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 {
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_foreach(void (*zfunc)(uma_zone_t, void *), void *);
289 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
290 static void zone_timeout(uma_zone_t zone, void *);
291 static int hash_alloc(struct uma_hash *, u_int);
292 static int hash_expand(struct uma_hash *, struct uma_hash *);
293 static void hash_free(struct uma_hash *hash);
294 static void uma_timeout(void *);
295 static void uma_startup3(void);
296 static void uma_shutdown(void);
297 static void *zone_alloc_item(uma_zone_t, void *, int, int);
298 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
299 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
300 static void zone_free_limit(uma_zone_t zone, int count);
301 static void bucket_enable(void);
302 static void bucket_init(void);
303 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
304 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
305 static void bucket_zone_drain(void);
306 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
307 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
308 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
309 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
310 uma_fini fini, int align, uint32_t flags);
311 static int zone_import(void *, void **, int, int, int);
312 static void zone_release(void *, void **, int);
313 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
314 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
316 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
317 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
318 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
319 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
320 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
321 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
322 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
324 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
326 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD, 0,
327 "Memory allocation debugging");
330 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
331 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
333 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
334 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
335 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
336 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
338 static u_int dbg_divisor = 1;
339 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
340 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
341 "Debug & thrash every this item in memory allocator");
343 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
344 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
345 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
346 &uma_dbg_cnt, "memory items debugged");
347 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
348 &uma_skip_cnt, "memory items skipped, not debugged");
351 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
353 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW, 0, "Universal Memory Allocator");
355 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
356 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
358 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
359 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
361 static int zone_warnings = 1;
362 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
363 "Warn when UMA zones becomes full");
365 static int multipage_slabs = 1;
366 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
367 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
368 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
369 "UMA may choose larger slab sizes for better efficiency");
372 * Select the slab zone for an offpage slab with the given maximum item count.
374 static inline uma_zone_t
378 return (slabzones[ipers > SLABZONE0_SETSIZE]);
382 * This routine checks to see whether or not it's safe to enable buckets.
388 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
389 bucketdisable = vm_page_count_min();
393 * Initialize bucket_zones, the array of zones of buckets of various sizes.
395 * For each zone, calculate the memory required for each bucket, consisting
396 * of the header and an array of pointers.
401 struct uma_bucket_zone *ubz;
404 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
405 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
406 size += sizeof(void *) * ubz->ubz_entries;
407 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
408 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
409 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
410 UMA_ZONE_FIRSTTOUCH);
415 * Given a desired number of entries for a bucket, return the zone from which
416 * to allocate the bucket.
418 static struct uma_bucket_zone *
419 bucket_zone_lookup(int entries)
421 struct uma_bucket_zone *ubz;
423 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
424 if (ubz->ubz_entries >= entries)
430 static struct uma_bucket_zone *
431 bucket_zone_max(uma_zone_t zone, int nitems)
433 struct uma_bucket_zone *ubz;
437 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
438 /* Count the cross-domain bucket. */
441 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
442 if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems)
444 if (ubz == &bucket_zones[0])
452 bucket_select(int size)
454 struct uma_bucket_zone *ubz;
456 ubz = &bucket_zones[0];
457 if (size > ubz->ubz_maxsize)
458 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
460 for (; ubz->ubz_entries != 0; ubz++)
461 if (ubz->ubz_maxsize < size)
464 return (ubz->ubz_entries);
468 bucket_alloc(uma_zone_t zone, void *udata, int flags)
470 struct uma_bucket_zone *ubz;
474 * Don't allocate buckets early in boot.
476 if (__predict_false(booted < BOOT_KVA))
480 * To limit bucket recursion we store the original zone flags
481 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
482 * NOVM flag to persist even through deep recursions. We also
483 * store ZFLAG_BUCKET once we have recursed attempting to allocate
484 * a bucket for a bucket zone so we do not allow infinite bucket
485 * recursion. This cookie will even persist to frees of unused
486 * buckets via the allocation path or bucket allocations in the
489 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
490 udata = (void *)(uintptr_t)zone->uz_flags;
492 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
494 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
496 if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
498 ubz = bucket_zone_lookup(zone->uz_bucket_size);
499 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
501 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
504 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
507 bucket->ub_entries = ubz->ubz_entries;
508 bucket->ub_seq = SMR_SEQ_INVALID;
509 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
510 zone->uz_name, zone, bucket);
517 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
519 struct uma_bucket_zone *ubz;
521 KASSERT(bucket->ub_cnt == 0,
522 ("bucket_free: Freeing a non free bucket."));
523 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
524 ("bucket_free: Freeing an SMR bucket."));
525 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
526 udata = (void *)(uintptr_t)zone->uz_flags;
527 ubz = bucket_zone_lookup(bucket->ub_entries);
528 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
532 bucket_zone_drain(void)
534 struct uma_bucket_zone *ubz;
536 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
537 uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN);
541 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
542 * zone's caches. If a bucket is found the zone is not locked on return.
545 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom)
551 ZONE_LOCK_ASSERT(zone);
553 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
556 /* SMR Buckets can not be re-used until readers expire. */
557 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
558 bucket->ub_seq != SMR_SEQ_INVALID) {
559 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
561 bucket->ub_seq = SMR_SEQ_INVALID;
562 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
564 MPASS(zdom->uzd_nitems >= bucket->ub_cnt);
565 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
566 zdom->uzd_nitems -= bucket->ub_cnt;
567 if (zdom->uzd_imin > zdom->uzd_nitems)
568 zdom->uzd_imin = zdom->uzd_nitems;
569 zone->uz_bkt_count -= bucket->ub_cnt;
572 for (i = 0; i < bucket->ub_cnt; i++)
573 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
580 * Insert a full bucket into the specified cache. The "ws" parameter indicates
581 * whether the bucket's contents should be counted as part of the zone's working
585 zone_put_bucket(uma_zone_t zone, uma_zone_domain_t zdom, uma_bucket_t bucket,
589 ZONE_LOCK_ASSERT(zone);
590 KASSERT(!ws || zone->uz_bkt_count < zone->uz_bkt_max,
591 ("%s: zone %p overflow", __func__, zone));
593 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
594 zdom->uzd_nitems += bucket->ub_cnt;
595 if (ws && zdom->uzd_imax < zdom->uzd_nitems)
596 zdom->uzd_imax = zdom->uzd_nitems;
597 zone->uz_bkt_count += bucket->ub_cnt;
600 /* Pops an item out of a per-cpu cache bucket. */
602 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
606 CRITICAL_ASSERT(curthread);
609 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
611 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
612 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
619 /* Pushes an item into a per-cpu cache bucket. */
621 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
624 CRITICAL_ASSERT(curthread);
625 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
626 ("uma_zfree: Freeing to non free bucket index."));
628 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
634 * Unload a UMA bucket from a per-cpu cache.
636 static inline uma_bucket_t
637 cache_bucket_unload(uma_cache_bucket_t bucket)
641 b = bucket->ucb_bucket;
643 MPASS(b->ub_entries == bucket->ucb_entries);
644 b->ub_cnt = bucket->ucb_cnt;
645 bucket->ucb_bucket = NULL;
646 bucket->ucb_entries = bucket->ucb_cnt = 0;
652 static inline uma_bucket_t
653 cache_bucket_unload_alloc(uma_cache_t cache)
656 return (cache_bucket_unload(&cache->uc_allocbucket));
659 static inline uma_bucket_t
660 cache_bucket_unload_free(uma_cache_t cache)
663 return (cache_bucket_unload(&cache->uc_freebucket));
666 static inline uma_bucket_t
667 cache_bucket_unload_cross(uma_cache_t cache)
670 return (cache_bucket_unload(&cache->uc_crossbucket));
674 * Load a bucket into a per-cpu cache bucket.
677 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
680 CRITICAL_ASSERT(curthread);
681 MPASS(bucket->ucb_bucket == NULL);
682 MPASS(b->ub_seq == SMR_SEQ_INVALID);
684 bucket->ucb_bucket = b;
685 bucket->ucb_cnt = b->ub_cnt;
686 bucket->ucb_entries = b->ub_entries;
690 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
693 cache_bucket_load(&cache->uc_allocbucket, b);
697 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
700 cache_bucket_load(&cache->uc_freebucket, b);
705 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
708 cache_bucket_load(&cache->uc_crossbucket, b);
713 * Copy and preserve ucb_spare.
716 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
719 b1->ucb_bucket = b2->ucb_bucket;
720 b1->ucb_entries = b2->ucb_entries;
721 b1->ucb_cnt = b2->ucb_cnt;
725 * Swap two cache buckets.
728 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
730 struct uma_cache_bucket b3;
732 CRITICAL_ASSERT(curthread);
734 cache_bucket_copy(&b3, b1);
735 cache_bucket_copy(b1, b2);
736 cache_bucket_copy(b2, &b3);
740 zone_log_warning(uma_zone_t zone)
742 static const struct timeval warninterval = { 300, 0 };
744 if (!zone_warnings || zone->uz_warning == NULL)
747 if (ratecheck(&zone->uz_ratecheck, &warninterval))
748 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
752 zone_maxaction(uma_zone_t zone)
755 if (zone->uz_maxaction.ta_func != NULL)
756 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
760 * Routine called by timeout which is used to fire off some time interval
761 * based calculations. (stats, hash size, etc.)
770 uma_timeout(void *unused)
773 zone_foreach(zone_timeout, NULL);
775 /* Reschedule this event */
776 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
780 * Update the working set size estimate for the zone's bucket cache.
781 * The constants chosen here are somewhat arbitrary. With an update period of
782 * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
786 zone_domain_update_wss(uma_zone_domain_t zdom)
790 MPASS(zdom->uzd_imax >= zdom->uzd_imin);
791 wss = zdom->uzd_imax - zdom->uzd_imin;
792 zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
793 zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
797 * Routine to perform timeout driven calculations. This expands the
798 * hashes and does per cpu statistics aggregation.
803 zone_timeout(uma_zone_t zone, void *unused)
808 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
814 * Hash zones are non-numa by definition so the first domain
815 * is the only one present.
818 pages = keg->uk_domain[0].ud_pages;
821 * Expand the keg hash table.
823 * This is done if the number of slabs is larger than the hash size.
824 * What I'm trying to do here is completely reduce collisions. This
825 * may be a little aggressive. Should I allow for two collisions max?
827 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
828 struct uma_hash newhash;
829 struct uma_hash oldhash;
833 * This is so involved because allocating and freeing
834 * while the keg lock is held will lead to deadlock.
835 * I have to do everything in stages and check for
839 ret = hash_alloc(&newhash, 1 << fls(slabs));
842 if (hash_expand(&keg->uk_hash, &newhash)) {
843 oldhash = keg->uk_hash;
844 keg->uk_hash = newhash;
857 for (int i = 0; i < vm_ndomains; i++)
858 zone_domain_update_wss(&zone->uz_domain[i]);
863 * Allocate and zero fill the next sized hash table from the appropriate
867 * hash A new hash structure with the old hash size in uh_hashsize
870 * 1 on success and 0 on failure.
873 hash_alloc(struct uma_hash *hash, u_int size)
877 KASSERT(powerof2(size), ("hash size must be power of 2"));
878 if (size > UMA_HASH_SIZE_INIT) {
879 hash->uh_hashsize = size;
880 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
881 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
883 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
884 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
885 UMA_ANYDOMAIN, M_WAITOK);
886 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
888 if (hash->uh_slab_hash) {
889 bzero(hash->uh_slab_hash, alloc);
890 hash->uh_hashmask = hash->uh_hashsize - 1;
898 * Expands the hash table for HASH zones. This is done from zone_timeout
899 * to reduce collisions. This must not be done in the regular allocation
900 * path, otherwise, we can recurse on the vm while allocating pages.
903 * oldhash The hash you want to expand
904 * newhash The hash structure for the new table
912 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
914 uma_hash_slab_t slab;
918 if (!newhash->uh_slab_hash)
921 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
925 * I need to investigate hash algorithms for resizing without a
929 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
930 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
931 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
932 LIST_REMOVE(slab, uhs_hlink);
933 hval = UMA_HASH(newhash, slab->uhs_data);
934 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
942 * Free the hash bucket to the appropriate backing store.
945 * slab_hash The hash bucket we're freeing
946 * hashsize The number of entries in that hash bucket
952 hash_free(struct uma_hash *hash)
954 if (hash->uh_slab_hash == NULL)
956 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
957 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
959 free(hash->uh_slab_hash, M_UMAHASH);
963 * Frees all outstanding items in a bucket
966 * zone The zone to free to, must be unlocked.
967 * bucket The free/alloc bucket with items.
974 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
978 if (bucket == NULL || bucket->ub_cnt == 0)
981 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
982 bucket->ub_seq != SMR_SEQ_INVALID) {
983 smr_wait(zone->uz_smr, bucket->ub_seq);
984 bucket->ub_seq = SMR_SEQ_INVALID;
985 for (i = 0; i < bucket->ub_cnt; i++)
986 item_dtor(zone, bucket->ub_bucket[i],
987 zone->uz_size, NULL, SKIP_NONE);
990 for (i = 0; i < bucket->ub_cnt; i++)
991 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
992 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
993 if (zone->uz_max_items > 0)
994 zone_free_limit(zone, bucket->ub_cnt);
996 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1002 * Drains the per cpu caches for a zone.
1004 * NOTE: This may only be called while the zone is being torn down, and not
1005 * during normal operation. This is necessary in order that we do not have
1006 * to migrate CPUs to drain the per-CPU caches.
1009 * zone The zone to drain, must be unlocked.
1015 cache_drain(uma_zone_t zone)
1018 uma_bucket_t bucket;
1023 * XXX: It is safe to not lock the per-CPU caches, because we're
1024 * tearing down the zone anyway. I.e., there will be no further use
1025 * of the caches at this point.
1027 * XXX: It would good to be able to assert that the zone is being
1028 * torn down to prevent improper use of cache_drain().
1030 seq = SMR_SEQ_INVALID;
1031 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1032 seq = smr_current(zone->uz_smr);
1034 cache = &zone->uz_cpu[cpu];
1035 bucket = cache_bucket_unload_alloc(cache);
1036 if (bucket != NULL) {
1037 bucket_drain(zone, bucket);
1038 bucket_free(zone, bucket, NULL);
1040 bucket = cache_bucket_unload_free(cache);
1041 if (bucket != NULL) {
1042 bucket->ub_seq = seq;
1043 bucket_drain(zone, bucket);
1044 bucket_free(zone, bucket, NULL);
1046 bucket = cache_bucket_unload_cross(cache);
1047 if (bucket != NULL) {
1048 bucket->ub_seq = seq;
1049 bucket_drain(zone, bucket);
1050 bucket_free(zone, bucket, NULL);
1053 bucket_cache_reclaim(zone, true);
1057 cache_shrink(uma_zone_t zone, void *unused)
1060 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1064 zone->uz_bucket_size =
1065 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1070 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1073 uma_bucket_t b1, b2, b3;
1076 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1079 b1 = b2 = b3 = NULL;
1081 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
1082 domain = PCPU_GET(domain);
1085 cache = &zone->uz_cpu[curcpu];
1086 b1 = cache_bucket_unload_alloc(cache);
1089 * Don't flush SMR zone buckets. This leaves the zone without a
1090 * bucket and forces every free to synchronize().
1092 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1093 b2 = cache_bucket_unload_free(cache);
1094 b3 = cache_bucket_unload_cross(cache);
1099 if (b1 != NULL && b1->ub_cnt != 0) {
1100 zone_put_bucket(zone, &zone->uz_domain[domain], b1, false);
1103 if (b2 != NULL && b2->ub_cnt != 0) {
1104 zone_put_bucket(zone, &zone->uz_domain[domain], b2, false);
1110 bucket_free(zone, b1, NULL);
1112 bucket_free(zone, b2, NULL);
1114 bucket_drain(zone, b3);
1115 bucket_free(zone, b3, NULL);
1120 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1121 * This is an expensive call because it needs to bind to all CPUs
1122 * one by one and enter a critical section on each of them in order
1123 * to safely access their cache buckets.
1124 * Zone lock must not be held on call this function.
1127 pcpu_cache_drain_safe(uma_zone_t zone)
1132 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1135 cache_shrink(zone, NULL);
1137 zone_foreach(cache_shrink, NULL);
1140 thread_lock(curthread);
1141 sched_bind(curthread, cpu);
1142 thread_unlock(curthread);
1145 cache_drain_safe_cpu(zone, NULL);
1147 zone_foreach(cache_drain_safe_cpu, NULL);
1149 thread_lock(curthread);
1150 sched_unbind(curthread);
1151 thread_unlock(curthread);
1155 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1156 * requested a drain, otherwise the per-domain caches are trimmed to either
1157 * estimated working set size.
1160 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1162 uma_zone_domain_t zdom;
1163 uma_bucket_t bucket;
1164 long target, tofree;
1167 for (i = 0; i < vm_ndomains; i++) {
1169 * The cross bucket is partially filled and not part of
1170 * the item count. Reclaim it individually here.
1172 zdom = &zone->uz_domain[i];
1173 ZONE_CROSS_LOCK(zone);
1174 bucket = zdom->uzd_cross;
1175 zdom->uzd_cross = NULL;
1176 ZONE_CROSS_UNLOCK(zone);
1177 if (bucket != NULL) {
1178 bucket_drain(zone, bucket);
1179 bucket_free(zone, bucket, NULL);
1183 * Shrink the zone bucket size to ensure that the per-CPU caches
1184 * don't grow too large.
1187 if (i == 0 && zone->uz_bucket_size > zone->uz_bucket_size_min)
1188 zone->uz_bucket_size--;
1191 * If we were asked to drain the zone, we are done only once
1192 * this bucket cache is empty. Otherwise, we reclaim items in
1193 * excess of the zone's estimated working set size. If the
1194 * difference nitems - imin is larger than the WSS estimate,
1195 * then the estimate will grow at the end of this interval and
1196 * we ignore the historical average.
1198 target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1200 while (zdom->uzd_nitems > target) {
1201 bucket = STAILQ_FIRST(&zdom->uzd_buckets);
1204 tofree = bucket->ub_cnt;
1205 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
1206 zdom->uzd_nitems -= tofree;
1207 zone->uz_bkt_count -= tofree;
1210 * Shift the bounds of the current WSS interval to avoid
1211 * perturbing the estimate.
1213 zdom->uzd_imax -= lmin(zdom->uzd_imax, tofree);
1214 zdom->uzd_imin -= lmin(zdom->uzd_imin, tofree);
1217 bucket_drain(zone, bucket);
1218 bucket_free(zone, bucket, NULL);
1226 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1232 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1233 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1235 mem = slab_data(slab, keg);
1236 flags = slab->us_flags;
1238 if (keg->uk_fini != NULL) {
1239 for (i--; i > -1; i--)
1242 * trash_fini implies that dtor was trash_dtor. trash_fini
1243 * would check that memory hasn't been modified since free,
1244 * which executed trash_dtor.
1245 * That's why we need to run uma_dbg_kskip() check here,
1246 * albeit we don't make skip check for other init/fini
1249 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1250 keg->uk_fini != trash_fini)
1252 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1254 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1255 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1257 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1258 uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1262 * Frees pages from a keg back to the system. This is done on demand from
1263 * the pageout daemon.
1268 keg_drain(uma_keg_t keg)
1270 struct slabhead freeslabs;
1272 uma_slab_t slab, tmp;
1275 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1278 for (i = 0; i < vm_ndomains; i++) {
1279 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1280 keg->uk_name, keg, i, dom->ud_free_items);
1281 dom = &keg->uk_domain[i];
1282 LIST_INIT(&freeslabs);
1285 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1286 LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
1287 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1289 n = dom->ud_free_slabs;
1290 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1291 dom->ud_free_slabs = 0;
1292 dom->ud_free_items -= n * keg->uk_ipers;
1293 dom->ud_pages -= n * keg->uk_ppera;
1296 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1297 keg_free_slab(keg, slab, keg->uk_ipers);
1302 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1306 * Set draining to interlock with zone_dtor() so we can release our
1307 * locks as we go. Only dtor() should do a WAITOK call since it
1308 * is the only call that knows the structure will still be available
1312 while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1313 if (waitok == M_NOWAIT)
1315 msleep(zone, &zone->uz_lock, PVM, "zonedrain", 1);
1317 zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1319 bucket_cache_reclaim(zone, drain);
1322 * The DRAINING flag protects us from being freed while
1323 * we're running. Normally the uma_rwlock would protect us but we
1324 * must be able to release and acquire the right lock for each keg.
1326 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1327 keg_drain(zone->uz_keg);
1329 zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1336 zone_drain(uma_zone_t zone, void *unused)
1339 zone_reclaim(zone, M_NOWAIT, true);
1343 zone_trim(uma_zone_t zone, void *unused)
1346 zone_reclaim(zone, M_NOWAIT, false);
1350 * Allocate a new slab for a keg and inserts it into the partial slab list.
1351 * The keg should be unlocked on entry. If the allocation succeeds it will
1352 * be locked on return.
1355 * flags Wait flags for the item initialization routine
1356 * aflags Wait flags for the slab allocation
1359 * The slab that was allocated or NULL if there is no memory and the
1360 * caller specified M_NOWAIT.
1363 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1374 KASSERT(domain >= 0 && domain < vm_ndomains,
1375 ("keg_alloc_slab: domain %d out of range", domain));
1377 allocf = keg->uk_allocf;
1380 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1381 uma_hash_slab_t hslab;
1382 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1386 slab = &hslab->uhs_slab;
1390 * This reproduces the old vm_zone behavior of zero filling pages the
1391 * first time they are added to a zone.
1393 * Malloced items are zeroed in uma_zalloc.
1396 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1401 if (keg->uk_flags & UMA_ZONE_NODUMP)
1404 /* zone is passed for legacy reasons. */
1405 size = keg->uk_ppera * PAGE_SIZE;
1406 mem = allocf(zone, size, domain, &sflags, aflags);
1408 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1409 zone_free_item(slabzone(keg->uk_ipers),
1410 slab_tohashslab(slab), NULL, SKIP_NONE);
1413 uma_total_inc(size);
1415 /* For HASH zones all pages go to the same uma_domain. */
1416 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1419 /* Point the slab into the allocated memory */
1420 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1421 slab = (uma_slab_t )(mem + keg->uk_pgoff);
1423 slab_tohashslab(slab)->uhs_data = mem;
1425 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1426 for (i = 0; i < keg->uk_ppera; i++)
1427 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1430 slab->us_freecount = keg->uk_ipers;
1431 slab->us_flags = sflags;
1432 slab->us_domain = domain;
1434 BIT_FILL(keg->uk_ipers, &slab->us_free);
1436 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1439 if (keg->uk_init != NULL) {
1440 for (i = 0; i < keg->uk_ipers; i++)
1441 if (keg->uk_init(slab_item(slab, keg, i),
1442 keg->uk_size, flags) != 0)
1444 if (i != keg->uk_ipers) {
1445 keg_free_slab(keg, slab, i);
1449 KEG_LOCK(keg, domain);
1451 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1452 slab, keg->uk_name, keg);
1454 if (keg->uk_flags & UMA_ZFLAG_HASH)
1455 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1458 * If we got a slab here it's safe to mark it partially used
1459 * and return. We assume that the caller is going to remove
1460 * at least one item.
1462 dom = &keg->uk_domain[domain];
1463 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1464 dom->ud_pages += keg->uk_ppera;
1465 dom->ud_free_items += keg->uk_ipers;
1474 * This function is intended to be used early on in place of page_alloc() so
1475 * that we may use the boot time page cache to satisfy allocations before
1479 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1488 pages = howmany(bytes, PAGE_SIZE);
1489 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1491 *pflag = UMA_SLAB_BOOT;
1492 m = vm_page_alloc_contig_domain(NULL, 0, domain,
1493 malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1494 (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1498 pa = VM_PAGE_TO_PHYS(m);
1499 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1500 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1501 defined(__riscv) || defined(__powerpc64__)
1502 if ((wait & M_NODUMP) == 0)
1506 /* Allocate KVA and indirectly advance bootmem. */
1507 mem = (void *)pmap_map(&bootmem, m->phys_addr,
1508 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1509 if ((wait & M_ZERO) != 0)
1510 bzero(mem, pages * PAGE_SIZE);
1516 startup_free(void *mem, vm_size_t bytes)
1521 va = (vm_offset_t)mem;
1522 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1523 pmap_remove(kernel_pmap, va, va + bytes);
1524 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1525 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1526 defined(__riscv) || defined(__powerpc64__)
1527 dump_drop_page(VM_PAGE_TO_PHYS(m));
1529 vm_page_unwire_noq(m);
1535 * Allocates a number of pages from the system
1538 * bytes The number of bytes requested
1539 * wait Shall we wait?
1542 * A pointer to the alloced memory or possibly
1543 * NULL if M_NOWAIT is set.
1546 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1549 void *p; /* Returned page */
1551 *pflag = UMA_SLAB_KERNEL;
1552 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1558 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1561 struct pglist alloctail;
1562 vm_offset_t addr, zkva;
1564 vm_page_t p, p_next;
1569 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1571 TAILQ_INIT(&alloctail);
1572 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1573 malloc2vm_flags(wait);
1574 *pflag = UMA_SLAB_KERNEL;
1575 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1576 if (CPU_ABSENT(cpu)) {
1577 p = vm_page_alloc(NULL, 0, flags);
1580 p = vm_page_alloc(NULL, 0, flags);
1582 pc = pcpu_find(cpu);
1583 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1586 p = vm_page_alloc_domain(NULL, 0,
1587 pc->pc_domain, flags);
1588 if (__predict_false(p == NULL))
1589 p = vm_page_alloc(NULL, 0, flags);
1592 if (__predict_false(p == NULL))
1594 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1596 if ((addr = kva_alloc(bytes)) == 0)
1599 TAILQ_FOREACH(p, &alloctail, listq) {
1600 pmap_qenter(zkva, &p, 1);
1603 return ((void*)addr);
1605 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1606 vm_page_unwire_noq(p);
1613 * Allocates a number of pages from within an object
1616 * bytes The number of bytes requested
1617 * wait Shall we wait?
1620 * A pointer to the alloced memory or possibly
1621 * NULL if M_NOWAIT is set.
1624 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1627 TAILQ_HEAD(, vm_page) alloctail;
1629 vm_offset_t retkva, zkva;
1630 vm_page_t p, p_next;
1633 TAILQ_INIT(&alloctail);
1636 npages = howmany(bytes, PAGE_SIZE);
1637 while (npages > 0) {
1638 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1639 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1640 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1644 * Since the page does not belong to an object, its
1647 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1652 * Page allocation failed, free intermediate pages and
1655 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1656 vm_page_unwire_noq(p);
1661 *flags = UMA_SLAB_PRIV;
1662 zkva = keg->uk_kva +
1663 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1665 TAILQ_FOREACH(p, &alloctail, listq) {
1666 pmap_qenter(zkva, &p, 1);
1670 return ((void *)retkva);
1674 * Allocate physically contiguous pages.
1677 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1681 *pflag = UMA_SLAB_KERNEL;
1682 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1683 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1687 * Frees a number of pages to the system
1690 * mem A pointer to the memory to be freed
1691 * size The size of the memory being freed
1692 * flags The original p->us_flags field
1698 page_free(void *mem, vm_size_t size, uint8_t flags)
1701 if ((flags & UMA_SLAB_BOOT) != 0) {
1702 startup_free(mem, size);
1706 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1707 ("UMA: page_free used with invalid flags %x", flags));
1709 kmem_free((vm_offset_t)mem, size);
1713 * Frees pcpu zone allocations
1716 * mem A pointer to the memory to be freed
1717 * size The size of the memory being freed
1718 * flags The original p->us_flags field
1724 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1726 vm_offset_t sva, curva;
1730 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1732 if ((flags & UMA_SLAB_BOOT) != 0) {
1733 startup_free(mem, size);
1737 sva = (vm_offset_t)mem;
1738 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1739 paddr = pmap_kextract(curva);
1740 m = PHYS_TO_VM_PAGE(paddr);
1741 vm_page_unwire_noq(m);
1744 pmap_qremove(sva, size >> PAGE_SHIFT);
1745 kva_free(sva, size);
1750 * Zero fill initializer
1752 * Arguments/Returns follow uma_init specifications
1755 zero_init(void *mem, int size, int flags)
1763 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1766 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1771 * Actual size of embedded struct slab (!OFFPAGE).
1774 slab_sizeof(int nitems)
1778 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1779 return (roundup(s, UMA_ALIGN_PTR + 1));
1783 * Size of memory for embedded slabs (!OFFPAGE).
1786 slab_space(int nitems)
1788 return (UMA_SLAB_SIZE - slab_sizeof(nitems));
1791 #define UMA_FIXPT_SHIFT 31
1792 #define UMA_FRAC_FIXPT(n, d) \
1793 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1794 #define UMA_FIXPT_PCT(f) \
1795 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1796 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
1797 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1800 * Compute the number of items that will fit in a slab. If hdr is true, the
1801 * item count may be limited to provide space in the slab for an inline slab
1802 * header. Otherwise, all slab space will be provided for item storage.
1805 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1810 /* The padding between items is not needed after the last item. */
1811 padpi = rsize - size;
1815 * Start with the maximum item count and remove items until
1816 * the slab header first alongside the allocatable memory.
1818 for (ipers = MIN(SLAB_MAX_SETSIZE,
1819 (slabsize + padpi - slab_sizeof(1)) / rsize);
1821 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1825 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1832 * Compute the number of items that will fit in a slab for a startup zone.
1835 slab_ipers(size_t size, int align)
1839 rsize = roundup(size, align + 1); /* Assume no CACHESPREAD */
1840 return (slab_ipers_hdr(size, rsize, UMA_SLAB_SIZE, true));
1843 struct keg_layout_result {
1851 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
1852 struct keg_layout_result *kl)
1857 kl->slabsize = slabsize;
1859 /* Handle INTERNAL as inline with an extra page. */
1860 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
1861 kl->format &= ~UMA_ZFLAG_INTERNAL;
1862 kl->slabsize += PAGE_SIZE;
1865 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
1866 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
1868 /* Account for memory used by an offpage slab header. */
1869 total = kl->slabsize;
1870 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
1871 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
1873 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
1877 * Determine the format of a uma keg. This determines where the slab header
1878 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
1881 * keg The zone we should initialize
1887 keg_layout(uma_keg_t keg)
1889 struct keg_layout_result kl = {}, kl_tmp;
1898 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1899 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
1900 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
1901 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
1902 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
1904 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
1905 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
1906 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
1909 alignsize = keg->uk_align + 1;
1912 * Calculate the size of each allocation (rsize) according to
1913 * alignment. If the requested size is smaller than we have
1914 * allocation bits for we round it up.
1916 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
1917 rsize = roundup2(rsize, alignsize);
1919 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
1921 * We want one item to start on every align boundary in a page.
1922 * To do this we will span pages. We will also extend the item
1923 * by the size of align if it is an even multiple of align.
1924 * Otherwise, it would fall on the same boundary every time.
1926 if ((rsize & alignsize) == 0)
1928 slabsize = rsize * (PAGE_SIZE / alignsize);
1929 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
1930 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
1931 slabsize = round_page(slabsize);
1934 * Start with a slab size of as many pages as it takes to
1935 * represent a single item. We will try to fit as many
1936 * additional items into the slab as possible.
1938 slabsize = round_page(keg->uk_size);
1941 /* Build a list of all of the available formats for this keg. */
1944 /* Evaluate an inline slab layout. */
1945 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
1948 /* TODO: vm_page-embedded slab. */
1951 * We can't do OFFPAGE if we're internal or if we've been
1952 * asked to not go to the VM for buckets. If we do this we
1953 * may end up going to the VM for slabs which we do not want
1954 * to do if we're UMA_ZONE_VM, which clearly forbids it.
1955 * In those cases, evaluate a pseudo-format called INTERNAL
1956 * which has an inline slab header and one extra page to
1957 * guarantee that it fits.
1959 * Otherwise, see if using an OFFPAGE slab will improve our
1962 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
1963 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
1965 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
1968 * Choose a slab size and format which satisfy the minimum efficiency.
1969 * Prefer the smallest slab size that meets the constraints.
1971 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
1972 * for small items (up to PAGE_SIZE), the iteration increment is one
1973 * page; and for large items, the increment is one item.
1975 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
1976 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
1977 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
1980 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
1981 round_page(rsize * (i - 1) + keg->uk_size);
1983 for (j = 0; j < nfmt; j++) {
1984 /* Only if we have no viable format yet. */
1985 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
1989 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
1990 if (kl_tmp.eff <= kl.eff)
1995 CTR6(KTR_UMA, "keg %s layout: format %#x "
1996 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
1997 keg->uk_name, kl.format, kl.ipers, rsize,
1998 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2000 /* Stop when we reach the minimum efficiency. */
2001 if (kl.eff >= UMA_MIN_EFF)
2005 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2006 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2007 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2011 pages = atop(kl.slabsize);
2012 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2013 pages *= mp_maxid + 1;
2015 keg->uk_rsize = rsize;
2016 keg->uk_ipers = kl.ipers;
2017 keg->uk_ppera = pages;
2018 keg->uk_flags |= kl.format;
2021 * How do we find the slab header if it is offpage or if not all item
2022 * start addresses are in the same page? We could solve the latter
2023 * case with vaddr alignment, but we don't.
2025 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2026 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2027 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2028 keg->uk_flags |= UMA_ZFLAG_HASH;
2030 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2033 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2034 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2036 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2037 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2038 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2039 keg->uk_ipers, pages));
2043 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2044 * the keg onto the global keg list.
2046 * Arguments/Returns follow uma_ctor specifications
2047 * udata Actually uma_kctor_args
2050 keg_ctor(void *mem, int size, void *udata, int flags)
2052 struct uma_kctor_args *arg = udata;
2053 uma_keg_t keg = mem;
2058 keg->uk_size = arg->size;
2059 keg->uk_init = arg->uminit;
2060 keg->uk_fini = arg->fini;
2061 keg->uk_align = arg->align;
2062 keg->uk_reserve = 0;
2063 keg->uk_flags = arg->flags;
2066 * We use a global round-robin policy by default. Zones with
2067 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2068 * case the iterator is never run.
2070 keg->uk_dr.dr_policy = DOMAINSET_RR();
2071 keg->uk_dr.dr_iter = 0;
2074 * The master zone is passed to us at keg-creation time.
2077 keg->uk_name = zone->uz_name;
2079 if (arg->flags & UMA_ZONE_ZINIT)
2080 keg->uk_init = zero_init;
2082 if (arg->flags & UMA_ZONE_MALLOC)
2083 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2086 keg->uk_flags &= ~UMA_ZONE_PCPU;
2092 * Use a first-touch NUMA policy for all kegs that pmap_extract()
2093 * will work on with the exception of critical VM structures
2094 * necessary for paging.
2096 * Zones may override the default by specifying either.
2099 if ((keg->uk_flags &
2100 (UMA_ZFLAG_HASH | UMA_ZONE_VM | UMA_ZONE_ROUNDROBIN)) == 0)
2101 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2102 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2103 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2107 * If we haven't booted yet we need allocations to go through the
2108 * startup cache until the vm is ready.
2110 #ifdef UMA_MD_SMALL_ALLOC
2111 if (keg->uk_ppera == 1)
2112 keg->uk_allocf = uma_small_alloc;
2115 if (booted < BOOT_KVA)
2116 keg->uk_allocf = startup_alloc;
2117 else if (keg->uk_flags & UMA_ZONE_PCPU)
2118 keg->uk_allocf = pcpu_page_alloc;
2119 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2120 keg->uk_allocf = contig_alloc;
2122 keg->uk_allocf = page_alloc;
2123 #ifdef UMA_MD_SMALL_ALLOC
2124 if (keg->uk_ppera == 1)
2125 keg->uk_freef = uma_small_free;
2128 if (keg->uk_flags & UMA_ZONE_PCPU)
2129 keg->uk_freef = pcpu_page_free;
2131 keg->uk_freef = page_free;
2134 * Initialize keg's locks.
2136 for (i = 0; i < vm_ndomains; i++)
2137 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2140 * If we're putting the slab header in the actual page we need to
2141 * figure out where in each page it goes. See slab_sizeof
2144 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2147 shsize = slab_sizeof(keg->uk_ipers);
2148 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2150 * The only way the following is possible is if with our
2151 * UMA_ALIGN_PTR adjustments we are now bigger than
2152 * UMA_SLAB_SIZE. I haven't checked whether this is
2153 * mathematically possible for all cases, so we make
2156 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2157 ("zone %s ipers %d rsize %d size %d slab won't fit",
2158 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2161 if (keg->uk_flags & UMA_ZFLAG_HASH)
2162 hash_alloc(&keg->uk_hash, 0);
2164 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2166 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2168 rw_wlock(&uma_rwlock);
2169 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2170 rw_wunlock(&uma_rwlock);
2175 zone_kva_available(uma_zone_t zone, void *unused)
2179 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2183 if (keg->uk_allocf == startup_alloc) {
2184 /* Switch to the real allocator. */
2185 if (keg->uk_flags & UMA_ZONE_PCPU)
2186 keg->uk_allocf = pcpu_page_alloc;
2187 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2189 keg->uk_allocf = contig_alloc;
2191 keg->uk_allocf = page_alloc;
2196 zone_alloc_counters(uma_zone_t zone, void *unused)
2199 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2200 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2201 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2205 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2207 uma_zone_domain_t zdom;
2210 struct sysctl_oid *oid, *domainoid;
2211 int domains, i, cnt;
2212 static const char *nokeg = "cache zone";
2216 * Make a sysctl safe copy of the zone name by removing
2217 * any special characters and handling dups by appending
2220 if (zone->uz_namecnt != 0) {
2221 /* Count the number of decimal digits and '_' separator. */
2222 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2224 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2226 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2229 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2230 for (c = zone->uz_ctlname; *c != '\0'; c++)
2231 if (strchr("./\\ -", *c) != NULL)
2235 * Basic parameters at the root.
2237 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2238 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD, NULL, "");
2240 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2241 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2242 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2243 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2244 zone, 0, sysctl_handle_uma_zone_flags, "A",
2245 "Allocator configuration flags");
2246 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2247 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2248 "Desired per-cpu cache size");
2249 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2250 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2251 "Maximum allowed per-cpu cache size");
2256 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2257 domains = vm_ndomains;
2260 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2261 "keg", CTLFLAG_RD, NULL, "");
2263 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2264 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2265 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2266 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2267 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2268 "Real object size with alignment");
2269 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2270 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2271 "pages per-slab allocation");
2272 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2273 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2274 "items available per-slab");
2275 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2276 "align", CTLFLAG_RD, &keg->uk_align, 0,
2277 "item alignment mask");
2278 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2279 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2280 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2281 "Slab utilization (100 - internal fragmentation %)");
2282 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2283 OID_AUTO, "domain", CTLFLAG_RD, NULL, "");
2284 for (i = 0; i < domains; i++) {
2285 dom = &keg->uk_domain[i];
2286 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2287 OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD,
2289 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2290 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2291 "Total pages currently allocated from VM");
2292 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2293 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2294 "items free in the slab layer");
2297 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2298 "name", CTLFLAG_RD, nokeg, "Keg name");
2301 * Information about zone limits.
2303 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2304 "limit", CTLFLAG_RD, NULL, "");
2305 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2306 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2307 zone, 0, sysctl_handle_uma_zone_items, "QU",
2308 "current number of allocated items if limit is set");
2309 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2310 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2311 "Maximum number of cached items");
2312 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2313 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2314 "Number of threads sleeping at limit");
2315 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2316 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2317 "Total zone limit sleeps");
2318 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2319 "bucket_max", CTLFLAG_RD, &zone->uz_bkt_max, 0,
2320 "Maximum number of items in the bucket cache");
2321 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2322 "bucket_cnt", CTLFLAG_RD, &zone->uz_bkt_count, 0,
2323 "Number of items in the bucket cache");
2326 * Per-domain zone information.
2328 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2329 OID_AUTO, "domain", CTLFLAG_RD, NULL, "");
2330 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2332 for (i = 0; i < domains; i++) {
2333 zdom = &zone->uz_domain[i];
2334 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2335 OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD, NULL, "");
2336 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2337 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2338 "number of items in this domain");
2339 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2340 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2341 "maximum item count in this period");
2342 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2343 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2344 "minimum item count in this period");
2345 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2346 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2347 "Working set size");
2351 * General statistics.
2353 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2354 "stats", CTLFLAG_RD, NULL, "");
2355 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2356 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2357 zone, 1, sysctl_handle_uma_zone_cur, "I",
2358 "Current number of allocated items");
2359 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2360 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2361 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2362 "Total allocation calls");
2363 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2364 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2365 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2366 "Total free calls");
2367 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2368 "fails", CTLFLAG_RD, &zone->uz_fails,
2369 "Number of allocation failures");
2370 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2371 "xdomain", CTLFLAG_RD, &zone->uz_xdomain, 0,
2372 "Free calls from the wrong domain");
2375 struct uma_zone_count {
2381 zone_count(uma_zone_t zone, void *arg)
2383 struct uma_zone_count *cnt;
2387 * Some zones are rapidly created with identical names and
2388 * destroyed out of order. This can lead to gaps in the count.
2389 * Use one greater than the maximum observed for this name.
2391 if (strcmp(zone->uz_name, cnt->name) == 0)
2392 cnt->count = MAX(cnt->count,
2393 zone->uz_namecnt + 1);
2397 zone_update_caches(uma_zone_t zone)
2401 for (i = 0; i <= mp_maxid; i++) {
2402 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2403 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2408 * Zone header ctor. This initializes all fields, locks, etc.
2410 * Arguments/Returns follow uma_ctor specifications
2411 * udata Actually uma_zctor_args
2414 zone_ctor(void *mem, int size, void *udata, int flags)
2416 struct uma_zone_count cnt;
2417 struct uma_zctor_args *arg = udata;
2418 uma_zone_t zone = mem;
2424 zone->uz_name = arg->name;
2425 zone->uz_ctor = arg->ctor;
2426 zone->uz_dtor = arg->dtor;
2427 zone->uz_init = NULL;
2428 zone->uz_fini = NULL;
2429 zone->uz_sleeps = 0;
2430 zone->uz_xdomain = 0;
2431 zone->uz_bucket_size = 0;
2432 zone->uz_bucket_size_min = 0;
2433 zone->uz_bucket_size_max = BUCKET_MAX;
2434 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2435 zone->uz_warning = NULL;
2436 /* The domain structures follow the cpu structures. */
2438 (struct uma_zone_domain *)&zone->uz_cpu[mp_maxid + 1];
2439 zone->uz_bkt_max = ULONG_MAX;
2440 timevalclear(&zone->uz_ratecheck);
2442 /* Count the number of duplicate names. */
2443 cnt.name = arg->name;
2445 zone_foreach(zone_count, &cnt);
2446 zone->uz_namecnt = cnt.count;
2447 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
2448 ZONE_CROSS_LOCK_INIT(zone);
2450 for (i = 0; i < vm_ndomains; i++)
2451 STAILQ_INIT(&zone->uz_domain[i].uzd_buckets);
2454 if (arg->uminit == trash_init && arg->fini == trash_fini)
2455 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2459 * This is a pure cache zone, no kegs.
2462 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2463 ("zone_ctor: Import specified for non-cache zone."));
2464 zone->uz_flags = arg->flags;
2465 zone->uz_size = arg->size;
2466 zone->uz_import = arg->import;
2467 zone->uz_release = arg->release;
2468 zone->uz_arg = arg->arg;
2469 rw_wlock(&uma_rwlock);
2470 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2471 rw_wunlock(&uma_rwlock);
2476 * Use the regular zone/keg/slab allocator.
2478 zone->uz_import = zone_import;
2479 zone->uz_release = zone_release;
2480 zone->uz_arg = zone;
2483 if (arg->flags & UMA_ZONE_SECONDARY) {
2484 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2485 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2486 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2487 zone->uz_init = arg->uminit;
2488 zone->uz_fini = arg->fini;
2489 zone->uz_flags |= UMA_ZONE_SECONDARY;
2490 rw_wlock(&uma_rwlock);
2492 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2493 if (LIST_NEXT(z, uz_link) == NULL) {
2494 LIST_INSERT_AFTER(z, zone, uz_link);
2499 rw_wunlock(&uma_rwlock);
2500 } else if (keg == NULL) {
2501 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2502 arg->align, arg->flags)) == NULL)
2505 struct uma_kctor_args karg;
2508 /* We should only be here from uma_startup() */
2509 karg.size = arg->size;
2510 karg.uminit = arg->uminit;
2511 karg.fini = arg->fini;
2512 karg.align = arg->align;
2513 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2515 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2521 /* Inherit properties from the keg. */
2523 zone->uz_size = keg->uk_size;
2524 zone->uz_flags |= (keg->uk_flags &
2525 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2528 if (__predict_true(booted >= BOOT_RUNNING)) {
2529 zone_alloc_counters(zone, NULL);
2530 zone_alloc_sysctl(zone, NULL);
2532 zone->uz_allocs = EARLY_COUNTER;
2533 zone->uz_frees = EARLY_COUNTER;
2534 zone->uz_fails = EARLY_COUNTER;
2537 /* Caller requests a private SMR context. */
2538 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2539 zone->uz_smr = smr_create(zone->uz_name);
2541 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2542 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2543 ("Invalid zone flag combination"));
2544 if (arg->flags & UMA_ZFLAG_INTERNAL)
2545 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2546 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2547 zone->uz_bucket_size = BUCKET_MAX;
2548 else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2549 zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2550 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2551 zone->uz_bucket_size = 0;
2553 zone->uz_bucket_size = bucket_select(zone->uz_size);
2554 zone->uz_bucket_size_min = zone->uz_bucket_size;
2555 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2556 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2557 zone_update_caches(zone);
2563 * Keg header dtor. This frees all data, destroys locks, frees the hash
2564 * table and removes the keg from the global list.
2566 * Arguments/Returns follow uma_dtor specifications
2570 keg_dtor(void *arg, int size, void *udata)
2573 uint32_t free, pages;
2576 keg = (uma_keg_t)arg;
2578 for (i = 0; i < vm_ndomains; i++) {
2579 free += keg->uk_domain[i].ud_free_items;
2580 pages += keg->uk_domain[i].ud_pages;
2581 KEG_LOCK_FINI(keg, i);
2584 printf("Freed UMA keg (%s) was not empty (%u items). "
2585 " Lost %u pages of memory.\n",
2586 keg->uk_name ? keg->uk_name : "",
2587 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2589 hash_free(&keg->uk_hash);
2595 * Arguments/Returns follow uma_dtor specifications
2599 zone_dtor(void *arg, int size, void *udata)
2604 zone = (uma_zone_t)arg;
2606 sysctl_remove_oid(zone->uz_oid, 1, 1);
2608 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2611 rw_wlock(&uma_rwlock);
2612 LIST_REMOVE(zone, uz_link);
2613 rw_wunlock(&uma_rwlock);
2615 * XXX there are some races here where
2616 * the zone can be drained but zone lock
2617 * released and then refilled before we
2618 * remove it... we dont care for now
2620 zone_reclaim(zone, M_WAITOK, true);
2622 * We only destroy kegs from non secondary/non cache zones.
2624 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2626 rw_wlock(&uma_rwlock);
2627 LIST_REMOVE(keg, uk_link);
2628 rw_wunlock(&uma_rwlock);
2629 zone_free_item(kegs, keg, NULL, SKIP_NONE);
2631 counter_u64_free(zone->uz_allocs);
2632 counter_u64_free(zone->uz_frees);
2633 counter_u64_free(zone->uz_fails);
2634 free(zone->uz_ctlname, M_UMA);
2635 ZONE_LOCK_FINI(zone);
2636 ZONE_CROSS_LOCK_FINI(zone);
2640 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2645 LIST_FOREACH(keg, &uma_kegs, uk_link) {
2646 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2649 LIST_FOREACH(zone, &uma_cachezones, uz_link)
2654 * Traverses every zone in the system and calls a callback
2657 * zfunc A pointer to a function which accepts a zone
2664 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2667 rw_rlock(&uma_rwlock);
2668 zone_foreach_unlocked(zfunc, arg);
2669 rw_runlock(&uma_rwlock);
2673 * Initialize the kernel memory allocator. This is done after pages can be
2674 * allocated but before general KVA is available.
2677 uma_startup1(vm_offset_t virtual_avail)
2679 struct uma_zctor_args args;
2680 size_t ksize, zsize, size;
2681 uma_keg_t masterkeg;
2685 bootstart = bootmem = virtual_avail;
2687 rw_init(&uma_rwlock, "UMA lock");
2688 sx_init(&uma_reclaim_lock, "umareclaim");
2690 ksize = sizeof(struct uma_keg) +
2691 (sizeof(struct uma_domain) * vm_ndomains);
2692 ksize = roundup(ksize, UMA_SUPER_ALIGN);
2693 zsize = sizeof(struct uma_zone) +
2694 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2695 (sizeof(struct uma_zone_domain) * vm_ndomains);
2696 zsize = roundup(zsize, UMA_SUPER_ALIGN);
2698 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2699 size = (zsize * 2) + ksize;
2700 m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO);
2701 zones = (uma_zone_t)m;
2703 kegs = (uma_zone_t)m;
2705 masterkeg = (uma_keg_t)m;
2707 /* "manually" create the initial zone */
2708 memset(&args, 0, sizeof(args));
2709 args.name = "UMA Kegs";
2711 args.ctor = keg_ctor;
2712 args.dtor = keg_dtor;
2713 args.uminit = zero_init;
2715 args.keg = masterkeg;
2716 args.align = UMA_SUPER_ALIGN - 1;
2717 args.flags = UMA_ZFLAG_INTERNAL;
2718 zone_ctor(kegs, zsize, &args, M_WAITOK);
2720 args.name = "UMA Zones";
2722 args.ctor = zone_ctor;
2723 args.dtor = zone_dtor;
2724 args.uminit = zero_init;
2727 args.align = UMA_SUPER_ALIGN - 1;
2728 args.flags = UMA_ZFLAG_INTERNAL;
2729 zone_ctor(zones, zsize, &args, M_WAITOK);
2731 /* Now make zones for slab headers */
2732 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2733 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2734 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2735 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2737 hashzone = uma_zcreate("UMA Hash",
2738 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2739 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2745 #ifndef UMA_MD_SMALL_ALLOC
2746 extern void vm_radix_reserve_kva(void);
2750 * Advertise the availability of normal kva allocations and switch to
2751 * the default back-end allocator. Marks the KVA we consumed on startup
2752 * as used in the map.
2758 if (bootstart != bootmem) {
2759 vm_map_lock(kernel_map);
2760 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2761 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2762 vm_map_unlock(kernel_map);
2765 #ifndef UMA_MD_SMALL_ALLOC
2766 /* Set up radix zone to use noobj_alloc. */
2767 vm_radix_reserve_kva();
2771 zone_foreach_unlocked(zone_kva_available, NULL);
2776 * Finish our initialization steps.
2783 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2784 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2785 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2787 zone_foreach_unlocked(zone_alloc_counters, NULL);
2788 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2789 callout_init(&uma_callout, 1);
2790 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2791 booted = BOOT_RUNNING;
2793 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2794 EVENTHANDLER_PRI_FIRST);
2801 booted = BOOT_SHUTDOWN;
2805 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2806 int align, uint32_t flags)
2808 struct uma_kctor_args args;
2811 args.uminit = uminit;
2813 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2816 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2819 /* Public functions */
2822 uma_set_align(int align)
2825 if (align != UMA_ALIGN_CACHE)
2826 uma_align_cache = align;
2831 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2832 uma_init uminit, uma_fini fini, int align, uint32_t flags)
2835 struct uma_zctor_args args;
2838 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2841 /* This stuff is essential for the zone ctor */
2842 memset(&args, 0, sizeof(args));
2847 args.uminit = uminit;
2851 * Inject procedures which check for memory use after free if we are
2852 * allowed to scramble the memory while it is not allocated. This
2853 * requires that: UMA is actually able to access the memory, no init
2854 * or fini procedures, no dependency on the initial value of the
2855 * memory, and no (legitimate) use of the memory after free. Note,
2856 * the ctor and dtor do not need to be empty.
2858 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
2859 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
2860 args.uminit = trash_init;
2861 args.fini = trash_fini;
2868 sx_slock(&uma_reclaim_lock);
2869 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
2870 sx_sunlock(&uma_reclaim_lock);
2877 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
2878 uma_init zinit, uma_fini zfini, uma_zone_t master)
2880 struct uma_zctor_args args;
2884 keg = master->uz_keg;
2885 memset(&args, 0, sizeof(args));
2887 args.size = keg->uk_size;
2890 args.uminit = zinit;
2892 args.align = keg->uk_align;
2893 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
2896 sx_slock(&uma_reclaim_lock);
2897 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
2898 sx_sunlock(&uma_reclaim_lock);
2905 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
2906 uma_init zinit, uma_fini zfini, uma_import zimport,
2907 uma_release zrelease, void *arg, int flags)
2909 struct uma_zctor_args args;
2911 memset(&args, 0, sizeof(args));
2916 args.uminit = zinit;
2918 args.import = zimport;
2919 args.release = zrelease;
2922 args.flags = flags | UMA_ZFLAG_CACHE;
2924 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
2929 uma_zdestroy(uma_zone_t zone)
2933 * Large slabs are expensive to reclaim, so don't bother doing
2934 * unnecessary work if we're shutting down.
2936 if (booted == BOOT_SHUTDOWN &&
2937 zone->uz_fini == NULL && zone->uz_release == zone_release)
2939 sx_slock(&uma_reclaim_lock);
2940 zone_free_item(zones, zone, NULL, SKIP_NONE);
2941 sx_sunlock(&uma_reclaim_lock);
2945 uma_zwait(uma_zone_t zone)
2948 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2949 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
2950 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
2951 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
2953 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
2957 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
2959 void *item, *pcpu_item;
2963 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
2965 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
2968 pcpu_item = zpcpu_base_to_offset(item);
2969 if (flags & M_ZERO) {
2971 for (i = 0; i <= mp_maxid; i++)
2972 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
2974 bzero(item, zone->uz_size);
2981 * A stub while both regular and pcpu cases are identical.
2984 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
2989 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
2991 item = zpcpu_offset_to_base(pcpu_item);
2992 uma_zfree_arg(zone, item, udata);
2995 static inline void *
2996 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3002 skipdbg = uma_dbg_zskip(zone, item);
3003 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3004 zone->uz_ctor != trash_ctor)
3005 trash_ctor(item, size, udata, flags);
3007 /* Check flags before loading ctor pointer. */
3008 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3009 __predict_false(zone->uz_ctor != NULL) &&
3010 zone->uz_ctor(item, size, udata, flags) != 0) {
3011 counter_u64_add(zone->uz_fails, 1);
3012 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3017 uma_dbg_alloc(zone, NULL, item);
3019 if (__predict_false(flags & M_ZERO))
3020 return (memset(item, 0, size));
3026 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3027 enum zfreeskip skip)
3032 skipdbg = uma_dbg_zskip(zone, item);
3033 if (skip == SKIP_NONE && !skipdbg) {
3034 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3035 uma_dbg_free(zone, udata, item);
3037 uma_dbg_free(zone, NULL, item);
3040 if (__predict_true(skip < SKIP_DTOR)) {
3041 if (zone->uz_dtor != NULL)
3042 zone->uz_dtor(item, size, udata);
3044 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3045 zone->uz_dtor != trash_dtor)
3046 trash_dtor(item, size, udata);
3051 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3052 #define UMA_ZALLOC_DEBUG
3054 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3060 if (flags & M_WAITOK) {
3061 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3062 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3067 KASSERT((flags & M_EXEC) == 0,
3068 ("uma_zalloc_debug: called with M_EXEC"));
3069 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3070 ("uma_zalloc_debug: called within spinlock or critical section"));
3071 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3072 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3075 #ifdef DEBUG_MEMGUARD
3076 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3078 item = memguard_alloc(zone->uz_size, flags);
3080 error = EJUSTRETURN;
3081 if (zone->uz_init != NULL &&
3082 zone->uz_init(item, zone->uz_size, flags) != 0) {
3086 if (zone->uz_ctor != NULL &&
3087 zone->uz_ctor(item, zone->uz_size, udata,
3089 counter_u64_add(zone->uz_fails, 1);
3090 zone->uz_fini(item, zone->uz_size);
3097 /* This is unfortunate but should not be fatal. */
3104 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3106 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3107 ("uma_zfree_debug: called with spinlock or critical section held"));
3109 #ifdef DEBUG_MEMGUARD
3110 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3111 if (zone->uz_dtor != NULL)
3112 zone->uz_dtor(item, zone->uz_size, udata);
3113 if (zone->uz_fini != NULL)
3114 zone->uz_fini(item, zone->uz_size);
3115 memguard_free(item);
3116 return (EJUSTRETURN);
3123 static inline void *
3124 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3125 void *udata, int flags)
3130 item = cache_bucket_pop(cache, bucket);
3131 size = cache_uz_size(cache);
3132 uz_flags = cache_uz_flags(cache);
3134 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3137 static __noinline void *
3138 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3140 uma_cache_bucket_t bucket;
3143 while (cache_alloc(zone, cache, udata, flags)) {
3144 cache = &zone->uz_cpu[curcpu];
3145 bucket = &cache->uc_allocbucket;
3146 if (__predict_false(bucket->ucb_cnt == 0))
3148 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3153 * We can not get a bucket so try to return a single item.
3155 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3156 domain = PCPU_GET(domain);
3158 domain = UMA_ANYDOMAIN;
3159 return (zone_alloc_item(zone, udata, domain, flags));
3164 uma_zalloc_smr(uma_zone_t zone, int flags)
3166 uma_cache_bucket_t bucket;
3169 #ifdef UMA_ZALLOC_DEBUG
3172 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3173 ("uma_zalloc_arg: called with non-SMR zone.\n"));
3174 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3179 cache = &zone->uz_cpu[curcpu];
3180 bucket = &cache->uc_allocbucket;
3181 if (__predict_false(bucket->ucb_cnt == 0))
3182 return (cache_alloc_retry(zone, cache, NULL, flags));
3183 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3188 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3190 uma_cache_bucket_t bucket;
3193 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3194 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3196 /* This is the fast path allocation */
3197 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3200 #ifdef UMA_ZALLOC_DEBUG
3203 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3204 ("uma_zalloc_arg: called with SMR zone.\n"));
3205 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3210 * If possible, allocate from the per-CPU cache. There are two
3211 * requirements for safe access to the per-CPU cache: (1) the thread
3212 * accessing the cache must not be preempted or yield during access,
3213 * and (2) the thread must not migrate CPUs without switching which
3214 * cache it accesses. We rely on a critical section to prevent
3215 * preemption and migration. We release the critical section in
3216 * order to acquire the zone mutex if we are unable to allocate from
3217 * the current cache; when we re-acquire the critical section, we
3218 * must detect and handle migration if it has occurred.
3221 cache = &zone->uz_cpu[curcpu];
3222 bucket = &cache->uc_allocbucket;
3223 if (__predict_false(bucket->ucb_cnt == 0))
3224 return (cache_alloc_retry(zone, cache, udata, flags));
3225 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3229 * Replenish an alloc bucket and possibly restore an old one. Called in
3230 * a critical section. Returns in a critical section.
3232 * A false return value indicates an allocation failure.
3233 * A true return value indicates success and the caller should retry.
3235 static __noinline bool
3236 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3238 uma_zone_domain_t zdom;
3239 uma_bucket_t bucket;
3243 CRITICAL_ASSERT(curthread);
3246 * If we have run out of items in our alloc bucket see
3247 * if we can switch with the free bucket.
3249 * SMR Zones can't re-use the free bucket until the sequence has
3252 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 &&
3253 cache->uc_freebucket.ucb_cnt != 0) {
3254 cache_bucket_swap(&cache->uc_freebucket,
3255 &cache->uc_allocbucket);
3260 * Discard any empty allocation bucket while we hold no locks.
3262 bucket = cache_bucket_unload_alloc(cache);
3265 bucket_free(zone, bucket, udata);
3267 /* Short-circuit for zones without buckets and low memory. */
3268 if (zone->uz_bucket_size == 0 || bucketdisable) {
3274 * Attempt to retrieve the item from the per-CPU cache has failed, so
3275 * we must go back to the zone. This requires the zone lock, so we
3276 * must drop the critical section, then re-acquire it when we go back
3277 * to the cache. Since the critical section is released, we may be
3278 * preempted or migrate. As such, make sure not to maintain any
3279 * thread-local state specific to the cache from prior to releasing
3280 * the critical section.
3283 if (ZONE_TRYLOCK(zone) == 0) {
3284 /* Record contention to size the buckets. */
3289 /* See if we lost the race to fill the cache. */
3291 cache = &zone->uz_cpu[curcpu];
3292 if (cache->uc_allocbucket.ucb_bucket != NULL) {
3298 * Check the zone's cache of buckets.
3300 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) {
3301 domain = PCPU_GET(domain);
3302 zdom = &zone->uz_domain[domain];
3304 domain = UMA_ANYDOMAIN;
3305 zdom = &zone->uz_domain[0];
3308 if ((bucket = zone_fetch_bucket(zone, zdom)) != NULL) {
3309 KASSERT(bucket->ub_cnt != 0,
3310 ("uma_zalloc_arg: Returning an empty bucket."));
3311 cache_bucket_load_alloc(cache, bucket);
3314 /* We are no longer associated with this CPU. */
3318 * We bump the uz count when the cache size is insufficient to
3319 * handle the working set.
3321 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
3322 zone->uz_bucket_size++;
3326 * Fill a bucket and attempt to use it as the alloc bucket.
3328 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3329 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3330 zone->uz_name, zone, bucket);
3331 if (bucket == NULL) {
3337 * See if we lost the race or were migrated. Cache the
3338 * initialized bucket to make this less likely or claim
3339 * the memory directly.
3343 cache = &zone->uz_cpu[curcpu];
3344 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3345 ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0 ||
3346 domain == PCPU_GET(domain))) {
3347 cache_bucket_load_alloc(cache, bucket);
3348 zdom->uzd_imax += bucket->ub_cnt;
3349 } else if (zone->uz_bkt_count >= zone->uz_bkt_max) {
3352 bucket_drain(zone, bucket);
3353 bucket_free(zone, bucket, udata);
3357 zone_put_bucket(zone, zdom, bucket, false);
3363 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3366 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3367 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3369 /* This is the fast path allocation */
3370 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3371 zone->uz_name, zone, domain, flags);
3373 if (flags & M_WAITOK) {
3374 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3375 "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3377 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3378 ("uma_zalloc_domain: called with spinlock or critical section held"));
3380 return (zone_alloc_item(zone, udata, domain, flags));
3384 * Find a slab with some space. Prefer slabs that are partially used over those
3385 * that are totally full. This helps to reduce fragmentation.
3387 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3391 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3397 KASSERT(domain >= 0 && domain < vm_ndomains,
3398 ("keg_first_slab: domain %d out of range", domain));
3399 KEG_LOCK_ASSERT(keg, domain);
3404 dom = &keg->uk_domain[domain];
3405 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3407 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3408 LIST_REMOVE(slab, us_link);
3409 dom->ud_free_slabs--;
3410 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3414 domain = (domain + 1) % vm_ndomains;
3415 } while (domain != start);
3421 * Fetch an existing slab from a free or partial list. Returns with the
3422 * keg domain lock held if a slab was found or unlocked if not.
3425 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3430 /* HASH has a single free list. */
3431 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3434 KEG_LOCK(keg, domain);
3435 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3436 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3437 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3438 KEG_UNLOCK(keg, domain);
3445 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3447 struct vm_domainset_iter di;
3454 * Use the keg's policy if upper layers haven't already specified a
3455 * domain (as happens with first-touch zones).
3457 * To avoid races we run the iterator with the keg lock held, but that
3458 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3459 * clear M_WAITOK and handle low memory conditions locally.
3461 rr = rdomain == UMA_ANYDOMAIN;
3463 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3464 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3472 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3477 * M_NOVM means don't ask at all!
3482 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3485 if (!rr && (flags & M_WAITOK) == 0)
3487 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3488 if ((flags & M_WAITOK) != 0) {
3489 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
3497 * We might not have been able to get a slab but another cpu
3498 * could have while we were unlocked. Check again before we
3501 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3508 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3514 KEG_LOCK_ASSERT(keg, slab->us_domain);
3516 dom = &keg->uk_domain[slab->us_domain];
3517 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3518 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3519 item = slab_item(slab, keg, freei);
3520 slab->us_freecount--;
3521 dom->ud_free_items--;
3524 * Move this slab to the full list. It must be on the partial list, so
3525 * we do not need to update the free slab count. In particular,
3526 * keg_fetch_slab() always returns slabs on the partial list.
3528 if (slab->us_freecount == 0) {
3529 LIST_REMOVE(slab, us_link);
3530 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3537 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3551 /* Try to keep the buckets totally full */
3552 for (i = 0; i < max; ) {
3553 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3556 stripe = howmany(max, vm_ndomains);
3558 dom = &keg->uk_domain[slab->us_domain];
3559 while (slab->us_freecount && i < max) {
3560 bucket[i++] = slab_alloc_item(keg, slab);
3561 if (dom->ud_free_items <= keg->uk_reserve)
3565 * If the zone is striped we pick a new slab for every
3566 * N allocations. Eliminating this conditional will
3567 * instead pick a new domain for each bucket rather
3568 * than stripe within each bucket. The current option
3569 * produces more fragmentation and requires more cpu
3570 * time but yields better distribution.
3572 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3573 vm_ndomains > 1 && --stripe == 0)
3577 KEG_UNLOCK(keg, slab->us_domain);
3578 /* Don't block if we allocated any successfully. */
3587 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3589 uint64_t old, new, total, max;
3592 * The hard case. We're going to sleep because there were existing
3593 * sleepers or because we ran out of items. This routine enforces
3594 * fairness by keeping fifo order.
3596 * First release our ill gotten gains and make some noise.
3599 zone_free_limit(zone, count);
3600 zone_log_warning(zone);
3601 zone_maxaction(zone);
3602 if (flags & M_NOWAIT)
3606 * We need to allocate an item or set ourself as a sleeper
3607 * while the sleepq lock is held to avoid wakeup races. This
3608 * is essentially a home rolled semaphore.
3610 sleepq_lock(&zone->uz_max_items);
3611 old = zone->uz_items;
3613 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3614 /* Cache the max since we will evaluate twice. */
3615 max = zone->uz_max_items;
3616 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3617 UZ_ITEMS_COUNT(old) >= max)
3618 new = old + UZ_ITEMS_SLEEPER;
3620 new = old + MIN(count, max - old);
3621 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3623 /* We may have successfully allocated under the sleepq lock. */
3624 if (UZ_ITEMS_SLEEPERS(new) == 0) {
3625 sleepq_release(&zone->uz_max_items);
3630 * This is in a different cacheline from uz_items so that we
3631 * don't constantly invalidate the fastpath cacheline when we
3632 * adjust item counts. This could be limited to toggling on
3635 atomic_add_32(&zone->uz_sleepers, 1);
3636 atomic_add_64(&zone->uz_sleeps, 1);
3639 * We have added ourselves as a sleeper. The sleepq lock
3640 * protects us from wakeup races. Sleep now and then retry.
3642 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3643 sleepq_wait(&zone->uz_max_items, PVM);
3646 * After wakeup, remove ourselves as a sleeper and try
3647 * again. We no longer have the sleepq lock for protection.
3649 * Subract ourselves as a sleeper while attempting to add
3652 atomic_subtract_32(&zone->uz_sleepers, 1);
3653 old = atomic_fetchadd_64(&zone->uz_items,
3654 -(UZ_ITEMS_SLEEPER - count));
3655 /* We're no longer a sleeper. */
3656 old -= UZ_ITEMS_SLEEPER;
3659 * If we're still at the limit, restart. Notably do not
3660 * block on other sleepers. Cache the max value to protect
3661 * against changes via sysctl.
3663 total = UZ_ITEMS_COUNT(old);
3664 max = zone->uz_max_items;
3667 /* Truncate if necessary, otherwise wake other sleepers. */
3668 if (total + count > max) {
3669 zone_free_limit(zone, total + count - max);
3670 count = max - total;
3671 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3672 wakeup_one(&zone->uz_max_items);
3679 * Allocate 'count' items from our max_items limit. Returns the number
3680 * available. If M_NOWAIT is not specified it will sleep until at least
3681 * one item can be allocated.
3684 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3689 max = zone->uz_max_items;
3693 * We expect normal allocations to succeed with a simple
3696 old = atomic_fetchadd_64(&zone->uz_items, count);
3697 if (__predict_true(old + count <= max))
3701 * If we had some items and no sleepers just return the
3702 * truncated value. We have to release the excess space
3703 * though because that may wake sleepers who weren't woken
3704 * because we were temporarily over the limit.
3707 zone_free_limit(zone, (old + count) - max);
3710 return (zone_alloc_limit_hard(zone, count, flags));
3714 * Free a number of items back to the limit.
3717 zone_free_limit(uma_zone_t zone, int count)
3724 * In the common case we either have no sleepers or
3725 * are still over the limit and can just return.
3727 old = atomic_fetchadd_64(&zone->uz_items, -count);
3728 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3729 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3733 * Moderate the rate of wakeups. Sleepers will continue
3734 * to generate wakeups if necessary.
3736 wakeup_one(&zone->uz_max_items);
3740 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3742 uma_bucket_t bucket;
3745 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3748 /* Avoid allocs targeting empty domains. */
3749 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3750 domain = UMA_ANYDOMAIN;
3752 if (zone->uz_max_items > 0)
3753 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3756 maxbucket = zone->uz_bucket_size;
3760 /* Don't wait for buckets, preserve caller's NOVM setting. */
3761 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3762 if (bucket == NULL) {
3767 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3768 MIN(maxbucket, bucket->ub_entries), domain, flags);
3771 * Initialize the memory if necessary.
3773 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3776 for (i = 0; i < bucket->ub_cnt; i++)
3777 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3781 * If we couldn't initialize the whole bucket, put the
3782 * rest back onto the freelist.
3784 if (i != bucket->ub_cnt) {
3785 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3786 bucket->ub_cnt - i);
3788 bzero(&bucket->ub_bucket[i],
3789 sizeof(void *) * (bucket->ub_cnt - i));
3795 cnt = bucket->ub_cnt;
3796 if (bucket->ub_cnt == 0) {
3797 bucket_free(zone, bucket, udata);
3798 counter_u64_add(zone->uz_fails, 1);
3802 if (zone->uz_max_items > 0 && cnt < maxbucket)
3803 zone_free_limit(zone, maxbucket - cnt);
3809 * Allocates a single item from a zone.
3812 * zone The zone to alloc for.
3813 * udata The data to be passed to the constructor.
3814 * domain The domain to allocate from or UMA_ANYDOMAIN.
3815 * flags M_WAITOK, M_NOWAIT, M_ZERO.
3818 * NULL if there is no memory and M_NOWAIT is set
3819 * An item if successful
3823 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
3827 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0)
3830 /* Avoid allocs targeting empty domains. */
3831 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3832 domain = UMA_ANYDOMAIN;
3834 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
3838 * We have to call both the zone's init (not the keg's init)
3839 * and the zone's ctor. This is because the item is going from
3840 * a keg slab directly to the user, and the user is expecting it
3841 * to be both zone-init'd as well as zone-ctor'd.
3843 if (zone->uz_init != NULL) {
3844 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
3845 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
3849 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
3854 counter_u64_add(zone->uz_allocs, 1);
3855 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
3856 zone->uz_name, zone);
3861 counter_u64_add(zone->uz_fails, 1);
3863 if (zone->uz_max_items > 0)
3864 zone_free_limit(zone, 1);
3865 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
3866 zone->uz_name, zone);
3873 uma_zfree_smr(uma_zone_t zone, void *item)
3876 uma_cache_bucket_t bucket;
3877 int domain, itemdomain, uz_flags;
3879 #ifdef UMA_ZALLOC_DEBUG
3880 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3881 ("uma_zfree_smr: called with non-SMR zone.\n"));
3882 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
3883 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
3886 cache = &zone->uz_cpu[curcpu];
3887 uz_flags = cache_uz_flags(cache);
3888 domain = itemdomain = 0;
3890 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
3891 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
3895 cache = &zone->uz_cpu[curcpu];
3896 /* SMR Zones must free to the free bucket. */
3897 bucket = &cache->uc_freebucket;
3899 domain = PCPU_GET(domain);
3900 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
3901 domain != itemdomain) {
3902 bucket = &cache->uc_crossbucket;
3905 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
3906 cache_bucket_push(cache, bucket, item);
3910 } while (cache_free(zone, cache, NULL, item, itemdomain));
3914 * If nothing else caught this, we'll just do an internal free.
3916 zone_free_item(zone, item, NULL, SKIP_NONE);
3921 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
3924 uma_cache_bucket_t bucket;
3925 int domain, itemdomain, uz_flags;
3927 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3928 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3930 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
3932 #ifdef UMA_ZALLOC_DEBUG
3933 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3934 ("uma_zfree_arg: called with SMR zone.\n"));
3935 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
3938 /* uma_zfree(..., NULL) does nothing, to match free(9). */
3943 * We are accessing the per-cpu cache without a critical section to
3944 * fetch size and flags. This is acceptable, if we are preempted we
3945 * will simply read another cpu's line.
3947 cache = &zone->uz_cpu[curcpu];
3948 uz_flags = cache_uz_flags(cache);
3949 if (UMA_ALWAYS_CTORDTOR ||
3950 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
3951 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
3954 * The race here is acceptable. If we miss it we'll just have to wait
3955 * a little longer for the limits to be reset.
3957 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
3958 if (zone->uz_sleepers > 0)
3963 * If possible, free to the per-CPU cache. There are two
3964 * requirements for safe access to the per-CPU cache: (1) the thread
3965 * accessing the cache must not be preempted or yield during access,
3966 * and (2) the thread must not migrate CPUs without switching which
3967 * cache it accesses. We rely on a critical section to prevent
3968 * preemption and migration. We release the critical section in
3969 * order to acquire the zone mutex if we are unable to free to the
3970 * current cache; when we re-acquire the critical section, we must
3971 * detect and handle migration if it has occurred.
3973 domain = itemdomain = 0;
3975 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
3976 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
3980 cache = &zone->uz_cpu[curcpu];
3982 * Try to free into the allocbucket first to give LIFO
3983 * ordering for cache-hot datastructures. Spill over
3984 * into the freebucket if necessary. Alloc will swap
3985 * them if one runs dry.
3987 bucket = &cache->uc_allocbucket;
3989 domain = PCPU_GET(domain);
3990 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
3991 domain != itemdomain) {
3992 bucket = &cache->uc_crossbucket;
3995 if (bucket->ucb_cnt >= bucket->ucb_entries)
3996 bucket = &cache->uc_freebucket;
3997 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
3998 cache_bucket_push(cache, bucket, item);
4002 } while (cache_free(zone, cache, udata, item, itemdomain));
4006 * If nothing else caught this, we'll just do an internal free.
4009 zone_free_item(zone, item, udata, SKIP_DTOR);
4014 * sort crossdomain free buckets to domain correct buckets and cache
4018 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4020 struct uma_bucketlist fullbuckets;
4021 uma_zone_domain_t zdom;
4028 "uma_zfree: zone %s(%p) draining cross bucket %p",
4029 zone->uz_name, zone, bucket);
4031 STAILQ_INIT(&fullbuckets);
4034 * To avoid having ndomain * ndomain buckets for sorting we have a
4035 * lock on the current crossfree bucket. A full matrix with
4036 * per-domain locking could be used if necessary.
4038 ZONE_CROSS_LOCK(zone);
4041 * It is possible for buckets to arrive here out of order so we fetch
4042 * the current smr seq rather than accepting the bucket's.
4044 seq = SMR_SEQ_INVALID;
4045 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4046 seq = smr_current(zone->uz_smr);
4047 while (bucket->ub_cnt > 0) {
4048 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4049 domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4050 zdom = &zone->uz_domain[domain];
4051 if (zdom->uzd_cross == NULL) {
4052 zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
4053 if (zdom->uzd_cross == NULL)
4056 b = zdom->uzd_cross;
4057 b->ub_bucket[b->ub_cnt++] = item;
4059 if (b->ub_cnt == b->ub_entries) {
4060 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4061 zdom->uzd_cross = NULL;
4065 ZONE_CROSS_UNLOCK(zone);
4066 if (!STAILQ_EMPTY(&fullbuckets)) {
4068 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4069 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4070 if (zone->uz_bkt_count >= zone->uz_bkt_max) {
4072 bucket_drain(zone, b);
4073 bucket_free(zone, b, udata);
4076 domain = _vm_phys_domain(
4078 (vm_offset_t)b->ub_bucket[0]));
4079 zdom = &zone->uz_domain[domain];
4080 zone_put_bucket(zone, zdom, b, true);
4085 if (bucket->ub_cnt != 0)
4086 bucket_drain(zone, bucket);
4087 bucket->ub_seq = SMR_SEQ_INVALID;
4088 bucket_free(zone, bucket, udata);
4093 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4094 int domain, int itemdomain)
4096 uma_zone_domain_t zdom;
4100 * Buckets coming from the wrong domain will be entirely for the
4101 * only other domain on two domain systems. In this case we can
4102 * simply cache them. Otherwise we need to sort them back to
4105 if (domain != itemdomain && vm_ndomains > 2) {
4106 zone_free_cross(zone, bucket, udata);
4112 * Attempt to save the bucket in the zone's domain bucket cache.
4114 * We bump the uz count when the cache size is insufficient to
4115 * handle the working set.
4117 if (ZONE_TRYLOCK(zone) == 0) {
4118 /* Record contention to size the buckets. */
4120 if (zone->uz_bucket_size < zone->uz_bucket_size_max)
4121 zone->uz_bucket_size++;
4125 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4126 zone->uz_name, zone, bucket);
4127 /* ub_cnt is pointing to the last free item */
4128 KASSERT(bucket->ub_cnt == bucket->ub_entries,
4129 ("uma_zfree: Attempting to insert partial bucket onto the full list.\n"));
4130 if (zone->uz_bkt_count >= zone->uz_bkt_max) {
4132 bucket_drain(zone, bucket);
4133 bucket_free(zone, bucket, udata);
4135 zdom = &zone->uz_domain[itemdomain];
4136 zone_put_bucket(zone, zdom, bucket, true);
4142 * Populate a free or cross bucket for the current cpu cache. Free any
4143 * existing full bucket either to the zone cache or back to the slab layer.
4145 * Enters and returns in a critical section. false return indicates that
4146 * we can not satisfy this free in the cache layer. true indicates that
4147 * the caller should retry.
4149 static __noinline bool
4150 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4153 uma_cache_bucket_t cbucket;
4154 uma_bucket_t newbucket, bucket;
4157 CRITICAL_ASSERT(curthread);
4159 if (zone->uz_bucket_size == 0)
4162 cache = &zone->uz_cpu[curcpu];
4166 * FIRSTTOUCH domains need to free to the correct zdom. When
4167 * enabled this is the zdom of the item. The bucket is the
4168 * cross bucket if the current domain and itemdomain do not match.
4170 cbucket = &cache->uc_freebucket;
4172 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) {
4173 domain = PCPU_GET(domain);
4174 if (domain != itemdomain) {
4175 cbucket = &cache->uc_crossbucket;
4176 if (cbucket->ucb_cnt != 0)
4177 atomic_add_64(&zone->uz_xdomain,
4182 itemdomain = domain = 0;
4183 bucket = cache_bucket_unload(cbucket);
4185 /* We are no longer associated with this CPU. */
4189 * Don't let SMR zones operate without a free bucket. Force
4190 * a synchronize and re-use this one. We will only degrade
4191 * to a synchronize every bucket_size items rather than every
4192 * item if we fail to allocate a bucket.
4194 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4196 bucket->ub_seq = smr_advance(zone->uz_smr);
4197 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4198 if (newbucket == NULL && bucket != NULL) {
4199 bucket_drain(zone, bucket);
4203 } else if (!bucketdisable)
4204 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4207 zone_free_bucket(zone, bucket, udata, domain, itemdomain);
4210 if ((bucket = newbucket) == NULL)
4212 cache = &zone->uz_cpu[curcpu];
4215 * Check to see if we should be populating the cross bucket. If it
4216 * is already populated we will fall through and attempt to populate
4219 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) {
4220 domain = PCPU_GET(domain);
4221 if (domain != itemdomain &&
4222 cache->uc_crossbucket.ucb_bucket == NULL) {
4223 cache_bucket_load_cross(cache, bucket);
4229 * We may have lost the race to fill the bucket or switched CPUs.
4231 if (cache->uc_freebucket.ucb_bucket != NULL) {
4233 bucket_free(zone, bucket, udata);
4236 cache_bucket_load_free(cache, bucket);
4242 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
4245 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4246 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4248 CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone);
4250 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
4251 ("uma_zfree_domain: called with spinlock or critical section held"));
4253 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4256 zone_free_item(zone, item, udata, SKIP_NONE);
4260 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4267 KEG_LOCK_ASSERT(keg, slab->us_domain);
4269 /* Do we need to remove from any lists? */
4270 dom = &keg->uk_domain[slab->us_domain];
4271 if (slab->us_freecount + 1 == keg->uk_ipers) {
4272 LIST_REMOVE(slab, us_link);
4273 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4274 dom->ud_free_slabs++;
4275 } else if (slab->us_freecount == 0) {
4276 LIST_REMOVE(slab, us_link);
4277 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4280 /* Slab management. */
4281 freei = slab_item_index(slab, keg, item);
4282 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4283 slab->us_freecount++;
4285 /* Keg statistics. */
4286 dom->ud_free_items++;
4290 zone_release(void *arg, void **bucket, int cnt)
4303 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4304 lock = KEG_LOCK(keg, 0);
4305 for (i = 0; i < cnt; i++) {
4307 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4308 slab = vtoslab((vm_offset_t)item);
4310 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4311 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4312 slab = hash_sfind(&keg->uk_hash, mem);
4314 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4316 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4319 lock = KEG_LOCK(keg, slab->us_domain);
4321 slab_free_item(zone, slab, item);
4328 * Frees a single item to any zone.
4331 * zone The zone to free to
4332 * item The item we're freeing
4333 * udata User supplied data for the dtor
4334 * skip Skip dtors and finis
4336 static __noinline void
4337 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4341 * If a free is sent directly to an SMR zone we have to
4342 * synchronize immediately because the item can instantly
4343 * be reallocated. This should only happen in degenerate
4344 * cases when no memory is available for per-cpu caches.
4346 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4347 smr_synchronize(zone->uz_smr);
4349 item_dtor(zone, item, zone->uz_size, udata, skip);
4351 if (skip < SKIP_FINI && zone->uz_fini)
4352 zone->uz_fini(item, zone->uz_size);
4354 zone->uz_release(zone->uz_arg, &item, 1);
4356 if (skip & SKIP_CNT)
4359 counter_u64_add(zone->uz_frees, 1);
4361 if (zone->uz_max_items > 0)
4362 zone_free_limit(zone, 1);
4367 uma_zone_set_max(uma_zone_t zone, int nitems)
4369 struct uma_bucket_zone *ubz;
4373 * XXX This can misbehave if the zone has any allocations with
4374 * no limit and a limit is imposed. There is currently no
4375 * way to clear a limit.
4378 ubz = bucket_zone_max(zone, nitems);
4379 count = ubz != NULL ? ubz->ubz_entries : 0;
4380 zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4381 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4382 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4383 zone->uz_max_items = nitems;
4384 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4385 zone_update_caches(zone);
4386 /* We may need to wake waiters. */
4387 wakeup(&zone->uz_max_items);
4395 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4397 struct uma_bucket_zone *ubz;
4401 ubz = bucket_zone_max(zone, nitems);
4404 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4405 /* Count the cross-domain bucket. */
4407 nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4408 zone->uz_bucket_size_max = ubz->ubz_entries;
4410 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4412 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4413 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4414 zone->uz_bkt_max = nitems;
4420 uma_zone_get_max(uma_zone_t zone)
4424 nitems = atomic_load_64(&zone->uz_max_items);
4431 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4434 ZONE_ASSERT_COLD(zone);
4435 zone->uz_warning = warning;
4440 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4443 ZONE_ASSERT_COLD(zone);
4444 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4449 uma_zone_get_cur(uma_zone_t zone)
4455 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4456 nitems = counter_u64_fetch(zone->uz_allocs) -
4457 counter_u64_fetch(zone->uz_frees);
4459 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4460 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4462 return (nitems < 0 ? 0 : nitems);
4466 uma_zone_get_allocs(uma_zone_t zone)
4472 if (zone->uz_allocs != EARLY_COUNTER)
4473 nitems = counter_u64_fetch(zone->uz_allocs);
4475 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4481 uma_zone_get_frees(uma_zone_t zone)
4487 if (zone->uz_frees != EARLY_COUNTER)
4488 nitems = counter_u64_fetch(zone->uz_frees);
4490 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4496 /* Used only for KEG_ASSERT_COLD(). */
4498 uma_keg_get_allocs(uma_keg_t keg)
4504 LIST_FOREACH(z, &keg->uk_zones, uz_link)
4505 nitems += uma_zone_get_allocs(z);
4513 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4518 KEG_ASSERT_COLD(keg);
4519 keg->uk_init = uminit;
4524 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4529 KEG_ASSERT_COLD(keg);
4530 keg->uk_fini = fini;
4535 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4538 ZONE_ASSERT_COLD(zone);
4539 zone->uz_init = zinit;
4544 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4547 ZONE_ASSERT_COLD(zone);
4548 zone->uz_fini = zfini;
4553 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4558 KEG_ASSERT_COLD(keg);
4559 keg->uk_freef = freef;
4564 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4569 KEG_ASSERT_COLD(keg);
4570 keg->uk_allocf = allocf;
4575 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4578 ZONE_ASSERT_COLD(zone);
4580 zone->uz_flags |= UMA_ZONE_SMR;
4582 zone_update_caches(zone);
4586 uma_zone_get_smr(uma_zone_t zone)
4589 return (zone->uz_smr);
4594 uma_zone_reserve(uma_zone_t zone, int items)
4599 KEG_ASSERT_COLD(keg);
4600 keg->uk_reserve = items;
4605 uma_zone_reserve_kva(uma_zone_t zone, int count)
4612 KEG_ASSERT_COLD(keg);
4613 ZONE_ASSERT_COLD(zone);
4615 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4617 #ifdef UMA_MD_SMALL_ALLOC
4618 if (keg->uk_ppera > 1) {
4622 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4629 MPASS(keg->uk_kva == 0);
4632 zone->uz_max_items = pages * keg->uk_ipers;
4633 #ifdef UMA_MD_SMALL_ALLOC
4634 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4636 keg->uk_allocf = noobj_alloc;
4638 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4639 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4640 zone_update_caches(zone);
4648 uma_prealloc(uma_zone_t zone, int items)
4650 struct vm_domainset_iter di;
4654 int aflags, domain, slabs;
4657 slabs = howmany(items, keg->uk_ipers);
4658 while (slabs-- > 0) {
4660 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4663 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4666 dom = &keg->uk_domain[slab->us_domain];
4668 * keg_alloc_slab() always returns a slab on the
4671 LIST_REMOVE(slab, us_link);
4672 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4674 dom->ud_free_slabs++;
4675 KEG_UNLOCK(keg, slab->us_domain);
4678 if (vm_domainset_iter_policy(&di, &domain) != 0)
4679 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
4686 uma_reclaim(int req)
4689 CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4690 sx_xlock(&uma_reclaim_lock);
4694 case UMA_RECLAIM_TRIM:
4695 zone_foreach(zone_trim, NULL);
4697 case UMA_RECLAIM_DRAIN:
4698 case UMA_RECLAIM_DRAIN_CPU:
4699 zone_foreach(zone_drain, NULL);
4700 if (req == UMA_RECLAIM_DRAIN_CPU) {
4701 pcpu_cache_drain_safe(NULL);
4702 zone_foreach(zone_drain, NULL);
4706 panic("unhandled reclamation request %d", req);
4710 * Some slabs may have been freed but this zone will be visited early
4711 * we visit again so that we can free pages that are empty once other
4712 * zones are drained. We have to do the same for buckets.
4714 zone_drain(slabzones[0], NULL);
4715 zone_drain(slabzones[1], NULL);
4716 bucket_zone_drain();
4717 sx_xunlock(&uma_reclaim_lock);
4720 static volatile int uma_reclaim_needed;
4723 uma_reclaim_wakeup(void)
4726 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4727 wakeup(uma_reclaim);
4731 uma_reclaim_worker(void *arg __unused)
4735 sx_xlock(&uma_reclaim_lock);
4736 while (atomic_load_int(&uma_reclaim_needed) == 0)
4737 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4739 sx_xunlock(&uma_reclaim_lock);
4740 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4741 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4742 atomic_store_int(&uma_reclaim_needed, 0);
4743 /* Don't fire more than once per-second. */
4744 pause("umarclslp", hz);
4750 uma_zone_reclaim(uma_zone_t zone, int req)
4754 case UMA_RECLAIM_TRIM:
4755 zone_trim(zone, NULL);
4757 case UMA_RECLAIM_DRAIN:
4758 zone_drain(zone, NULL);
4760 case UMA_RECLAIM_DRAIN_CPU:
4761 pcpu_cache_drain_safe(zone);
4762 zone_drain(zone, NULL);
4765 panic("unhandled reclamation request %d", req);
4771 uma_zone_exhausted(uma_zone_t zone)
4774 return (atomic_load_32(&zone->uz_sleepers) > 0);
4781 return (uma_kmem_limit);
4785 uma_set_limit(unsigned long limit)
4788 uma_kmem_limit = limit;
4795 return (atomic_load_long(&uma_kmem_total));
4802 return (uma_kmem_limit - uma_size());
4807 * Generate statistics across both the zone and its per-cpu cache's. Return
4808 * desired statistics if the pointer is non-NULL for that statistic.
4810 * Note: does not update the zone statistics, as it can't safely clear the
4811 * per-CPU cache statistic.
4815 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4816 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4819 uint64_t allocs, frees, sleeps, xdomain;
4822 allocs = frees = sleeps = xdomain = 0;
4825 cache = &z->uz_cpu[cpu];
4826 cachefree += cache->uc_allocbucket.ucb_cnt;
4827 cachefree += cache->uc_freebucket.ucb_cnt;
4828 xdomain += cache->uc_crossbucket.ucb_cnt;
4829 cachefree += cache->uc_crossbucket.ucb_cnt;
4830 allocs += cache->uc_allocs;
4831 frees += cache->uc_frees;
4833 allocs += counter_u64_fetch(z->uz_allocs);
4834 frees += counter_u64_fetch(z->uz_frees);
4835 sleeps += z->uz_sleeps;
4836 xdomain += z->uz_xdomain;
4837 if (cachefreep != NULL)
4838 *cachefreep = cachefree;
4839 if (allocsp != NULL)
4843 if (sleepsp != NULL)
4845 if (xdomainp != NULL)
4846 *xdomainp = xdomain;
4851 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
4858 rw_rlock(&uma_rwlock);
4859 LIST_FOREACH(kz, &uma_kegs, uk_link) {
4860 LIST_FOREACH(z, &kz->uk_zones, uz_link)
4863 LIST_FOREACH(z, &uma_cachezones, uz_link)
4866 rw_runlock(&uma_rwlock);
4867 return (sysctl_handle_int(oidp, &count, 0, req));
4871 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
4872 struct uma_percpu_stat *ups, bool internal)
4874 uma_zone_domain_t zdom;
4879 for (i = 0; i < vm_ndomains; i++) {
4880 zdom = &z->uz_domain[i];
4881 uth->uth_zone_free += zdom->uzd_nitems;
4883 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
4884 uth->uth_frees = counter_u64_fetch(z->uz_frees);
4885 uth->uth_fails = counter_u64_fetch(z->uz_fails);
4886 uth->uth_sleeps = z->uz_sleeps;
4887 uth->uth_xdomain = z->uz_xdomain;
4890 * While it is not normally safe to access the cache bucket pointers
4891 * while not on the CPU that owns the cache, we only allow the pointers
4892 * to be exchanged without the zone lock held, not invalidated, so
4893 * accept the possible race associated with bucket exchange during
4894 * monitoring. Use atomic_load_ptr() to ensure that the bucket pointers
4895 * are loaded only once.
4897 for (i = 0; i < mp_maxid + 1; i++) {
4898 bzero(&ups[i], sizeof(*ups));
4899 if (internal || CPU_ABSENT(i))
4901 cache = &z->uz_cpu[i];
4902 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
4903 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
4904 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
4905 ups[i].ups_allocs = cache->uc_allocs;
4906 ups[i].ups_frees = cache->uc_frees;
4911 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
4913 struct uma_stream_header ush;
4914 struct uma_type_header uth;
4915 struct uma_percpu_stat *ups;
4920 uint32_t kfree, pages;
4921 int count, error, i;
4923 error = sysctl_wire_old_buffer(req, 0);
4926 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
4927 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
4928 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
4931 rw_rlock(&uma_rwlock);
4932 LIST_FOREACH(kz, &uma_kegs, uk_link) {
4933 LIST_FOREACH(z, &kz->uk_zones, uz_link)
4937 LIST_FOREACH(z, &uma_cachezones, uz_link)
4941 * Insert stream header.
4943 bzero(&ush, sizeof(ush));
4944 ush.ush_version = UMA_STREAM_VERSION;
4945 ush.ush_maxcpus = (mp_maxid + 1);
4946 ush.ush_count = count;
4947 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
4949 LIST_FOREACH(kz, &uma_kegs, uk_link) {
4951 for (i = 0; i < vm_ndomains; i++) {
4952 kfree += kz->uk_domain[i].ud_free_items;
4953 pages += kz->uk_domain[i].ud_pages;
4955 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
4956 bzero(&uth, sizeof(uth));
4958 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
4959 uth.uth_align = kz->uk_align;
4960 uth.uth_size = kz->uk_size;
4961 uth.uth_rsize = kz->uk_rsize;
4962 if (z->uz_max_items > 0) {
4963 items = UZ_ITEMS_COUNT(z->uz_items);
4964 uth.uth_pages = (items / kz->uk_ipers) *
4967 uth.uth_pages = pages;
4968 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
4970 uth.uth_limit = z->uz_max_items;
4971 uth.uth_keg_free = kfree;
4974 * A zone is secondary is it is not the first entry
4975 * on the keg's zone list.
4977 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
4978 (LIST_FIRST(&kz->uk_zones) != z))
4979 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
4980 uma_vm_zone_stats(&uth, z, &sbuf, ups,
4981 kz->uk_flags & UMA_ZFLAG_INTERNAL);
4983 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
4984 for (i = 0; i < mp_maxid + 1; i++)
4985 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
4988 LIST_FOREACH(z, &uma_cachezones, uz_link) {
4989 bzero(&uth, sizeof(uth));
4991 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
4992 uth.uth_size = z->uz_size;
4993 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
4995 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
4996 for (i = 0; i < mp_maxid + 1; i++)
4997 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5000 rw_runlock(&uma_rwlock);
5001 error = sbuf_finish(&sbuf);
5008 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5010 uma_zone_t zone = *(uma_zone_t *)arg1;
5013 max = uma_zone_get_max(zone);
5014 error = sysctl_handle_int(oidp, &max, 0, req);
5015 if (error || !req->newptr)
5018 uma_zone_set_max(zone, max);
5024 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5030 * Some callers want to add sysctls for global zones that
5031 * may not yet exist so they pass a pointer to a pointer.
5034 zone = *(uma_zone_t *)arg1;
5037 cur = uma_zone_get_cur(zone);
5038 return (sysctl_handle_int(oidp, &cur, 0, req));
5042 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5044 uma_zone_t zone = arg1;
5047 cur = uma_zone_get_allocs(zone);
5048 return (sysctl_handle_64(oidp, &cur, 0, req));
5052 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5054 uma_zone_t zone = arg1;
5057 cur = uma_zone_get_frees(zone);
5058 return (sysctl_handle_64(oidp, &cur, 0, req));
5062 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5065 uma_zone_t zone = arg1;
5068 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5069 if (zone->uz_flags != 0)
5070 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5072 sbuf_printf(&sbuf, "0");
5073 error = sbuf_finish(&sbuf);
5080 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5082 uma_keg_t keg = arg1;
5083 int avail, effpct, total;
5085 total = keg->uk_ppera * PAGE_SIZE;
5086 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5087 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5089 * We consider the client's requested size and alignment here, not the
5090 * real size determination uk_rsize, because we also adjust the real
5091 * size for internal implementation reasons (max bitset size).
5093 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5094 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5095 avail *= mp_maxid + 1;
5096 effpct = 100 * avail / total;
5097 return (sysctl_handle_int(oidp, &effpct, 0, req));
5101 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5103 uma_zone_t zone = arg1;
5106 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5107 return (sysctl_handle_64(oidp, &cur, 0, req));
5112 uma_dbg_getslab(uma_zone_t zone, void *item)
5119 * It is safe to return the slab here even though the
5120 * zone is unlocked because the item's allocation state
5121 * essentially holds a reference.
5123 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5124 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5126 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5127 return (vtoslab((vm_offset_t)mem));
5129 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5130 return ((uma_slab_t)(mem + keg->uk_pgoff));
5132 slab = hash_sfind(&keg->uk_hash, mem);
5139 uma_dbg_zskip(uma_zone_t zone, void *mem)
5142 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5145 return (uma_dbg_kskip(zone->uz_keg, mem));
5149 uma_dbg_kskip(uma_keg_t keg, void *mem)
5153 if (dbg_divisor == 0)
5156 if (dbg_divisor == 1)
5159 idx = (uintptr_t)mem >> PAGE_SHIFT;
5160 if (keg->uk_ipers > 1) {
5161 idx *= keg->uk_ipers;
5162 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5165 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5166 counter_u64_add(uma_skip_cnt, 1);
5169 counter_u64_add(uma_dbg_cnt, 1);
5175 * Set up the slab's freei data such that uma_dbg_free can function.
5179 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5185 slab = uma_dbg_getslab(zone, item);
5187 panic("uma: item %p did not belong to zone %s\n",
5188 item, zone->uz_name);
5191 freei = slab_item_index(slab, keg, item);
5193 if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5194 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
5195 item, zone, zone->uz_name, slab, freei);
5196 BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5200 * Verifies freed addresses. Checks for alignment, valid slab membership
5201 * and duplicate frees.
5205 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5211 slab = uma_dbg_getslab(zone, item);
5213 panic("uma: Freed item %p did not belong to zone %s\n",
5214 item, zone->uz_name);
5217 freei = slab_item_index(slab, keg, item);
5219 if (freei >= keg->uk_ipers)
5220 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
5221 item, zone, zone->uz_name, slab, freei);
5223 if (slab_item(slab, keg, freei) != item)
5224 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
5225 item, zone, zone->uz_name, slab, freei);
5227 if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5228 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
5229 item, zone, zone->uz_name, slab, freei);
5231 BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5233 #endif /* INVARIANTS */
5237 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5238 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5243 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5244 *allocs = counter_u64_fetch(z->uz_allocs);
5245 frees = counter_u64_fetch(z->uz_frees);
5246 *sleeps = z->uz_sleeps;
5250 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5252 for (i = 0; i < vm_ndomains; i++) {
5253 *cachefree += z->uz_domain[i].uzd_nitems;
5254 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5255 (LIST_FIRST(&kz->uk_zones) != z)))
5256 *cachefree += kz->uk_domain[i].ud_free_items;
5258 *used = *allocs - frees;
5259 return (((int64_t)*used + *cachefree) * kz->uk_size);
5262 DB_SHOW_COMMAND(uma, db_show_uma)
5264 const char *fmt_hdr, *fmt_entry;
5267 uint64_t allocs, used, sleeps, xdomain;
5269 /* variables for sorting */
5271 uma_zone_t cur_zone, last_zone;
5272 int64_t cur_size, last_size, size;
5275 /* /i option produces machine-parseable CSV output */
5276 if (modif[0] == 'i') {
5277 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5278 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5280 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5281 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5284 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5285 "Sleeps", "Bucket", "Total Mem", "XFree");
5287 /* Sort the zones with largest size first. */
5289 last_size = INT64_MAX;
5294 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5295 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5297 * In the case of size ties, print out zones
5298 * in the order they are encountered. That is,
5299 * when we encounter the most recently output
5300 * zone, we have already printed all preceding
5301 * ties, and we must print all following ties.
5303 if (z == last_zone) {
5307 size = get_uma_stats(kz, z, &allocs, &used,
5308 &sleeps, &cachefree, &xdomain);
5309 if (size > cur_size && size < last_size + ties)
5317 if (cur_zone == NULL)
5320 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5321 &sleeps, &cachefree, &xdomain);
5322 db_printf(fmt_entry, cur_zone->uz_name,
5323 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5324 (uintmax_t)allocs, (uintmax_t)sleeps,
5325 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5330 last_zone = cur_zone;
5331 last_size = cur_size;
5335 DB_SHOW_COMMAND(umacache, db_show_umacache)
5338 uint64_t allocs, frees;
5342 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5343 "Requests", "Bucket");
5344 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5345 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5346 for (i = 0; i < vm_ndomains; i++)
5347 cachefree += z->uz_domain[i].uzd_nitems;
5348 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5349 z->uz_name, (uintmax_t)z->uz_size,
5350 (intmax_t)(allocs - frees), cachefree,
5351 (uintmax_t)allocs, z->uz_bucket_size);