2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2002-2019 Jeffrey Roberson <jeff@FreeBSD.org>
5 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
6 * Copyright (c) 2004-2006 Robert N. M. Watson
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice unmodified, this list of conditions, and the following
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
20 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
21 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
22 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
24 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
28 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32 * uma_core.c Implementation of the Universal Memory allocator
34 * This allocator is intended to replace the multitude of similar object caches
35 * in the standard FreeBSD kernel. The intent is to be flexible as well as
36 * efficient. A primary design goal is to return unused memory to the rest of
37 * the system. This will make the system as a whole more flexible due to the
38 * ability to move memory to subsystems which most need it instead of leaving
39 * pools of reserved memory unused.
41 * The basic ideas stem from similar slab/zone based allocators whose algorithms
48 * - Improve memory usage for large allocations
49 * - Investigate cache size adjustments
52 #include <sys/cdefs.h>
53 __FBSDID("$FreeBSD$");
56 #include "opt_param.h"
59 #include <sys/param.h>
60 #include <sys/systm.h>
61 #include <sys/bitset.h>
62 #include <sys/domainset.h>
63 #include <sys/eventhandler.h>
64 #include <sys/kernel.h>
65 #include <sys/types.h>
66 #include <sys/limits.h>
67 #include <sys/queue.h>
68 #include <sys/malloc.h>
71 #include <sys/sysctl.h>
72 #include <sys/mutex.h>
74 #include <sys/random.h>
75 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
81 #include <sys/taskqueue.h>
82 #include <sys/vmmeter.h>
85 #include <vm/vm_domainset.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_param.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>
96 #include <vm/uma_int.h>
97 #include <vm/uma_dbg.h>
101 #ifdef DEBUG_MEMGUARD
102 #include <vm/memguard.h>
105 #include <machine/md_var.h>
108 #define UMA_ALWAYS_CTORDTOR 1
110 #define UMA_ALWAYS_CTORDTOR 0
114 * This is the zone and keg from which all zones are spawned.
116 static uma_zone_t kegs;
117 static uma_zone_t zones;
120 * On INVARIANTS builds, the slab contains a second bitset of the same size,
121 * "dbg_bits", which is laid out immediately after us_free.
124 #define SLAB_BITSETS 2
126 #define SLAB_BITSETS 1
130 * These are the two zones from which all offpage uma_slab_ts are allocated.
132 * One zone is for slab headers that can represent a larger number of items,
133 * making the slabs themselves more efficient, and the other zone is for
134 * headers that are smaller and represent fewer items, making the headers more
137 #define SLABZONE_SIZE(setsize) \
138 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
139 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16)
140 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE
141 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE)
142 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE)
143 static uma_zone_t slabzones[2];
146 * The initial hash tables come out of this zone so they can be allocated
147 * prior to malloc coming up.
149 static uma_zone_t hashzone;
151 /* The boot-time adjusted value for cache line alignment. */
152 int uma_align_cache = 64 - 1;
154 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
155 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
158 * Are we allowed to allocate buckets?
160 static int bucketdisable = 1;
162 /* Linked list of all kegs in the system */
163 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
165 /* Linked list of all cache-only zones in the system */
166 static LIST_HEAD(,uma_zone) uma_cachezones =
167 LIST_HEAD_INITIALIZER(uma_cachezones);
169 /* This RW lock protects the keg list */
170 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
173 * First available virual address for boot time allocations.
175 static vm_offset_t bootstart;
176 static vm_offset_t bootmem;
178 static struct sx uma_reclaim_lock;
181 * kmem soft limit, initialized by uma_set_limit(). Ensure that early
182 * allocations don't trigger a wakeup of the reclaim thread.
184 unsigned long uma_kmem_limit = LONG_MAX;
185 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
186 "UMA kernel memory soft limit");
187 unsigned long uma_kmem_total;
188 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
189 "UMA kernel memory usage");
191 /* Is the VM done starting up? */
197 } booted = BOOT_COLD;
200 * This is the handle used to schedule events that need to happen
201 * outside of the allocation fast path.
203 static struct callout uma_callout;
204 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
207 * This structure is passed as the zone ctor arg so that I don't have to create
208 * a special allocation function just for zones.
210 struct uma_zctor_args {
225 struct uma_kctor_args {
234 struct uma_bucket_zone {
236 const char *ubz_name;
237 int ubz_entries; /* Number of items it can hold. */
238 int ubz_maxsize; /* Maximum allocation size per-item. */
242 * Compute the actual number of bucket entries to pack them in power
243 * of two sizes for more efficient space utilization.
245 #define BUCKET_SIZE(n) \
246 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
248 #define BUCKET_MAX BUCKET_SIZE(256)
251 struct uma_bucket_zone bucket_zones[] = {
252 /* Literal bucket sizes. */
253 { NULL, "2 Bucket", 2, 4096 },
254 { NULL, "4 Bucket", 4, 3072 },
255 { NULL, "8 Bucket", 8, 2048 },
256 { NULL, "16 Bucket", 16, 1024 },
257 /* Rounded down power of 2 sizes for efficiency. */
258 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
259 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
260 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
261 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
266 * Flags and enumerations to be passed to internal functions.
270 SKIP_CNT = 0x00000001,
271 SKIP_DTOR = 0x00010000,
272 SKIP_FINI = 0x00020000,
277 void uma_startup1(vm_offset_t);
278 void uma_startup2(void);
280 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
281 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
282 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
283 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
284 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
285 static void page_free(void *, vm_size_t, uint8_t);
286 static void pcpu_page_free(void *, vm_size_t, uint8_t);
287 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
288 static void cache_drain(uma_zone_t);
289 static void bucket_drain(uma_zone_t, uma_bucket_t);
290 static void bucket_cache_reclaim(uma_zone_t zone, bool);
291 static int keg_ctor(void *, int, void *, int);
292 static void keg_dtor(void *, int, void *);
293 static int zone_ctor(void *, int, void *, int);
294 static void zone_dtor(void *, int, void *);
295 static inline void item_dtor(uma_zone_t zone, void *item, int size,
296 void *udata, enum zfreeskip skip);
297 static int zero_init(void *, int, int);
298 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
299 int itemdomain, bool ws);
300 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
301 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
302 static void zone_timeout(uma_zone_t zone, void *);
303 static int hash_alloc(struct uma_hash *, u_int);
304 static int hash_expand(struct uma_hash *, struct uma_hash *);
305 static void hash_free(struct uma_hash *hash);
306 static void uma_timeout(void *);
307 static void uma_startup3(void);
308 static void uma_shutdown(void);
309 static void *zone_alloc_item(uma_zone_t, void *, int, int);
310 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
311 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
312 static void zone_free_limit(uma_zone_t zone, int count);
313 static void bucket_enable(void);
314 static void bucket_init(void);
315 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
316 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
317 static void bucket_zone_drain(void);
318 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
319 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
320 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
321 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
322 uma_fini fini, int align, uint32_t flags);
323 static int zone_import(void *, void **, int, int, int);
324 static void zone_release(void *, void **, int);
325 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
326 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
328 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
329 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
330 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
331 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
332 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
333 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
334 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
336 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
338 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
339 "Memory allocation debugging");
342 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
343 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
345 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
346 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
347 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
348 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
350 static u_int dbg_divisor = 1;
351 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
352 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
353 "Debug & thrash every this item in memory allocator");
355 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
356 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
357 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
358 &uma_dbg_cnt, "memory items debugged");
359 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
360 &uma_skip_cnt, "memory items skipped, not debugged");
363 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
365 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
366 "Universal Memory Allocator");
368 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
369 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
371 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
372 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
374 static int zone_warnings = 1;
375 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
376 "Warn when UMA zones becomes full");
378 static int multipage_slabs = 1;
379 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
380 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
381 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
382 "UMA may choose larger slab sizes for better efficiency");
385 * Select the slab zone for an offpage slab with the given maximum item count.
387 static inline uma_zone_t
391 return (slabzones[ipers > SLABZONE0_SETSIZE]);
395 * This routine checks to see whether or not it's safe to enable buckets.
401 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
402 bucketdisable = vm_page_count_min();
406 * Initialize bucket_zones, the array of zones of buckets of various sizes.
408 * For each zone, calculate the memory required for each bucket, consisting
409 * of the header and an array of pointers.
414 struct uma_bucket_zone *ubz;
417 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
418 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
419 size += sizeof(void *) * ubz->ubz_entries;
420 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
421 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
422 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
423 UMA_ZONE_FIRSTTOUCH);
428 * Given a desired number of entries for a bucket, return the zone from which
429 * to allocate the bucket.
431 static struct uma_bucket_zone *
432 bucket_zone_lookup(int entries)
434 struct uma_bucket_zone *ubz;
436 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
437 if (ubz->ubz_entries >= entries)
443 static struct uma_bucket_zone *
444 bucket_zone_max(uma_zone_t zone, int nitems)
446 struct uma_bucket_zone *ubz;
450 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
451 /* Count the cross-domain bucket. */
454 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
455 if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems)
457 if (ubz == &bucket_zones[0])
465 bucket_select(int size)
467 struct uma_bucket_zone *ubz;
469 ubz = &bucket_zones[0];
470 if (size > ubz->ubz_maxsize)
471 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
473 for (; ubz->ubz_entries != 0; ubz++)
474 if (ubz->ubz_maxsize < size)
477 return (ubz->ubz_entries);
481 bucket_alloc(uma_zone_t zone, void *udata, int flags)
483 struct uma_bucket_zone *ubz;
487 * Don't allocate buckets early in boot.
489 if (__predict_false(booted < BOOT_KVA))
493 * To limit bucket recursion we store the original zone flags
494 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
495 * NOVM flag to persist even through deep recursions. We also
496 * store ZFLAG_BUCKET once we have recursed attempting to allocate
497 * a bucket for a bucket zone so we do not allow infinite bucket
498 * recursion. This cookie will even persist to frees of unused
499 * buckets via the allocation path or bucket allocations in the
502 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
503 udata = (void *)(uintptr_t)zone->uz_flags;
505 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
507 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
509 if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
511 ubz = bucket_zone_lookup(zone->uz_bucket_size);
512 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
514 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
517 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
520 bucket->ub_entries = ubz->ubz_entries;
521 bucket->ub_seq = SMR_SEQ_INVALID;
522 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
523 zone->uz_name, zone, bucket);
530 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
532 struct uma_bucket_zone *ubz;
534 if (bucket->ub_cnt != 0)
535 bucket_drain(zone, bucket);
537 KASSERT(bucket->ub_cnt == 0,
538 ("bucket_free: Freeing a non free bucket."));
539 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
540 ("bucket_free: Freeing an SMR bucket."));
541 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
542 udata = (void *)(uintptr_t)zone->uz_flags;
543 ubz = bucket_zone_lookup(bucket->ub_entries);
544 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
548 bucket_zone_drain(void)
550 struct uma_bucket_zone *ubz;
552 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
553 uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN);
557 * Acquire the domain lock and record contention.
559 static uma_zone_domain_t
560 zone_domain_lock(uma_zone_t zone, int domain)
562 uma_zone_domain_t zdom;
565 zdom = ZDOM_GET(zone, domain);
567 if (ZDOM_OWNED(zdom))
570 /* This is unsynchronized. The counter does not need to be precise. */
571 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
572 zone->uz_bucket_size++;
577 * Search for the domain with the least cached items and return it if it
578 * is out of balance with the preferred domain.
580 static __noinline int
581 zone_domain_lowest(uma_zone_t zone, int pref)
583 long least, nitems, prefitems;
587 prefitems = least = LONG_MAX;
589 for (i = 0; i < vm_ndomains; i++) {
590 nitems = ZDOM_GET(zone, i)->uzd_nitems;
591 if (nitems < least) {
598 if (prefitems < least * 2)
605 * Search for the domain with the most cached items and return it or the
606 * preferred domain if it has enough to proceed.
608 static __noinline int
609 zone_domain_highest(uma_zone_t zone, int pref)
615 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
620 for (i = 0; i < vm_ndomains; i++) {
621 nitems = ZDOM_GET(zone, i)->uzd_nitems;
632 * Safely subtract cnt from imax.
635 zone_domain_imax_sub(uma_zone_domain_t zdom, int cnt)
640 old = zdom->uzd_imax;
646 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, new) == 0);
650 * Set the maximum imax value.
653 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
657 old = zdom->uzd_imax;
661 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
665 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
666 * zone's caches. If a bucket is found the zone is not locked on return.
669 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
675 ZDOM_LOCK_ASSERT(zdom);
677 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
680 /* SMR Buckets can not be re-used until readers expire. */
681 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
682 bucket->ub_seq != SMR_SEQ_INVALID) {
683 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
685 bucket->ub_seq = SMR_SEQ_INVALID;
686 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
687 if (STAILQ_NEXT(bucket, ub_link) != NULL)
688 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
690 MPASS(zdom->uzd_nitems >= bucket->ub_cnt);
691 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
692 zdom->uzd_nitems -= bucket->ub_cnt;
695 * Shift the bounds of the current WSS interval to avoid
696 * perturbing the estimate.
699 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
700 zone_domain_imax_sub(zdom, bucket->ub_cnt);
701 } else if (zdom->uzd_imin > zdom->uzd_nitems)
702 zdom->uzd_imin = zdom->uzd_nitems;
706 for (i = 0; i < bucket->ub_cnt; i++)
707 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
714 * Insert a full bucket into the specified cache. The "ws" parameter indicates
715 * whether the bucket's contents should be counted as part of the zone's working
716 * set. The bucket may be freed if it exceeds the bucket limit.
719 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
722 uma_zone_domain_t zdom;
724 /* We don't cache empty buckets. This can happen after a reclaim. */
725 if (bucket->ub_cnt == 0)
727 zdom = zone_domain_lock(zone, domain);
730 * Conditionally set the maximum number of items.
732 zdom->uzd_nitems += bucket->ub_cnt;
733 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
735 zone_domain_imax_set(zdom, zdom->uzd_nitems);
736 if (STAILQ_EMPTY(&zdom->uzd_buckets))
737 zdom->uzd_seq = bucket->ub_seq;
738 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
742 zdom->uzd_nitems -= bucket->ub_cnt;
745 bucket_free(zone, bucket, udata);
748 /* Pops an item out of a per-cpu cache bucket. */
750 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
754 CRITICAL_ASSERT(curthread);
757 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
759 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
760 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
767 /* Pushes an item into a per-cpu cache bucket. */
769 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
772 CRITICAL_ASSERT(curthread);
773 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
774 ("uma_zfree: Freeing to non free bucket index."));
776 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
782 * Unload a UMA bucket from a per-cpu cache.
784 static inline uma_bucket_t
785 cache_bucket_unload(uma_cache_bucket_t bucket)
789 b = bucket->ucb_bucket;
791 MPASS(b->ub_entries == bucket->ucb_entries);
792 b->ub_cnt = bucket->ucb_cnt;
793 bucket->ucb_bucket = NULL;
794 bucket->ucb_entries = bucket->ucb_cnt = 0;
800 static inline uma_bucket_t
801 cache_bucket_unload_alloc(uma_cache_t cache)
804 return (cache_bucket_unload(&cache->uc_allocbucket));
807 static inline uma_bucket_t
808 cache_bucket_unload_free(uma_cache_t cache)
811 return (cache_bucket_unload(&cache->uc_freebucket));
814 static inline uma_bucket_t
815 cache_bucket_unload_cross(uma_cache_t cache)
818 return (cache_bucket_unload(&cache->uc_crossbucket));
822 * Load a bucket into a per-cpu cache bucket.
825 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
828 CRITICAL_ASSERT(curthread);
829 MPASS(bucket->ucb_bucket == NULL);
830 MPASS(b->ub_seq == SMR_SEQ_INVALID);
832 bucket->ucb_bucket = b;
833 bucket->ucb_cnt = b->ub_cnt;
834 bucket->ucb_entries = b->ub_entries;
838 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
841 cache_bucket_load(&cache->uc_allocbucket, b);
845 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
848 cache_bucket_load(&cache->uc_freebucket, b);
853 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
856 cache_bucket_load(&cache->uc_crossbucket, b);
861 * Copy and preserve ucb_spare.
864 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
867 b1->ucb_bucket = b2->ucb_bucket;
868 b1->ucb_entries = b2->ucb_entries;
869 b1->ucb_cnt = b2->ucb_cnt;
873 * Swap two cache buckets.
876 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
878 struct uma_cache_bucket b3;
880 CRITICAL_ASSERT(curthread);
882 cache_bucket_copy(&b3, b1);
883 cache_bucket_copy(b1, b2);
884 cache_bucket_copy(b2, &b3);
888 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
891 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
893 uma_zone_domain_t zdom;
897 * Avoid the lock if possible.
899 zdom = ZDOM_GET(zone, domain);
900 if (zdom->uzd_nitems == 0)
903 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
904 !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
908 * Check the zone's cache of buckets.
910 zdom = zone_domain_lock(zone, domain);
911 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
912 KASSERT(bucket->ub_cnt != 0,
913 ("cache_fetch_bucket: Returning an empty bucket."));
922 zone_log_warning(uma_zone_t zone)
924 static const struct timeval warninterval = { 300, 0 };
926 if (!zone_warnings || zone->uz_warning == NULL)
929 if (ratecheck(&zone->uz_ratecheck, &warninterval))
930 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
934 zone_maxaction(uma_zone_t zone)
937 if (zone->uz_maxaction.ta_func != NULL)
938 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
942 * Routine called by timeout which is used to fire off some time interval
943 * based calculations. (stats, hash size, etc.)
952 uma_timeout(void *unused)
955 zone_foreach(zone_timeout, NULL);
957 /* Reschedule this event */
958 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
962 * Update the working set size estimate for the zone's bucket cache.
963 * The constants chosen here are somewhat arbitrary. With an update period of
964 * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
968 zone_domain_update_wss(uma_zone_domain_t zdom)
973 MPASS(zdom->uzd_imax >= zdom->uzd_imin);
974 wss = zdom->uzd_imax - zdom->uzd_imin;
975 zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
976 zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
981 * Routine to perform timeout driven calculations. This expands the
982 * hashes and does per cpu statistics aggregation.
987 zone_timeout(uma_zone_t zone, void *unused)
992 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
998 * Hash zones are non-numa by definition so the first domain
999 * is the only one present.
1002 pages = keg->uk_domain[0].ud_pages;
1005 * Expand the keg hash table.
1007 * This is done if the number of slabs is larger than the hash size.
1008 * What I'm trying to do here is completely reduce collisions. This
1009 * may be a little aggressive. Should I allow for two collisions max?
1011 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1012 struct uma_hash newhash;
1013 struct uma_hash oldhash;
1017 * This is so involved because allocating and freeing
1018 * while the keg lock is held will lead to deadlock.
1019 * I have to do everything in stages and check for
1023 ret = hash_alloc(&newhash, 1 << fls(slabs));
1026 if (hash_expand(&keg->uk_hash, &newhash)) {
1027 oldhash = keg->uk_hash;
1028 keg->uk_hash = newhash;
1033 hash_free(&oldhash);
1040 for (int i = 0; i < vm_ndomains; i++)
1041 zone_domain_update_wss(ZDOM_GET(zone, i));
1045 * Allocate and zero fill the next sized hash table from the appropriate
1049 * hash A new hash structure with the old hash size in uh_hashsize
1052 * 1 on success and 0 on failure.
1055 hash_alloc(struct uma_hash *hash, u_int size)
1059 KASSERT(powerof2(size), ("hash size must be power of 2"));
1060 if (size > UMA_HASH_SIZE_INIT) {
1061 hash->uh_hashsize = size;
1062 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1063 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1065 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1066 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1067 UMA_ANYDOMAIN, M_WAITOK);
1068 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1070 if (hash->uh_slab_hash) {
1071 bzero(hash->uh_slab_hash, alloc);
1072 hash->uh_hashmask = hash->uh_hashsize - 1;
1080 * Expands the hash table for HASH zones. This is done from zone_timeout
1081 * to reduce collisions. This must not be done in the regular allocation
1082 * path, otherwise, we can recurse on the vm while allocating pages.
1085 * oldhash The hash you want to expand
1086 * newhash The hash structure for the new table
1094 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1096 uma_hash_slab_t slab;
1100 if (!newhash->uh_slab_hash)
1103 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1107 * I need to investigate hash algorithms for resizing without a
1111 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1112 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1113 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1114 LIST_REMOVE(slab, uhs_hlink);
1115 hval = UMA_HASH(newhash, slab->uhs_data);
1116 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1124 * Free the hash bucket to the appropriate backing store.
1127 * slab_hash The hash bucket we're freeing
1128 * hashsize The number of entries in that hash bucket
1134 hash_free(struct uma_hash *hash)
1136 if (hash->uh_slab_hash == NULL)
1138 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1139 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1141 free(hash->uh_slab_hash, M_UMAHASH);
1145 * Frees all outstanding items in a bucket
1148 * zone The zone to free to, must be unlocked.
1149 * bucket The free/alloc bucket with items.
1155 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1159 if (bucket->ub_cnt == 0)
1162 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1163 bucket->ub_seq != SMR_SEQ_INVALID) {
1164 smr_wait(zone->uz_smr, bucket->ub_seq);
1165 bucket->ub_seq = SMR_SEQ_INVALID;
1166 for (i = 0; i < bucket->ub_cnt; i++)
1167 item_dtor(zone, bucket->ub_bucket[i],
1168 zone->uz_size, NULL, SKIP_NONE);
1171 for (i = 0; i < bucket->ub_cnt; i++)
1172 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1173 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1174 if (zone->uz_max_items > 0)
1175 zone_free_limit(zone, bucket->ub_cnt);
1177 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1183 * Drains the per cpu caches for a zone.
1185 * NOTE: This may only be called while the zone is being torn down, and not
1186 * during normal operation. This is necessary in order that we do not have
1187 * to migrate CPUs to drain the per-CPU caches.
1190 * zone The zone to drain, must be unlocked.
1196 cache_drain(uma_zone_t zone)
1199 uma_bucket_t bucket;
1204 * XXX: It is safe to not lock the per-CPU caches, because we're
1205 * tearing down the zone anyway. I.e., there will be no further use
1206 * of the caches at this point.
1208 * XXX: It would good to be able to assert that the zone is being
1209 * torn down to prevent improper use of cache_drain().
1211 seq = SMR_SEQ_INVALID;
1212 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1213 seq = smr_advance(zone->uz_smr);
1215 cache = &zone->uz_cpu[cpu];
1216 bucket = cache_bucket_unload_alloc(cache);
1218 bucket_free(zone, bucket, NULL);
1219 bucket = cache_bucket_unload_free(cache);
1220 if (bucket != NULL) {
1221 bucket->ub_seq = seq;
1222 bucket_free(zone, bucket, NULL);
1224 bucket = cache_bucket_unload_cross(cache);
1225 if (bucket != NULL) {
1226 bucket->ub_seq = seq;
1227 bucket_free(zone, bucket, NULL);
1230 bucket_cache_reclaim(zone, true);
1234 cache_shrink(uma_zone_t zone, void *unused)
1237 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1240 zone->uz_bucket_size =
1241 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1245 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1248 uma_bucket_t b1, b2, b3;
1251 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1254 b1 = b2 = b3 = NULL;
1256 cache = &zone->uz_cpu[curcpu];
1257 domain = PCPU_GET(domain);
1258 b1 = cache_bucket_unload_alloc(cache);
1261 * Don't flush SMR zone buckets. This leaves the zone without a
1262 * bucket and forces every free to synchronize().
1264 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1265 b2 = cache_bucket_unload_free(cache);
1266 b3 = cache_bucket_unload_cross(cache);
1271 zone_free_bucket(zone, b1, NULL, domain, false);
1273 zone_free_bucket(zone, b2, NULL, domain, false);
1275 /* Adjust the domain so it goes to zone_free_cross. */
1276 domain = (domain + 1) % vm_ndomains;
1277 zone_free_bucket(zone, b3, NULL, domain, false);
1282 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1283 * This is an expensive call because it needs to bind to all CPUs
1284 * one by one and enter a critical section on each of them in order
1285 * to safely access their cache buckets.
1286 * Zone lock must not be held on call this function.
1289 pcpu_cache_drain_safe(uma_zone_t zone)
1294 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1297 cache_shrink(zone, NULL);
1299 zone_foreach(cache_shrink, NULL);
1302 thread_lock(curthread);
1303 sched_bind(curthread, cpu);
1304 thread_unlock(curthread);
1307 cache_drain_safe_cpu(zone, NULL);
1309 zone_foreach(cache_drain_safe_cpu, NULL);
1311 thread_lock(curthread);
1312 sched_unbind(curthread);
1313 thread_unlock(curthread);
1317 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1318 * requested a drain, otherwise the per-domain caches are trimmed to either
1319 * estimated working set size.
1322 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1324 uma_zone_domain_t zdom;
1325 uma_bucket_t bucket;
1330 * Shrink the zone bucket size to ensure that the per-CPU caches
1331 * don't grow too large.
1333 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1334 zone->uz_bucket_size--;
1336 for (i = 0; i < vm_ndomains; i++) {
1338 * The cross bucket is partially filled and not part of
1339 * the item count. Reclaim it individually here.
1341 zdom = ZDOM_GET(zone, i);
1342 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1343 ZONE_CROSS_LOCK(zone);
1344 bucket = zdom->uzd_cross;
1345 zdom->uzd_cross = NULL;
1346 ZONE_CROSS_UNLOCK(zone);
1348 bucket_free(zone, bucket, NULL);
1352 * If we were asked to drain the zone, we are done only once
1353 * this bucket cache is empty. Otherwise, we reclaim items in
1354 * excess of the zone's estimated working set size. If the
1355 * difference nitems - imin is larger than the WSS estimate,
1356 * then the estimate will grow at the end of this interval and
1357 * we ignore the historical average.
1360 target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1362 while (zdom->uzd_nitems > target) {
1363 bucket = zone_fetch_bucket(zone, zdom, true);
1366 bucket_free(zone, bucket, NULL);
1374 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1380 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1381 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1383 mem = slab_data(slab, keg);
1384 flags = slab->us_flags;
1386 if (keg->uk_fini != NULL) {
1387 for (i--; i > -1; i--)
1390 * trash_fini implies that dtor was trash_dtor. trash_fini
1391 * would check that memory hasn't been modified since free,
1392 * which executed trash_dtor.
1393 * That's why we need to run uma_dbg_kskip() check here,
1394 * albeit we don't make skip check for other init/fini
1397 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1398 keg->uk_fini != trash_fini)
1400 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1402 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1403 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1405 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1406 uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1410 * Frees pages from a keg back to the system. This is done on demand from
1411 * the pageout daemon.
1416 keg_drain(uma_keg_t keg)
1418 struct slabhead freeslabs;
1420 uma_slab_t slab, tmp;
1423 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1426 for (i = 0; i < vm_ndomains; i++) {
1427 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1428 keg->uk_name, keg, i, dom->ud_free_items);
1429 dom = &keg->uk_domain[i];
1430 LIST_INIT(&freeslabs);
1433 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1434 LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
1435 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1437 n = dom->ud_free_slabs;
1438 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1439 dom->ud_free_slabs = 0;
1440 dom->ud_free_items -= n * keg->uk_ipers;
1441 dom->ud_pages -= n * keg->uk_ppera;
1444 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1445 keg_free_slab(keg, slab, keg->uk_ipers);
1450 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1454 * Set draining to interlock with zone_dtor() so we can release our
1455 * locks as we go. Only dtor() should do a WAITOK call since it
1456 * is the only call that knows the structure will still be available
1460 while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1461 if (waitok == M_NOWAIT)
1463 msleep(zone, &ZDOM_GET(zone, 0)->uzd_lock, PVM, "zonedrain",
1466 zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1468 bucket_cache_reclaim(zone, drain);
1471 * The DRAINING flag protects us from being freed while
1472 * we're running. Normally the uma_rwlock would protect us but we
1473 * must be able to release and acquire the right lock for each keg.
1475 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1476 keg_drain(zone->uz_keg);
1478 zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1485 zone_drain(uma_zone_t zone, void *unused)
1488 zone_reclaim(zone, M_NOWAIT, true);
1492 zone_trim(uma_zone_t zone, void *unused)
1495 zone_reclaim(zone, M_NOWAIT, false);
1499 * Allocate a new slab for a keg and inserts it into the partial slab list.
1500 * The keg should be unlocked on entry. If the allocation succeeds it will
1501 * be locked on return.
1504 * flags Wait flags for the item initialization routine
1505 * aflags Wait flags for the slab allocation
1508 * The slab that was allocated or NULL if there is no memory and the
1509 * caller specified M_NOWAIT.
1512 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1523 KASSERT(domain >= 0 && domain < vm_ndomains,
1524 ("keg_alloc_slab: domain %d out of range", domain));
1526 allocf = keg->uk_allocf;
1529 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1530 uma_hash_slab_t hslab;
1531 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1535 slab = &hslab->uhs_slab;
1539 * This reproduces the old vm_zone behavior of zero filling pages the
1540 * first time they are added to a zone.
1542 * Malloced items are zeroed in uma_zalloc.
1545 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1550 if (keg->uk_flags & UMA_ZONE_NODUMP)
1553 /* zone is passed for legacy reasons. */
1554 size = keg->uk_ppera * PAGE_SIZE;
1555 mem = allocf(zone, size, domain, &sflags, aflags);
1557 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1558 zone_free_item(slabzone(keg->uk_ipers),
1559 slab_tohashslab(slab), NULL, SKIP_NONE);
1562 uma_total_inc(size);
1564 /* For HASH zones all pages go to the same uma_domain. */
1565 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1568 /* Point the slab into the allocated memory */
1569 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1570 slab = (uma_slab_t )(mem + keg->uk_pgoff);
1572 slab_tohashslab(slab)->uhs_data = mem;
1574 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1575 for (i = 0; i < keg->uk_ppera; i++)
1576 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1579 slab->us_freecount = keg->uk_ipers;
1580 slab->us_flags = sflags;
1581 slab->us_domain = domain;
1583 BIT_FILL(keg->uk_ipers, &slab->us_free);
1585 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1588 if (keg->uk_init != NULL) {
1589 for (i = 0; i < keg->uk_ipers; i++)
1590 if (keg->uk_init(slab_item(slab, keg, i),
1591 keg->uk_size, flags) != 0)
1593 if (i != keg->uk_ipers) {
1594 keg_free_slab(keg, slab, i);
1598 KEG_LOCK(keg, domain);
1600 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1601 slab, keg->uk_name, keg);
1603 if (keg->uk_flags & UMA_ZFLAG_HASH)
1604 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1607 * If we got a slab here it's safe to mark it partially used
1608 * and return. We assume that the caller is going to remove
1609 * at least one item.
1611 dom = &keg->uk_domain[domain];
1612 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1613 dom->ud_pages += keg->uk_ppera;
1614 dom->ud_free_items += keg->uk_ipers;
1623 * This function is intended to be used early on in place of page_alloc() so
1624 * that we may use the boot time page cache to satisfy allocations before
1628 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1637 pages = howmany(bytes, PAGE_SIZE);
1638 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1640 *pflag = UMA_SLAB_BOOT;
1641 m = vm_page_alloc_contig_domain(NULL, 0, domain,
1642 malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1643 (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1647 pa = VM_PAGE_TO_PHYS(m);
1648 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1649 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1650 defined(__riscv) || defined(__powerpc64__)
1651 if ((wait & M_NODUMP) == 0)
1655 /* Allocate KVA and indirectly advance bootmem. */
1656 mem = (void *)pmap_map(&bootmem, m->phys_addr,
1657 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1658 if ((wait & M_ZERO) != 0)
1659 bzero(mem, pages * PAGE_SIZE);
1665 startup_free(void *mem, vm_size_t bytes)
1670 va = (vm_offset_t)mem;
1671 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1672 pmap_remove(kernel_pmap, va, va + bytes);
1673 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1674 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1675 defined(__riscv) || defined(__powerpc64__)
1676 dump_drop_page(VM_PAGE_TO_PHYS(m));
1678 vm_page_unwire_noq(m);
1684 * Allocates a number of pages from the system
1687 * bytes The number of bytes requested
1688 * wait Shall we wait?
1691 * A pointer to the alloced memory or possibly
1692 * NULL if M_NOWAIT is set.
1695 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1698 void *p; /* Returned page */
1700 *pflag = UMA_SLAB_KERNEL;
1701 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1707 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1710 struct pglist alloctail;
1711 vm_offset_t addr, zkva;
1713 vm_page_t p, p_next;
1718 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1720 TAILQ_INIT(&alloctail);
1721 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1722 malloc2vm_flags(wait);
1723 *pflag = UMA_SLAB_KERNEL;
1724 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1725 if (CPU_ABSENT(cpu)) {
1726 p = vm_page_alloc(NULL, 0, flags);
1729 p = vm_page_alloc(NULL, 0, flags);
1731 pc = pcpu_find(cpu);
1732 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1735 p = vm_page_alloc_domain(NULL, 0,
1736 pc->pc_domain, flags);
1737 if (__predict_false(p == NULL))
1738 p = vm_page_alloc(NULL, 0, flags);
1741 if (__predict_false(p == NULL))
1743 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1745 if ((addr = kva_alloc(bytes)) == 0)
1748 TAILQ_FOREACH(p, &alloctail, listq) {
1749 pmap_qenter(zkva, &p, 1);
1752 return ((void*)addr);
1754 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1755 vm_page_unwire_noq(p);
1762 * Allocates a number of pages from within an object
1765 * bytes The number of bytes requested
1766 * wait Shall we wait?
1769 * A pointer to the alloced memory or possibly
1770 * NULL if M_NOWAIT is set.
1773 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1776 TAILQ_HEAD(, vm_page) alloctail;
1778 vm_offset_t retkva, zkva;
1779 vm_page_t p, p_next;
1782 TAILQ_INIT(&alloctail);
1785 npages = howmany(bytes, PAGE_SIZE);
1786 while (npages > 0) {
1787 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1788 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1789 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1793 * Since the page does not belong to an object, its
1796 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1801 * Page allocation failed, free intermediate pages and
1804 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1805 vm_page_unwire_noq(p);
1810 *flags = UMA_SLAB_PRIV;
1811 zkva = keg->uk_kva +
1812 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1814 TAILQ_FOREACH(p, &alloctail, listq) {
1815 pmap_qenter(zkva, &p, 1);
1819 return ((void *)retkva);
1823 * Allocate physically contiguous pages.
1826 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1830 *pflag = UMA_SLAB_KERNEL;
1831 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1832 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1836 * Frees a number of pages to the system
1839 * mem A pointer to the memory to be freed
1840 * size The size of the memory being freed
1841 * flags The original p->us_flags field
1847 page_free(void *mem, vm_size_t size, uint8_t flags)
1850 if ((flags & UMA_SLAB_BOOT) != 0) {
1851 startup_free(mem, size);
1855 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1856 ("UMA: page_free used with invalid flags %x", flags));
1858 kmem_free((vm_offset_t)mem, size);
1862 * Frees pcpu zone allocations
1865 * mem A pointer to the memory to be freed
1866 * size The size of the memory being freed
1867 * flags The original p->us_flags field
1873 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1875 vm_offset_t sva, curva;
1879 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1881 if ((flags & UMA_SLAB_BOOT) != 0) {
1882 startup_free(mem, size);
1886 sva = (vm_offset_t)mem;
1887 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1888 paddr = pmap_kextract(curva);
1889 m = PHYS_TO_VM_PAGE(paddr);
1890 vm_page_unwire_noq(m);
1893 pmap_qremove(sva, size >> PAGE_SHIFT);
1894 kva_free(sva, size);
1899 * Zero fill initializer
1901 * Arguments/Returns follow uma_init specifications
1904 zero_init(void *mem, int size, int flags)
1911 static struct noslabbits *
1912 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1915 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1920 * Actual size of embedded struct slab (!OFFPAGE).
1923 slab_sizeof(int nitems)
1927 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1928 return (roundup(s, UMA_ALIGN_PTR + 1));
1931 #define UMA_FIXPT_SHIFT 31
1932 #define UMA_FRAC_FIXPT(n, d) \
1933 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1934 #define UMA_FIXPT_PCT(f) \
1935 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1936 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
1937 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1940 * Compute the number of items that will fit in a slab. If hdr is true, the
1941 * item count may be limited to provide space in the slab for an inline slab
1942 * header. Otherwise, all slab space will be provided for item storage.
1945 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1950 /* The padding between items is not needed after the last item. */
1951 padpi = rsize - size;
1955 * Start with the maximum item count and remove items until
1956 * the slab header first alongside the allocatable memory.
1958 for (ipers = MIN(SLAB_MAX_SETSIZE,
1959 (slabsize + padpi - slab_sizeof(1)) / rsize);
1961 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1965 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1971 struct keg_layout_result {
1979 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
1980 struct keg_layout_result *kl)
1985 kl->slabsize = slabsize;
1987 /* Handle INTERNAL as inline with an extra page. */
1988 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
1989 kl->format &= ~UMA_ZFLAG_INTERNAL;
1990 kl->slabsize += PAGE_SIZE;
1993 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
1994 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
1996 /* Account for memory used by an offpage slab header. */
1997 total = kl->slabsize;
1998 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
1999 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2001 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2005 * Determine the format of a uma keg. This determines where the slab header
2006 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2009 * keg The zone we should initialize
2015 keg_layout(uma_keg_t keg)
2017 struct keg_layout_result kl = {}, kl_tmp;
2026 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2027 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2028 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2029 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2030 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2032 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2033 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2034 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2037 alignsize = keg->uk_align + 1;
2040 * Calculate the size of each allocation (rsize) according to
2041 * alignment. If the requested size is smaller than we have
2042 * allocation bits for we round it up.
2044 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2045 rsize = roundup2(rsize, alignsize);
2047 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2049 * We want one item to start on every align boundary in a page.
2050 * To do this we will span pages. We will also extend the item
2051 * by the size of align if it is an even multiple of align.
2052 * Otherwise, it would fall on the same boundary every time.
2054 if ((rsize & alignsize) == 0)
2056 slabsize = rsize * (PAGE_SIZE / alignsize);
2057 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2058 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2059 slabsize = round_page(slabsize);
2062 * Start with a slab size of as many pages as it takes to
2063 * represent a single item. We will try to fit as many
2064 * additional items into the slab as possible.
2066 slabsize = round_page(keg->uk_size);
2069 /* Build a list of all of the available formats for this keg. */
2072 /* Evaluate an inline slab layout. */
2073 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2076 /* TODO: vm_page-embedded slab. */
2079 * We can't do OFFPAGE if we're internal or if we've been
2080 * asked to not go to the VM for buckets. If we do this we
2081 * may end up going to the VM for slabs which we do not want
2082 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2083 * In those cases, evaluate a pseudo-format called INTERNAL
2084 * which has an inline slab header and one extra page to
2085 * guarantee that it fits.
2087 * Otherwise, see if using an OFFPAGE slab will improve our
2090 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2091 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2093 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2096 * Choose a slab size and format which satisfy the minimum efficiency.
2097 * Prefer the smallest slab size that meets the constraints.
2099 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2100 * for small items (up to PAGE_SIZE), the iteration increment is one
2101 * page; and for large items, the increment is one item.
2103 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2104 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2105 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2108 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2109 round_page(rsize * (i - 1) + keg->uk_size);
2111 for (j = 0; j < nfmt; j++) {
2112 /* Only if we have no viable format yet. */
2113 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2117 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2118 if (kl_tmp.eff <= kl.eff)
2123 CTR6(KTR_UMA, "keg %s layout: format %#x "
2124 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2125 keg->uk_name, kl.format, kl.ipers, rsize,
2126 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2128 /* Stop when we reach the minimum efficiency. */
2129 if (kl.eff >= UMA_MIN_EFF)
2133 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2134 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2135 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2139 pages = atop(kl.slabsize);
2140 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2141 pages *= mp_maxid + 1;
2143 keg->uk_rsize = rsize;
2144 keg->uk_ipers = kl.ipers;
2145 keg->uk_ppera = pages;
2146 keg->uk_flags |= kl.format;
2149 * How do we find the slab header if it is offpage or if not all item
2150 * start addresses are in the same page? We could solve the latter
2151 * case with vaddr alignment, but we don't.
2153 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2154 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2155 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2156 keg->uk_flags |= UMA_ZFLAG_HASH;
2158 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2161 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2162 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2164 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2165 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2166 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2167 keg->uk_ipers, pages));
2171 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2172 * the keg onto the global keg list.
2174 * Arguments/Returns follow uma_ctor specifications
2175 * udata Actually uma_kctor_args
2178 keg_ctor(void *mem, int size, void *udata, int flags)
2180 struct uma_kctor_args *arg = udata;
2181 uma_keg_t keg = mem;
2186 keg->uk_size = arg->size;
2187 keg->uk_init = arg->uminit;
2188 keg->uk_fini = arg->fini;
2189 keg->uk_align = arg->align;
2190 keg->uk_reserve = 0;
2191 keg->uk_flags = arg->flags;
2194 * We use a global round-robin policy by default. Zones with
2195 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2196 * case the iterator is never run.
2198 keg->uk_dr.dr_policy = DOMAINSET_RR();
2199 keg->uk_dr.dr_iter = 0;
2202 * The master zone is passed to us at keg-creation time.
2205 keg->uk_name = zone->uz_name;
2207 if (arg->flags & UMA_ZONE_ZINIT)
2208 keg->uk_init = zero_init;
2210 if (arg->flags & UMA_ZONE_MALLOC)
2211 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2214 keg->uk_flags &= ~UMA_ZONE_PCPU;
2220 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2221 * work on. Use round-robin for everything else.
2223 * Zones may override the default by specifying either.
2226 if ((keg->uk_flags &
2227 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2228 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2229 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2230 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2234 * If we haven't booted yet we need allocations to go through the
2235 * startup cache until the vm is ready.
2237 #ifdef UMA_MD_SMALL_ALLOC
2238 if (keg->uk_ppera == 1)
2239 keg->uk_allocf = uma_small_alloc;
2242 if (booted < BOOT_KVA)
2243 keg->uk_allocf = startup_alloc;
2244 else if (keg->uk_flags & UMA_ZONE_PCPU)
2245 keg->uk_allocf = pcpu_page_alloc;
2246 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2247 keg->uk_allocf = contig_alloc;
2249 keg->uk_allocf = page_alloc;
2250 #ifdef UMA_MD_SMALL_ALLOC
2251 if (keg->uk_ppera == 1)
2252 keg->uk_freef = uma_small_free;
2255 if (keg->uk_flags & UMA_ZONE_PCPU)
2256 keg->uk_freef = pcpu_page_free;
2258 keg->uk_freef = page_free;
2261 * Initialize keg's locks.
2263 for (i = 0; i < vm_ndomains; i++)
2264 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2267 * If we're putting the slab header in the actual page we need to
2268 * figure out where in each page it goes. See slab_sizeof
2271 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2274 shsize = slab_sizeof(keg->uk_ipers);
2275 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2277 * The only way the following is possible is if with our
2278 * UMA_ALIGN_PTR adjustments we are now bigger than
2279 * UMA_SLAB_SIZE. I haven't checked whether this is
2280 * mathematically possible for all cases, so we make
2283 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2284 ("zone %s ipers %d rsize %d size %d slab won't fit",
2285 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2288 if (keg->uk_flags & UMA_ZFLAG_HASH)
2289 hash_alloc(&keg->uk_hash, 0);
2291 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2293 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2295 rw_wlock(&uma_rwlock);
2296 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2297 rw_wunlock(&uma_rwlock);
2302 zone_kva_available(uma_zone_t zone, void *unused)
2306 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2310 if (keg->uk_allocf == startup_alloc) {
2311 /* Switch to the real allocator. */
2312 if (keg->uk_flags & UMA_ZONE_PCPU)
2313 keg->uk_allocf = pcpu_page_alloc;
2314 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2316 keg->uk_allocf = contig_alloc;
2318 keg->uk_allocf = page_alloc;
2323 zone_alloc_counters(uma_zone_t zone, void *unused)
2326 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2327 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2328 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2329 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2333 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2335 uma_zone_domain_t zdom;
2338 struct sysctl_oid *oid, *domainoid;
2339 int domains, i, cnt;
2340 static const char *nokeg = "cache zone";
2344 * Make a sysctl safe copy of the zone name by removing
2345 * any special characters and handling dups by appending
2348 if (zone->uz_namecnt != 0) {
2349 /* Count the number of decimal digits and '_' separator. */
2350 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2352 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2354 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2357 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2358 for (c = zone->uz_ctlname; *c != '\0'; c++)
2359 if (strchr("./\\ -", *c) != NULL)
2363 * Basic parameters at the root.
2365 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2366 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2368 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2369 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2370 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2371 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2372 zone, 0, sysctl_handle_uma_zone_flags, "A",
2373 "Allocator configuration flags");
2374 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2375 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2376 "Desired per-cpu cache size");
2377 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2378 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2379 "Maximum allowed per-cpu cache size");
2384 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2385 domains = vm_ndomains;
2388 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2389 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2391 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2392 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2393 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2394 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2395 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2396 "Real object size with alignment");
2397 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2398 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2399 "pages per-slab allocation");
2400 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2401 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2402 "items available per-slab");
2403 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2404 "align", CTLFLAG_RD, &keg->uk_align, 0,
2405 "item alignment mask");
2406 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2407 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2408 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2409 "Slab utilization (100 - internal fragmentation %)");
2410 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2411 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2412 for (i = 0; i < domains; i++) {
2413 dom = &keg->uk_domain[i];
2414 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2415 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2416 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2417 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2418 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2419 "Total pages currently allocated from VM");
2420 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2421 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2422 "items free in the slab layer");
2425 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2426 "name", CTLFLAG_RD, nokeg, "Keg name");
2429 * Information about zone limits.
2431 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2432 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2433 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2434 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2435 zone, 0, sysctl_handle_uma_zone_items, "QU",
2436 "current number of allocated items if limit is set");
2437 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2438 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2439 "Maximum number of cached items");
2440 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2441 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2442 "Number of threads sleeping at limit");
2443 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2444 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2445 "Total zone limit sleeps");
2446 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2447 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2448 "Maximum number of items in each domain's bucket cache");
2451 * Per-domain zone information.
2453 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2454 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2455 for (i = 0; i < domains; i++) {
2456 zdom = ZDOM_GET(zone, i);
2457 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2458 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2459 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2460 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2461 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2462 "number of items in this domain");
2463 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2464 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2465 "maximum item count in this period");
2466 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2467 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2468 "minimum item count in this period");
2469 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2470 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2471 "Working set size");
2475 * General statistics.
2477 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2478 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2479 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2480 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2481 zone, 1, sysctl_handle_uma_zone_cur, "I",
2482 "Current number of allocated items");
2483 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2484 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2485 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2486 "Total allocation calls");
2487 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2488 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2489 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2490 "Total free calls");
2491 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2492 "fails", CTLFLAG_RD, &zone->uz_fails,
2493 "Number of allocation failures");
2494 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2495 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2496 "Free calls from the wrong domain");
2499 struct uma_zone_count {
2505 zone_count(uma_zone_t zone, void *arg)
2507 struct uma_zone_count *cnt;
2511 * Some zones are rapidly created with identical names and
2512 * destroyed out of order. This can lead to gaps in the count.
2513 * Use one greater than the maximum observed for this name.
2515 if (strcmp(zone->uz_name, cnt->name) == 0)
2516 cnt->count = MAX(cnt->count,
2517 zone->uz_namecnt + 1);
2521 zone_update_caches(uma_zone_t zone)
2525 for (i = 0; i <= mp_maxid; i++) {
2526 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2527 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2532 * Zone header ctor. This initializes all fields, locks, etc.
2534 * Arguments/Returns follow uma_ctor specifications
2535 * udata Actually uma_zctor_args
2538 zone_ctor(void *mem, int size, void *udata, int flags)
2540 struct uma_zone_count cnt;
2541 struct uma_zctor_args *arg = udata;
2542 uma_zone_domain_t zdom;
2543 uma_zone_t zone = mem;
2549 zone->uz_name = arg->name;
2550 zone->uz_ctor = arg->ctor;
2551 zone->uz_dtor = arg->dtor;
2552 zone->uz_init = NULL;
2553 zone->uz_fini = NULL;
2554 zone->uz_sleeps = 0;
2555 zone->uz_bucket_size = 0;
2556 zone->uz_bucket_size_min = 0;
2557 zone->uz_bucket_size_max = BUCKET_MAX;
2558 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2559 zone->uz_warning = NULL;
2560 /* The domain structures follow the cpu structures. */
2561 zone->uz_bucket_max = ULONG_MAX;
2562 timevalclear(&zone->uz_ratecheck);
2564 /* Count the number of duplicate names. */
2565 cnt.name = arg->name;
2567 zone_foreach(zone_count, &cnt);
2568 zone->uz_namecnt = cnt.count;
2569 ZONE_CROSS_LOCK_INIT(zone);
2571 for (i = 0; i < vm_ndomains; i++) {
2572 zdom = ZDOM_GET(zone, i);
2573 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2574 STAILQ_INIT(&zdom->uzd_buckets);
2578 if (arg->uminit == trash_init && arg->fini == trash_fini)
2579 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2583 * This is a pure cache zone, no kegs.
2586 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2587 ("zone_ctor: Import specified for non-cache zone."));
2588 zone->uz_flags = arg->flags;
2589 zone->uz_size = arg->size;
2590 zone->uz_import = arg->import;
2591 zone->uz_release = arg->release;
2592 zone->uz_arg = arg->arg;
2595 * Cache zones are round-robin unless a policy is
2596 * specified because they may have incompatible
2599 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2600 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2602 rw_wlock(&uma_rwlock);
2603 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2604 rw_wunlock(&uma_rwlock);
2609 * Use the regular zone/keg/slab allocator.
2611 zone->uz_import = zone_import;
2612 zone->uz_release = zone_release;
2613 zone->uz_arg = zone;
2616 if (arg->flags & UMA_ZONE_SECONDARY) {
2617 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2618 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2619 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2620 zone->uz_init = arg->uminit;
2621 zone->uz_fini = arg->fini;
2622 zone->uz_flags |= UMA_ZONE_SECONDARY;
2623 rw_wlock(&uma_rwlock);
2625 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2626 if (LIST_NEXT(z, uz_link) == NULL) {
2627 LIST_INSERT_AFTER(z, zone, uz_link);
2632 rw_wunlock(&uma_rwlock);
2633 } else if (keg == NULL) {
2634 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2635 arg->align, arg->flags)) == NULL)
2638 struct uma_kctor_args karg;
2641 /* We should only be here from uma_startup() */
2642 karg.size = arg->size;
2643 karg.uminit = arg->uminit;
2644 karg.fini = arg->fini;
2645 karg.align = arg->align;
2646 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2648 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2654 /* Inherit properties from the keg. */
2656 zone->uz_size = keg->uk_size;
2657 zone->uz_flags |= (keg->uk_flags &
2658 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2661 if (__predict_true(booted >= BOOT_RUNNING)) {
2662 zone_alloc_counters(zone, NULL);
2663 zone_alloc_sysctl(zone, NULL);
2665 zone->uz_allocs = EARLY_COUNTER;
2666 zone->uz_frees = EARLY_COUNTER;
2667 zone->uz_fails = EARLY_COUNTER;
2670 /* Caller requests a private SMR context. */
2671 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2672 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2674 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2675 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2676 ("Invalid zone flag combination"));
2677 if (arg->flags & UMA_ZFLAG_INTERNAL)
2678 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2679 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2680 zone->uz_bucket_size = BUCKET_MAX;
2681 else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2682 zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2683 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2684 zone->uz_bucket_size = 0;
2686 zone->uz_bucket_size = bucket_select(zone->uz_size);
2687 zone->uz_bucket_size_min = zone->uz_bucket_size;
2688 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2689 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2690 zone_update_caches(zone);
2696 * Keg header dtor. This frees all data, destroys locks, frees the hash
2697 * table and removes the keg from the global list.
2699 * Arguments/Returns follow uma_dtor specifications
2703 keg_dtor(void *arg, int size, void *udata)
2706 uint32_t free, pages;
2709 keg = (uma_keg_t)arg;
2711 for (i = 0; i < vm_ndomains; i++) {
2712 free += keg->uk_domain[i].ud_free_items;
2713 pages += keg->uk_domain[i].ud_pages;
2714 KEG_LOCK_FINI(keg, i);
2717 printf("Freed UMA keg (%s) was not empty (%u items). "
2718 " Lost %u pages of memory.\n",
2719 keg->uk_name ? keg->uk_name : "",
2720 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2722 hash_free(&keg->uk_hash);
2728 * Arguments/Returns follow uma_dtor specifications
2732 zone_dtor(void *arg, int size, void *udata)
2738 zone = (uma_zone_t)arg;
2740 sysctl_remove_oid(zone->uz_oid, 1, 1);
2742 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2745 rw_wlock(&uma_rwlock);
2746 LIST_REMOVE(zone, uz_link);
2747 rw_wunlock(&uma_rwlock);
2748 zone_reclaim(zone, M_WAITOK, true);
2751 * We only destroy kegs from non secondary/non cache zones.
2753 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2755 rw_wlock(&uma_rwlock);
2756 LIST_REMOVE(keg, uk_link);
2757 rw_wunlock(&uma_rwlock);
2758 zone_free_item(kegs, keg, NULL, SKIP_NONE);
2760 counter_u64_free(zone->uz_allocs);
2761 counter_u64_free(zone->uz_frees);
2762 counter_u64_free(zone->uz_fails);
2763 counter_u64_free(zone->uz_xdomain);
2764 free(zone->uz_ctlname, M_UMA);
2765 for (i = 0; i < vm_ndomains; i++)
2766 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2767 ZONE_CROSS_LOCK_FINI(zone);
2771 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2776 LIST_FOREACH(keg, &uma_kegs, uk_link) {
2777 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2780 LIST_FOREACH(zone, &uma_cachezones, uz_link)
2785 * Traverses every zone in the system and calls a callback
2788 * zfunc A pointer to a function which accepts a zone
2795 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2798 rw_rlock(&uma_rwlock);
2799 zone_foreach_unlocked(zfunc, arg);
2800 rw_runlock(&uma_rwlock);
2804 * Initialize the kernel memory allocator. This is done after pages can be
2805 * allocated but before general KVA is available.
2808 uma_startup1(vm_offset_t virtual_avail)
2810 struct uma_zctor_args args;
2811 size_t ksize, zsize, size;
2812 uma_keg_t masterkeg;
2816 bootstart = bootmem = virtual_avail;
2818 rw_init(&uma_rwlock, "UMA lock");
2819 sx_init(&uma_reclaim_lock, "umareclaim");
2821 ksize = sizeof(struct uma_keg) +
2822 (sizeof(struct uma_domain) * vm_ndomains);
2823 ksize = roundup(ksize, UMA_SUPER_ALIGN);
2824 zsize = sizeof(struct uma_zone) +
2825 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2826 (sizeof(struct uma_zone_domain) * vm_ndomains);
2827 zsize = roundup(zsize, UMA_SUPER_ALIGN);
2829 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2830 size = (zsize * 2) + ksize;
2831 m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO);
2832 zones = (uma_zone_t)m;
2834 kegs = (uma_zone_t)m;
2836 masterkeg = (uma_keg_t)m;
2838 /* "manually" create the initial zone */
2839 memset(&args, 0, sizeof(args));
2840 args.name = "UMA Kegs";
2842 args.ctor = keg_ctor;
2843 args.dtor = keg_dtor;
2844 args.uminit = zero_init;
2846 args.keg = masterkeg;
2847 args.align = UMA_SUPER_ALIGN - 1;
2848 args.flags = UMA_ZFLAG_INTERNAL;
2849 zone_ctor(kegs, zsize, &args, M_WAITOK);
2851 args.name = "UMA Zones";
2853 args.ctor = zone_ctor;
2854 args.dtor = zone_dtor;
2855 args.uminit = zero_init;
2858 args.align = UMA_SUPER_ALIGN - 1;
2859 args.flags = UMA_ZFLAG_INTERNAL;
2860 zone_ctor(zones, zsize, &args, M_WAITOK);
2862 /* Now make zones for slab headers */
2863 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2864 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2865 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2866 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2868 hashzone = uma_zcreate("UMA Hash",
2869 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2870 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2876 #ifndef UMA_MD_SMALL_ALLOC
2877 extern void vm_radix_reserve_kva(void);
2881 * Advertise the availability of normal kva allocations and switch to
2882 * the default back-end allocator. Marks the KVA we consumed on startup
2883 * as used in the map.
2889 if (bootstart != bootmem) {
2890 vm_map_lock(kernel_map);
2891 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2892 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2893 vm_map_unlock(kernel_map);
2896 #ifndef UMA_MD_SMALL_ALLOC
2897 /* Set up radix zone to use noobj_alloc. */
2898 vm_radix_reserve_kva();
2902 zone_foreach_unlocked(zone_kva_available, NULL);
2907 * Finish our initialization steps.
2914 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2915 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2916 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2918 zone_foreach_unlocked(zone_alloc_counters, NULL);
2919 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2920 callout_init(&uma_callout, 1);
2921 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2922 booted = BOOT_RUNNING;
2924 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2925 EVENTHANDLER_PRI_FIRST);
2932 booted = BOOT_SHUTDOWN;
2936 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2937 int align, uint32_t flags)
2939 struct uma_kctor_args args;
2942 args.uminit = uminit;
2944 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2947 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2950 /* Public functions */
2953 uma_set_align(int align)
2956 if (align != UMA_ALIGN_CACHE)
2957 uma_align_cache = align;
2962 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2963 uma_init uminit, uma_fini fini, int align, uint32_t flags)
2966 struct uma_zctor_args args;
2969 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2972 /* This stuff is essential for the zone ctor */
2973 memset(&args, 0, sizeof(args));
2978 args.uminit = uminit;
2982 * Inject procedures which check for memory use after free if we are
2983 * allowed to scramble the memory while it is not allocated. This
2984 * requires that: UMA is actually able to access the memory, no init
2985 * or fini procedures, no dependency on the initial value of the
2986 * memory, and no (legitimate) use of the memory after free. Note,
2987 * the ctor and dtor do not need to be empty.
2989 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
2990 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
2991 args.uminit = trash_init;
2992 args.fini = trash_fini;
2999 sx_slock(&uma_reclaim_lock);
3000 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3001 sx_sunlock(&uma_reclaim_lock);
3008 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3009 uma_init zinit, uma_fini zfini, uma_zone_t master)
3011 struct uma_zctor_args args;
3015 keg = master->uz_keg;
3016 memset(&args, 0, sizeof(args));
3018 args.size = keg->uk_size;
3021 args.uminit = zinit;
3023 args.align = keg->uk_align;
3024 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3027 sx_slock(&uma_reclaim_lock);
3028 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3029 sx_sunlock(&uma_reclaim_lock);
3036 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3037 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3038 void *arg, int flags)
3040 struct uma_zctor_args args;
3042 memset(&args, 0, sizeof(args));
3047 args.uminit = zinit;
3049 args.import = zimport;
3050 args.release = zrelease;
3053 args.flags = flags | UMA_ZFLAG_CACHE;
3055 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3060 uma_zdestroy(uma_zone_t zone)
3064 * Large slabs are expensive to reclaim, so don't bother doing
3065 * unnecessary work if we're shutting down.
3067 if (booted == BOOT_SHUTDOWN &&
3068 zone->uz_fini == NULL && zone->uz_release == zone_release)
3070 sx_slock(&uma_reclaim_lock);
3071 zone_free_item(zones, zone, NULL, SKIP_NONE);
3072 sx_sunlock(&uma_reclaim_lock);
3076 uma_zwait(uma_zone_t zone)
3079 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3080 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3081 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3082 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3084 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3088 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3090 void *item, *pcpu_item;
3094 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3096 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3099 pcpu_item = zpcpu_base_to_offset(item);
3100 if (flags & M_ZERO) {
3102 for (i = 0; i <= mp_maxid; i++)
3103 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3105 bzero(item, zone->uz_size);
3112 * A stub while both regular and pcpu cases are identical.
3115 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3120 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3122 item = zpcpu_offset_to_base(pcpu_item);
3123 uma_zfree_arg(zone, item, udata);
3126 static inline void *
3127 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3133 skipdbg = uma_dbg_zskip(zone, item);
3134 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3135 zone->uz_ctor != trash_ctor)
3136 trash_ctor(item, size, udata, flags);
3138 /* Check flags before loading ctor pointer. */
3139 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3140 __predict_false(zone->uz_ctor != NULL) &&
3141 zone->uz_ctor(item, size, udata, flags) != 0) {
3142 counter_u64_add(zone->uz_fails, 1);
3143 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3148 uma_dbg_alloc(zone, NULL, item);
3150 if (__predict_false(flags & M_ZERO))
3151 return (memset(item, 0, size));
3157 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3158 enum zfreeskip skip)
3163 skipdbg = uma_dbg_zskip(zone, item);
3164 if (skip == SKIP_NONE && !skipdbg) {
3165 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3166 uma_dbg_free(zone, udata, item);
3168 uma_dbg_free(zone, NULL, item);
3171 if (__predict_true(skip < SKIP_DTOR)) {
3172 if (zone->uz_dtor != NULL)
3173 zone->uz_dtor(item, size, udata);
3175 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3176 zone->uz_dtor != trash_dtor)
3177 trash_dtor(item, size, udata);
3182 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3183 #define UMA_ZALLOC_DEBUG
3185 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3191 if (flags & M_WAITOK) {
3192 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3193 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3198 KASSERT((flags & M_EXEC) == 0,
3199 ("uma_zalloc_debug: called with M_EXEC"));
3200 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3201 ("uma_zalloc_debug: called within spinlock or critical section"));
3202 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3203 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3206 #ifdef DEBUG_MEMGUARD
3207 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3209 item = memguard_alloc(zone->uz_size, flags);
3211 error = EJUSTRETURN;
3212 if (zone->uz_init != NULL &&
3213 zone->uz_init(item, zone->uz_size, flags) != 0) {
3217 if (zone->uz_ctor != NULL &&
3218 zone->uz_ctor(item, zone->uz_size, udata,
3220 counter_u64_add(zone->uz_fails, 1);
3221 zone->uz_fini(item, zone->uz_size);
3228 /* This is unfortunate but should not be fatal. */
3235 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3237 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3238 ("uma_zfree_debug: called with spinlock or critical section held"));
3240 #ifdef DEBUG_MEMGUARD
3241 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3242 if (zone->uz_dtor != NULL)
3243 zone->uz_dtor(item, zone->uz_size, udata);
3244 if (zone->uz_fini != NULL)
3245 zone->uz_fini(item, zone->uz_size);
3246 memguard_free(item);
3247 return (EJUSTRETURN);
3254 static inline void *
3255 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3256 void *udata, int flags)
3261 item = cache_bucket_pop(cache, bucket);
3262 size = cache_uz_size(cache);
3263 uz_flags = cache_uz_flags(cache);
3265 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3268 static __noinline void *
3269 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3271 uma_cache_bucket_t bucket;
3274 while (cache_alloc(zone, cache, udata, flags)) {
3275 cache = &zone->uz_cpu[curcpu];
3276 bucket = &cache->uc_allocbucket;
3277 if (__predict_false(bucket->ucb_cnt == 0))
3279 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3284 * We can not get a bucket so try to return a single item.
3286 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3287 domain = PCPU_GET(domain);
3289 domain = UMA_ANYDOMAIN;
3290 return (zone_alloc_item(zone, udata, domain, flags));
3295 uma_zalloc_smr(uma_zone_t zone, int flags)
3297 uma_cache_bucket_t bucket;
3300 #ifdef UMA_ZALLOC_DEBUG
3303 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3304 ("uma_zalloc_arg: called with non-SMR zone.\n"));
3305 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3310 cache = &zone->uz_cpu[curcpu];
3311 bucket = &cache->uc_allocbucket;
3312 if (__predict_false(bucket->ucb_cnt == 0))
3313 return (cache_alloc_retry(zone, cache, NULL, flags));
3314 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3319 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3321 uma_cache_bucket_t bucket;
3324 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3325 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3327 /* This is the fast path allocation */
3328 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3331 #ifdef UMA_ZALLOC_DEBUG
3334 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3335 ("uma_zalloc_arg: called with SMR zone.\n"));
3336 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3341 * If possible, allocate from the per-CPU cache. There are two
3342 * requirements for safe access to the per-CPU cache: (1) the thread
3343 * accessing the cache must not be preempted or yield during access,
3344 * and (2) the thread must not migrate CPUs without switching which
3345 * cache it accesses. We rely on a critical section to prevent
3346 * preemption and migration. We release the critical section in
3347 * order to acquire the zone mutex if we are unable to allocate from
3348 * the current cache; when we re-acquire the critical section, we
3349 * must detect and handle migration if it has occurred.
3352 cache = &zone->uz_cpu[curcpu];
3353 bucket = &cache->uc_allocbucket;
3354 if (__predict_false(bucket->ucb_cnt == 0))
3355 return (cache_alloc_retry(zone, cache, udata, flags));
3356 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3360 * Replenish an alloc bucket and possibly restore an old one. Called in
3361 * a critical section. Returns in a critical section.
3363 * A false return value indicates an allocation failure.
3364 * A true return value indicates success and the caller should retry.
3366 static __noinline bool
3367 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3369 uma_bucket_t bucket;
3373 CRITICAL_ASSERT(curthread);
3376 * If we have run out of items in our alloc bucket see
3377 * if we can switch with the free bucket.
3379 * SMR Zones can't re-use the free bucket until the sequence has
3382 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3383 cache->uc_freebucket.ucb_cnt != 0) {
3384 cache_bucket_swap(&cache->uc_freebucket,
3385 &cache->uc_allocbucket);
3390 * Discard any empty allocation bucket while we hold no locks.
3392 bucket = cache_bucket_unload_alloc(cache);
3395 if (bucket != NULL) {
3396 KASSERT(bucket->ub_cnt == 0,
3397 ("cache_alloc: Entered with non-empty alloc bucket."));
3398 bucket_free(zone, bucket, udata);
3401 /* Short-circuit for zones without buckets and low memory. */
3402 if (zone->uz_bucket_size == 0 || bucketdisable) {
3408 * Attempt to retrieve the item from the per-CPU cache has failed, so
3409 * we must go back to the zone. This requires the zdom lock, so we
3410 * must drop the critical section, then re-acquire it when we go back
3411 * to the cache. Since the critical section is released, we may be
3412 * preempted or migrate. As such, make sure not to maintain any
3413 * thread-local state specific to the cache from prior to releasing
3414 * the critical section.
3416 domain = PCPU_GET(domain);
3417 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0)
3418 domain = zone_domain_highest(zone, domain);
3419 bucket = cache_fetch_bucket(zone, cache, domain);
3420 if (bucket == NULL) {
3421 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3426 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3427 zone->uz_name, zone, bucket);
3428 if (bucket == NULL) {
3434 * See if we lost the race or were migrated. Cache the
3435 * initialized bucket to make this less likely or claim
3436 * the memory directly.
3439 cache = &zone->uz_cpu[curcpu];
3440 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3441 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3442 domain == PCPU_GET(domain))) {
3444 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3446 cache_bucket_load_alloc(cache, bucket);
3451 * We lost the race, release this bucket and start over.
3454 zone_put_bucket(zone, domain, bucket, udata, false);
3461 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3464 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3465 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3467 /* This is the fast path allocation */
3468 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3469 zone->uz_name, zone, domain, flags);
3471 if (flags & M_WAITOK) {
3472 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3473 "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3475 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3476 ("uma_zalloc_domain: called with spinlock or critical section held"));
3478 return (zone_alloc_item(zone, udata, domain, flags));
3482 * Find a slab with some space. Prefer slabs that are partially used over those
3483 * that are totally full. This helps to reduce fragmentation.
3485 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3489 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3495 KASSERT(domain >= 0 && domain < vm_ndomains,
3496 ("keg_first_slab: domain %d out of range", domain));
3497 KEG_LOCK_ASSERT(keg, domain);
3502 dom = &keg->uk_domain[domain];
3503 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3505 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3506 LIST_REMOVE(slab, us_link);
3507 dom->ud_free_slabs--;
3508 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3512 domain = (domain + 1) % vm_ndomains;
3513 } while (domain != start);
3519 * Fetch an existing slab from a free or partial list. Returns with the
3520 * keg domain lock held if a slab was found or unlocked if not.
3523 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3528 /* HASH has a single free list. */
3529 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3532 KEG_LOCK(keg, domain);
3533 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3534 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3535 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3536 KEG_UNLOCK(keg, domain);
3543 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3545 struct vm_domainset_iter di;
3552 * Use the keg's policy if upper layers haven't already specified a
3553 * domain (as happens with first-touch zones).
3555 * To avoid races we run the iterator with the keg lock held, but that
3556 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3557 * clear M_WAITOK and handle low memory conditions locally.
3559 rr = rdomain == UMA_ANYDOMAIN;
3561 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3562 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3570 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3575 * M_NOVM means don't ask at all!
3580 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3583 if (!rr && (flags & M_WAITOK) == 0)
3585 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3586 if ((flags & M_WAITOK) != 0) {
3587 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
3595 * We might not have been able to get a slab but another cpu
3596 * could have while we were unlocked. Check again before we
3599 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3606 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3612 KEG_LOCK_ASSERT(keg, slab->us_domain);
3614 dom = &keg->uk_domain[slab->us_domain];
3615 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3616 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3617 item = slab_item(slab, keg, freei);
3618 slab->us_freecount--;
3619 dom->ud_free_items--;
3622 * Move this slab to the full list. It must be on the partial list, so
3623 * we do not need to update the free slab count. In particular,
3624 * keg_fetch_slab() always returns slabs on the partial list.
3626 if (slab->us_freecount == 0) {
3627 LIST_REMOVE(slab, us_link);
3628 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3635 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3649 /* Try to keep the buckets totally full */
3650 for (i = 0; i < max; ) {
3651 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3654 stripe = howmany(max, vm_ndomains);
3656 dom = &keg->uk_domain[slab->us_domain];
3657 while (slab->us_freecount && i < max) {
3658 bucket[i++] = slab_alloc_item(keg, slab);
3659 if (dom->ud_free_items <= keg->uk_reserve)
3663 * If the zone is striped we pick a new slab for every
3664 * N allocations. Eliminating this conditional will
3665 * instead pick a new domain for each bucket rather
3666 * than stripe within each bucket. The current option
3667 * produces more fragmentation and requires more cpu
3668 * time but yields better distribution.
3670 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3671 vm_ndomains > 1 && --stripe == 0)
3675 KEG_UNLOCK(keg, slab->us_domain);
3676 /* Don't block if we allocated any successfully. */
3685 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3687 uint64_t old, new, total, max;
3690 * The hard case. We're going to sleep because there were existing
3691 * sleepers or because we ran out of items. This routine enforces
3692 * fairness by keeping fifo order.
3694 * First release our ill gotten gains and make some noise.
3697 zone_free_limit(zone, count);
3698 zone_log_warning(zone);
3699 zone_maxaction(zone);
3700 if (flags & M_NOWAIT)
3704 * We need to allocate an item or set ourself as a sleeper
3705 * while the sleepq lock is held to avoid wakeup races. This
3706 * is essentially a home rolled semaphore.
3708 sleepq_lock(&zone->uz_max_items);
3709 old = zone->uz_items;
3711 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3712 /* Cache the max since we will evaluate twice. */
3713 max = zone->uz_max_items;
3714 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3715 UZ_ITEMS_COUNT(old) >= max)
3716 new = old + UZ_ITEMS_SLEEPER;
3718 new = old + MIN(count, max - old);
3719 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3721 /* We may have successfully allocated under the sleepq lock. */
3722 if (UZ_ITEMS_SLEEPERS(new) == 0) {
3723 sleepq_release(&zone->uz_max_items);
3728 * This is in a different cacheline from uz_items so that we
3729 * don't constantly invalidate the fastpath cacheline when we
3730 * adjust item counts. This could be limited to toggling on
3733 atomic_add_32(&zone->uz_sleepers, 1);
3734 atomic_add_64(&zone->uz_sleeps, 1);
3737 * We have added ourselves as a sleeper. The sleepq lock
3738 * protects us from wakeup races. Sleep now and then retry.
3740 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3741 sleepq_wait(&zone->uz_max_items, PVM);
3744 * After wakeup, remove ourselves as a sleeper and try
3745 * again. We no longer have the sleepq lock for protection.
3747 * Subract ourselves as a sleeper while attempting to add
3750 atomic_subtract_32(&zone->uz_sleepers, 1);
3751 old = atomic_fetchadd_64(&zone->uz_items,
3752 -(UZ_ITEMS_SLEEPER - count));
3753 /* We're no longer a sleeper. */
3754 old -= UZ_ITEMS_SLEEPER;
3757 * If we're still at the limit, restart. Notably do not
3758 * block on other sleepers. Cache the max value to protect
3759 * against changes via sysctl.
3761 total = UZ_ITEMS_COUNT(old);
3762 max = zone->uz_max_items;
3765 /* Truncate if necessary, otherwise wake other sleepers. */
3766 if (total + count > max) {
3767 zone_free_limit(zone, total + count - max);
3768 count = max - total;
3769 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3770 wakeup_one(&zone->uz_max_items);
3777 * Allocate 'count' items from our max_items limit. Returns the number
3778 * available. If M_NOWAIT is not specified it will sleep until at least
3779 * one item can be allocated.
3782 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3787 max = zone->uz_max_items;
3791 * We expect normal allocations to succeed with a simple
3794 old = atomic_fetchadd_64(&zone->uz_items, count);
3795 if (__predict_true(old + count <= max))
3799 * If we had some items and no sleepers just return the
3800 * truncated value. We have to release the excess space
3801 * though because that may wake sleepers who weren't woken
3802 * because we were temporarily over the limit.
3805 zone_free_limit(zone, (old + count) - max);
3808 return (zone_alloc_limit_hard(zone, count, flags));
3812 * Free a number of items back to the limit.
3815 zone_free_limit(uma_zone_t zone, int count)
3822 * In the common case we either have no sleepers or
3823 * are still over the limit and can just return.
3825 old = atomic_fetchadd_64(&zone->uz_items, -count);
3826 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3827 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3831 * Moderate the rate of wakeups. Sleepers will continue
3832 * to generate wakeups if necessary.
3834 wakeup_one(&zone->uz_max_items);
3838 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3840 uma_bucket_t bucket;
3843 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3846 /* Avoid allocs targeting empty domains. */
3847 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3848 domain = UMA_ANYDOMAIN;
3849 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
3850 domain = UMA_ANYDOMAIN;
3852 if (zone->uz_max_items > 0)
3853 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3856 maxbucket = zone->uz_bucket_size;
3860 /* Don't wait for buckets, preserve caller's NOVM setting. */
3861 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3862 if (bucket == NULL) {
3867 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3868 MIN(maxbucket, bucket->ub_entries), domain, flags);
3871 * Initialize the memory if necessary.
3873 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3876 for (i = 0; i < bucket->ub_cnt; i++)
3877 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3881 * If we couldn't initialize the whole bucket, put the
3882 * rest back onto the freelist.
3884 if (i != bucket->ub_cnt) {
3885 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3886 bucket->ub_cnt - i);
3888 bzero(&bucket->ub_bucket[i],
3889 sizeof(void *) * (bucket->ub_cnt - i));
3895 cnt = bucket->ub_cnt;
3896 if (bucket->ub_cnt == 0) {
3897 bucket_free(zone, bucket, udata);
3898 counter_u64_add(zone->uz_fails, 1);
3902 if (zone->uz_max_items > 0 && cnt < maxbucket)
3903 zone_free_limit(zone, maxbucket - cnt);
3909 * Allocates a single item from a zone.
3912 * zone The zone to alloc for.
3913 * udata The data to be passed to the constructor.
3914 * domain The domain to allocate from or UMA_ANYDOMAIN.
3915 * flags M_WAITOK, M_NOWAIT, M_ZERO.
3918 * NULL if there is no memory and M_NOWAIT is set
3919 * An item if successful
3923 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
3927 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0)
3930 /* Avoid allocs targeting empty domains. */
3931 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3932 domain = UMA_ANYDOMAIN;
3934 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
3938 * We have to call both the zone's init (not the keg's init)
3939 * and the zone's ctor. This is because the item is going from
3940 * a keg slab directly to the user, and the user is expecting it
3941 * to be both zone-init'd as well as zone-ctor'd.
3943 if (zone->uz_init != NULL) {
3944 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
3945 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
3949 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
3954 counter_u64_add(zone->uz_allocs, 1);
3955 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
3956 zone->uz_name, zone);
3961 counter_u64_add(zone->uz_fails, 1);
3963 if (zone->uz_max_items > 0)
3964 zone_free_limit(zone, 1);
3965 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
3966 zone->uz_name, zone);
3973 uma_zfree_smr(uma_zone_t zone, void *item)
3976 uma_cache_bucket_t bucket;
3977 int itemdomain, uz_flags;
3979 #ifdef UMA_ZALLOC_DEBUG
3980 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3981 ("uma_zfree_smr: called with non-SMR zone.\n"));
3982 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
3983 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
3984 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
3987 cache = &zone->uz_cpu[curcpu];
3988 uz_flags = cache_uz_flags(cache);
3991 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
3992 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
3996 cache = &zone->uz_cpu[curcpu];
3997 /* SMR Zones must free to the free bucket. */
3998 bucket = &cache->uc_freebucket;
4000 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4001 PCPU_GET(domain) != itemdomain) {
4002 bucket = &cache->uc_crossbucket;
4005 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4006 cache_bucket_push(cache, bucket, item);
4010 } while (cache_free(zone, cache, NULL, item, itemdomain));
4014 * If nothing else caught this, we'll just do an internal free.
4016 zone_free_item(zone, item, NULL, SKIP_NONE);
4021 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4024 uma_cache_bucket_t bucket;
4025 int itemdomain, uz_flags;
4027 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4028 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4030 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4032 #ifdef UMA_ZALLOC_DEBUG
4033 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4034 ("uma_zfree_arg: called with SMR zone.\n"));
4035 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4038 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4043 * We are accessing the per-cpu cache without a critical section to
4044 * fetch size and flags. This is acceptable, if we are preempted we
4045 * will simply read another cpu's line.
4047 cache = &zone->uz_cpu[curcpu];
4048 uz_flags = cache_uz_flags(cache);
4049 if (UMA_ALWAYS_CTORDTOR ||
4050 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4051 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4054 * The race here is acceptable. If we miss it we'll just have to wait
4055 * a little longer for the limits to be reset.
4057 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4058 if (zone->uz_sleepers > 0)
4063 * If possible, free to the per-CPU cache. There are two
4064 * requirements for safe access to the per-CPU cache: (1) the thread
4065 * accessing the cache must not be preempted or yield during access,
4066 * and (2) the thread must not migrate CPUs without switching which
4067 * cache it accesses. We rely on a critical section to prevent
4068 * preemption and migration. We release the critical section in
4069 * order to acquire the zone mutex if we are unable to free to the
4070 * current cache; when we re-acquire the critical section, we must
4071 * detect and handle migration if it has occurred.
4075 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4076 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4080 cache = &zone->uz_cpu[curcpu];
4082 * Try to free into the allocbucket first to give LIFO
4083 * ordering for cache-hot datastructures. Spill over
4084 * into the freebucket if necessary. Alloc will swap
4085 * them if one runs dry.
4087 bucket = &cache->uc_allocbucket;
4089 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4090 PCPU_GET(domain) != itemdomain) {
4091 bucket = &cache->uc_crossbucket;
4094 if (bucket->ucb_cnt == bucket->ucb_entries &&
4095 cache->uc_freebucket.ucb_cnt <
4096 cache->uc_freebucket.ucb_entries)
4097 cache_bucket_swap(&cache->uc_freebucket,
4098 &cache->uc_allocbucket);
4099 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4100 cache_bucket_push(cache, bucket, item);
4104 } while (cache_free(zone, cache, udata, item, itemdomain));
4108 * If nothing else caught this, we'll just do an internal free.
4111 zone_free_item(zone, item, udata, SKIP_DTOR);
4116 * sort crossdomain free buckets to domain correct buckets and cache
4120 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4122 struct uma_bucketlist fullbuckets;
4123 uma_zone_domain_t zdom;
4130 "uma_zfree: zone %s(%p) draining cross bucket %p",
4131 zone->uz_name, zone, bucket);
4134 * It is possible for buckets to arrive here out of order so we fetch
4135 * the current smr seq rather than accepting the bucket's.
4137 seq = SMR_SEQ_INVALID;
4138 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4139 seq = smr_advance(zone->uz_smr);
4142 * To avoid having ndomain * ndomain buckets for sorting we have a
4143 * lock on the current crossfree bucket. A full matrix with
4144 * per-domain locking could be used if necessary.
4146 STAILQ_INIT(&fullbuckets);
4147 ZONE_CROSS_LOCK(zone);
4148 while (bucket->ub_cnt > 0) {
4149 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4150 domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4151 zdom = ZDOM_GET(zone, domain);
4152 if (zdom->uzd_cross == NULL) {
4153 zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
4154 if (zdom->uzd_cross == NULL)
4157 b = zdom->uzd_cross;
4158 b->ub_bucket[b->ub_cnt++] = item;
4160 if (b->ub_cnt == b->ub_entries) {
4161 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4162 zdom->uzd_cross = NULL;
4166 ZONE_CROSS_UNLOCK(zone);
4167 if (bucket->ub_cnt == 0)
4168 bucket->ub_seq = SMR_SEQ_INVALID;
4169 bucket_free(zone, bucket, udata);
4171 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4172 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4173 domain = _vm_phys_domain(pmap_kextract(
4174 (vm_offset_t)b->ub_bucket[0]));
4175 zone_put_bucket(zone, domain, b, udata, true);
4181 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4182 int itemdomain, bool ws)
4187 * Buckets coming from the wrong domain will be entirely for the
4188 * only other domain on two domain systems. In this case we can
4189 * simply cache them. Otherwise we need to sort them back to
4192 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4193 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4194 zone_free_cross(zone, bucket, udata);
4200 * Attempt to save the bucket in the zone's domain bucket cache.
4203 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4204 zone->uz_name, zone, bucket);
4205 /* ub_cnt is pointing to the last free item */
4206 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4207 itemdomain = zone_domain_lowest(zone, itemdomain);
4208 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4212 * Populate a free or cross bucket for the current cpu cache. Free any
4213 * existing full bucket either to the zone cache or back to the slab layer.
4215 * Enters and returns in a critical section. false return indicates that
4216 * we can not satisfy this free in the cache layer. true indicates that
4217 * the caller should retry.
4219 static __noinline bool
4220 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4223 uma_cache_bucket_t cbucket;
4224 uma_bucket_t newbucket, bucket;
4226 CRITICAL_ASSERT(curthread);
4228 if (zone->uz_bucket_size == 0)
4231 cache = &zone->uz_cpu[curcpu];
4235 * FIRSTTOUCH domains need to free to the correct zdom. When
4236 * enabled this is the zdom of the item. The bucket is the
4237 * cross bucket if the current domain and itemdomain do not match.
4239 cbucket = &cache->uc_freebucket;
4241 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4242 if (PCPU_GET(domain) != itemdomain) {
4243 cbucket = &cache->uc_crossbucket;
4244 if (cbucket->ucb_cnt != 0)
4245 counter_u64_add(zone->uz_xdomain,
4250 bucket = cache_bucket_unload(cbucket);
4251 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4252 ("cache_free: Entered with non-full free bucket."));
4254 /* We are no longer associated with this CPU. */
4258 * Don't let SMR zones operate without a free bucket. Force
4259 * a synchronize and re-use this one. We will only degrade
4260 * to a synchronize every bucket_size items rather than every
4261 * item if we fail to allocate a bucket.
4263 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4265 bucket->ub_seq = smr_advance(zone->uz_smr);
4266 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4267 if (newbucket == NULL && bucket != NULL) {
4268 bucket_drain(zone, bucket);
4272 } else if (!bucketdisable)
4273 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4276 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4279 if ((bucket = newbucket) == NULL)
4281 cache = &zone->uz_cpu[curcpu];
4284 * Check to see if we should be populating the cross bucket. If it
4285 * is already populated we will fall through and attempt to populate
4288 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4289 if (PCPU_GET(domain) != itemdomain &&
4290 cache->uc_crossbucket.ucb_bucket == NULL) {
4291 cache_bucket_load_cross(cache, bucket);
4297 * We may have lost the race to fill the bucket or switched CPUs.
4299 if (cache->uc_freebucket.ucb_bucket != NULL) {
4301 bucket_free(zone, bucket, udata);
4304 cache_bucket_load_free(cache, bucket);
4310 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
4313 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4314 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4316 CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone);
4318 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
4319 ("uma_zfree_domain: called with spinlock or critical section held"));
4321 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4324 zone_free_item(zone, item, udata, SKIP_NONE);
4328 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4335 KEG_LOCK_ASSERT(keg, slab->us_domain);
4337 /* Do we need to remove from any lists? */
4338 dom = &keg->uk_domain[slab->us_domain];
4339 if (slab->us_freecount + 1 == keg->uk_ipers) {
4340 LIST_REMOVE(slab, us_link);
4341 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4342 dom->ud_free_slabs++;
4343 } else if (slab->us_freecount == 0) {
4344 LIST_REMOVE(slab, us_link);
4345 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4348 /* Slab management. */
4349 freei = slab_item_index(slab, keg, item);
4350 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4351 slab->us_freecount++;
4353 /* Keg statistics. */
4354 dom->ud_free_items++;
4358 zone_release(void *arg, void **bucket, int cnt)
4371 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4372 lock = KEG_LOCK(keg, 0);
4373 for (i = 0; i < cnt; i++) {
4375 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4376 slab = vtoslab((vm_offset_t)item);
4378 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4379 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4380 slab = hash_sfind(&keg->uk_hash, mem);
4382 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4384 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4387 lock = KEG_LOCK(keg, slab->us_domain);
4389 slab_free_item(zone, slab, item);
4396 * Frees a single item to any zone.
4399 * zone The zone to free to
4400 * item The item we're freeing
4401 * udata User supplied data for the dtor
4402 * skip Skip dtors and finis
4404 static __noinline void
4405 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4409 * If a free is sent directly to an SMR zone we have to
4410 * synchronize immediately because the item can instantly
4411 * be reallocated. This should only happen in degenerate
4412 * cases when no memory is available for per-cpu caches.
4414 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4415 smr_synchronize(zone->uz_smr);
4417 item_dtor(zone, item, zone->uz_size, udata, skip);
4419 if (skip < SKIP_FINI && zone->uz_fini)
4420 zone->uz_fini(item, zone->uz_size);
4422 zone->uz_release(zone->uz_arg, &item, 1);
4424 if (skip & SKIP_CNT)
4427 counter_u64_add(zone->uz_frees, 1);
4429 if (zone->uz_max_items > 0)
4430 zone_free_limit(zone, 1);
4435 uma_zone_set_max(uma_zone_t zone, int nitems)
4437 struct uma_bucket_zone *ubz;
4441 * XXX This can misbehave if the zone has any allocations with
4442 * no limit and a limit is imposed. There is currently no
4443 * way to clear a limit.
4446 ubz = bucket_zone_max(zone, nitems);
4447 count = ubz != NULL ? ubz->ubz_entries : 0;
4448 zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4449 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4450 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4451 zone->uz_max_items = nitems;
4452 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4453 zone_update_caches(zone);
4454 /* We may need to wake waiters. */
4455 wakeup(&zone->uz_max_items);
4463 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4465 struct uma_bucket_zone *ubz;
4469 ubz = bucket_zone_max(zone, nitems);
4472 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4473 /* Count the cross-domain bucket. */
4475 nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4476 zone->uz_bucket_size_max = ubz->ubz_entries;
4478 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4480 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4481 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4482 zone->uz_bucket_max = nitems / vm_ndomains;
4488 uma_zone_get_max(uma_zone_t zone)
4492 nitems = atomic_load_64(&zone->uz_max_items);
4499 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4502 ZONE_ASSERT_COLD(zone);
4503 zone->uz_warning = warning;
4508 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4511 ZONE_ASSERT_COLD(zone);
4512 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4517 uma_zone_get_cur(uma_zone_t zone)
4523 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4524 nitems = counter_u64_fetch(zone->uz_allocs) -
4525 counter_u64_fetch(zone->uz_frees);
4527 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4528 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4530 return (nitems < 0 ? 0 : nitems);
4534 uma_zone_get_allocs(uma_zone_t zone)
4540 if (zone->uz_allocs != EARLY_COUNTER)
4541 nitems = counter_u64_fetch(zone->uz_allocs);
4543 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4549 uma_zone_get_frees(uma_zone_t zone)
4555 if (zone->uz_frees != EARLY_COUNTER)
4556 nitems = counter_u64_fetch(zone->uz_frees);
4558 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4564 /* Used only for KEG_ASSERT_COLD(). */
4566 uma_keg_get_allocs(uma_keg_t keg)
4572 LIST_FOREACH(z, &keg->uk_zones, uz_link)
4573 nitems += uma_zone_get_allocs(z);
4581 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4586 KEG_ASSERT_COLD(keg);
4587 keg->uk_init = uminit;
4592 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4597 KEG_ASSERT_COLD(keg);
4598 keg->uk_fini = fini;
4603 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4606 ZONE_ASSERT_COLD(zone);
4607 zone->uz_init = zinit;
4612 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4615 ZONE_ASSERT_COLD(zone);
4616 zone->uz_fini = zfini;
4621 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4626 KEG_ASSERT_COLD(keg);
4627 keg->uk_freef = freef;
4632 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4637 KEG_ASSERT_COLD(keg);
4638 keg->uk_allocf = allocf;
4643 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4646 ZONE_ASSERT_COLD(zone);
4648 KASSERT(smr != NULL, ("Got NULL smr"));
4649 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4650 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
4651 zone->uz_flags |= UMA_ZONE_SMR;
4653 zone_update_caches(zone);
4657 uma_zone_get_smr(uma_zone_t zone)
4660 return (zone->uz_smr);
4665 uma_zone_reserve(uma_zone_t zone, int items)
4670 KEG_ASSERT_COLD(keg);
4671 keg->uk_reserve = items;
4676 uma_zone_reserve_kva(uma_zone_t zone, int count)
4683 KEG_ASSERT_COLD(keg);
4684 ZONE_ASSERT_COLD(zone);
4686 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4688 #ifdef UMA_MD_SMALL_ALLOC
4689 if (keg->uk_ppera > 1) {
4693 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4699 MPASS(keg->uk_kva == 0);
4702 zone->uz_max_items = pages * keg->uk_ipers;
4703 #ifdef UMA_MD_SMALL_ALLOC
4704 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4706 keg->uk_allocf = noobj_alloc;
4708 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4709 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4710 zone_update_caches(zone);
4717 uma_prealloc(uma_zone_t zone, int items)
4719 struct vm_domainset_iter di;
4723 int aflags, domain, slabs;
4726 slabs = howmany(items, keg->uk_ipers);
4727 while (slabs-- > 0) {
4729 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4732 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4735 dom = &keg->uk_domain[slab->us_domain];
4737 * keg_alloc_slab() always returns a slab on the
4740 LIST_REMOVE(slab, us_link);
4741 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4743 dom->ud_free_slabs++;
4744 KEG_UNLOCK(keg, slab->us_domain);
4747 if (vm_domainset_iter_policy(&di, &domain) != 0)
4748 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
4754 * Returns a snapshot of memory consumption in bytes.
4757 uma_zone_memory(uma_zone_t zone)
4763 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
4764 for (i = 0; i < vm_ndomains; i++)
4765 sz += ZDOM_GET(zone, i)->uzd_nitems;
4766 return (sz * zone->uz_size);
4768 for (i = 0; i < vm_ndomains; i++)
4769 sz += zone->uz_keg->uk_domain[i].ud_pages;
4771 return (sz * PAGE_SIZE);
4776 uma_reclaim(int req)
4779 CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4780 sx_xlock(&uma_reclaim_lock);
4784 case UMA_RECLAIM_TRIM:
4785 zone_foreach(zone_trim, NULL);
4787 case UMA_RECLAIM_DRAIN:
4788 case UMA_RECLAIM_DRAIN_CPU:
4789 zone_foreach(zone_drain, NULL);
4790 if (req == UMA_RECLAIM_DRAIN_CPU) {
4791 pcpu_cache_drain_safe(NULL);
4792 zone_foreach(zone_drain, NULL);
4796 panic("unhandled reclamation request %d", req);
4800 * Some slabs may have been freed but this zone will be visited early
4801 * we visit again so that we can free pages that are empty once other
4802 * zones are drained. We have to do the same for buckets.
4804 zone_drain(slabzones[0], NULL);
4805 zone_drain(slabzones[1], NULL);
4806 bucket_zone_drain();
4807 sx_xunlock(&uma_reclaim_lock);
4810 static volatile int uma_reclaim_needed;
4813 uma_reclaim_wakeup(void)
4816 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4817 wakeup(uma_reclaim);
4821 uma_reclaim_worker(void *arg __unused)
4825 sx_xlock(&uma_reclaim_lock);
4826 while (atomic_load_int(&uma_reclaim_needed) == 0)
4827 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4829 sx_xunlock(&uma_reclaim_lock);
4830 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4831 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4832 atomic_store_int(&uma_reclaim_needed, 0);
4833 /* Don't fire more than once per-second. */
4834 pause("umarclslp", hz);
4840 uma_zone_reclaim(uma_zone_t zone, int req)
4844 case UMA_RECLAIM_TRIM:
4845 zone_trim(zone, NULL);
4847 case UMA_RECLAIM_DRAIN:
4848 zone_drain(zone, NULL);
4850 case UMA_RECLAIM_DRAIN_CPU:
4851 pcpu_cache_drain_safe(zone);
4852 zone_drain(zone, NULL);
4855 panic("unhandled reclamation request %d", req);
4861 uma_zone_exhausted(uma_zone_t zone)
4864 return (atomic_load_32(&zone->uz_sleepers) > 0);
4871 return (uma_kmem_limit);
4875 uma_set_limit(unsigned long limit)
4878 uma_kmem_limit = limit;
4885 return (atomic_load_long(&uma_kmem_total));
4892 return (uma_kmem_limit - uma_size());
4897 * Generate statistics across both the zone and its per-cpu cache's. Return
4898 * desired statistics if the pointer is non-NULL for that statistic.
4900 * Note: does not update the zone statistics, as it can't safely clear the
4901 * per-CPU cache statistic.
4905 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4906 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4909 uint64_t allocs, frees, sleeps, xdomain;
4912 allocs = frees = sleeps = xdomain = 0;
4915 cache = &z->uz_cpu[cpu];
4916 cachefree += cache->uc_allocbucket.ucb_cnt;
4917 cachefree += cache->uc_freebucket.ucb_cnt;
4918 xdomain += cache->uc_crossbucket.ucb_cnt;
4919 cachefree += cache->uc_crossbucket.ucb_cnt;
4920 allocs += cache->uc_allocs;
4921 frees += cache->uc_frees;
4923 allocs += counter_u64_fetch(z->uz_allocs);
4924 frees += counter_u64_fetch(z->uz_frees);
4925 xdomain += counter_u64_fetch(z->uz_xdomain);
4926 sleeps += z->uz_sleeps;
4927 if (cachefreep != NULL)
4928 *cachefreep = cachefree;
4929 if (allocsp != NULL)
4933 if (sleepsp != NULL)
4935 if (xdomainp != NULL)
4936 *xdomainp = xdomain;
4941 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
4948 rw_rlock(&uma_rwlock);
4949 LIST_FOREACH(kz, &uma_kegs, uk_link) {
4950 LIST_FOREACH(z, &kz->uk_zones, uz_link)
4953 LIST_FOREACH(z, &uma_cachezones, uz_link)
4956 rw_runlock(&uma_rwlock);
4957 return (sysctl_handle_int(oidp, &count, 0, req));
4961 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
4962 struct uma_percpu_stat *ups, bool internal)
4964 uma_zone_domain_t zdom;
4969 for (i = 0; i < vm_ndomains; i++) {
4970 zdom = ZDOM_GET(z, i);
4971 uth->uth_zone_free += zdom->uzd_nitems;
4973 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
4974 uth->uth_frees = counter_u64_fetch(z->uz_frees);
4975 uth->uth_fails = counter_u64_fetch(z->uz_fails);
4976 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
4977 uth->uth_sleeps = z->uz_sleeps;
4979 for (i = 0; i < mp_maxid + 1; i++) {
4980 bzero(&ups[i], sizeof(*ups));
4981 if (internal || CPU_ABSENT(i))
4983 cache = &z->uz_cpu[i];
4984 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
4985 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
4986 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
4987 ups[i].ups_allocs = cache->uc_allocs;
4988 ups[i].ups_frees = cache->uc_frees;
4993 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
4995 struct uma_stream_header ush;
4996 struct uma_type_header uth;
4997 struct uma_percpu_stat *ups;
5002 uint32_t kfree, pages;
5003 int count, error, i;
5005 error = sysctl_wire_old_buffer(req, 0);
5008 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5009 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5010 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5013 rw_rlock(&uma_rwlock);
5014 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5015 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5019 LIST_FOREACH(z, &uma_cachezones, uz_link)
5023 * Insert stream header.
5025 bzero(&ush, sizeof(ush));
5026 ush.ush_version = UMA_STREAM_VERSION;
5027 ush.ush_maxcpus = (mp_maxid + 1);
5028 ush.ush_count = count;
5029 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5031 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5033 for (i = 0; i < vm_ndomains; i++) {
5034 kfree += kz->uk_domain[i].ud_free_items;
5035 pages += kz->uk_domain[i].ud_pages;
5037 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5038 bzero(&uth, sizeof(uth));
5039 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5040 uth.uth_align = kz->uk_align;
5041 uth.uth_size = kz->uk_size;
5042 uth.uth_rsize = kz->uk_rsize;
5043 if (z->uz_max_items > 0) {
5044 items = UZ_ITEMS_COUNT(z->uz_items);
5045 uth.uth_pages = (items / kz->uk_ipers) *
5048 uth.uth_pages = pages;
5049 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5051 uth.uth_limit = z->uz_max_items;
5052 uth.uth_keg_free = kfree;
5055 * A zone is secondary is it is not the first entry
5056 * on the keg's zone list.
5058 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5059 (LIST_FIRST(&kz->uk_zones) != z))
5060 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5061 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5062 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5063 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5064 for (i = 0; i < mp_maxid + 1; i++)
5065 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5068 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5069 bzero(&uth, sizeof(uth));
5070 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5071 uth.uth_size = z->uz_size;
5072 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5073 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5074 for (i = 0; i < mp_maxid + 1; i++)
5075 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5078 rw_runlock(&uma_rwlock);
5079 error = sbuf_finish(&sbuf);
5086 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5088 uma_zone_t zone = *(uma_zone_t *)arg1;
5091 max = uma_zone_get_max(zone);
5092 error = sysctl_handle_int(oidp, &max, 0, req);
5093 if (error || !req->newptr)
5096 uma_zone_set_max(zone, max);
5102 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5108 * Some callers want to add sysctls for global zones that
5109 * may not yet exist so they pass a pointer to a pointer.
5112 zone = *(uma_zone_t *)arg1;
5115 cur = uma_zone_get_cur(zone);
5116 return (sysctl_handle_int(oidp, &cur, 0, req));
5120 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5122 uma_zone_t zone = arg1;
5125 cur = uma_zone_get_allocs(zone);
5126 return (sysctl_handle_64(oidp, &cur, 0, req));
5130 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5132 uma_zone_t zone = arg1;
5135 cur = uma_zone_get_frees(zone);
5136 return (sysctl_handle_64(oidp, &cur, 0, req));
5140 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5143 uma_zone_t zone = arg1;
5146 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5147 if (zone->uz_flags != 0)
5148 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5150 sbuf_printf(&sbuf, "0");
5151 error = sbuf_finish(&sbuf);
5158 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5160 uma_keg_t keg = arg1;
5161 int avail, effpct, total;
5163 total = keg->uk_ppera * PAGE_SIZE;
5164 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5165 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5167 * We consider the client's requested size and alignment here, not the
5168 * real size determination uk_rsize, because we also adjust the real
5169 * size for internal implementation reasons (max bitset size).
5171 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5172 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5173 avail *= mp_maxid + 1;
5174 effpct = 100 * avail / total;
5175 return (sysctl_handle_int(oidp, &effpct, 0, req));
5179 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5181 uma_zone_t zone = arg1;
5184 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5185 return (sysctl_handle_64(oidp, &cur, 0, req));
5190 uma_dbg_getslab(uma_zone_t zone, void *item)
5197 * It is safe to return the slab here even though the
5198 * zone is unlocked because the item's allocation state
5199 * essentially holds a reference.
5201 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5202 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5204 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5205 return (vtoslab((vm_offset_t)mem));
5207 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5208 return ((uma_slab_t)(mem + keg->uk_pgoff));
5210 slab = hash_sfind(&keg->uk_hash, mem);
5217 uma_dbg_zskip(uma_zone_t zone, void *mem)
5220 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5223 return (uma_dbg_kskip(zone->uz_keg, mem));
5227 uma_dbg_kskip(uma_keg_t keg, void *mem)
5231 if (dbg_divisor == 0)
5234 if (dbg_divisor == 1)
5237 idx = (uintptr_t)mem >> PAGE_SHIFT;
5238 if (keg->uk_ipers > 1) {
5239 idx *= keg->uk_ipers;
5240 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5243 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5244 counter_u64_add(uma_skip_cnt, 1);
5247 counter_u64_add(uma_dbg_cnt, 1);
5253 * Set up the slab's freei data such that uma_dbg_free can function.
5257 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5263 slab = uma_dbg_getslab(zone, item);
5265 panic("uma: item %p did not belong to zone %s\n",
5266 item, zone->uz_name);
5269 freei = slab_item_index(slab, keg, item);
5271 if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5272 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
5273 item, zone, zone->uz_name, slab, freei);
5274 BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5278 * Verifies freed addresses. Checks for alignment, valid slab membership
5279 * and duplicate frees.
5283 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5289 slab = uma_dbg_getslab(zone, item);
5291 panic("uma: Freed item %p did not belong to zone %s\n",
5292 item, zone->uz_name);
5295 freei = slab_item_index(slab, keg, item);
5297 if (freei >= keg->uk_ipers)
5298 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
5299 item, zone, zone->uz_name, slab, freei);
5301 if (slab_item(slab, keg, freei) != item)
5302 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
5303 item, zone, zone->uz_name, slab, freei);
5305 if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5306 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
5307 item, zone, zone->uz_name, slab, freei);
5309 BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5311 #endif /* INVARIANTS */
5315 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5316 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5321 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5322 *allocs = counter_u64_fetch(z->uz_allocs);
5323 frees = counter_u64_fetch(z->uz_frees);
5324 *sleeps = z->uz_sleeps;
5328 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5330 for (i = 0; i < vm_ndomains; i++) {
5331 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5332 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5333 (LIST_FIRST(&kz->uk_zones) != z)))
5334 *cachefree += kz->uk_domain[i].ud_free_items;
5336 *used = *allocs - frees;
5337 return (((int64_t)*used + *cachefree) * kz->uk_size);
5340 DB_SHOW_COMMAND(uma, db_show_uma)
5342 const char *fmt_hdr, *fmt_entry;
5345 uint64_t allocs, used, sleeps, xdomain;
5347 /* variables for sorting */
5349 uma_zone_t cur_zone, last_zone;
5350 int64_t cur_size, last_size, size;
5353 /* /i option produces machine-parseable CSV output */
5354 if (modif[0] == 'i') {
5355 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5356 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5358 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5359 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5362 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5363 "Sleeps", "Bucket", "Total Mem", "XFree");
5365 /* Sort the zones with largest size first. */
5367 last_size = INT64_MAX;
5372 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5373 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5375 * In the case of size ties, print out zones
5376 * in the order they are encountered. That is,
5377 * when we encounter the most recently output
5378 * zone, we have already printed all preceding
5379 * ties, and we must print all following ties.
5381 if (z == last_zone) {
5385 size = get_uma_stats(kz, z, &allocs, &used,
5386 &sleeps, &cachefree, &xdomain);
5387 if (size > cur_size && size < last_size + ties)
5395 if (cur_zone == NULL)
5398 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5399 &sleeps, &cachefree, &xdomain);
5400 db_printf(fmt_entry, cur_zone->uz_name,
5401 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5402 (uintmax_t)allocs, (uintmax_t)sleeps,
5403 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5408 last_zone = cur_zone;
5409 last_size = cur_size;
5413 DB_SHOW_COMMAND(umacache, db_show_umacache)
5416 uint64_t allocs, frees;
5420 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5421 "Requests", "Bucket");
5422 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5423 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5424 for (i = 0; i < vm_ndomains; i++)
5425 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5426 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5427 z->uz_name, (uintmax_t)z->uz_size,
5428 (intmax_t)(allocs - frees), cachefree,
5429 (uintmax_t)allocs, z->uz_bucket_size);