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);
327 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
328 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
330 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
331 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
332 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
333 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
335 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD, 0,
336 "Memory allocation debugging");
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");
366 * Select the slab zone for an offpage slab with the given maximum item count.
368 static inline uma_zone_t
372 return (slabzones[ipers > SLABZONE0_SETSIZE]);
376 * This routine checks to see whether or not it's safe to enable buckets.
382 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
383 bucketdisable = vm_page_count_min();
387 * Initialize bucket_zones, the array of zones of buckets of various sizes.
389 * For each zone, calculate the memory required for each bucket, consisting
390 * of the header and an array of pointers.
395 struct uma_bucket_zone *ubz;
398 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
399 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
400 size += sizeof(void *) * ubz->ubz_entries;
401 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
402 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
403 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
404 UMA_ZONE_FIRSTTOUCH);
409 * Given a desired number of entries for a bucket, return the zone from which
410 * to allocate the bucket.
412 static struct uma_bucket_zone *
413 bucket_zone_lookup(int entries)
415 struct uma_bucket_zone *ubz;
417 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
418 if (ubz->ubz_entries >= entries)
424 static struct uma_bucket_zone *
425 bucket_zone_max(uma_zone_t zone, int nitems)
427 struct uma_bucket_zone *ubz;
431 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
432 /* Count the cross-domain bucket. */
435 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
436 if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems)
438 if (ubz == &bucket_zones[0])
446 bucket_select(int size)
448 struct uma_bucket_zone *ubz;
450 ubz = &bucket_zones[0];
451 if (size > ubz->ubz_maxsize)
452 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
454 for (; ubz->ubz_entries != 0; ubz++)
455 if (ubz->ubz_maxsize < size)
458 return (ubz->ubz_entries);
462 bucket_alloc(uma_zone_t zone, void *udata, int flags)
464 struct uma_bucket_zone *ubz;
468 * Don't allocate buckets early in boot.
470 if (__predict_false(booted < BOOT_KVA))
474 * To limit bucket recursion we store the original zone flags
475 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
476 * NOVM flag to persist even through deep recursions. We also
477 * store ZFLAG_BUCKET once we have recursed attempting to allocate
478 * a bucket for a bucket zone so we do not allow infinite bucket
479 * recursion. This cookie will even persist to frees of unused
480 * buckets via the allocation path or bucket allocations in the
483 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
484 udata = (void *)(uintptr_t)zone->uz_flags;
486 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
488 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
490 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
492 ubz = bucket_zone_lookup(zone->uz_bucket_size);
493 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
495 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
498 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
501 bucket->ub_entries = ubz->ubz_entries;
502 bucket->ub_seq = SMR_SEQ_INVALID;
503 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
504 zone->uz_name, zone, bucket);
511 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
513 struct uma_bucket_zone *ubz;
515 KASSERT(bucket->ub_cnt == 0,
516 ("bucket_free: Freeing a non free bucket."));
517 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
518 ("bucket_free: Freeing an SMR bucket."));
519 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
520 udata = (void *)(uintptr_t)zone->uz_flags;
521 ubz = bucket_zone_lookup(bucket->ub_entries);
522 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
526 bucket_zone_drain(void)
528 struct uma_bucket_zone *ubz;
530 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
531 uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN);
535 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
536 * zone's caches. If a bucket is found the zone is not locked on return.
539 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom)
545 ZONE_LOCK_ASSERT(zone);
547 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
550 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
551 bucket->ub_seq != SMR_SEQ_INVALID) {
552 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
554 bucket->ub_seq = SMR_SEQ_INVALID;
555 dtor = (zone->uz_dtor != NULL) | UMA_ALWAYS_CTORDTOR;
557 MPASS(zdom->uzd_nitems >= bucket->ub_cnt);
558 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
559 zdom->uzd_nitems -= bucket->ub_cnt;
560 if (zdom->uzd_imin > zdom->uzd_nitems)
561 zdom->uzd_imin = zdom->uzd_nitems;
562 zone->uz_bkt_count -= bucket->ub_cnt;
565 for (i = 0; i < bucket->ub_cnt; i++)
566 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
573 * Insert a full bucket into the specified cache. The "ws" parameter indicates
574 * whether the bucket's contents should be counted as part of the zone's working
578 zone_put_bucket(uma_zone_t zone, uma_zone_domain_t zdom, uma_bucket_t bucket,
582 ZONE_LOCK_ASSERT(zone);
583 KASSERT(!ws || zone->uz_bkt_count < zone->uz_bkt_max,
584 ("%s: zone %p overflow", __func__, zone));
586 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
587 zdom->uzd_nitems += bucket->ub_cnt;
588 if (ws && zdom->uzd_imax < zdom->uzd_nitems)
589 zdom->uzd_imax = zdom->uzd_nitems;
590 zone->uz_bkt_count += bucket->ub_cnt;
593 /* Pops an item out of a per-cpu cache bucket. */
595 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
599 CRITICAL_ASSERT(curthread);
602 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
604 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
605 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
612 /* Pushes an item into a per-cpu cache bucket. */
614 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
617 CRITICAL_ASSERT(curthread);
618 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
619 ("uma_zfree: Freeing to non free bucket index."));
621 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
627 * Unload a UMA bucket from a per-cpu cache.
629 static inline uma_bucket_t
630 cache_bucket_unload(uma_cache_bucket_t bucket)
634 b = bucket->ucb_bucket;
636 MPASS(b->ub_entries == bucket->ucb_entries);
637 b->ub_cnt = bucket->ucb_cnt;
638 bucket->ucb_bucket = NULL;
639 bucket->ucb_entries = bucket->ucb_cnt = 0;
645 static inline uma_bucket_t
646 cache_bucket_unload_alloc(uma_cache_t cache)
649 return (cache_bucket_unload(&cache->uc_allocbucket));
652 static inline uma_bucket_t
653 cache_bucket_unload_free(uma_cache_t cache)
656 return (cache_bucket_unload(&cache->uc_freebucket));
659 static inline uma_bucket_t
660 cache_bucket_unload_cross(uma_cache_t cache)
663 return (cache_bucket_unload(&cache->uc_crossbucket));
667 * Load a bucket into a per-cpu cache bucket.
670 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
673 CRITICAL_ASSERT(curthread);
674 MPASS(bucket->ucb_bucket == NULL);
676 bucket->ucb_bucket = b;
677 bucket->ucb_cnt = b->ub_cnt;
678 bucket->ucb_entries = b->ub_entries;
682 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
685 cache_bucket_load(&cache->uc_allocbucket, b);
689 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
692 cache_bucket_load(&cache->uc_freebucket, b);
697 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
700 cache_bucket_load(&cache->uc_crossbucket, b);
705 * Copy and preserve ucb_spare.
708 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
711 b1->ucb_bucket = b2->ucb_bucket;
712 b1->ucb_entries = b2->ucb_entries;
713 b1->ucb_cnt = b2->ucb_cnt;
717 * Swap two cache buckets.
720 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
722 struct uma_cache_bucket b3;
724 CRITICAL_ASSERT(curthread);
726 cache_bucket_copy(&b3, b1);
727 cache_bucket_copy(b1, b2);
728 cache_bucket_copy(b2, &b3);
732 zone_log_warning(uma_zone_t zone)
734 static const struct timeval warninterval = { 300, 0 };
736 if (!zone_warnings || zone->uz_warning == NULL)
739 if (ratecheck(&zone->uz_ratecheck, &warninterval))
740 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
744 zone_maxaction(uma_zone_t zone)
747 if (zone->uz_maxaction.ta_func != NULL)
748 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
752 * Routine called by timeout which is used to fire off some time interval
753 * based calculations. (stats, hash size, etc.)
762 uma_timeout(void *unused)
765 zone_foreach(zone_timeout, NULL);
767 /* Reschedule this event */
768 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
772 * Update the working set size estimate for the zone's bucket cache.
773 * The constants chosen here are somewhat arbitrary. With an update period of
774 * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
778 zone_domain_update_wss(uma_zone_domain_t zdom)
782 MPASS(zdom->uzd_imax >= zdom->uzd_imin);
783 wss = zdom->uzd_imax - zdom->uzd_imin;
784 zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
785 zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
789 * Routine to perform timeout driven calculations. This expands the
790 * hashes and does per cpu statistics aggregation.
795 zone_timeout(uma_zone_t zone, void *unused)
800 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
806 * Hash zones are non-numa by definition so the first domain
807 * is the only one present.
810 pages = keg->uk_domain[0].ud_pages;
813 * Expand the keg hash table.
815 * This is done if the number of slabs is larger than the hash size.
816 * What I'm trying to do here is completely reduce collisions. This
817 * may be a little aggressive. Should I allow for two collisions max?
819 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
820 struct uma_hash newhash;
821 struct uma_hash oldhash;
825 * This is so involved because allocating and freeing
826 * while the keg lock is held will lead to deadlock.
827 * I have to do everything in stages and check for
831 ret = hash_alloc(&newhash, 1 << fls(slabs));
834 if (hash_expand(&keg->uk_hash, &newhash)) {
835 oldhash = keg->uk_hash;
836 keg->uk_hash = newhash;
849 for (int i = 0; i < vm_ndomains; i++)
850 zone_domain_update_wss(&zone->uz_domain[i]);
855 * Allocate and zero fill the next sized hash table from the appropriate
859 * hash A new hash structure with the old hash size in uh_hashsize
862 * 1 on success and 0 on failure.
865 hash_alloc(struct uma_hash *hash, u_int size)
869 KASSERT(powerof2(size), ("hash size must be power of 2"));
870 if (size > UMA_HASH_SIZE_INIT) {
871 hash->uh_hashsize = size;
872 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
873 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
875 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
876 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
877 UMA_ANYDOMAIN, M_WAITOK);
878 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
880 if (hash->uh_slab_hash) {
881 bzero(hash->uh_slab_hash, alloc);
882 hash->uh_hashmask = hash->uh_hashsize - 1;
890 * Expands the hash table for HASH zones. This is done from zone_timeout
891 * to reduce collisions. This must not be done in the regular allocation
892 * path, otherwise, we can recurse on the vm while allocating pages.
895 * oldhash The hash you want to expand
896 * newhash The hash structure for the new table
904 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
906 uma_hash_slab_t slab;
910 if (!newhash->uh_slab_hash)
913 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
917 * I need to investigate hash algorithms for resizing without a
921 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
922 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
923 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
924 LIST_REMOVE(slab, uhs_hlink);
925 hval = UMA_HASH(newhash, slab->uhs_data);
926 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
934 * Free the hash bucket to the appropriate backing store.
937 * slab_hash The hash bucket we're freeing
938 * hashsize The number of entries in that hash bucket
944 hash_free(struct uma_hash *hash)
946 if (hash->uh_slab_hash == NULL)
948 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
949 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
951 free(hash->uh_slab_hash, M_UMAHASH);
955 * Frees all outstanding items in a bucket
958 * zone The zone to free to, must be unlocked.
959 * bucket The free/alloc bucket with items.
966 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
970 if (bucket == NULL || bucket->ub_cnt == 0)
973 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
974 bucket->ub_seq != SMR_SEQ_INVALID) {
975 smr_wait(zone->uz_smr, bucket->ub_seq);
976 for (i = 0; i < bucket->ub_cnt; i++)
977 item_dtor(zone, bucket->ub_bucket[i],
978 zone->uz_size, NULL, SKIP_NONE);
979 bucket->ub_seq = SMR_SEQ_INVALID;
982 for (i = 0; i < bucket->ub_cnt; i++)
983 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
984 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
985 if (zone->uz_max_items > 0)
986 zone_free_limit(zone, bucket->ub_cnt);
988 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
994 * Drains the per cpu caches for a zone.
996 * NOTE: This may only be called while the zone is being torn down, and not
997 * during normal operation. This is necessary in order that we do not have
998 * to migrate CPUs to drain the per-CPU caches.
1001 * zone The zone to drain, must be unlocked.
1007 cache_drain(uma_zone_t zone)
1010 uma_bucket_t bucket;
1014 * XXX: It is safe to not lock the per-CPU caches, because we're
1015 * tearing down the zone anyway. I.e., there will be no further use
1016 * of the caches at this point.
1018 * XXX: It would good to be able to assert that the zone is being
1019 * torn down to prevent improper use of cache_drain().
1022 cache = &zone->uz_cpu[cpu];
1023 bucket = cache_bucket_unload_alloc(cache);
1024 if (bucket != NULL) {
1025 bucket_drain(zone, bucket);
1026 bucket_free(zone, bucket, NULL);
1028 bucket = cache_bucket_unload_free(cache);
1029 if (bucket != NULL) {
1030 bucket_drain(zone, bucket);
1031 bucket_free(zone, bucket, NULL);
1033 bucket = cache_bucket_unload_cross(cache);
1034 if (bucket != NULL) {
1035 bucket_drain(zone, bucket);
1036 bucket_free(zone, bucket, NULL);
1039 bucket_cache_reclaim(zone, true);
1043 cache_shrink(uma_zone_t zone, void *unused)
1046 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1050 zone->uz_bucket_size =
1051 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1056 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1059 uma_bucket_t b1, b2, b3;
1062 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1065 b1 = b2 = b3 = NULL;
1068 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
1069 domain = PCPU_GET(domain);
1072 cache = &zone->uz_cpu[curcpu];
1073 b1 = cache_bucket_unload_alloc(cache);
1074 if (b1 != NULL && b1->ub_cnt != 0) {
1075 zone_put_bucket(zone, &zone->uz_domain[domain], b1, false);
1080 * Don't flush SMR zone buckets. This leaves the zone without a
1081 * bucket and forces every free to synchronize().
1083 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1085 b2 = cache_bucket_unload_free(cache);
1086 if (b2 != NULL && b2->ub_cnt != 0) {
1087 zone_put_bucket(zone, &zone->uz_domain[domain], b2, false);
1090 b3 = cache_bucket_unload_cross(cache);
1096 bucket_free(zone, b1, NULL);
1098 bucket_free(zone, b2, NULL);
1100 bucket_drain(zone, b3);
1101 bucket_free(zone, b3, NULL);
1106 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1107 * This is an expensive call because it needs to bind to all CPUs
1108 * one by one and enter a critical section on each of them in order
1109 * to safely access their cache buckets.
1110 * Zone lock must not be held on call this function.
1113 pcpu_cache_drain_safe(uma_zone_t zone)
1118 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1121 cache_shrink(zone, NULL);
1123 zone_foreach(cache_shrink, NULL);
1126 thread_lock(curthread);
1127 sched_bind(curthread, cpu);
1128 thread_unlock(curthread);
1131 cache_drain_safe_cpu(zone, NULL);
1133 zone_foreach(cache_drain_safe_cpu, NULL);
1135 thread_lock(curthread);
1136 sched_unbind(curthread);
1137 thread_unlock(curthread);
1141 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1142 * requested a drain, otherwise the per-domain caches are trimmed to either
1143 * estimated working set size.
1146 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1148 uma_zone_domain_t zdom;
1149 uma_bucket_t bucket;
1150 long target, tofree;
1153 for (i = 0; i < vm_ndomains; i++) {
1155 * The cross bucket is partially filled and not part of
1156 * the item count. Reclaim it individually here.
1158 zdom = &zone->uz_domain[i];
1159 ZONE_CROSS_LOCK(zone);
1160 bucket = zdom->uzd_cross;
1161 zdom->uzd_cross = NULL;
1162 ZONE_CROSS_UNLOCK(zone);
1163 if (bucket != NULL) {
1164 bucket_drain(zone, bucket);
1165 bucket_free(zone, bucket, NULL);
1169 * Shrink the zone bucket size to ensure that the per-CPU caches
1170 * don't grow too large.
1173 if (i == 0 && zone->uz_bucket_size > zone->uz_bucket_size_min)
1174 zone->uz_bucket_size--;
1177 * If we were asked to drain the zone, we are done only once
1178 * this bucket cache is empty. Otherwise, we reclaim items in
1179 * excess of the zone's estimated working set size. If the
1180 * difference nitems - imin is larger than the WSS estimate,
1181 * then the estimate will grow at the end of this interval and
1182 * we ignore the historical average.
1184 target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1186 while (zdom->uzd_nitems > target) {
1187 bucket = STAILQ_FIRST(&zdom->uzd_buckets);
1190 tofree = bucket->ub_cnt;
1191 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
1192 zdom->uzd_nitems -= tofree;
1195 * Shift the bounds of the current WSS interval to avoid
1196 * perturbing the estimate.
1198 zdom->uzd_imax -= lmin(zdom->uzd_imax, tofree);
1199 zdom->uzd_imin -= lmin(zdom->uzd_imin, tofree);
1202 bucket_drain(zone, bucket);
1203 bucket_free(zone, bucket, NULL);
1211 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1217 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1218 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1220 mem = slab_data(slab, keg);
1221 flags = slab->us_flags;
1223 if (keg->uk_fini != NULL) {
1224 for (i--; i > -1; i--)
1227 * trash_fini implies that dtor was trash_dtor. trash_fini
1228 * would check that memory hasn't been modified since free,
1229 * which executed trash_dtor.
1230 * That's why we need to run uma_dbg_kskip() check here,
1231 * albeit we don't make skip check for other init/fini
1234 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1235 keg->uk_fini != trash_fini)
1237 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1239 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1240 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1242 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1243 uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1247 * Frees pages from a keg back to the system. This is done on demand from
1248 * the pageout daemon.
1253 keg_drain(uma_keg_t keg)
1255 struct slabhead freeslabs = { 0 };
1257 uma_slab_t slab, tmp;
1261 * We don't want to take pages from statically allocated kegs at this
1264 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1267 for (i = 0; i < vm_ndomains; i++) {
1268 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1269 keg->uk_name, keg, i, dom->ud_free);
1271 dom = &keg->uk_domain[i];
1273 LIST_FOREACH_SAFE(slab, &dom->ud_free_slab, us_link, tmp) {
1274 if (keg->uk_flags & UMA_ZFLAG_HASH)
1275 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1277 LIST_REMOVE(slab, us_link);
1278 LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1280 dom->ud_pages -= n * keg->uk_ppera;
1281 dom->ud_free -= n * keg->uk_ipers;
1285 while ((slab = LIST_FIRST(&freeslabs)) != NULL) {
1286 LIST_REMOVE(slab, us_link);
1287 keg_free_slab(keg, slab, keg->uk_ipers);
1292 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1296 * Set draining to interlock with zone_dtor() so we can release our
1297 * locks as we go. Only dtor() should do a WAITOK call since it
1298 * is the only call that knows the structure will still be available
1302 while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1303 if (waitok == M_NOWAIT)
1305 msleep(zone, &zone->uz_lock, PVM, "zonedrain", 1);
1307 zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1309 bucket_cache_reclaim(zone, drain);
1312 * The DRAINING flag protects us from being freed while
1313 * we're running. Normally the uma_rwlock would protect us but we
1314 * must be able to release and acquire the right lock for each keg.
1316 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1317 keg_drain(zone->uz_keg);
1319 zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1326 zone_drain(uma_zone_t zone, void *unused)
1329 zone_reclaim(zone, M_NOWAIT, true);
1333 zone_trim(uma_zone_t zone, void *unused)
1336 zone_reclaim(zone, M_NOWAIT, false);
1340 * Allocate a new slab for a keg and inserts it into the partial slab list.
1341 * The keg should be unlocked on entry. If the allocation succeeds it will
1342 * be locked on return.
1345 * flags Wait flags for the item initialization routine
1346 * aflags Wait flags for the slab allocation
1349 * The slab that was allocated or NULL if there is no memory and the
1350 * caller specified M_NOWAIT.
1353 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1364 KASSERT(domain >= 0 && domain < vm_ndomains,
1365 ("keg_alloc_slab: domain %d out of range", domain));
1367 allocf = keg->uk_allocf;
1370 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1371 uma_hash_slab_t hslab;
1372 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1376 slab = &hslab->uhs_slab;
1380 * This reproduces the old vm_zone behavior of zero filling pages the
1381 * first time they are added to a zone.
1383 * Malloced items are zeroed in uma_zalloc.
1386 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1391 if (keg->uk_flags & UMA_ZONE_NODUMP)
1394 /* zone is passed for legacy reasons. */
1395 size = keg->uk_ppera * PAGE_SIZE;
1396 mem = allocf(zone, size, domain, &sflags, aflags);
1398 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1399 zone_free_item(slabzone(keg->uk_ipers),
1400 slab_tohashslab(slab), NULL, SKIP_NONE);
1403 uma_total_inc(size);
1405 /* For HASH zones all pages go to the same uma_domain. */
1406 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1409 /* Point the slab into the allocated memory */
1410 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1411 slab = (uma_slab_t )(mem + keg->uk_pgoff);
1413 slab_tohashslab(slab)->uhs_data = mem;
1415 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1416 for (i = 0; i < keg->uk_ppera; i++)
1417 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1420 slab->us_freecount = keg->uk_ipers;
1421 slab->us_flags = sflags;
1422 slab->us_domain = domain;
1424 BIT_FILL(keg->uk_ipers, &slab->us_free);
1426 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1429 if (keg->uk_init != NULL) {
1430 for (i = 0; i < keg->uk_ipers; i++)
1431 if (keg->uk_init(slab_item(slab, keg, i),
1432 keg->uk_size, flags) != 0)
1434 if (i != keg->uk_ipers) {
1435 keg_free_slab(keg, slab, i);
1439 KEG_LOCK(keg, domain);
1441 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1442 slab, keg->uk_name, keg);
1444 if (keg->uk_flags & UMA_ZFLAG_HASH)
1445 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1448 * If we got a slab here it's safe to mark it partially used
1449 * and return. We assume that the caller is going to remove
1450 * at least one item.
1452 dom = &keg->uk_domain[domain];
1453 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1454 dom->ud_pages += keg->uk_ppera;
1455 dom->ud_free += keg->uk_ipers;
1464 * This function is intended to be used early on in place of page_alloc() so
1465 * that we may use the boot time page cache to satisfy allocations before
1469 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1478 pages = howmany(bytes, PAGE_SIZE);
1479 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1481 *pflag = UMA_SLAB_BOOT;
1482 m = vm_page_alloc_contig_domain(NULL, 0, domain,
1483 malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1484 (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1488 pa = VM_PAGE_TO_PHYS(m);
1489 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1490 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1491 defined(__riscv) || defined(__powerpc64__)
1492 if ((wait & M_NODUMP) == 0)
1496 /* Allocate KVA and indirectly advance bootmem. */
1497 mem = (void *)pmap_map(&bootmem, m->phys_addr,
1498 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1499 if ((wait & M_ZERO) != 0)
1500 bzero(mem, pages * PAGE_SIZE);
1506 startup_free(void *mem, vm_size_t bytes)
1511 va = (vm_offset_t)mem;
1512 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1513 pmap_remove(kernel_pmap, va, va + bytes);
1514 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1515 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1516 defined(__riscv) || defined(__powerpc64__)
1517 dump_drop_page(VM_PAGE_TO_PHYS(m));
1519 vm_page_unwire_noq(m);
1525 * Allocates a number of pages from the system
1528 * bytes The number of bytes requested
1529 * wait Shall we wait?
1532 * A pointer to the alloced memory or possibly
1533 * NULL if M_NOWAIT is set.
1536 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1539 void *p; /* Returned page */
1541 *pflag = UMA_SLAB_KERNEL;
1542 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1548 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1551 struct pglist alloctail;
1552 vm_offset_t addr, zkva;
1554 vm_page_t p, p_next;
1559 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1561 TAILQ_INIT(&alloctail);
1562 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1563 malloc2vm_flags(wait);
1564 *pflag = UMA_SLAB_KERNEL;
1565 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1566 if (CPU_ABSENT(cpu)) {
1567 p = vm_page_alloc(NULL, 0, flags);
1570 p = vm_page_alloc(NULL, 0, flags);
1572 pc = pcpu_find(cpu);
1573 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1576 p = vm_page_alloc_domain(NULL, 0,
1577 pc->pc_domain, flags);
1578 if (__predict_false(p == NULL))
1579 p = vm_page_alloc(NULL, 0, flags);
1582 if (__predict_false(p == NULL))
1584 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1586 if ((addr = kva_alloc(bytes)) == 0)
1589 TAILQ_FOREACH(p, &alloctail, listq) {
1590 pmap_qenter(zkva, &p, 1);
1593 return ((void*)addr);
1595 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1596 vm_page_unwire_noq(p);
1603 * Allocates a number of pages from within an object
1606 * bytes The number of bytes requested
1607 * wait Shall we wait?
1610 * A pointer to the alloced memory or possibly
1611 * NULL if M_NOWAIT is set.
1614 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1617 TAILQ_HEAD(, vm_page) alloctail;
1619 vm_offset_t retkva, zkva;
1620 vm_page_t p, p_next;
1623 TAILQ_INIT(&alloctail);
1626 npages = howmany(bytes, PAGE_SIZE);
1627 while (npages > 0) {
1628 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1629 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1630 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1634 * Since the page does not belong to an object, its
1637 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1642 * Page allocation failed, free intermediate pages and
1645 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1646 vm_page_unwire_noq(p);
1651 *flags = UMA_SLAB_PRIV;
1652 zkva = keg->uk_kva +
1653 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1655 TAILQ_FOREACH(p, &alloctail, listq) {
1656 pmap_qenter(zkva, &p, 1);
1660 return ((void *)retkva);
1664 * Allocate physically contiguous pages.
1667 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1671 *pflag = UMA_SLAB_KERNEL;
1672 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1673 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1677 * Frees a number of pages to the system
1680 * mem A pointer to the memory to be freed
1681 * size The size of the memory being freed
1682 * flags The original p->us_flags field
1688 page_free(void *mem, vm_size_t size, uint8_t flags)
1691 if ((flags & UMA_SLAB_BOOT) != 0) {
1692 startup_free(mem, size);
1696 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1697 ("UMA: page_free used with invalid flags %x", flags));
1699 kmem_free((vm_offset_t)mem, size);
1703 * Frees pcpu zone allocations
1706 * mem A pointer to the memory to be freed
1707 * size The size of the memory being freed
1708 * flags The original p->us_flags field
1714 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1716 vm_offset_t sva, curva;
1720 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1722 if ((flags & UMA_SLAB_BOOT) != 0) {
1723 startup_free(mem, size);
1727 sva = (vm_offset_t)mem;
1728 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1729 paddr = pmap_kextract(curva);
1730 m = PHYS_TO_VM_PAGE(paddr);
1731 vm_page_unwire_noq(m);
1734 pmap_qremove(sva, size >> PAGE_SHIFT);
1735 kva_free(sva, size);
1740 * Zero fill initializer
1742 * Arguments/Returns follow uma_init specifications
1745 zero_init(void *mem, int size, int flags)
1753 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1756 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1761 * Actual size of embedded struct slab (!OFFPAGE).
1764 slab_sizeof(int nitems)
1768 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1769 return (roundup(s, UMA_ALIGN_PTR + 1));
1773 * Size of memory for embedded slabs (!OFFPAGE).
1776 slab_space(int nitems)
1778 return (UMA_SLAB_SIZE - slab_sizeof(nitems));
1781 #define UMA_FIXPT_SHIFT 31
1782 #define UMA_FRAC_FIXPT(n, d) \
1783 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1784 #define UMA_FIXPT_PCT(f) \
1785 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1786 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
1787 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1790 * Compute the number of items that will fit in a slab. If hdr is true, the
1791 * item count may be limited to provide space in the slab for an inline slab
1792 * header. Otherwise, all slab space will be provided for item storage.
1795 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1800 /* The padding between items is not needed after the last item. */
1801 padpi = rsize - size;
1805 * Start with the maximum item count and remove items until
1806 * the slab header first alongside the allocatable memory.
1808 for (ipers = MIN(SLAB_MAX_SETSIZE,
1809 (slabsize + padpi - slab_sizeof(1)) / rsize);
1811 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1815 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1822 * Compute the number of items that will fit in a slab for a startup zone.
1825 slab_ipers(size_t size, int align)
1829 rsize = roundup(size, align + 1); /* Assume no CACHESPREAD */
1830 return (slab_ipers_hdr(size, rsize, UMA_SLAB_SIZE, true));
1834 * Determine the format of a uma keg. This determines where the slab header
1835 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
1838 * keg The zone we should initialize
1844 keg_layout(uma_keg_t keg)
1851 u_int ipers_offpage;
1856 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1857 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
1858 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
1859 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
1860 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
1862 KASSERT((keg->uk_flags &
1863 (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY)) == 0 ||
1864 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
1865 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
1868 alignsize = keg->uk_align + 1;
1873 * Calculate the size of each allocation (rsize) according to
1874 * alignment. If the requested size is smaller than we have
1875 * allocation bits for we round it up.
1877 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
1878 rsize = roundup2(rsize, alignsize);
1880 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0) {
1881 slabsize = UMA_PCPU_ALLOC_SIZE;
1882 pages = mp_maxid + 1;
1883 } else if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
1885 * We want one item to start on every align boundary in a page.
1886 * To do this we will span pages. We will also extend the item
1887 * by the size of align if it is an even multiple of align.
1888 * Otherwise, it would fall on the same boundary every time.
1890 if ((rsize & alignsize) == 0)
1892 slabsize = rsize * (PAGE_SIZE / alignsize);
1893 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
1894 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
1895 pages = howmany(slabsize, PAGE_SIZE);
1896 slabsize = ptoa(pages);
1899 * Choose a slab size of as many pages as it takes to represent
1900 * a single item. We will then try to fit as many additional
1901 * items into the slab as possible. At some point, we may want
1902 * to increase the slab size for awkward item sizes in order to
1903 * increase efficiency.
1905 pages = howmany(keg->uk_size, PAGE_SIZE);
1906 slabsize = ptoa(pages);
1909 /* Evaluate an inline slab layout. */
1910 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
1911 ipers = slab_ipers_hdr(keg->uk_size, rsize, slabsize, true);
1913 /* TODO: vm_page-embedded slab. */
1916 * We can't do OFFPAGE if we're internal or if we've been
1917 * asked to not go to the VM for buckets. If we do this we
1918 * may end up going to the VM for slabs which we do not
1919 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1920 * of UMA_ZONE_VM, which clearly forbids it.
1922 if ((keg->uk_flags &
1923 (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY)) != 0) {
1925 /* We need an extra page for the slab header. */
1927 slabsize = ptoa(pages);
1928 ipers = slab_ipers_hdr(keg->uk_size, rsize, slabsize,
1935 * See if using an OFFPAGE slab will improve our efficiency.
1936 * Only do this if we are below our efficiency threshold.
1938 * XXX We could try growing slabsize to limit max waste as well.
1939 * Historically this was not done because the VM could not
1940 * efficiently handle contiguous allocations.
1942 eff = UMA_FRAC_FIXPT(ipers * rsize, slabsize);
1943 ipers_offpage = slab_ipers_hdr(keg->uk_size, rsize, slabsize, false);
1944 eff_offpage = UMA_FRAC_FIXPT(ipers_offpage * rsize,
1945 slabsize + slabzone(ipers_offpage)->uz_keg->uk_rsize);
1946 if (ipers == 0 || (eff < UMA_MIN_EFF && eff < eff_offpage)) {
1947 CTR5(KTR_UMA, "UMA decided we need offpage slab headers for "
1948 "keg: %s(%p), minimum efficiency allowed = %u%%, "
1949 "old efficiency = %u%%, offpage efficiency = %u%%",
1950 keg->uk_name, keg, UMA_FIXPT_PCT(UMA_MIN_EFF),
1951 UMA_FIXPT_PCT(eff), UMA_FIXPT_PCT(eff_offpage));
1952 format = UMA_ZFLAG_OFFPAGE;
1953 ipers = ipers_offpage;
1958 * How do we find the slab header if it is offpage or if not all item
1959 * start addresses are in the same page? We could solve the latter
1960 * case with vaddr alignment, but we don't.
1962 if ((format & UMA_ZFLAG_OFFPAGE) != 0 ||
1963 (ipers - 1) * rsize >= PAGE_SIZE) {
1964 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
1965 format |= UMA_ZFLAG_HASH;
1967 format |= UMA_ZFLAG_VTOSLAB;
1969 keg->uk_ipers = ipers;
1970 keg->uk_rsize = rsize;
1971 keg->uk_flags |= format;
1972 keg->uk_ppera = pages;
1973 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
1974 __func__, keg->uk_name, keg->uk_flags, rsize, ipers, pages);
1975 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
1976 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
1977 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize, ipers,
1982 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1983 * the keg onto the global keg list.
1985 * Arguments/Returns follow uma_ctor specifications
1986 * udata Actually uma_kctor_args
1989 keg_ctor(void *mem, int size, void *udata, int flags)
1991 struct uma_kctor_args *arg = udata;
1992 uma_keg_t keg = mem;
1997 keg->uk_size = arg->size;
1998 keg->uk_init = arg->uminit;
1999 keg->uk_fini = arg->fini;
2000 keg->uk_align = arg->align;
2001 keg->uk_reserve = 0;
2002 keg->uk_flags = arg->flags;
2005 * We use a global round-robin policy by default. Zones with
2006 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2007 * case the iterator is never run.
2009 keg->uk_dr.dr_policy = DOMAINSET_RR();
2010 keg->uk_dr.dr_iter = 0;
2013 * The master zone is passed to us at keg-creation time.
2016 keg->uk_name = zone->uz_name;
2018 if (arg->flags & UMA_ZONE_VM)
2019 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
2021 if (arg->flags & UMA_ZONE_ZINIT)
2022 keg->uk_init = zero_init;
2024 if (arg->flags & UMA_ZONE_MALLOC)
2025 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2028 keg->uk_flags &= ~UMA_ZONE_PCPU;
2034 * Use a first-touch NUMA policy for all kegs that pmap_extract()
2035 * will work on with the exception of critical VM structures
2036 * necessary for paging.
2038 * Zones may override the default by specifying either.
2041 if ((keg->uk_flags &
2042 (UMA_ZFLAG_HASH | UMA_ZONE_VM | UMA_ZONE_ROUNDROBIN)) == 0)
2043 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2044 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2045 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2049 * If we haven't booted yet we need allocations to go through the
2050 * startup cache until the vm is ready.
2052 #ifdef UMA_MD_SMALL_ALLOC
2053 if (keg->uk_ppera == 1)
2054 keg->uk_allocf = uma_small_alloc;
2057 if (booted < BOOT_KVA)
2058 keg->uk_allocf = startup_alloc;
2059 else if (keg->uk_flags & UMA_ZONE_PCPU)
2060 keg->uk_allocf = pcpu_page_alloc;
2061 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2062 keg->uk_allocf = contig_alloc;
2064 keg->uk_allocf = page_alloc;
2065 #ifdef UMA_MD_SMALL_ALLOC
2066 if (keg->uk_ppera == 1)
2067 keg->uk_freef = uma_small_free;
2070 if (keg->uk_flags & UMA_ZONE_PCPU)
2071 keg->uk_freef = pcpu_page_free;
2073 keg->uk_freef = page_free;
2076 * Initialize keg's locks.
2078 for (i = 0; i < vm_ndomains; i++)
2079 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2082 * If we're putting the slab header in the actual page we need to
2083 * figure out where in each page it goes. See slab_sizeof
2086 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2089 shsize = slab_sizeof(keg->uk_ipers);
2090 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2092 * The only way the following is possible is if with our
2093 * UMA_ALIGN_PTR adjustments we are now bigger than
2094 * UMA_SLAB_SIZE. I haven't checked whether this is
2095 * mathematically possible for all cases, so we make
2098 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2099 ("zone %s ipers %d rsize %d size %d slab won't fit",
2100 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2103 if (keg->uk_flags & UMA_ZFLAG_HASH)
2104 hash_alloc(&keg->uk_hash, 0);
2106 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2108 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2110 rw_wlock(&uma_rwlock);
2111 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2112 rw_wunlock(&uma_rwlock);
2117 zone_kva_available(uma_zone_t zone, void *unused)
2121 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2125 if (keg->uk_allocf == startup_alloc) {
2126 /* Switch to the real allocator. */
2127 if (keg->uk_flags & UMA_ZONE_PCPU)
2128 keg->uk_allocf = pcpu_page_alloc;
2129 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2131 keg->uk_allocf = contig_alloc;
2133 keg->uk_allocf = page_alloc;
2138 zone_alloc_counters(uma_zone_t zone, void *unused)
2141 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2142 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2143 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2147 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2149 uma_zone_domain_t zdom;
2152 struct sysctl_oid *oid, *domainoid;
2153 int domains, i, cnt;
2154 static const char *nokeg = "cache zone";
2158 * Make a sysctl safe copy of the zone name by removing
2159 * any special characters and handling dups by appending
2162 if (zone->uz_namecnt != 0) {
2163 /* Count the number of decimal digits and '_' separator. */
2164 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2166 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2168 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2171 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2172 for (c = zone->uz_ctlname; *c != '\0'; c++)
2173 if (strchr("./\\ -", *c) != NULL)
2177 * Basic parameters at the root.
2179 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2180 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD, NULL, "");
2182 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2183 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2184 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2185 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2186 zone, 0, sysctl_handle_uma_zone_flags, "A",
2187 "Allocator configuration flags");
2188 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2189 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2190 "Desired per-cpu cache size");
2191 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2192 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2193 "Maximum allowed per-cpu cache size");
2198 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2199 domains = vm_ndomains;
2202 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2203 "keg", CTLFLAG_RD, NULL, "");
2205 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2206 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2207 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2208 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2209 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2210 "Real object size with alignment");
2211 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2212 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2213 "pages per-slab allocation");
2214 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2215 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2216 "items available per-slab");
2217 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2218 "align", CTLFLAG_RD, &keg->uk_align, 0,
2219 "item alignment mask");
2220 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2221 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2222 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2223 "Slab utilization (100 - internal fragmentation %)");
2224 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2225 OID_AUTO, "domain", CTLFLAG_RD, NULL, "");
2226 for (i = 0; i < domains; i++) {
2227 dom = &keg->uk_domain[i];
2228 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2229 OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD,
2231 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2232 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2233 "Total pages currently allocated from VM");
2234 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2235 "free", CTLFLAG_RD, &dom->ud_free, 0,
2236 "items free in the slab layer");
2239 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2240 "name", CTLFLAG_RD, nokeg, "Keg name");
2243 * Information about zone limits.
2245 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2246 "limit", CTLFLAG_RD, NULL, "");
2247 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2248 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2249 zone, 0, sysctl_handle_uma_zone_items, "QU",
2250 "current number of allocated items if limit is set");
2251 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2252 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2253 "Maximum number of cached items");
2254 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2255 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2256 "Number of threads sleeping at limit");
2257 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2258 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2259 "Total zone limit sleeps");
2260 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2261 "bucket_max", CTLFLAG_RD, &zone->uz_bkt_max, 0,
2262 "Maximum number of items in the bucket cache");
2263 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2264 "bucket_cnt", CTLFLAG_RD, &zone->uz_bkt_count, 0,
2265 "Number of items in the bucket cache");
2268 * Per-domain zone information.
2270 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2271 OID_AUTO, "domain", CTLFLAG_RD, NULL, "");
2272 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2274 for (i = 0; i < domains; i++) {
2275 zdom = &zone->uz_domain[i];
2276 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2277 OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD, NULL, "");
2278 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2279 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2280 "number of items in this domain");
2281 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2282 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2283 "maximum item count in this period");
2284 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2285 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2286 "minimum item count in this period");
2287 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2288 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2289 "Working set size");
2293 * General statistics.
2295 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2296 "stats", CTLFLAG_RD, NULL, "");
2297 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2298 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2299 zone, 1, sysctl_handle_uma_zone_cur, "I",
2300 "Current number of allocated items");
2301 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2302 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2303 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2304 "Total allocation calls");
2305 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2306 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2307 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2308 "Total free calls");
2309 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2310 "fails", CTLFLAG_RD, &zone->uz_fails,
2311 "Number of allocation failures");
2312 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2313 "xdomain", CTLFLAG_RD, &zone->uz_xdomain, 0,
2314 "Free calls from the wrong domain");
2317 struct uma_zone_count {
2323 zone_count(uma_zone_t zone, void *arg)
2325 struct uma_zone_count *cnt;
2329 * Some zones are rapidly created with identical names and
2330 * destroyed out of order. This can lead to gaps in the count.
2331 * Use one greater than the maximum observed for this name.
2333 if (strcmp(zone->uz_name, cnt->name) == 0)
2334 cnt->count = MAX(cnt->count,
2335 zone->uz_namecnt + 1);
2339 zone_update_caches(uma_zone_t zone)
2343 for (i = 0; i <= mp_maxid; i++) {
2344 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2345 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2350 * Zone header ctor. This initializes all fields, locks, etc.
2352 * Arguments/Returns follow uma_ctor specifications
2353 * udata Actually uma_zctor_args
2356 zone_ctor(void *mem, int size, void *udata, int flags)
2358 struct uma_zone_count cnt;
2359 struct uma_zctor_args *arg = udata;
2360 uma_zone_t zone = mem;
2366 zone->uz_name = arg->name;
2367 zone->uz_ctor = arg->ctor;
2368 zone->uz_dtor = arg->dtor;
2369 zone->uz_init = NULL;
2370 zone->uz_fini = NULL;
2371 zone->uz_sleeps = 0;
2372 zone->uz_xdomain = 0;
2373 zone->uz_bucket_size = 0;
2374 zone->uz_bucket_size_min = 0;
2375 zone->uz_bucket_size_max = BUCKET_MAX;
2376 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2377 zone->uz_warning = NULL;
2378 /* The domain structures follow the cpu structures. */
2380 (struct uma_zone_domain *)&zone->uz_cpu[mp_maxid + 1];
2381 zone->uz_bkt_max = ULONG_MAX;
2382 timevalclear(&zone->uz_ratecheck);
2384 /* Count the number of duplicate names. */
2385 cnt.name = arg->name;
2387 zone_foreach(zone_count, &cnt);
2388 zone->uz_namecnt = cnt.count;
2389 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
2390 ZONE_CROSS_LOCK_INIT(zone);
2392 for (i = 0; i < vm_ndomains; i++)
2393 STAILQ_INIT(&zone->uz_domain[i].uzd_buckets);
2396 if (arg->uminit == trash_init && arg->fini == trash_fini)
2397 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2401 * This is a pure cache zone, no kegs.
2404 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2405 ("zone_ctor: Import specified for non-cache zone."));
2406 if (arg->flags & UMA_ZONE_VM)
2407 arg->flags |= UMA_ZFLAG_CACHEONLY;
2408 zone->uz_flags = arg->flags;
2409 zone->uz_size = arg->size;
2410 zone->uz_import = arg->import;
2411 zone->uz_release = arg->release;
2412 zone->uz_arg = arg->arg;
2413 rw_wlock(&uma_rwlock);
2414 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2415 rw_wunlock(&uma_rwlock);
2420 * Use the regular zone/keg/slab allocator.
2422 zone->uz_import = zone_import;
2423 zone->uz_release = zone_release;
2424 zone->uz_arg = zone;
2427 if (arg->flags & UMA_ZONE_SECONDARY) {
2428 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2429 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2430 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2431 zone->uz_init = arg->uminit;
2432 zone->uz_fini = arg->fini;
2433 zone->uz_flags |= UMA_ZONE_SECONDARY;
2434 rw_wlock(&uma_rwlock);
2436 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2437 if (LIST_NEXT(z, uz_link) == NULL) {
2438 LIST_INSERT_AFTER(z, zone, uz_link);
2443 rw_wunlock(&uma_rwlock);
2444 } else if (keg == NULL) {
2445 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2446 arg->align, arg->flags)) == NULL)
2449 struct uma_kctor_args karg;
2452 /* We should only be here from uma_startup() */
2453 karg.size = arg->size;
2454 karg.uminit = arg->uminit;
2455 karg.fini = arg->fini;
2456 karg.align = arg->align;
2457 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2459 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2465 /* Inherit properties from the keg. */
2467 zone->uz_size = keg->uk_size;
2468 zone->uz_flags |= (keg->uk_flags &
2469 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2472 if (__predict_true(booted >= BOOT_RUNNING)) {
2473 zone_alloc_counters(zone, NULL);
2474 zone_alloc_sysctl(zone, NULL);
2476 zone->uz_allocs = EARLY_COUNTER;
2477 zone->uz_frees = EARLY_COUNTER;
2478 zone->uz_fails = EARLY_COUNTER;
2481 /* Caller requests a private SMR context. */
2482 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2483 zone->uz_smr = smr_create(zone->uz_name);
2485 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2486 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2487 ("Invalid zone flag combination"));
2488 if (arg->flags & UMA_ZFLAG_INTERNAL)
2489 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2490 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2491 zone->uz_bucket_size = BUCKET_MAX;
2492 else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2493 zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2494 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2495 zone->uz_bucket_size = 0;
2497 zone->uz_bucket_size = bucket_select(zone->uz_size);
2498 zone->uz_bucket_size_min = zone->uz_bucket_size;
2499 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2500 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2501 zone_update_caches(zone);
2507 * Keg header dtor. This frees all data, destroys locks, frees the hash
2508 * table and removes the keg from the global list.
2510 * Arguments/Returns follow uma_dtor specifications
2514 keg_dtor(void *arg, int size, void *udata)
2517 uint32_t free, pages;
2520 keg = (uma_keg_t)arg;
2522 for (i = 0; i < vm_ndomains; i++) {
2523 free += keg->uk_domain[i].ud_free;
2524 pages += keg->uk_domain[i].ud_pages;
2525 KEG_LOCK_FINI(keg, i);
2528 printf("Freed UMA keg (%s) was not empty (%u items). "
2529 " Lost %u pages of memory.\n",
2530 keg->uk_name ? keg->uk_name : "",
2531 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2533 hash_free(&keg->uk_hash);
2539 * Arguments/Returns follow uma_dtor specifications
2543 zone_dtor(void *arg, int size, void *udata)
2548 zone = (uma_zone_t)arg;
2550 sysctl_remove_oid(zone->uz_oid, 1, 1);
2552 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2555 rw_wlock(&uma_rwlock);
2556 LIST_REMOVE(zone, uz_link);
2557 rw_wunlock(&uma_rwlock);
2559 * XXX there are some races here where
2560 * the zone can be drained but zone lock
2561 * released and then refilled before we
2562 * remove it... we dont care for now
2564 zone_reclaim(zone, M_WAITOK, true);
2566 * We only destroy kegs from non secondary/non cache zones.
2568 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2570 rw_wlock(&uma_rwlock);
2571 LIST_REMOVE(keg, uk_link);
2572 rw_wunlock(&uma_rwlock);
2573 zone_free_item(kegs, keg, NULL, SKIP_NONE);
2575 counter_u64_free(zone->uz_allocs);
2576 counter_u64_free(zone->uz_frees);
2577 counter_u64_free(zone->uz_fails);
2578 free(zone->uz_ctlname, M_UMA);
2579 ZONE_LOCK_FINI(zone);
2580 ZONE_CROSS_LOCK_FINI(zone);
2584 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2589 LIST_FOREACH(keg, &uma_kegs, uk_link) {
2590 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2593 LIST_FOREACH(zone, &uma_cachezones, uz_link)
2598 * Traverses every zone in the system and calls a callback
2601 * zfunc A pointer to a function which accepts a zone
2608 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2611 rw_rlock(&uma_rwlock);
2612 zone_foreach_unlocked(zfunc, arg);
2613 rw_runlock(&uma_rwlock);
2617 * Initialize the kernel memory allocator. This is done after pages can be
2618 * allocated but before general KVA is available.
2621 uma_startup1(vm_offset_t virtual_avail)
2623 struct uma_zctor_args args;
2624 size_t ksize, zsize, size;
2625 uma_keg_t masterkeg;
2629 bootstart = bootmem = virtual_avail;
2631 rw_init(&uma_rwlock, "UMA lock");
2632 sx_init(&uma_reclaim_lock, "umareclaim");
2634 ksize = sizeof(struct uma_keg) +
2635 (sizeof(struct uma_domain) * vm_ndomains);
2636 ksize = roundup(ksize, UMA_SUPER_ALIGN);
2637 zsize = sizeof(struct uma_zone) +
2638 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2639 (sizeof(struct uma_zone_domain) * vm_ndomains);
2640 zsize = roundup(zsize, UMA_SUPER_ALIGN);
2642 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2643 size = (zsize * 2) + ksize;
2644 m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO);
2645 zones = (uma_zone_t)m;
2647 kegs = (uma_zone_t)m;
2649 masterkeg = (uma_keg_t)m;
2651 /* "manually" create the initial zone */
2652 memset(&args, 0, sizeof(args));
2653 args.name = "UMA Kegs";
2655 args.ctor = keg_ctor;
2656 args.dtor = keg_dtor;
2657 args.uminit = zero_init;
2659 args.keg = masterkeg;
2660 args.align = UMA_SUPER_ALIGN - 1;
2661 args.flags = UMA_ZFLAG_INTERNAL;
2662 zone_ctor(kegs, zsize, &args, M_WAITOK);
2664 args.name = "UMA Zones";
2666 args.ctor = zone_ctor;
2667 args.dtor = zone_dtor;
2668 args.uminit = zero_init;
2671 args.align = UMA_SUPER_ALIGN - 1;
2672 args.flags = UMA_ZFLAG_INTERNAL;
2673 zone_ctor(zones, zsize, &args, M_WAITOK);
2675 /* Now make zones for slab headers */
2676 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2677 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2678 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2679 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2681 hashzone = uma_zcreate("UMA Hash",
2682 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2683 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2689 #ifndef UMA_MD_SMALL_ALLOC
2690 extern void vm_radix_reserve_kva(void);
2694 * Advertise the availability of normal kva allocations and switch to
2695 * the default back-end allocator. Marks the KVA we consumed on startup
2696 * as used in the map.
2702 if (bootstart != bootmem) {
2703 vm_map_lock(kernel_map);
2704 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2705 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2706 vm_map_unlock(kernel_map);
2709 #ifndef UMA_MD_SMALL_ALLOC
2710 /* Set up radix zone to use noobj_alloc. */
2711 vm_radix_reserve_kva();
2715 zone_foreach_unlocked(zone_kva_available, NULL);
2720 * Finish our initialization steps.
2727 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2728 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2729 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2731 zone_foreach_unlocked(zone_alloc_counters, NULL);
2732 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2733 callout_init(&uma_callout, 1);
2734 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2735 booted = BOOT_RUNNING;
2737 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2738 EVENTHANDLER_PRI_FIRST);
2745 booted = BOOT_SHUTDOWN;
2749 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2750 int align, uint32_t flags)
2752 struct uma_kctor_args args;
2755 args.uminit = uminit;
2757 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2760 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2763 /* Public functions */
2766 uma_set_align(int align)
2769 if (align != UMA_ALIGN_CACHE)
2770 uma_align_cache = align;
2775 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2776 uma_init uminit, uma_fini fini, int align, uint32_t flags)
2779 struct uma_zctor_args args;
2782 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2785 /* This stuff is essential for the zone ctor */
2786 memset(&args, 0, sizeof(args));
2791 args.uminit = uminit;
2795 * Inject procedures which check for memory use after free if we are
2796 * allowed to scramble the memory while it is not allocated. This
2797 * requires that: UMA is actually able to access the memory, no init
2798 * or fini procedures, no dependency on the initial value of the
2799 * memory, and no (legitimate) use of the memory after free. Note,
2800 * the ctor and dtor do not need to be empty.
2802 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
2803 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
2804 args.uminit = trash_init;
2805 args.fini = trash_fini;
2812 sx_slock(&uma_reclaim_lock);
2813 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
2814 sx_sunlock(&uma_reclaim_lock);
2821 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
2822 uma_init zinit, uma_fini zfini, uma_zone_t master)
2824 struct uma_zctor_args args;
2828 keg = master->uz_keg;
2829 memset(&args, 0, sizeof(args));
2831 args.size = keg->uk_size;
2834 args.uminit = zinit;
2836 args.align = keg->uk_align;
2837 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
2840 sx_slock(&uma_reclaim_lock);
2841 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
2842 sx_sunlock(&uma_reclaim_lock);
2849 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
2850 uma_init zinit, uma_fini zfini, uma_import zimport,
2851 uma_release zrelease, void *arg, int flags)
2853 struct uma_zctor_args args;
2855 memset(&args, 0, sizeof(args));
2860 args.uminit = zinit;
2862 args.import = zimport;
2863 args.release = zrelease;
2866 args.flags = flags | UMA_ZFLAG_CACHE;
2868 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
2873 uma_zdestroy(uma_zone_t zone)
2877 * Large slabs are expensive to reclaim, so don't bother doing
2878 * unnecessary work if we're shutting down.
2880 if (booted == BOOT_SHUTDOWN &&
2881 zone->uz_fini == NULL && zone->uz_release == zone_release)
2883 sx_slock(&uma_reclaim_lock);
2884 zone_free_item(zones, zone, NULL, SKIP_NONE);
2885 sx_sunlock(&uma_reclaim_lock);
2889 uma_zwait(uma_zone_t zone)
2893 item = uma_zalloc_arg(zone, NULL, M_WAITOK);
2894 uma_zfree(zone, item);
2898 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
2904 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
2906 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
2907 if (item != NULL && (flags & M_ZERO)) {
2909 for (i = 0; i <= mp_maxid; i++)
2910 bzero(zpcpu_get_cpu(item, i), zone->uz_size);
2912 bzero(item, zone->uz_size);
2919 * A stub while both regular and pcpu cases are identical.
2922 uma_zfree_pcpu_arg(uma_zone_t zone, void *item, void *udata)
2926 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
2928 uma_zfree_arg(zone, item, udata);
2931 static inline void *
2932 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
2938 skipdbg = uma_dbg_zskip(zone, item);
2939 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
2940 zone->uz_ctor != trash_ctor)
2941 trash_ctor(item, size, udata, flags);
2943 /* Check flags before loading ctor pointer. */
2944 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
2945 __predict_false(zone->uz_ctor != NULL) &&
2946 zone->uz_ctor(item, size, udata, flags) != 0) {
2947 counter_u64_add(zone->uz_fails, 1);
2948 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
2953 uma_dbg_alloc(zone, NULL, item);
2962 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
2963 enum zfreeskip skip)
2968 skipdbg = uma_dbg_zskip(zone, item);
2969 if (skip == SKIP_NONE && !skipdbg) {
2970 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
2971 uma_dbg_free(zone, udata, item);
2973 uma_dbg_free(zone, NULL, item);
2976 if (__predict_true(skip < SKIP_DTOR)) {
2977 if (zone->uz_dtor != NULL)
2978 zone->uz_dtor(item, size, udata);
2980 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
2981 zone->uz_dtor != trash_dtor)
2982 trash_dtor(item, size, udata);
2987 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
2988 #define UMA_ZALLOC_DEBUG
2990 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
2996 if (flags & M_WAITOK) {
2997 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2998 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3003 KASSERT((flags & M_EXEC) == 0,
3004 ("uma_zalloc_debug: called with M_EXEC"));
3005 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3006 ("uma_zalloc_debug: called within spinlock or critical section"));
3007 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3008 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3011 #ifdef DEBUG_MEMGUARD
3012 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3014 item = memguard_alloc(zone->uz_size, flags);
3016 error = EJUSTRETURN;
3017 if (zone->uz_init != NULL &&
3018 zone->uz_init(item, zone->uz_size, flags) != 0) {
3022 if (zone->uz_ctor != NULL &&
3023 zone->uz_ctor(item, zone->uz_size, udata,
3025 counter_u64_add(zone->uz_fails, 1);
3026 zone->uz_fini(item, zone->uz_size);
3033 /* This is unfortunate but should not be fatal. */
3040 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3042 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3043 ("uma_zfree_debug: called with spinlock or critical section held"));
3045 #ifdef DEBUG_MEMGUARD
3046 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3047 if (zone->uz_dtor != NULL)
3048 zone->uz_dtor(item, zone->uz_size, udata);
3049 if (zone->uz_fini != NULL)
3050 zone->uz_fini(item, zone->uz_size);
3051 memguard_free(item);
3052 return (EJUSTRETURN);
3059 static __noinline void *
3060 uma_zalloc_single(uma_zone_t zone, void *udata, int flags)
3065 * We can not get a bucket so try to return a single item.
3067 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3068 domain = PCPU_GET(domain);
3070 domain = UMA_ANYDOMAIN;
3071 return (zone_alloc_item(zone, udata, domain, flags));
3076 uma_zalloc_smr(uma_zone_t zone, int flags)
3078 uma_cache_bucket_t bucket;
3083 #ifdef UMA_ZALLOC_DEBUG
3084 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3085 ("uma_zalloc_arg: called with non-SMR zone.\n"));
3086 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3092 cache = &zone->uz_cpu[curcpu];
3093 bucket = &cache->uc_allocbucket;
3094 size = cache_uz_size(cache);
3095 uz_flags = cache_uz_flags(cache);
3096 if (__predict_true(bucket->ucb_cnt != 0)) {
3097 item = cache_bucket_pop(cache, bucket);
3099 return (item_ctor(zone, uz_flags, size, NULL, flags,
3102 } while (cache_alloc(zone, cache, NULL, flags));
3105 return (uma_zalloc_single(zone, NULL, flags));
3110 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3112 uma_cache_bucket_t bucket;
3117 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3118 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3120 /* This is the fast path allocation */
3121 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3124 #ifdef UMA_ZALLOC_DEBUG
3125 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3126 ("uma_zalloc_arg: called with SMR zone.\n"));
3127 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3132 * If possible, allocate from the per-CPU cache. There are two
3133 * requirements for safe access to the per-CPU cache: (1) the thread
3134 * accessing the cache must not be preempted or yield during access,
3135 * and (2) the thread must not migrate CPUs without switching which
3136 * cache it accesses. We rely on a critical section to prevent
3137 * preemption and migration. We release the critical section in
3138 * order to acquire the zone mutex if we are unable to allocate from
3139 * the current cache; when we re-acquire the critical section, we
3140 * must detect and handle migration if it has occurred.
3144 cache = &zone->uz_cpu[curcpu];
3145 bucket = &cache->uc_allocbucket;
3146 size = cache_uz_size(cache);
3147 uz_flags = cache_uz_flags(cache);
3148 if (__predict_true(bucket->ucb_cnt != 0)) {
3149 item = cache_bucket_pop(cache, bucket);
3151 return (item_ctor(zone, uz_flags, size, udata, flags,
3154 } while (cache_alloc(zone, cache, udata, flags));
3157 return (uma_zalloc_single(zone, udata, flags));
3161 * Replenish an alloc bucket and possibly restore an old one. Called in
3162 * a critical section. Returns in a critical section.
3164 * A false return value indicates an allocation failure.
3165 * A true return value indicates success and the caller should retry.
3167 static __noinline bool
3168 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3170 uma_zone_domain_t zdom;
3171 uma_bucket_t bucket;
3175 CRITICAL_ASSERT(curthread);
3178 * If we have run out of items in our alloc bucket see
3179 * if we can switch with the free bucket.
3181 * SMR Zones can't re-use the free bucket until the sequence has
3184 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 &&
3185 cache->uc_freebucket.ucb_cnt != 0) {
3186 cache_bucket_swap(&cache->uc_freebucket,
3187 &cache->uc_allocbucket);
3192 * Discard any empty allocation bucket while we hold no locks.
3194 bucket = cache_bucket_unload_alloc(cache);
3197 bucket_free(zone, bucket, udata);
3199 /* Short-circuit for zones without buckets and low memory. */
3200 if (zone->uz_bucket_size == 0 || bucketdisable) {
3206 * Attempt to retrieve the item from the per-CPU cache has failed, so
3207 * we must go back to the zone. This requires the zone lock, so we
3208 * must drop the critical section, then re-acquire it when we go back
3209 * to the cache. Since the critical section is released, we may be
3210 * preempted or migrate. As such, make sure not to maintain any
3211 * thread-local state specific to the cache from prior to releasing
3212 * the critical section.
3215 if (ZONE_TRYLOCK(zone) == 0) {
3216 /* Record contention to size the buckets. */
3221 /* See if we lost the race to fill the cache. */
3223 cache = &zone->uz_cpu[curcpu];
3224 if (cache->uc_allocbucket.ucb_bucket != NULL) {
3230 * Check the zone's cache of buckets.
3232 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) {
3233 domain = PCPU_GET(domain);
3234 zdom = &zone->uz_domain[domain];
3236 domain = UMA_ANYDOMAIN;
3237 zdom = &zone->uz_domain[0];
3240 if ((bucket = zone_fetch_bucket(zone, zdom)) != NULL) {
3241 KASSERT(bucket->ub_cnt != 0,
3242 ("uma_zalloc_arg: Returning an empty bucket."));
3243 cache_bucket_load_alloc(cache, bucket);
3246 /* We are no longer associated with this CPU. */
3250 * We bump the uz count when the cache size is insufficient to
3251 * handle the working set.
3253 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
3254 zone->uz_bucket_size++;
3258 * Fill a bucket and attempt to use it as the alloc bucket.
3260 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3261 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3262 zone->uz_name, zone, bucket);
3263 if (bucket == NULL) {
3269 * See if we lost the race or were migrated. Cache the
3270 * initialized bucket to make this less likely or claim
3271 * the memory directly.
3275 cache = &zone->uz_cpu[curcpu];
3276 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3277 ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0 ||
3278 domain == PCPU_GET(domain))) {
3279 cache_bucket_load_alloc(cache, bucket);
3280 zdom->uzd_imax += bucket->ub_cnt;
3281 } else if (zone->uz_bkt_count >= zone->uz_bkt_max) {
3284 bucket_drain(zone, bucket);
3285 bucket_free(zone, bucket, udata);
3289 zone_put_bucket(zone, zdom, bucket, false);
3295 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3298 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3299 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3301 /* This is the fast path allocation */
3302 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3303 zone->uz_name, zone, domain, flags);
3305 if (flags & M_WAITOK) {
3306 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3307 "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3309 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3310 ("uma_zalloc_domain: called with spinlock or critical section held"));
3312 return (zone_alloc_item(zone, udata, domain, flags));
3316 * Find a slab with some space. Prefer slabs that are partially used over those
3317 * that are totally full. This helps to reduce fragmentation.
3319 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3323 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3329 KASSERT(domain >= 0 && domain < vm_ndomains,
3330 ("keg_first_slab: domain %d out of range", domain));
3331 KEG_LOCK_ASSERT(keg, domain);
3336 dom = &keg->uk_domain[domain];
3337 if (!LIST_EMPTY(&dom->ud_part_slab))
3338 return (LIST_FIRST(&dom->ud_part_slab));
3339 if (!LIST_EMPTY(&dom->ud_free_slab)) {
3340 slab = LIST_FIRST(&dom->ud_free_slab);
3341 LIST_REMOVE(slab, us_link);
3342 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3346 domain = (domain + 1) % vm_ndomains;
3347 } while (domain != start);
3353 * Fetch an existing slab from a free or partial list. Returns with the
3354 * keg domain lock held if a slab was found or unlocked if not.
3357 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3362 /* HASH has a single free list. */
3363 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3366 KEG_LOCK(keg, domain);
3367 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3368 if (keg->uk_domain[domain].ud_free <= reserve ||
3369 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3370 KEG_UNLOCK(keg, domain);
3377 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3379 struct vm_domainset_iter di;
3386 * Use the keg's policy if upper layers haven't already specified a
3387 * domain (as happens with first-touch zones).
3389 * To avoid races we run the iterator with the keg lock held, but that
3390 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3391 * clear M_WAITOK and handle low memory conditions locally.
3393 rr = rdomain == UMA_ANYDOMAIN;
3395 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3396 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3404 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3409 * M_NOVM means don't ask at all!
3414 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3417 if (!rr && (flags & M_WAITOK) == 0)
3419 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3420 if ((flags & M_WAITOK) != 0) {
3421 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
3429 * We might not have been able to get a slab but another cpu
3430 * could have while we were unlocked. Check again before we
3433 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3440 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3446 KEG_LOCK_ASSERT(keg, slab->us_domain);
3448 dom = &keg->uk_domain[slab->us_domain];
3449 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3450 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3451 item = slab_item(slab, keg, freei);
3452 slab->us_freecount--;
3455 /* Move this slab to the full list */
3456 if (slab->us_freecount == 0) {
3457 LIST_REMOVE(slab, us_link);
3458 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3465 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3479 /* Try to keep the buckets totally full */
3480 for (i = 0; i < max; ) {
3481 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3484 stripe = howmany(max, vm_ndomains);
3486 dom = &keg->uk_domain[slab->us_domain];
3487 while (slab->us_freecount && i < max) {
3488 bucket[i++] = slab_alloc_item(keg, slab);
3489 if (dom->ud_free <= keg->uk_reserve)
3493 * If the zone is striped we pick a new slab for every
3494 * N allocations. Eliminating this conditional will
3495 * instead pick a new domain for each bucket rather
3496 * than stripe within each bucket. The current option
3497 * produces more fragmentation and requires more cpu
3498 * time but yields better distribution.
3500 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3501 vm_ndomains > 1 && --stripe == 0)
3505 KEG_UNLOCK(keg, slab->us_domain);
3506 /* Don't block if we allocated any successfully. */
3515 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3517 uint64_t old, new, total, max;
3520 * The hard case. We're going to sleep because there were existing
3521 * sleepers or because we ran out of items. This routine enforces
3522 * fairness by keeping fifo order.
3524 * First release our ill gotten gains and make some noise.
3527 zone_free_limit(zone, count);
3528 zone_log_warning(zone);
3529 zone_maxaction(zone);
3530 if (flags & M_NOWAIT)
3534 * We need to allocate an item or set ourself as a sleeper
3535 * while the sleepq lock is held to avoid wakeup races. This
3536 * is essentially a home rolled semaphore.
3538 sleepq_lock(&zone->uz_max_items);
3539 old = zone->uz_items;
3541 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3542 /* Cache the max since we will evaluate twice. */
3543 max = zone->uz_max_items;
3544 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3545 UZ_ITEMS_COUNT(old) >= max)
3546 new = old + UZ_ITEMS_SLEEPER;
3548 new = old + MIN(count, max - old);
3549 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3551 /* We may have successfully allocated under the sleepq lock. */
3552 if (UZ_ITEMS_SLEEPERS(new) == 0) {
3553 sleepq_release(&zone->uz_max_items);
3558 * This is in a different cacheline from uz_items so that we
3559 * don't constantly invalidate the fastpath cacheline when we
3560 * adjust item counts. This could be limited to toggling on
3563 atomic_add_32(&zone->uz_sleepers, 1);
3564 atomic_add_64(&zone->uz_sleeps, 1);
3567 * We have added ourselves as a sleeper. The sleepq lock
3568 * protects us from wakeup races. Sleep now and then retry.
3570 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3571 sleepq_wait(&zone->uz_max_items, PVM);
3574 * After wakeup, remove ourselves as a sleeper and try
3575 * again. We no longer have the sleepq lock for protection.
3577 * Subract ourselves as a sleeper while attempting to add
3580 atomic_subtract_32(&zone->uz_sleepers, 1);
3581 old = atomic_fetchadd_64(&zone->uz_items,
3582 -(UZ_ITEMS_SLEEPER - count));
3583 /* We're no longer a sleeper. */
3584 old -= UZ_ITEMS_SLEEPER;
3587 * If we're still at the limit, restart. Notably do not
3588 * block on other sleepers. Cache the max value to protect
3589 * against changes via sysctl.
3591 total = UZ_ITEMS_COUNT(old);
3592 max = zone->uz_max_items;
3595 /* Truncate if necessary, otherwise wake other sleepers. */
3596 if (total + count > max) {
3597 zone_free_limit(zone, total + count - max);
3598 count = max - total;
3599 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3600 wakeup_one(&zone->uz_max_items);
3607 * Allocate 'count' items from our max_items limit. Returns the number
3608 * available. If M_NOWAIT is not specified it will sleep until at least
3609 * one item can be allocated.
3612 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3617 max = zone->uz_max_items;
3621 * We expect normal allocations to succeed with a simple
3624 old = atomic_fetchadd_64(&zone->uz_items, count);
3625 if (__predict_true(old + count <= max))
3629 * If we had some items and no sleepers just return the
3630 * truncated value. We have to release the excess space
3631 * though because that may wake sleepers who weren't woken
3632 * because we were temporarily over the limit.
3635 zone_free_limit(zone, (old + count) - max);
3638 return (zone_alloc_limit_hard(zone, count, flags));
3642 * Free a number of items back to the limit.
3645 zone_free_limit(uma_zone_t zone, int count)
3652 * In the common case we either have no sleepers or
3653 * are still over the limit and can just return.
3655 old = atomic_fetchadd_64(&zone->uz_items, -count);
3656 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3657 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3661 * Moderate the rate of wakeups. Sleepers will continue
3662 * to generate wakeups if necessary.
3664 wakeup_one(&zone->uz_max_items);
3668 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3670 uma_bucket_t bucket;
3673 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3676 /* Avoid allocs targeting empty domains. */
3677 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3678 domain = UMA_ANYDOMAIN;
3680 if (zone->uz_max_items > 0)
3681 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3684 maxbucket = zone->uz_bucket_size;
3688 /* Don't wait for buckets, preserve caller's NOVM setting. */
3689 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3690 if (bucket == NULL) {
3695 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3696 MIN(maxbucket, bucket->ub_entries), domain, flags);
3699 * Initialize the memory if necessary.
3701 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3704 for (i = 0; i < bucket->ub_cnt; i++)
3705 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3709 * If we couldn't initialize the whole bucket, put the
3710 * rest back onto the freelist.
3712 if (i != bucket->ub_cnt) {
3713 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3714 bucket->ub_cnt - i);
3716 bzero(&bucket->ub_bucket[i],
3717 sizeof(void *) * (bucket->ub_cnt - i));
3723 cnt = bucket->ub_cnt;
3724 if (bucket->ub_cnt == 0) {
3725 bucket_free(zone, bucket, udata);
3726 counter_u64_add(zone->uz_fails, 1);
3730 if (zone->uz_max_items > 0 && cnt < maxbucket)
3731 zone_free_limit(zone, maxbucket - cnt);
3737 * Allocates a single item from a zone.
3740 * zone The zone to alloc for.
3741 * udata The data to be passed to the constructor.
3742 * domain The domain to allocate from or UMA_ANYDOMAIN.
3743 * flags M_WAITOK, M_NOWAIT, M_ZERO.
3746 * NULL if there is no memory and M_NOWAIT is set
3747 * An item if successful
3751 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
3755 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0)
3758 /* Avoid allocs targeting empty domains. */
3759 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3760 domain = UMA_ANYDOMAIN;
3762 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
3766 * We have to call both the zone's init (not the keg's init)
3767 * and the zone's ctor. This is because the item is going from
3768 * a keg slab directly to the user, and the user is expecting it
3769 * to be both zone-init'd as well as zone-ctor'd.
3771 if (zone->uz_init != NULL) {
3772 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
3773 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
3777 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
3782 counter_u64_add(zone->uz_allocs, 1);
3783 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
3784 zone->uz_name, zone);
3789 counter_u64_add(zone->uz_fails, 1);
3791 if (zone->uz_max_items > 0)
3792 zone_free_limit(zone, 1);
3793 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
3794 zone->uz_name, zone);
3801 uma_zfree_smr(uma_zone_t zone, void *item)
3804 uma_cache_bucket_t bucket;
3805 int domain, itemdomain, uz_flags;
3807 #ifdef UMA_ZALLOC_DEBUG
3808 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3809 ("uma_zfree_smr: called with non-SMR zone.\n"));
3810 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
3811 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
3814 cache = &zone->uz_cpu[curcpu];
3815 uz_flags = cache_uz_flags(cache);
3816 domain = itemdomain = 0;
3818 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
3819 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
3823 cache = &zone->uz_cpu[curcpu];
3824 /* SMR Zones must free to the free bucket. */
3825 bucket = &cache->uc_freebucket;
3827 domain = PCPU_GET(domain);
3828 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
3829 domain != itemdomain) {
3830 bucket = &cache->uc_crossbucket;
3833 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
3834 cache_bucket_push(cache, bucket, item);
3838 } while (cache_free(zone, cache, NULL, item, itemdomain));
3842 * If nothing else caught this, we'll just do an internal free.
3844 zone_free_item(zone, item, NULL, SKIP_NONE);
3849 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
3852 uma_cache_bucket_t bucket;
3853 int domain, itemdomain, uz_flags;
3855 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3856 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3858 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
3860 #ifdef UMA_ZALLOC_DEBUG
3861 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3862 ("uma_zfree_arg: called with SMR zone.\n"));
3863 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
3866 /* uma_zfree(..., NULL) does nothing, to match free(9). */
3871 * We are accessing the per-cpu cache without a critical section to
3872 * fetch size and flags. This is acceptable, if we are preempted we
3873 * will simply read another cpu's line.
3875 cache = &zone->uz_cpu[curcpu];
3876 uz_flags = cache_uz_flags(cache);
3877 if (UMA_ALWAYS_CTORDTOR ||
3878 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
3879 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
3882 * The race here is acceptable. If we miss it we'll just have to wait
3883 * a little longer for the limits to be reset.
3885 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
3886 if (zone->uz_sleepers > 0)
3891 * If possible, free to the per-CPU cache. There are two
3892 * requirements for safe access to the per-CPU cache: (1) the thread
3893 * accessing the cache must not be preempted or yield during access,
3894 * and (2) the thread must not migrate CPUs without switching which
3895 * cache it accesses. We rely on a critical section to prevent
3896 * preemption and migration. We release the critical section in
3897 * order to acquire the zone mutex if we are unable to free to the
3898 * current cache; when we re-acquire the critical section, we must
3899 * detect and handle migration if it has occurred.
3901 domain = itemdomain = 0;
3903 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
3904 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
3908 cache = &zone->uz_cpu[curcpu];
3910 * Try to free into the allocbucket first to give LIFO
3911 * ordering for cache-hot datastructures. Spill over
3912 * into the freebucket if necessary. Alloc will swap
3913 * them if one runs dry.
3915 bucket = &cache->uc_allocbucket;
3917 domain = PCPU_GET(domain);
3918 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
3919 domain != itemdomain) {
3920 bucket = &cache->uc_crossbucket;
3923 if (bucket->ucb_cnt >= bucket->ucb_entries)
3924 bucket = &cache->uc_freebucket;
3925 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
3926 cache_bucket_push(cache, bucket, item);
3930 } while (cache_free(zone, cache, udata, item, itemdomain));
3934 * If nothing else caught this, we'll just do an internal free.
3937 zone_free_item(zone, item, udata, SKIP_DTOR);
3942 * sort crossdomain free buckets to domain correct buckets and cache
3946 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
3948 struct uma_bucketlist fullbuckets;
3949 uma_zone_domain_t zdom;
3955 "uma_zfree: zone %s(%p) draining cross bucket %p",
3956 zone->uz_name, zone, bucket);
3958 STAILQ_INIT(&fullbuckets);
3961 * To avoid having ndomain * ndomain buckets for sorting we have a
3962 * lock on the current crossfree bucket. A full matrix with
3963 * per-domain locking could be used if necessary.
3965 ZONE_CROSS_LOCK(zone);
3966 while (bucket->ub_cnt > 0) {
3967 item = bucket->ub_bucket[bucket->ub_cnt - 1];
3968 domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
3969 zdom = &zone->uz_domain[domain];
3970 if (zdom->uzd_cross == NULL) {
3971 zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
3972 if (zdom->uzd_cross == NULL)
3975 zdom->uzd_cross->ub_bucket[zdom->uzd_cross->ub_cnt++] = item;
3976 if (zdom->uzd_cross->ub_cnt == zdom->uzd_cross->ub_entries) {
3977 STAILQ_INSERT_HEAD(&fullbuckets, zdom->uzd_cross,
3979 zdom->uzd_cross = NULL;
3983 ZONE_CROSS_UNLOCK(zone);
3984 if (!STAILQ_EMPTY(&fullbuckets)) {
3986 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
3987 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3988 bucket->ub_seq = smr_current(zone->uz_smr);
3989 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
3990 if (zone->uz_bkt_count >= zone->uz_bkt_max) {
3992 bucket_drain(zone, b);
3993 bucket_free(zone, b, udata);
3996 domain = _vm_phys_domain(
3998 (vm_offset_t)b->ub_bucket[0]));
3999 zdom = &zone->uz_domain[domain];
4000 zone_put_bucket(zone, zdom, b, true);
4005 if (bucket->ub_cnt != 0)
4006 bucket_drain(zone, bucket);
4007 bucket->ub_seq = SMR_SEQ_INVALID;
4008 bucket_free(zone, bucket, udata);
4013 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4014 int domain, int itemdomain)
4016 uma_zone_domain_t zdom;
4020 * Buckets coming from the wrong domain will be entirely for the
4021 * only other domain on two domain systems. In this case we can
4022 * simply cache them. Otherwise we need to sort them back to
4025 if (domain != itemdomain && vm_ndomains > 2) {
4026 zone_free_cross(zone, bucket, udata);
4032 * Attempt to save the bucket in the zone's domain bucket cache.
4034 * We bump the uz count when the cache size is insufficient to
4035 * handle the working set.
4037 if (ZONE_TRYLOCK(zone) == 0) {
4038 /* Record contention to size the buckets. */
4040 if (zone->uz_bucket_size < zone->uz_bucket_size_max)
4041 zone->uz_bucket_size++;
4045 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4046 zone->uz_name, zone, bucket);
4047 /* ub_cnt is pointing to the last free item */
4048 KASSERT(bucket->ub_cnt == bucket->ub_entries,
4049 ("uma_zfree: Attempting to insert partial bucket onto the full list.\n"));
4050 if (zone->uz_bkt_count >= zone->uz_bkt_max) {
4052 bucket_drain(zone, bucket);
4053 bucket_free(zone, bucket, udata);
4055 zdom = &zone->uz_domain[itemdomain];
4056 zone_put_bucket(zone, zdom, bucket, true);
4062 * Populate a free or cross bucket for the current cpu cache. Free any
4063 * existing full bucket either to the zone cache or back to the slab layer.
4065 * Enters and returns in a critical section. false return indicates that
4066 * we can not satisfy this free in the cache layer. true indicates that
4067 * the caller should retry.
4069 static __noinline bool
4070 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4073 uma_cache_bucket_t cbucket;
4074 uma_bucket_t newbucket, bucket;
4077 CRITICAL_ASSERT(curthread);
4079 if (zone->uz_bucket_size == 0)
4082 cache = &zone->uz_cpu[curcpu];
4086 * FIRSTTOUCH domains need to free to the correct zdom. When
4087 * enabled this is the zdom of the item. The bucket is the
4088 * cross bucket if the current domain and itemdomain do not match.
4090 cbucket = &cache->uc_freebucket;
4092 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) {
4093 domain = PCPU_GET(domain);
4094 if (domain != itemdomain) {
4095 cbucket = &cache->uc_crossbucket;
4096 if (cbucket->ucb_cnt != 0)
4097 atomic_add_64(&zone->uz_xdomain,
4102 itemdomain = domain = 0;
4103 bucket = cache_bucket_unload(cbucket);
4105 /* We are no longer associated with this CPU. */
4109 * Don't let SMR zones operate without a free bucket. Force
4110 * a synchronize and re-use this one. We will only degrade
4111 * to a synchronize every bucket_size items rather than every
4112 * item if we fail to allocate a bucket.
4114 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4116 bucket->ub_seq = smr_advance(zone->uz_smr);
4117 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4118 if (newbucket == NULL && bucket != NULL) {
4119 bucket_drain(zone, bucket);
4123 } else if (!bucketdisable)
4124 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4127 zone_free_bucket(zone, bucket, udata, domain, itemdomain);
4130 if ((bucket = newbucket) == NULL)
4132 cache = &zone->uz_cpu[curcpu];
4135 * Check to see if we should be populating the cross bucket. If it
4136 * is already populated we will fall through and attempt to populate
4139 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) {
4140 domain = PCPU_GET(domain);
4141 if (domain != itemdomain &&
4142 cache->uc_crossbucket.ucb_bucket == NULL) {
4143 cache_bucket_load_cross(cache, bucket);
4149 * We may have lost the race to fill the bucket or switched CPUs.
4151 if (cache->uc_freebucket.ucb_bucket != NULL) {
4153 bucket_free(zone, bucket, udata);
4156 cache_bucket_load_free(cache, bucket);
4162 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
4165 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4166 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4168 CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone);
4170 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
4171 ("uma_zfree_domain: called with spinlock or critical section held"));
4173 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4176 zone_free_item(zone, item, udata, SKIP_NONE);
4180 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4187 KEG_LOCK_ASSERT(keg, slab->us_domain);
4189 /* Do we need to remove from any lists? */
4190 dom = &keg->uk_domain[slab->us_domain];
4191 if (slab->us_freecount+1 == keg->uk_ipers) {
4192 LIST_REMOVE(slab, us_link);
4193 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4194 } else if (slab->us_freecount == 0) {
4195 LIST_REMOVE(slab, us_link);
4196 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4199 /* Slab management. */
4200 freei = slab_item_index(slab, keg, item);
4201 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4202 slab->us_freecount++;
4204 /* Keg statistics. */
4209 zone_release(void *arg, void **bucket, int cnt)
4222 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4223 lock = KEG_LOCK(keg, 0);
4224 for (i = 0; i < cnt; i++) {
4226 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4227 slab = vtoslab((vm_offset_t)item);
4229 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4230 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4231 slab = hash_sfind(&keg->uk_hash, mem);
4233 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4235 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4238 lock = KEG_LOCK(keg, slab->us_domain);
4240 slab_free_item(zone, slab, item);
4247 * Frees a single item to any zone.
4250 * zone The zone to free to
4251 * item The item we're freeing
4252 * udata User supplied data for the dtor
4253 * skip Skip dtors and finis
4256 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4260 * If a free is sent directly to an SMR zone we have to
4261 * synchronize immediately because the item can instantly
4262 * be reallocated. This should only happen in degenerate
4263 * cases when no memory is available for per-cpu caches.
4265 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4266 smr_synchronize(zone->uz_smr);
4268 item_dtor(zone, item, zone->uz_size, udata, skip);
4270 if (skip < SKIP_FINI && zone->uz_fini)
4271 zone->uz_fini(item, zone->uz_size);
4273 zone->uz_release(zone->uz_arg, &item, 1);
4275 if (skip & SKIP_CNT)
4278 counter_u64_add(zone->uz_frees, 1);
4280 if (zone->uz_max_items > 0)
4281 zone_free_limit(zone, 1);
4286 uma_zone_set_max(uma_zone_t zone, int nitems)
4288 struct uma_bucket_zone *ubz;
4292 * XXX This can misbehave if the zone has any allocations with
4293 * no limit and a limit is imposed. There is currently no
4294 * way to clear a limit.
4297 ubz = bucket_zone_max(zone, nitems);
4298 count = ubz != NULL ? ubz->ubz_entries : 0;
4299 zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4300 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4301 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4302 zone->uz_max_items = nitems;
4303 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4304 zone_update_caches(zone);
4305 /* We may need to wake waiters. */
4306 wakeup(&zone->uz_max_items);
4314 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4316 struct uma_bucket_zone *ubz;
4320 ubz = bucket_zone_max(zone, nitems);
4323 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4324 /* Count the cross-domain bucket. */
4326 nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4327 zone->uz_bucket_size_max = ubz->ubz_entries;
4329 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4331 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4332 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4333 zone->uz_bkt_max = nitems;
4339 uma_zone_get_max(uma_zone_t zone)
4343 nitems = atomic_load_64(&zone->uz_max_items);
4350 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4353 ZONE_ASSERT_COLD(zone);
4354 zone->uz_warning = warning;
4359 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4362 ZONE_ASSERT_COLD(zone);
4363 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4368 uma_zone_get_cur(uma_zone_t zone)
4374 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4375 nitems = counter_u64_fetch(zone->uz_allocs) -
4376 counter_u64_fetch(zone->uz_frees);
4378 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4379 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4381 return (nitems < 0 ? 0 : nitems);
4385 uma_zone_get_allocs(uma_zone_t zone)
4391 if (zone->uz_allocs != EARLY_COUNTER)
4392 nitems = counter_u64_fetch(zone->uz_allocs);
4394 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4400 uma_zone_get_frees(uma_zone_t zone)
4406 if (zone->uz_frees != EARLY_COUNTER)
4407 nitems = counter_u64_fetch(zone->uz_frees);
4409 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4415 /* Used only for KEG_ASSERT_COLD(). */
4417 uma_keg_get_allocs(uma_keg_t keg)
4423 LIST_FOREACH(z, &keg->uk_zones, uz_link)
4424 nitems += uma_zone_get_allocs(z);
4432 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4437 KEG_ASSERT_COLD(keg);
4438 keg->uk_init = uminit;
4443 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4448 KEG_ASSERT_COLD(keg);
4449 keg->uk_fini = fini;
4454 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4457 ZONE_ASSERT_COLD(zone);
4458 zone->uz_init = zinit;
4463 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4466 ZONE_ASSERT_COLD(zone);
4467 zone->uz_fini = zfini;
4472 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4477 KEG_ASSERT_COLD(keg);
4478 keg->uk_freef = freef;
4483 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4488 KEG_ASSERT_COLD(keg);
4489 keg->uk_allocf = allocf;
4494 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4497 ZONE_ASSERT_COLD(zone);
4499 zone->uz_flags |= UMA_ZONE_SMR;
4501 zone_update_caches(zone);
4505 uma_zone_get_smr(uma_zone_t zone)
4508 return (zone->uz_smr);
4513 uma_zone_reserve(uma_zone_t zone, int items)
4518 KEG_ASSERT_COLD(keg);
4519 keg->uk_reserve = items;
4524 uma_zone_reserve_kva(uma_zone_t zone, int count)
4531 KEG_ASSERT_COLD(keg);
4532 ZONE_ASSERT_COLD(zone);
4534 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4536 #ifdef UMA_MD_SMALL_ALLOC
4537 if (keg->uk_ppera > 1) {
4541 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4548 MPASS(keg->uk_kva == 0);
4551 zone->uz_max_items = pages * keg->uk_ipers;
4552 #ifdef UMA_MD_SMALL_ALLOC
4553 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4555 keg->uk_allocf = noobj_alloc;
4557 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4558 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4559 zone_update_caches(zone);
4567 uma_prealloc(uma_zone_t zone, int items)
4569 struct vm_domainset_iter di;
4573 int aflags, domain, slabs;
4576 slabs = howmany(items, keg->uk_ipers);
4577 while (slabs-- > 0) {
4579 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4582 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4585 dom = &keg->uk_domain[slab->us_domain];
4586 LIST_REMOVE(slab, us_link);
4587 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4589 KEG_UNLOCK(keg, slab->us_domain);
4592 if (vm_domainset_iter_policy(&di, &domain) != 0)
4593 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
4600 uma_reclaim(int req)
4603 CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4604 sx_xlock(&uma_reclaim_lock);
4608 case UMA_RECLAIM_TRIM:
4609 zone_foreach(zone_trim, NULL);
4611 case UMA_RECLAIM_DRAIN:
4612 case UMA_RECLAIM_DRAIN_CPU:
4613 zone_foreach(zone_drain, NULL);
4614 if (req == UMA_RECLAIM_DRAIN_CPU) {
4615 pcpu_cache_drain_safe(NULL);
4616 zone_foreach(zone_drain, NULL);
4620 panic("unhandled reclamation request %d", req);
4624 * Some slabs may have been freed but this zone will be visited early
4625 * we visit again so that we can free pages that are empty once other
4626 * zones are drained. We have to do the same for buckets.
4628 zone_drain(slabzones[0], NULL);
4629 zone_drain(slabzones[1], NULL);
4630 bucket_zone_drain();
4631 sx_xunlock(&uma_reclaim_lock);
4634 static volatile int uma_reclaim_needed;
4637 uma_reclaim_wakeup(void)
4640 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4641 wakeup(uma_reclaim);
4645 uma_reclaim_worker(void *arg __unused)
4649 sx_xlock(&uma_reclaim_lock);
4650 while (atomic_load_int(&uma_reclaim_needed) == 0)
4651 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4653 sx_xunlock(&uma_reclaim_lock);
4654 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4655 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4656 atomic_store_int(&uma_reclaim_needed, 0);
4657 /* Don't fire more than once per-second. */
4658 pause("umarclslp", hz);
4664 uma_zone_reclaim(uma_zone_t zone, int req)
4668 case UMA_RECLAIM_TRIM:
4669 zone_trim(zone, NULL);
4671 case UMA_RECLAIM_DRAIN:
4672 zone_drain(zone, NULL);
4674 case UMA_RECLAIM_DRAIN_CPU:
4675 pcpu_cache_drain_safe(zone);
4676 zone_drain(zone, NULL);
4679 panic("unhandled reclamation request %d", req);
4685 uma_zone_exhausted(uma_zone_t zone)
4688 return (atomic_load_32(&zone->uz_sleepers) > 0);
4695 return (uma_kmem_limit);
4699 uma_set_limit(unsigned long limit)
4702 uma_kmem_limit = limit;
4709 return (atomic_load_long(&uma_kmem_total));
4716 return (uma_kmem_limit - uma_size());
4721 * Generate statistics across both the zone and its per-cpu cache's. Return
4722 * desired statistics if the pointer is non-NULL for that statistic.
4724 * Note: does not update the zone statistics, as it can't safely clear the
4725 * per-CPU cache statistic.
4729 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4730 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4733 uint64_t allocs, frees, sleeps, xdomain;
4736 allocs = frees = sleeps = xdomain = 0;
4739 cache = &z->uz_cpu[cpu];
4740 cachefree += cache->uc_allocbucket.ucb_cnt;
4741 cachefree += cache->uc_freebucket.ucb_cnt;
4742 xdomain += cache->uc_crossbucket.ucb_cnt;
4743 cachefree += cache->uc_crossbucket.ucb_cnt;
4744 allocs += cache->uc_allocs;
4745 frees += cache->uc_frees;
4747 allocs += counter_u64_fetch(z->uz_allocs);
4748 frees += counter_u64_fetch(z->uz_frees);
4749 sleeps += z->uz_sleeps;
4750 xdomain += z->uz_xdomain;
4751 if (cachefreep != NULL)
4752 *cachefreep = cachefree;
4753 if (allocsp != NULL)
4757 if (sleepsp != NULL)
4759 if (xdomainp != NULL)
4760 *xdomainp = xdomain;
4765 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
4772 rw_rlock(&uma_rwlock);
4773 LIST_FOREACH(kz, &uma_kegs, uk_link) {
4774 LIST_FOREACH(z, &kz->uk_zones, uz_link)
4777 LIST_FOREACH(z, &uma_cachezones, uz_link)
4780 rw_runlock(&uma_rwlock);
4781 return (sysctl_handle_int(oidp, &count, 0, req));
4785 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
4786 struct uma_percpu_stat *ups, bool internal)
4788 uma_zone_domain_t zdom;
4793 for (i = 0; i < vm_ndomains; i++) {
4794 zdom = &z->uz_domain[i];
4795 uth->uth_zone_free += zdom->uzd_nitems;
4797 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
4798 uth->uth_frees = counter_u64_fetch(z->uz_frees);
4799 uth->uth_fails = counter_u64_fetch(z->uz_fails);
4800 uth->uth_sleeps = z->uz_sleeps;
4801 uth->uth_xdomain = z->uz_xdomain;
4804 * While it is not normally safe to access the cache bucket pointers
4805 * while not on the CPU that owns the cache, we only allow the pointers
4806 * to be exchanged without the zone lock held, not invalidated, so
4807 * accept the possible race associated with bucket exchange during
4808 * monitoring. Use atomic_load_ptr() to ensure that the bucket pointers
4809 * are loaded only once.
4811 for (i = 0; i < mp_maxid + 1; i++) {
4812 bzero(&ups[i], sizeof(*ups));
4813 if (internal || CPU_ABSENT(i))
4815 cache = &z->uz_cpu[i];
4816 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
4817 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
4818 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
4819 ups[i].ups_allocs = cache->uc_allocs;
4820 ups[i].ups_frees = cache->uc_frees;
4825 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
4827 struct uma_stream_header ush;
4828 struct uma_type_header uth;
4829 struct uma_percpu_stat *ups;
4834 uint32_t kfree, pages;
4835 int count, error, i;
4837 error = sysctl_wire_old_buffer(req, 0);
4840 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
4841 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
4842 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
4845 rw_rlock(&uma_rwlock);
4846 LIST_FOREACH(kz, &uma_kegs, uk_link) {
4847 LIST_FOREACH(z, &kz->uk_zones, uz_link)
4851 LIST_FOREACH(z, &uma_cachezones, uz_link)
4855 * Insert stream header.
4857 bzero(&ush, sizeof(ush));
4858 ush.ush_version = UMA_STREAM_VERSION;
4859 ush.ush_maxcpus = (mp_maxid + 1);
4860 ush.ush_count = count;
4861 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
4863 LIST_FOREACH(kz, &uma_kegs, uk_link) {
4865 for (i = 0; i < vm_ndomains; i++) {
4866 kfree += kz->uk_domain[i].ud_free;
4867 pages += kz->uk_domain[i].ud_pages;
4869 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
4870 bzero(&uth, sizeof(uth));
4872 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
4873 uth.uth_align = kz->uk_align;
4874 uth.uth_size = kz->uk_size;
4875 uth.uth_rsize = kz->uk_rsize;
4876 if (z->uz_max_items > 0) {
4877 items = UZ_ITEMS_COUNT(z->uz_items);
4878 uth.uth_pages = (items / kz->uk_ipers) *
4881 uth.uth_pages = pages;
4882 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
4884 uth.uth_limit = z->uz_max_items;
4885 uth.uth_keg_free = kfree;
4888 * A zone is secondary is it is not the first entry
4889 * on the keg's zone list.
4891 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
4892 (LIST_FIRST(&kz->uk_zones) != z))
4893 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
4894 uma_vm_zone_stats(&uth, z, &sbuf, ups,
4895 kz->uk_flags & UMA_ZFLAG_INTERNAL);
4897 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
4898 for (i = 0; i < mp_maxid + 1; i++)
4899 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
4902 LIST_FOREACH(z, &uma_cachezones, uz_link) {
4903 bzero(&uth, sizeof(uth));
4905 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
4906 uth.uth_size = z->uz_size;
4907 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
4909 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
4910 for (i = 0; i < mp_maxid + 1; i++)
4911 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
4914 rw_runlock(&uma_rwlock);
4915 error = sbuf_finish(&sbuf);
4922 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
4924 uma_zone_t zone = *(uma_zone_t *)arg1;
4927 max = uma_zone_get_max(zone);
4928 error = sysctl_handle_int(oidp, &max, 0, req);
4929 if (error || !req->newptr)
4932 uma_zone_set_max(zone, max);
4938 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
4944 * Some callers want to add sysctls for global zones that
4945 * may not yet exist so they pass a pointer to a pointer.
4948 zone = *(uma_zone_t *)arg1;
4951 cur = uma_zone_get_cur(zone);
4952 return (sysctl_handle_int(oidp, &cur, 0, req));
4956 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
4958 uma_zone_t zone = arg1;
4961 cur = uma_zone_get_allocs(zone);
4962 return (sysctl_handle_64(oidp, &cur, 0, req));
4966 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
4968 uma_zone_t zone = arg1;
4971 cur = uma_zone_get_frees(zone);
4972 return (sysctl_handle_64(oidp, &cur, 0, req));
4976 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
4979 uma_zone_t zone = arg1;
4982 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
4983 if (zone->uz_flags != 0)
4984 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
4986 sbuf_printf(&sbuf, "0");
4987 error = sbuf_finish(&sbuf);
4994 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
4996 uma_keg_t keg = arg1;
4997 int avail, effpct, total;
4999 total = keg->uk_ppera * PAGE_SIZE;
5000 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5001 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5003 * We consider the client's requested size and alignment here, not the
5004 * real size determination uk_rsize, because we also adjust the real
5005 * size for internal implementation reasons (max bitset size).
5007 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5008 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5009 avail *= mp_maxid + 1;
5010 effpct = 100 * avail / total;
5011 return (sysctl_handle_int(oidp, &effpct, 0, req));
5015 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5017 uma_zone_t zone = arg1;
5020 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5021 return (sysctl_handle_64(oidp, &cur, 0, req));
5026 uma_dbg_getslab(uma_zone_t zone, void *item)
5033 * It is safe to return the slab here even though the
5034 * zone is unlocked because the item's allocation state
5035 * essentially holds a reference.
5037 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5038 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5040 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5041 return (vtoslab((vm_offset_t)mem));
5043 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5044 return ((uma_slab_t)(mem + keg->uk_pgoff));
5046 slab = hash_sfind(&keg->uk_hash, mem);
5053 uma_dbg_zskip(uma_zone_t zone, void *mem)
5056 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5059 return (uma_dbg_kskip(zone->uz_keg, mem));
5063 uma_dbg_kskip(uma_keg_t keg, void *mem)
5067 if (dbg_divisor == 0)
5070 if (dbg_divisor == 1)
5073 idx = (uintptr_t)mem >> PAGE_SHIFT;
5074 if (keg->uk_ipers > 1) {
5075 idx *= keg->uk_ipers;
5076 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5079 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5080 counter_u64_add(uma_skip_cnt, 1);
5083 counter_u64_add(uma_dbg_cnt, 1);
5089 * Set up the slab's freei data such that uma_dbg_free can function.
5093 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5099 slab = uma_dbg_getslab(zone, item);
5101 panic("uma: item %p did not belong to zone %s\n",
5102 item, zone->uz_name);
5105 freei = slab_item_index(slab, keg, item);
5107 if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5108 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
5109 item, zone, zone->uz_name, slab, freei);
5110 BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5114 * Verifies freed addresses. Checks for alignment, valid slab membership
5115 * and duplicate frees.
5119 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5125 slab = uma_dbg_getslab(zone, item);
5127 panic("uma: Freed item %p did not belong to zone %s\n",
5128 item, zone->uz_name);
5131 freei = slab_item_index(slab, keg, item);
5133 if (freei >= keg->uk_ipers)
5134 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
5135 item, zone, zone->uz_name, slab, freei);
5137 if (slab_item(slab, keg, freei) != item)
5138 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
5139 item, zone, zone->uz_name, slab, freei);
5141 if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5142 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
5143 item, zone, zone->uz_name, slab, freei);
5145 BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5147 #endif /* INVARIANTS */
5151 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5152 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5157 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5158 *allocs = counter_u64_fetch(z->uz_allocs);
5159 frees = counter_u64_fetch(z->uz_frees);
5160 *sleeps = z->uz_sleeps;
5164 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5166 for (i = 0; i < vm_ndomains; i++) {
5167 *cachefree += z->uz_domain[i].uzd_nitems;
5168 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5169 (LIST_FIRST(&kz->uk_zones) != z)))
5170 *cachefree += kz->uk_domain[i].ud_free;
5172 *used = *allocs - frees;
5173 return (((int64_t)*used + *cachefree) * kz->uk_size);
5176 DB_SHOW_COMMAND(uma, db_show_uma)
5178 const char *fmt_hdr, *fmt_entry;
5181 uint64_t allocs, used, sleeps, xdomain;
5183 /* variables for sorting */
5185 uma_zone_t cur_zone, last_zone;
5186 int64_t cur_size, last_size, size;
5189 /* /i option produces machine-parseable CSV output */
5190 if (modif[0] == 'i') {
5191 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5192 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5194 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5195 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5198 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5199 "Sleeps", "Bucket", "Total Mem", "XFree");
5201 /* Sort the zones with largest size first. */
5203 last_size = INT64_MAX;
5208 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5209 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5211 * In the case of size ties, print out zones
5212 * in the order they are encountered. That is,
5213 * when we encounter the most recently output
5214 * zone, we have already printed all preceding
5215 * ties, and we must print all following ties.
5217 if (z == last_zone) {
5221 size = get_uma_stats(kz, z, &allocs, &used,
5222 &sleeps, &cachefree, &xdomain);
5223 if (size > cur_size && size < last_size + ties)
5231 if (cur_zone == NULL)
5234 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5235 &sleeps, &cachefree, &xdomain);
5236 db_printf(fmt_entry, cur_zone->uz_name,
5237 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5238 (uintmax_t)allocs, (uintmax_t)sleeps,
5239 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5244 last_zone = cur_zone;
5245 last_size = cur_size;
5249 DB_SHOW_COMMAND(umacache, db_show_umacache)
5252 uint64_t allocs, frees;
5256 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5257 "Requests", "Bucket");
5258 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5259 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5260 for (i = 0; i < vm_ndomains; i++)
5261 cachefree += z->uz_domain[i].uzd_nitems;
5262 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5263 z->uz_name, (uintmax_t)z->uz_size,
5264 (intmax_t)(allocs - frees), cachefree,
5265 (uintmax_t)allocs, z->uz_bucket_size);