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 if it
568 * is out of balance with the preferred domain.
570 static __noinline int
571 zone_domain_lowest(uma_zone_t zone, int pref)
573 long least, nitems, prefitems;
577 prefitems = least = LONG_MAX;
579 for (i = 0; i < vm_ndomains; i++) {
580 nitems = ZDOM_GET(zone, i)->uzd_nitems;
581 if (nitems < least) {
588 if (prefitems < least * 2)
595 * Search for the domain with the most cached items and return it or the
596 * preferred domain if it has enough to proceed.
598 static __noinline int
599 zone_domain_highest(uma_zone_t zone, int pref)
605 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
610 for (i = 0; i < vm_ndomains; i++) {
611 nitems = ZDOM_GET(zone, i)->uzd_nitems;
622 * Safely subtract cnt from imax.
625 zone_domain_imax_sub(uma_zone_domain_t zdom, int cnt)
630 old = zdom->uzd_imax;
636 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, new) == 0);
640 * Set the maximum imax value.
643 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
647 old = zdom->uzd_imax;
651 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
655 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
656 * zone's caches. If a bucket is found the zone is not locked on return.
659 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
665 ZDOM_LOCK_ASSERT(zdom);
667 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
670 /* SMR Buckets can not be re-used until readers expire. */
671 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
672 bucket->ub_seq != SMR_SEQ_INVALID) {
673 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
675 bucket->ub_seq = SMR_SEQ_INVALID;
676 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
677 if (STAILQ_NEXT(bucket, ub_link) != NULL)
678 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
680 MPASS(zdom->uzd_nitems >= bucket->ub_cnt);
681 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
682 zdom->uzd_nitems -= bucket->ub_cnt;
685 * Shift the bounds of the current WSS interval to avoid
686 * perturbing the estimate.
689 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
690 zone_domain_imax_sub(zdom, bucket->ub_cnt);
691 } else if (zdom->uzd_imin > zdom->uzd_nitems)
692 zdom->uzd_imin = zdom->uzd_nitems;
696 for (i = 0; i < bucket->ub_cnt; i++)
697 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
704 * Insert a full bucket into the specified cache. The "ws" parameter indicates
705 * whether the bucket's contents should be counted as part of the zone's working
706 * set. The bucket may be freed if it exceeds the bucket limit.
709 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
712 uma_zone_domain_t zdom;
714 /* We don't cache empty buckets. This can happen after a reclaim. */
715 if (bucket->ub_cnt == 0)
717 zdom = zone_domain_lock(zone, domain);
719 KASSERT(!ws || zdom->uzd_nitems < zone->uz_bucket_max,
720 ("%s: zone %p overflow", __func__, zone));
723 * Conditionally set the maximum number of items.
725 zdom->uzd_nitems += bucket->ub_cnt;
726 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
728 zone_domain_imax_set(zdom, zdom->uzd_nitems);
729 if (STAILQ_EMPTY(&zdom->uzd_buckets))
730 zdom->uzd_seq = bucket->ub_seq;
731 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
735 zdom->uzd_nitems -= bucket->ub_cnt;
738 bucket_free(zone, bucket, udata);
741 /* Pops an item out of a per-cpu cache bucket. */
743 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
747 CRITICAL_ASSERT(curthread);
750 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
752 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
753 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
760 /* Pushes an item into a per-cpu cache bucket. */
762 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
765 CRITICAL_ASSERT(curthread);
766 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
767 ("uma_zfree: Freeing to non free bucket index."));
769 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
775 * Unload a UMA bucket from a per-cpu cache.
777 static inline uma_bucket_t
778 cache_bucket_unload(uma_cache_bucket_t bucket)
782 b = bucket->ucb_bucket;
784 MPASS(b->ub_entries == bucket->ucb_entries);
785 b->ub_cnt = bucket->ucb_cnt;
786 bucket->ucb_bucket = NULL;
787 bucket->ucb_entries = bucket->ucb_cnt = 0;
793 static inline uma_bucket_t
794 cache_bucket_unload_alloc(uma_cache_t cache)
797 return (cache_bucket_unload(&cache->uc_allocbucket));
800 static inline uma_bucket_t
801 cache_bucket_unload_free(uma_cache_t cache)
804 return (cache_bucket_unload(&cache->uc_freebucket));
807 static inline uma_bucket_t
808 cache_bucket_unload_cross(uma_cache_t cache)
811 return (cache_bucket_unload(&cache->uc_crossbucket));
815 * Load a bucket into a per-cpu cache bucket.
818 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
821 CRITICAL_ASSERT(curthread);
822 MPASS(bucket->ucb_bucket == NULL);
823 MPASS(b->ub_seq == SMR_SEQ_INVALID);
825 bucket->ucb_bucket = b;
826 bucket->ucb_cnt = b->ub_cnt;
827 bucket->ucb_entries = b->ub_entries;
831 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
834 cache_bucket_load(&cache->uc_allocbucket, b);
838 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
841 cache_bucket_load(&cache->uc_freebucket, b);
846 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
849 cache_bucket_load(&cache->uc_crossbucket, b);
854 * Copy and preserve ucb_spare.
857 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
860 b1->ucb_bucket = b2->ucb_bucket;
861 b1->ucb_entries = b2->ucb_entries;
862 b1->ucb_cnt = b2->ucb_cnt;
866 * Swap two cache buckets.
869 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
871 struct uma_cache_bucket b3;
873 CRITICAL_ASSERT(curthread);
875 cache_bucket_copy(&b3, b1);
876 cache_bucket_copy(b1, b2);
877 cache_bucket_copy(b2, &b3);
881 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
884 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
886 uma_zone_domain_t zdom;
890 * Avoid the lock if possible.
892 zdom = ZDOM_GET(zone, domain);
893 if (zdom->uzd_nitems == 0)
896 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
897 !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
901 * Check the zone's cache of buckets.
903 zdom = zone_domain_lock(zone, domain);
904 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
905 KASSERT(bucket->ub_cnt != 0,
906 ("cache_fetch_bucket: Returning an empty bucket."));
915 zone_log_warning(uma_zone_t zone)
917 static const struct timeval warninterval = { 300, 0 };
919 if (!zone_warnings || zone->uz_warning == NULL)
922 if (ratecheck(&zone->uz_ratecheck, &warninterval))
923 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
927 zone_maxaction(uma_zone_t zone)
930 if (zone->uz_maxaction.ta_func != NULL)
931 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
935 * Routine called by timeout which is used to fire off some time interval
936 * based calculations. (stats, hash size, etc.)
945 uma_timeout(void *unused)
948 zone_foreach(zone_timeout, NULL);
950 /* Reschedule this event */
951 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
955 * Update the working set size estimate for the zone's bucket cache.
956 * The constants chosen here are somewhat arbitrary. With an update period of
957 * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
961 zone_domain_update_wss(uma_zone_domain_t zdom)
966 MPASS(zdom->uzd_imax >= zdom->uzd_imin);
967 wss = zdom->uzd_imax - zdom->uzd_imin;
968 zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
969 zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
974 * Routine to perform timeout driven calculations. This expands the
975 * hashes and does per cpu statistics aggregation.
980 zone_timeout(uma_zone_t zone, void *unused)
985 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
991 * Hash zones are non-numa by definition so the first domain
992 * is the only one present.
995 pages = keg->uk_domain[0].ud_pages;
998 * Expand the keg hash table.
1000 * This is done if the number of slabs is larger than the hash size.
1001 * What I'm trying to do here is completely reduce collisions. This
1002 * may be a little aggressive. Should I allow for two collisions max?
1004 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1005 struct uma_hash newhash;
1006 struct uma_hash oldhash;
1010 * This is so involved because allocating and freeing
1011 * while the keg lock is held will lead to deadlock.
1012 * I have to do everything in stages and check for
1016 ret = hash_alloc(&newhash, 1 << fls(slabs));
1019 if (hash_expand(&keg->uk_hash, &newhash)) {
1020 oldhash = keg->uk_hash;
1021 keg->uk_hash = newhash;
1026 hash_free(&oldhash);
1033 for (int i = 0; i < vm_ndomains; i++)
1034 zone_domain_update_wss(ZDOM_GET(zone, i));
1038 * Allocate and zero fill the next sized hash table from the appropriate
1042 * hash A new hash structure with the old hash size in uh_hashsize
1045 * 1 on success and 0 on failure.
1048 hash_alloc(struct uma_hash *hash, u_int size)
1052 KASSERT(powerof2(size), ("hash size must be power of 2"));
1053 if (size > UMA_HASH_SIZE_INIT) {
1054 hash->uh_hashsize = size;
1055 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1056 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1058 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1059 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1060 UMA_ANYDOMAIN, M_WAITOK);
1061 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1063 if (hash->uh_slab_hash) {
1064 bzero(hash->uh_slab_hash, alloc);
1065 hash->uh_hashmask = hash->uh_hashsize - 1;
1073 * Expands the hash table for HASH zones. This is done from zone_timeout
1074 * to reduce collisions. This must not be done in the regular allocation
1075 * path, otherwise, we can recurse on the vm while allocating pages.
1078 * oldhash The hash you want to expand
1079 * newhash The hash structure for the new table
1087 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1089 uma_hash_slab_t slab;
1093 if (!newhash->uh_slab_hash)
1096 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1100 * I need to investigate hash algorithms for resizing without a
1104 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1105 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1106 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1107 LIST_REMOVE(slab, uhs_hlink);
1108 hval = UMA_HASH(newhash, slab->uhs_data);
1109 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1117 * Free the hash bucket to the appropriate backing store.
1120 * slab_hash The hash bucket we're freeing
1121 * hashsize The number of entries in that hash bucket
1127 hash_free(struct uma_hash *hash)
1129 if (hash->uh_slab_hash == NULL)
1131 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1132 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1134 free(hash->uh_slab_hash, M_UMAHASH);
1138 * Frees all outstanding items in a bucket
1141 * zone The zone to free to, must be unlocked.
1142 * bucket The free/alloc bucket with items.
1148 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1152 if (bucket->ub_cnt == 0)
1155 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1156 bucket->ub_seq != SMR_SEQ_INVALID) {
1157 smr_wait(zone->uz_smr, bucket->ub_seq);
1158 bucket->ub_seq = SMR_SEQ_INVALID;
1159 for (i = 0; i < bucket->ub_cnt; i++)
1160 item_dtor(zone, bucket->ub_bucket[i],
1161 zone->uz_size, NULL, SKIP_NONE);
1164 for (i = 0; i < bucket->ub_cnt; i++)
1165 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1166 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1167 if (zone->uz_max_items > 0)
1168 zone_free_limit(zone, bucket->ub_cnt);
1170 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1176 * Drains the per cpu caches for a zone.
1178 * NOTE: This may only be called while the zone is being torn down, and not
1179 * during normal operation. This is necessary in order that we do not have
1180 * to migrate CPUs to drain the per-CPU caches.
1183 * zone The zone to drain, must be unlocked.
1189 cache_drain(uma_zone_t zone)
1192 uma_bucket_t bucket;
1197 * XXX: It is safe to not lock the per-CPU caches, because we're
1198 * tearing down the zone anyway. I.e., there will be no further use
1199 * of the caches at this point.
1201 * XXX: It would good to be able to assert that the zone is being
1202 * torn down to prevent improper use of cache_drain().
1204 seq = SMR_SEQ_INVALID;
1205 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1206 seq = smr_advance(zone->uz_smr);
1208 cache = &zone->uz_cpu[cpu];
1209 bucket = cache_bucket_unload_alloc(cache);
1211 bucket_free(zone, bucket, NULL);
1212 bucket = cache_bucket_unload_free(cache);
1213 if (bucket != NULL) {
1214 bucket->ub_seq = seq;
1215 bucket_free(zone, bucket, NULL);
1217 bucket = cache_bucket_unload_cross(cache);
1218 if (bucket != NULL) {
1219 bucket->ub_seq = seq;
1220 bucket_free(zone, bucket, NULL);
1223 bucket_cache_reclaim(zone, true);
1227 cache_shrink(uma_zone_t zone, void *unused)
1230 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1233 zone->uz_bucket_size =
1234 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1238 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1241 uma_bucket_t b1, b2, b3;
1244 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1247 b1 = b2 = b3 = NULL;
1249 cache = &zone->uz_cpu[curcpu];
1250 domain = PCPU_GET(domain);
1251 b1 = cache_bucket_unload_alloc(cache);
1254 * Don't flush SMR zone buckets. This leaves the zone without a
1255 * bucket and forces every free to synchronize().
1257 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1258 b2 = cache_bucket_unload_free(cache);
1259 b3 = cache_bucket_unload_cross(cache);
1264 zone_free_bucket(zone, b1, NULL, domain, false);
1266 zone_free_bucket(zone, b2, NULL, domain, false);
1268 /* Adjust the domain so it goes to zone_free_cross. */
1269 domain = (domain + 1) % vm_ndomains;
1270 zone_free_bucket(zone, b3, NULL, domain, false);
1275 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1276 * This is an expensive call because it needs to bind to all CPUs
1277 * one by one and enter a critical section on each of them in order
1278 * to safely access their cache buckets.
1279 * Zone lock must not be held on call this function.
1282 pcpu_cache_drain_safe(uma_zone_t zone)
1287 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1290 cache_shrink(zone, NULL);
1292 zone_foreach(cache_shrink, NULL);
1295 thread_lock(curthread);
1296 sched_bind(curthread, cpu);
1297 thread_unlock(curthread);
1300 cache_drain_safe_cpu(zone, NULL);
1302 zone_foreach(cache_drain_safe_cpu, NULL);
1304 thread_lock(curthread);
1305 sched_unbind(curthread);
1306 thread_unlock(curthread);
1310 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1311 * requested a drain, otherwise the per-domain caches are trimmed to either
1312 * estimated working set size.
1315 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1317 uma_zone_domain_t zdom;
1318 uma_bucket_t bucket;
1323 * Shrink the zone bucket size to ensure that the per-CPU caches
1324 * don't grow too large.
1326 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1327 zone->uz_bucket_size--;
1329 for (i = 0; i < vm_ndomains; i++) {
1331 * The cross bucket is partially filled and not part of
1332 * the item count. Reclaim it individually here.
1334 zdom = ZDOM_GET(zone, i);
1335 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1336 ZONE_CROSS_LOCK(zone);
1337 bucket = zdom->uzd_cross;
1338 zdom->uzd_cross = NULL;
1339 ZONE_CROSS_UNLOCK(zone);
1341 bucket_free(zone, bucket, NULL);
1345 * If we were asked to drain the zone, we are done only once
1346 * this bucket cache is empty. Otherwise, we reclaim items in
1347 * excess of the zone's estimated working set size. If the
1348 * difference nitems - imin is larger than the WSS estimate,
1349 * then the estimate will grow at the end of this interval and
1350 * we ignore the historical average.
1353 target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1355 while (zdom->uzd_nitems > target) {
1356 bucket = zone_fetch_bucket(zone, zdom, true);
1359 bucket_free(zone, bucket, NULL);
1367 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1373 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1374 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1376 mem = slab_data(slab, keg);
1377 flags = slab->us_flags;
1379 if (keg->uk_fini != NULL) {
1380 for (i--; i > -1; i--)
1383 * trash_fini implies that dtor was trash_dtor. trash_fini
1384 * would check that memory hasn't been modified since free,
1385 * which executed trash_dtor.
1386 * That's why we need to run uma_dbg_kskip() check here,
1387 * albeit we don't make skip check for other init/fini
1390 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1391 keg->uk_fini != trash_fini)
1393 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1395 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1396 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1398 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1399 uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1403 * Frees pages from a keg back to the system. This is done on demand from
1404 * the pageout daemon.
1409 keg_drain(uma_keg_t keg)
1411 struct slabhead freeslabs;
1413 uma_slab_t slab, tmp;
1416 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1419 for (i = 0; i < vm_ndomains; i++) {
1420 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1421 keg->uk_name, keg, i, dom->ud_free_items);
1422 dom = &keg->uk_domain[i];
1423 LIST_INIT(&freeslabs);
1426 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1427 LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
1428 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1430 n = dom->ud_free_slabs;
1431 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1432 dom->ud_free_slabs = 0;
1433 dom->ud_free_items -= n * keg->uk_ipers;
1434 dom->ud_pages -= n * keg->uk_ppera;
1437 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1438 keg_free_slab(keg, slab, keg->uk_ipers);
1443 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1447 * Set draining to interlock with zone_dtor() so we can release our
1448 * locks as we go. Only dtor() should do a WAITOK call since it
1449 * is the only call that knows the structure will still be available
1453 while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1454 if (waitok == M_NOWAIT)
1456 msleep(zone, &ZDOM_GET(zone, 0)->uzd_lock, PVM, "zonedrain",
1459 zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1461 bucket_cache_reclaim(zone, drain);
1464 * The DRAINING flag protects us from being freed while
1465 * we're running. Normally the uma_rwlock would protect us but we
1466 * must be able to release and acquire the right lock for each keg.
1468 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1469 keg_drain(zone->uz_keg);
1471 zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1478 zone_drain(uma_zone_t zone, void *unused)
1481 zone_reclaim(zone, M_NOWAIT, true);
1485 zone_trim(uma_zone_t zone, void *unused)
1488 zone_reclaim(zone, M_NOWAIT, false);
1492 * Allocate a new slab for a keg and inserts it into the partial slab list.
1493 * The keg should be unlocked on entry. If the allocation succeeds it will
1494 * be locked on return.
1497 * flags Wait flags for the item initialization routine
1498 * aflags Wait flags for the slab allocation
1501 * The slab that was allocated or NULL if there is no memory and the
1502 * caller specified M_NOWAIT.
1505 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1516 KASSERT(domain >= 0 && domain < vm_ndomains,
1517 ("keg_alloc_slab: domain %d out of range", domain));
1519 allocf = keg->uk_allocf;
1522 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1523 uma_hash_slab_t hslab;
1524 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1528 slab = &hslab->uhs_slab;
1532 * This reproduces the old vm_zone behavior of zero filling pages the
1533 * first time they are added to a zone.
1535 * Malloced items are zeroed in uma_zalloc.
1538 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1543 if (keg->uk_flags & UMA_ZONE_NODUMP)
1546 /* zone is passed for legacy reasons. */
1547 size = keg->uk_ppera * PAGE_SIZE;
1548 mem = allocf(zone, size, domain, &sflags, aflags);
1550 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1551 zone_free_item(slabzone(keg->uk_ipers),
1552 slab_tohashslab(slab), NULL, SKIP_NONE);
1555 uma_total_inc(size);
1557 /* For HASH zones all pages go to the same uma_domain. */
1558 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1561 /* Point the slab into the allocated memory */
1562 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1563 slab = (uma_slab_t )(mem + keg->uk_pgoff);
1565 slab_tohashslab(slab)->uhs_data = mem;
1567 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1568 for (i = 0; i < keg->uk_ppera; i++)
1569 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1572 slab->us_freecount = keg->uk_ipers;
1573 slab->us_flags = sflags;
1574 slab->us_domain = domain;
1576 BIT_FILL(keg->uk_ipers, &slab->us_free);
1578 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1581 if (keg->uk_init != NULL) {
1582 for (i = 0; i < keg->uk_ipers; i++)
1583 if (keg->uk_init(slab_item(slab, keg, i),
1584 keg->uk_size, flags) != 0)
1586 if (i != keg->uk_ipers) {
1587 keg_free_slab(keg, slab, i);
1591 KEG_LOCK(keg, domain);
1593 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1594 slab, keg->uk_name, keg);
1596 if (keg->uk_flags & UMA_ZFLAG_HASH)
1597 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1600 * If we got a slab here it's safe to mark it partially used
1601 * and return. We assume that the caller is going to remove
1602 * at least one item.
1604 dom = &keg->uk_domain[domain];
1605 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1606 dom->ud_pages += keg->uk_ppera;
1607 dom->ud_free_items += keg->uk_ipers;
1616 * This function is intended to be used early on in place of page_alloc() so
1617 * that we may use the boot time page cache to satisfy allocations before
1621 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1630 pages = howmany(bytes, PAGE_SIZE);
1631 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1633 *pflag = UMA_SLAB_BOOT;
1634 m = vm_page_alloc_contig_domain(NULL, 0, domain,
1635 malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1636 (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1640 pa = VM_PAGE_TO_PHYS(m);
1641 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1642 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1643 defined(__riscv) || defined(__powerpc64__)
1644 if ((wait & M_NODUMP) == 0)
1648 /* Allocate KVA and indirectly advance bootmem. */
1649 mem = (void *)pmap_map(&bootmem, m->phys_addr,
1650 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1651 if ((wait & M_ZERO) != 0)
1652 bzero(mem, pages * PAGE_SIZE);
1658 startup_free(void *mem, vm_size_t bytes)
1663 va = (vm_offset_t)mem;
1664 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1665 pmap_remove(kernel_pmap, va, va + bytes);
1666 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1667 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1668 defined(__riscv) || defined(__powerpc64__)
1669 dump_drop_page(VM_PAGE_TO_PHYS(m));
1671 vm_page_unwire_noq(m);
1677 * Allocates a number of pages from the system
1680 * bytes The number of bytes requested
1681 * wait Shall we wait?
1684 * A pointer to the alloced memory or possibly
1685 * NULL if M_NOWAIT is set.
1688 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1691 void *p; /* Returned page */
1693 *pflag = UMA_SLAB_KERNEL;
1694 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1700 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1703 struct pglist alloctail;
1704 vm_offset_t addr, zkva;
1706 vm_page_t p, p_next;
1711 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1713 TAILQ_INIT(&alloctail);
1714 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1715 malloc2vm_flags(wait);
1716 *pflag = UMA_SLAB_KERNEL;
1717 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1718 if (CPU_ABSENT(cpu)) {
1719 p = vm_page_alloc(NULL, 0, flags);
1722 p = vm_page_alloc(NULL, 0, flags);
1724 pc = pcpu_find(cpu);
1725 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1728 p = vm_page_alloc_domain(NULL, 0,
1729 pc->pc_domain, flags);
1730 if (__predict_false(p == NULL))
1731 p = vm_page_alloc(NULL, 0, flags);
1734 if (__predict_false(p == NULL))
1736 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1738 if ((addr = kva_alloc(bytes)) == 0)
1741 TAILQ_FOREACH(p, &alloctail, listq) {
1742 pmap_qenter(zkva, &p, 1);
1745 return ((void*)addr);
1747 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1748 vm_page_unwire_noq(p);
1755 * Allocates a number of pages from within an object
1758 * bytes The number of bytes requested
1759 * wait Shall we wait?
1762 * A pointer to the alloced memory or possibly
1763 * NULL if M_NOWAIT is set.
1766 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1769 TAILQ_HEAD(, vm_page) alloctail;
1771 vm_offset_t retkva, zkva;
1772 vm_page_t p, p_next;
1775 TAILQ_INIT(&alloctail);
1778 npages = howmany(bytes, PAGE_SIZE);
1779 while (npages > 0) {
1780 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1781 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1782 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1786 * Since the page does not belong to an object, its
1789 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1794 * Page allocation failed, free intermediate pages and
1797 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1798 vm_page_unwire_noq(p);
1803 *flags = UMA_SLAB_PRIV;
1804 zkva = keg->uk_kva +
1805 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1807 TAILQ_FOREACH(p, &alloctail, listq) {
1808 pmap_qenter(zkva, &p, 1);
1812 return ((void *)retkva);
1816 * Allocate physically contiguous pages.
1819 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1823 *pflag = UMA_SLAB_KERNEL;
1824 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1825 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1829 * Frees a number of pages to the system
1832 * mem A pointer to the memory to be freed
1833 * size The size of the memory being freed
1834 * flags The original p->us_flags field
1840 page_free(void *mem, vm_size_t size, uint8_t flags)
1843 if ((flags & UMA_SLAB_BOOT) != 0) {
1844 startup_free(mem, size);
1848 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1849 ("UMA: page_free used with invalid flags %x", flags));
1851 kmem_free((vm_offset_t)mem, size);
1855 * Frees pcpu zone allocations
1858 * mem A pointer to the memory to be freed
1859 * size The size of the memory being freed
1860 * flags The original p->us_flags field
1866 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1868 vm_offset_t sva, curva;
1872 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1874 if ((flags & UMA_SLAB_BOOT) != 0) {
1875 startup_free(mem, size);
1879 sva = (vm_offset_t)mem;
1880 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1881 paddr = pmap_kextract(curva);
1882 m = PHYS_TO_VM_PAGE(paddr);
1883 vm_page_unwire_noq(m);
1886 pmap_qremove(sva, size >> PAGE_SHIFT);
1887 kva_free(sva, size);
1892 * Zero fill initializer
1894 * Arguments/Returns follow uma_init specifications
1897 zero_init(void *mem, int size, int flags)
1905 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1908 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1913 * Actual size of embedded struct slab (!OFFPAGE).
1916 slab_sizeof(int nitems)
1920 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1921 return (roundup(s, UMA_ALIGN_PTR + 1));
1925 * Size of memory for embedded slabs (!OFFPAGE).
1928 slab_space(int nitems)
1930 return (UMA_SLAB_SIZE - slab_sizeof(nitems));
1933 #define UMA_FIXPT_SHIFT 31
1934 #define UMA_FRAC_FIXPT(n, d) \
1935 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1936 #define UMA_FIXPT_PCT(f) \
1937 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1938 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
1939 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1942 * Compute the number of items that will fit in a slab. If hdr is true, the
1943 * item count may be limited to provide space in the slab for an inline slab
1944 * header. Otherwise, all slab space will be provided for item storage.
1947 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1952 /* The padding between items is not needed after the last item. */
1953 padpi = rsize - size;
1957 * Start with the maximum item count and remove items until
1958 * the slab header first alongside the allocatable memory.
1960 for (ipers = MIN(SLAB_MAX_SETSIZE,
1961 (slabsize + padpi - slab_sizeof(1)) / rsize);
1963 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1967 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1974 * Compute the number of items that will fit in a slab for a startup zone.
1977 slab_ipers(size_t size, int align)
1981 rsize = roundup(size, align + 1); /* Assume no CACHESPREAD */
1982 return (slab_ipers_hdr(size, rsize, UMA_SLAB_SIZE, true));
1985 struct keg_layout_result {
1993 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
1994 struct keg_layout_result *kl)
1999 kl->slabsize = slabsize;
2001 /* Handle INTERNAL as inline with an extra page. */
2002 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2003 kl->format &= ~UMA_ZFLAG_INTERNAL;
2004 kl->slabsize += PAGE_SIZE;
2007 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2008 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2010 /* Account for memory used by an offpage slab header. */
2011 total = kl->slabsize;
2012 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2013 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2015 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2019 * Determine the format of a uma keg. This determines where the slab header
2020 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2023 * keg The zone we should initialize
2029 keg_layout(uma_keg_t keg)
2031 struct keg_layout_result kl = {}, kl_tmp;
2040 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2041 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2042 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2043 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2044 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2046 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2047 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2048 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2051 alignsize = keg->uk_align + 1;
2054 * Calculate the size of each allocation (rsize) according to
2055 * alignment. If the requested size is smaller than we have
2056 * allocation bits for we round it up.
2058 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2059 rsize = roundup2(rsize, alignsize);
2061 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2063 * We want one item to start on every align boundary in a page.
2064 * To do this we will span pages. We will also extend the item
2065 * by the size of align if it is an even multiple of align.
2066 * Otherwise, it would fall on the same boundary every time.
2068 if ((rsize & alignsize) == 0)
2070 slabsize = rsize * (PAGE_SIZE / alignsize);
2071 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2072 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2073 slabsize = round_page(slabsize);
2076 * Start with a slab size of as many pages as it takes to
2077 * represent a single item. We will try to fit as many
2078 * additional items into the slab as possible.
2080 slabsize = round_page(keg->uk_size);
2083 /* Build a list of all of the available formats for this keg. */
2086 /* Evaluate an inline slab layout. */
2087 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2090 /* TODO: vm_page-embedded slab. */
2093 * We can't do OFFPAGE if we're internal or if we've been
2094 * asked to not go to the VM for buckets. If we do this we
2095 * may end up going to the VM for slabs which we do not want
2096 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2097 * In those cases, evaluate a pseudo-format called INTERNAL
2098 * which has an inline slab header and one extra page to
2099 * guarantee that it fits.
2101 * Otherwise, see if using an OFFPAGE slab will improve our
2104 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2105 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2107 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2110 * Choose a slab size and format which satisfy the minimum efficiency.
2111 * Prefer the smallest slab size that meets the constraints.
2113 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2114 * for small items (up to PAGE_SIZE), the iteration increment is one
2115 * page; and for large items, the increment is one item.
2117 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2118 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2119 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2122 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2123 round_page(rsize * (i - 1) + keg->uk_size);
2125 for (j = 0; j < nfmt; j++) {
2126 /* Only if we have no viable format yet. */
2127 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2131 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2132 if (kl_tmp.eff <= kl.eff)
2137 CTR6(KTR_UMA, "keg %s layout: format %#x "
2138 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2139 keg->uk_name, kl.format, kl.ipers, rsize,
2140 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2142 /* Stop when we reach the minimum efficiency. */
2143 if (kl.eff >= UMA_MIN_EFF)
2147 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2148 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2149 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2153 pages = atop(kl.slabsize);
2154 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2155 pages *= mp_maxid + 1;
2157 keg->uk_rsize = rsize;
2158 keg->uk_ipers = kl.ipers;
2159 keg->uk_ppera = pages;
2160 keg->uk_flags |= kl.format;
2163 * How do we find the slab header if it is offpage or if not all item
2164 * start addresses are in the same page? We could solve the latter
2165 * case with vaddr alignment, but we don't.
2167 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2168 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2169 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2170 keg->uk_flags |= UMA_ZFLAG_HASH;
2172 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2175 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2176 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2178 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2179 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2180 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2181 keg->uk_ipers, pages));
2185 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2186 * the keg onto the global keg list.
2188 * Arguments/Returns follow uma_ctor specifications
2189 * udata Actually uma_kctor_args
2192 keg_ctor(void *mem, int size, void *udata, int flags)
2194 struct uma_kctor_args *arg = udata;
2195 uma_keg_t keg = mem;
2200 keg->uk_size = arg->size;
2201 keg->uk_init = arg->uminit;
2202 keg->uk_fini = arg->fini;
2203 keg->uk_align = arg->align;
2204 keg->uk_reserve = 0;
2205 keg->uk_flags = arg->flags;
2208 * We use a global round-robin policy by default. Zones with
2209 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2210 * case the iterator is never run.
2212 keg->uk_dr.dr_policy = DOMAINSET_RR();
2213 keg->uk_dr.dr_iter = 0;
2216 * The master zone is passed to us at keg-creation time.
2219 keg->uk_name = zone->uz_name;
2221 if (arg->flags & UMA_ZONE_ZINIT)
2222 keg->uk_init = zero_init;
2224 if (arg->flags & UMA_ZONE_MALLOC)
2225 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2228 keg->uk_flags &= ~UMA_ZONE_PCPU;
2234 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2235 * work on. Use round-robin for everything else.
2237 * Zones may override the default by specifying either.
2240 if ((keg->uk_flags &
2241 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2242 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2243 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2244 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2248 * If we haven't booted yet we need allocations to go through the
2249 * startup cache until the vm is ready.
2251 #ifdef UMA_MD_SMALL_ALLOC
2252 if (keg->uk_ppera == 1)
2253 keg->uk_allocf = uma_small_alloc;
2256 if (booted < BOOT_KVA)
2257 keg->uk_allocf = startup_alloc;
2258 else if (keg->uk_flags & UMA_ZONE_PCPU)
2259 keg->uk_allocf = pcpu_page_alloc;
2260 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2261 keg->uk_allocf = contig_alloc;
2263 keg->uk_allocf = page_alloc;
2264 #ifdef UMA_MD_SMALL_ALLOC
2265 if (keg->uk_ppera == 1)
2266 keg->uk_freef = uma_small_free;
2269 if (keg->uk_flags & UMA_ZONE_PCPU)
2270 keg->uk_freef = pcpu_page_free;
2272 keg->uk_freef = page_free;
2275 * Initialize keg's locks.
2277 for (i = 0; i < vm_ndomains; i++)
2278 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2281 * If we're putting the slab header in the actual page we need to
2282 * figure out where in each page it goes. See slab_sizeof
2285 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2288 shsize = slab_sizeof(keg->uk_ipers);
2289 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2291 * The only way the following is possible is if with our
2292 * UMA_ALIGN_PTR adjustments we are now bigger than
2293 * UMA_SLAB_SIZE. I haven't checked whether this is
2294 * mathematically possible for all cases, so we make
2297 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2298 ("zone %s ipers %d rsize %d size %d slab won't fit",
2299 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2302 if (keg->uk_flags & UMA_ZFLAG_HASH)
2303 hash_alloc(&keg->uk_hash, 0);
2305 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2307 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2309 rw_wlock(&uma_rwlock);
2310 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2311 rw_wunlock(&uma_rwlock);
2316 zone_kva_available(uma_zone_t zone, void *unused)
2320 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2324 if (keg->uk_allocf == startup_alloc) {
2325 /* Switch to the real allocator. */
2326 if (keg->uk_flags & UMA_ZONE_PCPU)
2327 keg->uk_allocf = pcpu_page_alloc;
2328 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2330 keg->uk_allocf = contig_alloc;
2332 keg->uk_allocf = page_alloc;
2337 zone_alloc_counters(uma_zone_t zone, void *unused)
2340 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2341 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2342 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2343 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2347 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2349 uma_zone_domain_t zdom;
2352 struct sysctl_oid *oid, *domainoid;
2353 int domains, i, cnt;
2354 static const char *nokeg = "cache zone";
2358 * Make a sysctl safe copy of the zone name by removing
2359 * any special characters and handling dups by appending
2362 if (zone->uz_namecnt != 0) {
2363 /* Count the number of decimal digits and '_' separator. */
2364 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2366 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2368 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2371 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2372 for (c = zone->uz_ctlname; *c != '\0'; c++)
2373 if (strchr("./\\ -", *c) != NULL)
2377 * Basic parameters at the root.
2379 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2380 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2382 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2383 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2384 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2385 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2386 zone, 0, sysctl_handle_uma_zone_flags, "A",
2387 "Allocator configuration flags");
2388 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2389 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2390 "Desired per-cpu cache size");
2391 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2392 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2393 "Maximum allowed per-cpu cache size");
2398 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2399 domains = vm_ndomains;
2402 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2403 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2405 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2406 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2407 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2408 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2409 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2410 "Real object size with alignment");
2411 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2412 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2413 "pages per-slab allocation");
2414 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2415 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2416 "items available per-slab");
2417 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2418 "align", CTLFLAG_RD, &keg->uk_align, 0,
2419 "item alignment mask");
2420 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2421 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2422 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2423 "Slab utilization (100 - internal fragmentation %)");
2424 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2425 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2426 for (i = 0; i < domains; i++) {
2427 dom = &keg->uk_domain[i];
2428 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2429 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2430 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2431 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2432 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2433 "Total pages currently allocated from VM");
2434 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2435 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2436 "items free in the slab layer");
2439 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2440 "name", CTLFLAG_RD, nokeg, "Keg name");
2443 * Information about zone limits.
2445 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2446 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2447 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2448 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2449 zone, 0, sysctl_handle_uma_zone_items, "QU",
2450 "current number of allocated items if limit is set");
2451 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2452 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2453 "Maximum number of cached items");
2454 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2455 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2456 "Number of threads sleeping at limit");
2457 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2458 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2459 "Total zone limit sleeps");
2460 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2461 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2462 "Maximum number of items in each domain's bucket cache");
2465 * Per-domain zone information.
2467 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2468 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2469 for (i = 0; i < domains; i++) {
2470 zdom = ZDOM_GET(zone, i);
2471 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2472 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2473 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2474 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2475 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2476 "number of items in this domain");
2477 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2478 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2479 "maximum item count in this period");
2480 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2481 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2482 "minimum item count in this period");
2483 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2484 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2485 "Working set size");
2489 * General statistics.
2491 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2492 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2493 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2494 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2495 zone, 1, sysctl_handle_uma_zone_cur, "I",
2496 "Current number of allocated items");
2497 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2498 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2499 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2500 "Total allocation calls");
2501 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2502 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2503 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2504 "Total free calls");
2505 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2506 "fails", CTLFLAG_RD, &zone->uz_fails,
2507 "Number of allocation failures");
2508 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2509 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2510 "Free calls from the wrong domain");
2513 struct uma_zone_count {
2519 zone_count(uma_zone_t zone, void *arg)
2521 struct uma_zone_count *cnt;
2525 * Some zones are rapidly created with identical names and
2526 * destroyed out of order. This can lead to gaps in the count.
2527 * Use one greater than the maximum observed for this name.
2529 if (strcmp(zone->uz_name, cnt->name) == 0)
2530 cnt->count = MAX(cnt->count,
2531 zone->uz_namecnt + 1);
2535 zone_update_caches(uma_zone_t zone)
2539 for (i = 0; i <= mp_maxid; i++) {
2540 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2541 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2546 * Zone header ctor. This initializes all fields, locks, etc.
2548 * Arguments/Returns follow uma_ctor specifications
2549 * udata Actually uma_zctor_args
2552 zone_ctor(void *mem, int size, void *udata, int flags)
2554 struct uma_zone_count cnt;
2555 struct uma_zctor_args *arg = udata;
2556 uma_zone_domain_t zdom;
2557 uma_zone_t zone = mem;
2563 zone->uz_name = arg->name;
2564 zone->uz_ctor = arg->ctor;
2565 zone->uz_dtor = arg->dtor;
2566 zone->uz_init = NULL;
2567 zone->uz_fini = NULL;
2568 zone->uz_sleeps = 0;
2569 zone->uz_bucket_size = 0;
2570 zone->uz_bucket_size_min = 0;
2571 zone->uz_bucket_size_max = BUCKET_MAX;
2572 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2573 zone->uz_warning = NULL;
2574 /* The domain structures follow the cpu structures. */
2575 zone->uz_bucket_max = ULONG_MAX;
2576 timevalclear(&zone->uz_ratecheck);
2578 /* Count the number of duplicate names. */
2579 cnt.name = arg->name;
2581 zone_foreach(zone_count, &cnt);
2582 zone->uz_namecnt = cnt.count;
2583 ZONE_CROSS_LOCK_INIT(zone);
2585 for (i = 0; i < vm_ndomains; i++) {
2586 zdom = ZDOM_GET(zone, i);
2587 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2588 STAILQ_INIT(&zdom->uzd_buckets);
2592 if (arg->uminit == trash_init && arg->fini == trash_fini)
2593 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2597 * This is a pure cache zone, no kegs.
2600 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2601 ("zone_ctor: Import specified for non-cache zone."));
2602 zone->uz_flags = arg->flags;
2603 zone->uz_size = arg->size;
2604 zone->uz_import = arg->import;
2605 zone->uz_release = arg->release;
2606 zone->uz_arg = arg->arg;
2609 * Cache zones are round-robin unless a policy is
2610 * specified because they may have incompatible
2613 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2614 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2616 rw_wlock(&uma_rwlock);
2617 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2618 rw_wunlock(&uma_rwlock);
2623 * Use the regular zone/keg/slab allocator.
2625 zone->uz_import = zone_import;
2626 zone->uz_release = zone_release;
2627 zone->uz_arg = zone;
2630 if (arg->flags & UMA_ZONE_SECONDARY) {
2631 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2632 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2633 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2634 zone->uz_init = arg->uminit;
2635 zone->uz_fini = arg->fini;
2636 zone->uz_flags |= UMA_ZONE_SECONDARY;
2637 rw_wlock(&uma_rwlock);
2639 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2640 if (LIST_NEXT(z, uz_link) == NULL) {
2641 LIST_INSERT_AFTER(z, zone, uz_link);
2646 rw_wunlock(&uma_rwlock);
2647 } else if (keg == NULL) {
2648 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2649 arg->align, arg->flags)) == NULL)
2652 struct uma_kctor_args karg;
2655 /* We should only be here from uma_startup() */
2656 karg.size = arg->size;
2657 karg.uminit = arg->uminit;
2658 karg.fini = arg->fini;
2659 karg.align = arg->align;
2660 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2662 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2668 /* Inherit properties from the keg. */
2670 zone->uz_size = keg->uk_size;
2671 zone->uz_flags |= (keg->uk_flags &
2672 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2675 if (__predict_true(booted >= BOOT_RUNNING)) {
2676 zone_alloc_counters(zone, NULL);
2677 zone_alloc_sysctl(zone, NULL);
2679 zone->uz_allocs = EARLY_COUNTER;
2680 zone->uz_frees = EARLY_COUNTER;
2681 zone->uz_fails = EARLY_COUNTER;
2684 /* Caller requests a private SMR context. */
2685 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2686 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2688 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2689 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2690 ("Invalid zone flag combination"));
2691 if (arg->flags & UMA_ZFLAG_INTERNAL)
2692 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2693 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2694 zone->uz_bucket_size = BUCKET_MAX;
2695 else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2696 zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2697 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2698 zone->uz_bucket_size = 0;
2700 zone->uz_bucket_size = bucket_select(zone->uz_size);
2701 zone->uz_bucket_size_min = zone->uz_bucket_size;
2702 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2703 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2704 zone_update_caches(zone);
2710 * Keg header dtor. This frees all data, destroys locks, frees the hash
2711 * table and removes the keg from the global list.
2713 * Arguments/Returns follow uma_dtor specifications
2717 keg_dtor(void *arg, int size, void *udata)
2720 uint32_t free, pages;
2723 keg = (uma_keg_t)arg;
2725 for (i = 0; i < vm_ndomains; i++) {
2726 free += keg->uk_domain[i].ud_free_items;
2727 pages += keg->uk_domain[i].ud_pages;
2728 KEG_LOCK_FINI(keg, i);
2731 printf("Freed UMA keg (%s) was not empty (%u items). "
2732 " Lost %u pages of memory.\n",
2733 keg->uk_name ? keg->uk_name : "",
2734 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2736 hash_free(&keg->uk_hash);
2742 * Arguments/Returns follow uma_dtor specifications
2746 zone_dtor(void *arg, int size, void *udata)
2752 zone = (uma_zone_t)arg;
2754 sysctl_remove_oid(zone->uz_oid, 1, 1);
2756 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2759 rw_wlock(&uma_rwlock);
2760 LIST_REMOVE(zone, uz_link);
2761 rw_wunlock(&uma_rwlock);
2762 zone_reclaim(zone, M_WAITOK, true);
2765 * We only destroy kegs from non secondary/non cache zones.
2767 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2769 rw_wlock(&uma_rwlock);
2770 LIST_REMOVE(keg, uk_link);
2771 rw_wunlock(&uma_rwlock);
2772 zone_free_item(kegs, keg, NULL, SKIP_NONE);
2774 counter_u64_free(zone->uz_allocs);
2775 counter_u64_free(zone->uz_frees);
2776 counter_u64_free(zone->uz_fails);
2777 counter_u64_free(zone->uz_xdomain);
2778 free(zone->uz_ctlname, M_UMA);
2779 for (i = 0; i < vm_ndomains; i++)
2780 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2781 ZONE_CROSS_LOCK_FINI(zone);
2785 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2790 LIST_FOREACH(keg, &uma_kegs, uk_link) {
2791 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2794 LIST_FOREACH(zone, &uma_cachezones, uz_link)
2799 * Traverses every zone in the system and calls a callback
2802 * zfunc A pointer to a function which accepts a zone
2809 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2812 rw_rlock(&uma_rwlock);
2813 zone_foreach_unlocked(zfunc, arg);
2814 rw_runlock(&uma_rwlock);
2818 * Initialize the kernel memory allocator. This is done after pages can be
2819 * allocated but before general KVA is available.
2822 uma_startup1(vm_offset_t virtual_avail)
2824 struct uma_zctor_args args;
2825 size_t ksize, zsize, size;
2826 uma_keg_t masterkeg;
2830 bootstart = bootmem = virtual_avail;
2832 rw_init(&uma_rwlock, "UMA lock");
2833 sx_init(&uma_reclaim_lock, "umareclaim");
2835 ksize = sizeof(struct uma_keg) +
2836 (sizeof(struct uma_domain) * vm_ndomains);
2837 ksize = roundup(ksize, UMA_SUPER_ALIGN);
2838 zsize = sizeof(struct uma_zone) +
2839 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2840 (sizeof(struct uma_zone_domain) * vm_ndomains);
2841 zsize = roundup(zsize, UMA_SUPER_ALIGN);
2843 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2844 size = (zsize * 2) + ksize;
2845 m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO);
2846 zones = (uma_zone_t)m;
2848 kegs = (uma_zone_t)m;
2850 masterkeg = (uma_keg_t)m;
2852 /* "manually" create the initial zone */
2853 memset(&args, 0, sizeof(args));
2854 args.name = "UMA Kegs";
2856 args.ctor = keg_ctor;
2857 args.dtor = keg_dtor;
2858 args.uminit = zero_init;
2860 args.keg = masterkeg;
2861 args.align = UMA_SUPER_ALIGN - 1;
2862 args.flags = UMA_ZFLAG_INTERNAL;
2863 zone_ctor(kegs, zsize, &args, M_WAITOK);
2865 args.name = "UMA Zones";
2867 args.ctor = zone_ctor;
2868 args.dtor = zone_dtor;
2869 args.uminit = zero_init;
2872 args.align = UMA_SUPER_ALIGN - 1;
2873 args.flags = UMA_ZFLAG_INTERNAL;
2874 zone_ctor(zones, zsize, &args, M_WAITOK);
2876 /* Now make zones for slab headers */
2877 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2878 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2879 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2880 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2882 hashzone = uma_zcreate("UMA Hash",
2883 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2884 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2890 #ifndef UMA_MD_SMALL_ALLOC
2891 extern void vm_radix_reserve_kva(void);
2895 * Advertise the availability of normal kva allocations and switch to
2896 * the default back-end allocator. Marks the KVA we consumed on startup
2897 * as used in the map.
2903 if (bootstart != bootmem) {
2904 vm_map_lock(kernel_map);
2905 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2906 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2907 vm_map_unlock(kernel_map);
2910 #ifndef UMA_MD_SMALL_ALLOC
2911 /* Set up radix zone to use noobj_alloc. */
2912 vm_radix_reserve_kva();
2916 zone_foreach_unlocked(zone_kva_available, NULL);
2921 * Finish our initialization steps.
2928 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2929 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2930 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2932 zone_foreach_unlocked(zone_alloc_counters, NULL);
2933 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2934 callout_init(&uma_callout, 1);
2935 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2936 booted = BOOT_RUNNING;
2938 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2939 EVENTHANDLER_PRI_FIRST);
2946 booted = BOOT_SHUTDOWN;
2950 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2951 int align, uint32_t flags)
2953 struct uma_kctor_args args;
2956 args.uminit = uminit;
2958 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2961 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2964 /* Public functions */
2967 uma_set_align(int align)
2970 if (align != UMA_ALIGN_CACHE)
2971 uma_align_cache = align;
2976 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2977 uma_init uminit, uma_fini fini, int align, uint32_t flags)
2980 struct uma_zctor_args args;
2983 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2986 /* This stuff is essential for the zone ctor */
2987 memset(&args, 0, sizeof(args));
2992 args.uminit = uminit;
2996 * Inject procedures which check for memory use after free if we are
2997 * allowed to scramble the memory while it is not allocated. This
2998 * requires that: UMA is actually able to access the memory, no init
2999 * or fini procedures, no dependency on the initial value of the
3000 * memory, and no (legitimate) use of the memory after free. Note,
3001 * the ctor and dtor do not need to be empty.
3003 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3004 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3005 args.uminit = trash_init;
3006 args.fini = trash_fini;
3013 sx_slock(&uma_reclaim_lock);
3014 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3015 sx_sunlock(&uma_reclaim_lock);
3022 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3023 uma_init zinit, uma_fini zfini, uma_zone_t master)
3025 struct uma_zctor_args args;
3029 keg = master->uz_keg;
3030 memset(&args, 0, sizeof(args));
3032 args.size = keg->uk_size;
3035 args.uminit = zinit;
3037 args.align = keg->uk_align;
3038 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3041 sx_slock(&uma_reclaim_lock);
3042 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3043 sx_sunlock(&uma_reclaim_lock);
3050 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3051 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3052 void *arg, int flags)
3054 struct uma_zctor_args args;
3056 memset(&args, 0, sizeof(args));
3061 args.uminit = zinit;
3063 args.import = zimport;
3064 args.release = zrelease;
3067 args.flags = flags | UMA_ZFLAG_CACHE;
3069 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3074 uma_zdestroy(uma_zone_t zone)
3078 * Large slabs are expensive to reclaim, so don't bother doing
3079 * unnecessary work if we're shutting down.
3081 if (booted == BOOT_SHUTDOWN &&
3082 zone->uz_fini == NULL && zone->uz_release == zone_release)
3084 sx_slock(&uma_reclaim_lock);
3085 zone_free_item(zones, zone, NULL, SKIP_NONE);
3086 sx_sunlock(&uma_reclaim_lock);
3090 uma_zwait(uma_zone_t zone)
3093 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3094 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3095 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3096 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3098 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3102 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3104 void *item, *pcpu_item;
3108 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3110 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3113 pcpu_item = zpcpu_base_to_offset(item);
3114 if (flags & M_ZERO) {
3116 for (i = 0; i <= mp_maxid; i++)
3117 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3119 bzero(item, zone->uz_size);
3126 * A stub while both regular and pcpu cases are identical.
3129 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3134 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3136 item = zpcpu_offset_to_base(pcpu_item);
3137 uma_zfree_arg(zone, item, udata);
3140 static inline void *
3141 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3147 skipdbg = uma_dbg_zskip(zone, item);
3148 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3149 zone->uz_ctor != trash_ctor)
3150 trash_ctor(item, size, udata, flags);
3152 /* Check flags before loading ctor pointer. */
3153 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3154 __predict_false(zone->uz_ctor != NULL) &&
3155 zone->uz_ctor(item, size, udata, flags) != 0) {
3156 counter_u64_add(zone->uz_fails, 1);
3157 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3162 uma_dbg_alloc(zone, NULL, item);
3164 if (__predict_false(flags & M_ZERO))
3165 return (memset(item, 0, size));
3171 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3172 enum zfreeskip skip)
3177 skipdbg = uma_dbg_zskip(zone, item);
3178 if (skip == SKIP_NONE && !skipdbg) {
3179 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3180 uma_dbg_free(zone, udata, item);
3182 uma_dbg_free(zone, NULL, item);
3185 if (__predict_true(skip < SKIP_DTOR)) {
3186 if (zone->uz_dtor != NULL)
3187 zone->uz_dtor(item, size, udata);
3189 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3190 zone->uz_dtor != trash_dtor)
3191 trash_dtor(item, size, udata);
3196 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3197 #define UMA_ZALLOC_DEBUG
3199 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3205 if (flags & M_WAITOK) {
3206 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3207 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3212 KASSERT((flags & M_EXEC) == 0,
3213 ("uma_zalloc_debug: called with M_EXEC"));
3214 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3215 ("uma_zalloc_debug: called within spinlock or critical section"));
3216 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3217 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3220 #ifdef DEBUG_MEMGUARD
3221 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3223 item = memguard_alloc(zone->uz_size, flags);
3225 error = EJUSTRETURN;
3226 if (zone->uz_init != NULL &&
3227 zone->uz_init(item, zone->uz_size, flags) != 0) {
3231 if (zone->uz_ctor != NULL &&
3232 zone->uz_ctor(item, zone->uz_size, udata,
3234 counter_u64_add(zone->uz_fails, 1);
3235 zone->uz_fini(item, zone->uz_size);
3242 /* This is unfortunate but should not be fatal. */
3249 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3251 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3252 ("uma_zfree_debug: called with spinlock or critical section held"));
3254 #ifdef DEBUG_MEMGUARD
3255 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3256 if (zone->uz_dtor != NULL)
3257 zone->uz_dtor(item, zone->uz_size, udata);
3258 if (zone->uz_fini != NULL)
3259 zone->uz_fini(item, zone->uz_size);
3260 memguard_free(item);
3261 return (EJUSTRETURN);
3268 static inline void *
3269 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3270 void *udata, int flags)
3275 item = cache_bucket_pop(cache, bucket);
3276 size = cache_uz_size(cache);
3277 uz_flags = cache_uz_flags(cache);
3279 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3282 static __noinline void *
3283 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3285 uma_cache_bucket_t bucket;
3288 while (cache_alloc(zone, cache, udata, flags)) {
3289 cache = &zone->uz_cpu[curcpu];
3290 bucket = &cache->uc_allocbucket;
3291 if (__predict_false(bucket->ucb_cnt == 0))
3293 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3298 * We can not get a bucket so try to return a single item.
3300 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3301 domain = PCPU_GET(domain);
3303 domain = UMA_ANYDOMAIN;
3304 return (zone_alloc_item(zone, udata, domain, flags));
3309 uma_zalloc_smr(uma_zone_t zone, int flags)
3311 uma_cache_bucket_t bucket;
3314 #ifdef UMA_ZALLOC_DEBUG
3317 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3318 ("uma_zalloc_arg: called with non-SMR zone.\n"));
3319 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3324 cache = &zone->uz_cpu[curcpu];
3325 bucket = &cache->uc_allocbucket;
3326 if (__predict_false(bucket->ucb_cnt == 0))
3327 return (cache_alloc_retry(zone, cache, NULL, flags));
3328 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3333 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3335 uma_cache_bucket_t bucket;
3338 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3339 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3341 /* This is the fast path allocation */
3342 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3345 #ifdef UMA_ZALLOC_DEBUG
3348 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3349 ("uma_zalloc_arg: called with SMR zone.\n"));
3350 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3355 * If possible, allocate from the per-CPU cache. There are two
3356 * requirements for safe access to the per-CPU cache: (1) the thread
3357 * accessing the cache must not be preempted or yield during access,
3358 * and (2) the thread must not migrate CPUs without switching which
3359 * cache it accesses. We rely on a critical section to prevent
3360 * preemption and migration. We release the critical section in
3361 * order to acquire the zone mutex if we are unable to allocate from
3362 * the current cache; when we re-acquire the critical section, we
3363 * must detect and handle migration if it has occurred.
3366 cache = &zone->uz_cpu[curcpu];
3367 bucket = &cache->uc_allocbucket;
3368 if (__predict_false(bucket->ucb_cnt == 0))
3369 return (cache_alloc_retry(zone, cache, udata, flags));
3370 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3374 * Replenish an alloc bucket and possibly restore an old one. Called in
3375 * a critical section. Returns in a critical section.
3377 * A false return value indicates an allocation failure.
3378 * A true return value indicates success and the caller should retry.
3380 static __noinline bool
3381 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3383 uma_bucket_t bucket;
3387 CRITICAL_ASSERT(curthread);
3390 * If we have run out of items in our alloc bucket see
3391 * if we can switch with the free bucket.
3393 * SMR Zones can't re-use the free bucket until the sequence has
3396 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3397 cache->uc_freebucket.ucb_cnt != 0) {
3398 cache_bucket_swap(&cache->uc_freebucket,
3399 &cache->uc_allocbucket);
3404 * Discard any empty allocation bucket while we hold no locks.
3406 bucket = cache_bucket_unload_alloc(cache);
3409 if (bucket != NULL) {
3410 KASSERT(bucket->ub_cnt == 0,
3411 ("cache_alloc: Entered with non-empty alloc bucket."));
3412 bucket_free(zone, bucket, udata);
3415 /* Short-circuit for zones without buckets and low memory. */
3416 if (zone->uz_bucket_size == 0 || bucketdisable) {
3422 * Attempt to retrieve the item from the per-CPU cache has failed, so
3423 * we must go back to the zone. This requires the zdom lock, so we
3424 * must drop the critical section, then re-acquire it when we go back
3425 * to the cache. Since the critical section is released, we may be
3426 * preempted or migrate. As such, make sure not to maintain any
3427 * thread-local state specific to the cache from prior to releasing
3428 * the critical section.
3430 domain = PCPU_GET(domain);
3431 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0)
3432 domain = zone_domain_highest(zone, domain);
3433 bucket = cache_fetch_bucket(zone, cache, domain);
3434 if (bucket == NULL) {
3435 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3440 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3441 zone->uz_name, zone, bucket);
3442 if (bucket == NULL) {
3448 * See if we lost the race or were migrated. Cache the
3449 * initialized bucket to make this less likely or claim
3450 * the memory directly.
3453 cache = &zone->uz_cpu[curcpu];
3454 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3455 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3456 domain == PCPU_GET(domain))) {
3458 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3460 cache_bucket_load_alloc(cache, bucket);
3465 * We lost the race, release this bucket and start over.
3468 zone_put_bucket(zone, domain, bucket, udata, false);
3475 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3478 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3479 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3481 /* This is the fast path allocation */
3482 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3483 zone->uz_name, zone, domain, flags);
3485 if (flags & M_WAITOK) {
3486 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3487 "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3489 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3490 ("uma_zalloc_domain: called with spinlock or critical section held"));
3492 return (zone_alloc_item(zone, udata, domain, flags));
3496 * Find a slab with some space. Prefer slabs that are partially used over those
3497 * that are totally full. This helps to reduce fragmentation.
3499 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3503 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3509 KASSERT(domain >= 0 && domain < vm_ndomains,
3510 ("keg_first_slab: domain %d out of range", domain));
3511 KEG_LOCK_ASSERT(keg, domain);
3516 dom = &keg->uk_domain[domain];
3517 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3519 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3520 LIST_REMOVE(slab, us_link);
3521 dom->ud_free_slabs--;
3522 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3526 domain = (domain + 1) % vm_ndomains;
3527 } while (domain != start);
3533 * Fetch an existing slab from a free or partial list. Returns with the
3534 * keg domain lock held if a slab was found or unlocked if not.
3537 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3542 /* HASH has a single free list. */
3543 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3546 KEG_LOCK(keg, domain);
3547 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3548 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3549 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3550 KEG_UNLOCK(keg, domain);
3557 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3559 struct vm_domainset_iter di;
3566 * Use the keg's policy if upper layers haven't already specified a
3567 * domain (as happens with first-touch zones).
3569 * To avoid races we run the iterator with the keg lock held, but that
3570 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3571 * clear M_WAITOK and handle low memory conditions locally.
3573 rr = rdomain == UMA_ANYDOMAIN;
3575 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3576 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3584 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3589 * M_NOVM means don't ask at all!
3594 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3597 if (!rr && (flags & M_WAITOK) == 0)
3599 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3600 if ((flags & M_WAITOK) != 0) {
3601 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
3609 * We might not have been able to get a slab but another cpu
3610 * could have while we were unlocked. Check again before we
3613 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3620 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3626 KEG_LOCK_ASSERT(keg, slab->us_domain);
3628 dom = &keg->uk_domain[slab->us_domain];
3629 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3630 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3631 item = slab_item(slab, keg, freei);
3632 slab->us_freecount--;
3633 dom->ud_free_items--;
3636 * Move this slab to the full list. It must be on the partial list, so
3637 * we do not need to update the free slab count. In particular,
3638 * keg_fetch_slab() always returns slabs on the partial list.
3640 if (slab->us_freecount == 0) {
3641 LIST_REMOVE(slab, us_link);
3642 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3649 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3663 /* Try to keep the buckets totally full */
3664 for (i = 0; i < max; ) {
3665 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3668 stripe = howmany(max, vm_ndomains);
3670 dom = &keg->uk_domain[slab->us_domain];
3671 while (slab->us_freecount && i < max) {
3672 bucket[i++] = slab_alloc_item(keg, slab);
3673 if (dom->ud_free_items <= keg->uk_reserve)
3677 * If the zone is striped we pick a new slab for every
3678 * N allocations. Eliminating this conditional will
3679 * instead pick a new domain for each bucket rather
3680 * than stripe within each bucket. The current option
3681 * produces more fragmentation and requires more cpu
3682 * time but yields better distribution.
3684 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3685 vm_ndomains > 1 && --stripe == 0)
3689 KEG_UNLOCK(keg, slab->us_domain);
3690 /* Don't block if we allocated any successfully. */
3699 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3701 uint64_t old, new, total, max;
3704 * The hard case. We're going to sleep because there were existing
3705 * sleepers or because we ran out of items. This routine enforces
3706 * fairness by keeping fifo order.
3708 * First release our ill gotten gains and make some noise.
3711 zone_free_limit(zone, count);
3712 zone_log_warning(zone);
3713 zone_maxaction(zone);
3714 if (flags & M_NOWAIT)
3718 * We need to allocate an item or set ourself as a sleeper
3719 * while the sleepq lock is held to avoid wakeup races. This
3720 * is essentially a home rolled semaphore.
3722 sleepq_lock(&zone->uz_max_items);
3723 old = zone->uz_items;
3725 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3726 /* Cache the max since we will evaluate twice. */
3727 max = zone->uz_max_items;
3728 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3729 UZ_ITEMS_COUNT(old) >= max)
3730 new = old + UZ_ITEMS_SLEEPER;
3732 new = old + MIN(count, max - old);
3733 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3735 /* We may have successfully allocated under the sleepq lock. */
3736 if (UZ_ITEMS_SLEEPERS(new) == 0) {
3737 sleepq_release(&zone->uz_max_items);
3742 * This is in a different cacheline from uz_items so that we
3743 * don't constantly invalidate the fastpath cacheline when we
3744 * adjust item counts. This could be limited to toggling on
3747 atomic_add_32(&zone->uz_sleepers, 1);
3748 atomic_add_64(&zone->uz_sleeps, 1);
3751 * We have added ourselves as a sleeper. The sleepq lock
3752 * protects us from wakeup races. Sleep now and then retry.
3754 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3755 sleepq_wait(&zone->uz_max_items, PVM);
3758 * After wakeup, remove ourselves as a sleeper and try
3759 * again. We no longer have the sleepq lock for protection.
3761 * Subract ourselves as a sleeper while attempting to add
3764 atomic_subtract_32(&zone->uz_sleepers, 1);
3765 old = atomic_fetchadd_64(&zone->uz_items,
3766 -(UZ_ITEMS_SLEEPER - count));
3767 /* We're no longer a sleeper. */
3768 old -= UZ_ITEMS_SLEEPER;
3771 * If we're still at the limit, restart. Notably do not
3772 * block on other sleepers. Cache the max value to protect
3773 * against changes via sysctl.
3775 total = UZ_ITEMS_COUNT(old);
3776 max = zone->uz_max_items;
3779 /* Truncate if necessary, otherwise wake other sleepers. */
3780 if (total + count > max) {
3781 zone_free_limit(zone, total + count - max);
3782 count = max - total;
3783 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3784 wakeup_one(&zone->uz_max_items);
3791 * Allocate 'count' items from our max_items limit. Returns the number
3792 * available. If M_NOWAIT is not specified it will sleep until at least
3793 * one item can be allocated.
3796 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3801 max = zone->uz_max_items;
3805 * We expect normal allocations to succeed with a simple
3808 old = atomic_fetchadd_64(&zone->uz_items, count);
3809 if (__predict_true(old + count <= max))
3813 * If we had some items and no sleepers just return the
3814 * truncated value. We have to release the excess space
3815 * though because that may wake sleepers who weren't woken
3816 * because we were temporarily over the limit.
3819 zone_free_limit(zone, (old + count) - max);
3822 return (zone_alloc_limit_hard(zone, count, flags));
3826 * Free a number of items back to the limit.
3829 zone_free_limit(uma_zone_t zone, int count)
3836 * In the common case we either have no sleepers or
3837 * are still over the limit and can just return.
3839 old = atomic_fetchadd_64(&zone->uz_items, -count);
3840 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3841 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3845 * Moderate the rate of wakeups. Sleepers will continue
3846 * to generate wakeups if necessary.
3848 wakeup_one(&zone->uz_max_items);
3852 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3854 uma_bucket_t bucket;
3857 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3860 /* Avoid allocs targeting empty domains. */
3861 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3862 domain = UMA_ANYDOMAIN;
3863 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
3864 domain = UMA_ANYDOMAIN;
3866 if (zone->uz_max_items > 0)
3867 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3870 maxbucket = zone->uz_bucket_size;
3874 /* Don't wait for buckets, preserve caller's NOVM setting. */
3875 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3876 if (bucket == NULL) {
3881 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3882 MIN(maxbucket, bucket->ub_entries), domain, flags);
3885 * Initialize the memory if necessary.
3887 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3890 for (i = 0; i < bucket->ub_cnt; i++)
3891 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3895 * If we couldn't initialize the whole bucket, put the
3896 * rest back onto the freelist.
3898 if (i != bucket->ub_cnt) {
3899 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3900 bucket->ub_cnt - i);
3902 bzero(&bucket->ub_bucket[i],
3903 sizeof(void *) * (bucket->ub_cnt - i));
3909 cnt = bucket->ub_cnt;
3910 if (bucket->ub_cnt == 0) {
3911 bucket_free(zone, bucket, udata);
3912 counter_u64_add(zone->uz_fails, 1);
3916 if (zone->uz_max_items > 0 && cnt < maxbucket)
3917 zone_free_limit(zone, maxbucket - cnt);
3923 * Allocates a single item from a zone.
3926 * zone The zone to alloc for.
3927 * udata The data to be passed to the constructor.
3928 * domain The domain to allocate from or UMA_ANYDOMAIN.
3929 * flags M_WAITOK, M_NOWAIT, M_ZERO.
3932 * NULL if there is no memory and M_NOWAIT is set
3933 * An item if successful
3937 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
3941 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0)
3944 /* Avoid allocs targeting empty domains. */
3945 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3946 domain = UMA_ANYDOMAIN;
3948 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
3952 * We have to call both the zone's init (not the keg's init)
3953 * and the zone's ctor. This is because the item is going from
3954 * a keg slab directly to the user, and the user is expecting it
3955 * to be both zone-init'd as well as zone-ctor'd.
3957 if (zone->uz_init != NULL) {
3958 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
3959 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
3963 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
3968 counter_u64_add(zone->uz_allocs, 1);
3969 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
3970 zone->uz_name, zone);
3975 counter_u64_add(zone->uz_fails, 1);
3977 if (zone->uz_max_items > 0)
3978 zone_free_limit(zone, 1);
3979 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
3980 zone->uz_name, zone);
3987 uma_zfree_smr(uma_zone_t zone, void *item)
3990 uma_cache_bucket_t bucket;
3991 int itemdomain, uz_flags;
3993 #ifdef UMA_ZALLOC_DEBUG
3994 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3995 ("uma_zfree_smr: called with non-SMR zone.\n"));
3996 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
3997 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
3998 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4001 cache = &zone->uz_cpu[curcpu];
4002 uz_flags = cache_uz_flags(cache);
4005 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4006 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4010 cache = &zone->uz_cpu[curcpu];
4011 /* SMR Zones must free to the free bucket. */
4012 bucket = &cache->uc_freebucket;
4014 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4015 PCPU_GET(domain) != itemdomain) {
4016 bucket = &cache->uc_crossbucket;
4019 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4020 cache_bucket_push(cache, bucket, item);
4024 } while (cache_free(zone, cache, NULL, item, itemdomain));
4028 * If nothing else caught this, we'll just do an internal free.
4030 zone_free_item(zone, item, NULL, SKIP_NONE);
4035 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4038 uma_cache_bucket_t bucket;
4039 int itemdomain, uz_flags;
4041 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4042 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4044 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4046 #ifdef UMA_ZALLOC_DEBUG
4047 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4048 ("uma_zfree_arg: called with SMR zone.\n"));
4049 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4052 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4057 * We are accessing the per-cpu cache without a critical section to
4058 * fetch size and flags. This is acceptable, if we are preempted we
4059 * will simply read another cpu's line.
4061 cache = &zone->uz_cpu[curcpu];
4062 uz_flags = cache_uz_flags(cache);
4063 if (UMA_ALWAYS_CTORDTOR ||
4064 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4065 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4068 * The race here is acceptable. If we miss it we'll just have to wait
4069 * a little longer for the limits to be reset.
4071 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4072 if (zone->uz_sleepers > 0)
4077 * If possible, free to the per-CPU cache. There are two
4078 * requirements for safe access to the per-CPU cache: (1) the thread
4079 * accessing the cache must not be preempted or yield during access,
4080 * and (2) the thread must not migrate CPUs without switching which
4081 * cache it accesses. We rely on a critical section to prevent
4082 * preemption and migration. We release the critical section in
4083 * order to acquire the zone mutex if we are unable to free to the
4084 * current cache; when we re-acquire the critical section, we must
4085 * detect and handle migration if it has occurred.
4089 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4090 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4094 cache = &zone->uz_cpu[curcpu];
4096 * Try to free into the allocbucket first to give LIFO
4097 * ordering for cache-hot datastructures. Spill over
4098 * into the freebucket if necessary. Alloc will swap
4099 * them if one runs dry.
4101 bucket = &cache->uc_allocbucket;
4103 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4104 PCPU_GET(domain) != itemdomain) {
4105 bucket = &cache->uc_crossbucket;
4108 if (bucket->ucb_cnt == bucket->ucb_entries &&
4109 cache->uc_freebucket.ucb_cnt <
4110 cache->uc_freebucket.ucb_entries)
4111 cache_bucket_swap(&cache->uc_freebucket,
4112 &cache->uc_allocbucket);
4113 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4114 cache_bucket_push(cache, bucket, item);
4118 } while (cache_free(zone, cache, udata, item, itemdomain));
4122 * If nothing else caught this, we'll just do an internal free.
4125 zone_free_item(zone, item, udata, SKIP_DTOR);
4130 * sort crossdomain free buckets to domain correct buckets and cache
4134 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4136 struct uma_bucketlist fullbuckets;
4137 uma_zone_domain_t zdom;
4144 "uma_zfree: zone %s(%p) draining cross bucket %p",
4145 zone->uz_name, zone, bucket);
4148 * It is possible for buckets to arrive here out of order so we fetch
4149 * the current smr seq rather than accepting the bucket's.
4151 seq = SMR_SEQ_INVALID;
4152 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4153 seq = smr_advance(zone->uz_smr);
4156 * To avoid having ndomain * ndomain buckets for sorting we have a
4157 * lock on the current crossfree bucket. A full matrix with
4158 * per-domain locking could be used if necessary.
4160 STAILQ_INIT(&fullbuckets);
4161 ZONE_CROSS_LOCK(zone);
4162 while (bucket->ub_cnt > 0) {
4163 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4164 domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4165 zdom = ZDOM_GET(zone, domain);
4166 if (zdom->uzd_cross == NULL) {
4167 zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
4168 if (zdom->uzd_cross == NULL)
4171 b = zdom->uzd_cross;
4172 b->ub_bucket[b->ub_cnt++] = item;
4174 if (b->ub_cnt == b->ub_entries) {
4175 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4176 zdom->uzd_cross = NULL;
4180 ZONE_CROSS_UNLOCK(zone);
4181 if (bucket->ub_cnt == 0)
4182 bucket->ub_seq = SMR_SEQ_INVALID;
4183 bucket_free(zone, bucket, udata);
4185 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4186 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4187 domain = _vm_phys_domain(pmap_kextract(
4188 (vm_offset_t)b->ub_bucket[0]));
4189 zone_put_bucket(zone, domain, b, udata, true);
4195 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4196 int itemdomain, bool ws)
4201 * Buckets coming from the wrong domain will be entirely for the
4202 * only other domain on two domain systems. In this case we can
4203 * simply cache them. Otherwise we need to sort them back to
4206 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4207 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4208 zone_free_cross(zone, bucket, udata);
4214 * Attempt to save the bucket in the zone's domain bucket cache.
4217 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4218 zone->uz_name, zone, bucket);
4219 /* ub_cnt is pointing to the last free item */
4220 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4221 itemdomain = zone_domain_lowest(zone, itemdomain);
4222 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4226 * Populate a free or cross bucket for the current cpu cache. Free any
4227 * existing full bucket either to the zone cache or back to the slab layer.
4229 * Enters and returns in a critical section. false return indicates that
4230 * we can not satisfy this free in the cache layer. true indicates that
4231 * the caller should retry.
4233 static __noinline bool
4234 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4237 uma_cache_bucket_t cbucket;
4238 uma_bucket_t newbucket, bucket;
4240 CRITICAL_ASSERT(curthread);
4242 if (zone->uz_bucket_size == 0)
4245 cache = &zone->uz_cpu[curcpu];
4249 * FIRSTTOUCH domains need to free to the correct zdom. When
4250 * enabled this is the zdom of the item. The bucket is the
4251 * cross bucket if the current domain and itemdomain do not match.
4253 cbucket = &cache->uc_freebucket;
4255 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4256 if (PCPU_GET(domain) != itemdomain) {
4257 cbucket = &cache->uc_crossbucket;
4258 if (cbucket->ucb_cnt != 0)
4259 counter_u64_add(zone->uz_xdomain,
4264 bucket = cache_bucket_unload(cbucket);
4265 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4266 ("cache_free: Entered with non-full free bucket."));
4268 /* We are no longer associated with this CPU. */
4272 * Don't let SMR zones operate without a free bucket. Force
4273 * a synchronize and re-use this one. We will only degrade
4274 * to a synchronize every bucket_size items rather than every
4275 * item if we fail to allocate a bucket.
4277 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4279 bucket->ub_seq = smr_advance(zone->uz_smr);
4280 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4281 if (newbucket == NULL && bucket != NULL) {
4282 bucket_drain(zone, bucket);
4286 } else if (!bucketdisable)
4287 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4290 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4293 if ((bucket = newbucket) == NULL)
4295 cache = &zone->uz_cpu[curcpu];
4298 * Check to see if we should be populating the cross bucket. If it
4299 * is already populated we will fall through and attempt to populate
4302 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4303 if (PCPU_GET(domain) != itemdomain &&
4304 cache->uc_crossbucket.ucb_bucket == NULL) {
4305 cache_bucket_load_cross(cache, bucket);
4311 * We may have lost the race to fill the bucket or switched CPUs.
4313 if (cache->uc_freebucket.ucb_bucket != NULL) {
4315 bucket_free(zone, bucket, udata);
4318 cache_bucket_load_free(cache, bucket);
4324 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
4327 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4328 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4330 CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone);
4332 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
4333 ("uma_zfree_domain: called with spinlock or critical section held"));
4335 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4338 zone_free_item(zone, item, udata, SKIP_NONE);
4342 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4349 KEG_LOCK_ASSERT(keg, slab->us_domain);
4351 /* Do we need to remove from any lists? */
4352 dom = &keg->uk_domain[slab->us_domain];
4353 if (slab->us_freecount + 1 == keg->uk_ipers) {
4354 LIST_REMOVE(slab, us_link);
4355 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4356 dom->ud_free_slabs++;
4357 } else if (slab->us_freecount == 0) {
4358 LIST_REMOVE(slab, us_link);
4359 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4362 /* Slab management. */
4363 freei = slab_item_index(slab, keg, item);
4364 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4365 slab->us_freecount++;
4367 /* Keg statistics. */
4368 dom->ud_free_items++;
4372 zone_release(void *arg, void **bucket, int cnt)
4385 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4386 lock = KEG_LOCK(keg, 0);
4387 for (i = 0; i < cnt; i++) {
4389 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4390 slab = vtoslab((vm_offset_t)item);
4392 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4393 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4394 slab = hash_sfind(&keg->uk_hash, mem);
4396 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4398 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4401 lock = KEG_LOCK(keg, slab->us_domain);
4403 slab_free_item(zone, slab, item);
4410 * Frees a single item to any zone.
4413 * zone The zone to free to
4414 * item The item we're freeing
4415 * udata User supplied data for the dtor
4416 * skip Skip dtors and finis
4418 static __noinline void
4419 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4423 * If a free is sent directly to an SMR zone we have to
4424 * synchronize immediately because the item can instantly
4425 * be reallocated. This should only happen in degenerate
4426 * cases when no memory is available for per-cpu caches.
4428 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4429 smr_synchronize(zone->uz_smr);
4431 item_dtor(zone, item, zone->uz_size, udata, skip);
4433 if (skip < SKIP_FINI && zone->uz_fini)
4434 zone->uz_fini(item, zone->uz_size);
4436 zone->uz_release(zone->uz_arg, &item, 1);
4438 if (skip & SKIP_CNT)
4441 counter_u64_add(zone->uz_frees, 1);
4443 if (zone->uz_max_items > 0)
4444 zone_free_limit(zone, 1);
4449 uma_zone_set_max(uma_zone_t zone, int nitems)
4451 struct uma_bucket_zone *ubz;
4455 * XXX This can misbehave if the zone has any allocations with
4456 * no limit and a limit is imposed. There is currently no
4457 * way to clear a limit.
4460 ubz = bucket_zone_max(zone, nitems);
4461 count = ubz != NULL ? ubz->ubz_entries : 0;
4462 zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4463 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4464 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4465 zone->uz_max_items = nitems;
4466 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4467 zone_update_caches(zone);
4468 /* We may need to wake waiters. */
4469 wakeup(&zone->uz_max_items);
4477 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4479 struct uma_bucket_zone *ubz;
4483 ubz = bucket_zone_max(zone, nitems);
4486 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4487 /* Count the cross-domain bucket. */
4489 nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4490 zone->uz_bucket_size_max = ubz->ubz_entries;
4492 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4494 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4495 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4496 zone->uz_bucket_max = nitems / vm_ndomains;
4502 uma_zone_get_max(uma_zone_t zone)
4506 nitems = atomic_load_64(&zone->uz_max_items);
4513 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4516 ZONE_ASSERT_COLD(zone);
4517 zone->uz_warning = warning;
4522 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4525 ZONE_ASSERT_COLD(zone);
4526 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4531 uma_zone_get_cur(uma_zone_t zone)
4537 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4538 nitems = counter_u64_fetch(zone->uz_allocs) -
4539 counter_u64_fetch(zone->uz_frees);
4541 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4542 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4544 return (nitems < 0 ? 0 : nitems);
4548 uma_zone_get_allocs(uma_zone_t zone)
4554 if (zone->uz_allocs != EARLY_COUNTER)
4555 nitems = counter_u64_fetch(zone->uz_allocs);
4557 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4563 uma_zone_get_frees(uma_zone_t zone)
4569 if (zone->uz_frees != EARLY_COUNTER)
4570 nitems = counter_u64_fetch(zone->uz_frees);
4572 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4578 /* Used only for KEG_ASSERT_COLD(). */
4580 uma_keg_get_allocs(uma_keg_t keg)
4586 LIST_FOREACH(z, &keg->uk_zones, uz_link)
4587 nitems += uma_zone_get_allocs(z);
4595 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4600 KEG_ASSERT_COLD(keg);
4601 keg->uk_init = uminit;
4606 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4611 KEG_ASSERT_COLD(keg);
4612 keg->uk_fini = fini;
4617 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4620 ZONE_ASSERT_COLD(zone);
4621 zone->uz_init = zinit;
4626 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4629 ZONE_ASSERT_COLD(zone);
4630 zone->uz_fini = zfini;
4635 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4640 KEG_ASSERT_COLD(keg);
4641 keg->uk_freef = freef;
4646 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4651 KEG_ASSERT_COLD(keg);
4652 keg->uk_allocf = allocf;
4657 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4660 ZONE_ASSERT_COLD(zone);
4662 KASSERT(smr != NULL, ("Got NULL smr"));
4663 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4664 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
4665 zone->uz_flags |= UMA_ZONE_SMR;
4667 zone_update_caches(zone);
4671 uma_zone_get_smr(uma_zone_t zone)
4674 return (zone->uz_smr);
4679 uma_zone_reserve(uma_zone_t zone, int items)
4684 KEG_ASSERT_COLD(keg);
4685 keg->uk_reserve = items;
4690 uma_zone_reserve_kva(uma_zone_t zone, int count)
4697 KEG_ASSERT_COLD(keg);
4698 ZONE_ASSERT_COLD(zone);
4700 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4702 #ifdef UMA_MD_SMALL_ALLOC
4703 if (keg->uk_ppera > 1) {
4707 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4713 MPASS(keg->uk_kva == 0);
4716 zone->uz_max_items = pages * keg->uk_ipers;
4717 #ifdef UMA_MD_SMALL_ALLOC
4718 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4720 keg->uk_allocf = noobj_alloc;
4722 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4723 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4724 zone_update_caches(zone);
4731 uma_prealloc(uma_zone_t zone, int items)
4733 struct vm_domainset_iter di;
4737 int aflags, domain, slabs;
4740 slabs = howmany(items, keg->uk_ipers);
4741 while (slabs-- > 0) {
4743 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4746 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4749 dom = &keg->uk_domain[slab->us_domain];
4751 * keg_alloc_slab() always returns a slab on the
4754 LIST_REMOVE(slab, us_link);
4755 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4757 dom->ud_free_slabs++;
4758 KEG_UNLOCK(keg, slab->us_domain);
4761 if (vm_domainset_iter_policy(&di, &domain) != 0)
4762 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
4768 * Returns a snapshot of memory consumption in bytes.
4771 uma_zone_memory(uma_zone_t zone)
4777 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
4778 for (i = 0; i < vm_ndomains; i++)
4779 sz += ZDOM_GET(zone, i)->uzd_nitems;
4780 return (sz * zone->uz_size);
4782 for (i = 0; i < vm_ndomains; i++)
4783 sz += zone->uz_keg->uk_domain[i].ud_pages;
4785 return (sz * PAGE_SIZE);
4790 uma_reclaim(int req)
4793 CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4794 sx_xlock(&uma_reclaim_lock);
4798 case UMA_RECLAIM_TRIM:
4799 zone_foreach(zone_trim, NULL);
4801 case UMA_RECLAIM_DRAIN:
4802 case UMA_RECLAIM_DRAIN_CPU:
4803 zone_foreach(zone_drain, NULL);
4804 if (req == UMA_RECLAIM_DRAIN_CPU) {
4805 pcpu_cache_drain_safe(NULL);
4806 zone_foreach(zone_drain, NULL);
4810 panic("unhandled reclamation request %d", req);
4814 * Some slabs may have been freed but this zone will be visited early
4815 * we visit again so that we can free pages that are empty once other
4816 * zones are drained. We have to do the same for buckets.
4818 zone_drain(slabzones[0], NULL);
4819 zone_drain(slabzones[1], NULL);
4820 bucket_zone_drain();
4821 sx_xunlock(&uma_reclaim_lock);
4824 static volatile int uma_reclaim_needed;
4827 uma_reclaim_wakeup(void)
4830 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4831 wakeup(uma_reclaim);
4835 uma_reclaim_worker(void *arg __unused)
4839 sx_xlock(&uma_reclaim_lock);
4840 while (atomic_load_int(&uma_reclaim_needed) == 0)
4841 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4843 sx_xunlock(&uma_reclaim_lock);
4844 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4845 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4846 atomic_store_int(&uma_reclaim_needed, 0);
4847 /* Don't fire more than once per-second. */
4848 pause("umarclslp", hz);
4854 uma_zone_reclaim(uma_zone_t zone, int req)
4858 case UMA_RECLAIM_TRIM:
4859 zone_trim(zone, NULL);
4861 case UMA_RECLAIM_DRAIN:
4862 zone_drain(zone, NULL);
4864 case UMA_RECLAIM_DRAIN_CPU:
4865 pcpu_cache_drain_safe(zone);
4866 zone_drain(zone, NULL);
4869 panic("unhandled reclamation request %d", req);
4875 uma_zone_exhausted(uma_zone_t zone)
4878 return (atomic_load_32(&zone->uz_sleepers) > 0);
4885 return (uma_kmem_limit);
4889 uma_set_limit(unsigned long limit)
4892 uma_kmem_limit = limit;
4899 return (atomic_load_long(&uma_kmem_total));
4906 return (uma_kmem_limit - uma_size());
4911 * Generate statistics across both the zone and its per-cpu cache's. Return
4912 * desired statistics if the pointer is non-NULL for that statistic.
4914 * Note: does not update the zone statistics, as it can't safely clear the
4915 * per-CPU cache statistic.
4919 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4920 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4923 uint64_t allocs, frees, sleeps, xdomain;
4926 allocs = frees = sleeps = xdomain = 0;
4929 cache = &z->uz_cpu[cpu];
4930 cachefree += cache->uc_allocbucket.ucb_cnt;
4931 cachefree += cache->uc_freebucket.ucb_cnt;
4932 xdomain += cache->uc_crossbucket.ucb_cnt;
4933 cachefree += cache->uc_crossbucket.ucb_cnt;
4934 allocs += cache->uc_allocs;
4935 frees += cache->uc_frees;
4937 allocs += counter_u64_fetch(z->uz_allocs);
4938 frees += counter_u64_fetch(z->uz_frees);
4939 xdomain += counter_u64_fetch(z->uz_xdomain);
4940 sleeps += z->uz_sleeps;
4941 if (cachefreep != NULL)
4942 *cachefreep = cachefree;
4943 if (allocsp != NULL)
4947 if (sleepsp != NULL)
4949 if (xdomainp != NULL)
4950 *xdomainp = xdomain;
4955 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
4962 rw_rlock(&uma_rwlock);
4963 LIST_FOREACH(kz, &uma_kegs, uk_link) {
4964 LIST_FOREACH(z, &kz->uk_zones, uz_link)
4967 LIST_FOREACH(z, &uma_cachezones, uz_link)
4970 rw_runlock(&uma_rwlock);
4971 return (sysctl_handle_int(oidp, &count, 0, req));
4975 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
4976 struct uma_percpu_stat *ups, bool internal)
4978 uma_zone_domain_t zdom;
4983 for (i = 0; i < vm_ndomains; i++) {
4984 zdom = ZDOM_GET(z, i);
4985 uth->uth_zone_free += zdom->uzd_nitems;
4987 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
4988 uth->uth_frees = counter_u64_fetch(z->uz_frees);
4989 uth->uth_fails = counter_u64_fetch(z->uz_fails);
4990 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
4991 uth->uth_sleeps = z->uz_sleeps;
4993 for (i = 0; i < mp_maxid + 1; i++) {
4994 bzero(&ups[i], sizeof(*ups));
4995 if (internal || CPU_ABSENT(i))
4997 cache = &z->uz_cpu[i];
4998 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
4999 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5000 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5001 ups[i].ups_allocs = cache->uc_allocs;
5002 ups[i].ups_frees = cache->uc_frees;
5007 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5009 struct uma_stream_header ush;
5010 struct uma_type_header uth;
5011 struct uma_percpu_stat *ups;
5016 uint32_t kfree, pages;
5017 int count, error, i;
5019 error = sysctl_wire_old_buffer(req, 0);
5022 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5023 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5024 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5027 rw_rlock(&uma_rwlock);
5028 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5029 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5033 LIST_FOREACH(z, &uma_cachezones, uz_link)
5037 * Insert stream header.
5039 bzero(&ush, sizeof(ush));
5040 ush.ush_version = UMA_STREAM_VERSION;
5041 ush.ush_maxcpus = (mp_maxid + 1);
5042 ush.ush_count = count;
5043 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5045 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5047 for (i = 0; i < vm_ndomains; i++) {
5048 kfree += kz->uk_domain[i].ud_free_items;
5049 pages += kz->uk_domain[i].ud_pages;
5051 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5052 bzero(&uth, sizeof(uth));
5053 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5054 uth.uth_align = kz->uk_align;
5055 uth.uth_size = kz->uk_size;
5056 uth.uth_rsize = kz->uk_rsize;
5057 if (z->uz_max_items > 0) {
5058 items = UZ_ITEMS_COUNT(z->uz_items);
5059 uth.uth_pages = (items / kz->uk_ipers) *
5062 uth.uth_pages = pages;
5063 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5065 uth.uth_limit = z->uz_max_items;
5066 uth.uth_keg_free = kfree;
5069 * A zone is secondary is it is not the first entry
5070 * on the keg's zone list.
5072 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5073 (LIST_FIRST(&kz->uk_zones) != z))
5074 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5075 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5076 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5077 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5078 for (i = 0; i < mp_maxid + 1; i++)
5079 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5082 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5083 bzero(&uth, sizeof(uth));
5084 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5085 uth.uth_size = z->uz_size;
5086 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5087 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5088 for (i = 0; i < mp_maxid + 1; i++)
5089 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5092 rw_runlock(&uma_rwlock);
5093 error = sbuf_finish(&sbuf);
5100 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5102 uma_zone_t zone = *(uma_zone_t *)arg1;
5105 max = uma_zone_get_max(zone);
5106 error = sysctl_handle_int(oidp, &max, 0, req);
5107 if (error || !req->newptr)
5110 uma_zone_set_max(zone, max);
5116 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5122 * Some callers want to add sysctls for global zones that
5123 * may not yet exist so they pass a pointer to a pointer.
5126 zone = *(uma_zone_t *)arg1;
5129 cur = uma_zone_get_cur(zone);
5130 return (sysctl_handle_int(oidp, &cur, 0, req));
5134 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5136 uma_zone_t zone = arg1;
5139 cur = uma_zone_get_allocs(zone);
5140 return (sysctl_handle_64(oidp, &cur, 0, req));
5144 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5146 uma_zone_t zone = arg1;
5149 cur = uma_zone_get_frees(zone);
5150 return (sysctl_handle_64(oidp, &cur, 0, req));
5154 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5157 uma_zone_t zone = arg1;
5160 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5161 if (zone->uz_flags != 0)
5162 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5164 sbuf_printf(&sbuf, "0");
5165 error = sbuf_finish(&sbuf);
5172 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5174 uma_keg_t keg = arg1;
5175 int avail, effpct, total;
5177 total = keg->uk_ppera * PAGE_SIZE;
5178 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5179 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5181 * We consider the client's requested size and alignment here, not the
5182 * real size determination uk_rsize, because we also adjust the real
5183 * size for internal implementation reasons (max bitset size).
5185 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5186 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5187 avail *= mp_maxid + 1;
5188 effpct = 100 * avail / total;
5189 return (sysctl_handle_int(oidp, &effpct, 0, req));
5193 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5195 uma_zone_t zone = arg1;
5198 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5199 return (sysctl_handle_64(oidp, &cur, 0, req));
5204 uma_dbg_getslab(uma_zone_t zone, void *item)
5211 * It is safe to return the slab here even though the
5212 * zone is unlocked because the item's allocation state
5213 * essentially holds a reference.
5215 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5216 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5218 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5219 return (vtoslab((vm_offset_t)mem));
5221 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5222 return ((uma_slab_t)(mem + keg->uk_pgoff));
5224 slab = hash_sfind(&keg->uk_hash, mem);
5231 uma_dbg_zskip(uma_zone_t zone, void *mem)
5234 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5237 return (uma_dbg_kskip(zone->uz_keg, mem));
5241 uma_dbg_kskip(uma_keg_t keg, void *mem)
5245 if (dbg_divisor == 0)
5248 if (dbg_divisor == 1)
5251 idx = (uintptr_t)mem >> PAGE_SHIFT;
5252 if (keg->uk_ipers > 1) {
5253 idx *= keg->uk_ipers;
5254 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5257 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5258 counter_u64_add(uma_skip_cnt, 1);
5261 counter_u64_add(uma_dbg_cnt, 1);
5267 * Set up the slab's freei data such that uma_dbg_free can function.
5271 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5277 slab = uma_dbg_getslab(zone, item);
5279 panic("uma: item %p did not belong to zone %s\n",
5280 item, zone->uz_name);
5283 freei = slab_item_index(slab, keg, item);
5285 if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5286 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
5287 item, zone, zone->uz_name, slab, freei);
5288 BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5292 * Verifies freed addresses. Checks for alignment, valid slab membership
5293 * and duplicate frees.
5297 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5303 slab = uma_dbg_getslab(zone, item);
5305 panic("uma: Freed item %p did not belong to zone %s\n",
5306 item, zone->uz_name);
5309 freei = slab_item_index(slab, keg, item);
5311 if (freei >= keg->uk_ipers)
5312 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
5313 item, zone, zone->uz_name, slab, freei);
5315 if (slab_item(slab, keg, freei) != item)
5316 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
5317 item, zone, zone->uz_name, slab, freei);
5319 if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5320 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
5321 item, zone, zone->uz_name, slab, freei);
5323 BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5325 #endif /* INVARIANTS */
5329 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5330 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5335 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5336 *allocs = counter_u64_fetch(z->uz_allocs);
5337 frees = counter_u64_fetch(z->uz_frees);
5338 *sleeps = z->uz_sleeps;
5342 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5344 for (i = 0; i < vm_ndomains; i++) {
5345 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5346 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5347 (LIST_FIRST(&kz->uk_zones) != z)))
5348 *cachefree += kz->uk_domain[i].ud_free_items;
5350 *used = *allocs - frees;
5351 return (((int64_t)*used + *cachefree) * kz->uk_size);
5354 DB_SHOW_COMMAND(uma, db_show_uma)
5356 const char *fmt_hdr, *fmt_entry;
5359 uint64_t allocs, used, sleeps, xdomain;
5361 /* variables for sorting */
5363 uma_zone_t cur_zone, last_zone;
5364 int64_t cur_size, last_size, size;
5367 /* /i option produces machine-parseable CSV output */
5368 if (modif[0] == 'i') {
5369 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5370 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5372 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5373 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5376 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5377 "Sleeps", "Bucket", "Total Mem", "XFree");
5379 /* Sort the zones with largest size first. */
5381 last_size = INT64_MAX;
5386 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5387 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5389 * In the case of size ties, print out zones
5390 * in the order they are encountered. That is,
5391 * when we encounter the most recently output
5392 * zone, we have already printed all preceding
5393 * ties, and we must print all following ties.
5395 if (z == last_zone) {
5399 size = get_uma_stats(kz, z, &allocs, &used,
5400 &sleeps, &cachefree, &xdomain);
5401 if (size > cur_size && size < last_size + ties)
5409 if (cur_zone == NULL)
5412 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5413 &sleeps, &cachefree, &xdomain);
5414 db_printf(fmt_entry, cur_zone->uz_name,
5415 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5416 (uintmax_t)allocs, (uintmax_t)sleeps,
5417 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5422 last_zone = cur_zone;
5423 last_size = cur_size;
5427 DB_SHOW_COMMAND(umacache, db_show_umacache)
5430 uint64_t allocs, frees;
5434 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5435 "Requests", "Bucket");
5436 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5437 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5438 for (i = 0; i < vm_ndomains; i++)
5439 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5440 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5441 z->uz_name, (uintmax_t)z->uz_size,
5442 (intmax_t)(allocs - frees), cachefree,
5443 (uintmax_t)allocs, z->uz_bucket_size);