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 * These are the two zones from which all offpage uma_slab_ts are allocated.
122 * One zone is for slab headers that can represent a larger number of items,
123 * making the slabs themselves more efficient, and the other zone is for
124 * headers that are smaller and represent fewer items, making the headers more
127 #define SLABZONE_SIZE(setsize) \
128 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
129 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16)
130 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE
131 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE)
132 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE)
133 static uma_zone_t slabzones[2];
136 * The initial hash tables come out of this zone so they can be allocated
137 * prior to malloc coming up.
139 static uma_zone_t hashzone;
141 /* The boot-time adjusted value for cache line alignment. */
142 int uma_align_cache = 64 - 1;
144 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
145 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
148 * Are we allowed to allocate buckets?
150 static int bucketdisable = 1;
152 /* Linked list of all kegs in the system */
153 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
155 /* Linked list of all cache-only zones in the system */
156 static LIST_HEAD(,uma_zone) uma_cachezones =
157 LIST_HEAD_INITIALIZER(uma_cachezones);
159 /* This RW lock protects the keg list */
160 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
163 * First available virual address for boot time allocations.
165 static vm_offset_t bootstart;
166 static vm_offset_t bootmem;
168 static struct sx uma_reclaim_lock;
171 * kmem soft limit, initialized by uma_set_limit(). Ensure that early
172 * allocations don't trigger a wakeup of the reclaim thread.
174 unsigned long uma_kmem_limit = LONG_MAX;
175 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
176 "UMA kernel memory soft limit");
177 unsigned long uma_kmem_total;
178 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
179 "UMA kernel memory usage");
181 /* Is the VM done starting up? */
187 } booted = BOOT_COLD;
190 * This is the handle used to schedule events that need to happen
191 * outside of the allocation fast path.
193 static struct callout uma_callout;
194 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
197 * This structure is passed as the zone ctor arg so that I don't have to create
198 * a special allocation function just for zones.
200 struct uma_zctor_args {
215 struct uma_kctor_args {
224 struct uma_bucket_zone {
226 const char *ubz_name;
227 int ubz_entries; /* Number of items it can hold. */
228 int ubz_maxsize; /* Maximum allocation size per-item. */
232 * Compute the actual number of bucket entries to pack them in power
233 * of two sizes for more efficient space utilization.
235 #define BUCKET_SIZE(n) \
236 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
238 #define BUCKET_MAX BUCKET_SIZE(256)
241 struct uma_bucket_zone bucket_zones[] = {
242 /* Literal bucket sizes. */
243 { NULL, "2 Bucket", 2, 4096 },
244 { NULL, "4 Bucket", 4, 3072 },
245 { NULL, "8 Bucket", 8, 2048 },
246 { NULL, "16 Bucket", 16, 1024 },
247 /* Rounded down power of 2 sizes for efficiency. */
248 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
249 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
250 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
251 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
256 * Flags and enumerations to be passed to internal functions.
260 SKIP_CNT = 0x00000001,
261 SKIP_DTOR = 0x00010000,
262 SKIP_FINI = 0x00020000,
267 void uma_startup1(vm_offset_t);
268 void uma_startup2(void);
270 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
271 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
272 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
273 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
274 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
275 static void page_free(void *, vm_size_t, uint8_t);
276 static void pcpu_page_free(void *, vm_size_t, uint8_t);
277 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
278 static void cache_drain(uma_zone_t);
279 static void bucket_drain(uma_zone_t, uma_bucket_t);
280 static void bucket_cache_reclaim(uma_zone_t zone, bool);
281 static int keg_ctor(void *, int, void *, int);
282 static void keg_dtor(void *, int, void *);
283 static int zone_ctor(void *, int, void *, int);
284 static void zone_dtor(void *, int, void *);
285 static inline void item_dtor(uma_zone_t zone, void *item, int size,
286 void *udata, enum zfreeskip skip);
287 static int zero_init(void *, int, int);
288 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
289 int itemdomain, bool ws);
290 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
291 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
292 static void zone_timeout(uma_zone_t zone, void *);
293 static int hash_alloc(struct uma_hash *, u_int);
294 static int hash_expand(struct uma_hash *, struct uma_hash *);
295 static void hash_free(struct uma_hash *hash);
296 static void uma_timeout(void *);
297 static void uma_startup3(void);
298 static void uma_shutdown(void);
299 static void *zone_alloc_item(uma_zone_t, void *, int, int);
300 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
301 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
302 static void zone_free_limit(uma_zone_t zone, int count);
303 static void bucket_enable(void);
304 static void bucket_init(void);
305 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
306 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
307 static void bucket_zone_drain(void);
308 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
309 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
310 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
311 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
312 uma_fini fini, int align, uint32_t flags);
313 static int zone_import(void *, void **, int, int, int);
314 static void zone_release(void *, void **, int);
315 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
316 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
318 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
319 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
320 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
321 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
322 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
323 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
324 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
326 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
328 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
329 "Memory allocation debugging");
332 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
333 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
335 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
336 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
337 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
338 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
340 static u_int dbg_divisor = 1;
341 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
342 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
343 "Debug & thrash every this item in memory allocator");
345 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
346 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
347 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
348 &uma_dbg_cnt, "memory items debugged");
349 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
350 &uma_skip_cnt, "memory items skipped, not debugged");
353 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
355 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
356 "Universal Memory Allocator");
358 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
359 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
361 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
362 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
364 static int zone_warnings = 1;
365 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
366 "Warn when UMA zones becomes full");
368 static int multipage_slabs = 1;
369 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
370 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
371 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
372 "UMA may choose larger slab sizes for better efficiency");
375 * Select the slab zone for an offpage slab with the given maximum item count.
377 static inline uma_zone_t
381 return (slabzones[ipers > SLABZONE0_SETSIZE]);
385 * This routine checks to see whether or not it's safe to enable buckets.
391 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
392 bucketdisable = vm_page_count_min();
396 * Initialize bucket_zones, the array of zones of buckets of various sizes.
398 * For each zone, calculate the memory required for each bucket, consisting
399 * of the header and an array of pointers.
404 struct uma_bucket_zone *ubz;
407 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
408 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
409 size += sizeof(void *) * ubz->ubz_entries;
410 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
411 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
412 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
413 UMA_ZONE_FIRSTTOUCH);
418 * Given a desired number of entries for a bucket, return the zone from which
419 * to allocate the bucket.
421 static struct uma_bucket_zone *
422 bucket_zone_lookup(int entries)
424 struct uma_bucket_zone *ubz;
426 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
427 if (ubz->ubz_entries >= entries)
433 static struct uma_bucket_zone *
434 bucket_zone_max(uma_zone_t zone, int nitems)
436 struct uma_bucket_zone *ubz;
440 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
441 /* Count the cross-domain bucket. */
444 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
445 if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems)
447 if (ubz == &bucket_zones[0])
455 bucket_select(int size)
457 struct uma_bucket_zone *ubz;
459 ubz = &bucket_zones[0];
460 if (size > ubz->ubz_maxsize)
461 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
463 for (; ubz->ubz_entries != 0; ubz++)
464 if (ubz->ubz_maxsize < size)
467 return (ubz->ubz_entries);
471 bucket_alloc(uma_zone_t zone, void *udata, int flags)
473 struct uma_bucket_zone *ubz;
477 * Don't allocate buckets early in boot.
479 if (__predict_false(booted < BOOT_KVA))
483 * To limit bucket recursion we store the original zone flags
484 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
485 * NOVM flag to persist even through deep recursions. We also
486 * store ZFLAG_BUCKET once we have recursed attempting to allocate
487 * a bucket for a bucket zone so we do not allow infinite bucket
488 * recursion. This cookie will even persist to frees of unused
489 * buckets via the allocation path or bucket allocations in the
492 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
493 udata = (void *)(uintptr_t)zone->uz_flags;
495 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
497 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
499 if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
501 ubz = bucket_zone_lookup(zone->uz_bucket_size);
502 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
504 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
507 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
510 bucket->ub_entries = ubz->ubz_entries;
511 bucket->ub_seq = SMR_SEQ_INVALID;
512 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
513 zone->uz_name, zone, bucket);
520 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
522 struct uma_bucket_zone *ubz;
524 if (bucket->ub_cnt != 0)
525 bucket_drain(zone, bucket);
527 KASSERT(bucket->ub_cnt == 0,
528 ("bucket_free: Freeing a non free bucket."));
529 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
530 ("bucket_free: Freeing an SMR bucket."));
531 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
532 udata = (void *)(uintptr_t)zone->uz_flags;
533 ubz = bucket_zone_lookup(bucket->ub_entries);
534 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
538 bucket_zone_drain(void)
540 struct uma_bucket_zone *ubz;
542 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
543 uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN);
547 * Acquire the domain lock and record contention.
549 static uma_zone_domain_t
550 zone_domain_lock(uma_zone_t zone, int domain)
552 uma_zone_domain_t zdom;
555 zdom = ZDOM_GET(zone, domain);
557 if (ZDOM_OWNED(zdom))
560 /* This is unsynchronized. The counter does not need to be precise. */
561 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
562 zone->uz_bucket_size++;
567 * Search for the domain with the least cached items and return it, breaking
568 * ties with a preferred domain by returning it.
570 static __noinline int
571 zone_domain_lowest(uma_zone_t zone, int pref)
579 for (i = 0; i < vm_ndomains; i++) {
580 nitems = ZDOM_GET(zone, i)->uzd_nitems;
581 if (nitems < least) {
584 } else if (nitems == least && (i == pref || domain == pref))
592 * Search for the domain with the most cached items and return it or the
593 * preferred domain if it has enough to proceed.
595 static __noinline int
596 zone_domain_highest(uma_zone_t zone, int pref)
602 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
607 for (i = 0; i < vm_ndomains; i++) {
608 nitems = ZDOM_GET(zone, i)->uzd_nitems;
619 * Safely subtract cnt from imax.
622 zone_domain_imax_sub(uma_zone_domain_t zdom, int cnt)
627 old = zdom->uzd_imax;
633 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, new) == 0);
637 * Set the maximum imax value.
640 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
644 old = zdom->uzd_imax;
648 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
652 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
653 * zone's caches. If a bucket is found the zone is not locked on return.
656 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
662 ZDOM_LOCK_ASSERT(zdom);
664 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
667 /* SMR Buckets can not be re-used until readers expire. */
668 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
669 bucket->ub_seq != SMR_SEQ_INVALID) {
670 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
672 bucket->ub_seq = SMR_SEQ_INVALID;
673 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
674 if (STAILQ_NEXT(bucket, ub_link) != NULL)
675 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
677 MPASS(zdom->uzd_nitems >= bucket->ub_cnt);
678 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
679 zdom->uzd_nitems -= bucket->ub_cnt;
682 * Shift the bounds of the current WSS interval to avoid
683 * perturbing the estimate.
686 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
687 zone_domain_imax_sub(zdom, bucket->ub_cnt);
688 } else if (zdom->uzd_imin > zdom->uzd_nitems)
689 zdom->uzd_imin = zdom->uzd_nitems;
693 for (i = 0; i < bucket->ub_cnt; i++)
694 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
701 * Insert a full bucket into the specified cache. The "ws" parameter indicates
702 * whether the bucket's contents should be counted as part of the zone's working
703 * set. The bucket may be freed if it exceeds the bucket limit.
706 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
709 uma_zone_domain_t zdom;
711 /* We don't cache empty buckets. This can happen after a reclaim. */
712 if (bucket->ub_cnt == 0)
714 zdom = zone_domain_lock(zone, domain);
716 KASSERT(!ws || zdom->uzd_nitems < zone->uz_bucket_max,
717 ("%s: zone %p overflow", __func__, zone));
720 * Conditionally set the maximum number of items.
722 zdom->uzd_nitems += bucket->ub_cnt;
723 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
725 zone_domain_imax_set(zdom, zdom->uzd_nitems);
726 if (STAILQ_EMPTY(&zdom->uzd_buckets))
727 zdom->uzd_seq = bucket->ub_seq;
728 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
732 zdom->uzd_nitems -= bucket->ub_cnt;
735 bucket_free(zone, bucket, udata);
738 /* Pops an item out of a per-cpu cache bucket. */
740 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
744 CRITICAL_ASSERT(curthread);
747 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
749 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
750 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
757 /* Pushes an item into a per-cpu cache bucket. */
759 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
762 CRITICAL_ASSERT(curthread);
763 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
764 ("uma_zfree: Freeing to non free bucket index."));
766 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
772 * Unload a UMA bucket from a per-cpu cache.
774 static inline uma_bucket_t
775 cache_bucket_unload(uma_cache_bucket_t bucket)
779 b = bucket->ucb_bucket;
781 MPASS(b->ub_entries == bucket->ucb_entries);
782 b->ub_cnt = bucket->ucb_cnt;
783 bucket->ucb_bucket = NULL;
784 bucket->ucb_entries = bucket->ucb_cnt = 0;
790 static inline uma_bucket_t
791 cache_bucket_unload_alloc(uma_cache_t cache)
794 return (cache_bucket_unload(&cache->uc_allocbucket));
797 static inline uma_bucket_t
798 cache_bucket_unload_free(uma_cache_t cache)
801 return (cache_bucket_unload(&cache->uc_freebucket));
804 static inline uma_bucket_t
805 cache_bucket_unload_cross(uma_cache_t cache)
808 return (cache_bucket_unload(&cache->uc_crossbucket));
812 * Load a bucket into a per-cpu cache bucket.
815 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
818 CRITICAL_ASSERT(curthread);
819 MPASS(bucket->ucb_bucket == NULL);
820 MPASS(b->ub_seq == SMR_SEQ_INVALID);
822 bucket->ucb_bucket = b;
823 bucket->ucb_cnt = b->ub_cnt;
824 bucket->ucb_entries = b->ub_entries;
828 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
831 cache_bucket_load(&cache->uc_allocbucket, b);
835 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
838 cache_bucket_load(&cache->uc_freebucket, b);
843 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
846 cache_bucket_load(&cache->uc_crossbucket, b);
851 * Copy and preserve ucb_spare.
854 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
857 b1->ucb_bucket = b2->ucb_bucket;
858 b1->ucb_entries = b2->ucb_entries;
859 b1->ucb_cnt = b2->ucb_cnt;
863 * Swap two cache buckets.
866 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
868 struct uma_cache_bucket b3;
870 CRITICAL_ASSERT(curthread);
872 cache_bucket_copy(&b3, b1);
873 cache_bucket_copy(b1, b2);
874 cache_bucket_copy(b2, &b3);
878 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
881 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
883 uma_zone_domain_t zdom;
887 * Avoid the lock if possible.
889 zdom = ZDOM_GET(zone, domain);
890 if (zdom->uzd_nitems == 0)
893 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
894 !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
898 * Check the zone's cache of buckets.
900 zdom = zone_domain_lock(zone, domain);
901 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
902 KASSERT(bucket->ub_cnt != 0,
903 ("cache_fetch_bucket: Returning an empty bucket."));
912 zone_log_warning(uma_zone_t zone)
914 static const struct timeval warninterval = { 300, 0 };
916 if (!zone_warnings || zone->uz_warning == NULL)
919 if (ratecheck(&zone->uz_ratecheck, &warninterval))
920 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
924 zone_maxaction(uma_zone_t zone)
927 if (zone->uz_maxaction.ta_func != NULL)
928 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
932 * Routine called by timeout which is used to fire off some time interval
933 * based calculations. (stats, hash size, etc.)
942 uma_timeout(void *unused)
945 zone_foreach(zone_timeout, NULL);
947 /* Reschedule this event */
948 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
952 * Update the working set size estimate for the zone's bucket cache.
953 * The constants chosen here are somewhat arbitrary. With an update period of
954 * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
958 zone_domain_update_wss(uma_zone_domain_t zdom)
963 MPASS(zdom->uzd_imax >= zdom->uzd_imin);
964 wss = zdom->uzd_imax - zdom->uzd_imin;
965 zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
966 zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
971 * Routine to perform timeout driven calculations. This expands the
972 * hashes and does per cpu statistics aggregation.
977 zone_timeout(uma_zone_t zone, void *unused)
982 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
988 * Hash zones are non-numa by definition so the first domain
989 * is the only one present.
992 pages = keg->uk_domain[0].ud_pages;
995 * Expand the keg hash table.
997 * This is done if the number of slabs is larger than the hash size.
998 * What I'm trying to do here is completely reduce collisions. This
999 * may be a little aggressive. Should I allow for two collisions max?
1001 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1002 struct uma_hash newhash;
1003 struct uma_hash oldhash;
1007 * This is so involved because allocating and freeing
1008 * while the keg lock is held will lead to deadlock.
1009 * I have to do everything in stages and check for
1013 ret = hash_alloc(&newhash, 1 << fls(slabs));
1016 if (hash_expand(&keg->uk_hash, &newhash)) {
1017 oldhash = keg->uk_hash;
1018 keg->uk_hash = newhash;
1023 hash_free(&oldhash);
1030 for (int i = 0; i < vm_ndomains; i++)
1031 zone_domain_update_wss(ZDOM_GET(zone, i));
1035 * Allocate and zero fill the next sized hash table from the appropriate
1039 * hash A new hash structure with the old hash size in uh_hashsize
1042 * 1 on success and 0 on failure.
1045 hash_alloc(struct uma_hash *hash, u_int size)
1049 KASSERT(powerof2(size), ("hash size must be power of 2"));
1050 if (size > UMA_HASH_SIZE_INIT) {
1051 hash->uh_hashsize = size;
1052 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1053 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1055 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1056 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1057 UMA_ANYDOMAIN, M_WAITOK);
1058 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1060 if (hash->uh_slab_hash) {
1061 bzero(hash->uh_slab_hash, alloc);
1062 hash->uh_hashmask = hash->uh_hashsize - 1;
1070 * Expands the hash table for HASH zones. This is done from zone_timeout
1071 * to reduce collisions. This must not be done in the regular allocation
1072 * path, otherwise, we can recurse on the vm while allocating pages.
1075 * oldhash The hash you want to expand
1076 * newhash The hash structure for the new table
1084 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1086 uma_hash_slab_t slab;
1090 if (!newhash->uh_slab_hash)
1093 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1097 * I need to investigate hash algorithms for resizing without a
1101 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1102 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1103 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1104 LIST_REMOVE(slab, uhs_hlink);
1105 hval = UMA_HASH(newhash, slab->uhs_data);
1106 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1114 * Free the hash bucket to the appropriate backing store.
1117 * slab_hash The hash bucket we're freeing
1118 * hashsize The number of entries in that hash bucket
1124 hash_free(struct uma_hash *hash)
1126 if (hash->uh_slab_hash == NULL)
1128 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1129 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1131 free(hash->uh_slab_hash, M_UMAHASH);
1135 * Frees all outstanding items in a bucket
1138 * zone The zone to free to, must be unlocked.
1139 * bucket The free/alloc bucket with items.
1145 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1149 if (bucket->ub_cnt == 0)
1152 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1153 bucket->ub_seq != SMR_SEQ_INVALID) {
1154 smr_wait(zone->uz_smr, bucket->ub_seq);
1155 bucket->ub_seq = SMR_SEQ_INVALID;
1156 for (i = 0; i < bucket->ub_cnt; i++)
1157 item_dtor(zone, bucket->ub_bucket[i],
1158 zone->uz_size, NULL, SKIP_NONE);
1161 for (i = 0; i < bucket->ub_cnt; i++)
1162 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1163 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1164 if (zone->uz_max_items > 0)
1165 zone_free_limit(zone, bucket->ub_cnt);
1167 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1173 * Drains the per cpu caches for a zone.
1175 * NOTE: This may only be called while the zone is being torn down, and not
1176 * during normal operation. This is necessary in order that we do not have
1177 * to migrate CPUs to drain the per-CPU caches.
1180 * zone The zone to drain, must be unlocked.
1186 cache_drain(uma_zone_t zone)
1189 uma_bucket_t bucket;
1194 * XXX: It is safe to not lock the per-CPU caches, because we're
1195 * tearing down the zone anyway. I.e., there will be no further use
1196 * of the caches at this point.
1198 * XXX: It would good to be able to assert that the zone is being
1199 * torn down to prevent improper use of cache_drain().
1201 seq = SMR_SEQ_INVALID;
1202 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1203 seq = smr_advance(zone->uz_smr);
1205 cache = &zone->uz_cpu[cpu];
1206 bucket = cache_bucket_unload_alloc(cache);
1208 bucket_free(zone, bucket, NULL);
1209 bucket = cache_bucket_unload_free(cache);
1210 if (bucket != NULL) {
1211 bucket->ub_seq = seq;
1212 bucket_free(zone, bucket, NULL);
1214 bucket = cache_bucket_unload_cross(cache);
1215 if (bucket != NULL) {
1216 bucket->ub_seq = seq;
1217 bucket_free(zone, bucket, NULL);
1220 bucket_cache_reclaim(zone, true);
1224 cache_shrink(uma_zone_t zone, void *unused)
1227 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1230 zone->uz_bucket_size =
1231 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1235 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1238 uma_bucket_t b1, b2, b3;
1241 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1244 b1 = b2 = b3 = NULL;
1246 cache = &zone->uz_cpu[curcpu];
1247 domain = PCPU_GET(domain);
1248 b1 = cache_bucket_unload_alloc(cache);
1251 * Don't flush SMR zone buckets. This leaves the zone without a
1252 * bucket and forces every free to synchronize().
1254 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1255 b2 = cache_bucket_unload_free(cache);
1256 b3 = cache_bucket_unload_cross(cache);
1261 zone_free_bucket(zone, b1, NULL, domain, false);
1263 zone_free_bucket(zone, b2, NULL, domain, false);
1265 /* Adjust the domain so it goes to zone_free_cross. */
1266 domain = (domain + 1) % vm_ndomains;
1267 zone_free_bucket(zone, b3, NULL, domain, false);
1272 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1273 * This is an expensive call because it needs to bind to all CPUs
1274 * one by one and enter a critical section on each of them in order
1275 * to safely access their cache buckets.
1276 * Zone lock must not be held on call this function.
1279 pcpu_cache_drain_safe(uma_zone_t zone)
1284 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1287 cache_shrink(zone, NULL);
1289 zone_foreach(cache_shrink, NULL);
1292 thread_lock(curthread);
1293 sched_bind(curthread, cpu);
1294 thread_unlock(curthread);
1297 cache_drain_safe_cpu(zone, NULL);
1299 zone_foreach(cache_drain_safe_cpu, NULL);
1301 thread_lock(curthread);
1302 sched_unbind(curthread);
1303 thread_unlock(curthread);
1307 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1308 * requested a drain, otherwise the per-domain caches are trimmed to either
1309 * estimated working set size.
1312 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1314 uma_zone_domain_t zdom;
1315 uma_bucket_t bucket;
1320 * Shrink the zone bucket size to ensure that the per-CPU caches
1321 * don't grow too large.
1323 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1324 zone->uz_bucket_size--;
1326 for (i = 0; i < vm_ndomains; i++) {
1328 * The cross bucket is partially filled and not part of
1329 * the item count. Reclaim it individually here.
1331 zdom = ZDOM_GET(zone, i);
1332 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1333 ZONE_CROSS_LOCK(zone);
1334 bucket = zdom->uzd_cross;
1335 zdom->uzd_cross = NULL;
1336 ZONE_CROSS_UNLOCK(zone);
1338 bucket_free(zone, bucket, NULL);
1342 * If we were asked to drain the zone, we are done only once
1343 * this bucket cache is empty. Otherwise, we reclaim items in
1344 * excess of the zone's estimated working set size. If the
1345 * difference nitems - imin is larger than the WSS estimate,
1346 * then the estimate will grow at the end of this interval and
1347 * we ignore the historical average.
1350 target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1352 while (zdom->uzd_nitems > target) {
1353 bucket = zone_fetch_bucket(zone, zdom, true);
1356 bucket_free(zone, bucket, NULL);
1364 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1370 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1371 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1373 mem = slab_data(slab, keg);
1374 flags = slab->us_flags;
1376 if (keg->uk_fini != NULL) {
1377 for (i--; i > -1; i--)
1380 * trash_fini implies that dtor was trash_dtor. trash_fini
1381 * would check that memory hasn't been modified since free,
1382 * which executed trash_dtor.
1383 * That's why we need to run uma_dbg_kskip() check here,
1384 * albeit we don't make skip check for other init/fini
1387 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1388 keg->uk_fini != trash_fini)
1390 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1392 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1393 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1395 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1396 uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1400 * Frees pages from a keg back to the system. This is done on demand from
1401 * the pageout daemon.
1406 keg_drain(uma_keg_t keg)
1408 struct slabhead freeslabs;
1410 uma_slab_t slab, tmp;
1413 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1416 for (i = 0; i < vm_ndomains; i++) {
1417 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1418 keg->uk_name, keg, i, dom->ud_free_items);
1419 dom = &keg->uk_domain[i];
1420 LIST_INIT(&freeslabs);
1423 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1424 LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
1425 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1427 n = dom->ud_free_slabs;
1428 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1429 dom->ud_free_slabs = 0;
1430 dom->ud_free_items -= n * keg->uk_ipers;
1431 dom->ud_pages -= n * keg->uk_ppera;
1434 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1435 keg_free_slab(keg, slab, keg->uk_ipers);
1440 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1444 * Set draining to interlock with zone_dtor() so we can release our
1445 * locks as we go. Only dtor() should do a WAITOK call since it
1446 * is the only call that knows the structure will still be available
1450 while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1451 if (waitok == M_NOWAIT)
1453 msleep(zone, &ZDOM_GET(zone, 0)->uzd_lock, PVM, "zonedrain",
1456 zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1458 bucket_cache_reclaim(zone, drain);
1461 * The DRAINING flag protects us from being freed while
1462 * we're running. Normally the uma_rwlock would protect us but we
1463 * must be able to release and acquire the right lock for each keg.
1465 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1466 keg_drain(zone->uz_keg);
1468 zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1475 zone_drain(uma_zone_t zone, void *unused)
1478 zone_reclaim(zone, M_NOWAIT, true);
1482 zone_trim(uma_zone_t zone, void *unused)
1485 zone_reclaim(zone, M_NOWAIT, false);
1489 * Allocate a new slab for a keg and inserts it into the partial slab list.
1490 * The keg should be unlocked on entry. If the allocation succeeds it will
1491 * be locked on return.
1494 * flags Wait flags for the item initialization routine
1495 * aflags Wait flags for the slab allocation
1498 * The slab that was allocated or NULL if there is no memory and the
1499 * caller specified M_NOWAIT.
1502 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1513 KASSERT(domain >= 0 && domain < vm_ndomains,
1514 ("keg_alloc_slab: domain %d out of range", domain));
1516 allocf = keg->uk_allocf;
1519 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1520 uma_hash_slab_t hslab;
1521 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1525 slab = &hslab->uhs_slab;
1529 * This reproduces the old vm_zone behavior of zero filling pages the
1530 * first time they are added to a zone.
1532 * Malloced items are zeroed in uma_zalloc.
1535 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1540 if (keg->uk_flags & UMA_ZONE_NODUMP)
1543 /* zone is passed for legacy reasons. */
1544 size = keg->uk_ppera * PAGE_SIZE;
1545 mem = allocf(zone, size, domain, &sflags, aflags);
1547 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1548 zone_free_item(slabzone(keg->uk_ipers),
1549 slab_tohashslab(slab), NULL, SKIP_NONE);
1552 uma_total_inc(size);
1554 /* For HASH zones all pages go to the same uma_domain. */
1555 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1558 /* Point the slab into the allocated memory */
1559 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1560 slab = (uma_slab_t )(mem + keg->uk_pgoff);
1562 slab_tohashslab(slab)->uhs_data = mem;
1564 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1565 for (i = 0; i < keg->uk_ppera; i++)
1566 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1569 slab->us_freecount = keg->uk_ipers;
1570 slab->us_flags = sflags;
1571 slab->us_domain = domain;
1573 BIT_FILL(keg->uk_ipers, &slab->us_free);
1575 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1578 if (keg->uk_init != NULL) {
1579 for (i = 0; i < keg->uk_ipers; i++)
1580 if (keg->uk_init(slab_item(slab, keg, i),
1581 keg->uk_size, flags) != 0)
1583 if (i != keg->uk_ipers) {
1584 keg_free_slab(keg, slab, i);
1588 KEG_LOCK(keg, domain);
1590 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1591 slab, keg->uk_name, keg);
1593 if (keg->uk_flags & UMA_ZFLAG_HASH)
1594 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1597 * If we got a slab here it's safe to mark it partially used
1598 * and return. We assume that the caller is going to remove
1599 * at least one item.
1601 dom = &keg->uk_domain[domain];
1602 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1603 dom->ud_pages += keg->uk_ppera;
1604 dom->ud_free_items += keg->uk_ipers;
1613 * This function is intended to be used early on in place of page_alloc() so
1614 * that we may use the boot time page cache to satisfy allocations before
1618 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1627 pages = howmany(bytes, PAGE_SIZE);
1628 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1630 *pflag = UMA_SLAB_BOOT;
1631 m = vm_page_alloc_contig_domain(NULL, 0, domain,
1632 malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1633 (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1637 pa = VM_PAGE_TO_PHYS(m);
1638 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1639 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1640 defined(__riscv) || defined(__powerpc64__)
1641 if ((wait & M_NODUMP) == 0)
1645 /* Allocate KVA and indirectly advance bootmem. */
1646 mem = (void *)pmap_map(&bootmem, m->phys_addr,
1647 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1648 if ((wait & M_ZERO) != 0)
1649 bzero(mem, pages * PAGE_SIZE);
1655 startup_free(void *mem, vm_size_t bytes)
1660 va = (vm_offset_t)mem;
1661 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1662 pmap_remove(kernel_pmap, va, va + bytes);
1663 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1664 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1665 defined(__riscv) || defined(__powerpc64__)
1666 dump_drop_page(VM_PAGE_TO_PHYS(m));
1668 vm_page_unwire_noq(m);
1674 * Allocates a number of pages from the system
1677 * bytes The number of bytes requested
1678 * wait Shall we wait?
1681 * A pointer to the alloced memory or possibly
1682 * NULL if M_NOWAIT is set.
1685 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1688 void *p; /* Returned page */
1690 *pflag = UMA_SLAB_KERNEL;
1691 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1697 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1700 struct pglist alloctail;
1701 vm_offset_t addr, zkva;
1703 vm_page_t p, p_next;
1708 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1710 TAILQ_INIT(&alloctail);
1711 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1712 malloc2vm_flags(wait);
1713 *pflag = UMA_SLAB_KERNEL;
1714 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1715 if (CPU_ABSENT(cpu)) {
1716 p = vm_page_alloc(NULL, 0, flags);
1719 p = vm_page_alloc(NULL, 0, flags);
1721 pc = pcpu_find(cpu);
1722 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1725 p = vm_page_alloc_domain(NULL, 0,
1726 pc->pc_domain, flags);
1727 if (__predict_false(p == NULL))
1728 p = vm_page_alloc(NULL, 0, flags);
1731 if (__predict_false(p == NULL))
1733 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1735 if ((addr = kva_alloc(bytes)) == 0)
1738 TAILQ_FOREACH(p, &alloctail, listq) {
1739 pmap_qenter(zkva, &p, 1);
1742 return ((void*)addr);
1744 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1745 vm_page_unwire_noq(p);
1752 * Allocates a number of pages from within an object
1755 * bytes The number of bytes requested
1756 * wait Shall we wait?
1759 * A pointer to the alloced memory or possibly
1760 * NULL if M_NOWAIT is set.
1763 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1766 TAILQ_HEAD(, vm_page) alloctail;
1768 vm_offset_t retkva, zkva;
1769 vm_page_t p, p_next;
1772 TAILQ_INIT(&alloctail);
1775 npages = howmany(bytes, PAGE_SIZE);
1776 while (npages > 0) {
1777 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1778 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1779 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1783 * Since the page does not belong to an object, its
1786 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1791 * Page allocation failed, free intermediate pages and
1794 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1795 vm_page_unwire_noq(p);
1800 *flags = UMA_SLAB_PRIV;
1801 zkva = keg->uk_kva +
1802 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1804 TAILQ_FOREACH(p, &alloctail, listq) {
1805 pmap_qenter(zkva, &p, 1);
1809 return ((void *)retkva);
1813 * Allocate physically contiguous pages.
1816 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1820 *pflag = UMA_SLAB_KERNEL;
1821 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1822 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1826 * Frees a number of pages to the system
1829 * mem A pointer to the memory to be freed
1830 * size The size of the memory being freed
1831 * flags The original p->us_flags field
1837 page_free(void *mem, vm_size_t size, uint8_t flags)
1840 if ((flags & UMA_SLAB_BOOT) != 0) {
1841 startup_free(mem, size);
1845 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1846 ("UMA: page_free used with invalid flags %x", flags));
1848 kmem_free((vm_offset_t)mem, size);
1852 * Frees pcpu zone allocations
1855 * mem A pointer to the memory to be freed
1856 * size The size of the memory being freed
1857 * flags The original p->us_flags field
1863 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1865 vm_offset_t sva, curva;
1869 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1871 if ((flags & UMA_SLAB_BOOT) != 0) {
1872 startup_free(mem, size);
1876 sva = (vm_offset_t)mem;
1877 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1878 paddr = pmap_kextract(curva);
1879 m = PHYS_TO_VM_PAGE(paddr);
1880 vm_page_unwire_noq(m);
1883 pmap_qremove(sva, size >> PAGE_SHIFT);
1884 kva_free(sva, size);
1889 * Zero fill initializer
1891 * Arguments/Returns follow uma_init specifications
1894 zero_init(void *mem, int size, int flags)
1902 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1905 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1910 * Actual size of embedded struct slab (!OFFPAGE).
1913 slab_sizeof(int nitems)
1917 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1918 return (roundup(s, UMA_ALIGN_PTR + 1));
1922 * Size of memory for embedded slabs (!OFFPAGE).
1925 slab_space(int nitems)
1927 return (UMA_SLAB_SIZE - slab_sizeof(nitems));
1930 #define UMA_FIXPT_SHIFT 31
1931 #define UMA_FRAC_FIXPT(n, d) \
1932 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1933 #define UMA_FIXPT_PCT(f) \
1934 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1935 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
1936 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1939 * Compute the number of items that will fit in a slab. If hdr is true, the
1940 * item count may be limited to provide space in the slab for an inline slab
1941 * header. Otherwise, all slab space will be provided for item storage.
1944 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1949 /* The padding between items is not needed after the last item. */
1950 padpi = rsize - size;
1954 * Start with the maximum item count and remove items until
1955 * the slab header first alongside the allocatable memory.
1957 for (ipers = MIN(SLAB_MAX_SETSIZE,
1958 (slabsize + padpi - slab_sizeof(1)) / rsize);
1960 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1964 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1971 * Compute the number of items that will fit in a slab for a startup zone.
1974 slab_ipers(size_t size, int align)
1978 rsize = roundup(size, align + 1); /* Assume no CACHESPREAD */
1979 return (slab_ipers_hdr(size, rsize, UMA_SLAB_SIZE, true));
1982 struct keg_layout_result {
1990 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
1991 struct keg_layout_result *kl)
1996 kl->slabsize = slabsize;
1998 /* Handle INTERNAL as inline with an extra page. */
1999 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2000 kl->format &= ~UMA_ZFLAG_INTERNAL;
2001 kl->slabsize += PAGE_SIZE;
2004 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2005 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2007 /* Account for memory used by an offpage slab header. */
2008 total = kl->slabsize;
2009 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2010 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2012 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2016 * Determine the format of a uma keg. This determines where the slab header
2017 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2020 * keg The zone we should initialize
2026 keg_layout(uma_keg_t keg)
2028 struct keg_layout_result kl = {}, kl_tmp;
2037 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2038 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2039 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2040 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2041 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2043 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2044 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2045 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2048 alignsize = keg->uk_align + 1;
2051 * Calculate the size of each allocation (rsize) according to
2052 * alignment. If the requested size is smaller than we have
2053 * allocation bits for we round it up.
2055 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2056 rsize = roundup2(rsize, alignsize);
2058 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2060 * We want one item to start on every align boundary in a page.
2061 * To do this we will span pages. We will also extend the item
2062 * by the size of align if it is an even multiple of align.
2063 * Otherwise, it would fall on the same boundary every time.
2065 if ((rsize & alignsize) == 0)
2067 slabsize = rsize * (PAGE_SIZE / alignsize);
2068 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2069 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2070 slabsize = round_page(slabsize);
2073 * Start with a slab size of as many pages as it takes to
2074 * represent a single item. We will try to fit as many
2075 * additional items into the slab as possible.
2077 slabsize = round_page(keg->uk_size);
2080 /* Build a list of all of the available formats for this keg. */
2083 /* Evaluate an inline slab layout. */
2084 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2087 /* TODO: vm_page-embedded slab. */
2090 * We can't do OFFPAGE if we're internal or if we've been
2091 * asked to not go to the VM for buckets. If we do this we
2092 * may end up going to the VM for slabs which we do not want
2093 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2094 * In those cases, evaluate a pseudo-format called INTERNAL
2095 * which has an inline slab header and one extra page to
2096 * guarantee that it fits.
2098 * Otherwise, see if using an OFFPAGE slab will improve our
2101 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2102 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2104 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2107 * Choose a slab size and format which satisfy the minimum efficiency.
2108 * Prefer the smallest slab size that meets the constraints.
2110 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2111 * for small items (up to PAGE_SIZE), the iteration increment is one
2112 * page; and for large items, the increment is one item.
2114 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2115 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2116 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2119 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2120 round_page(rsize * (i - 1) + keg->uk_size);
2122 for (j = 0; j < nfmt; j++) {
2123 /* Only if we have no viable format yet. */
2124 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2128 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2129 if (kl_tmp.eff <= kl.eff)
2134 CTR6(KTR_UMA, "keg %s layout: format %#x "
2135 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2136 keg->uk_name, kl.format, kl.ipers, rsize,
2137 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2139 /* Stop when we reach the minimum efficiency. */
2140 if (kl.eff >= UMA_MIN_EFF)
2144 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2145 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2146 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2150 pages = atop(kl.slabsize);
2151 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2152 pages *= mp_maxid + 1;
2154 keg->uk_rsize = rsize;
2155 keg->uk_ipers = kl.ipers;
2156 keg->uk_ppera = pages;
2157 keg->uk_flags |= kl.format;
2160 * How do we find the slab header if it is offpage or if not all item
2161 * start addresses are in the same page? We could solve the latter
2162 * case with vaddr alignment, but we don't.
2164 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2165 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2166 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2167 keg->uk_flags |= UMA_ZFLAG_HASH;
2169 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2172 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2173 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2175 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2176 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2177 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2178 keg->uk_ipers, pages));
2182 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2183 * the keg onto the global keg list.
2185 * Arguments/Returns follow uma_ctor specifications
2186 * udata Actually uma_kctor_args
2189 keg_ctor(void *mem, int size, void *udata, int flags)
2191 struct uma_kctor_args *arg = udata;
2192 uma_keg_t keg = mem;
2197 keg->uk_size = arg->size;
2198 keg->uk_init = arg->uminit;
2199 keg->uk_fini = arg->fini;
2200 keg->uk_align = arg->align;
2201 keg->uk_reserve = 0;
2202 keg->uk_flags = arg->flags;
2205 * We use a global round-robin policy by default. Zones with
2206 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2207 * case the iterator is never run.
2209 keg->uk_dr.dr_policy = DOMAINSET_RR();
2210 keg->uk_dr.dr_iter = 0;
2213 * The master zone is passed to us at keg-creation time.
2216 keg->uk_name = zone->uz_name;
2218 if (arg->flags & UMA_ZONE_ZINIT)
2219 keg->uk_init = zero_init;
2221 if (arg->flags & UMA_ZONE_MALLOC)
2222 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2225 keg->uk_flags &= ~UMA_ZONE_PCPU;
2231 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2232 * work on. Use round-robin for everything else.
2234 * Zones may override the default by specifying either.
2237 if ((keg->uk_flags &
2238 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2239 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2240 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2241 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2245 * If we haven't booted yet we need allocations to go through the
2246 * startup cache until the vm is ready.
2248 #ifdef UMA_MD_SMALL_ALLOC
2249 if (keg->uk_ppera == 1)
2250 keg->uk_allocf = uma_small_alloc;
2253 if (booted < BOOT_KVA)
2254 keg->uk_allocf = startup_alloc;
2255 else if (keg->uk_flags & UMA_ZONE_PCPU)
2256 keg->uk_allocf = pcpu_page_alloc;
2257 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2258 keg->uk_allocf = contig_alloc;
2260 keg->uk_allocf = page_alloc;
2261 #ifdef UMA_MD_SMALL_ALLOC
2262 if (keg->uk_ppera == 1)
2263 keg->uk_freef = uma_small_free;
2266 if (keg->uk_flags & UMA_ZONE_PCPU)
2267 keg->uk_freef = pcpu_page_free;
2269 keg->uk_freef = page_free;
2272 * Initialize keg's locks.
2274 for (i = 0; i < vm_ndomains; i++)
2275 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2278 * If we're putting the slab header in the actual page we need to
2279 * figure out where in each page it goes. See slab_sizeof
2282 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2285 shsize = slab_sizeof(keg->uk_ipers);
2286 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2288 * The only way the following is possible is if with our
2289 * UMA_ALIGN_PTR adjustments we are now bigger than
2290 * UMA_SLAB_SIZE. I haven't checked whether this is
2291 * mathematically possible for all cases, so we make
2294 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2295 ("zone %s ipers %d rsize %d size %d slab won't fit",
2296 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2299 if (keg->uk_flags & UMA_ZFLAG_HASH)
2300 hash_alloc(&keg->uk_hash, 0);
2302 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2304 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2306 rw_wlock(&uma_rwlock);
2307 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2308 rw_wunlock(&uma_rwlock);
2313 zone_kva_available(uma_zone_t zone, void *unused)
2317 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2321 if (keg->uk_allocf == startup_alloc) {
2322 /* Switch to the real allocator. */
2323 if (keg->uk_flags & UMA_ZONE_PCPU)
2324 keg->uk_allocf = pcpu_page_alloc;
2325 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2327 keg->uk_allocf = contig_alloc;
2329 keg->uk_allocf = page_alloc;
2334 zone_alloc_counters(uma_zone_t zone, void *unused)
2337 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2338 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2339 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2340 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2344 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2346 uma_zone_domain_t zdom;
2349 struct sysctl_oid *oid, *domainoid;
2350 int domains, i, cnt;
2351 static const char *nokeg = "cache zone";
2355 * Make a sysctl safe copy of the zone name by removing
2356 * any special characters and handling dups by appending
2359 if (zone->uz_namecnt != 0) {
2360 /* Count the number of decimal digits and '_' separator. */
2361 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2363 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2365 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2368 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2369 for (c = zone->uz_ctlname; *c != '\0'; c++)
2370 if (strchr("./\\ -", *c) != NULL)
2374 * Basic parameters at the root.
2376 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2377 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2379 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2380 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2381 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2382 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2383 zone, 0, sysctl_handle_uma_zone_flags, "A",
2384 "Allocator configuration flags");
2385 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2386 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2387 "Desired per-cpu cache size");
2388 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2389 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2390 "Maximum allowed per-cpu cache size");
2395 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2396 domains = vm_ndomains;
2399 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2400 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2402 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2403 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2404 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2405 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2406 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2407 "Real object size with alignment");
2408 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2409 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2410 "pages per-slab allocation");
2411 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2412 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2413 "items available per-slab");
2414 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2415 "align", CTLFLAG_RD, &keg->uk_align, 0,
2416 "item alignment mask");
2417 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2418 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2419 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2420 "Slab utilization (100 - internal fragmentation %)");
2421 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2422 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2423 for (i = 0; i < domains; i++) {
2424 dom = &keg->uk_domain[i];
2425 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2426 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2427 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2428 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2429 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2430 "Total pages currently allocated from VM");
2431 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2432 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2433 "items free in the slab layer");
2436 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2437 "name", CTLFLAG_RD, nokeg, "Keg name");
2440 * Information about zone limits.
2442 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2443 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2444 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2445 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2446 zone, 0, sysctl_handle_uma_zone_items, "QU",
2447 "current number of allocated items if limit is set");
2448 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2449 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2450 "Maximum number of cached items");
2451 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2452 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2453 "Number of threads sleeping at limit");
2454 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2455 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2456 "Total zone limit sleeps");
2457 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2458 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2459 "Maximum number of items in each domain's bucket cache");
2462 * Per-domain zone information.
2464 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2465 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2466 for (i = 0; i < domains; i++) {
2467 zdom = ZDOM_GET(zone, i);
2468 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2469 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2470 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2471 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2472 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2473 "number of items in this domain");
2474 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2475 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2476 "maximum item count in this period");
2477 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2478 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2479 "minimum item count in this period");
2480 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2481 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2482 "Working set size");
2486 * General statistics.
2488 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2489 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2490 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2491 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2492 zone, 1, sysctl_handle_uma_zone_cur, "I",
2493 "Current number of allocated items");
2494 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2495 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2496 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2497 "Total allocation calls");
2498 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2499 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2500 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2501 "Total free calls");
2502 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2503 "fails", CTLFLAG_RD, &zone->uz_fails,
2504 "Number of allocation failures");
2505 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2506 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2507 "Free calls from the wrong domain");
2510 struct uma_zone_count {
2516 zone_count(uma_zone_t zone, void *arg)
2518 struct uma_zone_count *cnt;
2522 * Some zones are rapidly created with identical names and
2523 * destroyed out of order. This can lead to gaps in the count.
2524 * Use one greater than the maximum observed for this name.
2526 if (strcmp(zone->uz_name, cnt->name) == 0)
2527 cnt->count = MAX(cnt->count,
2528 zone->uz_namecnt + 1);
2532 zone_update_caches(uma_zone_t zone)
2536 for (i = 0; i <= mp_maxid; i++) {
2537 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2538 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2543 * Zone header ctor. This initializes all fields, locks, etc.
2545 * Arguments/Returns follow uma_ctor specifications
2546 * udata Actually uma_zctor_args
2549 zone_ctor(void *mem, int size, void *udata, int flags)
2551 struct uma_zone_count cnt;
2552 struct uma_zctor_args *arg = udata;
2553 uma_zone_domain_t zdom;
2554 uma_zone_t zone = mem;
2560 zone->uz_name = arg->name;
2561 zone->uz_ctor = arg->ctor;
2562 zone->uz_dtor = arg->dtor;
2563 zone->uz_init = NULL;
2564 zone->uz_fini = NULL;
2565 zone->uz_sleeps = 0;
2566 zone->uz_bucket_size = 0;
2567 zone->uz_bucket_size_min = 0;
2568 zone->uz_bucket_size_max = BUCKET_MAX;
2569 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2570 zone->uz_warning = NULL;
2571 /* The domain structures follow the cpu structures. */
2572 zone->uz_bucket_max = ULONG_MAX;
2573 timevalclear(&zone->uz_ratecheck);
2575 /* Count the number of duplicate names. */
2576 cnt.name = arg->name;
2578 zone_foreach(zone_count, &cnt);
2579 zone->uz_namecnt = cnt.count;
2580 ZONE_CROSS_LOCK_INIT(zone);
2582 for (i = 0; i < vm_ndomains; i++) {
2583 zdom = ZDOM_GET(zone, i);
2584 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2585 STAILQ_INIT(&zdom->uzd_buckets);
2589 if (arg->uminit == trash_init && arg->fini == trash_fini)
2590 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2594 * This is a pure cache zone, no kegs.
2597 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2598 ("zone_ctor: Import specified for non-cache zone."));
2599 zone->uz_flags = arg->flags;
2600 zone->uz_size = arg->size;
2601 zone->uz_import = arg->import;
2602 zone->uz_release = arg->release;
2603 zone->uz_arg = arg->arg;
2606 * Cache zones are round-robin unless a policy is
2607 * specified because they may have incompatible
2610 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2611 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2613 rw_wlock(&uma_rwlock);
2614 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2615 rw_wunlock(&uma_rwlock);
2620 * Use the regular zone/keg/slab allocator.
2622 zone->uz_import = zone_import;
2623 zone->uz_release = zone_release;
2624 zone->uz_arg = zone;
2627 if (arg->flags & UMA_ZONE_SECONDARY) {
2628 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2629 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2630 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2631 zone->uz_init = arg->uminit;
2632 zone->uz_fini = arg->fini;
2633 zone->uz_flags |= UMA_ZONE_SECONDARY;
2634 rw_wlock(&uma_rwlock);
2636 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2637 if (LIST_NEXT(z, uz_link) == NULL) {
2638 LIST_INSERT_AFTER(z, zone, uz_link);
2643 rw_wunlock(&uma_rwlock);
2644 } else if (keg == NULL) {
2645 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2646 arg->align, arg->flags)) == NULL)
2649 struct uma_kctor_args karg;
2652 /* We should only be here from uma_startup() */
2653 karg.size = arg->size;
2654 karg.uminit = arg->uminit;
2655 karg.fini = arg->fini;
2656 karg.align = arg->align;
2657 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2659 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2665 /* Inherit properties from the keg. */
2667 zone->uz_size = keg->uk_size;
2668 zone->uz_flags |= (keg->uk_flags &
2669 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2672 if (__predict_true(booted >= BOOT_RUNNING)) {
2673 zone_alloc_counters(zone, NULL);
2674 zone_alloc_sysctl(zone, NULL);
2676 zone->uz_allocs = EARLY_COUNTER;
2677 zone->uz_frees = EARLY_COUNTER;
2678 zone->uz_fails = EARLY_COUNTER;
2681 /* Caller requests a private SMR context. */
2682 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2683 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2685 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2686 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2687 ("Invalid zone flag combination"));
2688 if (arg->flags & UMA_ZFLAG_INTERNAL)
2689 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2690 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2691 zone->uz_bucket_size = BUCKET_MAX;
2692 else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2693 zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2694 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2695 zone->uz_bucket_size = 0;
2697 zone->uz_bucket_size = bucket_select(zone->uz_size);
2698 zone->uz_bucket_size_min = zone->uz_bucket_size;
2699 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2700 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2701 zone_update_caches(zone);
2707 * Keg header dtor. This frees all data, destroys locks, frees the hash
2708 * table and removes the keg from the global list.
2710 * Arguments/Returns follow uma_dtor specifications
2714 keg_dtor(void *arg, int size, void *udata)
2717 uint32_t free, pages;
2720 keg = (uma_keg_t)arg;
2722 for (i = 0; i < vm_ndomains; i++) {
2723 free += keg->uk_domain[i].ud_free_items;
2724 pages += keg->uk_domain[i].ud_pages;
2725 KEG_LOCK_FINI(keg, i);
2728 printf("Freed UMA keg (%s) was not empty (%u items). "
2729 " Lost %u pages of memory.\n",
2730 keg->uk_name ? keg->uk_name : "",
2731 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2733 hash_free(&keg->uk_hash);
2739 * Arguments/Returns follow uma_dtor specifications
2743 zone_dtor(void *arg, int size, void *udata)
2749 zone = (uma_zone_t)arg;
2751 sysctl_remove_oid(zone->uz_oid, 1, 1);
2753 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2756 rw_wlock(&uma_rwlock);
2757 LIST_REMOVE(zone, uz_link);
2758 rw_wunlock(&uma_rwlock);
2759 zone_reclaim(zone, M_WAITOK, true);
2762 * We only destroy kegs from non secondary/non cache zones.
2764 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2766 rw_wlock(&uma_rwlock);
2767 LIST_REMOVE(keg, uk_link);
2768 rw_wunlock(&uma_rwlock);
2769 zone_free_item(kegs, keg, NULL, SKIP_NONE);
2771 counter_u64_free(zone->uz_allocs);
2772 counter_u64_free(zone->uz_frees);
2773 counter_u64_free(zone->uz_fails);
2774 counter_u64_free(zone->uz_xdomain);
2775 free(zone->uz_ctlname, M_UMA);
2776 for (i = 0; i < vm_ndomains; i++)
2777 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2778 ZONE_CROSS_LOCK_FINI(zone);
2782 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2787 LIST_FOREACH(keg, &uma_kegs, uk_link) {
2788 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2791 LIST_FOREACH(zone, &uma_cachezones, uz_link)
2796 * Traverses every zone in the system and calls a callback
2799 * zfunc A pointer to a function which accepts a zone
2806 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2809 rw_rlock(&uma_rwlock);
2810 zone_foreach_unlocked(zfunc, arg);
2811 rw_runlock(&uma_rwlock);
2815 * Initialize the kernel memory allocator. This is done after pages can be
2816 * allocated but before general KVA is available.
2819 uma_startup1(vm_offset_t virtual_avail)
2821 struct uma_zctor_args args;
2822 size_t ksize, zsize, size;
2823 uma_keg_t masterkeg;
2827 bootstart = bootmem = virtual_avail;
2829 rw_init(&uma_rwlock, "UMA lock");
2830 sx_init(&uma_reclaim_lock, "umareclaim");
2832 ksize = sizeof(struct uma_keg) +
2833 (sizeof(struct uma_domain) * vm_ndomains);
2834 ksize = roundup(ksize, UMA_SUPER_ALIGN);
2835 zsize = sizeof(struct uma_zone) +
2836 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2837 (sizeof(struct uma_zone_domain) * vm_ndomains);
2838 zsize = roundup(zsize, UMA_SUPER_ALIGN);
2840 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2841 size = (zsize * 2) + ksize;
2842 m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO);
2843 zones = (uma_zone_t)m;
2845 kegs = (uma_zone_t)m;
2847 masterkeg = (uma_keg_t)m;
2849 /* "manually" create the initial zone */
2850 memset(&args, 0, sizeof(args));
2851 args.name = "UMA Kegs";
2853 args.ctor = keg_ctor;
2854 args.dtor = keg_dtor;
2855 args.uminit = zero_init;
2857 args.keg = masterkeg;
2858 args.align = UMA_SUPER_ALIGN - 1;
2859 args.flags = UMA_ZFLAG_INTERNAL;
2860 zone_ctor(kegs, zsize, &args, M_WAITOK);
2862 args.name = "UMA Zones";
2864 args.ctor = zone_ctor;
2865 args.dtor = zone_dtor;
2866 args.uminit = zero_init;
2869 args.align = UMA_SUPER_ALIGN - 1;
2870 args.flags = UMA_ZFLAG_INTERNAL;
2871 zone_ctor(zones, zsize, &args, M_WAITOK);
2873 /* Now make zones for slab headers */
2874 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2875 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2876 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2877 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2879 hashzone = uma_zcreate("UMA Hash",
2880 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2881 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2887 #ifndef UMA_MD_SMALL_ALLOC
2888 extern void vm_radix_reserve_kva(void);
2892 * Advertise the availability of normal kva allocations and switch to
2893 * the default back-end allocator. Marks the KVA we consumed on startup
2894 * as used in the map.
2900 if (bootstart != bootmem) {
2901 vm_map_lock(kernel_map);
2902 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2903 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2904 vm_map_unlock(kernel_map);
2907 #ifndef UMA_MD_SMALL_ALLOC
2908 /* Set up radix zone to use noobj_alloc. */
2909 vm_radix_reserve_kva();
2913 zone_foreach_unlocked(zone_kva_available, NULL);
2918 * Finish our initialization steps.
2925 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2926 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2927 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2929 zone_foreach_unlocked(zone_alloc_counters, NULL);
2930 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2931 callout_init(&uma_callout, 1);
2932 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2933 booted = BOOT_RUNNING;
2935 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2936 EVENTHANDLER_PRI_FIRST);
2943 booted = BOOT_SHUTDOWN;
2947 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2948 int align, uint32_t flags)
2950 struct uma_kctor_args args;
2953 args.uminit = uminit;
2955 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2958 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2961 /* Public functions */
2964 uma_set_align(int align)
2967 if (align != UMA_ALIGN_CACHE)
2968 uma_align_cache = align;
2973 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2974 uma_init uminit, uma_fini fini, int align, uint32_t flags)
2977 struct uma_zctor_args args;
2980 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2983 /* This stuff is essential for the zone ctor */
2984 memset(&args, 0, sizeof(args));
2989 args.uminit = uminit;
2993 * Inject procedures which check for memory use after free if we are
2994 * allowed to scramble the memory while it is not allocated. This
2995 * requires that: UMA is actually able to access the memory, no init
2996 * or fini procedures, no dependency on the initial value of the
2997 * memory, and no (legitimate) use of the memory after free. Note,
2998 * the ctor and dtor do not need to be empty.
3000 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3001 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3002 args.uminit = trash_init;
3003 args.fini = trash_fini;
3010 sx_slock(&uma_reclaim_lock);
3011 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3012 sx_sunlock(&uma_reclaim_lock);
3019 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3020 uma_init zinit, uma_fini zfini, uma_zone_t master)
3022 struct uma_zctor_args args;
3026 keg = master->uz_keg;
3027 memset(&args, 0, sizeof(args));
3029 args.size = keg->uk_size;
3032 args.uminit = zinit;
3034 args.align = keg->uk_align;
3035 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3038 sx_slock(&uma_reclaim_lock);
3039 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3040 sx_sunlock(&uma_reclaim_lock);
3047 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3048 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3049 void *arg, int flags)
3051 struct uma_zctor_args args;
3053 memset(&args, 0, sizeof(args));
3058 args.uminit = zinit;
3060 args.import = zimport;
3061 args.release = zrelease;
3064 args.flags = flags | UMA_ZFLAG_CACHE;
3066 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3071 uma_zdestroy(uma_zone_t zone)
3075 * Large slabs are expensive to reclaim, so don't bother doing
3076 * unnecessary work if we're shutting down.
3078 if (booted == BOOT_SHUTDOWN &&
3079 zone->uz_fini == NULL && zone->uz_release == zone_release)
3081 sx_slock(&uma_reclaim_lock);
3082 zone_free_item(zones, zone, NULL, SKIP_NONE);
3083 sx_sunlock(&uma_reclaim_lock);
3087 uma_zwait(uma_zone_t zone)
3090 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3091 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3092 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3093 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3095 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3099 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3101 void *item, *pcpu_item;
3105 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3107 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3110 pcpu_item = zpcpu_base_to_offset(item);
3111 if (flags & M_ZERO) {
3113 for (i = 0; i <= mp_maxid; i++)
3114 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3116 bzero(item, zone->uz_size);
3123 * A stub while both regular and pcpu cases are identical.
3126 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3131 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3133 item = zpcpu_offset_to_base(pcpu_item);
3134 uma_zfree_arg(zone, item, udata);
3137 static inline void *
3138 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3144 skipdbg = uma_dbg_zskip(zone, item);
3145 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3146 zone->uz_ctor != trash_ctor)
3147 trash_ctor(item, size, udata, flags);
3149 /* Check flags before loading ctor pointer. */
3150 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3151 __predict_false(zone->uz_ctor != NULL) &&
3152 zone->uz_ctor(item, size, udata, flags) != 0) {
3153 counter_u64_add(zone->uz_fails, 1);
3154 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3159 uma_dbg_alloc(zone, NULL, item);
3161 if (__predict_false(flags & M_ZERO))
3162 return (memset(item, 0, size));
3168 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3169 enum zfreeskip skip)
3174 skipdbg = uma_dbg_zskip(zone, item);
3175 if (skip == SKIP_NONE && !skipdbg) {
3176 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3177 uma_dbg_free(zone, udata, item);
3179 uma_dbg_free(zone, NULL, item);
3182 if (__predict_true(skip < SKIP_DTOR)) {
3183 if (zone->uz_dtor != NULL)
3184 zone->uz_dtor(item, size, udata);
3186 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3187 zone->uz_dtor != trash_dtor)
3188 trash_dtor(item, size, udata);
3193 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3194 #define UMA_ZALLOC_DEBUG
3196 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3202 if (flags & M_WAITOK) {
3203 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3204 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3209 KASSERT((flags & M_EXEC) == 0,
3210 ("uma_zalloc_debug: called with M_EXEC"));
3211 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3212 ("uma_zalloc_debug: called within spinlock or critical section"));
3213 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3214 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3217 #ifdef DEBUG_MEMGUARD
3218 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3220 item = memguard_alloc(zone->uz_size, flags);
3222 error = EJUSTRETURN;
3223 if (zone->uz_init != NULL &&
3224 zone->uz_init(item, zone->uz_size, flags) != 0) {
3228 if (zone->uz_ctor != NULL &&
3229 zone->uz_ctor(item, zone->uz_size, udata,
3231 counter_u64_add(zone->uz_fails, 1);
3232 zone->uz_fini(item, zone->uz_size);
3239 /* This is unfortunate but should not be fatal. */
3246 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3248 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3249 ("uma_zfree_debug: called with spinlock or critical section held"));
3251 #ifdef DEBUG_MEMGUARD
3252 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3253 if (zone->uz_dtor != NULL)
3254 zone->uz_dtor(item, zone->uz_size, udata);
3255 if (zone->uz_fini != NULL)
3256 zone->uz_fini(item, zone->uz_size);
3257 memguard_free(item);
3258 return (EJUSTRETURN);
3265 static inline void *
3266 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3267 void *udata, int flags)
3272 item = cache_bucket_pop(cache, bucket);
3273 size = cache_uz_size(cache);
3274 uz_flags = cache_uz_flags(cache);
3276 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3279 static __noinline void *
3280 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3282 uma_cache_bucket_t bucket;
3285 while (cache_alloc(zone, cache, udata, flags)) {
3286 cache = &zone->uz_cpu[curcpu];
3287 bucket = &cache->uc_allocbucket;
3288 if (__predict_false(bucket->ucb_cnt == 0))
3290 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3295 * We can not get a bucket so try to return a single item.
3297 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3298 domain = PCPU_GET(domain);
3300 domain = UMA_ANYDOMAIN;
3301 return (zone_alloc_item(zone, udata, domain, flags));
3306 uma_zalloc_smr(uma_zone_t zone, int flags)
3308 uma_cache_bucket_t bucket;
3311 #ifdef UMA_ZALLOC_DEBUG
3314 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3315 ("uma_zalloc_arg: called with non-SMR zone.\n"));
3316 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3321 cache = &zone->uz_cpu[curcpu];
3322 bucket = &cache->uc_allocbucket;
3323 if (__predict_false(bucket->ucb_cnt == 0))
3324 return (cache_alloc_retry(zone, cache, NULL, flags));
3325 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3330 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3332 uma_cache_bucket_t bucket;
3335 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3336 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3338 /* This is the fast path allocation */
3339 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3342 #ifdef UMA_ZALLOC_DEBUG
3345 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3346 ("uma_zalloc_arg: called with SMR zone.\n"));
3347 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3352 * If possible, allocate from the per-CPU cache. There are two
3353 * requirements for safe access to the per-CPU cache: (1) the thread
3354 * accessing the cache must not be preempted or yield during access,
3355 * and (2) the thread must not migrate CPUs without switching which
3356 * cache it accesses. We rely on a critical section to prevent
3357 * preemption and migration. We release the critical section in
3358 * order to acquire the zone mutex if we are unable to allocate from
3359 * the current cache; when we re-acquire the critical section, we
3360 * must detect and handle migration if it has occurred.
3363 cache = &zone->uz_cpu[curcpu];
3364 bucket = &cache->uc_allocbucket;
3365 if (__predict_false(bucket->ucb_cnt == 0))
3366 return (cache_alloc_retry(zone, cache, udata, flags));
3367 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3371 * Replenish an alloc bucket and possibly restore an old one. Called in
3372 * a critical section. Returns in a critical section.
3374 * A false return value indicates an allocation failure.
3375 * A true return value indicates success and the caller should retry.
3377 static __noinline bool
3378 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3380 uma_bucket_t bucket;
3384 CRITICAL_ASSERT(curthread);
3387 * If we have run out of items in our alloc bucket see
3388 * if we can switch with the free bucket.
3390 * SMR Zones can't re-use the free bucket until the sequence has
3393 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3394 cache->uc_freebucket.ucb_cnt != 0) {
3395 cache_bucket_swap(&cache->uc_freebucket,
3396 &cache->uc_allocbucket);
3401 * Discard any empty allocation bucket while we hold no locks.
3403 bucket = cache_bucket_unload_alloc(cache);
3406 if (bucket != NULL) {
3407 KASSERT(bucket->ub_cnt == 0,
3408 ("cache_alloc: Entered with non-empty alloc bucket."));
3409 bucket_free(zone, bucket, udata);
3412 /* Short-circuit for zones without buckets and low memory. */
3413 if (zone->uz_bucket_size == 0 || bucketdisable) {
3419 * Attempt to retrieve the item from the per-CPU cache has failed, so
3420 * we must go back to the zone. This requires the zdom lock, so we
3421 * must drop the critical section, then re-acquire it when we go back
3422 * to the cache. Since the critical section is released, we may be
3423 * preempted or migrate. As such, make sure not to maintain any
3424 * thread-local state specific to the cache from prior to releasing
3425 * the critical section.
3427 domain = PCPU_GET(domain);
3428 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0)
3429 domain = zone_domain_highest(zone, domain);
3430 bucket = cache_fetch_bucket(zone, cache, domain);
3431 if (bucket == NULL) {
3432 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3437 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3438 zone->uz_name, zone, bucket);
3439 if (bucket == NULL) {
3445 * See if we lost the race or were migrated. Cache the
3446 * initialized bucket to make this less likely or claim
3447 * the memory directly.
3450 cache = &zone->uz_cpu[curcpu];
3451 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3452 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3453 domain == PCPU_GET(domain))) {
3455 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3457 cache_bucket_load_alloc(cache, bucket);
3462 * We lost the race, release this bucket and start over.
3465 zone_put_bucket(zone, domain, bucket, udata, false);
3472 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3475 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3476 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3478 /* This is the fast path allocation */
3479 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3480 zone->uz_name, zone, domain, flags);
3482 if (flags & M_WAITOK) {
3483 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3484 "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3486 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3487 ("uma_zalloc_domain: called with spinlock or critical section held"));
3489 return (zone_alloc_item(zone, udata, domain, flags));
3493 * Find a slab with some space. Prefer slabs that are partially used over those
3494 * that are totally full. This helps to reduce fragmentation.
3496 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3500 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3506 KASSERT(domain >= 0 && domain < vm_ndomains,
3507 ("keg_first_slab: domain %d out of range", domain));
3508 KEG_LOCK_ASSERT(keg, domain);
3513 dom = &keg->uk_domain[domain];
3514 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3516 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3517 LIST_REMOVE(slab, us_link);
3518 dom->ud_free_slabs--;
3519 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3523 domain = (domain + 1) % vm_ndomains;
3524 } while (domain != start);
3530 * Fetch an existing slab from a free or partial list. Returns with the
3531 * keg domain lock held if a slab was found or unlocked if not.
3534 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3539 /* HASH has a single free list. */
3540 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3543 KEG_LOCK(keg, domain);
3544 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3545 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3546 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3547 KEG_UNLOCK(keg, domain);
3554 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3556 struct vm_domainset_iter di;
3563 * Use the keg's policy if upper layers haven't already specified a
3564 * domain (as happens with first-touch zones).
3566 * To avoid races we run the iterator with the keg lock held, but that
3567 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3568 * clear M_WAITOK and handle low memory conditions locally.
3570 rr = rdomain == UMA_ANYDOMAIN;
3572 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3573 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3581 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3586 * M_NOVM means don't ask at all!
3591 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3594 if (!rr && (flags & M_WAITOK) == 0)
3596 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3597 if ((flags & M_WAITOK) != 0) {
3598 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
3606 * We might not have been able to get a slab but another cpu
3607 * could have while we were unlocked. Check again before we
3610 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3617 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3623 KEG_LOCK_ASSERT(keg, slab->us_domain);
3625 dom = &keg->uk_domain[slab->us_domain];
3626 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3627 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3628 item = slab_item(slab, keg, freei);
3629 slab->us_freecount--;
3630 dom->ud_free_items--;
3633 * Move this slab to the full list. It must be on the partial list, so
3634 * we do not need to update the free slab count. In particular,
3635 * keg_fetch_slab() always returns slabs on the partial list.
3637 if (slab->us_freecount == 0) {
3638 LIST_REMOVE(slab, us_link);
3639 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3646 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3660 /* Try to keep the buckets totally full */
3661 for (i = 0; i < max; ) {
3662 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3665 stripe = howmany(max, vm_ndomains);
3667 dom = &keg->uk_domain[slab->us_domain];
3668 while (slab->us_freecount && i < max) {
3669 bucket[i++] = slab_alloc_item(keg, slab);
3670 if (dom->ud_free_items <= keg->uk_reserve)
3674 * If the zone is striped we pick a new slab for every
3675 * N allocations. Eliminating this conditional will
3676 * instead pick a new domain for each bucket rather
3677 * than stripe within each bucket. The current option
3678 * produces more fragmentation and requires more cpu
3679 * time but yields better distribution.
3681 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3682 vm_ndomains > 1 && --stripe == 0)
3686 KEG_UNLOCK(keg, slab->us_domain);
3687 /* Don't block if we allocated any successfully. */
3696 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3698 uint64_t old, new, total, max;
3701 * The hard case. We're going to sleep because there were existing
3702 * sleepers or because we ran out of items. This routine enforces
3703 * fairness by keeping fifo order.
3705 * First release our ill gotten gains and make some noise.
3708 zone_free_limit(zone, count);
3709 zone_log_warning(zone);
3710 zone_maxaction(zone);
3711 if (flags & M_NOWAIT)
3715 * We need to allocate an item or set ourself as a sleeper
3716 * while the sleepq lock is held to avoid wakeup races. This
3717 * is essentially a home rolled semaphore.
3719 sleepq_lock(&zone->uz_max_items);
3720 old = zone->uz_items;
3722 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3723 /* Cache the max since we will evaluate twice. */
3724 max = zone->uz_max_items;
3725 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3726 UZ_ITEMS_COUNT(old) >= max)
3727 new = old + UZ_ITEMS_SLEEPER;
3729 new = old + MIN(count, max - old);
3730 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3732 /* We may have successfully allocated under the sleepq lock. */
3733 if (UZ_ITEMS_SLEEPERS(new) == 0) {
3734 sleepq_release(&zone->uz_max_items);
3739 * This is in a different cacheline from uz_items so that we
3740 * don't constantly invalidate the fastpath cacheline when we
3741 * adjust item counts. This could be limited to toggling on
3744 atomic_add_32(&zone->uz_sleepers, 1);
3745 atomic_add_64(&zone->uz_sleeps, 1);
3748 * We have added ourselves as a sleeper. The sleepq lock
3749 * protects us from wakeup races. Sleep now and then retry.
3751 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3752 sleepq_wait(&zone->uz_max_items, PVM);
3755 * After wakeup, remove ourselves as a sleeper and try
3756 * again. We no longer have the sleepq lock for protection.
3758 * Subract ourselves as a sleeper while attempting to add
3761 atomic_subtract_32(&zone->uz_sleepers, 1);
3762 old = atomic_fetchadd_64(&zone->uz_items,
3763 -(UZ_ITEMS_SLEEPER - count));
3764 /* We're no longer a sleeper. */
3765 old -= UZ_ITEMS_SLEEPER;
3768 * If we're still at the limit, restart. Notably do not
3769 * block on other sleepers. Cache the max value to protect
3770 * against changes via sysctl.
3772 total = UZ_ITEMS_COUNT(old);
3773 max = zone->uz_max_items;
3776 /* Truncate if necessary, otherwise wake other sleepers. */
3777 if (total + count > max) {
3778 zone_free_limit(zone, total + count - max);
3779 count = max - total;
3780 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3781 wakeup_one(&zone->uz_max_items);
3788 * Allocate 'count' items from our max_items limit. Returns the number
3789 * available. If M_NOWAIT is not specified it will sleep until at least
3790 * one item can be allocated.
3793 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3798 max = zone->uz_max_items;
3802 * We expect normal allocations to succeed with a simple
3805 old = atomic_fetchadd_64(&zone->uz_items, count);
3806 if (__predict_true(old + count <= max))
3810 * If we had some items and no sleepers just return the
3811 * truncated value. We have to release the excess space
3812 * though because that may wake sleepers who weren't woken
3813 * because we were temporarily over the limit.
3816 zone_free_limit(zone, (old + count) - max);
3819 return (zone_alloc_limit_hard(zone, count, flags));
3823 * Free a number of items back to the limit.
3826 zone_free_limit(uma_zone_t zone, int count)
3833 * In the common case we either have no sleepers or
3834 * are still over the limit and can just return.
3836 old = atomic_fetchadd_64(&zone->uz_items, -count);
3837 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3838 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3842 * Moderate the rate of wakeups. Sleepers will continue
3843 * to generate wakeups if necessary.
3845 wakeup_one(&zone->uz_max_items);
3849 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3851 uma_bucket_t bucket;
3854 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3857 /* Avoid allocs targeting empty domains. */
3858 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3859 domain = UMA_ANYDOMAIN;
3860 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
3861 domain = UMA_ANYDOMAIN;
3863 if (zone->uz_max_items > 0)
3864 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3867 maxbucket = zone->uz_bucket_size;
3871 /* Don't wait for buckets, preserve caller's NOVM setting. */
3872 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3873 if (bucket == NULL) {
3878 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3879 MIN(maxbucket, bucket->ub_entries), domain, flags);
3882 * Initialize the memory if necessary.
3884 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3887 for (i = 0; i < bucket->ub_cnt; i++)
3888 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3892 * If we couldn't initialize the whole bucket, put the
3893 * rest back onto the freelist.
3895 if (i != bucket->ub_cnt) {
3896 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3897 bucket->ub_cnt - i);
3899 bzero(&bucket->ub_bucket[i],
3900 sizeof(void *) * (bucket->ub_cnt - i));
3906 cnt = bucket->ub_cnt;
3907 if (bucket->ub_cnt == 0) {
3908 bucket_free(zone, bucket, udata);
3909 counter_u64_add(zone->uz_fails, 1);
3913 if (zone->uz_max_items > 0 && cnt < maxbucket)
3914 zone_free_limit(zone, maxbucket - cnt);
3920 * Allocates a single item from a zone.
3923 * zone The zone to alloc for.
3924 * udata The data to be passed to the constructor.
3925 * domain The domain to allocate from or UMA_ANYDOMAIN.
3926 * flags M_WAITOK, M_NOWAIT, M_ZERO.
3929 * NULL if there is no memory and M_NOWAIT is set
3930 * An item if successful
3934 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
3938 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0)
3941 /* Avoid allocs targeting empty domains. */
3942 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3943 domain = UMA_ANYDOMAIN;
3945 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
3949 * We have to call both the zone's init (not the keg's init)
3950 * and the zone's ctor. This is because the item is going from
3951 * a keg slab directly to the user, and the user is expecting it
3952 * to be both zone-init'd as well as zone-ctor'd.
3954 if (zone->uz_init != NULL) {
3955 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
3956 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
3960 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
3965 counter_u64_add(zone->uz_allocs, 1);
3966 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
3967 zone->uz_name, zone);
3972 counter_u64_add(zone->uz_fails, 1);
3974 if (zone->uz_max_items > 0)
3975 zone_free_limit(zone, 1);
3976 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
3977 zone->uz_name, zone);
3984 uma_zfree_smr(uma_zone_t zone, void *item)
3987 uma_cache_bucket_t bucket;
3988 int itemdomain, uz_flags;
3990 #ifdef UMA_ZALLOC_DEBUG
3991 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3992 ("uma_zfree_smr: called with non-SMR zone.\n"));
3993 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
3994 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
3995 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
3998 cache = &zone->uz_cpu[curcpu];
3999 uz_flags = cache_uz_flags(cache);
4002 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4003 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4007 cache = &zone->uz_cpu[curcpu];
4008 /* SMR Zones must free to the free bucket. */
4009 bucket = &cache->uc_freebucket;
4011 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4012 PCPU_GET(domain) != itemdomain) {
4013 bucket = &cache->uc_crossbucket;
4016 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4017 cache_bucket_push(cache, bucket, item);
4021 } while (cache_free(zone, cache, NULL, item, itemdomain));
4025 * If nothing else caught this, we'll just do an internal free.
4027 zone_free_item(zone, item, NULL, SKIP_NONE);
4032 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4035 uma_cache_bucket_t bucket;
4036 int itemdomain, uz_flags;
4038 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4039 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4041 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4043 #ifdef UMA_ZALLOC_DEBUG
4044 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4045 ("uma_zfree_arg: called with SMR zone.\n"));
4046 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4049 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4054 * We are accessing the per-cpu cache without a critical section to
4055 * fetch size and flags. This is acceptable, if we are preempted we
4056 * will simply read another cpu's line.
4058 cache = &zone->uz_cpu[curcpu];
4059 uz_flags = cache_uz_flags(cache);
4060 if (UMA_ALWAYS_CTORDTOR ||
4061 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4062 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4065 * The race here is acceptable. If we miss it we'll just have to wait
4066 * a little longer for the limits to be reset.
4068 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4069 if (zone->uz_sleepers > 0)
4074 * If possible, free to the per-CPU cache. There are two
4075 * requirements for safe access to the per-CPU cache: (1) the thread
4076 * accessing the cache must not be preempted or yield during access,
4077 * and (2) the thread must not migrate CPUs without switching which
4078 * cache it accesses. We rely on a critical section to prevent
4079 * preemption and migration. We release the critical section in
4080 * order to acquire the zone mutex if we are unable to free to the
4081 * current cache; when we re-acquire the critical section, we must
4082 * detect and handle migration if it has occurred.
4086 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4087 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4091 cache = &zone->uz_cpu[curcpu];
4093 * Try to free into the allocbucket first to give LIFO
4094 * ordering for cache-hot datastructures. Spill over
4095 * into the freebucket if necessary. Alloc will swap
4096 * them if one runs dry.
4098 bucket = &cache->uc_allocbucket;
4100 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4101 PCPU_GET(domain) != itemdomain) {
4102 bucket = &cache->uc_crossbucket;
4105 if (bucket->ucb_cnt >= bucket->ucb_entries)
4106 bucket = &cache->uc_freebucket;
4107 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4108 cache_bucket_push(cache, bucket, item);
4112 } while (cache_free(zone, cache, udata, item, itemdomain));
4116 * If nothing else caught this, we'll just do an internal free.
4119 zone_free_item(zone, item, udata, SKIP_DTOR);
4124 * sort crossdomain free buckets to domain correct buckets and cache
4128 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4130 struct uma_bucketlist fullbuckets;
4131 uma_zone_domain_t zdom;
4138 "uma_zfree: zone %s(%p) draining cross bucket %p",
4139 zone->uz_name, zone, bucket);
4142 * It is possible for buckets to arrive here out of order so we fetch
4143 * the current smr seq rather than accepting the bucket's.
4145 seq = SMR_SEQ_INVALID;
4146 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4147 seq = smr_advance(zone->uz_smr);
4150 * To avoid having ndomain * ndomain buckets for sorting we have a
4151 * lock on the current crossfree bucket. A full matrix with
4152 * per-domain locking could be used if necessary.
4154 STAILQ_INIT(&fullbuckets);
4155 ZONE_CROSS_LOCK(zone);
4156 while (bucket->ub_cnt > 0) {
4157 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4158 domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4159 zdom = ZDOM_GET(zone, domain);
4160 if (zdom->uzd_cross == NULL) {
4161 zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
4162 if (zdom->uzd_cross == NULL)
4165 b = zdom->uzd_cross;
4166 b->ub_bucket[b->ub_cnt++] = item;
4168 if (b->ub_cnt == b->ub_entries) {
4169 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4170 zdom->uzd_cross = NULL;
4174 ZONE_CROSS_UNLOCK(zone);
4175 if (bucket->ub_cnt == 0)
4176 bucket->ub_seq = SMR_SEQ_INVALID;
4177 bucket_free(zone, bucket, udata);
4179 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4180 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4181 domain = _vm_phys_domain(pmap_kextract(
4182 (vm_offset_t)b->ub_bucket[0]));
4183 zone_put_bucket(zone, domain, b, udata, true);
4189 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4190 int itemdomain, bool ws)
4195 * Buckets coming from the wrong domain will be entirely for the
4196 * only other domain on two domain systems. In this case we can
4197 * simply cache them. Otherwise we need to sort them back to
4200 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4201 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4202 zone_free_cross(zone, bucket, udata);
4208 * Attempt to save the bucket in the zone's domain bucket cache.
4211 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4212 zone->uz_name, zone, bucket);
4213 /* ub_cnt is pointing to the last free item */
4214 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4215 itemdomain = zone_domain_lowest(zone, itemdomain);
4216 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4220 * Populate a free or cross bucket for the current cpu cache. Free any
4221 * existing full bucket either to the zone cache or back to the slab layer.
4223 * Enters and returns in a critical section. false return indicates that
4224 * we can not satisfy this free in the cache layer. true indicates that
4225 * the caller should retry.
4227 static __noinline bool
4228 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4231 uma_cache_bucket_t cbucket;
4232 uma_bucket_t newbucket, bucket;
4234 CRITICAL_ASSERT(curthread);
4236 if (zone->uz_bucket_size == 0)
4239 cache = &zone->uz_cpu[curcpu];
4243 * FIRSTTOUCH domains need to free to the correct zdom. When
4244 * enabled this is the zdom of the item. The bucket is the
4245 * cross bucket if the current domain and itemdomain do not match.
4247 cbucket = &cache->uc_freebucket;
4249 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4250 if (PCPU_GET(domain) != itemdomain) {
4251 cbucket = &cache->uc_crossbucket;
4252 if (cbucket->ucb_cnt != 0)
4253 counter_u64_add(zone->uz_xdomain,
4258 bucket = cache_bucket_unload(cbucket);
4259 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4260 ("cache_free: Entered with non-full free bucket."));
4262 /* We are no longer associated with this CPU. */
4266 * Don't let SMR zones operate without a free bucket. Force
4267 * a synchronize and re-use this one. We will only degrade
4268 * to a synchronize every bucket_size items rather than every
4269 * item if we fail to allocate a bucket.
4271 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4273 bucket->ub_seq = smr_advance(zone->uz_smr);
4274 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4275 if (newbucket == NULL && bucket != NULL) {
4276 bucket_drain(zone, bucket);
4280 } else if (!bucketdisable)
4281 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4284 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4287 if ((bucket = newbucket) == NULL)
4289 cache = &zone->uz_cpu[curcpu];
4292 * Check to see if we should be populating the cross bucket. If it
4293 * is already populated we will fall through and attempt to populate
4296 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4297 if (PCPU_GET(domain) != itemdomain &&
4298 cache->uc_crossbucket.ucb_bucket == NULL) {
4299 cache_bucket_load_cross(cache, bucket);
4305 * We may have lost the race to fill the bucket or switched CPUs.
4307 if (cache->uc_freebucket.ucb_bucket != NULL) {
4309 bucket_free(zone, bucket, udata);
4312 cache_bucket_load_free(cache, bucket);
4318 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
4321 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4322 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4324 CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone);
4326 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
4327 ("uma_zfree_domain: called with spinlock or critical section held"));
4329 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4332 zone_free_item(zone, item, udata, SKIP_NONE);
4336 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4343 KEG_LOCK_ASSERT(keg, slab->us_domain);
4345 /* Do we need to remove from any lists? */
4346 dom = &keg->uk_domain[slab->us_domain];
4347 if (slab->us_freecount + 1 == keg->uk_ipers) {
4348 LIST_REMOVE(slab, us_link);
4349 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4350 dom->ud_free_slabs++;
4351 } else if (slab->us_freecount == 0) {
4352 LIST_REMOVE(slab, us_link);
4353 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4356 /* Slab management. */
4357 freei = slab_item_index(slab, keg, item);
4358 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4359 slab->us_freecount++;
4361 /* Keg statistics. */
4362 dom->ud_free_items++;
4366 zone_release(void *arg, void **bucket, int cnt)
4379 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4380 lock = KEG_LOCK(keg, 0);
4381 for (i = 0; i < cnt; i++) {
4383 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4384 slab = vtoslab((vm_offset_t)item);
4386 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4387 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4388 slab = hash_sfind(&keg->uk_hash, mem);
4390 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4392 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4395 lock = KEG_LOCK(keg, slab->us_domain);
4397 slab_free_item(zone, slab, item);
4404 * Frees a single item to any zone.
4407 * zone The zone to free to
4408 * item The item we're freeing
4409 * udata User supplied data for the dtor
4410 * skip Skip dtors and finis
4412 static __noinline void
4413 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4417 * If a free is sent directly to an SMR zone we have to
4418 * synchronize immediately because the item can instantly
4419 * be reallocated. This should only happen in degenerate
4420 * cases when no memory is available for per-cpu caches.
4422 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4423 smr_synchronize(zone->uz_smr);
4425 item_dtor(zone, item, zone->uz_size, udata, skip);
4427 if (skip < SKIP_FINI && zone->uz_fini)
4428 zone->uz_fini(item, zone->uz_size);
4430 zone->uz_release(zone->uz_arg, &item, 1);
4432 if (skip & SKIP_CNT)
4435 counter_u64_add(zone->uz_frees, 1);
4437 if (zone->uz_max_items > 0)
4438 zone_free_limit(zone, 1);
4443 uma_zone_set_max(uma_zone_t zone, int nitems)
4445 struct uma_bucket_zone *ubz;
4449 * XXX This can misbehave if the zone has any allocations with
4450 * no limit and a limit is imposed. There is currently no
4451 * way to clear a limit.
4454 ubz = bucket_zone_max(zone, nitems);
4455 count = ubz != NULL ? ubz->ubz_entries : 0;
4456 zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4457 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4458 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4459 zone->uz_max_items = nitems;
4460 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4461 zone_update_caches(zone);
4462 /* We may need to wake waiters. */
4463 wakeup(&zone->uz_max_items);
4471 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4473 struct uma_bucket_zone *ubz;
4477 ubz = bucket_zone_max(zone, nitems);
4480 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4481 /* Count the cross-domain bucket. */
4483 nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4484 zone->uz_bucket_size_max = ubz->ubz_entries;
4486 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4488 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4489 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4490 zone->uz_bucket_max = nitems / vm_ndomains;
4496 uma_zone_get_max(uma_zone_t zone)
4500 nitems = atomic_load_64(&zone->uz_max_items);
4507 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4510 ZONE_ASSERT_COLD(zone);
4511 zone->uz_warning = warning;
4516 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4519 ZONE_ASSERT_COLD(zone);
4520 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4525 uma_zone_get_cur(uma_zone_t zone)
4531 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4532 nitems = counter_u64_fetch(zone->uz_allocs) -
4533 counter_u64_fetch(zone->uz_frees);
4535 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4536 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4538 return (nitems < 0 ? 0 : nitems);
4542 uma_zone_get_allocs(uma_zone_t zone)
4548 if (zone->uz_allocs != EARLY_COUNTER)
4549 nitems = counter_u64_fetch(zone->uz_allocs);
4551 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4557 uma_zone_get_frees(uma_zone_t zone)
4563 if (zone->uz_frees != EARLY_COUNTER)
4564 nitems = counter_u64_fetch(zone->uz_frees);
4566 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4572 /* Used only for KEG_ASSERT_COLD(). */
4574 uma_keg_get_allocs(uma_keg_t keg)
4580 LIST_FOREACH(z, &keg->uk_zones, uz_link)
4581 nitems += uma_zone_get_allocs(z);
4589 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4594 KEG_ASSERT_COLD(keg);
4595 keg->uk_init = uminit;
4600 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4605 KEG_ASSERT_COLD(keg);
4606 keg->uk_fini = fini;
4611 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4614 ZONE_ASSERT_COLD(zone);
4615 zone->uz_init = zinit;
4620 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4623 ZONE_ASSERT_COLD(zone);
4624 zone->uz_fini = zfini;
4629 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4634 KEG_ASSERT_COLD(keg);
4635 keg->uk_freef = freef;
4640 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4645 KEG_ASSERT_COLD(keg);
4646 keg->uk_allocf = allocf;
4651 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4654 ZONE_ASSERT_COLD(zone);
4656 zone->uz_flags |= UMA_ZONE_SMR;
4658 zone_update_caches(zone);
4662 uma_zone_get_smr(uma_zone_t zone)
4665 return (zone->uz_smr);
4670 uma_zone_reserve(uma_zone_t zone, int items)
4675 KEG_ASSERT_COLD(keg);
4676 keg->uk_reserve = items;
4681 uma_zone_reserve_kva(uma_zone_t zone, int count)
4688 KEG_ASSERT_COLD(keg);
4689 ZONE_ASSERT_COLD(zone);
4691 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4693 #ifdef UMA_MD_SMALL_ALLOC
4694 if (keg->uk_ppera > 1) {
4698 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4704 MPASS(keg->uk_kva == 0);
4707 zone->uz_max_items = pages * keg->uk_ipers;
4708 #ifdef UMA_MD_SMALL_ALLOC
4709 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4711 keg->uk_allocf = noobj_alloc;
4713 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4714 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4715 zone_update_caches(zone);
4722 uma_prealloc(uma_zone_t zone, int items)
4724 struct vm_domainset_iter di;
4728 int aflags, domain, slabs;
4731 slabs = howmany(items, keg->uk_ipers);
4732 while (slabs-- > 0) {
4734 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4737 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4740 dom = &keg->uk_domain[slab->us_domain];
4742 * keg_alloc_slab() always returns a slab on the
4745 LIST_REMOVE(slab, us_link);
4746 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4748 dom->ud_free_slabs++;
4749 KEG_UNLOCK(keg, slab->us_domain);
4752 if (vm_domainset_iter_policy(&di, &domain) != 0)
4753 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
4759 * Returns a snapshot of memory consumption in bytes.
4762 uma_zone_memory(uma_zone_t zone)
4768 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
4769 for (i = 0; i < vm_ndomains; i++)
4770 sz += ZDOM_GET(zone, i)->uzd_nitems;
4771 return (sz * zone->uz_size);
4773 for (i = 0; i < vm_ndomains; i++)
4774 sz += zone->uz_keg->uk_domain[i].ud_pages;
4776 return (sz * PAGE_SIZE);
4781 uma_reclaim(int req)
4784 CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4785 sx_xlock(&uma_reclaim_lock);
4789 case UMA_RECLAIM_TRIM:
4790 zone_foreach(zone_trim, NULL);
4792 case UMA_RECLAIM_DRAIN:
4793 case UMA_RECLAIM_DRAIN_CPU:
4794 zone_foreach(zone_drain, NULL);
4795 if (req == UMA_RECLAIM_DRAIN_CPU) {
4796 pcpu_cache_drain_safe(NULL);
4797 zone_foreach(zone_drain, NULL);
4801 panic("unhandled reclamation request %d", req);
4805 * Some slabs may have been freed but this zone will be visited early
4806 * we visit again so that we can free pages that are empty once other
4807 * zones are drained. We have to do the same for buckets.
4809 zone_drain(slabzones[0], NULL);
4810 zone_drain(slabzones[1], NULL);
4811 bucket_zone_drain();
4812 sx_xunlock(&uma_reclaim_lock);
4815 static volatile int uma_reclaim_needed;
4818 uma_reclaim_wakeup(void)
4821 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4822 wakeup(uma_reclaim);
4826 uma_reclaim_worker(void *arg __unused)
4830 sx_xlock(&uma_reclaim_lock);
4831 while (atomic_load_int(&uma_reclaim_needed) == 0)
4832 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4834 sx_xunlock(&uma_reclaim_lock);
4835 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4836 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4837 atomic_store_int(&uma_reclaim_needed, 0);
4838 /* Don't fire more than once per-second. */
4839 pause("umarclslp", hz);
4845 uma_zone_reclaim(uma_zone_t zone, int req)
4849 case UMA_RECLAIM_TRIM:
4850 zone_trim(zone, NULL);
4852 case UMA_RECLAIM_DRAIN:
4853 zone_drain(zone, NULL);
4855 case UMA_RECLAIM_DRAIN_CPU:
4856 pcpu_cache_drain_safe(zone);
4857 zone_drain(zone, NULL);
4860 panic("unhandled reclamation request %d", req);
4866 uma_zone_exhausted(uma_zone_t zone)
4869 return (atomic_load_32(&zone->uz_sleepers) > 0);
4876 return (uma_kmem_limit);
4880 uma_set_limit(unsigned long limit)
4883 uma_kmem_limit = limit;
4890 return (atomic_load_long(&uma_kmem_total));
4897 return (uma_kmem_limit - uma_size());
4902 * Generate statistics across both the zone and its per-cpu cache's. Return
4903 * desired statistics if the pointer is non-NULL for that statistic.
4905 * Note: does not update the zone statistics, as it can't safely clear the
4906 * per-CPU cache statistic.
4910 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4911 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4914 uint64_t allocs, frees, sleeps, xdomain;
4917 allocs = frees = sleeps = xdomain = 0;
4920 cache = &z->uz_cpu[cpu];
4921 cachefree += cache->uc_allocbucket.ucb_cnt;
4922 cachefree += cache->uc_freebucket.ucb_cnt;
4923 xdomain += cache->uc_crossbucket.ucb_cnt;
4924 cachefree += cache->uc_crossbucket.ucb_cnt;
4925 allocs += cache->uc_allocs;
4926 frees += cache->uc_frees;
4928 allocs += counter_u64_fetch(z->uz_allocs);
4929 frees += counter_u64_fetch(z->uz_frees);
4930 xdomain += counter_u64_fetch(z->uz_xdomain);
4931 sleeps += z->uz_sleeps;
4932 if (cachefreep != NULL)
4933 *cachefreep = cachefree;
4934 if (allocsp != NULL)
4938 if (sleepsp != NULL)
4940 if (xdomainp != NULL)
4941 *xdomainp = xdomain;
4946 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
4953 rw_rlock(&uma_rwlock);
4954 LIST_FOREACH(kz, &uma_kegs, uk_link) {
4955 LIST_FOREACH(z, &kz->uk_zones, uz_link)
4958 LIST_FOREACH(z, &uma_cachezones, uz_link)
4961 rw_runlock(&uma_rwlock);
4962 return (sysctl_handle_int(oidp, &count, 0, req));
4966 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
4967 struct uma_percpu_stat *ups, bool internal)
4969 uma_zone_domain_t zdom;
4974 for (i = 0; i < vm_ndomains; i++) {
4975 zdom = ZDOM_GET(z, i);
4976 uth->uth_zone_free += zdom->uzd_nitems;
4978 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
4979 uth->uth_frees = counter_u64_fetch(z->uz_frees);
4980 uth->uth_fails = counter_u64_fetch(z->uz_fails);
4981 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
4982 uth->uth_sleeps = z->uz_sleeps;
4984 for (i = 0; i < mp_maxid + 1; i++) {
4985 bzero(&ups[i], sizeof(*ups));
4986 if (internal || CPU_ABSENT(i))
4988 cache = &z->uz_cpu[i];
4989 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
4990 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
4991 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
4992 ups[i].ups_allocs = cache->uc_allocs;
4993 ups[i].ups_frees = cache->uc_frees;
4998 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5000 struct uma_stream_header ush;
5001 struct uma_type_header uth;
5002 struct uma_percpu_stat *ups;
5007 uint32_t kfree, pages;
5008 int count, error, i;
5010 error = sysctl_wire_old_buffer(req, 0);
5013 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5014 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5015 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5018 rw_rlock(&uma_rwlock);
5019 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5020 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5024 LIST_FOREACH(z, &uma_cachezones, uz_link)
5028 * Insert stream header.
5030 bzero(&ush, sizeof(ush));
5031 ush.ush_version = UMA_STREAM_VERSION;
5032 ush.ush_maxcpus = (mp_maxid + 1);
5033 ush.ush_count = count;
5034 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5036 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5038 for (i = 0; i < vm_ndomains; i++) {
5039 kfree += kz->uk_domain[i].ud_free_items;
5040 pages += kz->uk_domain[i].ud_pages;
5042 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5043 bzero(&uth, sizeof(uth));
5044 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5045 uth.uth_align = kz->uk_align;
5046 uth.uth_size = kz->uk_size;
5047 uth.uth_rsize = kz->uk_rsize;
5048 if (z->uz_max_items > 0) {
5049 items = UZ_ITEMS_COUNT(z->uz_items);
5050 uth.uth_pages = (items / kz->uk_ipers) *
5053 uth.uth_pages = pages;
5054 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5056 uth.uth_limit = z->uz_max_items;
5057 uth.uth_keg_free = kfree;
5060 * A zone is secondary is it is not the first entry
5061 * on the keg's zone list.
5063 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5064 (LIST_FIRST(&kz->uk_zones) != z))
5065 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5066 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5067 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5068 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5069 for (i = 0; i < mp_maxid + 1; i++)
5070 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5073 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5074 bzero(&uth, sizeof(uth));
5075 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5076 uth.uth_size = z->uz_size;
5077 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5078 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5079 for (i = 0; i < mp_maxid + 1; i++)
5080 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5083 rw_runlock(&uma_rwlock);
5084 error = sbuf_finish(&sbuf);
5091 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5093 uma_zone_t zone = *(uma_zone_t *)arg1;
5096 max = uma_zone_get_max(zone);
5097 error = sysctl_handle_int(oidp, &max, 0, req);
5098 if (error || !req->newptr)
5101 uma_zone_set_max(zone, max);
5107 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5113 * Some callers want to add sysctls for global zones that
5114 * may not yet exist so they pass a pointer to a pointer.
5117 zone = *(uma_zone_t *)arg1;
5120 cur = uma_zone_get_cur(zone);
5121 return (sysctl_handle_int(oidp, &cur, 0, req));
5125 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5127 uma_zone_t zone = arg1;
5130 cur = uma_zone_get_allocs(zone);
5131 return (sysctl_handle_64(oidp, &cur, 0, req));
5135 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5137 uma_zone_t zone = arg1;
5140 cur = uma_zone_get_frees(zone);
5141 return (sysctl_handle_64(oidp, &cur, 0, req));
5145 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5148 uma_zone_t zone = arg1;
5151 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5152 if (zone->uz_flags != 0)
5153 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5155 sbuf_printf(&sbuf, "0");
5156 error = sbuf_finish(&sbuf);
5163 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5165 uma_keg_t keg = arg1;
5166 int avail, effpct, total;
5168 total = keg->uk_ppera * PAGE_SIZE;
5169 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5170 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5172 * We consider the client's requested size and alignment here, not the
5173 * real size determination uk_rsize, because we also adjust the real
5174 * size for internal implementation reasons (max bitset size).
5176 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5177 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5178 avail *= mp_maxid + 1;
5179 effpct = 100 * avail / total;
5180 return (sysctl_handle_int(oidp, &effpct, 0, req));
5184 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5186 uma_zone_t zone = arg1;
5189 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5190 return (sysctl_handle_64(oidp, &cur, 0, req));
5195 uma_dbg_getslab(uma_zone_t zone, void *item)
5202 * It is safe to return the slab here even though the
5203 * zone is unlocked because the item's allocation state
5204 * essentially holds a reference.
5206 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5207 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5209 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5210 return (vtoslab((vm_offset_t)mem));
5212 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5213 return ((uma_slab_t)(mem + keg->uk_pgoff));
5215 slab = hash_sfind(&keg->uk_hash, mem);
5222 uma_dbg_zskip(uma_zone_t zone, void *mem)
5225 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5228 return (uma_dbg_kskip(zone->uz_keg, mem));
5232 uma_dbg_kskip(uma_keg_t keg, void *mem)
5236 if (dbg_divisor == 0)
5239 if (dbg_divisor == 1)
5242 idx = (uintptr_t)mem >> PAGE_SHIFT;
5243 if (keg->uk_ipers > 1) {
5244 idx *= keg->uk_ipers;
5245 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5248 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5249 counter_u64_add(uma_skip_cnt, 1);
5252 counter_u64_add(uma_dbg_cnt, 1);
5258 * Set up the slab's freei data such that uma_dbg_free can function.
5262 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5268 slab = uma_dbg_getslab(zone, item);
5270 panic("uma: item %p did not belong to zone %s\n",
5271 item, zone->uz_name);
5274 freei = slab_item_index(slab, keg, item);
5276 if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5277 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
5278 item, zone, zone->uz_name, slab, freei);
5279 BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5283 * Verifies freed addresses. Checks for alignment, valid slab membership
5284 * and duplicate frees.
5288 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5294 slab = uma_dbg_getslab(zone, item);
5296 panic("uma: Freed item %p did not belong to zone %s\n",
5297 item, zone->uz_name);
5300 freei = slab_item_index(slab, keg, item);
5302 if (freei >= keg->uk_ipers)
5303 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
5304 item, zone, zone->uz_name, slab, freei);
5306 if (slab_item(slab, keg, freei) != item)
5307 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
5308 item, zone, zone->uz_name, slab, freei);
5310 if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5311 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
5312 item, zone, zone->uz_name, slab, freei);
5314 BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5316 #endif /* INVARIANTS */
5320 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5321 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5326 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5327 *allocs = counter_u64_fetch(z->uz_allocs);
5328 frees = counter_u64_fetch(z->uz_frees);
5329 *sleeps = z->uz_sleeps;
5333 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5335 for (i = 0; i < vm_ndomains; i++) {
5336 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5337 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5338 (LIST_FIRST(&kz->uk_zones) != z)))
5339 *cachefree += kz->uk_domain[i].ud_free_items;
5341 *used = *allocs - frees;
5342 return (((int64_t)*used + *cachefree) * kz->uk_size);
5345 DB_SHOW_COMMAND(uma, db_show_uma)
5347 const char *fmt_hdr, *fmt_entry;
5350 uint64_t allocs, used, sleeps, xdomain;
5352 /* variables for sorting */
5354 uma_zone_t cur_zone, last_zone;
5355 int64_t cur_size, last_size, size;
5358 /* /i option produces machine-parseable CSV output */
5359 if (modif[0] == 'i') {
5360 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5361 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5363 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5364 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5367 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5368 "Sleeps", "Bucket", "Total Mem", "XFree");
5370 /* Sort the zones with largest size first. */
5372 last_size = INT64_MAX;
5377 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5378 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5380 * In the case of size ties, print out zones
5381 * in the order they are encountered. That is,
5382 * when we encounter the most recently output
5383 * zone, we have already printed all preceding
5384 * ties, and we must print all following ties.
5386 if (z == last_zone) {
5390 size = get_uma_stats(kz, z, &allocs, &used,
5391 &sleeps, &cachefree, &xdomain);
5392 if (size > cur_size && size < last_size + ties)
5400 if (cur_zone == NULL)
5403 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5404 &sleeps, &cachefree, &xdomain);
5405 db_printf(fmt_entry, cur_zone->uz_name,
5406 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5407 (uintmax_t)allocs, (uintmax_t)sleeps,
5408 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5413 last_zone = cur_zone;
5414 last_size = cur_size;
5418 DB_SHOW_COMMAND(umacache, db_show_umacache)
5421 uint64_t allocs, frees;
5425 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5426 "Requests", "Bucket");
5427 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5428 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5429 for (i = 0; i < vm_ndomains; i++)
5430 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5431 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5432 z->uz_name, (uintmax_t)z->uz_size,
5433 (intmax_t)(allocs - frees), cachefree,
5434 (uintmax_t)allocs, z->uz_bucket_size);