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_param.h>
86 #include <vm/vm_domainset.h>
87 #include <vm/vm_object.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_pageout.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>
95 #include <vm/vm_dumpset.h>
97 #include <vm/uma_int.h>
98 #include <vm/uma_dbg.h>
102 #ifdef DEBUG_MEMGUARD
103 #include <vm/memguard.h>
106 #include <machine/md_var.h>
109 #define UMA_ALWAYS_CTORDTOR 1
111 #define UMA_ALWAYS_CTORDTOR 0
115 * This is the zone and keg from which all zones are spawned.
117 static uma_zone_t kegs;
118 static uma_zone_t zones;
121 * On INVARIANTS builds, the slab contains a second bitset of the same size,
122 * "dbg_bits", which is laid out immediately after us_free.
125 #define SLAB_BITSETS 2
127 #define SLAB_BITSETS 1
131 * These are the two zones from which all offpage uma_slab_ts are allocated.
133 * One zone is for slab headers that can represent a larger number of items,
134 * making the slabs themselves more efficient, and the other zone is for
135 * headers that are smaller and represent fewer items, making the headers more
138 #define SLABZONE_SIZE(setsize) \
139 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
140 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16)
141 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE
142 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE)
143 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE)
144 static uma_zone_t slabzones[2];
147 * The initial hash tables come out of this zone so they can be allocated
148 * prior to malloc coming up.
150 static uma_zone_t hashzone;
152 /* The boot-time adjusted value for cache line alignment. */
153 int uma_align_cache = 64 - 1;
155 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
156 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
159 * Are we allowed to allocate buckets?
161 static int bucketdisable = 1;
163 /* Linked list of all kegs in the system */
164 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
166 /* Linked list of all cache-only zones in the system */
167 static LIST_HEAD(,uma_zone) uma_cachezones =
168 LIST_HEAD_INITIALIZER(uma_cachezones);
171 * Mutex for global lists: uma_kegs, uma_cachezones, and the per-keg list of
174 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
176 static struct sx uma_reclaim_lock;
179 * First available virual address for boot time allocations.
181 static vm_offset_t bootstart;
182 static vm_offset_t bootmem;
185 * kmem soft limit, initialized by uma_set_limit(). Ensure that early
186 * allocations don't trigger a wakeup of the reclaim thread.
188 unsigned long uma_kmem_limit = LONG_MAX;
189 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
190 "UMA kernel memory soft limit");
191 unsigned long uma_kmem_total;
192 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
193 "UMA kernel memory usage");
195 /* Is the VM done starting up? */
202 } booted = BOOT_COLD;
205 * This is the handle used to schedule events that need to happen
206 * outside of the allocation fast path.
208 static struct callout uma_callout;
209 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
212 * This structure is passed as the zone ctor arg so that I don't have to create
213 * a special allocation function just for zones.
215 struct uma_zctor_args {
230 struct uma_kctor_args {
239 struct uma_bucket_zone {
241 const char *ubz_name;
242 int ubz_entries; /* Number of items it can hold. */
243 int ubz_maxsize; /* Maximum allocation size per-item. */
247 * Compute the actual number of bucket entries to pack them in power
248 * of two sizes for more efficient space utilization.
250 #define BUCKET_SIZE(n) \
251 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
253 #define BUCKET_MAX BUCKET_SIZE(256)
255 struct uma_bucket_zone bucket_zones[] = {
256 /* Literal bucket sizes. */
257 { NULL, "2 Bucket", 2, 4096 },
258 { NULL, "4 Bucket", 4, 3072 },
259 { NULL, "8 Bucket", 8, 2048 },
260 { NULL, "16 Bucket", 16, 1024 },
261 /* Rounded down power of 2 sizes for efficiency. */
262 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
263 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
264 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
265 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
270 * Flags and enumerations to be passed to internal functions.
274 SKIP_CNT = 0x00000001,
275 SKIP_DTOR = 0x00010000,
276 SKIP_FINI = 0x00020000,
281 void uma_startup1(vm_offset_t);
282 void uma_startup2(void);
284 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
285 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
286 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
287 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
288 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
289 static void page_free(void *, vm_size_t, uint8_t);
290 static void pcpu_page_free(void *, vm_size_t, uint8_t);
291 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
292 static void cache_drain(uma_zone_t);
293 static void bucket_drain(uma_zone_t, uma_bucket_t);
294 static void bucket_cache_reclaim(uma_zone_t zone, bool, int);
295 static bool bucket_cache_reclaim_domain(uma_zone_t, bool, bool, int);
296 static int keg_ctor(void *, int, void *, int);
297 static void keg_dtor(void *, int, void *);
298 static void keg_drain(uma_keg_t keg, int domain);
299 static int zone_ctor(void *, int, void *, int);
300 static void zone_dtor(void *, int, void *);
301 static inline void item_dtor(uma_zone_t zone, void *item, int size,
302 void *udata, enum zfreeskip skip);
303 static int zero_init(void *, int, int);
304 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
305 int itemdomain, bool ws);
306 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
307 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
308 static void zone_timeout(uma_zone_t zone, void *);
309 static int hash_alloc(struct uma_hash *, u_int);
310 static int hash_expand(struct uma_hash *, struct uma_hash *);
311 static void hash_free(struct uma_hash *hash);
312 static void uma_timeout(void *);
313 static void uma_shutdown(void);
314 static void *zone_alloc_item(uma_zone_t, void *, int, int);
315 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
316 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
317 static void zone_free_limit(uma_zone_t zone, int count);
318 static void bucket_enable(void);
319 static void bucket_init(void);
320 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
321 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
322 static void bucket_zone_drain(int domain);
323 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
324 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
325 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
326 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
327 uma_fini fini, int align, uint32_t flags);
328 static int zone_import(void *, void **, int, int, int);
329 static void zone_release(void *, void **, int);
330 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
331 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
333 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
334 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
335 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
336 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
337 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
338 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
339 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
341 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
343 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
344 "Memory allocation debugging");
347 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
348 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
350 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
351 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
352 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
353 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
355 static u_int dbg_divisor = 1;
356 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
357 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
358 "Debug & thrash every this item in memory allocator");
360 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
361 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
362 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
363 &uma_dbg_cnt, "memory items debugged");
364 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
365 &uma_skip_cnt, "memory items skipped, not debugged");
368 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
369 "Universal Memory Allocator");
371 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
372 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
374 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
375 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
377 static int zone_warnings = 1;
378 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
379 "Warn when UMA zones becomes full");
381 static int multipage_slabs = 1;
382 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
383 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
384 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
385 "UMA may choose larger slab sizes for better efficiency");
388 * Select the slab zone for an offpage slab with the given maximum item count.
390 static inline uma_zone_t
394 return (slabzones[ipers > SLABZONE0_SETSIZE]);
398 * This routine checks to see whether or not it's safe to enable buckets.
404 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
405 bucketdisable = vm_page_count_min();
409 * Initialize bucket_zones, the array of zones of buckets of various sizes.
411 * For each zone, calculate the memory required for each bucket, consisting
412 * of the header and an array of pointers.
417 struct uma_bucket_zone *ubz;
420 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
421 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
422 size += sizeof(void *) * ubz->ubz_entries;
423 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
424 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
425 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
426 UMA_ZONE_FIRSTTOUCH);
431 * Given a desired number of entries for a bucket, return the zone from which
432 * to allocate the bucket.
434 static struct uma_bucket_zone *
435 bucket_zone_lookup(int entries)
437 struct uma_bucket_zone *ubz;
439 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
440 if (ubz->ubz_entries >= entries)
447 bucket_select(int size)
449 struct uma_bucket_zone *ubz;
451 ubz = &bucket_zones[0];
452 if (size > ubz->ubz_maxsize)
453 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
455 for (; ubz->ubz_entries != 0; ubz++)
456 if (ubz->ubz_maxsize < size)
459 return (ubz->ubz_entries);
463 bucket_alloc(uma_zone_t zone, void *udata, int flags)
465 struct uma_bucket_zone *ubz;
469 * Don't allocate buckets early in boot.
471 if (__predict_false(booted < BOOT_KVA))
475 * To limit bucket recursion we store the original zone flags
476 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
477 * NOVM flag to persist even through deep recursions. We also
478 * store ZFLAG_BUCKET once we have recursed attempting to allocate
479 * a bucket for a bucket zone so we do not allow infinite bucket
480 * recursion. This cookie will even persist to frees of unused
481 * buckets via the allocation path or bucket allocations in the
484 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
485 udata = (void *)(uintptr_t)zone->uz_flags;
487 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
489 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
491 if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
493 ubz = bucket_zone_lookup(atomic_load_16(&zone->uz_bucket_size));
494 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
496 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
499 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
502 bucket->ub_entries = min(ubz->ubz_entries,
503 zone->uz_bucket_size_max);
504 bucket->ub_seq = SMR_SEQ_INVALID;
505 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
506 zone->uz_name, zone, bucket);
513 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
515 struct uma_bucket_zone *ubz;
517 if (bucket->ub_cnt != 0)
518 bucket_drain(zone, bucket);
520 KASSERT(bucket->ub_cnt == 0,
521 ("bucket_free: Freeing a non free bucket."));
522 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
523 ("bucket_free: Freeing an SMR bucket."));
524 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
525 udata = (void *)(uintptr_t)zone->uz_flags;
526 ubz = bucket_zone_lookup(bucket->ub_entries);
527 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
531 bucket_zone_drain(int domain)
533 struct uma_bucket_zone *ubz;
535 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
536 uma_zone_reclaim_domain(ubz->ubz_zone, UMA_RECLAIM_DRAIN,
541 * Acquire the domain lock and record contention.
543 static uma_zone_domain_t
544 zone_domain_lock(uma_zone_t zone, int domain)
546 uma_zone_domain_t zdom;
549 zdom = ZDOM_GET(zone, domain);
551 if (ZDOM_OWNED(zdom))
554 /* This is unsynchronized. The counter does not need to be precise. */
555 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
556 zone->uz_bucket_size++;
561 * Search for the domain with the least cached items and return it if it
562 * is out of balance with the preferred domain.
564 static __noinline int
565 zone_domain_lowest(uma_zone_t zone, int pref)
567 long least, nitems, prefitems;
571 prefitems = least = LONG_MAX;
573 for (i = 0; i < vm_ndomains; i++) {
574 nitems = ZDOM_GET(zone, i)->uzd_nitems;
575 if (nitems < least) {
582 if (prefitems < least * 2)
589 * Search for the domain with the most cached items and return it or the
590 * preferred domain if it has enough to proceed.
592 static __noinline int
593 zone_domain_highest(uma_zone_t zone, int pref)
599 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
604 for (i = 0; i < vm_ndomains; i++) {
605 nitems = ZDOM_GET(zone, i)->uzd_nitems;
616 * Set the maximum imax value.
619 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
623 old = zdom->uzd_imax;
627 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
630 * We are at new maximum, so do the last WSS update for the old
631 * bimin and prepare to measure next allocation batch.
633 if (zdom->uzd_wss < old - zdom->uzd_bimin)
634 zdom->uzd_wss = old - zdom->uzd_bimin;
635 zdom->uzd_bimin = nitems;
639 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
640 * zone's caches. If a bucket is found the zone is not locked on return.
643 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
650 ZDOM_LOCK_ASSERT(zdom);
652 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
655 /* SMR Buckets can not be re-used until readers expire. */
656 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
657 bucket->ub_seq != SMR_SEQ_INVALID) {
658 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
660 bucket->ub_seq = SMR_SEQ_INVALID;
661 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
662 if (STAILQ_NEXT(bucket, ub_link) != NULL)
663 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
665 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
667 KASSERT(zdom->uzd_nitems >= bucket->ub_cnt,
668 ("%s: item count underflow (%ld, %d)",
669 __func__, zdom->uzd_nitems, bucket->ub_cnt));
670 KASSERT(bucket->ub_cnt > 0,
671 ("%s: empty bucket in bucket cache", __func__));
672 zdom->uzd_nitems -= bucket->ub_cnt;
676 * Shift the bounds of the current WSS interval to avoid
677 * perturbing the estimates.
679 cnt = lmin(zdom->uzd_bimin, bucket->ub_cnt);
680 atomic_subtract_long(&zdom->uzd_imax, cnt);
681 zdom->uzd_bimin -= cnt;
682 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
683 if (zdom->uzd_limin >= bucket->ub_cnt) {
684 zdom->uzd_limin -= bucket->ub_cnt;
689 } else if (zdom->uzd_bimin > zdom->uzd_nitems) {
690 zdom->uzd_bimin = zdom->uzd_nitems;
691 if (zdom->uzd_imin > zdom->uzd_nitems)
692 zdom->uzd_imin = zdom->uzd_nitems;
697 for (i = 0; i < bucket->ub_cnt; i++)
698 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
705 * Insert a full bucket into the specified cache. The "ws" parameter indicates
706 * whether the bucket's contents should be counted as part of the zone's working
707 * set. The bucket may be freed if it exceeds the bucket limit.
710 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
713 uma_zone_domain_t zdom;
715 /* We don't cache empty buckets. This can happen after a reclaim. */
716 if (bucket->ub_cnt == 0)
718 zdom = zone_domain_lock(zone, domain);
721 * Conditionally set the maximum number of items.
723 zdom->uzd_nitems += bucket->ub_cnt;
724 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
726 zone_domain_imax_set(zdom, zdom->uzd_nitems);
729 * Shift the bounds of the current WSS interval to
730 * avoid perturbing the estimates.
732 atomic_add_long(&zdom->uzd_imax, bucket->ub_cnt);
733 zdom->uzd_imin += bucket->ub_cnt;
734 zdom->uzd_bimin += bucket->ub_cnt;
735 zdom->uzd_limin += bucket->ub_cnt;
737 if (STAILQ_EMPTY(&zdom->uzd_buckets))
738 zdom->uzd_seq = bucket->ub_seq;
741 * Try to promote reuse of recently used items. For items
742 * protected by SMR, try to defer reuse to minimize polling.
744 if (bucket->ub_seq == SMR_SEQ_INVALID)
745 STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
747 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
751 zdom->uzd_nitems -= bucket->ub_cnt;
754 bucket_free(zone, bucket, udata);
757 /* Pops an item out of a per-cpu cache bucket. */
759 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
763 CRITICAL_ASSERT(curthread);
766 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
768 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
769 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
776 /* Pushes an item into a per-cpu cache bucket. */
778 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
781 CRITICAL_ASSERT(curthread);
782 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
783 ("uma_zfree: Freeing to non free bucket index."));
785 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
791 * Unload a UMA bucket from a per-cpu cache.
793 static inline uma_bucket_t
794 cache_bucket_unload(uma_cache_bucket_t bucket)
798 b = bucket->ucb_bucket;
800 MPASS(b->ub_entries == bucket->ucb_entries);
801 b->ub_cnt = bucket->ucb_cnt;
802 bucket->ucb_bucket = NULL;
803 bucket->ucb_entries = bucket->ucb_cnt = 0;
809 static inline uma_bucket_t
810 cache_bucket_unload_alloc(uma_cache_t cache)
813 return (cache_bucket_unload(&cache->uc_allocbucket));
816 static inline uma_bucket_t
817 cache_bucket_unload_free(uma_cache_t cache)
820 return (cache_bucket_unload(&cache->uc_freebucket));
823 static inline uma_bucket_t
824 cache_bucket_unload_cross(uma_cache_t cache)
827 return (cache_bucket_unload(&cache->uc_crossbucket));
831 * Load a bucket into a per-cpu cache bucket.
834 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
837 CRITICAL_ASSERT(curthread);
838 MPASS(bucket->ucb_bucket == NULL);
839 MPASS(b->ub_seq == SMR_SEQ_INVALID);
841 bucket->ucb_bucket = b;
842 bucket->ucb_cnt = b->ub_cnt;
843 bucket->ucb_entries = b->ub_entries;
847 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
850 cache_bucket_load(&cache->uc_allocbucket, b);
854 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
857 cache_bucket_load(&cache->uc_freebucket, b);
862 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
865 cache_bucket_load(&cache->uc_crossbucket, b);
870 * Copy and preserve ucb_spare.
873 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
876 b1->ucb_bucket = b2->ucb_bucket;
877 b1->ucb_entries = b2->ucb_entries;
878 b1->ucb_cnt = b2->ucb_cnt;
882 * Swap two cache buckets.
885 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
887 struct uma_cache_bucket b3;
889 CRITICAL_ASSERT(curthread);
891 cache_bucket_copy(&b3, b1);
892 cache_bucket_copy(b1, b2);
893 cache_bucket_copy(b2, &b3);
897 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
900 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
902 uma_zone_domain_t zdom;
906 * Avoid the lock if possible.
908 zdom = ZDOM_GET(zone, domain);
909 if (zdom->uzd_nitems == 0)
912 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
913 !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
917 * Check the zone's cache of buckets.
919 zdom = zone_domain_lock(zone, domain);
920 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL)
928 zone_log_warning(uma_zone_t zone)
930 static const struct timeval warninterval = { 300, 0 };
932 if (!zone_warnings || zone->uz_warning == NULL)
935 if (ratecheck(&zone->uz_ratecheck, &warninterval))
936 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
940 zone_maxaction(uma_zone_t zone)
943 if (zone->uz_maxaction.ta_func != NULL)
944 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
948 * Routine called by timeout which is used to fire off some time interval
949 * based calculations. (stats, hash size, etc.)
958 uma_timeout(void *unused)
961 zone_foreach(zone_timeout, NULL);
963 /* Reschedule this event */
964 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
968 * Update the working set size estimates for the zone's bucket cache.
969 * The constants chosen here are somewhat arbitrary.
972 zone_domain_update_wss(uma_zone_domain_t zdom)
976 ZDOM_LOCK_ASSERT(zdom);
977 MPASS(zdom->uzd_imax >= zdom->uzd_nitems);
978 MPASS(zdom->uzd_nitems >= zdom->uzd_bimin);
979 MPASS(zdom->uzd_bimin >= zdom->uzd_imin);
982 * Estimate WSS as modified moving average of biggest allocation
983 * batches for each period over few minutes (UMA_TIMEOUT of 20s).
985 zdom->uzd_wss = lmax(zdom->uzd_wss * 3 / 4,
986 zdom->uzd_imax - zdom->uzd_bimin);
989 * Estimate longtime minimum item count as a combination of recent
990 * minimum item count, adjusted by WSS for safety, and the modified
991 * moving average over the last several hours (UMA_TIMEOUT of 20s).
992 * timin measures time since limin tried to go negative, that means
993 * we were dangerously close to or got out of cache.
995 m = zdom->uzd_imin - zdom->uzd_wss;
997 if (zdom->uzd_limin >= m)
1000 zdom->uzd_limin = (m + zdom->uzd_limin * 255) / 256;
1003 zdom->uzd_limin = 0;
1004 zdom->uzd_timin = 0;
1007 /* To reduce period edge effects on WSS keep half of the imax. */
1008 atomic_subtract_long(&zdom->uzd_imax,
1009 (zdom->uzd_imax - zdom->uzd_nitems + 1) / 2);
1010 zdom->uzd_imin = zdom->uzd_bimin = zdom->uzd_nitems;
1014 * Routine to perform timeout driven calculations. This expands the
1015 * hashes and does per cpu statistics aggregation.
1020 zone_timeout(uma_zone_t zone, void *unused)
1025 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
1031 * Hash zones are non-numa by definition so the first domain
1032 * is the only one present.
1035 pages = keg->uk_domain[0].ud_pages;
1038 * Expand the keg hash table.
1040 * This is done if the number of slabs is larger than the hash size.
1041 * What I'm trying to do here is completely reduce collisions. This
1042 * may be a little aggressive. Should I allow for two collisions max?
1044 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1045 struct uma_hash newhash;
1046 struct uma_hash oldhash;
1050 * This is so involved because allocating and freeing
1051 * while the keg lock is held will lead to deadlock.
1052 * I have to do everything in stages and check for
1056 ret = hash_alloc(&newhash, 1 << fls(slabs));
1059 if (hash_expand(&keg->uk_hash, &newhash)) {
1060 oldhash = keg->uk_hash;
1061 keg->uk_hash = newhash;
1066 hash_free(&oldhash);
1073 /* Trim caches not used for a long time. */
1074 for (int i = 0; i < vm_ndomains; i++) {
1075 if (bucket_cache_reclaim_domain(zone, false, false, i) &&
1076 (zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1077 keg_drain(zone->uz_keg, i);
1082 * Allocate and zero fill the next sized hash table from the appropriate
1086 * hash A new hash structure with the old hash size in uh_hashsize
1089 * 1 on success and 0 on failure.
1092 hash_alloc(struct uma_hash *hash, u_int size)
1096 KASSERT(powerof2(size), ("hash size must be power of 2"));
1097 if (size > UMA_HASH_SIZE_INIT) {
1098 hash->uh_hashsize = size;
1099 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1100 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1102 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1103 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1104 UMA_ANYDOMAIN, M_WAITOK);
1105 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1107 if (hash->uh_slab_hash) {
1108 bzero(hash->uh_slab_hash, alloc);
1109 hash->uh_hashmask = hash->uh_hashsize - 1;
1117 * Expands the hash table for HASH zones. This is done from zone_timeout
1118 * to reduce collisions. This must not be done in the regular allocation
1119 * path, otherwise, we can recurse on the vm while allocating pages.
1122 * oldhash The hash you want to expand
1123 * newhash The hash structure for the new table
1131 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1133 uma_hash_slab_t slab;
1137 if (!newhash->uh_slab_hash)
1140 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1144 * I need to investigate hash algorithms for resizing without a
1148 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1149 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1150 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1151 LIST_REMOVE(slab, uhs_hlink);
1152 hval = UMA_HASH(newhash, slab->uhs_data);
1153 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1161 * Free the hash bucket to the appropriate backing store.
1164 * slab_hash The hash bucket we're freeing
1165 * hashsize The number of entries in that hash bucket
1171 hash_free(struct uma_hash *hash)
1173 if (hash->uh_slab_hash == NULL)
1175 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1176 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1178 free(hash->uh_slab_hash, M_UMAHASH);
1182 * Frees all outstanding items in a bucket
1185 * zone The zone to free to, must be unlocked.
1186 * bucket The free/alloc bucket with items.
1192 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1196 if (bucket->ub_cnt == 0)
1199 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1200 bucket->ub_seq != SMR_SEQ_INVALID) {
1201 smr_wait(zone->uz_smr, bucket->ub_seq);
1202 bucket->ub_seq = SMR_SEQ_INVALID;
1203 for (i = 0; i < bucket->ub_cnt; i++)
1204 item_dtor(zone, bucket->ub_bucket[i],
1205 zone->uz_size, NULL, SKIP_NONE);
1208 for (i = 0; i < bucket->ub_cnt; i++)
1209 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1210 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1211 if (zone->uz_max_items > 0)
1212 zone_free_limit(zone, bucket->ub_cnt);
1214 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1220 * Drains the per cpu caches for a zone.
1222 * NOTE: This may only be called while the zone is being torn down, and not
1223 * during normal operation. This is necessary in order that we do not have
1224 * to migrate CPUs to drain the per-CPU caches.
1227 * zone The zone to drain, must be unlocked.
1233 cache_drain(uma_zone_t zone)
1236 uma_bucket_t bucket;
1241 * XXX: It is safe to not lock the per-CPU caches, because we're
1242 * tearing down the zone anyway. I.e., there will be no further use
1243 * of the caches at this point.
1245 * XXX: It would good to be able to assert that the zone is being
1246 * torn down to prevent improper use of cache_drain().
1248 seq = SMR_SEQ_INVALID;
1249 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1250 seq = smr_advance(zone->uz_smr);
1252 cache = &zone->uz_cpu[cpu];
1253 bucket = cache_bucket_unload_alloc(cache);
1255 bucket_free(zone, bucket, NULL);
1256 bucket = cache_bucket_unload_free(cache);
1257 if (bucket != NULL) {
1258 bucket->ub_seq = seq;
1259 bucket_free(zone, bucket, NULL);
1261 bucket = cache_bucket_unload_cross(cache);
1262 if (bucket != NULL) {
1263 bucket->ub_seq = seq;
1264 bucket_free(zone, bucket, NULL);
1267 bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN);
1271 cache_shrink(uma_zone_t zone, void *unused)
1274 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1278 zone->uz_bucket_size =
1279 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1284 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1287 uma_bucket_t b1, b2, b3;
1290 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1293 b1 = b2 = b3 = NULL;
1295 cache = &zone->uz_cpu[curcpu];
1296 domain = PCPU_GET(domain);
1297 b1 = cache_bucket_unload_alloc(cache);
1300 * Don't flush SMR zone buckets. This leaves the zone without a
1301 * bucket and forces every free to synchronize().
1303 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1304 b2 = cache_bucket_unload_free(cache);
1305 b3 = cache_bucket_unload_cross(cache);
1310 zone_free_bucket(zone, b1, NULL, domain, false);
1312 zone_free_bucket(zone, b2, NULL, domain, false);
1314 /* Adjust the domain so it goes to zone_free_cross. */
1315 domain = (domain + 1) % vm_ndomains;
1316 zone_free_bucket(zone, b3, NULL, domain, false);
1321 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1322 * This is an expensive call because it needs to bind to all CPUs
1323 * one by one and enter a critical section on each of them in order
1324 * to safely access their cache buckets.
1325 * Zone lock must not be held on call this function.
1328 pcpu_cache_drain_safe(uma_zone_t zone)
1333 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1336 cache_shrink(zone, NULL);
1338 zone_foreach(cache_shrink, NULL);
1341 thread_lock(curthread);
1342 sched_bind(curthread, cpu);
1343 thread_unlock(curthread);
1346 cache_drain_safe_cpu(zone, NULL);
1348 zone_foreach(cache_drain_safe_cpu, NULL);
1350 thread_lock(curthread);
1351 sched_unbind(curthread);
1352 thread_unlock(curthread);
1356 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1357 * requested a drain, otherwise the per-domain caches are trimmed to either
1358 * estimated working set size.
1361 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain)
1363 uma_zone_domain_t zdom;
1364 uma_bucket_t bucket;
1369 * The cross bucket is partially filled and not part of
1370 * the item count. Reclaim it individually here.
1372 zdom = ZDOM_GET(zone, domain);
1373 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1374 ZONE_CROSS_LOCK(zone);
1375 bucket = zdom->uzd_cross;
1376 zdom->uzd_cross = NULL;
1377 ZONE_CROSS_UNLOCK(zone);
1379 bucket_free(zone, bucket, NULL);
1383 * If we were asked to drain the zone, we are done only once
1384 * this bucket cache is empty. If trim, we reclaim items in
1385 * excess of the zone's estimated working set size. Multiple
1386 * consecutive calls will shrink the WSS and so reclaim more.
1387 * If neither drain nor trim, then voluntarily reclaim 1/4
1388 * (to reduce first spike) of items not used for a long time.
1391 zone_domain_update_wss(zdom);
1395 target = zdom->uzd_wss;
1396 else if (zdom->uzd_timin > 900 / UMA_TIMEOUT)
1397 target = zdom->uzd_nitems - zdom->uzd_limin / 4;
1402 while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL &&
1403 zdom->uzd_nitems >= target + bucket->ub_cnt) {
1404 bucket = zone_fetch_bucket(zone, zdom, true);
1407 bucket_free(zone, bucket, NULL);
1416 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain)
1421 * Shrink the zone bucket size to ensure that the per-CPU caches
1422 * don't grow too large.
1424 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1425 zone->uz_bucket_size--;
1427 if (domain != UMA_ANYDOMAIN &&
1428 (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) {
1429 bucket_cache_reclaim_domain(zone, drain, true, domain);
1431 for (i = 0; i < vm_ndomains; i++)
1432 bucket_cache_reclaim_domain(zone, drain, true, i);
1437 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1443 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1444 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1446 mem = slab_data(slab, keg);
1447 flags = slab->us_flags;
1449 if (keg->uk_fini != NULL) {
1450 for (i--; i > -1; i--)
1453 * trash_fini implies that dtor was trash_dtor. trash_fini
1454 * would check that memory hasn't been modified since free,
1455 * which executed trash_dtor.
1456 * That's why we need to run uma_dbg_kskip() check here,
1457 * albeit we don't make skip check for other init/fini
1460 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1461 keg->uk_fini != trash_fini)
1463 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1465 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1466 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1468 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1469 uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1473 keg_drain_domain(uma_keg_t keg, int domain)
1475 struct slabhead freeslabs;
1477 uma_slab_t slab, tmp;
1478 uint32_t i, stofree, stokeep, partial;
1480 dom = &keg->uk_domain[domain];
1481 LIST_INIT(&freeslabs);
1483 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1484 keg->uk_name, keg, domain, dom->ud_free_items);
1486 KEG_LOCK(keg, domain);
1489 * Are the free items in partially allocated slabs sufficient to meet
1490 * the reserve? If not, compute the number of fully free slabs that must
1493 partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1494 if (partial < keg->uk_reserve) {
1495 stokeep = min(dom->ud_free_slabs,
1496 howmany(keg->uk_reserve - partial, keg->uk_ipers));
1500 stofree = dom->ud_free_slabs - stokeep;
1503 * Partition the free slabs into two sets: those that must be kept in
1504 * order to maintain the reserve, and those that may be released back to
1505 * the system. Since one set may be much larger than the other,
1506 * populate the smaller of the two sets and swap them if necessary.
1508 for (i = min(stofree, stokeep); i > 0; i--) {
1509 slab = LIST_FIRST(&dom->ud_free_slab);
1510 LIST_REMOVE(slab, us_link);
1511 LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1513 if (stofree > stokeep)
1514 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1516 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1517 LIST_FOREACH(slab, &freeslabs, us_link)
1518 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1520 dom->ud_free_items -= stofree * keg->uk_ipers;
1521 dom->ud_free_slabs -= stofree;
1522 dom->ud_pages -= stofree * keg->uk_ppera;
1523 KEG_UNLOCK(keg, domain);
1525 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1526 keg_free_slab(keg, slab, keg->uk_ipers);
1530 * Frees pages from a keg back to the system. This is done on demand from
1531 * the pageout daemon.
1536 keg_drain(uma_keg_t keg, int domain)
1540 if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1542 if (domain != UMA_ANYDOMAIN) {
1543 keg_drain_domain(keg, domain);
1545 for (i = 0; i < vm_ndomains; i++)
1546 keg_drain_domain(keg, i);
1551 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain)
1554 * Count active reclaim operations in order to interlock with
1555 * zone_dtor(), which removes the zone from global lists before
1556 * attempting to reclaim items itself.
1558 * The zone may be destroyed while sleeping, so only zone_dtor() should
1562 if (waitok == M_WAITOK) {
1563 while (zone->uz_reclaimers > 0)
1564 msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1);
1566 zone->uz_reclaimers++;
1568 bucket_cache_reclaim(zone, drain, domain);
1570 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1571 keg_drain(zone->uz_keg, domain);
1573 zone->uz_reclaimers--;
1574 if (zone->uz_reclaimers == 0)
1580 zone_drain(uma_zone_t zone, void *arg)
1584 domain = (int)(uintptr_t)arg;
1585 zone_reclaim(zone, domain, M_NOWAIT, true);
1589 zone_trim(uma_zone_t zone, void *arg)
1593 domain = (int)(uintptr_t)arg;
1594 zone_reclaim(zone, domain, M_NOWAIT, false);
1598 * Allocate a new slab for a keg and inserts it into the partial slab list.
1599 * The keg should be unlocked on entry. If the allocation succeeds it will
1600 * be locked on return.
1603 * flags Wait flags for the item initialization routine
1604 * aflags Wait flags for the slab allocation
1607 * The slab that was allocated or NULL if there is no memory and the
1608 * caller specified M_NOWAIT.
1611 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1622 KASSERT(domain >= 0 && domain < vm_ndomains,
1623 ("keg_alloc_slab: domain %d out of range", domain));
1625 allocf = keg->uk_allocf;
1628 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1629 uma_hash_slab_t hslab;
1630 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1634 slab = &hslab->uhs_slab;
1638 * This reproduces the old vm_zone behavior of zero filling pages the
1639 * first time they are added to a zone.
1641 * Malloced items are zeroed in uma_zalloc.
1644 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1649 if (keg->uk_flags & UMA_ZONE_NODUMP)
1652 /* zone is passed for legacy reasons. */
1653 size = keg->uk_ppera * PAGE_SIZE;
1654 mem = allocf(zone, size, domain, &sflags, aflags);
1656 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1657 zone_free_item(slabzone(keg->uk_ipers),
1658 slab_tohashslab(slab), NULL, SKIP_NONE);
1661 uma_total_inc(size);
1663 /* For HASH zones all pages go to the same uma_domain. */
1664 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1667 /* Point the slab into the allocated memory */
1668 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1669 slab = (uma_slab_t )(mem + keg->uk_pgoff);
1671 slab_tohashslab(slab)->uhs_data = mem;
1673 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1674 for (i = 0; i < keg->uk_ppera; i++)
1675 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1678 slab->us_freecount = keg->uk_ipers;
1679 slab->us_flags = sflags;
1680 slab->us_domain = domain;
1682 BIT_FILL(keg->uk_ipers, &slab->us_free);
1684 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1687 if (keg->uk_init != NULL) {
1688 for (i = 0; i < keg->uk_ipers; i++)
1689 if (keg->uk_init(slab_item(slab, keg, i),
1690 keg->uk_size, flags) != 0)
1692 if (i != keg->uk_ipers) {
1693 keg_free_slab(keg, slab, i);
1697 KEG_LOCK(keg, domain);
1699 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1700 slab, keg->uk_name, keg);
1702 if (keg->uk_flags & UMA_ZFLAG_HASH)
1703 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1706 * If we got a slab here it's safe to mark it partially used
1707 * and return. We assume that the caller is going to remove
1708 * at least one item.
1710 dom = &keg->uk_domain[domain];
1711 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1712 dom->ud_pages += keg->uk_ppera;
1713 dom->ud_free_items += keg->uk_ipers;
1722 * This function is intended to be used early on in place of page_alloc(). It
1723 * performs contiguous physical memory allocations and uses a bump allocator for
1724 * KVA, so is usable before the kernel map is initialized.
1727 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1736 pages = howmany(bytes, PAGE_SIZE);
1737 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1739 *pflag = UMA_SLAB_BOOT;
1740 m = vm_page_alloc_contig_domain(NULL, 0, domain,
1741 malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1742 (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1746 pa = VM_PAGE_TO_PHYS(m);
1747 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1748 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1749 defined(__riscv) || defined(__powerpc64__)
1750 if ((wait & M_NODUMP) == 0)
1754 /* Allocate KVA and indirectly advance bootmem. */
1755 mem = (void *)pmap_map(&bootmem, m->phys_addr,
1756 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1757 if ((wait & M_ZERO) != 0)
1758 bzero(mem, pages * PAGE_SIZE);
1764 startup_free(void *mem, vm_size_t bytes)
1769 va = (vm_offset_t)mem;
1770 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1773 * startup_alloc() returns direct-mapped slabs on some platforms. Avoid
1774 * unmapping ranges of the direct map.
1776 if (va >= bootstart && va + bytes <= bootmem)
1777 pmap_remove(kernel_pmap, va, va + bytes);
1778 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1779 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1780 defined(__riscv) || defined(__powerpc64__)
1781 dump_drop_page(VM_PAGE_TO_PHYS(m));
1783 vm_page_unwire_noq(m);
1789 * Allocates a number of pages from the system
1792 * bytes The number of bytes requested
1793 * wait Shall we wait?
1796 * A pointer to the alloced memory or possibly
1797 * NULL if M_NOWAIT is set.
1800 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1803 void *p; /* Returned page */
1805 *pflag = UMA_SLAB_KERNEL;
1806 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1812 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1815 struct pglist alloctail;
1816 vm_offset_t addr, zkva;
1818 vm_page_t p, p_next;
1823 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1825 TAILQ_INIT(&alloctail);
1826 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1827 malloc2vm_flags(wait);
1828 *pflag = UMA_SLAB_KERNEL;
1829 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1830 if (CPU_ABSENT(cpu)) {
1831 p = vm_page_alloc(NULL, 0, flags);
1834 p = vm_page_alloc(NULL, 0, flags);
1836 pc = pcpu_find(cpu);
1837 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1840 p = vm_page_alloc_domain(NULL, 0,
1841 pc->pc_domain, flags);
1842 if (__predict_false(p == NULL))
1843 p = vm_page_alloc(NULL, 0, flags);
1846 if (__predict_false(p == NULL))
1848 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1850 if ((addr = kva_alloc(bytes)) == 0)
1853 TAILQ_FOREACH(p, &alloctail, listq) {
1854 pmap_qenter(zkva, &p, 1);
1857 return ((void*)addr);
1859 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1860 vm_page_unwire_noq(p);
1867 * Allocates a number of pages from within an object
1870 * bytes The number of bytes requested
1871 * wait Shall we wait?
1874 * A pointer to the alloced memory or possibly
1875 * NULL if M_NOWAIT is set.
1878 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1881 TAILQ_HEAD(, vm_page) alloctail;
1883 vm_offset_t retkva, zkva;
1884 vm_page_t p, p_next;
1887 TAILQ_INIT(&alloctail);
1890 npages = howmany(bytes, PAGE_SIZE);
1891 while (npages > 0) {
1892 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1893 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1894 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1898 * Since the page does not belong to an object, its
1901 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1906 * Page allocation failed, free intermediate pages and
1909 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1910 vm_page_unwire_noq(p);
1915 *flags = UMA_SLAB_PRIV;
1916 zkva = keg->uk_kva +
1917 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1919 TAILQ_FOREACH(p, &alloctail, listq) {
1920 pmap_qenter(zkva, &p, 1);
1924 return ((void *)retkva);
1928 * Allocate physically contiguous pages.
1931 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1935 *pflag = UMA_SLAB_KERNEL;
1936 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1937 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1941 * Frees a number of pages to the system
1944 * mem A pointer to the memory to be freed
1945 * size The size of the memory being freed
1946 * flags The original p->us_flags field
1952 page_free(void *mem, vm_size_t size, uint8_t flags)
1955 if ((flags & UMA_SLAB_BOOT) != 0) {
1956 startup_free(mem, size);
1960 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1961 ("UMA: page_free used with invalid flags %x", flags));
1963 kmem_free((vm_offset_t)mem, size);
1967 * Frees pcpu zone allocations
1970 * mem A pointer to the memory to be freed
1971 * size The size of the memory being freed
1972 * flags The original p->us_flags field
1978 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1980 vm_offset_t sva, curva;
1984 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1986 if ((flags & UMA_SLAB_BOOT) != 0) {
1987 startup_free(mem, size);
1991 sva = (vm_offset_t)mem;
1992 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1993 paddr = pmap_kextract(curva);
1994 m = PHYS_TO_VM_PAGE(paddr);
1995 vm_page_unwire_noq(m);
1998 pmap_qremove(sva, size >> PAGE_SHIFT);
1999 kva_free(sva, size);
2003 * Zero fill initializer
2005 * Arguments/Returns follow uma_init specifications
2008 zero_init(void *mem, int size, int flags)
2015 static struct noslabbits *
2016 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2019 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2024 * Actual size of embedded struct slab (!OFFPAGE).
2027 slab_sizeof(int nitems)
2031 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2032 return (roundup(s, UMA_ALIGN_PTR + 1));
2035 #define UMA_FIXPT_SHIFT 31
2036 #define UMA_FRAC_FIXPT(n, d) \
2037 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2038 #define UMA_FIXPT_PCT(f) \
2039 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2040 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
2041 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2044 * Compute the number of items that will fit in a slab. If hdr is true, the
2045 * item count may be limited to provide space in the slab for an inline slab
2046 * header. Otherwise, all slab space will be provided for item storage.
2049 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2054 /* The padding between items is not needed after the last item. */
2055 padpi = rsize - size;
2059 * Start with the maximum item count and remove items until
2060 * the slab header first alongside the allocatable memory.
2062 for (ipers = MIN(SLAB_MAX_SETSIZE,
2063 (slabsize + padpi - slab_sizeof(1)) / rsize);
2065 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2069 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2075 struct keg_layout_result {
2083 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2084 struct keg_layout_result *kl)
2089 kl->slabsize = slabsize;
2091 /* Handle INTERNAL as inline with an extra page. */
2092 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2093 kl->format &= ~UMA_ZFLAG_INTERNAL;
2094 kl->slabsize += PAGE_SIZE;
2097 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2098 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2100 /* Account for memory used by an offpage slab header. */
2101 total = kl->slabsize;
2102 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2103 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2105 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2109 * Determine the format of a uma keg. This determines where the slab header
2110 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2113 * keg The zone we should initialize
2119 keg_layout(uma_keg_t keg)
2121 struct keg_layout_result kl = {}, kl_tmp;
2130 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2131 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2132 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2133 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2134 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2136 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2137 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2138 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2141 alignsize = keg->uk_align + 1;
2144 * Calculate the size of each allocation (rsize) according to
2145 * alignment. If the requested size is smaller than we have
2146 * allocation bits for we round it up.
2148 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2149 rsize = roundup2(rsize, alignsize);
2151 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2153 * We want one item to start on every align boundary in a page.
2154 * To do this we will span pages. We will also extend the item
2155 * by the size of align if it is an even multiple of align.
2156 * Otherwise, it would fall on the same boundary every time.
2158 if ((rsize & alignsize) == 0)
2160 slabsize = rsize * (PAGE_SIZE / alignsize);
2161 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2162 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2163 slabsize = round_page(slabsize);
2166 * Start with a slab size of as many pages as it takes to
2167 * represent a single item. We will try to fit as many
2168 * additional items into the slab as possible.
2170 slabsize = round_page(keg->uk_size);
2173 /* Build a list of all of the available formats for this keg. */
2176 /* Evaluate an inline slab layout. */
2177 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2180 /* TODO: vm_page-embedded slab. */
2183 * We can't do OFFPAGE if we're internal or if we've been
2184 * asked to not go to the VM for buckets. If we do this we
2185 * may end up going to the VM for slabs which we do not want
2186 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2187 * In those cases, evaluate a pseudo-format called INTERNAL
2188 * which has an inline slab header and one extra page to
2189 * guarantee that it fits.
2191 * Otherwise, see if using an OFFPAGE slab will improve our
2194 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2195 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2197 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2200 * Choose a slab size and format which satisfy the minimum efficiency.
2201 * Prefer the smallest slab size that meets the constraints.
2203 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2204 * for small items (up to PAGE_SIZE), the iteration increment is one
2205 * page; and for large items, the increment is one item.
2207 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2208 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2209 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2212 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2213 round_page(rsize * (i - 1) + keg->uk_size);
2215 for (j = 0; j < nfmt; j++) {
2216 /* Only if we have no viable format yet. */
2217 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2221 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2222 if (kl_tmp.eff <= kl.eff)
2227 CTR6(KTR_UMA, "keg %s layout: format %#x "
2228 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2229 keg->uk_name, kl.format, kl.ipers, rsize,
2230 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2232 /* Stop when we reach the minimum efficiency. */
2233 if (kl.eff >= UMA_MIN_EFF)
2237 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2238 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2239 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2243 pages = atop(kl.slabsize);
2244 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2245 pages *= mp_maxid + 1;
2247 keg->uk_rsize = rsize;
2248 keg->uk_ipers = kl.ipers;
2249 keg->uk_ppera = pages;
2250 keg->uk_flags |= kl.format;
2253 * How do we find the slab header if it is offpage or if not all item
2254 * start addresses are in the same page? We could solve the latter
2255 * case with vaddr alignment, but we don't.
2257 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2258 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2259 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2260 keg->uk_flags |= UMA_ZFLAG_HASH;
2262 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2265 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2266 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2268 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2269 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2270 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2271 keg->uk_ipers, pages));
2275 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2276 * the keg onto the global keg list.
2278 * Arguments/Returns follow uma_ctor specifications
2279 * udata Actually uma_kctor_args
2282 keg_ctor(void *mem, int size, void *udata, int flags)
2284 struct uma_kctor_args *arg = udata;
2285 uma_keg_t keg = mem;
2290 keg->uk_size = arg->size;
2291 keg->uk_init = arg->uminit;
2292 keg->uk_fini = arg->fini;
2293 keg->uk_align = arg->align;
2294 keg->uk_reserve = 0;
2295 keg->uk_flags = arg->flags;
2298 * We use a global round-robin policy by default. Zones with
2299 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2300 * case the iterator is never run.
2302 keg->uk_dr.dr_policy = DOMAINSET_RR();
2303 keg->uk_dr.dr_iter = 0;
2306 * The primary zone is passed to us at keg-creation time.
2309 keg->uk_name = zone->uz_name;
2311 if (arg->flags & UMA_ZONE_ZINIT)
2312 keg->uk_init = zero_init;
2314 if (arg->flags & UMA_ZONE_MALLOC)
2315 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2318 keg->uk_flags &= ~UMA_ZONE_PCPU;
2324 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2325 * work on. Use round-robin for everything else.
2327 * Zones may override the default by specifying either.
2330 if ((keg->uk_flags &
2331 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2332 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2333 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2334 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2338 * If we haven't booted yet we need allocations to go through the
2339 * startup cache until the vm is ready.
2341 #ifdef UMA_MD_SMALL_ALLOC
2342 if (keg->uk_ppera == 1)
2343 keg->uk_allocf = uma_small_alloc;
2346 if (booted < BOOT_KVA)
2347 keg->uk_allocf = startup_alloc;
2348 else if (keg->uk_flags & UMA_ZONE_PCPU)
2349 keg->uk_allocf = pcpu_page_alloc;
2350 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2351 keg->uk_allocf = contig_alloc;
2353 keg->uk_allocf = page_alloc;
2354 #ifdef UMA_MD_SMALL_ALLOC
2355 if (keg->uk_ppera == 1)
2356 keg->uk_freef = uma_small_free;
2359 if (keg->uk_flags & UMA_ZONE_PCPU)
2360 keg->uk_freef = pcpu_page_free;
2362 keg->uk_freef = page_free;
2365 * Initialize keg's locks.
2367 for (i = 0; i < vm_ndomains; i++)
2368 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2371 * If we're putting the slab header in the actual page we need to
2372 * figure out where in each page it goes. See slab_sizeof
2375 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2378 shsize = slab_sizeof(keg->uk_ipers);
2379 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2381 * The only way the following is possible is if with our
2382 * UMA_ALIGN_PTR adjustments we are now bigger than
2383 * UMA_SLAB_SIZE. I haven't checked whether this is
2384 * mathematically possible for all cases, so we make
2387 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2388 ("zone %s ipers %d rsize %d size %d slab won't fit",
2389 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2392 if (keg->uk_flags & UMA_ZFLAG_HASH)
2393 hash_alloc(&keg->uk_hash, 0);
2395 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2397 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2399 rw_wlock(&uma_rwlock);
2400 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2401 rw_wunlock(&uma_rwlock);
2406 zone_kva_available(uma_zone_t zone, void *unused)
2410 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2414 if (keg->uk_allocf == startup_alloc) {
2415 /* Switch to the real allocator. */
2416 if (keg->uk_flags & UMA_ZONE_PCPU)
2417 keg->uk_allocf = pcpu_page_alloc;
2418 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2420 keg->uk_allocf = contig_alloc;
2422 keg->uk_allocf = page_alloc;
2427 zone_alloc_counters(uma_zone_t zone, void *unused)
2430 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2431 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2432 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2433 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2437 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2439 uma_zone_domain_t zdom;
2442 struct sysctl_oid *oid, *domainoid;
2443 int domains, i, cnt;
2444 static const char *nokeg = "cache zone";
2448 * Make a sysctl safe copy of the zone name by removing
2449 * any special characters and handling dups by appending
2452 if (zone->uz_namecnt != 0) {
2453 /* Count the number of decimal digits and '_' separator. */
2454 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2456 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2458 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2461 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2462 for (c = zone->uz_ctlname; *c != '\0'; c++)
2463 if (strchr("./\\ -", *c) != NULL)
2467 * Basic parameters at the root.
2469 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2470 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2472 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2473 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2474 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2475 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2476 zone, 0, sysctl_handle_uma_zone_flags, "A",
2477 "Allocator configuration flags");
2478 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2479 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2480 "Desired per-cpu cache size");
2481 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2482 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2483 "Maximum allowed per-cpu cache size");
2488 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2489 domains = vm_ndomains;
2492 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2493 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2495 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2496 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2497 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2498 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2499 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2500 "Real object size with alignment");
2501 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2502 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2503 "pages per-slab allocation");
2504 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2505 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2506 "items available per-slab");
2507 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2508 "align", CTLFLAG_RD, &keg->uk_align, 0,
2509 "item alignment mask");
2510 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2511 "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2512 "number of reserved items");
2513 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2514 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2515 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2516 "Slab utilization (100 - internal fragmentation %)");
2517 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2518 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2519 for (i = 0; i < domains; i++) {
2520 dom = &keg->uk_domain[i];
2521 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2522 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2523 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2524 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2525 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2526 "Total pages currently allocated from VM");
2527 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2528 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2529 "items free in the slab layer");
2532 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2533 "name", CTLFLAG_RD, nokeg, "Keg name");
2536 * Information about zone limits.
2538 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2539 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2540 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2541 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2542 zone, 0, sysctl_handle_uma_zone_items, "QU",
2543 "Current number of allocated items if limit is set");
2544 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2545 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2546 "Maximum number of allocated and cached items");
2547 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2548 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2549 "Number of threads sleeping at limit");
2550 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2551 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2552 "Total zone limit sleeps");
2553 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2554 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2555 "Maximum number of items in each domain's bucket cache");
2558 * Per-domain zone information.
2560 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2561 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2562 for (i = 0; i < domains; i++) {
2563 zdom = ZDOM_GET(zone, i);
2564 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2565 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2566 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2567 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2568 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2569 "number of items in this domain");
2570 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2571 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2572 "maximum item count in this period");
2573 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2574 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2575 "minimum item count in this period");
2576 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2577 "bimin", CTLFLAG_RD, &zdom->uzd_bimin,
2578 "Minimum item count in this batch");
2579 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2580 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2581 "Working set size");
2582 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2583 "limin", CTLFLAG_RD, &zdom->uzd_limin,
2584 "Long time minimum item count");
2585 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2586 "timin", CTLFLAG_RD, &zdom->uzd_timin, 0,
2587 "Time since zero long time minimum item count");
2591 * General statistics.
2593 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2594 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2595 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2596 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2597 zone, 1, sysctl_handle_uma_zone_cur, "I",
2598 "Current number of allocated items");
2599 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2600 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2601 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2602 "Total allocation calls");
2603 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2604 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2605 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2606 "Total free calls");
2607 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2608 "fails", CTLFLAG_RD, &zone->uz_fails,
2609 "Number of allocation failures");
2610 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2611 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2612 "Free calls from the wrong domain");
2615 struct uma_zone_count {
2621 zone_count(uma_zone_t zone, void *arg)
2623 struct uma_zone_count *cnt;
2627 * Some zones are rapidly created with identical names and
2628 * destroyed out of order. This can lead to gaps in the count.
2629 * Use one greater than the maximum observed for this name.
2631 if (strcmp(zone->uz_name, cnt->name) == 0)
2632 cnt->count = MAX(cnt->count,
2633 zone->uz_namecnt + 1);
2637 zone_update_caches(uma_zone_t zone)
2641 for (i = 0; i <= mp_maxid; i++) {
2642 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2643 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2648 * Zone header ctor. This initializes all fields, locks, etc.
2650 * Arguments/Returns follow uma_ctor specifications
2651 * udata Actually uma_zctor_args
2654 zone_ctor(void *mem, int size, void *udata, int flags)
2656 struct uma_zone_count cnt;
2657 struct uma_zctor_args *arg = udata;
2658 uma_zone_domain_t zdom;
2659 uma_zone_t zone = mem;
2665 zone->uz_name = arg->name;
2666 zone->uz_ctor = arg->ctor;
2667 zone->uz_dtor = arg->dtor;
2668 zone->uz_init = NULL;
2669 zone->uz_fini = NULL;
2670 zone->uz_sleeps = 0;
2671 zone->uz_bucket_size = 0;
2672 zone->uz_bucket_size_min = 0;
2673 zone->uz_bucket_size_max = BUCKET_MAX;
2674 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2675 zone->uz_warning = NULL;
2676 /* The domain structures follow the cpu structures. */
2677 zone->uz_bucket_max = ULONG_MAX;
2678 timevalclear(&zone->uz_ratecheck);
2680 /* Count the number of duplicate names. */
2681 cnt.name = arg->name;
2683 zone_foreach(zone_count, &cnt);
2684 zone->uz_namecnt = cnt.count;
2685 ZONE_CROSS_LOCK_INIT(zone);
2687 for (i = 0; i < vm_ndomains; i++) {
2688 zdom = ZDOM_GET(zone, i);
2689 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2690 STAILQ_INIT(&zdom->uzd_buckets);
2694 if (arg->uminit == trash_init && arg->fini == trash_fini)
2695 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2699 * This is a pure cache zone, no kegs.
2702 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2703 ("zone_ctor: Import specified for non-cache zone."));
2704 zone->uz_flags = arg->flags;
2705 zone->uz_size = arg->size;
2706 zone->uz_import = arg->import;
2707 zone->uz_release = arg->release;
2708 zone->uz_arg = arg->arg;
2711 * Cache zones are round-robin unless a policy is
2712 * specified because they may have incompatible
2715 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2716 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2718 rw_wlock(&uma_rwlock);
2719 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2720 rw_wunlock(&uma_rwlock);
2725 * Use the regular zone/keg/slab allocator.
2727 zone->uz_import = zone_import;
2728 zone->uz_release = zone_release;
2729 zone->uz_arg = zone;
2732 if (arg->flags & UMA_ZONE_SECONDARY) {
2733 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2734 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2735 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2736 zone->uz_init = arg->uminit;
2737 zone->uz_fini = arg->fini;
2738 zone->uz_flags |= UMA_ZONE_SECONDARY;
2739 rw_wlock(&uma_rwlock);
2741 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2742 if (LIST_NEXT(z, uz_link) == NULL) {
2743 LIST_INSERT_AFTER(z, zone, uz_link);
2748 rw_wunlock(&uma_rwlock);
2749 } else if (keg == NULL) {
2750 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2751 arg->align, arg->flags)) == NULL)
2754 struct uma_kctor_args karg;
2757 /* We should only be here from uma_startup() */
2758 karg.size = arg->size;
2759 karg.uminit = arg->uminit;
2760 karg.fini = arg->fini;
2761 karg.align = arg->align;
2762 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2764 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2770 /* Inherit properties from the keg. */
2772 zone->uz_size = keg->uk_size;
2773 zone->uz_flags |= (keg->uk_flags &
2774 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2777 if (booted >= BOOT_PCPU) {
2778 zone_alloc_counters(zone, NULL);
2779 if (booted >= BOOT_RUNNING)
2780 zone_alloc_sysctl(zone, NULL);
2782 zone->uz_allocs = EARLY_COUNTER;
2783 zone->uz_frees = EARLY_COUNTER;
2784 zone->uz_fails = EARLY_COUNTER;
2787 /* Caller requests a private SMR context. */
2788 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2789 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2791 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2792 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2793 ("Invalid zone flag combination"));
2794 if (arg->flags & UMA_ZFLAG_INTERNAL)
2795 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2796 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2797 zone->uz_bucket_size = BUCKET_MAX;
2798 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2799 zone->uz_bucket_size = 0;
2801 zone->uz_bucket_size = bucket_select(zone->uz_size);
2802 zone->uz_bucket_size_min = zone->uz_bucket_size;
2803 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2804 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2805 zone_update_caches(zone);
2811 * Keg header dtor. This frees all data, destroys locks, frees the hash
2812 * table and removes the keg from the global list.
2814 * Arguments/Returns follow uma_dtor specifications
2818 keg_dtor(void *arg, int size, void *udata)
2821 uint32_t free, pages;
2824 keg = (uma_keg_t)arg;
2826 for (i = 0; i < vm_ndomains; i++) {
2827 free += keg->uk_domain[i].ud_free_items;
2828 pages += keg->uk_domain[i].ud_pages;
2829 KEG_LOCK_FINI(keg, i);
2832 printf("Freed UMA keg (%s) was not empty (%u items). "
2833 " Lost %u pages of memory.\n",
2834 keg->uk_name ? keg->uk_name : "",
2835 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2837 hash_free(&keg->uk_hash);
2843 * Arguments/Returns follow uma_dtor specifications
2847 zone_dtor(void *arg, int size, void *udata)
2853 zone = (uma_zone_t)arg;
2855 sysctl_remove_oid(zone->uz_oid, 1, 1);
2857 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2860 rw_wlock(&uma_rwlock);
2861 LIST_REMOVE(zone, uz_link);
2862 rw_wunlock(&uma_rwlock);
2863 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2865 keg->uk_reserve = 0;
2867 zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true);
2870 * We only destroy kegs from non secondary/non cache zones.
2872 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2874 rw_wlock(&uma_rwlock);
2875 LIST_REMOVE(keg, uk_link);
2876 rw_wunlock(&uma_rwlock);
2877 zone_free_item(kegs, keg, NULL, SKIP_NONE);
2879 counter_u64_free(zone->uz_allocs);
2880 counter_u64_free(zone->uz_frees);
2881 counter_u64_free(zone->uz_fails);
2882 counter_u64_free(zone->uz_xdomain);
2883 free(zone->uz_ctlname, M_UMA);
2884 for (i = 0; i < vm_ndomains; i++)
2885 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2886 ZONE_CROSS_LOCK_FINI(zone);
2890 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2895 LIST_FOREACH(keg, &uma_kegs, uk_link) {
2896 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2899 LIST_FOREACH(zone, &uma_cachezones, uz_link)
2904 * Traverses every zone in the system and calls a callback
2907 * zfunc A pointer to a function which accepts a zone
2914 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2917 rw_rlock(&uma_rwlock);
2918 zone_foreach_unlocked(zfunc, arg);
2919 rw_runlock(&uma_rwlock);
2923 * Initialize the kernel memory allocator. This is done after pages can be
2924 * allocated but before general KVA is available.
2927 uma_startup1(vm_offset_t virtual_avail)
2929 struct uma_zctor_args args;
2930 size_t ksize, zsize, size;
2931 uma_keg_t primarykeg;
2936 bootstart = bootmem = virtual_avail;
2938 rw_init(&uma_rwlock, "UMA lock");
2939 sx_init(&uma_reclaim_lock, "umareclaim");
2941 ksize = sizeof(struct uma_keg) +
2942 (sizeof(struct uma_domain) * vm_ndomains);
2943 ksize = roundup(ksize, UMA_SUPER_ALIGN);
2944 zsize = sizeof(struct uma_zone) +
2945 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2946 (sizeof(struct uma_zone_domain) * vm_ndomains);
2947 zsize = roundup(zsize, UMA_SUPER_ALIGN);
2949 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2950 size = (zsize * 2) + ksize;
2951 for (domain = 0; domain < vm_ndomains; domain++) {
2952 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
2957 zones = (uma_zone_t)m;
2959 kegs = (uma_zone_t)m;
2961 primarykeg = (uma_keg_t)m;
2963 /* "manually" create the initial zone */
2964 memset(&args, 0, sizeof(args));
2965 args.name = "UMA Kegs";
2967 args.ctor = keg_ctor;
2968 args.dtor = keg_dtor;
2969 args.uminit = zero_init;
2971 args.keg = primarykeg;
2972 args.align = UMA_SUPER_ALIGN - 1;
2973 args.flags = UMA_ZFLAG_INTERNAL;
2974 zone_ctor(kegs, zsize, &args, M_WAITOK);
2976 args.name = "UMA Zones";
2978 args.ctor = zone_ctor;
2979 args.dtor = zone_dtor;
2980 args.uminit = zero_init;
2983 args.align = UMA_SUPER_ALIGN - 1;
2984 args.flags = UMA_ZFLAG_INTERNAL;
2985 zone_ctor(zones, zsize, &args, M_WAITOK);
2987 /* Now make zones for slab headers */
2988 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2989 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2990 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2991 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2993 hashzone = uma_zcreate("UMA Hash",
2994 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2995 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3001 #ifndef UMA_MD_SMALL_ALLOC
3002 extern void vm_radix_reserve_kva(void);
3006 * Advertise the availability of normal kva allocations and switch to
3007 * the default back-end allocator. Marks the KVA we consumed on startup
3008 * as used in the map.
3014 if (bootstart != bootmem) {
3015 vm_map_lock(kernel_map);
3016 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
3017 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
3018 vm_map_unlock(kernel_map);
3021 #ifndef UMA_MD_SMALL_ALLOC
3022 /* Set up radix zone to use noobj_alloc. */
3023 vm_radix_reserve_kva();
3027 zone_foreach_unlocked(zone_kva_available, NULL);
3032 * Allocate counters as early as possible so that boot-time allocations are
3033 * accounted more precisely.
3036 uma_startup_pcpu(void *arg __unused)
3039 zone_foreach_unlocked(zone_alloc_counters, NULL);
3042 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
3045 * Finish our initialization steps.
3048 uma_startup3(void *arg __unused)
3052 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
3053 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
3054 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
3056 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
3057 callout_init(&uma_callout, 1);
3058 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
3059 booted = BOOT_RUNNING;
3061 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
3062 EVENTHANDLER_PRI_FIRST);
3064 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
3070 booted = BOOT_SHUTDOWN;
3074 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3075 int align, uint32_t flags)
3077 struct uma_kctor_args args;
3080 args.uminit = uminit;
3082 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
3085 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3088 /* Public functions */
3091 uma_set_align(int align)
3094 if (align != UMA_ALIGN_CACHE)
3095 uma_align_cache = align;
3100 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3101 uma_init uminit, uma_fini fini, int align, uint32_t flags)
3104 struct uma_zctor_args args;
3107 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
3110 /* This stuff is essential for the zone ctor */
3111 memset(&args, 0, sizeof(args));
3116 args.uminit = uminit;
3120 * Inject procedures which check for memory use after free if we are
3121 * allowed to scramble the memory while it is not allocated. This
3122 * requires that: UMA is actually able to access the memory, no init
3123 * or fini procedures, no dependency on the initial value of the
3124 * memory, and no (legitimate) use of the memory after free. Note,
3125 * the ctor and dtor do not need to be empty.
3127 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3128 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3129 args.uminit = trash_init;
3130 args.fini = trash_fini;
3137 sx_xlock(&uma_reclaim_lock);
3138 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3139 sx_xunlock(&uma_reclaim_lock);
3146 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3147 uma_init zinit, uma_fini zfini, uma_zone_t primary)
3149 struct uma_zctor_args args;
3153 keg = primary->uz_keg;
3154 memset(&args, 0, sizeof(args));
3156 args.size = keg->uk_size;
3159 args.uminit = zinit;
3161 args.align = keg->uk_align;
3162 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3165 sx_xlock(&uma_reclaim_lock);
3166 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3167 sx_xunlock(&uma_reclaim_lock);
3174 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3175 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3176 void *arg, int flags)
3178 struct uma_zctor_args args;
3180 memset(&args, 0, sizeof(args));
3185 args.uminit = zinit;
3187 args.import = zimport;
3188 args.release = zrelease;
3191 args.flags = flags | UMA_ZFLAG_CACHE;
3193 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3198 uma_zdestroy(uma_zone_t zone)
3202 * Large slabs are expensive to reclaim, so don't bother doing
3203 * unnecessary work if we're shutting down.
3205 if (booted == BOOT_SHUTDOWN &&
3206 zone->uz_fini == NULL && zone->uz_release == zone_release)
3208 sx_xlock(&uma_reclaim_lock);
3209 zone_free_item(zones, zone, NULL, SKIP_NONE);
3210 sx_xunlock(&uma_reclaim_lock);
3214 uma_zwait(uma_zone_t zone)
3217 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3218 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3219 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3220 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3222 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3226 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3228 void *item, *pcpu_item;
3232 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3234 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3237 pcpu_item = zpcpu_base_to_offset(item);
3238 if (flags & M_ZERO) {
3240 for (i = 0; i <= mp_maxid; i++)
3241 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3243 bzero(item, zone->uz_size);
3250 * A stub while both regular and pcpu cases are identical.
3253 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3258 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3261 /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */
3262 if (pcpu_item == NULL)
3265 item = zpcpu_offset_to_base(pcpu_item);
3266 uma_zfree_arg(zone, item, udata);
3269 static inline void *
3270 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3276 skipdbg = uma_dbg_zskip(zone, item);
3277 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3278 zone->uz_ctor != trash_ctor)
3279 trash_ctor(item, size, udata, flags);
3281 /* Check flags before loading ctor pointer. */
3282 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3283 __predict_false(zone->uz_ctor != NULL) &&
3284 zone->uz_ctor(item, size, udata, flags) != 0) {
3285 counter_u64_add(zone->uz_fails, 1);
3286 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3291 uma_dbg_alloc(zone, NULL, item);
3293 if (__predict_false(flags & M_ZERO))
3294 return (memset(item, 0, size));
3300 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3301 enum zfreeskip skip)
3306 skipdbg = uma_dbg_zskip(zone, item);
3307 if (skip == SKIP_NONE && !skipdbg) {
3308 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3309 uma_dbg_free(zone, udata, item);
3311 uma_dbg_free(zone, NULL, item);
3314 if (__predict_true(skip < SKIP_DTOR)) {
3315 if (zone->uz_dtor != NULL)
3316 zone->uz_dtor(item, size, udata);
3318 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3319 zone->uz_dtor != trash_dtor)
3320 trash_dtor(item, size, udata);
3327 item_domain(void *item)
3331 domain = vm_phys_domain(vtophys(item));
3332 KASSERT(domain >= 0 && domain < vm_ndomains,
3333 ("%s: unknown domain for item %p", __func__, item));
3338 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3339 #define UMA_ZALLOC_DEBUG
3341 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3347 if (flags & M_WAITOK) {
3348 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3349 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3354 KASSERT((flags & M_EXEC) == 0,
3355 ("uma_zalloc_debug: called with M_EXEC"));
3356 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3357 ("uma_zalloc_debug: called within spinlock or critical section"));
3358 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3359 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3362 #ifdef DEBUG_MEMGUARD
3363 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3365 item = memguard_alloc(zone->uz_size, flags);
3367 error = EJUSTRETURN;
3368 if (zone->uz_init != NULL &&
3369 zone->uz_init(item, zone->uz_size, flags) != 0) {
3373 if (zone->uz_ctor != NULL &&
3374 zone->uz_ctor(item, zone->uz_size, udata,
3376 counter_u64_add(zone->uz_fails, 1);
3377 zone->uz_fini(item, zone->uz_size);
3384 /* This is unfortunate but should not be fatal. */
3391 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3393 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3394 ("uma_zfree_debug: called with spinlock or critical section held"));
3396 #ifdef DEBUG_MEMGUARD
3397 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3398 if (zone->uz_dtor != NULL)
3399 zone->uz_dtor(item, zone->uz_size, udata);
3400 if (zone->uz_fini != NULL)
3401 zone->uz_fini(item, zone->uz_size);
3402 memguard_free(item);
3403 return (EJUSTRETURN);
3410 static inline void *
3411 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3412 void *udata, int flags)
3417 item = cache_bucket_pop(cache, bucket);
3418 size = cache_uz_size(cache);
3419 uz_flags = cache_uz_flags(cache);
3421 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3424 static __noinline void *
3425 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3427 uma_cache_bucket_t bucket;
3430 while (cache_alloc(zone, cache, udata, flags)) {
3431 cache = &zone->uz_cpu[curcpu];
3432 bucket = &cache->uc_allocbucket;
3433 if (__predict_false(bucket->ucb_cnt == 0))
3435 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3440 * We can not get a bucket so try to return a single item.
3442 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3443 domain = PCPU_GET(domain);
3445 domain = UMA_ANYDOMAIN;
3446 return (zone_alloc_item(zone, udata, domain, flags));
3451 uma_zalloc_smr(uma_zone_t zone, int flags)
3453 uma_cache_bucket_t bucket;
3456 #ifdef UMA_ZALLOC_DEBUG
3459 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3460 ("uma_zalloc_arg: called with non-SMR zone."));
3461 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3466 cache = &zone->uz_cpu[curcpu];
3467 bucket = &cache->uc_allocbucket;
3468 if (__predict_false(bucket->ucb_cnt == 0))
3469 return (cache_alloc_retry(zone, cache, NULL, flags));
3470 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3475 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3477 uma_cache_bucket_t bucket;
3480 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3481 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3483 /* This is the fast path allocation */
3484 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3487 #ifdef UMA_ZALLOC_DEBUG
3490 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3491 ("uma_zalloc_arg: called with SMR zone."));
3492 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3497 * If possible, allocate from the per-CPU cache. There are two
3498 * requirements for safe access to the per-CPU cache: (1) the thread
3499 * accessing the cache must not be preempted or yield during access,
3500 * and (2) the thread must not migrate CPUs without switching which
3501 * cache it accesses. We rely on a critical section to prevent
3502 * preemption and migration. We release the critical section in
3503 * order to acquire the zone mutex if we are unable to allocate from
3504 * the current cache; when we re-acquire the critical section, we
3505 * must detect and handle migration if it has occurred.
3508 cache = &zone->uz_cpu[curcpu];
3509 bucket = &cache->uc_allocbucket;
3510 if (__predict_false(bucket->ucb_cnt == 0))
3511 return (cache_alloc_retry(zone, cache, udata, flags));
3512 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3516 * Replenish an alloc bucket and possibly restore an old one. Called in
3517 * a critical section. Returns in a critical section.
3519 * A false return value indicates an allocation failure.
3520 * A true return value indicates success and the caller should retry.
3522 static __noinline bool
3523 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3525 uma_bucket_t bucket;
3526 int curdomain, domain;
3529 CRITICAL_ASSERT(curthread);
3532 * If we have run out of items in our alloc bucket see
3533 * if we can switch with the free bucket.
3535 * SMR Zones can't re-use the free bucket until the sequence has
3538 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3539 cache->uc_freebucket.ucb_cnt != 0) {
3540 cache_bucket_swap(&cache->uc_freebucket,
3541 &cache->uc_allocbucket);
3546 * Discard any empty allocation bucket while we hold no locks.
3548 bucket = cache_bucket_unload_alloc(cache);
3551 if (bucket != NULL) {
3552 KASSERT(bucket->ub_cnt == 0,
3553 ("cache_alloc: Entered with non-empty alloc bucket."));
3554 bucket_free(zone, bucket, udata);
3558 * Attempt to retrieve the item from the per-CPU cache has failed, so
3559 * we must go back to the zone. This requires the zdom lock, so we
3560 * must drop the critical section, then re-acquire it when we go back
3561 * to the cache. Since the critical section is released, we may be
3562 * preempted or migrate. As such, make sure not to maintain any
3563 * thread-local state specific to the cache from prior to releasing
3564 * the critical section.
3566 domain = PCPU_GET(domain);
3567 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3568 VM_DOMAIN_EMPTY(domain))
3569 domain = zone_domain_highest(zone, domain);
3570 bucket = cache_fetch_bucket(zone, cache, domain);
3571 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3572 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3578 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3579 zone->uz_name, zone, bucket);
3580 if (bucket == NULL) {
3586 * See if we lost the race or were migrated. Cache the
3587 * initialized bucket to make this less likely or claim
3588 * the memory directly.
3591 cache = &zone->uz_cpu[curcpu];
3592 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3593 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3594 (curdomain = PCPU_GET(domain)) == domain ||
3595 VM_DOMAIN_EMPTY(curdomain))) {
3597 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3599 cache_bucket_load_alloc(cache, bucket);
3604 * We lost the race, release this bucket and start over.
3607 zone_put_bucket(zone, domain, bucket, udata, !new);
3614 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3617 uma_bucket_t bucket;
3618 uma_zone_domain_t zdom;
3622 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3623 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3625 /* This is the fast path allocation */
3626 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3627 zone->uz_name, zone, domain, flags);
3629 if (flags & M_WAITOK) {
3630 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3631 "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3633 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3634 ("uma_zalloc_domain: called with spinlock or critical section held"));
3635 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3636 ("uma_zalloc_domain: called with SMR zone."));
3638 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3639 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3641 if (vm_ndomains == 1)
3642 return (uma_zalloc_arg(zone, udata, flags));
3645 * Try to allocate from the bucket cache before falling back to the keg.
3646 * We could try harder and attempt to allocate from per-CPU caches or
3647 * the per-domain cross-domain buckets, but the complexity is probably
3648 * not worth it. It is more important that frees of previous
3649 * cross-domain allocations do not blow up the cache.
3651 zdom = zone_domain_lock(zone, domain);
3652 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3653 item = bucket->ub_bucket[bucket->ub_cnt - 1];
3655 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3658 zone_put_bucket(zone, domain, bucket, udata, true);
3659 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3662 KASSERT(item_domain(item) == domain,
3663 ("%s: bucket cache item %p from wrong domain",
3665 counter_u64_add(zone->uz_allocs, 1);
3670 return (zone_alloc_item(zone, udata, domain, flags));
3672 return (uma_zalloc_arg(zone, udata, flags));
3677 * Find a slab with some space. Prefer slabs that are partially used over those
3678 * that are totally full. This helps to reduce fragmentation.
3680 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3684 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3690 KASSERT(domain >= 0 && domain < vm_ndomains,
3691 ("keg_first_slab: domain %d out of range", domain));
3692 KEG_LOCK_ASSERT(keg, domain);
3697 dom = &keg->uk_domain[domain];
3698 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3700 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3701 LIST_REMOVE(slab, us_link);
3702 dom->ud_free_slabs--;
3703 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3707 domain = (domain + 1) % vm_ndomains;
3708 } while (domain != start);
3714 * Fetch an existing slab from a free or partial list. Returns with the
3715 * keg domain lock held if a slab was found or unlocked if not.
3718 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3723 /* HASH has a single free list. */
3724 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3727 KEG_LOCK(keg, domain);
3728 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3729 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3730 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3731 KEG_UNLOCK(keg, domain);
3738 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3740 struct vm_domainset_iter di;
3747 * Use the keg's policy if upper layers haven't already specified a
3748 * domain (as happens with first-touch zones).
3750 * To avoid races we run the iterator with the keg lock held, but that
3751 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3752 * clear M_WAITOK and handle low memory conditions locally.
3754 rr = rdomain == UMA_ANYDOMAIN;
3756 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3757 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3765 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3770 * M_NOVM means don't ask at all!
3775 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3778 if (!rr && (flags & M_WAITOK) == 0)
3780 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3781 if ((flags & M_WAITOK) != 0) {
3782 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
3790 * We might not have been able to get a slab but another cpu
3791 * could have while we were unlocked. Check again before we
3794 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3801 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3807 KEG_LOCK_ASSERT(keg, slab->us_domain);
3809 dom = &keg->uk_domain[slab->us_domain];
3810 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3811 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3812 item = slab_item(slab, keg, freei);
3813 slab->us_freecount--;
3814 dom->ud_free_items--;
3817 * Move this slab to the full list. It must be on the partial list, so
3818 * we do not need to update the free slab count. In particular,
3819 * keg_fetch_slab() always returns slabs on the partial list.
3821 if (slab->us_freecount == 0) {
3822 LIST_REMOVE(slab, us_link);
3823 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3830 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3844 /* Try to keep the buckets totally full */
3845 for (i = 0; i < max; ) {
3846 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3849 stripe = howmany(max, vm_ndomains);
3851 dom = &keg->uk_domain[slab->us_domain];
3853 bucket[i++] = slab_alloc_item(keg, slab);
3854 if (dom->ud_free_items <= keg->uk_reserve) {
3856 * Avoid depleting the reserve after a
3857 * successful item allocation, even if
3858 * M_USE_RESERVE is specified.
3860 KEG_UNLOCK(keg, slab->us_domain);
3865 * If the zone is striped we pick a new slab for every
3866 * N allocations. Eliminating this conditional will
3867 * instead pick a new domain for each bucket rather
3868 * than stripe within each bucket. The current option
3869 * produces more fragmentation and requires more cpu
3870 * time but yields better distribution.
3872 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3873 vm_ndomains > 1 && --stripe == 0)
3876 } while (slab->us_freecount != 0 && i < max);
3877 KEG_UNLOCK(keg, slab->us_domain);
3879 /* Don't block if we allocated any successfully. */
3888 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3890 uint64_t old, new, total, max;
3893 * The hard case. We're going to sleep because there were existing
3894 * sleepers or because we ran out of items. This routine enforces
3895 * fairness by keeping fifo order.
3897 * First release our ill gotten gains and make some noise.
3900 zone_free_limit(zone, count);
3901 zone_log_warning(zone);
3902 zone_maxaction(zone);
3903 if (flags & M_NOWAIT)
3907 * We need to allocate an item or set ourself as a sleeper
3908 * while the sleepq lock is held to avoid wakeup races. This
3909 * is essentially a home rolled semaphore.
3911 sleepq_lock(&zone->uz_max_items);
3912 old = zone->uz_items;
3914 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3915 /* Cache the max since we will evaluate twice. */
3916 max = zone->uz_max_items;
3917 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3918 UZ_ITEMS_COUNT(old) >= max)
3919 new = old + UZ_ITEMS_SLEEPER;
3921 new = old + MIN(count, max - old);
3922 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3924 /* We may have successfully allocated under the sleepq lock. */
3925 if (UZ_ITEMS_SLEEPERS(new) == 0) {
3926 sleepq_release(&zone->uz_max_items);
3931 * This is in a different cacheline from uz_items so that we
3932 * don't constantly invalidate the fastpath cacheline when we
3933 * adjust item counts. This could be limited to toggling on
3936 atomic_add_32(&zone->uz_sleepers, 1);
3937 atomic_add_64(&zone->uz_sleeps, 1);
3940 * We have added ourselves as a sleeper. The sleepq lock
3941 * protects us from wakeup races. Sleep now and then retry.
3943 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3944 sleepq_wait(&zone->uz_max_items, PVM);
3947 * After wakeup, remove ourselves as a sleeper and try
3948 * again. We no longer have the sleepq lock for protection.
3950 * Subract ourselves as a sleeper while attempting to add
3953 atomic_subtract_32(&zone->uz_sleepers, 1);
3954 old = atomic_fetchadd_64(&zone->uz_items,
3955 -(UZ_ITEMS_SLEEPER - count));
3956 /* We're no longer a sleeper. */
3957 old -= UZ_ITEMS_SLEEPER;
3960 * If we're still at the limit, restart. Notably do not
3961 * block on other sleepers. Cache the max value to protect
3962 * against changes via sysctl.
3964 total = UZ_ITEMS_COUNT(old);
3965 max = zone->uz_max_items;
3968 /* Truncate if necessary, otherwise wake other sleepers. */
3969 if (total + count > max) {
3970 zone_free_limit(zone, total + count - max);
3971 count = max - total;
3972 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3973 wakeup_one(&zone->uz_max_items);
3980 * Allocate 'count' items from our max_items limit. Returns the number
3981 * available. If M_NOWAIT is not specified it will sleep until at least
3982 * one item can be allocated.
3985 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3990 max = zone->uz_max_items;
3994 * We expect normal allocations to succeed with a simple
3997 old = atomic_fetchadd_64(&zone->uz_items, count);
3998 if (__predict_true(old + count <= max))
4002 * If we had some items and no sleepers just return the
4003 * truncated value. We have to release the excess space
4004 * though because that may wake sleepers who weren't woken
4005 * because we were temporarily over the limit.
4008 zone_free_limit(zone, (old + count) - max);
4011 return (zone_alloc_limit_hard(zone, count, flags));
4015 * Free a number of items back to the limit.
4018 zone_free_limit(uma_zone_t zone, int count)
4025 * In the common case we either have no sleepers or
4026 * are still over the limit and can just return.
4028 old = atomic_fetchadd_64(&zone->uz_items, -count);
4029 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4030 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4034 * Moderate the rate of wakeups. Sleepers will continue
4035 * to generate wakeups if necessary.
4037 wakeup_one(&zone->uz_max_items);
4041 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4043 uma_bucket_t bucket;
4046 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4049 /* Avoid allocs targeting empty domains. */
4050 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4051 domain = UMA_ANYDOMAIN;
4052 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4053 domain = UMA_ANYDOMAIN;
4055 if (zone->uz_max_items > 0)
4056 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4059 maxbucket = zone->uz_bucket_size;
4063 /* Don't wait for buckets, preserve caller's NOVM setting. */
4064 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4065 if (bucket == NULL) {
4070 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4071 MIN(maxbucket, bucket->ub_entries), domain, flags);
4074 * Initialize the memory if necessary.
4076 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4079 for (i = 0; i < bucket->ub_cnt; i++)
4080 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
4084 * If we couldn't initialize the whole bucket, put the
4085 * rest back onto the freelist.
4087 if (i != bucket->ub_cnt) {
4088 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4089 bucket->ub_cnt - i);
4091 bzero(&bucket->ub_bucket[i],
4092 sizeof(void *) * (bucket->ub_cnt - i));
4098 cnt = bucket->ub_cnt;
4099 if (bucket->ub_cnt == 0) {
4100 bucket_free(zone, bucket, udata);
4101 counter_u64_add(zone->uz_fails, 1);
4105 if (zone->uz_max_items > 0 && cnt < maxbucket)
4106 zone_free_limit(zone, maxbucket - cnt);
4112 * Allocates a single item from a zone.
4115 * zone The zone to alloc for.
4116 * udata The data to be passed to the constructor.
4117 * domain The domain to allocate from or UMA_ANYDOMAIN.
4118 * flags M_WAITOK, M_NOWAIT, M_ZERO.
4121 * NULL if there is no memory and M_NOWAIT is set
4122 * An item if successful
4126 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4130 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4131 counter_u64_add(zone->uz_fails, 1);
4135 /* Avoid allocs targeting empty domains. */
4136 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4137 domain = UMA_ANYDOMAIN;
4139 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4143 * We have to call both the zone's init (not the keg's init)
4144 * and the zone's ctor. This is because the item is going from
4145 * a keg slab directly to the user, and the user is expecting it
4146 * to be both zone-init'd as well as zone-ctor'd.
4148 if (zone->uz_init != NULL) {
4149 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
4150 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4154 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4159 counter_u64_add(zone->uz_allocs, 1);
4160 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4161 zone->uz_name, zone);
4166 counter_u64_add(zone->uz_fails, 1);
4168 if (zone->uz_max_items > 0)
4169 zone_free_limit(zone, 1);
4170 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4171 zone->uz_name, zone);
4178 uma_zfree_smr(uma_zone_t zone, void *item)
4181 uma_cache_bucket_t bucket;
4182 int itemdomain, uz_flags;
4184 #ifdef UMA_ZALLOC_DEBUG
4185 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4186 ("uma_zfree_smr: called with non-SMR zone."));
4187 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4188 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4189 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4192 cache = &zone->uz_cpu[curcpu];
4193 uz_flags = cache_uz_flags(cache);
4196 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4197 itemdomain = item_domain(item);
4201 cache = &zone->uz_cpu[curcpu];
4202 /* SMR Zones must free to the free bucket. */
4203 bucket = &cache->uc_freebucket;
4205 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4206 PCPU_GET(domain) != itemdomain) {
4207 bucket = &cache->uc_crossbucket;
4210 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4211 cache_bucket_push(cache, bucket, item);
4215 } while (cache_free(zone, cache, NULL, item, itemdomain));
4219 * If nothing else caught this, we'll just do an internal free.
4221 zone_free_item(zone, item, NULL, SKIP_NONE);
4226 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4229 uma_cache_bucket_t bucket;
4230 int itemdomain, uz_flags;
4232 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4233 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4235 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4237 #ifdef UMA_ZALLOC_DEBUG
4238 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4239 ("uma_zfree_arg: called with SMR zone."));
4240 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4243 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4248 * We are accessing the per-cpu cache without a critical section to
4249 * fetch size and flags. This is acceptable, if we are preempted we
4250 * will simply read another cpu's line.
4252 cache = &zone->uz_cpu[curcpu];
4253 uz_flags = cache_uz_flags(cache);
4254 if (UMA_ALWAYS_CTORDTOR ||
4255 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4256 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4259 * The race here is acceptable. If we miss it we'll just have to wait
4260 * a little longer for the limits to be reset.
4262 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4263 if (atomic_load_32(&zone->uz_sleepers) > 0)
4268 * If possible, free to the per-CPU cache. There are two
4269 * requirements for safe access to the per-CPU cache: (1) the thread
4270 * accessing the cache must not be preempted or yield during access,
4271 * and (2) the thread must not migrate CPUs without switching which
4272 * cache it accesses. We rely on a critical section to prevent
4273 * preemption and migration. We release the critical section in
4274 * order to acquire the zone mutex if we are unable to free to the
4275 * current cache; when we re-acquire the critical section, we must
4276 * detect and handle migration if it has occurred.
4280 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4281 itemdomain = item_domain(item);
4285 cache = &zone->uz_cpu[curcpu];
4287 * Try to free into the allocbucket first to give LIFO
4288 * ordering for cache-hot datastructures. Spill over
4289 * into the freebucket if necessary. Alloc will swap
4290 * them if one runs dry.
4292 bucket = &cache->uc_allocbucket;
4294 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4295 PCPU_GET(domain) != itemdomain) {
4296 bucket = &cache->uc_crossbucket;
4299 if (bucket->ucb_cnt == bucket->ucb_entries &&
4300 cache->uc_freebucket.ucb_cnt <
4301 cache->uc_freebucket.ucb_entries)
4302 cache_bucket_swap(&cache->uc_freebucket,
4303 &cache->uc_allocbucket);
4304 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4305 cache_bucket_push(cache, bucket, item);
4309 } while (cache_free(zone, cache, udata, item, itemdomain));
4313 * If nothing else caught this, we'll just do an internal free.
4316 zone_free_item(zone, item, udata, SKIP_DTOR);
4321 * sort crossdomain free buckets to domain correct buckets and cache
4325 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4327 struct uma_bucketlist emptybuckets, fullbuckets;
4328 uma_zone_domain_t zdom;
4335 "uma_zfree: zone %s(%p) draining cross bucket %p",
4336 zone->uz_name, zone, bucket);
4339 * It is possible for buckets to arrive here out of order so we fetch
4340 * the current smr seq rather than accepting the bucket's.
4342 seq = SMR_SEQ_INVALID;
4343 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4344 seq = smr_advance(zone->uz_smr);
4347 * To avoid having ndomain * ndomain buckets for sorting we have a
4348 * lock on the current crossfree bucket. A full matrix with
4349 * per-domain locking could be used if necessary.
4351 STAILQ_INIT(&emptybuckets);
4352 STAILQ_INIT(&fullbuckets);
4353 ZONE_CROSS_LOCK(zone);
4354 for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4355 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4356 domain = item_domain(item);
4357 zdom = ZDOM_GET(zone, domain);
4358 if (zdom->uzd_cross == NULL) {
4359 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4360 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4361 zdom->uzd_cross = b;
4364 * Avoid allocating a bucket with the cross lock
4365 * held, since allocation can trigger a
4366 * cross-domain free and bucket zones may
4367 * allocate from each other.
4369 ZONE_CROSS_UNLOCK(zone);
4370 b = bucket_alloc(zone, udata, M_NOWAIT);
4373 ZONE_CROSS_LOCK(zone);
4374 if (zdom->uzd_cross != NULL) {
4375 STAILQ_INSERT_HEAD(&emptybuckets, b,
4378 zdom->uzd_cross = b;
4382 b = zdom->uzd_cross;
4383 b->ub_bucket[b->ub_cnt++] = item;
4385 if (b->ub_cnt == b->ub_entries) {
4386 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4387 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4388 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4389 zdom->uzd_cross = b;
4392 ZONE_CROSS_UNLOCK(zone);
4394 if (bucket->ub_cnt == 0)
4395 bucket->ub_seq = SMR_SEQ_INVALID;
4396 bucket_free(zone, bucket, udata);
4398 while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4399 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4400 bucket_free(zone, b, udata);
4402 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4403 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4404 domain = item_domain(b->ub_bucket[0]);
4405 zone_put_bucket(zone, domain, b, udata, true);
4411 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4412 int itemdomain, bool ws)
4417 * Buckets coming from the wrong domain will be entirely for the
4418 * only other domain on two domain systems. In this case we can
4419 * simply cache them. Otherwise we need to sort them back to
4422 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4423 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4424 zone_free_cross(zone, bucket, udata);
4430 * Attempt to save the bucket in the zone's domain bucket cache.
4433 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4434 zone->uz_name, zone, bucket);
4435 /* ub_cnt is pointing to the last free item */
4436 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4437 itemdomain = zone_domain_lowest(zone, itemdomain);
4438 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4442 * Populate a free or cross bucket for the current cpu cache. Free any
4443 * existing full bucket either to the zone cache or back to the slab layer.
4445 * Enters and returns in a critical section. false return indicates that
4446 * we can not satisfy this free in the cache layer. true indicates that
4447 * the caller should retry.
4449 static __noinline bool
4450 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4453 uma_cache_bucket_t cbucket;
4454 uma_bucket_t newbucket, bucket;
4456 CRITICAL_ASSERT(curthread);
4458 if (zone->uz_bucket_size == 0)
4461 cache = &zone->uz_cpu[curcpu];
4465 * FIRSTTOUCH domains need to free to the correct zdom. When
4466 * enabled this is the zdom of the item. The bucket is the
4467 * cross bucket if the current domain and itemdomain do not match.
4469 cbucket = &cache->uc_freebucket;
4471 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4472 if (PCPU_GET(domain) != itemdomain) {
4473 cbucket = &cache->uc_crossbucket;
4474 if (cbucket->ucb_cnt != 0)
4475 counter_u64_add(zone->uz_xdomain,
4480 bucket = cache_bucket_unload(cbucket);
4481 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4482 ("cache_free: Entered with non-full free bucket."));
4484 /* We are no longer associated with this CPU. */
4488 * Don't let SMR zones operate without a free bucket. Force
4489 * a synchronize and re-use this one. We will only degrade
4490 * to a synchronize every bucket_size items rather than every
4491 * item if we fail to allocate a bucket.
4493 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4495 bucket->ub_seq = smr_advance(zone->uz_smr);
4496 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4497 if (newbucket == NULL && bucket != NULL) {
4498 bucket_drain(zone, bucket);
4502 } else if (!bucketdisable)
4503 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4506 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4509 if ((bucket = newbucket) == NULL)
4511 cache = &zone->uz_cpu[curcpu];
4514 * Check to see if we should be populating the cross bucket. If it
4515 * is already populated we will fall through and attempt to populate
4518 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4519 if (PCPU_GET(domain) != itemdomain &&
4520 cache->uc_crossbucket.ucb_bucket == NULL) {
4521 cache_bucket_load_cross(cache, bucket);
4527 * We may have lost the race to fill the bucket or switched CPUs.
4529 if (cache->uc_freebucket.ucb_bucket != NULL) {
4531 bucket_free(zone, bucket, udata);
4534 cache_bucket_load_free(cache, bucket);
4540 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4547 KEG_LOCK_ASSERT(keg, slab->us_domain);
4549 /* Do we need to remove from any lists? */
4550 dom = &keg->uk_domain[slab->us_domain];
4551 if (slab->us_freecount + 1 == keg->uk_ipers) {
4552 LIST_REMOVE(slab, us_link);
4553 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4554 dom->ud_free_slabs++;
4555 } else if (slab->us_freecount == 0) {
4556 LIST_REMOVE(slab, us_link);
4557 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4560 /* Slab management. */
4561 freei = slab_item_index(slab, keg, item);
4562 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4563 slab->us_freecount++;
4565 /* Keg statistics. */
4566 dom->ud_free_items++;
4570 zone_release(void *arg, void **bucket, int cnt)
4583 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4584 lock = KEG_LOCK(keg, 0);
4585 for (i = 0; i < cnt; i++) {
4587 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4588 slab = vtoslab((vm_offset_t)item);
4590 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4591 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4592 slab = hash_sfind(&keg->uk_hash, mem);
4594 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4596 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4599 lock = KEG_LOCK(keg, slab->us_domain);
4601 slab_free_item(zone, slab, item);
4608 * Frees a single item to any zone.
4611 * zone The zone to free to
4612 * item The item we're freeing
4613 * udata User supplied data for the dtor
4614 * skip Skip dtors and finis
4616 static __noinline void
4617 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4621 * If a free is sent directly to an SMR zone we have to
4622 * synchronize immediately because the item can instantly
4623 * be reallocated. This should only happen in degenerate
4624 * cases when no memory is available for per-cpu caches.
4626 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4627 smr_synchronize(zone->uz_smr);
4629 item_dtor(zone, item, zone->uz_size, udata, skip);
4631 if (skip < SKIP_FINI && zone->uz_fini)
4632 zone->uz_fini(item, zone->uz_size);
4634 zone->uz_release(zone->uz_arg, &item, 1);
4636 if (skip & SKIP_CNT)
4639 counter_u64_add(zone->uz_frees, 1);
4641 if (zone->uz_max_items > 0)
4642 zone_free_limit(zone, 1);
4647 uma_zone_set_max(uma_zone_t zone, int nitems)
4651 * If the limit is small, we may need to constrain the maximum per-CPU
4652 * cache size, or disable caching entirely.
4654 uma_zone_set_maxcache(zone, nitems);
4657 * XXX This can misbehave if the zone has any allocations with
4658 * no limit and a limit is imposed. There is currently no
4659 * way to clear a limit.
4662 zone->uz_max_items = nitems;
4663 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4664 zone_update_caches(zone);
4665 /* We may need to wake waiters. */
4666 wakeup(&zone->uz_max_items);
4674 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4676 int bpcpu, bpdom, bsize, nb;
4681 * Compute a lower bound on the number of items that may be cached in
4682 * the zone. Each CPU gets at least two buckets, and for cross-domain
4683 * frees we use an additional bucket per CPU and per domain. Select the
4684 * largest bucket size that does not exceed half of the requested limit,
4685 * with the left over space given to the full bucket cache.
4690 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4695 nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4696 bsize = nitems / nb / 2;
4697 if (bsize > BUCKET_MAX)
4699 else if (bsize == 0 && nitems / nb > 0)
4701 zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
4702 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4703 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4704 zone->uz_bucket_max = nitems - nb * bsize;
4710 uma_zone_get_max(uma_zone_t zone)
4714 nitems = atomic_load_64(&zone->uz_max_items);
4721 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4724 ZONE_ASSERT_COLD(zone);
4725 zone->uz_warning = warning;
4730 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4733 ZONE_ASSERT_COLD(zone);
4734 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4739 uma_zone_get_cur(uma_zone_t zone)
4745 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4746 nitems = counter_u64_fetch(zone->uz_allocs) -
4747 counter_u64_fetch(zone->uz_frees);
4749 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4750 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4752 return (nitems < 0 ? 0 : nitems);
4756 uma_zone_get_allocs(uma_zone_t zone)
4762 if (zone->uz_allocs != EARLY_COUNTER)
4763 nitems = counter_u64_fetch(zone->uz_allocs);
4765 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4771 uma_zone_get_frees(uma_zone_t zone)
4777 if (zone->uz_frees != EARLY_COUNTER)
4778 nitems = counter_u64_fetch(zone->uz_frees);
4780 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4786 /* Used only for KEG_ASSERT_COLD(). */
4788 uma_keg_get_allocs(uma_keg_t keg)
4794 LIST_FOREACH(z, &keg->uk_zones, uz_link)
4795 nitems += uma_zone_get_allocs(z);
4803 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4808 KEG_ASSERT_COLD(keg);
4809 keg->uk_init = uminit;
4814 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4819 KEG_ASSERT_COLD(keg);
4820 keg->uk_fini = fini;
4825 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4828 ZONE_ASSERT_COLD(zone);
4829 zone->uz_init = zinit;
4834 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4837 ZONE_ASSERT_COLD(zone);
4838 zone->uz_fini = zfini;
4843 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4848 KEG_ASSERT_COLD(keg);
4849 keg->uk_freef = freef;
4854 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4859 KEG_ASSERT_COLD(keg);
4860 keg->uk_allocf = allocf;
4865 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4868 ZONE_ASSERT_COLD(zone);
4870 KASSERT(smr != NULL, ("Got NULL smr"));
4871 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4872 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
4873 zone->uz_flags |= UMA_ZONE_SMR;
4875 zone_update_caches(zone);
4879 uma_zone_get_smr(uma_zone_t zone)
4882 return (zone->uz_smr);
4887 uma_zone_reserve(uma_zone_t zone, int items)
4892 KEG_ASSERT_COLD(keg);
4893 keg->uk_reserve = items;
4898 uma_zone_reserve_kva(uma_zone_t zone, int count)
4905 KEG_ASSERT_COLD(keg);
4906 ZONE_ASSERT_COLD(zone);
4908 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4910 #ifdef UMA_MD_SMALL_ALLOC
4911 if (keg->uk_ppera > 1) {
4915 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4921 MPASS(keg->uk_kva == 0);
4924 zone->uz_max_items = pages * keg->uk_ipers;
4925 #ifdef UMA_MD_SMALL_ALLOC
4926 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4928 keg->uk_allocf = noobj_alloc;
4930 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4931 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4932 zone_update_caches(zone);
4939 uma_prealloc(uma_zone_t zone, int items)
4941 struct vm_domainset_iter di;
4945 int aflags, domain, slabs;
4948 slabs = howmany(items, keg->uk_ipers);
4949 while (slabs-- > 0) {
4951 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4954 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4957 dom = &keg->uk_domain[slab->us_domain];
4959 * keg_alloc_slab() always returns a slab on the
4962 LIST_REMOVE(slab, us_link);
4963 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4965 dom->ud_free_slabs++;
4966 KEG_UNLOCK(keg, slab->us_domain);
4969 if (vm_domainset_iter_policy(&di, &domain) != 0)
4970 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
4976 * Returns a snapshot of memory consumption in bytes.
4979 uma_zone_memory(uma_zone_t zone)
4985 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
4986 for (i = 0; i < vm_ndomains; i++)
4987 sz += ZDOM_GET(zone, i)->uzd_nitems;
4988 return (sz * zone->uz_size);
4990 for (i = 0; i < vm_ndomains; i++)
4991 sz += zone->uz_keg->uk_domain[i].ud_pages;
4993 return (sz * PAGE_SIZE);
4998 uma_reclaim(int req)
5000 uma_reclaim_domain(req, UMA_ANYDOMAIN);
5004 uma_reclaim_domain(int req, int domain)
5010 arg = (void *)(uintptr_t)domain;
5011 sx_slock(&uma_reclaim_lock);
5013 case UMA_RECLAIM_TRIM:
5014 zone_foreach(zone_trim, arg);
5016 case UMA_RECLAIM_DRAIN:
5017 zone_foreach(zone_drain, arg);
5019 case UMA_RECLAIM_DRAIN_CPU:
5020 zone_foreach(zone_drain, arg);
5021 pcpu_cache_drain_safe(NULL);
5022 zone_foreach(zone_drain, arg);
5025 panic("unhandled reclamation request %d", req);
5029 * Some slabs may have been freed but this zone will be visited early
5030 * we visit again so that we can free pages that are empty once other
5031 * zones are drained. We have to do the same for buckets.
5033 zone_drain(slabzones[0], arg);
5034 zone_drain(slabzones[1], arg);
5035 bucket_zone_drain(domain);
5036 sx_sunlock(&uma_reclaim_lock);
5039 static volatile int uma_reclaim_needed;
5042 uma_reclaim_wakeup(void)
5045 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5046 wakeup(uma_reclaim);
5050 uma_reclaim_worker(void *arg __unused)
5054 sx_xlock(&uma_reclaim_lock);
5055 while (atomic_load_int(&uma_reclaim_needed) == 0)
5056 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5058 sx_xunlock(&uma_reclaim_lock);
5059 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5060 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5061 atomic_store_int(&uma_reclaim_needed, 0);
5062 /* Don't fire more than once per-second. */
5063 pause("umarclslp", hz);
5069 uma_zone_reclaim(uma_zone_t zone, int req)
5071 uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5075 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5079 arg = (void *)(uintptr_t)domain;
5081 case UMA_RECLAIM_TRIM:
5082 zone_trim(zone, arg);
5084 case UMA_RECLAIM_DRAIN:
5085 zone_drain(zone, arg);
5087 case UMA_RECLAIM_DRAIN_CPU:
5088 pcpu_cache_drain_safe(zone);
5089 zone_drain(zone, arg);
5092 panic("unhandled reclamation request %d", req);
5098 uma_zone_exhausted(uma_zone_t zone)
5101 return (atomic_load_32(&zone->uz_sleepers) > 0);
5108 return (uma_kmem_limit);
5112 uma_set_limit(unsigned long limit)
5115 uma_kmem_limit = limit;
5122 return (atomic_load_long(&uma_kmem_total));
5129 return (uma_kmem_limit - uma_size());
5134 * Generate statistics across both the zone and its per-cpu cache's. Return
5135 * desired statistics if the pointer is non-NULL for that statistic.
5137 * Note: does not update the zone statistics, as it can't safely clear the
5138 * per-CPU cache statistic.
5142 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5143 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5146 uint64_t allocs, frees, sleeps, xdomain;
5149 allocs = frees = sleeps = xdomain = 0;
5152 cache = &z->uz_cpu[cpu];
5153 cachefree += cache->uc_allocbucket.ucb_cnt;
5154 cachefree += cache->uc_freebucket.ucb_cnt;
5155 xdomain += cache->uc_crossbucket.ucb_cnt;
5156 cachefree += cache->uc_crossbucket.ucb_cnt;
5157 allocs += cache->uc_allocs;
5158 frees += cache->uc_frees;
5160 allocs += counter_u64_fetch(z->uz_allocs);
5161 frees += counter_u64_fetch(z->uz_frees);
5162 xdomain += counter_u64_fetch(z->uz_xdomain);
5163 sleeps += z->uz_sleeps;
5164 if (cachefreep != NULL)
5165 *cachefreep = cachefree;
5166 if (allocsp != NULL)
5170 if (sleepsp != NULL)
5172 if (xdomainp != NULL)
5173 *xdomainp = xdomain;
5178 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5185 rw_rlock(&uma_rwlock);
5186 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5187 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5190 LIST_FOREACH(z, &uma_cachezones, uz_link)
5193 rw_runlock(&uma_rwlock);
5194 return (sysctl_handle_int(oidp, &count, 0, req));
5198 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5199 struct uma_percpu_stat *ups, bool internal)
5201 uma_zone_domain_t zdom;
5205 for (i = 0; i < vm_ndomains; i++) {
5206 zdom = ZDOM_GET(z, i);
5207 uth->uth_zone_free += zdom->uzd_nitems;
5209 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5210 uth->uth_frees = counter_u64_fetch(z->uz_frees);
5211 uth->uth_fails = counter_u64_fetch(z->uz_fails);
5212 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5213 uth->uth_sleeps = z->uz_sleeps;
5215 for (i = 0; i < mp_maxid + 1; i++) {
5216 bzero(&ups[i], sizeof(*ups));
5217 if (internal || CPU_ABSENT(i))
5219 cache = &z->uz_cpu[i];
5220 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5221 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5222 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5223 ups[i].ups_allocs = cache->uc_allocs;
5224 ups[i].ups_frees = cache->uc_frees;
5229 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5231 struct uma_stream_header ush;
5232 struct uma_type_header uth;
5233 struct uma_percpu_stat *ups;
5238 uint32_t kfree, pages;
5239 int count, error, i;
5241 error = sysctl_wire_old_buffer(req, 0);
5244 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5245 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5246 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5249 rw_rlock(&uma_rwlock);
5250 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5251 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5255 LIST_FOREACH(z, &uma_cachezones, uz_link)
5259 * Insert stream header.
5261 bzero(&ush, sizeof(ush));
5262 ush.ush_version = UMA_STREAM_VERSION;
5263 ush.ush_maxcpus = (mp_maxid + 1);
5264 ush.ush_count = count;
5265 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5267 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5269 for (i = 0; i < vm_ndomains; i++) {
5270 kfree += kz->uk_domain[i].ud_free_items;
5271 pages += kz->uk_domain[i].ud_pages;
5273 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5274 bzero(&uth, sizeof(uth));
5275 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5276 uth.uth_align = kz->uk_align;
5277 uth.uth_size = kz->uk_size;
5278 uth.uth_rsize = kz->uk_rsize;
5279 if (z->uz_max_items > 0) {
5280 items = UZ_ITEMS_COUNT(z->uz_items);
5281 uth.uth_pages = (items / kz->uk_ipers) *
5284 uth.uth_pages = pages;
5285 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5287 uth.uth_limit = z->uz_max_items;
5288 uth.uth_keg_free = kfree;
5291 * A zone is secondary is it is not the first entry
5292 * on the keg's zone list.
5294 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5295 (LIST_FIRST(&kz->uk_zones) != z))
5296 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5297 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5298 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5299 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5300 for (i = 0; i < mp_maxid + 1; i++)
5301 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5304 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5305 bzero(&uth, sizeof(uth));
5306 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5307 uth.uth_size = z->uz_size;
5308 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5309 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5310 for (i = 0; i < mp_maxid + 1; i++)
5311 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5314 rw_runlock(&uma_rwlock);
5315 error = sbuf_finish(&sbuf);
5322 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5324 uma_zone_t zone = *(uma_zone_t *)arg1;
5327 max = uma_zone_get_max(zone);
5328 error = sysctl_handle_int(oidp, &max, 0, req);
5329 if (error || !req->newptr)
5332 uma_zone_set_max(zone, max);
5338 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5344 * Some callers want to add sysctls for global zones that
5345 * may not yet exist so they pass a pointer to a pointer.
5348 zone = *(uma_zone_t *)arg1;
5351 cur = uma_zone_get_cur(zone);
5352 return (sysctl_handle_int(oidp, &cur, 0, req));
5356 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5358 uma_zone_t zone = arg1;
5361 cur = uma_zone_get_allocs(zone);
5362 return (sysctl_handle_64(oidp, &cur, 0, req));
5366 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5368 uma_zone_t zone = arg1;
5371 cur = uma_zone_get_frees(zone);
5372 return (sysctl_handle_64(oidp, &cur, 0, req));
5376 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5379 uma_zone_t zone = arg1;
5382 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5383 if (zone->uz_flags != 0)
5384 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5386 sbuf_printf(&sbuf, "0");
5387 error = sbuf_finish(&sbuf);
5394 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5396 uma_keg_t keg = arg1;
5397 int avail, effpct, total;
5399 total = keg->uk_ppera * PAGE_SIZE;
5400 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5401 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5403 * We consider the client's requested size and alignment here, not the
5404 * real size determination uk_rsize, because we also adjust the real
5405 * size for internal implementation reasons (max bitset size).
5407 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5408 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5409 avail *= mp_maxid + 1;
5410 effpct = 100 * avail / total;
5411 return (sysctl_handle_int(oidp, &effpct, 0, req));
5415 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5417 uma_zone_t zone = arg1;
5420 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5421 return (sysctl_handle_64(oidp, &cur, 0, req));
5426 uma_dbg_getslab(uma_zone_t zone, void *item)
5433 * It is safe to return the slab here even though the
5434 * zone is unlocked because the item's allocation state
5435 * essentially holds a reference.
5437 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5438 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5440 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5441 return (vtoslab((vm_offset_t)mem));
5443 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5444 return ((uma_slab_t)(mem + keg->uk_pgoff));
5446 slab = hash_sfind(&keg->uk_hash, mem);
5453 uma_dbg_zskip(uma_zone_t zone, void *mem)
5456 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5459 return (uma_dbg_kskip(zone->uz_keg, mem));
5463 uma_dbg_kskip(uma_keg_t keg, void *mem)
5467 if (dbg_divisor == 0)
5470 if (dbg_divisor == 1)
5473 idx = (uintptr_t)mem >> PAGE_SHIFT;
5474 if (keg->uk_ipers > 1) {
5475 idx *= keg->uk_ipers;
5476 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5479 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5480 counter_u64_add(uma_skip_cnt, 1);
5483 counter_u64_add(uma_dbg_cnt, 1);
5489 * Set up the slab's freei data such that uma_dbg_free can function.
5493 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5499 slab = uma_dbg_getslab(zone, item);
5501 panic("uma: item %p did not belong to zone %s",
5502 item, zone->uz_name);
5505 freei = slab_item_index(slab, keg, item);
5507 if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5508 slab_dbg_bits(slab, keg)))
5509 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5510 item, zone, zone->uz_name, slab, freei);
5514 * Verifies freed addresses. Checks for alignment, valid slab membership
5515 * and duplicate frees.
5519 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5525 slab = uma_dbg_getslab(zone, item);
5527 panic("uma: Freed item %p did not belong to zone %s",
5528 item, zone->uz_name);
5531 freei = slab_item_index(slab, keg, item);
5533 if (freei >= keg->uk_ipers)
5534 panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5535 item, zone, zone->uz_name, slab, freei);
5537 if (slab_item(slab, keg, freei) != item)
5538 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5539 item, zone, zone->uz_name, slab, freei);
5541 if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5542 slab_dbg_bits(slab, keg)))
5543 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5544 item, zone, zone->uz_name, slab, freei);
5546 #endif /* INVARIANTS */
5550 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5551 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5556 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5557 *allocs = counter_u64_fetch(z->uz_allocs);
5558 frees = counter_u64_fetch(z->uz_frees);
5559 *sleeps = z->uz_sleeps;
5563 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5565 for (i = 0; i < vm_ndomains; i++) {
5566 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5567 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5568 (LIST_FIRST(&kz->uk_zones) != z)))
5569 *cachefree += kz->uk_domain[i].ud_free_items;
5571 *used = *allocs - frees;
5572 return (((int64_t)*used + *cachefree) * kz->uk_size);
5575 DB_SHOW_COMMAND(uma, db_show_uma)
5577 const char *fmt_hdr, *fmt_entry;
5580 uint64_t allocs, used, sleeps, xdomain;
5582 /* variables for sorting */
5584 uma_zone_t cur_zone, last_zone;
5585 int64_t cur_size, last_size, size;
5588 /* /i option produces machine-parseable CSV output */
5589 if (modif[0] == 'i') {
5590 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5591 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5593 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5594 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5597 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5598 "Sleeps", "Bucket", "Total Mem", "XFree");
5600 /* Sort the zones with largest size first. */
5602 last_size = INT64_MAX;
5607 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5608 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5610 * In the case of size ties, print out zones
5611 * in the order they are encountered. That is,
5612 * when we encounter the most recently output
5613 * zone, we have already printed all preceding
5614 * ties, and we must print all following ties.
5616 if (z == last_zone) {
5620 size = get_uma_stats(kz, z, &allocs, &used,
5621 &sleeps, &cachefree, &xdomain);
5622 if (size > cur_size && size < last_size + ties)
5630 if (cur_zone == NULL)
5633 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5634 &sleeps, &cachefree, &xdomain);
5635 db_printf(fmt_entry, cur_zone->uz_name,
5636 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5637 (uintmax_t)allocs, (uintmax_t)sleeps,
5638 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5643 last_zone = cur_zone;
5644 last_size = cur_size;
5648 DB_SHOW_COMMAND(umacache, db_show_umacache)
5651 uint64_t allocs, frees;
5655 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5656 "Requests", "Bucket");
5657 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5658 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5659 for (i = 0; i < vm_ndomains; i++)
5660 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5661 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5662 z->uz_name, (uintmax_t)z->uz_size,
5663 (intmax_t)(allocs - frees), cachefree,
5664 (uintmax_t)allocs, z->uz_bucket_size);