done starting up? */
202 } booted = BOOT_COLD;
205 * This is the handle used to schedule events that need to happen
206 * outside of the allocation fast path.
208 static struct timeout_task uma_timeout_task;
209 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
212 * This structure is passed as the zone ctor arg so that I don't have to create
213 * a special allocation function just for zones.
215 struct uma_zctor_args {
230 struct uma_kctor_args {
239 struct uma_bucket_zone {
241 const char *ubz_name;
242 int ubz_entries; /* Number of items it can hold. */
243 int ubz_maxsize; /* Maximum allocation size per-item. */
247 * Compute the actual number of bucket entries to pack them in power
248 * of two sizes for more efficient space utilization.
250 #define BUCKET_SIZE(n) \
251 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
253 #define BUCKET_MAX BUCKET_SIZE(256)
255 struct uma_bucket_zone bucket_zones[] = {
256 /* Literal bucket sizes. */
257 { NULL, "2 Bucket", 2, 4096 },
258 { NULL, "4 Bucket", 4, 3072 },
259 { NULL, "8 Bucket", 8, 2048 },
260 { NULL, "16 Bucket", 16, 1024 },
261 /* Rounded down power of 2 sizes for efficiency. */
262 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
263 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
264 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
265 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
270 * Flags and enumerations to be passed to internal functions.
274 SKIP_CNT = 0x00000001,
275 SKIP_DTOR = 0x00010000,
276 SKIP_FINI = 0x00020000,
281 void uma_startup1(vm_offset_t);
282 void uma_startup2(void);
284 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
285 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
286 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
287 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
288 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
289 static void page_free(void *, vm_size_t, uint8_t);
290 static void pcpu_page_free(void *, vm_size_t, uint8_t);
291 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
292 static void cache_drain(uma_zone_t);
293 static void bucket_drain(uma_zone_t, uma_bucket_t);
294 static void bucket_cache_reclaim(uma_zone_t zone, bool, int);
295 static bool bucket_cache_reclaim_domain(uma_zone_t, bool, bool, int);
296 static int keg_ctor(void *, int, void *, int);
297 static void keg_dtor(void *, int, void *);
298 static void keg_drain(uma_keg_t keg, int domain);
299 static int zone_ctor(void *, int, void *, int);
300 static void zone_dtor(void *, int, void *);
301 static inline void item_dtor(uma_zone_t zone, void *item, int size,
302 void *udata, enum zfreeskip skip);
303 static int zero_init(void *, int, int);
304 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
305 int itemdomain, bool ws);
306 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
307 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
308 static void zone_timeout(uma_zone_t zone, void *);
309 static int hash_alloc(struct uma_hash *, u_int);
310 static int hash_expand(struct uma_hash *, struct uma_hash *);
311 static void hash_free(struct uma_hash *hash);
312 static void uma_timeout(void *, int);
313 static void uma_shutdown(void);
314 static void *zone_alloc_item(uma_zone_t, void *, int, int);
315 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
316 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
317 static void zone_free_limit(uma_zone_t zone, int count);
318 static void bucket_enable(void);
319 static void bucket_init(void);
320 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
321 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
322 static void bucket_zone_drain(int domain);
323 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
324 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
325 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
326 static size_t slab_sizeof(int nitems);
327 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
328 uma_fini fini, int align, uint32_t flags);
329 static int zone_import(void *, void **, int, int, int);
330 static void zone_release(void *, void **, int);
331 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
332 static bool cache_free(uma_zone_t, uma_cache_t, void *, int);
334 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
335 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
336 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
337 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
338 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
339 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
340 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
342 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
344 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
345 "Memory allocation debugging");
348 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
349 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
351 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
352 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
353 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
354 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
356 static u_int dbg_divisor = 1;
357 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
358 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
359 "Debug & thrash every this item in memory allocator");
361 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
362 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
363 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
364 &uma_dbg_cnt, "memory items debugged");
365 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
366 &uma_skip_cnt, "memory items skipped, not debugged");
369 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
370 "Universal Memory Allocator");
372 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
373 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
375 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
376 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
378 static int zone_warnings = 1;
379 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
380 "Warn when UMA zones becomes full");
382 static int multipage_slabs = 1;
383 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
384 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
385 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
386 "UMA may choose larger slab sizes for better efficiency");
389 * Select the slab zone for an offpage slab with the given maximum item count.
391 static inline uma_zone_t
395 return (slabzones[ipers > SLABZONE0_SETSIZE]);
399 * This routine checks to see whether or not it's safe to enable buckets.
405 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
406 bucketdisable = vm_page_count_min();
410 * Initialize bucket_zones, the array of zones of buckets of various sizes.
412 * For each zone, calculate the memory required for each bucket, consisting
413 * of the header and an array of pointers.
418 struct uma_bucket_zone *ubz;
421 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
422 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
423 size += sizeof(void *) * ubz->ubz_entries;
424 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
425 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
426 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
427 UMA_ZONE_FIRSTTOUCH);
432 * Given a desired number of entries for a bucket, return the zone from which
433 * to allocate the bucket.
435 static struct uma_bucket_zone *
436 bucket_zone_lookup(int entries)
438 struct uma_bucket_zone *ubz;
440 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
441 if (ubz->ubz_entries >= entries)
448 bucket_select(int size)
450 struct uma_bucket_zone *ubz;
452 ubz = &bucket_zones[0];
453 if (size > ubz->ubz_maxsize)
454 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
456 for (; ubz->ubz_entries != 0; ubz++)
457 if (ubz->ubz_maxsize < size)
460 return (ubz->ubz_entries);
464 bucket_alloc(uma_zone_t zone, void *udata, int flags)
466 struct uma_bucket_zone *ubz;
470 * Don't allocate buckets early in boot.
472 if (__predict_false(booted < BOOT_KVA))
476 * To limit bucket recursion we store the original zone flags
477 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
478 * NOVM flag to persist even through deep recursions. We also
479 * store ZFLAG_BUCKET once we have recursed attempting to allocate
480 * a bucket for a bucket zone so we do not allow infinite bucket
481 * recursion. This cookie will even persist to frees of unused
482 * buckets via the allocation path or bucket allocations in the
485 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
486 udata = (void *)(uintptr_t)zone->uz_flags;
488 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
490 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
492 if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
494 ubz = bucket_zone_lookup(atomic_load_16(&zone->uz_bucket_size));
495 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
497 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
500 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
503 bucket->ub_entries = min(ubz->ubz_entries,
504 zone->uz_bucket_size_max);
505 bucket->ub_seq = SMR_SEQ_INVALID;
506 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
507 zone->uz_name, zone, bucket);
514 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
516 struct uma_bucket_zone *ubz;
518 if (bucket->ub_cnt != 0)
519 bucket_drain(zone, bucket);
521 KASSERT(bucket->ub_cnt == 0,
522 ("bucket_free: Freeing a non free bucket."));
523 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
524 ("bucket_free: Freeing an SMR bucket."));
525 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
526 udata = (void *)(uintptr_t)zone->uz_flags;
527 ubz = bucket_zone_lookup(bucket->ub_entries);
528 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
532 bucket_zone_drain(int domain)
534 struct uma_bucket_zone *ubz;
536 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
537 uma_zone_reclaim_domain(ubz->ubz_zone, UMA_RECLAIM_DRAIN,
542 _Static_assert(UMA_SMALLEST_UNIT % KASAN_SHADOW_SCALE == 0,
543 "Base UMA allocation size not a multiple of the KASAN scale factor");
546 kasan_mark_item_valid(uma_zone_t zone, void *item)
552 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
556 rsz = roundup2(sz, KASAN_SHADOW_SCALE);
557 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
558 kasan_mark(item, sz, rsz, KASAN_GENERIC_REDZONE);
560 pcpu_item = zpcpu_base_to_offset(item);
561 for (i = 0; i <= mp_maxid; i++)
562 kasan_mark(zpcpu_get_cpu(pcpu_item, i), sz, rsz,
563 KASAN_GENERIC_REDZONE);
568 kasan_mark_item_invalid(uma_zone_t zone, void *item)
574 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
577 sz = roundup2(zone->uz_size, KASAN_SHADOW_SCALE);
578 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
579 kasan_mark(item, 0, sz, KASAN_UMA_FREED);
581 pcpu_item = zpcpu_base_to_offset(item);
582 for (i = 0; i <= mp_maxid; i++)
583 kasan_mark(zpcpu_get_cpu(pcpu_item, i), 0, sz,
589 kasan_mark_slab_valid(uma_keg_t keg, void *mem)
593 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
594 sz = keg->uk_ppera * PAGE_SIZE;
595 kasan_mark(mem, sz, sz, 0);
600 kasan_mark_slab_invalid(uma_keg_t keg, void *mem)
604 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
605 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
606 sz = keg->uk_ppera * PAGE_SIZE;
609 kasan_mark(mem, 0, sz, KASAN_UMA_FREED);
614 kasan_mark_item_valid(uma_zone_t zone __unused, void *item __unused)
619 kasan_mark_item_invalid(uma_zone_t zone __unused, void *item __unused)
624 kasan_mark_slab_valid(uma_keg_t keg __unused, void *mem __unused)
629 kasan_mark_slab_invalid(uma_keg_t keg __unused, void *mem __unused)
636 kmsan_mark_item_uninitialized(uma_zone_t zone, void *item)
642 if ((zone->uz_flags &
643 (UMA_ZFLAG_CACHE | UMA_ZONE_SECONDARY | UMA_ZONE_MALLOC)) != 0) {
645 * Cache zones should not be instrumented by default, as UMA
646 * does not have enough information to do so correctly.
647 * Consumers can mark items themselves if it makes sense to do
650 * Items from secondary zones are initialized by the parent
651 * zone and thus cannot safely be marked by UMA.
653 * malloc zones are handled directly by malloc(9) and friends,
654 * since they can provide more precise origin tracking.
658 if (zone->uz_keg->uk_init != NULL) {
660 * By definition, initialized items cannot be marked. The
661 * best we can do is mark items from these zones after they
662 * are freed to the keg.
668 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
669 kmsan_orig(item, sz, KMSAN_TYPE_UMA, KMSAN_RET_ADDR);
670 kmsan_mark(item, sz, KMSAN_STATE_UNINIT);
672 pcpu_item = zpcpu_base_to_offset(item);
673 for (i = 0; i <= mp_maxid; i++) {
674 kmsan_orig(zpcpu_get_cpu(pcpu_item, i), sz,
675 KMSAN_TYPE_UMA, KMSAN_RET_ADDR);
676 kmsan_mark(zpcpu_get_cpu(pcpu_item, i), sz,
683 kmsan_mark_item_uninitialized(uma_zone_t zone __unused, void *item __unused)
689 * Acquire the domain lock and record contention.
691 static uma_zone_domain_t
692 zone_domain_lock(uma_zone_t zone, int domain)
694 uma_zone_domain_t zdom;
697 zdom = ZDOM_GET(zone, domain);
699 if (ZDOM_OWNED(zdom))
702 /* This is unsynchronized. The counter does not need to be precise. */
703 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
704 zone->uz_bucket_size++;
709 * Search for the domain with the least cached items and return it if it
710 * is out of balance with the preferred domain.
712 static __noinline int
713 zone_domain_lowest(uma_zone_t zone, int pref)
715 long least, nitems, prefitems;
719 prefitems = least = LONG_MAX;
721 for (i = 0; i < vm_ndomains; i++) {
722 nitems = ZDOM_GET(zone, i)->uzd_nitems;
723 if (nitems < least) {
730 if (prefitems < least * 2)
737 * Search for the domain with the most cached items and return it or the
738 * preferred domain if it has enough to proceed.
740 static __noinline int
741 zone_domain_highest(uma_zone_t zone, int pref)
747 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
752 for (i = 0; i < vm_ndomains; i++) {
753 nitems = ZDOM_GET(zone, i)->uzd_nitems;
764 * Set the maximum imax value.
767 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
771 old = zdom->uzd_imax;
775 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
778 * We are at new maximum, so do the last WSS update for the old
779 * bimin and prepare to measure next allocation batch.
781 if (zdom->uzd_wss < old - zdom->uzd_bimin)
782 zdom->uzd_wss = old - zdom->uzd_bimin;
783 zdom->uzd_bimin = nitems;
787 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
788 * zone's caches. If a bucket is found the zone is not locked on return.
791 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
798 ZDOM_LOCK_ASSERT(zdom);
800 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
803 /* SMR Buckets can not be re-used until readers expire. */
804 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
805 bucket->ub_seq != SMR_SEQ_INVALID) {
806 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
808 bucket->ub_seq = SMR_SEQ_INVALID;
809 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
810 if (STAILQ_NEXT(bucket, ub_link) != NULL)
811 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
813 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
815 KASSERT(zdom->uzd_nitems >= bucket->ub_cnt,
816 ("%s: item count underflow (%ld, %d)",
817 __func__, zdom->uzd_nitems, bucket->ub_cnt));
818 KASSERT(bucket->ub_cnt > 0,
819 ("%s: empty bucket in bucket cache", __func__));
820 zdom->uzd_nitems -= bucket->ub_cnt;
824 * Shift the bounds of the current WSS interval to avoid
825 * perturbing the estimates.
827 cnt = lmin(zdom->uzd_bimin, bucket->ub_cnt);
828 atomic_subtract_long(&zdom->uzd_imax, cnt);
829 zdom->uzd_bimin -= cnt;
830 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
831 if (zdom->uzd_limin >= bucket->ub_cnt) {
832 zdom->uzd_limin -= bucket->ub_cnt;
837 } else if (zdom->uzd_bimin > zdom->uzd_nitems) {
838 zdom->uzd_bimin = zdom->uzd_nitems;
839 if (zdom->uzd_imin > zdom->uzd_nitems)
840 zdom->uzd_imin = zdom->uzd_nitems;
845 for (i = 0; i < bucket->ub_cnt; i++)
846 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
853 * Insert a full bucket into the specified cache. The "ws" parameter indicates
854 * whether the bucket's contents should be counted as part of the zone's working
855 * set. The bucket may be freed if it exceeds the bucket limit.
858 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
861 uma_zone_domain_t zdom;
863 /* We don't cache empty buckets. This can happen after a reclaim. */
864 if (bucket->ub_cnt == 0)
866 zdom = zone_domain_lock(zone, domain);
869 * Conditionally set the maximum number of items.
871 zdom->uzd_nitems += bucket->ub_cnt;
872 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
874 zone_domain_imax_set(zdom, zdom->uzd_nitems);
877 * Shift the bounds of the current WSS interval to
878 * avoid perturbing the estimates.
880 atomic_add_long(&zdom->uzd_imax, bucket->ub_cnt);
881 zdom->uzd_imin += bucket->ub_cnt;
882 zdom->uzd_bimin += bucket->ub_cnt;
883 zdom->uzd_limin += bucket->ub_cnt;
885 if (STAILQ_EMPTY(&zdom->uzd_buckets))
886 zdom->uzd_seq = bucket->ub_seq;
889 * Try to promote reuse of recently used items. For items
890 * protected by SMR, try to defer reuse to minimize polling.
892 if (bucket->ub_seq == SMR_SEQ_INVALID)
893 STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
895 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
899 zdom->uzd_nitems -= bucket->ub_cnt;
902 bucket_free(zone, bucket, udata);
905 /* Pops an item out of a per-cpu cache bucket. */
907 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
911 CRITICAL_ASSERT(curthread);
914 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
916 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
917 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
924 /* Pushes an item into a per-cpu cache bucket. */
926 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
929 CRITICAL_ASSERT(curthread);
930 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
931 ("uma_zfree: Freeing to non free bucket index."));
933 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
939 * Unload a UMA bucket from a per-cpu cache.
941 static inline uma_bucket_t
942 cache_bucket_unload(uma_cache_bucket_t bucket)
946 b = bucket->ucb_bucket;
948 MPASS(b->ub_entries == bucket->ucb_entries);
949 b->ub_cnt = bucket->ucb_cnt;
950 bucket->ucb_bucket = NULL;
951 bucket->ucb_entries = bucket->ucb_cnt = 0;
957 static inline uma_bucket_t
958 cache_bucket_unload_alloc(uma_cache_t cache)
961 return (cache_bucket_unload(&cache->uc_allocbucket));
964 static inline uma_bucket_t
965 cache_bucket_unload_free(uma_cache_t cache)
968 return (cache_bucket_unload(&cache->uc_freebucket));
971 static inline uma_bucket_t
972 cache_bucket_unload_cross(uma_cache_t cache)
975 return (cache_bucket_unload(&cache->uc_crossbucket));
979 * Load a bucket into a per-cpu cache bucket.
982 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
985 CRITICAL_ASSERT(curthread);
986 MPASS(bucket->ucb_bucket == NULL);
987 MPASS(b->ub_seq == SMR_SEQ_INVALID);
989 bucket->ucb_bucket = b;
990 bucket->ucb_cnt = b->ub_cnt;
991 bucket->ucb_entries = b->ub_entries;
995 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
998 cache_bucket_load(&cache->uc_allocbucket, b);
1002 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
1005 cache_bucket_load(&cache->uc_freebucket, b);
1010 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
1013 cache_bucket_load(&cache->uc_crossbucket, b);
1018 * Copy and preserve ucb_spare.
1021 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
1024 b1->ucb_bucket = b2->ucb_bucket;
1025 b1->ucb_entries = b2->ucb_entries;
1026 b1->ucb_cnt = b2->ucb_cnt;
1030 * Swap two cache buckets.
1033 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
1035 struct uma_cache_bucket b3;
1037 CRITICAL_ASSERT(curthread);
1039 cache_bucket_copy(&b3, b1);
1040 cache_bucket_copy(b1, b2);
1041 cache_bucket_copy(b2, &b3);
1045 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
1048 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
1050 uma_zone_domain_t zdom;
1051 uma_bucket_t bucket;
1055 * Avoid the lock if possible.
1057 zdom = ZDOM_GET(zone, domain);
1058 if (zdom->uzd_nitems == 0)
1061 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
1062 (seq = atomic_load_32(&zdom->uzd_seq)) != SMR_SEQ_INVALID &&
1063 !smr_poll(zone->uz_smr, seq, false))
1067 * Check the zone's cache of buckets.
1069 zdom = zone_domain_lock(zone, domain);
1070 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL)
1078 zone_log_warning(uma_zone_t zone)
1080 static const struct timeval warninterval = { 300, 0 };
1082 if (!zone_warnings || zone->uz_warning == NULL)
1085 if (ratecheck(&zone->uz_ratecheck, &warninterval))
1086 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
1090 zone_maxaction(uma_zone_t zone)
1093 if (zone->uz_maxaction.ta_func != NULL)
1094 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
1098 * Routine called by timeout which is used to fire off some time interval
1099 * based calculations. (stats, hash size, etc.)
1108 uma_timeout(void *context __unused, int pending __unused)
1111 zone_foreach(zone_timeout, NULL);
1113 /* Reschedule this event */
1114 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task,
1119 * Update the working set size estimates for the zone's bucket cache.
1120 * The constants chosen here are somewhat arbitrary.
1123 zone_domain_update_wss(uma_zone_domain_t zdom)
1127 ZDOM_LOCK_ASSERT(zdom);
1128 MPASS(zdom->uzd_imax >= zdom->uzd_nitems);
1129 MPASS(zdom->uzd_nitems >= zdom->uzd_bimin);
1130 MPASS(zdom->uzd_bimin >= zdom->uzd_imin);
1133 * Estimate WSS as modified moving average of biggest allocation
1134 * batches for each period over few minutes (UMA_TIMEOUT of 20s).
1136 zdom->uzd_wss = lmax(zdom->uzd_wss * 3 / 4,
1137 zdom->uzd_imax - zdom->uzd_bimin);
1140 * Estimate longtime minimum item count as a combination of recent
1141 * minimum item count, adjusted by WSS for safety, and the modified
1142 * moving average over the last several hours (UMA_TIMEOUT of 20s).
1143 * timin measures time since limin tried to go negative, that means
1144 * we were dangerously close to or got out of cache.
1146 m = zdom->uzd_imin - zdom->uzd_wss;
1148 if (zdom->uzd_limin >= m)
1149 zdom->uzd_limin = m;
1151 zdom->uzd_limin = (m + zdom->uzd_limin * 255) / 256;
1154 zdom->uzd_limin = 0;
1155 zdom->uzd_timin = 0;
1158 /* To reduce period edge effects on WSS keep half of the imax. */
1159 atomic_subtract_long(&zdom->uzd_imax,
1160 (zdom->uzd_imax - zdom->uzd_nitems + 1) / 2);
1161 zdom->uzd_imin = zdom->uzd_bimin = zdom->uzd_nitems;
1165 * Routine to perform timeout driven calculations. This expands the
1166 * hashes and does per cpu statistics aggregation.
1171 zone_timeout(uma_zone_t zone, void *unused)
1176 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
1182 * Hash zones are non-numa by definition so the first domain
1183 * is the only one present.
1186 pages = keg->uk_domain[0].ud_pages;
1189 * Expand the keg hash table.
1191 * This is done if the number of slabs is larger than the hash size.
1192 * What I'm trying to do here is completely reduce collisions. This
1193 * may be a little aggressive. Should I allow for two collisions max?
1195 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1196 struct uma_hash newhash;
1197 struct uma_hash oldhash;
1201 * This is so involved because allocating and freeing
1202 * while the keg lock is held will lead to deadlock.
1203 * I have to do everything in stages and check for
1207 ret = hash_alloc(&newhash, 1 << fls(slabs));
1210 if (hash_expand(&keg->uk_hash, &newhash)) {
1211 oldhash = keg->uk_hash;
1212 keg->uk_hash = newhash;
1217 hash_free(&oldhash);
1224 /* Trim caches not used for a long time. */
1225 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0) {
1226 for (int i = 0; i < vm_ndomains; i++) {
1227 if (bucket_cache_reclaim_domain(zone, false, false, i) &&
1228 (zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1229 keg_drain(zone->uz_keg, i);
1235 * Allocate and zero fill the next sized hash table from the appropriate
1239 * hash A new hash structure with the old hash size in uh_hashsize
1242 * 1 on success and 0 on failure.
1245 hash_alloc(struct uma_hash *hash, u_int size)
1249 KASSERT(powerof2(size), ("hash size must be power of 2"));
1250 if (size > UMA_HASH_SIZE_INIT) {
1251 hash->uh_hashsize = size;
1252 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1253 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1255 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1256 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1257 UMA_ANYDOMAIN, M_WAITOK);
1258 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1260 if (hash->uh_slab_hash) {
1261 bzero(hash->uh_slab_hash, alloc);
1262 hash->uh_hashmask = hash->uh_hashsize - 1;
1270 * Expands the hash table for HASH zones. This is done from zone_timeout
1271 * to reduce collisions. This must not be done in the regular allocation
1272 * path, otherwise, we can recurse on the vm while allocating pages.
1275 * oldhash The hash you want to expand
1276 * newhash The hash structure for the new table
1284 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1286 uma_hash_slab_t slab;
1290 if (!newhash->uh_slab_hash)
1293 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1297 * I need to investigate hash algorithms for resizing without a
1301 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1302 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1303 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1304 LIST_REMOVE(slab, uhs_hlink);
1305 hval = UMA_HASH(newhash, slab->uhs_data);
1306 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1314 * Free the hash bucket to the appropriate backing store.
1317 * slab_hash The hash bucket we're freeing
1318 * hashsize The number of entries in that hash bucket
1324 hash_free(struct uma_hash *hash)
1326 if (hash->uh_slab_hash == NULL)
1328 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1329 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1331 free(hash->uh_slab_hash, M_UMAHASH);
1335 * Frees all outstanding items in a bucket
1338 * zone The zone to free to, must be unlocked.
1339 * bucket The free/alloc bucket with items.
1345 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1349 if (bucket->ub_cnt == 0)
1352 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1353 bucket->ub_seq != SMR_SEQ_INVALID) {
1354 smr_wait(zone->uz_smr, bucket->ub_seq);
1355 bucket->ub_seq = SMR_SEQ_INVALID;
1356 for (i = 0; i < bucket->ub_cnt; i++)
1357 item_dtor(zone, bucket->ub_bucket[i],
1358 zone->uz_size, NULL, SKIP_NONE);
1361 for (i = 0; i < bucket->ub_cnt; i++) {
1362 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
1363 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1364 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
1366 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1367 if (zone->uz_max_items > 0)
1368 zone_free_limit(zone, bucket->ub_cnt);
1370 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1376 * Drains the per cpu caches for a zone.
1378 * NOTE: This may only be called while the zone is being torn down, and not
1379 * during normal operation. This is necessary in order that we do not have
1380 * to migrate CPUs to drain the per-CPU caches.
1383 * zone The zone to drain, must be unlocked.
1389 cache_drain(uma_zone_t zone)
1392 uma_bucket_t bucket;
1397 * XXX: It is safe to not lock the per-CPU caches, because we're
1398 * tearing down the zone anyway. I.e., there will be no further use
1399 * of the caches at this point.
1401 * XXX: It would good to be able to assert that the zone is being
1402 * torn down to prevent improper use of cache_drain().
1404 seq = SMR_SEQ_INVALID;
1405 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1406 seq = smr_advance(zone->uz_smr);
1408 cache = &zone->uz_cpu[cpu];
1409 bucket = cache_bucket_unload_alloc(cache);
1411 bucket_free(zone, bucket, NULL);
1412 bucket = cache_bucket_unload_free(cache);
1413 if (bucket != NULL) {
1414 bucket->ub_seq = seq;
1415 bucket_free(zone, bucket, NULL);
1417 bucket = cache_bucket_unload_cross(cache);
1418 if (bucket != NULL) {
1419 bucket->ub_seq = seq;
1420 bucket_free(zone, bucket, NULL);
1423 bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN);
1427 cache_shrink(uma_zone_t zone, void *unused)
1430 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1434 zone->uz_bucket_size =
1435 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1440 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1443 uma_bucket_t b1, b2, b3;
1446 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1449 b1 = b2 = b3 = NULL;
1451 cache = &zone->uz_cpu[curcpu];
1452 domain = PCPU_GET(domain);
1453 b1 = cache_bucket_unload_alloc(cache);
1456 * Don't flush SMR zone buckets. This leaves the zone without a
1457 * bucket and forces every free to synchronize().
1459 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1460 b2 = cache_bucket_unload_free(cache);
1461 b3 = cache_bucket_unload_cross(cache);
1466 zone_free_bucket(zone, b1, NULL, domain, false);
1468 zone_free_bucket(zone, b2, NULL, domain, false);
1470 /* Adjust the domain so it goes to zone_free_cross. */
1471 domain = (domain + 1) % vm_ndomains;
1472 zone_free_bucket(zone, b3, NULL, domain, false);
1477 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1478 * This is an expensive call because it needs to bind to all CPUs
1479 * one by one and enter a critical section on each of them in order
1480 * to safely access their cache buckets.
1481 * Zone lock must not be held on call this function.
1484 pcpu_cache_drain_safe(uma_zone_t zone)
1489 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1492 cache_shrink(zone, NULL);
1494 zone_foreach(cache_shrink, NULL);
1497 thread_lock(curthread);
1498 sched_bind(curthread, cpu);
1499 thread_unlock(curthread);
1502 cache_drain_safe_cpu(zone, NULL);
1504 zone_foreach(cache_drain_safe_cpu, NULL);
1506 thread_lock(curthread);
1507 sched_unbind(curthread);
1508 thread_unlock(curthread);
1512 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1513 * requested a drain, otherwise the per-domain caches are trimmed to either
1514 * estimated working set size.
1517 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain)
1519 uma_zone_domain_t zdom;
1520 uma_bucket_t bucket;
1525 * The cross bucket is partially filled and not part of
1526 * the item count. Reclaim it individually here.
1528 zdom = ZDOM_GET(zone, domain);
1529 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1530 ZONE_CROSS_LOCK(zone);
1531 bucket = zdom->uzd_cross;
1532 zdom->uzd_cross = NULL;
1533 ZONE_CROSS_UNLOCK(zone);
1535 bucket_free(zone, bucket, NULL);
1539 * If we were asked to drain the zone, we are done only once
1540 * this bucket cache is empty. If trim, we reclaim items in
1541 * excess of the zone's estimated working set size. Multiple
1542 * consecutive calls will shrink the WSS and so reclaim more.
1543 * If neither drain nor trim, then voluntarily reclaim 1/4
1544 * (to reduce first spike) of items not used for a long time.
1547 zone_domain_update_wss(zdom);
1551 target = zdom->uzd_wss;
1552 else if (zdom->uzd_timin > 900 / UMA_TIMEOUT)
1553 target = zdom->uzd_nitems - zdom->uzd_limin / 4;
1558 while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL &&
1559 zdom->uzd_nitems >= target + bucket->ub_cnt) {
1560 bucket = zone_fetch_bucket(zone, zdom, true);
1563 bucket_free(zone, bucket, NULL);
1572 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain)
1577 * Shrink the zone bucket size to ensure that the per-CPU caches
1578 * don't grow too large.
1580 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1581 zone->uz_bucket_size--;
1583 if (domain != UMA_ANYDOMAIN &&
1584 (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) {
1585 bucket_cache_reclaim_domain(zone, drain, true, domain);
1587 for (i = 0; i < vm_ndomains; i++)
1588 bucket_cache_reclaim_domain(zone, drain, true, i);
1593 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1600 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1601 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1603 mem = slab_data(slab, keg);
1604 size = PAGE_SIZE * keg->uk_ppera;
1606 kasan_mark_slab_valid(keg, mem);
1607 if (keg->uk_fini != NULL) {
1608 for (i = start - 1; i > -1; i--)
1611 * trash_fini implies that dtor was trash_dtor. trash_fini
1612 * would check that memory hasn't been modified since free,
1613 * which executed trash_dtor.
1614 * That's why we need to run uma_dbg_kskip() check here,
1615 * albeit we don't make skip check for other init/fini
1618 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1619 keg->uk_fini != trash_fini)
1621 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1623 flags = slab->us_flags;
1624 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1625 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1628 keg->uk_freef(mem, size, flags);
1629 uma_total_dec(size);
1633 keg_drain_domain(uma_keg_t keg, int domain)
1635 struct slabhead freeslabs;
1637 uma_slab_t slab, tmp;
1638 uint32_t i, stofree, stokeep, partial;
1640 dom = &keg->uk_domain[domain];
1641 LIST_INIT(&freeslabs);
1643 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1644 keg->uk_name, keg, domain, dom->ud_free_items);
1646 KEG_LOCK(keg, domain);
1649 * Are the free items in partially allocated slabs sufficient to meet
1650 * the reserve? If not, compute the number of fully free slabs that must
1653 partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1654 if (partial < keg->uk_reserve) {
1655 stokeep = min(dom->ud_free_slabs,
1656 howmany(keg->uk_reserve - partial, keg->uk_ipers));
1660 stofree = dom->ud_free_slabs - stokeep;
1663 * Partition the free slabs into two sets: those that must be kept in
1664 * order to maintain the reserve, and those that may be released back to
1665 * the system. Since one set may be much larger than the other,
1666 * populate the smaller of the two sets and swap them if necessary.
1668 for (i = min(stofree, stokeep); i > 0; i--) {
1669 slab = LIST_FIRST(&dom->ud_free_slab);
1670 LIST_REMOVE(slab, us_link);
1671 LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1673 if (stofree > stokeep)
1674 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1676 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1677 LIST_FOREACH(slab, &freeslabs, us_link)
1678 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1680 dom->ud_free_items -= stofree * keg->uk_ipers;
1681 dom->ud_free_slabs -= stofree;
1682 dom->ud_pages -= stofree * keg->uk_ppera;
1683 KEG_UNLOCK(keg, domain);
1685 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1686 keg_free_slab(keg, slab, keg->uk_ipers);
1690 * Frees pages from a keg back to the system. This is done on demand from
1691 * the pageout daemon.
1696 keg_drain(uma_keg_t keg, int domain)
1700 if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1702 if (domain != UMA_ANYDOMAIN) {
1703 keg_drain_domain(keg, domain);
1705 for (i = 0; i < vm_ndomains; i++)
1706 keg_drain_domain(keg, i);
1711 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain)
1714 * Count active reclaim operations in order to interlock with
1715 * zone_dtor(), which removes the zone from global lists before
1716 * attempting to reclaim items itself.
1718 * The zone may be destroyed while sleeping, so only zone_dtor() should
1722 if (waitok == M_WAITOK) {
1723 while (zone->uz_reclaimers > 0)
1724 msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1);
1726 zone->uz_reclaimers++;
1728 bucket_cache_reclaim(zone, drain, domain);
1730 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1731 keg_drain(zone->uz_keg, domain);
1733 zone->uz_reclaimers--;
1734 if (zone->uz_reclaimers == 0)
1740 * Allocate a new slab for a keg and inserts it into the partial slab list.
1741 * The keg should be unlocked on entry. If the allocation succeeds it will
1742 * be locked on return.
1745 * flags Wait flags for the item initialization routine
1746 * aflags Wait flags for the slab allocation
1749 * The slab that was allocated or NULL if there is no memory and the
1750 * caller specified M_NOWAIT.
1753 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;
1869 * This function is intended to be used early on in place of page_alloc(). It
1870 * performs contiguous physical memory allocations and uses a bump allocator for
1871 * KVA, so is usable before the kernel map is initialized.
1874 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1881 pages = howmany(bytes, PAGE_SIZE);
1882 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1884 *pflag = UMA_SLAB_BOOT;
1885 m = vm_page_alloc_noobj_contig_domain(domain, malloc2vm_flags(wait) |
1886 VM_ALLOC_WIRED, pages, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
1887 VM_MEMATTR_DEFAULT);
1891 pa = VM_PAGE_TO_PHYS(m);
1892 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1893 #if defined(__aarch64__) || defined(__amd64__) || \
1894 defined(__riscv) || defined(__powerpc64__)
1895 if ((wait & M_NODUMP) == 0)
1900 /* Allocate KVA and indirectly advance bootmem. */
1901 return ((void *)pmap_map(&bootmem, m->phys_addr,
1902 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE));
1906 startup_free(void *mem, vm_size_t bytes)
1911 va = (vm_offset_t)mem;
1912 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1915 * startup_alloc() returns direct-mapped slabs on some platforms. Avoid
1916 * unmapping ranges of the direct map.
1918 if (va >= bootstart && va + bytes <= bootmem)
1919 pmap_remove(kernel_pmap, va, va + bytes);
1920 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1921 #if defined(__aarch64__) || defined(__amd64__) || \
1922 defined(__riscv) || defined(__powerpc64__)
1923 dump_drop_page(VM_PAGE_TO_PHYS(m));
1925 vm_page_unwire_noq(m);
1931 * Allocates a number of pages from the system
1934 * bytes The number of bytes requested
1935 * wait Shall we wait?
1938 * A pointer to the alloced memory or possibly
1939 * NULL if M_NOWAIT is set.
1942 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1945 void *p; /* Returned page */
1947 *pflag = UMA_SLAB_KERNEL;
1948 p = kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1954 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1957 struct pglist alloctail;
1958 vm_offset_t addr, zkva;
1960 vm_page_t p, p_next;
1965 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1967 TAILQ_INIT(&alloctail);
1968 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | malloc2vm_flags(wait);
1969 *pflag = UMA_SLAB_KERNEL;
1970 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1971 if (CPU_ABSENT(cpu)) {
1972 p = vm_page_alloc_noobj(flags);
1975 p = vm_page_alloc_noobj(flags);
1977 pc = pcpu_find(cpu);
1978 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1981 p = vm_page_alloc_noobj_domain(pc->pc_domain,
1983 if (__predict_false(p == NULL))
1984 p = vm_page_alloc_noobj(flags);
1987 if (__predict_false(p == NULL))
1989 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1991 if ((addr = kva_alloc(bytes)) == 0)
1994 TAILQ_FOREACH(p, &alloctail, listq) {
1995 pmap_qenter(zkva, &p, 1);
1998 return ((void*)addr);
2000 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2001 vm_page_unwire_noq(p);
2008 * Allocates a number of pages not belonging to a VM object
2011 * bytes The number of bytes requested
2012 * wait Shall we wait?
2015 * A pointer to the alloced memory or possibly
2016 * NULL if M_NOWAIT is set.
2019 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
2022 TAILQ_HEAD(, vm_page) alloctail;
2024 vm_offset_t retkva, zkva;
2025 vm_page_t p, p_next;
2029 TAILQ_INIT(&alloctail);
2031 req = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
2032 if ((wait & M_WAITOK) != 0)
2033 req |= VM_ALLOC_WAITOK;
2035 npages = howmany(bytes, PAGE_SIZE);
2036 while (npages > 0) {
2037 p = vm_page_alloc_noobj_domain(domain, req);
2040 * Since the page does not belong to an object, its
2043 TAILQ_INSERT_TAIL(&alloctail, p, listq);
2048 * Page allocation failed, free intermediate pages and
2051 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2052 vm_page_unwire_noq(p);
2057 *flags = UMA_SLAB_PRIV;
2058 zkva = keg->uk_kva +
2059 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
2061 TAILQ_FOREACH(p, &alloctail, listq) {
2062 pmap_qenter(zkva, &p, 1);
2066 return ((void *)retkva);
2070 * Allocate physically contiguous pages.
2073 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
2077 *pflag = UMA_SLAB_KERNEL;
2078 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
2079 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
2083 * Frees a number of pages to the system
2086 * mem A pointer to the memory to be freed
2087 * size The size of the memory being freed
2088 * flags The original p->us_flags field
2094 page_free(void *mem, vm_size_t size, uint8_t flags)
2097 if ((flags & UMA_SLAB_BOOT) != 0) {
2098 startup_free(mem, size);
2102 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
2103 ("UMA: page_free used with invalid flags %x", flags));
2105 kmem_free(mem, size);
2109 * Frees pcpu zone allocations
2112 * mem A pointer to the memory to be freed
2113 * size The size of the memory being freed
2114 * flags The original p->us_flags field
2120 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
2122 vm_offset_t sva, curva;
2126 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
2128 if ((flags & UMA_SLAB_BOOT) != 0) {
2129 startup_free(mem, size);
2133 sva = (vm_offset_t)mem;
2134 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
2135 paddr = pmap_kextract(curva);
2136 m = PHYS_TO_VM_PAGE(paddr);
2137 vm_page_unwire_noq(m);
2140 pmap_qremove(sva, size >> PAGE_SHIFT);
2141 kva_free(sva, size);
2145 * Zero fill initializer
2147 * Arguments/Returns follow uma_init specifications
2150 zero_init(void *mem, int size, int flags)
2157 static struct noslabbits *
2158 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2161 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2166 * Actual size of embedded struct slab (!OFFPAGE).
2169 slab_sizeof(int nitems)
2173 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2174 return (roundup(s, UMA_ALIGN_PTR + 1));
2177 #define UMA_FIXPT_SHIFT 31
2178 #define UMA_FRAC_FIXPT(n, d) \
2179 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2180 #define UMA_FIXPT_PCT(f) \
2181 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2182 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
2183 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2186 * Compute the number of items that will fit in a slab. If hdr is true, the
2187 * item count may be limited to provide space in the slab for an inline slab
2188 * header. Otherwise, all slab space will be provided for item storage.
2191 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2196 /* The padding between items is not needed after the last item. */
2197 padpi = rsize - size;
2201 * Start with the maximum item count and remove items until
2202 * the slab header first alongside the allocatable memory.
2204 for (ipers = MIN(SLAB_MAX_SETSIZE,
2205 (slabsize + padpi - slab_sizeof(1)) / rsize);
2207 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2211 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2217 struct keg_layout_result {
2225 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2226 struct keg_layout_result *kl)
2231 kl->slabsize = slabsize;
2233 /* Handle INTERNAL as inline with an extra page. */
2234 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2235 kl->format &= ~UMA_ZFLAG_INTERNAL;
2236 kl->slabsize += PAGE_SIZE;
2239 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2240 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2242 /* Account for memory used by an offpage slab header. */
2243 total = kl->slabsize;
2244 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2245 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2247 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2251 * Determine the format of a uma keg. This determines where the slab header
2252 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2255 * keg The zone we should initialize
2261 keg_layout(uma_keg_t keg)
2263 struct keg_layout_result kl = {}, kl_tmp;
2272 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2273 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2274 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2275 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2276 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2278 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2279 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2280 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2283 alignsize = keg->uk_align + 1;
2286 * ASAN requires that each allocation be aligned to the shadow map
2289 if (alignsize < KASAN_SHADOW_SCALE)
2290 alignsize = KASAN_SHADOW_SCALE;
2294 * Calculate the size of each allocation (rsize) according to
2295 * alignment. If the requested size is smaller than we have
2296 * allocation bits for we round it up.
2298 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2299 rsize = roundup2(rsize, alignsize);
2301 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2303 * We want one item to start on every align boundary in a page.
2304 * To do this we will span pages. We will also extend the item
2305 * by the size of align if it is an even multiple of align.
2306 * Otherwise, it would fall on the same boundary every time.
2308 if ((rsize & alignsize) == 0)
2310 slabsize = rsize * (PAGE_SIZE / alignsize);
2311 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2312 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2313 slabsize = round_page(slabsize);
2316 * Start with a slab size of as many pages as it takes to
2317 * represent a single item. We will try to fit as many
2318 * additional items into the slab as possible.
2320 slabsize = round_page(keg->uk_size);
2323 /* Build a list of all of the available formats for this keg. */
2326 /* Evaluate an inline slab layout. */
2327 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2330 /* TODO: vm_page-embedded slab. */
2333 * We can't do OFFPAGE if we're internal or if we've been
2334 * asked to not go to the VM for buckets. If we do this we
2335 * may end up going to the VM for slabs which we do not want
2336 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2337 * In those cases, evaluate a pseudo-format called INTERNAL
2338 * which has an inline slab header and one extra page to
2339 * guarantee that it fits.
2341 * Otherwise, see if using an OFFPAGE slab will improve our
2344 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2345 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2347 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2350 * Choose a slab size and format which satisfy the minimum efficiency.
2351 * Prefer the smallest slab size that meets the constraints.
2353 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2354 * for small items (up to PAGE_SIZE), the iteration increment is one
2355 * page; and for large items, the increment is one item.
2357 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2358 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2359 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2362 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2363 round_page(rsize * (i - 1) + keg->uk_size);
2365 for (j = 0; j < nfmt; j++) {
2366 /* Only if we have no viable format yet. */
2367 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2371 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2372 if (kl_tmp.eff <= kl.eff)
2377 CTR6(KTR_UMA, "keg %s layout: format %#x "
2378 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2379 keg->uk_name, kl.format, kl.ipers, rsize,
2380 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2382 /* Stop when we reach the minimum efficiency. */
2383 if (kl.eff >= UMA_MIN_EFF)
2387 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2388 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2389 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2393 pages = atop(kl.slabsize);
2394 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2395 pages *= mp_maxid + 1;
2397 keg->uk_rsize = rsize;
2398 keg->uk_ipers = kl.ipers;
2399 keg->uk_ppera = pages;
2400 keg->uk_flags |= kl.format;
2403 * How do we find the slab header if it is offpage or if not all item
2404 * start addresses are in the same page? We could solve the latter
2405 * case with vaddr alignment, but we don't.
2407 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2408 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2409 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2410 keg->uk_flags |= UMA_ZFLAG_HASH;
2412 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2415 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2416 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2418 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2419 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2420 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2421 keg->uk_ipers, pages));
2425 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2426 * the keg onto the global keg list.
2428 * Arguments/Returns follow uma_ctor specifications
2429 * udata Actually uma_kctor_args
2432 keg_ctor(void *mem, int size, void *udata, int flags)
2434 struct uma_kctor_args *arg = udata;
2435 uma_keg_t keg = mem;
2440 keg->uk_size = arg->size;
2441 keg->uk_init = arg->uminit;
2442 keg->uk_fini = arg->fini;
2443 keg->uk_align = arg->align;
2444 keg->uk_reserve = 0;
2445 keg->uk_flags = arg->flags;
2448 * We use a global round-robin policy by default. Zones with
2449 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2450 * case the iterator is never run.
2452 keg->uk_dr.dr_policy = DOMAINSET_RR();
2453 keg->uk_dr.dr_iter = 0;
2456 * The primary zone is passed to us at keg-creation time.
2459 keg->uk_name = zone->uz_name;
2461 if (arg->flags & UMA_ZONE_ZINIT)
2462 keg->uk_init = zero_init;
2464 if (arg->flags & UMA_ZONE_MALLOC)
2465 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2468 keg->uk_flags &= ~UMA_ZONE_PCPU;
2474 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2475 * work on. Use round-robin for everything else.
2477 * Zones may override the default by specifying either.
2480 if ((keg->uk_flags &
2481 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2482 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2483 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2484 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2488 * If we haven't booted yet we need allocations to go through the
2489 * startup cache until the vm is ready.
2491 #ifdef UMA_MD_SMALL_ALLOC
2492 if (keg->uk_ppera == 1)
2493 keg->uk_allocf = uma_small_alloc;
2496 if (booted < BOOT_KVA)
2497 keg->uk_allocf = startup_alloc;
2498 else if (keg->uk_flags & UMA_ZONE_PCPU)
2499 keg->uk_allocf = pcpu_page_alloc;
2500 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2501 keg->uk_allocf = contig_alloc;
2503 keg->uk_allocf = page_alloc;
2504 #ifdef UMA_MD_SMALL_ALLOC
2505 if (keg->uk_ppera == 1)
2506 keg->uk_freef = uma_small_free;
2509 if (keg->uk_flags & UMA_ZONE_PCPU)
2510 keg->uk_freef = pcpu_page_free;
2512 keg->uk_freef = page_free;
2515 * Initialize keg's locks.
2517 for (i = 0; i < vm_ndomains; i++)
2518 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2521 * If we're putting the slab header in the actual page we need to
2522 * figure out where in each page it goes. See slab_sizeof
2525 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2528 shsize = slab_sizeof(keg->uk_ipers);
2529 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2531 * The only way the following is possible is if with our
2532 * UMA_ALIGN_PTR adjustments we are now bigger than
2533 * UMA_SLAB_SIZE. I haven't checked whether this is
2534 * mathematically possible for all cases, so we make
2537 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2538 ("zone %s ipers %d rsize %d size %d slab won't fit",
2539 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2542 if (keg->uk_flags & UMA_ZFLAG_HASH)
2543 hash_alloc(&keg->uk_hash, 0);
2545 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2547 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2549 rw_wlock(&uma_rwlock);
2550 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2551 rw_wunlock(&uma_rwlock);
2556 zone_kva_available(uma_zone_t zone, void *unused)
2560 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2564 if (keg->uk_allocf == startup_alloc) {
2565 /* Switch to the real allocator. */
2566 if (keg->uk_flags & UMA_ZONE_PCPU)
2567 keg->uk_allocf = pcpu_page_alloc;
2568 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2570 keg->uk_allocf = contig_alloc;
2572 keg->uk_allocf = page_alloc;
2577 zone_alloc_counters(uma_zone_t zone, void *unused)
2580 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2581 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2582 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2583 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2587 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2589 uma_zone_domain_t zdom;
2592 struct sysctl_oid *oid, *domainoid;
2593 int domains, i, cnt;
2594 static const char *nokeg = "cache zone";
2598 * Make a sysctl safe copy of the zone name by removing
2599 * any special characters and handling dups by appending
2602 if (zone->uz_namecnt != 0) {
2603 /* Count the number of decimal digits and '_' separator. */
2604 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2606 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2608 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2611 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2612 for (c = zone->uz_ctlname; *c != '\0'; c++)
2613 if (strchr("./\\ -", *c) != NULL)
2617 * Basic parameters at the root.
2619 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2620 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2622 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2623 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2624 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2625 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2626 zone, 0, sysctl_handle_uma_zone_flags, "A",
2627 "Allocator configuration flags");
2628 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2629 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2630 "Desired per-cpu cache size");
2631 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2632 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2633 "Maximum allowed per-cpu cache size");
2638 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2639 domains = vm_ndomains;
2642 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2643 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2645 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2646 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2647 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2648 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2649 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2650 "Real object size with alignment");
2651 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2652 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2653 "pages per-slab allocation");
2654 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2655 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2656 "items available per-slab");
2657 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2658 "align", CTLFLAG_RD, &keg->uk_align, 0,
2659 "item alignment mask");
2660 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2661 "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2662 "number of reserved items");
2663 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2664 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2665 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2666 "Slab utilization (100 - internal fragmentation %)");
2667 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2668 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2669 for (i = 0; i < domains; i++) {
2670 dom = &keg->uk_domain[i];
2671 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2672 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2673 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2674 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2675 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2676 "Total pages currently allocated from VM");
2677 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2678 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2679 "Items free in the slab layer");
2680 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2681 "free_slabs", CTLFLAG_RD, &dom->ud_free_slabs, 0,
2685 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2686 "name", CTLFLAG_RD, nokeg, "Keg name");
2689 * Information about zone limits.
2691 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2692 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2693 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2694 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2695 zone, 0, sysctl_handle_uma_zone_items, "QU",
2696 "Current number of allocated items if limit is set");
2697 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2698 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2699 "Maximum number of allocated and cached items");
2700 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2701 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2702 "Number of threads sleeping at limit");
2703 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2704 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2705 "Total zone limit sleeps");
2706 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2707 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2708 "Maximum number of items in each domain's bucket cache");
2711 * Per-domain zone information.
2713 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2714 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2715 for (i = 0; i < domains; i++) {
2716 zdom = ZDOM_GET(zone, i);
2717 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2718 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2719 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2720 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2721 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2722 "number of items in this domain");
2723 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2724 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2725 "maximum item count in this period");
2726 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2727 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2728 "minimum item count in this period");
2729 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2730 "bimin", CTLFLAG_RD, &zdom->uzd_bimin,
2731 "Minimum item count in this batch");
2732 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2733 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2734 "Working set size");
2735 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2736 "limin", CTLFLAG_RD, &zdom->uzd_limin,
2737 "Long time minimum item count");
2738 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2739 "timin", CTLFLAG_RD, &zdom->uzd_timin, 0,
2740 "Time since zero long time minimum item count");
2744 * General statistics.
2746 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2747 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2748 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2749 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2750 zone, 1, sysctl_handle_uma_zone_cur, "I",
2751 "Current number of allocated items");
2752 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2753 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2754 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2755 "Total allocation calls");
2756 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2757 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2758 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2759 "Total free calls");
2760 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2761 "fails", CTLFLAG_RD, &zone->uz_fails,
2762 "Number of allocation failures");
2763 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2764 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2765 "Free calls from the wrong domain");
2768 struct uma_zone_count {
2774 zone_count(uma_zone_t zone, void *arg)
2776 struct uma_zone_count *cnt;
2780 * Some zones are rapidly created with identical names and
2781 * destroyed out of order. This can lead to gaps in the count.
2782 * Use one greater than the maximum observed for this name.
2784 if (strcmp(zone->uz_name, cnt->name) == 0)
2785 cnt->count = MAX(cnt->count,
2786 zone->uz_namecnt + 1);
2790 zone_update_caches(uma_zone_t zone)
2794 for (i = 0; i <= mp_maxid; i++) {
2795 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2796 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2801 * Zone header ctor. This initializes all fields, locks, etc.
2803 * Arguments/Returns follow uma_ctor specifications
2804 * udata Actually uma_zctor_args
2807 zone_ctor(void *mem, int size, void *udata, int flags)
2809 struct uma_zone_count cnt;
2810 struct uma_zctor_args *arg = udata;
2811 uma_zone_domain_t zdom;
2812 uma_zone_t zone = mem;
2818 zone->uz_name = arg->name;
2819 zone->uz_ctor = arg->ctor;
2820 zone->uz_dtor = arg->dtor;
2821 zone->uz_init = NULL;
2822 zone->uz_fini = NULL;
2823 zone->uz_sleeps = 0;
2824 zone->uz_bucket_size = 0;
2825 zone->uz_bucket_size_min = 0;
2826 zone->uz_bucket_size_max = BUCKET_MAX;
2827 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2828 zone->uz_warning = NULL;
2829 /* The domain structures follow the cpu structures. */
2830 zone->uz_bucket_max = ULONG_MAX;
2831 timevalclear(&zone->uz_ratecheck);
2833 /* Count the number of duplicate names. */
2834 cnt.name = arg->name;
2836 zone_foreach(zone_count, &cnt);
2837 zone->uz_namecnt = cnt.count;
2838 ZONE_CROSS_LOCK_INIT(zone);
2840 for (i = 0; i < vm_ndomains; i++) {
2841 zdom = ZDOM_GET(zone, i);
2842 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2843 STAILQ_INIT(&zdom->uzd_buckets);
2846 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
2847 if (arg->uminit == trash_init && arg->fini == trash_fini)
2848 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2849 #elif defined(KASAN)
2850 if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0)
2851 arg->flags |= UMA_ZONE_NOKASAN;
2855 * This is a pure cache zone, no kegs.
2858 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2859 ("zone_ctor: Import specified for non-cache zone."));
2860 zone->uz_flags = arg->flags;
2861 zone->uz_size = arg->size;
2862 zone->uz_import = arg->import;
2863 zone->uz_release = arg->release;
2864 zone->uz_arg = arg->arg;
2867 * Cache zones are round-robin unless a policy is
2868 * specified because they may have incompatible
2871 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2872 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2874 rw_wlock(&uma_rwlock);
2875 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2876 rw_wunlock(&uma_rwlock);
2881 * Use the regular zone/keg/slab allocator.
2883 zone->uz_import = zone_import;
2884 zone->uz_release = zone_release;
2885 zone->uz_arg = zone;
2888 if (arg->flags & UMA_ZONE_SECONDARY) {
2889 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2890 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2891 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2892 zone->uz_init = arg->uminit;
2893 zone->uz_fini = arg->fini;
2894 zone->uz_flags |= UMA_ZONE_SECONDARY;
2895 rw_wlock(&uma_rwlock);
2897 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2898 if (LIST_NEXT(z, uz_link) == NULL) {
2899 LIST_INSERT_AFTER(z, zone, uz_link);
2904 rw_wunlock(&uma_rwlock);
2905 } else if (keg == NULL) {
2906 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2907 arg->align, arg->flags)) == NULL)
2910 struct uma_kctor_args karg;
2913 /* We should only be here from uma_startup() */
2914 karg.size = arg->size;
2915 karg.uminit = arg->uminit;
2916 karg.fini = arg->fini;
2917 karg.align = arg->align;
2918 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2920 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2926 /* Inherit properties from the keg. */
2928 zone->uz_size = keg->uk_size;
2929 zone->uz_flags |= (keg->uk_flags &
2930 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2933 if (booted >= BOOT_PCPU) {
2934 zone_alloc_counters(zone, NULL);
2935 if (booted >= BOOT_RUNNING)
2936 zone_alloc_sysctl(zone, NULL);
2938 zone->uz_allocs = EARLY_COUNTER;
2939 zone->uz_frees = EARLY_COUNTER;
2940 zone->uz_fails = EARLY_COUNTER;
2943 /* Caller requests a private SMR context. */
2944 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2945 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2947 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2948 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2949 ("Invalid zone flag combination"));
2950 if (arg->flags & UMA_ZFLAG_INTERNAL)
2951 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2952 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2953 zone->uz_bucket_size = BUCKET_MAX;
2954 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2955 zone->uz_bucket_size = 0;
2957 zone->uz_bucket_size = bucket_select(zone->uz_size);
2958 zone->uz_bucket_size_min = zone->uz_bucket_size;
2959 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2960 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2961 zone_update_caches(zone);
2967 * Keg header dtor. This frees all data, destroys locks, frees the hash
2968 * table and removes the keg from the global list.
2970 * Arguments/Returns follow uma_dtor specifications
2974 keg_dtor(void *arg, int size, void *udata)
2977 uint32_t free, pages;
2980 keg = (uma_keg_t)arg;
2982 for (i = 0; i < vm_ndomains; i++) {
2983 free += keg->uk_domain[i].ud_free_items;
2984 pages += keg->uk_domain[i].ud_pages;
2985 KEG_LOCK_FINI(keg, i);
2988 printf("Freed UMA keg (%s) was not empty (%u items). "
2989 " Lost %u pages of memory.\n",
2990 keg->uk_name ? keg->uk_name : "",
2991 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2993 hash_free(&keg->uk_hash);
2999 * Arguments/Returns follow uma_dtor specifications
3003 zone_dtor(void *arg, int size, void *udata)
3009 zone = (uma_zone_t)arg;
3011 sysctl_remove_oid(zone->uz_oid, 1, 1);
3013 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
3016 rw_wlock(&uma_rwlock);
3017 LIST_REMOVE(zone, uz_link);
3018 rw_wunlock(&uma_rwlock);
3019 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3021 keg->uk_reserve = 0;
3023 zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true);
3026 * We only destroy kegs from non secondary/non cache zones.
3028 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3030 rw_wlock(&uma_rwlock);
3031 LIST_REMOVE(keg, uk_link);
3032 rw_wunlock(&uma_rwlock);
3033 zone_free_item(kegs, keg, NULL, SKIP_NONE);
3035 counter_u64_free(zone->uz_allocs);
3036 counter_u64_free(zone->uz_frees);
3037 counter_u64_free(zone->uz_fails);
3038 counter_u64_free(zone->uz_xdomain);
3039 free(zone->uz_ctlname, M_UMA);
3040 for (i = 0; i < vm_ndomains; i++)
3041 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
3042 ZONE_CROSS_LOCK_FINI(zone);
3046 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3051 LIST_FOREACH(keg, &uma_kegs, uk_link) {
3052 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
3055 LIST_FOREACH(zone, &uma_cachezones, uz_link)
3060 * Traverses every zone in the system and calls a callback
3063 * zfunc A pointer to a function which accepts a zone
3070 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3073 rw_rlock(&uma_rwlock);
3074 zone_foreach_unlocked(zfunc, arg);
3075 rw_runlock(&uma_rwlock);
3079 * Initialize the kernel memory allocator. This is done after pages can be
3080 * allocated but before general KVA is available.
3083 uma_startup1(vm_offset_t virtual_avail)
3085 struct uma_zctor_args args;
3086 size_t ksize, zsize, size;
3087 uma_keg_t primarykeg;
3092 bootstart = bootmem = virtual_avail;
3094 rw_init(&uma_rwlock, "UMA lock");
3095 sx_init(&uma_reclaim_lock, "umareclaim");
3097 ksize = sizeof(struct uma_keg) +
3098 (sizeof(struct uma_domain) * vm_ndomains);
3099 ksize = roundup(ksize, UMA_SUPER_ALIGN);
3100 zsize = sizeof(struct uma_zone) +
3101 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
3102 (sizeof(struct uma_zone_domain) * vm_ndomains);
3103 zsize = roundup(zsize, UMA_SUPER_ALIGN);
3105 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
3106 size = (zsize * 2) + ksize;
3107 for (domain = 0; domain < vm_ndomains; domain++) {
3108 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
3113 zones = (uma_zone_t)m;
3115 kegs = (uma_zone_t)m;
3117 primarykeg = (uma_keg_t)m;
3119 /* "manually" create the initial zone */
3120 memset(&args, 0, sizeof(args));
3121 args.name = "UMA Kegs";
3123 args.ctor = keg_ctor;
3124 args.dtor = keg_dtor;
3125 args.uminit = zero_init;
3127 args.keg = primarykeg;
3128 args.align = UMA_SUPER_ALIGN - 1;
3129 args.flags = UMA_ZFLAG_INTERNAL;
3130 zone_ctor(kegs, zsize, &args, M_WAITOK);
3132 args.name = "UMA Zones";
3134 args.ctor = zone_ctor;
3135 args.dtor = zone_dtor;
3136 args.uminit = zero_init;
3139 args.align = UMA_SUPER_ALIGN - 1;
3140 args.flags = UMA_ZFLAG_INTERNAL;
3141 zone_ctor(zones, zsize, &args, M_WAITOK);
3143 /* Now make zones for slab headers */
3144 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
3145 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3146 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
3147 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3149 hashzone = uma_zcreate("UMA Hash",
3150 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
3151 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3157 #ifndef UMA_MD_SMALL_ALLOC
3158 extern void vm_radix_reserve_kva(void);
3162 * Advertise the availability of normal kva allocations and switch to
3163 * the default back-end allocator. Marks the KVA we consumed on startup
3164 * as used in the map.
3170 if (bootstart != bootmem) {
3171 vm_map_lock(kernel_map);
3172 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
3173 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
3174 vm_map_unlock(kernel_map);
3177 #ifndef UMA_MD_SMALL_ALLOC
3178 /* Set up radix zone to use noobj_alloc. */
3179 vm_radix_reserve_kva();
3183 zone_foreach_unlocked(zone_kva_available, NULL);
3188 * Allocate counters as early as possible so that boot-time allocations are
3189 * accounted more precisely.
3192 uma_startup_pcpu(void *arg __unused)
3195 zone_foreach_unlocked(zone_alloc_counters, NULL);
3198 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
3201 * Finish our initialization steps.
3204 uma_startup3(void *arg __unused)
3208 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
3209 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
3210 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
3212 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
3213 booted = BOOT_RUNNING;
3215 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
3216 EVENTHANDLER_PRI_FIRST);
3218 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
3221 uma_startup4(void *arg __unused)
3223 TIMEOUT_TASK_INIT(taskqueue_thread, &uma_timeout_task, 0, uma_timeout,
3225 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task,
3228 SYSINIT(uma_startup4, SI_SUB_TASKQ, SI_ORDER_ANY, uma_startup4, NULL);
3234 booted = BOOT_SHUTDOWN;
3238 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3239 int align, uint32_t flags)
3241 struct uma_kctor_args args;
3244 args.uminit = uminit;
3246 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
3249 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3252 /* Public functions */
3255 uma_set_align(int align)
3258 if (align != UMA_ALIGN_CACHE)
3259 uma_align_cache = align;
3264 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3265 uma_init uminit, uma_fini fini, int align, uint32_t flags)
3268 struct uma_zctor_args args;
3271 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
3274 /* This stuff is essential for the zone ctor */
3275 memset(&args, 0, sizeof(args));
3280 args.uminit = uminit;
3282 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
3284 * Inject procedures which check for memory use after free if we are
3285 * allowed to scramble the memory while it is not allocated. This
3286 * requires that: UMA is actually able to access the memory, no init
3287 * or fini procedures, no dependency on the initial value of the
3288 * memory, and no (legitimate) use of the memory after free. Note,
3289 * the ctor and dtor do not need to be empty.
3291 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3292 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3293 args.uminit = trash_init;
3294 args.fini = trash_fini;
3301 sx_xlock(&uma_reclaim_lock);
3302 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3303 sx_xunlock(&uma_reclaim_lock);
3310 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3311 uma_init zinit, uma_fini zfini, uma_zone_t primary)
3313 struct uma_zctor_args args;
3317 keg = primary->uz_keg;
3318 memset(&args, 0, sizeof(args));
3320 args.size = keg->uk_size;
3323 args.uminit = zinit;
3325 args.align = keg->uk_align;
3326 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3329 sx_xlock(&uma_reclaim_lock);
3330 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3331 sx_xunlock(&uma_reclaim_lock);
3338 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3339 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3340 void *arg, int flags)
3342 struct uma_zctor_args args;
3344 memset(&args, 0, sizeof(args));
3349 args.uminit = zinit;
3351 args.import = zimport;
3352 args.release = zrelease;
3355 args.flags = flags | UMA_ZFLAG_CACHE;
3357 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3362 uma_zdestroy(uma_zone_t zone)
3366 * Large slabs are expensive to reclaim, so don't bother doing
3367 * unnecessary work if we're shutting down.
3369 if (booted == BOOT_SHUTDOWN &&
3370 zone->uz_fini == NULL && zone->uz_release == zone_release)
3372 sx_xlock(&uma_reclaim_lock);
3373 zone_free_item(zones, zone, NULL, SKIP_NONE);
3374 sx_xunlock(&uma_reclaim_lock);
3378 uma_zwait(uma_zone_t zone)
3381 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3382 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3383 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3384 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3386 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3390 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3392 void *item, *pcpu_item;
3396 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3398 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3401 pcpu_item = zpcpu_base_to_offset(item);
3402 if (flags & M_ZERO) {
3404 for (i = 0; i <= mp_maxid; i++)
3405 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3407 bzero(item, zone->uz_size);
3414 * A stub while both regular and pcpu cases are identical.
3417 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3422 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3425 /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */
3426 if (pcpu_item == NULL)
3429 item = zpcpu_offset_to_base(pcpu_item);
3430 uma_zfree_arg(zone, item, udata);
3433 static inline void *
3434 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3441 kasan_mark_item_valid(zone, item);
3442 kmsan_mark_item_uninitialized(zone, item);
3445 skipdbg = uma_dbg_zskip(zone, item);
3446 if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3447 zone->uz_ctor != trash_ctor)
3448 trash_ctor(item, size, udata, flags);
3451 /* Check flags before loading ctor pointer. */
3452 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3453 __predict_false(zone->uz_ctor != NULL) &&
3454 zone->uz_ctor(item, size, udata, flags) != 0) {
3455 counter_u64_add(zone->uz_fails, 1);
3456 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3461 uma_dbg_alloc(zone, NULL, item);
3463 if (__predict_false(flags & M_ZERO))
3464 return (memset(item, 0, size));
3470 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3471 enum zfreeskip skip)
3476 skipdbg = uma_dbg_zskip(zone, item);
3477 if (skip == SKIP_NONE && !skipdbg) {
3478 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3479 uma_dbg_free(zone, udata, item);
3481 uma_dbg_free(zone, NULL, item);
3484 if (__predict_true(skip < SKIP_DTOR)) {
3485 if (zone->uz_dtor != NULL)
3486 zone->uz_dtor(item, size, udata);
3488 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3489 zone->uz_dtor != trash_dtor)
3490 trash_dtor(item, size, udata);
3493 kasan_mark_item_invalid(zone, item);
3498 item_domain(void *item)
3502 domain = vm_phys_domain(vtophys(item));
3503 KASSERT(domain >= 0 && domain < vm_ndomains,
3504 ("%s: unknown domain for item %p", __func__, item));
3509 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3510 #if defined(INVARIANTS) && (defined(DDB) || defined(STACK))
3511 #include <sys/stack.h>
3513 #define UMA_ZALLOC_DEBUG
3515 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3521 if (flags & M_WAITOK) {
3522 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3523 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3528 KASSERT((flags & M_EXEC) == 0,
3529 ("uma_zalloc_debug: called with M_EXEC"));
3530 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3531 ("uma_zalloc_debug: called within spinlock or critical section"));
3532 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3533 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3535 _Static_assert(M_NOWAIT != 0 && M_WAITOK != 0,
3536 "M_NOWAIT and M_WAITOK must be non-zero for this assertion:");
3539 * Give the #elif clause time to find problems, then remove it
3540 * and enable this. (Remove <sys/stack.h> above, too.)
3542 KASSERT((flags & (M_NOWAIT|M_WAITOK)) == M_NOWAIT ||
3543 (flags & (M_NOWAIT|M_WAITOK)) == M_WAITOK,
3544 ("uma_zalloc_debug: must pass one of M_NOWAIT or M_WAITOK"));
3545 #elif defined(DDB) || defined(STACK)
3546 if (__predict_false((flags & (M_NOWAIT|M_WAITOK)) != M_NOWAIT &&
3547 (flags & (M_NOWAIT|M_WAITOK)) != M_WAITOK)) {
3548 static int stack_count;
3551 if (stack_count < 10) {
3553 printf("uma_zalloc* called with bad WAIT flags:\n");
3561 #ifdef DEBUG_MEMGUARD
3562 if ((zone->uz_flags & (UMA_ZONE_SMR | UMA_ZFLAG_CACHE)) == 0 &&
3563 memguard_cmp_zone(zone)) {
3565 item = memguard_alloc(zone->uz_size, flags);
3567 error = EJUSTRETURN;
3568 if (zone->uz_init != NULL &&
3569 zone->uz_init(item, zone->uz_size, flags) != 0) {
3573 if (zone->uz_ctor != NULL &&
3574 zone->uz_ctor(item, zone->uz_size, udata,
3576 counter_u64_add(zone->uz_fails, 1);
3577 if (zone->uz_fini != NULL)
3578 zone->uz_fini(item, zone->uz_size);
3585 /* This is unfortunate but should not be fatal. */
3592 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3594 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3595 ("uma_zfree_debug: called with spinlock or critical section held"));
3597 #ifdef DEBUG_MEMGUARD
3598 if ((zone->uz_flags & (UMA_ZONE_SMR | UMA_ZFLAG_CACHE)) == 0 &&
3599 is_memguard_addr(item)) {
3600 if (zone->uz_dtor != NULL)
3601 zone->uz_dtor(item, zone->uz_size, udata);
3602 if (zone->uz_fini != NULL)
3603 zone->uz_fini(item, zone->uz_size);
3604 memguard_free(item);
3605 return (EJUSTRETURN);
3612 static inline void *
3613 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3614 void *udata, int flags)
3619 item = cache_bucket_pop(cache, bucket);
3620 size = cache_uz_size(cache);
3621 uz_flags = cache_uz_flags(cache);
3623 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3626 static __noinline void *
3627 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3629 uma_cache_bucket_t bucket;
3632 while (cache_alloc(zone, cache, udata, flags)) {
3633 cache = &zone->uz_cpu[curcpu];
3634 bucket = &cache->uc_allocbucket;
3635 if (__predict_false(bucket->ucb_cnt == 0))
3637 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3642 * We can not get a bucket so try to return a single item.
3644 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3645 domain = PCPU_GET(domain);
3647 domain = UMA_ANYDOMAIN;
3648 return (zone_alloc_item(zone, udata, domain, flags));
3653 uma_zalloc_smr(uma_zone_t zone, int flags)
3655 uma_cache_bucket_t bucket;
3658 CTR3(KTR_UMA, "uma_zalloc_smr zone %s(%p) flags %d", zone->uz_name,
3661 #ifdef UMA_ZALLOC_DEBUG
3664 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3665 ("uma_zalloc_arg: called with non-SMR zone."));
3666 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3671 cache = &zone->uz_cpu[curcpu];
3672 bucket = &cache->uc_allocbucket;
3673 if (__predict_false(bucket->ucb_cnt == 0))
3674 return (cache_alloc_retry(zone, cache, NULL, flags));
3675 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3680 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3682 uma_cache_bucket_t bucket;
3685 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3686 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3688 /* This is the fast path allocation */
3689 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3692 #ifdef UMA_ZALLOC_DEBUG
3695 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3696 ("uma_zalloc_arg: called with SMR zone."));
3697 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3702 * If possible, allocate from the per-CPU cache. There are two
3703 * requirements for safe access to the per-CPU cache: (1) the thread
3704 * accessing the cache must not be preempted or yield during access,
3705 * and (2) the thread must not migrate CPUs without switching which
3706 * cache it accesses. We rely on a critical section to prevent
3707 * preemption and migration. We release the critical section in
3708 * order to acquire the zone mutex if we are unable to allocate from
3709 * the current cache; when we re-acquire the critical section, we
3710 * must detect and handle migration if it has occurred.
3713 cache = &zone->uz_cpu[curcpu];
3714 bucket = &cache->uc_allocbucket;
3715 if (__predict_false(bucket->ucb_cnt == 0))
3716 return (cache_alloc_retry(zone, cache, udata, flags));
3717 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3721 * Replenish an alloc bucket and possibly restore an old one. Called in
3722 * a critical section. Returns in a critical section.
3724 * A false return value indicates an allocation failure.
3725 * A true return value indicates success and the caller should retry.
3727 static __noinline bool
3728 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3730 uma_bucket_t bucket;
3731 int curdomain, domain;
3734 CRITICAL_ASSERT(curthread);
3737 * If we have run out of items in our alloc bucket see
3738 * if we can switch with the free bucket.
3740 * SMR Zones can't re-use the free bucket until the sequence has
3743 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3744 cache->uc_freebucket.ucb_cnt != 0) {
3745 cache_bucket_swap(&cache->uc_freebucket,
3746 &cache->uc_allocbucket);
3751 * Discard any empty allocation bucket while we hold no locks.
3753 bucket = cache_bucket_unload_alloc(cache);
3756 if (bucket != NULL) {
3757 KASSERT(bucket->ub_cnt == 0,
3758 ("cache_alloc: Entered with non-empty alloc bucket."));
3759 bucket_free(zone, bucket, udata);
3763 * Attempt to retrieve the item from the per-CPU cache has failed, so
3764 * we must go back to the zone. This requires the zdom lock, so we
3765 * must drop the critical section, then re-acquire it when we go back
3766 * to the cache. Since the critical section is released, we may be
3767 * preempted or migrate. As such, make sure not to maintain any
3768 * thread-local state specific to the cache from prior to releasing
3769 * the critical section.
3771 domain = PCPU_GET(domain);
3772 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3773 VM_DOMAIN_EMPTY(domain))
3774 domain = zone_domain_highest(zone, domain);
3775 bucket = cache_fetch_bucket(zone, cache, domain);
3776 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3777 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3783 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3784 zone->uz_name, zone, bucket);
3785 if (bucket == NULL) {
3791 * See if we lost the race or were migrated. Cache the
3792 * initialized bucket to make this less likely or claim
3793 * the memory directly.
3796 cache = &zone->uz_cpu[curcpu];
3797 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3798 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3799 (curdomain = PCPU_GET(domain)) == domain ||
3800 VM_DOMAIN_EMPTY(curdomain))) {
3802 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3804 cache_bucket_load_alloc(cache, bucket);
3809 * We lost the race, release this bucket and start over.
3812 zone_put_bucket(zone, domain, bucket, udata, !new);
3819 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3822 uma_bucket_t bucket;
3823 uma_zone_domain_t zdom;
3827 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3828 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3830 /* This is the fast path allocation */
3831 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3832 zone->uz_name, zone, domain, flags);
3834 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3835 ("uma_zalloc_domain: called with SMR zone."));
3837 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3838 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3840 if (vm_ndomains == 1)
3841 return (uma_zalloc_arg(zone, udata, flags));
3843 #ifdef UMA_ZALLOC_DEBUG
3844 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3849 * Try to allocate from the bucket cache before falling back to the keg.
3850 * We could try harder and attempt to allocate from per-CPU caches or
3851 * the per-domain cross-domain buckets, but the complexity is probably
3852 * not worth it. It is more important that frees of previous
3853 * cross-domain allocations do not blow up the cache.
3855 zdom = zone_domain_lock(zone, domain);
3856 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3857 item = bucket->ub_bucket[bucket->ub_cnt - 1];
3859 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3862 zone_put_bucket(zone, domain, bucket, udata, true);
3863 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3866 KASSERT(item_domain(item) == domain,
3867 ("%s: bucket cache item %p from wrong domain",
3869 counter_u64_add(zone->uz_allocs, 1);
3874 return (zone_alloc_item(zone, udata, domain, flags));
3876 return (uma_zalloc_arg(zone, udata, flags));
3881 * Find a slab with some space. Prefer slabs that are partially used over those
3882 * that are totally full. This helps to reduce fragmentation.
3884 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3888 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3894 KASSERT(domain >= 0 && domain < vm_ndomains,
3895 ("keg_first_slab: domain %d out of range", domain));
3896 KEG_LOCK_ASSERT(keg, domain);
3901 dom = &keg->uk_domain[domain];
3902 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3904 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3905 LIST_REMOVE(slab, us_link);
3906 dom->ud_free_slabs--;
3907 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3911 domain = (domain + 1) % vm_ndomains;
3912 } while (domain != start);
3918 * Fetch an existing slab from a free or partial list. Returns with the
3919 * keg domain lock held if a slab was found or unlocked if not.
3922 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3927 /* HASH has a single free list. */
3928 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3931 KEG_LOCK(keg, domain);
3932 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3933 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3934 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3935 KEG_UNLOCK(keg, domain);
3942 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3944 struct vm_domainset_iter di;
3949 KASSERT((flags & (M_WAITOK | M_NOVM)) != (M_WAITOK | M_NOVM),
3950 ("%s: invalid flags %#x", __func__, flags));
3954 * Use the keg's policy if upper layers haven't already specified a
3955 * domain (as happens with first-touch zones).
3957 * To avoid races we run the iterator with the keg lock held, but that
3958 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3959 * clear M_WAITOK and handle low memory conditions locally.
3961 rr = rdomain == UMA_ANYDOMAIN;
3963 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3964 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3972 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3977 * M_NOVM is used to break the recursion that can otherwise
3978 * occur if low-level memory management routines use UMA.
3980 if ((flags & M_NOVM) == 0) {
3981 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3987 if ((flags & M_USE_RESERVE) != 0) {
3989 * Drain reserves from other domains before
3990 * giving up or sleeping. It may be useful to
3991 * support per-domain reserves eventually.
3993 rdomain = UMA_ANYDOMAIN;
3996 if ((flags & M_WAITOK) == 0)
3998 vm_wait_domain(domain);
3999 } else if (vm_domainset_iter_policy(&di, &domain) != 0) {
4000 if ((flags & M_WAITOK) != 0) {
4001 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
4009 * We might not have been able to get a slab but another cpu
4010 * could have while we were unlocked. Check again before we
4013 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
4020 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
4026 KEG_LOCK_ASSERT(keg, slab->us_domain);
4028 dom = &keg->uk_domain[slab->us_domain];
4029 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
4030 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
4031 item = slab_item(slab, keg, freei);
4032 slab->us_freecount--;
4033 dom->ud_free_items--;
4036 * Move this slab to the full list. It must be on the partial list, so
4037 * we do not need to update the free slab count. In particular,
4038 * keg_fetch_slab() always returns slabs on the partial list.
4040 if (slab->us_freecount == 0) {
4041 LIST_REMOVE(slab, us_link);
4042 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
4049 zone_import(void *arg, void **bucket, int max, int domain, int flags)
4063 /* Try to keep the buckets totally full */
4064 for (i = 0; i < max; ) {
4065 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
4068 stripe = howmany(max, vm_ndomains);
4070 dom = &keg->uk_domain[slab->us_domain];
4072 bucket[i++] = slab_alloc_item(keg, slab);
4073 if (keg->uk_reserve > 0 &&
4074 dom->ud_free_items <= keg->uk_reserve) {
4076 * Avoid depleting the reserve after a
4077 * successful item allocation, even if
4078 * M_USE_RESERVE is specified.
4080 KEG_UNLOCK(keg, slab->us_domain);
4085 * If the zone is striped we pick a new slab for every
4086 * N allocations. Eliminating this conditional will
4087 * instead pick a new domain for each bucket rather
4088 * than stripe within each bucket. The current option
4089 * produces more fragmentation and requires more cpu
4090 * time but yields better distribution.
4092 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
4093 vm_ndomains > 1 && --stripe == 0)
4096 } while (slab->us_freecount != 0 && i < max);
4097 KEG_UNLOCK(keg, slab->us_domain);
4099 /* Don't block if we allocated any successfully. */
4108 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
4110 uint64_t old, new, total, max;
4113 * The hard case. We're going to sleep because there were existing
4114 * sleepers or because we ran out of items. This routine enforces
4115 * fairness by keeping fifo order.
4117 * First release our ill gotten gains and make some noise.
4120 zone_free_limit(zone, count);
4121 zone_log_warning(zone);
4122 zone_maxaction(zone);
4123 if (flags & M_NOWAIT)
4127 * We need to allocate an item or set ourself as a sleeper
4128 * while the sleepq lock is held to avoid wakeup races. This
4129 * is essentially a home rolled semaphore.
4131 sleepq_lock(&zone->uz_max_items);
4132 old = zone->uz_items;
4134 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
4135 /* Cache the max since we will evaluate twice. */
4136 max = zone->uz_max_items;
4137 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
4138 UZ_ITEMS_COUNT(old) >= max)
4139 new = old + UZ_ITEMS_SLEEPER;
4141 new = old + MIN(count, max - old);
4142 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
4144 /* We may have successfully allocated under the sleepq lock. */
4145 if (UZ_ITEMS_SLEEPERS(new) == 0) {
4146 sleepq_release(&zone->uz_max_items);
4151 * This is in a different cacheline from uz_items so that we
4152 * don't constantly invalidate the fastpath cacheline when we
4153 * adjust item counts. This could be limited to toggling on
4156 atomic_add_32(&zone->uz_sleepers, 1);
4157 atomic_add_64(&zone->uz_sleeps, 1);
4160 * We have added ourselves as a sleeper. The sleepq lock
4161 * protects us from wakeup races. Sleep now and then retry.
4163 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
4164 sleepq_wait(&zone->uz_max_items, PVM);
4167 * After wakeup, remove ourselves as a sleeper and try
4168 * again. We no longer have the sleepq lock for protection.
4170 * Subract ourselves as a sleeper while attempting to add
4173 atomic_subtract_32(&zone->uz_sleepers, 1);
4174 old = atomic_fetchadd_64(&zone->uz_items,
4175 -(UZ_ITEMS_SLEEPER - count));
4176 /* We're no longer a sleeper. */
4177 old -= UZ_ITEMS_SLEEPER;
4180 * If we're still at the limit, restart. Notably do not
4181 * block on other sleepers. Cache the max value to protect
4182 * against changes via sysctl.
4184 total = UZ_ITEMS_COUNT(old);
4185 max = zone->uz_max_items;
4188 /* Truncate if necessary, otherwise wake other sleepers. */
4189 if (total + count > max) {
4190 zone_free_limit(zone, total + count - max);
4191 count = max - total;
4192 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
4193 wakeup_one(&zone->uz_max_items);
4200 * Allocate 'count' items from our max_items limit. Returns the number
4201 * available. If M_NOWAIT is not specified it will sleep until at least
4202 * one item can be allocated.
4205 zone_alloc_limit(uma_zone_t zone, int count, int flags)
4210 max = zone->uz_max_items;
4214 * We expect normal allocations to succeed with a simple
4217 old = atomic_fetchadd_64(&zone->uz_items, count);
4218 if (__predict_true(old + count <= max))
4222 * If we had some items and no sleepers just return the
4223 * truncated value. We have to release the excess space
4224 * though because that may wake sleepers who weren't woken
4225 * because we were temporarily over the limit.
4228 zone_free_limit(zone, (old + count) - max);
4231 return (zone_alloc_limit_hard(zone, count, flags));
4235 * Free a number of items back to the limit.
4238 zone_free_limit(uma_zone_t zone, int count)
4245 * In the common case we either have no sleepers or
4246 * are still over the limit and can just return.
4248 old = atomic_fetchadd_64(&zone->uz_items, -count);
4249 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4250 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4254 * Moderate the rate of wakeups. Sleepers will continue
4255 * to generate wakeups if necessary.
4257 wakeup_one(&zone->uz_max_items);
4261 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4263 uma_bucket_t bucket;
4264 int error, maxbucket, cnt;
4266 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4269 /* Avoid allocs targeting empty domains. */
4270 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4271 domain = UMA_ANYDOMAIN;
4272 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4273 domain = UMA_ANYDOMAIN;
4275 if (zone->uz_max_items > 0)
4276 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4279 maxbucket = zone->uz_bucket_size;
4283 /* Don't wait for buckets, preserve caller's NOVM setting. */
4284 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4285 if (bucket == NULL) {
4290 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4291 MIN(maxbucket, bucket->ub_entries), domain, flags);
4294 * Initialize the memory if necessary.
4296 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4299 for (i = 0; i < bucket->ub_cnt; i++) {
4300 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
4301 error = zone->uz_init(bucket->ub_bucket[i],
4302 zone->uz_size, flags);
4303 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
4309 * If we couldn't initialize the whole bucket, put the
4310 * rest back onto the freelist.
4312 if (i != bucket->ub_cnt) {
4313 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4314 bucket->ub_cnt - i);
4316 bzero(&bucket->ub_bucket[i],
4317 sizeof(void *) * (bucket->ub_cnt - i));
4323 cnt = bucket->ub_cnt;
4324 if (bucket->ub_cnt == 0) {
4325 bucket_free(zone, bucket, udata);
4326 counter_u64_add(zone->uz_fails, 1);
4330 if (zone->uz_max_items > 0 && cnt < maxbucket)
4331 zone_free_limit(zone, maxbucket - cnt);
4337 * Allocates a single item from a zone.
4340 * zone The zone to alloc for.
4341 * udata The data to be passed to the constructor.
4342 * domain The domain to allocate from or UMA_ANYDOMAIN.
4343 * flags M_WAITOK, M_NOWAIT, M_ZERO.
4346 * NULL if there is no memory and M_NOWAIT is set
4347 * An item if successful
4351 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4355 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4356 counter_u64_add(zone->uz_fails, 1);
4360 /* Avoid allocs targeting empty domains. */
4361 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4362 domain = UMA_ANYDOMAIN;
4364 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4368 * We have to call both the zone's init (not the keg's init)
4369 * and the zone's ctor. This is because the item is going from
4370 * a keg slab directly to the user, and the user is expecting it
4371 * to be both zone-init'd as well as zone-ctor'd.
4373 if (zone->uz_init != NULL) {
4376 kasan_mark_item_valid(zone, item);
4377 error = zone->uz_init(item, zone->uz_size, flags);
4378 kasan_mark_item_invalid(zone, item);
4380 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4384 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4389 counter_u64_add(zone->uz_allocs, 1);
4390 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4391 zone->uz_name, zone);
4396 counter_u64_add(zone->uz_fails, 1);
4398 if (zone->uz_max_items > 0)
4399 zone_free_limit(zone, 1);
4400 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4401 zone->uz_name, zone);
4408 uma_zfree_smr(uma_zone_t zone, void *item)
4411 uma_cache_bucket_t bucket;
4417 CTR3(KTR_UMA, "uma_zfree_smr zone %s(%p) item %p",
4418 zone->uz_name, zone, item);
4420 #ifdef UMA_ZALLOC_DEBUG
4421 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4422 ("uma_zfree_smr: called with non-SMR zone."));
4423 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4424 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4425 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4428 cache = &zone->uz_cpu[curcpu];
4431 uz_flags = cache_uz_flags(cache);
4432 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4433 itemdomain = item_domain(item);
4437 cache = &zone->uz_cpu[curcpu];
4438 /* SMR Zones must free to the free bucket. */
4439 bucket = &cache->uc_freebucket;
4441 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4442 PCPU_GET(domain) != itemdomain) {
4443 bucket = &cache->uc_crossbucket;
4446 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4447 cache_bucket_push(cache, bucket, item);
4451 } while (cache_free(zone, cache, NULL, itemdomain));
4455 * If nothing else caught this, we'll just do an internal free.
4457 zone_free_item(zone, item, NULL, SKIP_NONE);
4462 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4465 uma_cache_bucket_t bucket;
4466 int itemdomain, uz_flags;
4468 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4469 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4471 CTR3(KTR_UMA, "uma_zfree_arg zone %s(%p) item %p",
4472 zone->uz_name, zone, item);
4474 #ifdef UMA_ZALLOC_DEBUG
4475 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4476 ("uma_zfree_arg: called with SMR zone."));
4477 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4480 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4485 * We are accessing the per-cpu cache without a critical section to
4486 * fetch size and flags. This is acceptable, if we are preempted we
4487 * will simply read another cpu's line.
4489 cache = &zone->uz_cpu[curcpu];
4490 uz_flags = cache_uz_flags(cache);
4491 if (UMA_ALWAYS_CTORDTOR ||
4492 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4493 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4496 * The race here is acceptable. If we miss it we'll just have to wait
4497 * a little longer for the limits to be reset.
4499 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4500 if (atomic_load_32(&zone->uz_sleepers) > 0)
4505 * If possible, free to the per-CPU cache. There are two
4506 * requirements for safe access to the per-CPU cache: (1) the thread
4507 * accessing the cache must not be preempted or yield during access,
4508 * and (2) the thread must not migrate CPUs without switching which
4509 * cache it accesses. We rely on a critical section to prevent
4510 * preemption and migration. We release the critical section in
4511 * order to acquire the zone mutex if we are unable to free to the
4512 * current cache; when we re-acquire the critical section, we must
4513 * detect and handle migration if it has occurred.
4517 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4518 itemdomain = item_domain(item);
4522 cache = &zone->uz_cpu[curcpu];
4524 * Try to free into the allocbucket first to give LIFO
4525 * ordering for cache-hot datastructures. Spill over
4526 * into the freebucket if necessary. Alloc will swap
4527 * them if one runs dry.
4529 bucket = &cache->uc_allocbucket;
4531 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4532 PCPU_GET(domain) != itemdomain) {
4533 bucket = &cache->uc_crossbucket;
4536 if (bucket->ucb_cnt == bucket->ucb_entries &&
4537 cache->uc_freebucket.ucb_cnt <
4538 cache->uc_freebucket.ucb_entries)
4539 cache_bucket_swap(&cache->uc_freebucket,
4540 &cache->uc_allocbucket);
4541 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4542 cache_bucket_push(cache, bucket, item);
4546 } while (cache_free(zone, cache, udata, itemdomain));
4550 * If nothing else caught this, we'll just do an internal free.
4553 zone_free_item(zone, item, udata, SKIP_DTOR);
4558 * sort crossdomain free buckets to domain correct buckets and cache
4562 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4564 struct uma_bucketlist emptybuckets, fullbuckets;
4565 uma_zone_domain_t zdom;
4572 "uma_zfree: zone %s(%p) draining cross bucket %p",
4573 zone->uz_name, zone, bucket);
4576 * It is possible for buckets to arrive here out of order so we fetch
4577 * the current smr seq rather than accepting the bucket's.
4579 seq = SMR_SEQ_INVALID;
4580 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4581 seq = smr_advance(zone->uz_smr);
4584 * To avoid having ndomain * ndomain buckets for sorting we have a
4585 * lock on the current crossfree bucket. A full matrix with
4586 * per-domain locking could be used if necessary.
4588 STAILQ_INIT(&emptybuckets);
4589 STAILQ_INIT(&fullbuckets);
4590 ZONE_CROSS_LOCK(zone);
4591 for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4592 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4593 domain = item_domain(item);
4594 zdom = ZDOM_GET(zone, domain);
4595 if (zdom->uzd_cross == NULL) {
4596 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4597 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4598 zdom->uzd_cross = b;
4601 * Avoid allocating a bucket with the cross lock
4602 * held, since allocation can trigger a
4603 * cross-domain free and bucket zones may
4604 * allocate from each other.
4606 ZONE_CROSS_UNLOCK(zone);
4607 b = bucket_alloc(zone, udata, M_NOWAIT);
4610 ZONE_CROSS_LOCK(zone);
4611 if (zdom->uzd_cross != NULL) {
4612 STAILQ_INSERT_HEAD(&emptybuckets, b,
4615 zdom->uzd_cross = b;
4619 b = zdom->uzd_cross;
4620 b->ub_bucket[b->ub_cnt++] = item;
4622 if (b->ub_cnt == b->ub_entries) {
4623 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4624 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4625 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4626 zdom->uzd_cross = b;
4629 ZONE_CROSS_UNLOCK(zone);
4631 if (bucket->ub_cnt == 0)
4632 bucket->ub_seq = SMR_SEQ_INVALID;
4633 bucket_free(zone, bucket, udata);
4635 while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4636 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4637 bucket_free(zone, b, udata);
4639 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4640 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4641 domain = item_domain(b->ub_bucket[0]);
4642 zone_put_bucket(zone, domain, b, udata, true);
4648 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4649 int itemdomain, bool ws)
4654 * Buckets coming from the wrong domain will be entirely for the
4655 * only other domain on two domain systems. In this case we can
4656 * simply cache them. Otherwise we need to sort them back to
4659 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4660 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4661 zone_free_cross(zone, bucket, udata);
4667 * Attempt to save the bucket in the zone's domain bucket cache.
4670 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4671 zone->uz_name, zone, bucket);
4672 /* ub_cnt is pointing to the last free item */
4673 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4674 itemdomain = zone_domain_lowest(zone, itemdomain);
4675 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4679 * Populate a free or cross bucket for the current cpu cache. Free any
4680 * existing full bucket either to the zone cache or back to the slab layer.
4682 * Enters and returns in a critical section. false return indicates that
4683 * we can not satisfy this free in the cache layer. true indicates that
4684 * the caller should retry.
4686 static __noinline bool
4687 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, int itemdomain)
4689 uma_cache_bucket_t cbucket;
4690 uma_bucket_t newbucket, bucket;
4692 CRITICAL_ASSERT(curthread);
4694 if (zone->uz_bucket_size == 0)
4697 cache = &zone->uz_cpu[curcpu];
4701 * FIRSTTOUCH domains need to free to the correct zdom. When
4702 * enabled this is the zdom of the item. The bucket is the
4703 * cross bucket if the current domain and itemdomain do not match.
4705 cbucket = &cache->uc_freebucket;
4707 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4708 if (PCPU_GET(domain) != itemdomain) {
4709 cbucket = &cache->uc_crossbucket;
4710 if (cbucket->ucb_cnt != 0)
4711 counter_u64_add(zone->uz_xdomain,
4716 bucket = cache_bucket_unload(cbucket);
4717 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4718 ("cache_free: Entered with non-full free bucket."));
4720 /* We are no longer associated with this CPU. */
4724 * Don't let SMR zones operate without a free bucket. Force
4725 * a synchronize and re-use this one. We will only degrade
4726 * to a synchronize every bucket_size items rather than every
4727 * item if we fail to allocate a bucket.
4729 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4731 bucket->ub_seq = smr_advance(zone->uz_smr);
4732 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4733 if (newbucket == NULL && bucket != NULL) {
4734 bucket_drain(zone, bucket);
4738 } else if (!bucketdisable)
4739 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4742 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4745 if ((bucket = newbucket) == NULL)
4747 cache = &zone->uz_cpu[curcpu];
4750 * Check to see if we should be populating the cross bucket. If it
4751 * is already populated we will fall through and attempt to populate
4754 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4755 if (PCPU_GET(domain) != itemdomain &&
4756 cache->uc_crossbucket.ucb_bucket == NULL) {
4757 cache_bucket_load_cross(cache, bucket);
4763 * We may have lost the race to fill the bucket or switched CPUs.
4765 if (cache->uc_freebucket.ucb_bucket != NULL) {
4767 bucket_free(zone, bucket, udata);
4770 cache_bucket_load_free(cache, bucket);
4776 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4783 KEG_LOCK_ASSERT(keg, slab->us_domain);
4785 /* Do we need to remove from any lists? */
4786 dom = &keg->uk_domain[slab->us_domain];
4787 if (slab->us_freecount + 1 == keg->uk_ipers) {
4788 LIST_REMOVE(slab, us_link);
4789 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4790 dom->ud_free_slabs++;
4791 } else if (slab->us_freecount == 0) {
4792 LIST_REMOVE(slab, us_link);
4793 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4796 /* Slab management. */
4797 freei = slab_item_index(slab, keg, item);
4798 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4799 slab->us_freecount++;
4801 /* Keg statistics. */
4802 dom->ud_free_items++;
4806 zone_release(void *arg, void **bucket, int cnt)
4819 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4820 lock = KEG_LOCK(keg, 0);
4821 for (i = 0; i < cnt; i++) {
4823 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4824 slab = vtoslab((vm_offset_t)item);
4826 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4827 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4828 slab = hash_sfind(&keg->uk_hash, mem);
4830 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4832 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4835 lock = KEG_LOCK(keg, slab->us_domain);
4837 slab_free_item(zone, slab, item);
4844 * Frees a single item to any zone.
4847 * zone The zone to free to
4848 * item The item we're freeing
4849 * udata User supplied data for the dtor
4850 * skip Skip dtors and finis
4852 static __noinline void
4853 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4857 * If a free is sent directly to an SMR zone we have to
4858 * synchronize immediately because the item can instantly
4859 * be reallocated. This should only happen in degenerate
4860 * cases when no memory is available for per-cpu caches.
4862 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4863 smr_synchronize(zone->uz_smr);
4865 item_dtor(zone, item, zone->uz_size, udata, skip);
4867 if (skip < SKIP_FINI && zone->uz_fini) {
4868 kasan_mark_item_valid(zone, item);
4869 zone->uz_fini(item, zone->uz_size);
4870 kasan_mark_item_invalid(zone, item);
4873 zone->uz_release(zone->uz_arg, &item, 1);
4875 if (skip & SKIP_CNT)
4878 counter_u64_add(zone->uz_frees, 1);
4880 if (zone->uz_max_items > 0)
4881 zone_free_limit(zone, 1);
4886 uma_zone_set_max(uma_zone_t zone, int nitems)
4890 * If the limit is small, we may need to constrain the maximum per-CPU
4891 * cache size, or disable caching entirely.
4893 uma_zone_set_maxcache(zone, nitems);
4896 * XXX This can misbehave if the zone has any allocations with
4897 * no limit and a limit is imposed. There is currently no
4898 * way to clear a limit.
4901 if (zone->uz_max_items == 0)
4902 ZONE_ASSERT_COLD(zone);
4903 zone->uz_max_items = nitems;
4904 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4905 zone_update_caches(zone);
4906 /* We may need to wake waiters. */
4907 wakeup(&zone->uz_max_items);
4915 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4917 int bpcpu, bpdom, bsize, nb;
4922 * Compute a lower bound on the number of items that may be cached in
4923 * the zone. Each CPU gets at least two buckets, and for cross-domain
4924 * frees we use an additional bucket per CPU and per domain. Select the
4925 * largest bucket size that does not exceed half of the requested limit,
4926 * with the left over space given to the full bucket cache.
4931 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4936 nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4937 bsize = nitems / nb / 2;
4938 if (bsize > BUCKET_MAX)
4940 else if (bsize == 0 && nitems / nb > 0)
4942 zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
4943 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4944 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4945 zone->uz_bucket_max = nitems - nb * bsize;
4951 uma_zone_get_max(uma_zone_t zone)
4955 nitems = atomic_load_64(&zone->uz_max_items);
4962 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4965 ZONE_ASSERT_COLD(zone);
4966 zone->uz_warning = warning;
4971 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4974 ZONE_ASSERT_COLD(zone);
4975 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4980 uma_zone_get_cur(uma_zone_t zone)
4986 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4987 nitems = counter_u64_fetch(zone->uz_allocs) -
4988 counter_u64_fetch(zone->uz_frees);
4990 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4991 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4993 return (nitems < 0 ? 0 : nitems);
4997 uma_zone_get_allocs(uma_zone_t zone)
5003 if (zone->uz_allocs != EARLY_COUNTER)
5004 nitems = counter_u64_fetch(zone->uz_allocs);
5006 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
5012 uma_zone_get_frees(uma_zone_t zone)
5018 if (zone->uz_frees != EARLY_COUNTER)
5019 nitems = counter_u64_fetch(zone->uz_frees);
5021 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
5027 /* Used only for KEG_ASSERT_COLD(). */
5029 uma_keg_get_allocs(uma_keg_t keg)
5035 LIST_FOREACH(z, &keg->uk_zones, uz_link)
5036 nitems += uma_zone_get_allocs(z);
5044 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
5049 KEG_ASSERT_COLD(keg);
5050 keg->uk_init = uminit;
5055 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
5060 KEG_ASSERT_COLD(keg);
5061 keg->uk_fini = fini;
5066 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
5069 ZONE_ASSERT_COLD(zone);
5070 zone->uz_init = zinit;
5075 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
5078 ZONE_ASSERT_COLD(zone);
5079 zone->uz_fini = zfini;
5084 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
5089 KEG_ASSERT_COLD(keg);
5090 keg->uk_freef = freef;
5095 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
5100 KEG_ASSERT_COLD(keg);
5101 keg->uk_allocf = allocf;
5106 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
5109 ZONE_ASSERT_COLD(zone);
5111 KASSERT(smr != NULL, ("Got NULL smr"));
5112 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
5113 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
5114 zone->uz_flags |= UMA_ZONE_SMR;
5116 zone_update_caches(zone);
5120 uma_zone_get_smr(uma_zone_t zone)
5123 return (zone->uz_smr);
5128 uma_zone_reserve(uma_zone_t zone, int items)
5133 KEG_ASSERT_COLD(keg);
5134 keg->uk_reserve = items;
5139 uma_zone_reserve_kva(uma_zone_t zone, int count)
5146 KEG_ASSERT_COLD(keg);
5147 ZONE_ASSERT_COLD(zone);
5149 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
5151 #ifdef UMA_MD_SMALL_ALLOC
5152 if (keg->uk_ppera > 1) {
5156 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
5162 MPASS(keg->uk_kva == 0);
5165 zone->uz_max_items = pages * keg->uk_ipers;
5166 #ifdef UMA_MD_SMALL_ALLOC
5167 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
5169 keg->uk_allocf = noobj_alloc;
5171 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5172 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5173 zone_update_caches(zone);
5180 uma_prealloc(uma_zone_t zone, int items)
5182 struct vm_domainset_iter di;
5186 int aflags, domain, slabs;
5189 slabs = howmany(items, keg->uk_ipers);
5190 while (slabs-- > 0) {
5192 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
5195 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
5198 dom = &keg->uk_domain[slab->us_domain];
5200 * keg_alloc_slab() always returns a slab on the
5203 LIST_REMOVE(slab, us_link);
5204 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
5206 dom->ud_free_slabs++;
5207 KEG_UNLOCK(keg, slab->us_domain);
5210 if (vm_domainset_iter_policy(&di, &domain) != 0)
5211 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
5217 * Returns a snapshot of memory consumption in bytes.
5220 uma_zone_memory(uma_zone_t zone)
5226 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
5227 for (i = 0; i < vm_ndomains; i++)
5228 sz += ZDOM_GET(zone, i)->uzd_nitems;
5229 return (sz * zone->uz_size);
5231 for (i = 0; i < vm_ndomains; i++)
5232 sz += zone->uz_keg->uk_domain[i].ud_pages;
5234 return (sz * PAGE_SIZE);
5237 struct uma_reclaim_args {
5243 uma_reclaim_domain_cb(uma_zone_t zone, void *arg)
5245 struct uma_reclaim_args *args;
5248 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0)
5249 uma_zone_reclaim_domain(zone, args->req, args->domain);
5254 uma_reclaim(int req)
5256 uma_reclaim_domain(req, UMA_ANYDOMAIN);
5260 uma_reclaim_domain(int req, int domain)
5262 struct uma_reclaim_args args;
5266 args.domain = domain;
5269 sx_slock(&uma_reclaim_lock);
5271 case UMA_RECLAIM_TRIM:
5272 case UMA_RECLAIM_DRAIN:
5273 zone_foreach(uma_reclaim_domain_cb, &args);
5275 case UMA_RECLAIM_DRAIN_CPU:
5276 zone_foreach(uma_reclaim_domain_cb, &args);
5277 pcpu_cache_drain_safe(NULL);
5278 zone_foreach(uma_reclaim_domain_cb, &args);
5281 panic("unhandled reclamation request %d", req);
5285 * Some slabs may have been freed but this zone will be visited early
5286 * we visit again so that we can free pages that are empty once other
5287 * zones are drained. We have to do the same for buckets.
5289 uma_zone_reclaim_domain(slabzones[0], UMA_RECLAIM_DRAIN, domain);
5290 uma_zone_reclaim_domain(slabzones[1], UMA_RECLAIM_DRAIN, domain);
5291 bucket_zone_drain(domain);
5292 sx_sunlock(&uma_reclaim_lock);
5295 static volatile int uma_reclaim_needed;
5298 uma_reclaim_wakeup(void)
5301 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5302 wakeup(uma_reclaim);
5306 uma_reclaim_worker(void *arg __unused)
5310 sx_xlock(&uma_reclaim_lock);
5311 while (atomic_load_int(&uma_reclaim_needed) == 0)
5312 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5314 sx_xunlock(&uma_reclaim_lock);
5315 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5316 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5317 atomic_store_int(&uma_reclaim_needed, 0);
5318 /* Don't fire more than once per-second. */
5319 pause("umarclslp", hz);
5325 uma_zone_reclaim(uma_zone_t zone, int req)
5327 uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5331 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5334 case UMA_RECLAIM_TRIM:
5335 zone_reclaim(zone, domain, M_NOWAIT, false);
5337 case UMA_RECLAIM_DRAIN:
5338 zone_reclaim(zone, domain, M_NOWAIT, true);
5340 case UMA_RECLAIM_DRAIN_CPU:
5341 pcpu_cache_drain_safe(zone);
5342 zone_reclaim(zone, domain, M_NOWAIT, true);
5345 panic("unhandled reclamation request %d", req);
5351 uma_zone_exhausted(uma_zone_t zone)
5354 return (atomic_load_32(&zone->uz_sleepers) > 0);
5361 return (uma_kmem_limit);
5365 uma_set_limit(unsigned long limit)
5368 uma_kmem_limit = limit;
5375 return (atomic_load_long(&uma_kmem_total));
5382 return (uma_kmem_limit - uma_size());
5387 * Generate statistics across both the zone and its per-cpu cache's. Return
5388 * desired statistics if the pointer is non-NULL for that statistic.
5390 * Note: does not update the zone statistics, as it can't safely clear the
5391 * per-CPU cache statistic.
5395 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5396 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5399 uint64_t allocs, frees, sleeps, xdomain;
5402 allocs = frees = sleeps = xdomain = 0;
5405 cache = &z->uz_cpu[cpu];
5406 cachefree += cache->uc_allocbucket.ucb_cnt;
5407 cachefree += cache->uc_freebucket.ucb_cnt;
5408 xdomain += cache->uc_crossbucket.ucb_cnt;
5409 cachefree += cache->uc_crossbucket.ucb_cnt;
5410 allocs += cache->uc_allocs;
5411 frees += cache->uc_frees;
5413 allocs += counter_u64_fetch(z->uz_allocs);
5414 frees += counter_u64_fetch(z->uz_frees);
5415 xdomain += counter_u64_fetch(z->uz_xdomain);
5416 sleeps += z->uz_sleeps;
5417 if (cachefreep != NULL)
5418 *cachefreep = cachefree;
5419 if (allocsp != NULL)
5423 if (sleepsp != NULL)
5425 if (xdomainp != NULL)
5426 *xdomainp = xdomain;
5431 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5438 rw_rlock(&uma_rwlock);
5439 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5440 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5443 LIST_FOREACH(z, &uma_cachezones, uz_link)
5446 rw_runlock(&uma_rwlock);
5447 return (sysctl_handle_int(oidp, &count, 0, req));
5451 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5452 struct uma_percpu_stat *ups, bool internal)
5454 uma_zone_domain_t zdom;
5458 for (i = 0; i < vm_ndomains; i++) {
5459 zdom = ZDOM_GET(z, i);
5460 uth->uth_zone_free += zdom->uzd_nitems;
5462 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5463 uth->uth_frees = counter_u64_fetch(z->uz_frees);
5464 uth->uth_fails = counter_u64_fetch(z->uz_fails);
5465 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5466 uth->uth_sleeps = z->uz_sleeps;
5468 for (i = 0; i < mp_maxid + 1; i++) {
5469 bzero(&ups[i], sizeof(*ups));
5470 if (internal || CPU_ABSENT(i))
5472 cache = &z->uz_cpu[i];
5473 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5474 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5475 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5476 ups[i].ups_allocs = cache->uc_allocs;
5477 ups[i].ups_frees = cache->uc_frees;
5482 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5484 struct uma_stream_header ush;
5485 struct uma_type_header uth;
5486 struct uma_percpu_stat *ups;
5491 uint32_t kfree, pages;
5492 int count, error, i;
5494 error = sysctl_wire_old_buffer(req, 0);
5497 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5498 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5499 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5502 rw_rlock(&uma_rwlock);
5503 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5504 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5508 LIST_FOREACH(z, &uma_cachezones, uz_link)
5512 * Insert stream header.
5514 bzero(&ush, sizeof(ush));
5515 ush.ush_version = UMA_STREAM_VERSION;
5516 ush.ush_maxcpus = (mp_maxid + 1);
5517 ush.ush_count = count;
5518 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5520 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5522 for (i = 0; i < vm_ndomains; i++) {
5523 kfree += kz->uk_domain[i].ud_free_items;
5524 pages += kz->uk_domain[i].ud_pages;
5526 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5527 bzero(&uth, sizeof(uth));
5528 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5529 uth.uth_align = kz->uk_align;
5530 uth.uth_size = kz->uk_size;
5531 uth.uth_rsize = kz->uk_rsize;
5532 if (z->uz_max_items > 0) {
5533 items = UZ_ITEMS_COUNT(z->uz_items);
5534 uth.uth_pages = (items / kz->uk_ipers) *
5537 uth.uth_pages = pages;
5538 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5540 uth.uth_limit = z->uz_max_items;
5541 uth.uth_keg_free = kfree;
5544 * A zone is secondary is it is not the first entry
5545 * on the keg's zone list.
5547 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5548 (LIST_FIRST(&kz->uk_zones) != z))
5549 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5550 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5551 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5552 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5553 for (i = 0; i < mp_maxid + 1; i++)
5554 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5557 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5558 bzero(&uth, sizeof(uth));
5559 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5560 uth.uth_size = z->uz_size;
5561 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5562 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5563 for (i = 0; i < mp_maxid + 1; i++)
5564 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5567 rw_runlock(&uma_rwlock);
5568 error = sbuf_finish(&sbuf);
5575 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5577 uma_zone_t zone = *(uma_zone_t *)arg1;
5580 max = uma_zone_get_max(zone);
5581 error = sysctl_handle_int(oidp, &max, 0, req);
5582 if (error || !req->newptr)
5585 uma_zone_set_max(zone, max);
5591 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5597 * Some callers want to add sysctls for global zones that
5598 * may not yet exist so they pass a pointer to a pointer.
5601 zone = *(uma_zone_t *)arg1;
5604 cur = uma_zone_get_cur(zone);
5605 return (sysctl_handle_int(oidp, &cur, 0, req));
5609 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5611 uma_zone_t zone = arg1;
5614 cur = uma_zone_get_allocs(zone);
5615 return (sysctl_handle_64(oidp, &cur, 0, req));
5619 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5621 uma_zone_t zone = arg1;
5624 cur = uma_zone_get_frees(zone);
5625 return (sysctl_handle_64(oidp, &cur, 0, req));
5629 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5632 uma_zone_t zone = arg1;
5635 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5636 if (zone->uz_flags != 0)
5637 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5639 sbuf_printf(&sbuf, "0");
5640 error = sbuf_finish(&sbuf);
5647 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5649 uma_keg_t keg = arg1;
5650 int avail, effpct, total;
5652 total = keg->uk_ppera * PAGE_SIZE;
5653 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5654 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5656 * We consider the client's requested size and alignment here, not the
5657 * real size determination uk_rsize, because we also adjust the real
5658 * size for internal implementation reasons (max bitset size).
5660 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5661 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5662 avail *= mp_maxid + 1;
5663 effpct = 100 * avail / total;
5664 return (sysctl_handle_int(oidp, &effpct, 0, req));
5668 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5670 uma_zone_t zone = arg1;
5673 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5674 return (sysctl_handle_64(oidp, &cur, 0, req));
5679 uma_dbg_getslab(uma_zone_t zone, void *item)
5686 * It is safe to return the slab here even though the
5687 * zone is unlocked because the item's allocation state
5688 * essentially holds a reference.
5690 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5691 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5693 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5694 return (vtoslab((vm_offset_t)mem));
5696 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5697 return ((uma_slab_t)(mem + keg->uk_pgoff));
5699 slab = hash_sfind(&keg->uk_hash, mem);
5706 uma_dbg_zskip(uma_zone_t zone, void *mem)
5709 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5712 return (uma_dbg_kskip(zone->uz_keg, mem));
5716 uma_dbg_kskip(uma_keg_t keg, void *mem)
5720 if (dbg_divisor == 0)
5723 if (dbg_divisor == 1)
5726 idx = (uintptr_t)mem >> PAGE_SHIFT;
5727 if (keg->uk_ipers > 1) {
5728 idx *= keg->uk_ipers;
5729 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5732 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5733 counter_u64_add(uma_skip_cnt, 1);
5736 counter_u64_add(uma_dbg_cnt, 1);
5742 * Set up the slab's freei data such that uma_dbg_free can function.
5746 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5752 slab = uma_dbg_getslab(zone, item);
5754 panic("uma: item %p did not belong to zone %s",
5755 item, zone->uz_name);
5758 freei = slab_item_index(slab, keg, item);
5760 if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5761 slab_dbg_bits(slab, keg)))
5762 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5763 item, zone, zone->uz_name, slab, freei);
5767 * Verifies freed addresses. Checks for alignment, valid slab membership
5768 * and duplicate frees.
5772 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5778 slab = uma_dbg_getslab(zone, item);
5780 panic("uma: Freed item %p did not belong to zone %s",
5781 item, zone->uz_name);
5784 freei = slab_item_index(slab, keg, item);
5786 if (freei >= keg->uk_ipers)
5787 panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5788 item, zone, zone->uz_name, slab, freei);
5790 if (slab_item(slab, keg, freei) != item)
5791 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5792 item, zone, zone->uz_name, slab, freei);
5794 if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5795 slab_dbg_bits(slab, keg)))
5796 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5797 item, zone, zone->uz_name, slab, freei);
5799 #endif /* INVARIANTS */
5803 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5804 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5809 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5810 *allocs = counter_u64_fetch(z->uz_allocs);
5811 frees = counter_u64_fetch(z->uz_frees);
5812 *sleeps = z->uz_sleeps;
5816 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5818 for (i = 0; i < vm_ndomains; i++) {
5819 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5820 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5821 (LIST_FIRST(&kz->uk_zones) != z)))
5822 *cachefree += kz->uk_domain[i].ud_free_items;
5824 *used = *allocs - frees;
5825 return (((int64_t)*used + *cachefree) * kz->uk_size);
5828 DB_SHOW_COMMAND_FLAGS(uma, db_show_uma, DB_CMD_MEMSAFE)
5830 const char *fmt_hdr, *fmt_entry;
5833 uint64_t allocs, used, sleeps, xdomain;
5835 /* variables for sorting */
5837 uma_zone_t cur_zone, last_zone;
5838 int64_t cur_size, last_size, size;
5841 /* /i option produces machine-parseable CSV output */
5842 if (modif[0] == 'i') {
5843 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5844 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5846 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5847 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5850 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5851 "Sleeps", "Bucket", "Total Mem", "XFree");
5853 /* Sort the zones with largest size first. */
5855 last_size = INT64_MAX;
5860 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5861 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5863 * In the case of size ties, print out zones
5864 * in the order they are encountered. That is,
5865 * when we encounter the most recently output
5866 * zone, we have already printed all preceding
5867 * ties, and we must print all following ties.
5869 if (z == last_zone) {
5873 size = get_uma_stats(kz, z, &allocs, &used,
5874 &sleeps, &cachefree, &xdomain);
5875 if (size > cur_size && size < last_size + ties)
5883 if (cur_zone == NULL)
5886 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5887 &sleeps, &cachefree, &xdomain);
5888 db_printf(fmt_entry, cur_zone->uz_name,
5889 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5890 (uintmax_t)allocs, (uintmax_t)sleeps,
5891 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5896 last_zone = cur_zone;
5897 last_size = cur_size;
5901 DB_SHOW_COMMAND_FLAGS(umacache, db_show_umacache, DB_CMD_MEMSAFE)
5904 uint64_t allocs, frees;
5908 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5909 "Requests", "Bucket");
5910 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5911 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5912 for (i = 0; i < vm_ndomains; i++)
5913 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5914 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5915 z->uz_name, (uintmax_t)z->uz_size,
5916 (intmax_t)(allocs - frees), cachefree,
5917 (uintmax_t)allocs, z->uz_bucket_size);