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>
62 #include <sys/bitset.h>
63 #include <sys/domainset.h>
64 #include <sys/eventhandler.h>
65 #include <sys/kernel.h>
66 #include <sys/types.h>
67 #include <sys/limits.h>
68 #include <sys/queue.h>
69 #include <sys/malloc.h>
72 #include <sys/sysctl.h>
73 #include <sys/mutex.h>
75 #include <sys/random.h>
76 #include <sys/rwlock.h>
78 #include <sys/sched.h>
79 #include <sys/sleepqueue.h>
82 #include <sys/taskqueue.h>
83 #include <sys/vmmeter.h>
86 #include <vm/vm_param.h>
87 #include <vm/vm_domainset.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_phys.h>
92 #include <vm/vm_pagequeue.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_dumpset.h>
98 #include <vm/uma_int.h>
99 #include <vm/uma_dbg.h>
103 #ifdef DEBUG_MEMGUARD
104 #include <vm/memguard.h>
107 #include <machine/md_var.h>
110 #define UMA_ALWAYS_CTORDTOR 1
112 #define UMA_ALWAYS_CTORDTOR 0
116 * This is the zone and keg from which all zones are spawned.
118 static uma_zone_t kegs;
119 static uma_zone_t zones;
122 * On INVARIANTS builds, the slab contains a second bitset of the same size,
123 * "dbg_bits", which is laid out immediately after us_free.
126 #define SLAB_BITSETS 2
128 #define SLAB_BITSETS 1
132 * These are the two zones from which all offpage uma_slab_ts are allocated.
134 * One zone is for slab headers that can represent a larger number of items,
135 * making the slabs themselves more efficient, and the other zone is for
136 * headers that are smaller and represent fewer items, making the headers more
139 #define SLABZONE_SIZE(setsize) \
140 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
141 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16)
142 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE
143 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE)
144 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE)
145 static uma_zone_t slabzones[2];
148 * The initial hash tables come out of this zone so they can be allocated
149 * prior to malloc coming up.
151 static uma_zone_t hashzone;
153 /* The boot-time adjusted value for cache line alignment. */
154 int uma_align_cache = 64 - 1;
156 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
157 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
160 * Are we allowed to allocate buckets?
162 static int bucketdisable = 1;
164 /* Linked list of all kegs in the system */
165 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
167 /* Linked list of all cache-only zones in the system */
168 static LIST_HEAD(,uma_zone) uma_cachezones =
169 LIST_HEAD_INITIALIZER(uma_cachezones);
172 * Mutex for global lists: uma_kegs, uma_cachezones, and the per-keg list of
175 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
177 static struct sx uma_reclaim_lock;
180 * First available virual address for boot time allocations.
182 static vm_offset_t bootstart;
183 static vm_offset_t bootmem;
186 * kmem soft limit, initialized by uma_set_limit(). Ensure that early
187 * allocations don't trigger a wakeup of the reclaim thread.
189 unsigned long uma_kmem_limit = LONG_MAX;
190 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
191 "UMA kernel memory soft limit");
192 unsigned long uma_kmem_total;
193 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
194 "UMA kernel memory usage");
196 /* Is the VM done starting up? */
203 } booted = BOOT_COLD;
206 * This is the handle used to schedule events that need to happen
207 * outside of the allocation fast path.
209 static struct callout uma_callout;
210 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
213 * This structure is passed as the zone ctor arg so that I don't have to create
214 * a special allocation function just for zones.
216 struct uma_zctor_args {
231 struct uma_kctor_args {
240 struct uma_bucket_zone {
242 const char *ubz_name;
243 int ubz_entries; /* Number of items it can hold. */
244 int ubz_maxsize; /* Maximum allocation size per-item. */
248 * Compute the actual number of bucket entries to pack them in power
249 * of two sizes for more efficient space utilization.
251 #define BUCKET_SIZE(n) \
252 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
254 #define BUCKET_MAX BUCKET_SIZE(256)
256 struct uma_bucket_zone bucket_zones[] = {
257 /* Literal bucket sizes. */
258 { NULL, "2 Bucket", 2, 4096 },
259 { NULL, "4 Bucket", 4, 3072 },
260 { NULL, "8 Bucket", 8, 2048 },
261 { NULL, "16 Bucket", 16, 1024 },
262 /* Rounded down power of 2 sizes for efficiency. */
263 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
264 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
265 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
266 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
271 * Flags and enumerations to be passed to internal functions.
275 SKIP_CNT = 0x00000001,
276 SKIP_DTOR = 0x00010000,
277 SKIP_FINI = 0x00020000,
282 void uma_startup1(vm_offset_t);
283 void uma_startup2(void);
285 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
286 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
287 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
288 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
289 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
290 static void page_free(void *, vm_size_t, uint8_t);
291 static void pcpu_page_free(void *, vm_size_t, uint8_t);
292 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
293 static void cache_drain(uma_zone_t);
294 static void bucket_drain(uma_zone_t, uma_bucket_t);
295 static void bucket_cache_reclaim(uma_zone_t zone, bool, int);
296 static bool bucket_cache_reclaim_domain(uma_zone_t, bool, bool, int);
297 static int keg_ctor(void *, int, void *, int);
298 static void keg_dtor(void *, int, void *);
299 static void keg_drain(uma_keg_t keg, int domain);
300 static int zone_ctor(void *, int, void *, int);
301 static void zone_dtor(void *, int, void *);
302 static inline void item_dtor(uma_zone_t zone, void *item, int size,
303 void *udata, enum zfreeskip skip);
304 static int zero_init(void *, int, int);
305 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
306 int itemdomain, bool ws);
307 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
308 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
309 static void zone_timeout(uma_zone_t zone, void *);
310 static int hash_alloc(struct uma_hash *, u_int);
311 static int hash_expand(struct uma_hash *, struct uma_hash *);
312 static void hash_free(struct uma_hash *hash);
313 static void uma_timeout(void *);
314 static void uma_shutdown(void);
315 static void *zone_alloc_item(uma_zone_t, void *, int, int);
316 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
317 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
318 static void zone_free_limit(uma_zone_t zone, int count);
319 static void bucket_enable(void);
320 static void bucket_init(void);
321 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
322 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
323 static void bucket_zone_drain(int domain);
324 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
325 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
326 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
327 static size_t slab_sizeof(int nitems);
328 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
329 uma_fini fini, int align, uint32_t flags);
330 static int zone_import(void *, void **, int, int, int);
331 static void zone_release(void *, void **, int);
332 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
333 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
335 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
336 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
337 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
338 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
339 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
340 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
341 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
343 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
345 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
346 "Memory allocation debugging");
349 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
350 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
352 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
353 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
354 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
355 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
357 static u_int dbg_divisor = 1;
358 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
359 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
360 "Debug & thrash every this item in memory allocator");
362 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
363 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
364 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
365 &uma_dbg_cnt, "memory items debugged");
366 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
367 &uma_skip_cnt, "memory items skipped, not debugged");
370 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
371 "Universal Memory Allocator");
373 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
374 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
376 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
377 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
379 static int zone_warnings = 1;
380 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
381 "Warn when UMA zones becomes full");
383 static int multipage_slabs = 1;
384 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
385 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
386 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
387 "UMA may choose larger slab sizes for better efficiency");
390 * Select the slab zone for an offpage slab with the given maximum item count.
392 static inline uma_zone_t
396 return (slabzones[ipers > SLABZONE0_SETSIZE]);
400 * This routine checks to see whether or not it's safe to enable buckets.
406 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
407 bucketdisable = vm_page_count_min();
411 * Initialize bucket_zones, the array of zones of buckets of various sizes.
413 * For each zone, calculate the memory required for each bucket, consisting
414 * of the header and an array of pointers.
419 struct uma_bucket_zone *ubz;
422 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
423 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
424 size += sizeof(void *) * ubz->ubz_entries;
425 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
426 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
427 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
428 UMA_ZONE_FIRSTTOUCH);
433 * Given a desired number of entries for a bucket, return the zone from which
434 * to allocate the bucket.
436 static struct uma_bucket_zone *
437 bucket_zone_lookup(int entries)
439 struct uma_bucket_zone *ubz;
441 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
442 if (ubz->ubz_entries >= entries)
449 bucket_select(int size)
451 struct uma_bucket_zone *ubz;
453 ubz = &bucket_zones[0];
454 if (size > ubz->ubz_maxsize)
455 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
457 for (; ubz->ubz_entries != 0; ubz++)
458 if (ubz->ubz_maxsize < size)
461 return (ubz->ubz_entries);
465 bucket_alloc(uma_zone_t zone, void *udata, int flags)
467 struct uma_bucket_zone *ubz;
471 * Don't allocate buckets early in boot.
473 if (__predict_false(booted < BOOT_KVA))
477 * To limit bucket recursion we store the original zone flags
478 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
479 * NOVM flag to persist even through deep recursions. We also
480 * store ZFLAG_BUCKET once we have recursed attempting to allocate
481 * a bucket for a bucket zone so we do not allow infinite bucket
482 * recursion. This cookie will even persist to frees of unused
483 * buckets via the allocation path or bucket allocations in the
486 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
487 udata = (void *)(uintptr_t)zone->uz_flags;
489 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
491 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
493 if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
495 ubz = bucket_zone_lookup(atomic_load_16(&zone->uz_bucket_size));
496 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
498 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
501 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
504 bucket->ub_entries = min(ubz->ubz_entries,
505 zone->uz_bucket_size_max);
506 bucket->ub_seq = SMR_SEQ_INVALID;
507 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
508 zone->uz_name, zone, bucket);
515 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
517 struct uma_bucket_zone *ubz;
519 if (bucket->ub_cnt != 0)
520 bucket_drain(zone, bucket);
522 KASSERT(bucket->ub_cnt == 0,
523 ("bucket_free: Freeing a non free bucket."));
524 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
525 ("bucket_free: Freeing an SMR bucket."));
526 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
527 udata = (void *)(uintptr_t)zone->uz_flags;
528 ubz = bucket_zone_lookup(bucket->ub_entries);
529 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
533 bucket_zone_drain(int domain)
535 struct uma_bucket_zone *ubz;
537 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
538 uma_zone_reclaim_domain(ubz->ubz_zone, UMA_RECLAIM_DRAIN,
543 _Static_assert(UMA_SMALLEST_UNIT % KASAN_SHADOW_SCALE == 0,
544 "Base UMA allocation size not a multiple of the KASAN scale factor");
547 kasan_mark_item_valid(uma_zone_t zone, void *item)
553 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
557 rsz = roundup2(sz, KASAN_SHADOW_SCALE);
558 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
559 kasan_mark(item, sz, rsz, KASAN_GENERIC_REDZONE);
561 pcpu_item = zpcpu_base_to_offset(item);
562 for (i = 0; i <= mp_maxid; i++)
563 kasan_mark(zpcpu_get_cpu(pcpu_item, i), sz, rsz,
564 KASAN_GENERIC_REDZONE);
569 kasan_mark_item_invalid(uma_zone_t zone, void *item)
575 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
578 sz = roundup2(zone->uz_size, KASAN_SHADOW_SCALE);
579 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
580 kasan_mark(item, 0, sz, KASAN_UMA_FREED);
582 pcpu_item = zpcpu_base_to_offset(item);
583 for (i = 0; i <= mp_maxid; i++)
584 kasan_mark(zpcpu_get_cpu(pcpu_item, i), 0, sz,
590 kasan_mark_slab_valid(uma_keg_t keg, void *mem)
594 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
595 sz = keg->uk_ppera * PAGE_SIZE;
596 kasan_mark(mem, sz, sz, 0);
601 kasan_mark_slab_invalid(uma_keg_t keg, void *mem)
605 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
606 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
607 sz = keg->uk_ppera * PAGE_SIZE;
610 kasan_mark(mem, 0, sz, KASAN_UMA_FREED);
615 kasan_mark_item_valid(uma_zone_t zone __unused, void *item __unused)
620 kasan_mark_item_invalid(uma_zone_t zone __unused, void *item __unused)
625 kasan_mark_slab_valid(uma_keg_t keg __unused, void *mem __unused)
630 kasan_mark_slab_invalid(uma_keg_t keg __unused, void *mem __unused)
636 * Acquire the domain lock and record contention.
638 static uma_zone_domain_t
639 zone_domain_lock(uma_zone_t zone, int domain)
641 uma_zone_domain_t zdom;
644 zdom = ZDOM_GET(zone, domain);
646 if (ZDOM_OWNED(zdom))
649 /* This is unsynchronized. The counter does not need to be precise. */
650 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
651 zone->uz_bucket_size++;
656 * Search for the domain with the least cached items and return it if it
657 * is out of balance with the preferred domain.
659 static __noinline int
660 zone_domain_lowest(uma_zone_t zone, int pref)
662 long least, nitems, prefitems;
666 prefitems = least = LONG_MAX;
668 for (i = 0; i < vm_ndomains; i++) {
669 nitems = ZDOM_GET(zone, i)->uzd_nitems;
670 if (nitems < least) {
677 if (prefitems < least * 2)
684 * Search for the domain with the most cached items and return it or the
685 * preferred domain if it has enough to proceed.
687 static __noinline int
688 zone_domain_highest(uma_zone_t zone, int pref)
694 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
699 for (i = 0; i < vm_ndomains; i++) {
700 nitems = ZDOM_GET(zone, i)->uzd_nitems;
711 * Set the maximum imax value.
714 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
718 old = zdom->uzd_imax;
722 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
725 * We are at new maximum, so do the last WSS update for the old
726 * bimin and prepare to measure next allocation batch.
728 if (zdom->uzd_wss < old - zdom->uzd_bimin)
729 zdom->uzd_wss = old - zdom->uzd_bimin;
730 zdom->uzd_bimin = nitems;
734 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
735 * zone's caches. If a bucket is found the zone is not locked on return.
738 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
745 ZDOM_LOCK_ASSERT(zdom);
747 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
750 /* SMR Buckets can not be re-used until readers expire. */
751 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
752 bucket->ub_seq != SMR_SEQ_INVALID) {
753 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
755 bucket->ub_seq = SMR_SEQ_INVALID;
756 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
757 if (STAILQ_NEXT(bucket, ub_link) != NULL)
758 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
760 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
762 KASSERT(zdom->uzd_nitems >= bucket->ub_cnt,
763 ("%s: item count underflow (%ld, %d)",
764 __func__, zdom->uzd_nitems, bucket->ub_cnt));
765 KASSERT(bucket->ub_cnt > 0,
766 ("%s: empty bucket in bucket cache", __func__));
767 zdom->uzd_nitems -= bucket->ub_cnt;
771 * Shift the bounds of the current WSS interval to avoid
772 * perturbing the estimates.
774 cnt = lmin(zdom->uzd_bimin, bucket->ub_cnt);
775 atomic_subtract_long(&zdom->uzd_imax, cnt);
776 zdom->uzd_bimin -= cnt;
777 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
778 if (zdom->uzd_limin >= bucket->ub_cnt) {
779 zdom->uzd_limin -= bucket->ub_cnt;
784 } else if (zdom->uzd_bimin > zdom->uzd_nitems) {
785 zdom->uzd_bimin = zdom->uzd_nitems;
786 if (zdom->uzd_imin > zdom->uzd_nitems)
787 zdom->uzd_imin = zdom->uzd_nitems;
792 for (i = 0; i < bucket->ub_cnt; i++)
793 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
800 * Insert a full bucket into the specified cache. The "ws" parameter indicates
801 * whether the bucket's contents should be counted as part of the zone's working
802 * set. The bucket may be freed if it exceeds the bucket limit.
805 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
808 uma_zone_domain_t zdom;
810 /* We don't cache empty buckets. This can happen after a reclaim. */
811 if (bucket->ub_cnt == 0)
813 zdom = zone_domain_lock(zone, domain);
816 * Conditionally set the maximum number of items.
818 zdom->uzd_nitems += bucket->ub_cnt;
819 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
821 zone_domain_imax_set(zdom, zdom->uzd_nitems);
824 * Shift the bounds of the current WSS interval to
825 * avoid perturbing the estimates.
827 atomic_add_long(&zdom->uzd_imax, bucket->ub_cnt);
828 zdom->uzd_imin += bucket->ub_cnt;
829 zdom->uzd_bimin += bucket->ub_cnt;
830 zdom->uzd_limin += bucket->ub_cnt;
832 if (STAILQ_EMPTY(&zdom->uzd_buckets))
833 zdom->uzd_seq = bucket->ub_seq;
836 * Try to promote reuse of recently used items. For items
837 * protected by SMR, try to defer reuse to minimize polling.
839 if (bucket->ub_seq == SMR_SEQ_INVALID)
840 STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
842 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
846 zdom->uzd_nitems -= bucket->ub_cnt;
849 bucket_free(zone, bucket, udata);
852 /* Pops an item out of a per-cpu cache bucket. */
854 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
858 CRITICAL_ASSERT(curthread);
861 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
863 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
864 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
871 /* Pushes an item into a per-cpu cache bucket. */
873 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
876 CRITICAL_ASSERT(curthread);
877 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
878 ("uma_zfree: Freeing to non free bucket index."));
880 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
886 * Unload a UMA bucket from a per-cpu cache.
888 static inline uma_bucket_t
889 cache_bucket_unload(uma_cache_bucket_t bucket)
893 b = bucket->ucb_bucket;
895 MPASS(b->ub_entries == bucket->ucb_entries);
896 b->ub_cnt = bucket->ucb_cnt;
897 bucket->ucb_bucket = NULL;
898 bucket->ucb_entries = bucket->ucb_cnt = 0;
904 static inline uma_bucket_t
905 cache_bucket_unload_alloc(uma_cache_t cache)
908 return (cache_bucket_unload(&cache->uc_allocbucket));
911 static inline uma_bucket_t
912 cache_bucket_unload_free(uma_cache_t cache)
915 return (cache_bucket_unload(&cache->uc_freebucket));
918 static inline uma_bucket_t
919 cache_bucket_unload_cross(uma_cache_t cache)
922 return (cache_bucket_unload(&cache->uc_crossbucket));
926 * Load a bucket into a per-cpu cache bucket.
929 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
932 CRITICAL_ASSERT(curthread);
933 MPASS(bucket->ucb_bucket == NULL);
934 MPASS(b->ub_seq == SMR_SEQ_INVALID);
936 bucket->ucb_bucket = b;
937 bucket->ucb_cnt = b->ub_cnt;
938 bucket->ucb_entries = b->ub_entries;
942 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
945 cache_bucket_load(&cache->uc_allocbucket, b);
949 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
952 cache_bucket_load(&cache->uc_freebucket, b);
957 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
960 cache_bucket_load(&cache->uc_crossbucket, b);
965 * Copy and preserve ucb_spare.
968 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
971 b1->ucb_bucket = b2->ucb_bucket;
972 b1->ucb_entries = b2->ucb_entries;
973 b1->ucb_cnt = b2->ucb_cnt;
977 * Swap two cache buckets.
980 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
982 struct uma_cache_bucket b3;
984 CRITICAL_ASSERT(curthread);
986 cache_bucket_copy(&b3, b1);
987 cache_bucket_copy(b1, b2);
988 cache_bucket_copy(b2, &b3);
992 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
995 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
997 uma_zone_domain_t zdom;
1001 * Avoid the lock if possible.
1003 zdom = ZDOM_GET(zone, domain);
1004 if (zdom->uzd_nitems == 0)
1007 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
1008 !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
1012 * Check the zone's cache of buckets.
1014 zdom = zone_domain_lock(zone, domain);
1015 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL)
1023 zone_log_warning(uma_zone_t zone)
1025 static const struct timeval warninterval = { 300, 0 };
1027 if (!zone_warnings || zone->uz_warning == NULL)
1030 if (ratecheck(&zone->uz_ratecheck, &warninterval))
1031 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
1035 zone_maxaction(uma_zone_t zone)
1038 if (zone->uz_maxaction.ta_func != NULL)
1039 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
1043 * Routine called by timeout which is used to fire off some time interval
1044 * based calculations. (stats, hash size, etc.)
1053 uma_timeout(void *unused)
1056 zone_foreach(zone_timeout, NULL);
1058 /* Reschedule this event */
1059 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1063 * Update the working set size estimates for the zone's bucket cache.
1064 * The constants chosen here are somewhat arbitrary.
1067 zone_domain_update_wss(uma_zone_domain_t zdom)
1071 ZDOM_LOCK_ASSERT(zdom);
1072 MPASS(zdom->uzd_imax >= zdom->uzd_nitems);
1073 MPASS(zdom->uzd_nitems >= zdom->uzd_bimin);
1074 MPASS(zdom->uzd_bimin >= zdom->uzd_imin);
1077 * Estimate WSS as modified moving average of biggest allocation
1078 * batches for each period over few minutes (UMA_TIMEOUT of 20s).
1080 zdom->uzd_wss = lmax(zdom->uzd_wss * 3 / 4,
1081 zdom->uzd_imax - zdom->uzd_bimin);
1084 * Estimate longtime minimum item count as a combination of recent
1085 * minimum item count, adjusted by WSS for safety, and the modified
1086 * moving average over the last several hours (UMA_TIMEOUT of 20s).
1087 * timin measures time since limin tried to go negative, that means
1088 * we were dangerously close to or got out of cache.
1090 m = zdom->uzd_imin - zdom->uzd_wss;
1092 if (zdom->uzd_limin >= m)
1093 zdom->uzd_limin = m;
1095 zdom->uzd_limin = (m + zdom->uzd_limin * 255) / 256;
1098 zdom->uzd_limin = 0;
1099 zdom->uzd_timin = 0;
1102 /* To reduce period edge effects on WSS keep half of the imax. */
1103 atomic_subtract_long(&zdom->uzd_imax,
1104 (zdom->uzd_imax - zdom->uzd_nitems + 1) / 2);
1105 zdom->uzd_imin = zdom->uzd_bimin = zdom->uzd_nitems;
1109 * Routine to perform timeout driven calculations. This expands the
1110 * hashes and does per cpu statistics aggregation.
1115 zone_timeout(uma_zone_t zone, void *unused)
1120 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
1126 * Hash zones are non-numa by definition so the first domain
1127 * is the only one present.
1130 pages = keg->uk_domain[0].ud_pages;
1133 * Expand the keg hash table.
1135 * This is done if the number of slabs is larger than the hash size.
1136 * What I'm trying to do here is completely reduce collisions. This
1137 * may be a little aggressive. Should I allow for two collisions max?
1139 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1140 struct uma_hash newhash;
1141 struct uma_hash oldhash;
1145 * This is so involved because allocating and freeing
1146 * while the keg lock is held will lead to deadlock.
1147 * I have to do everything in stages and check for
1151 ret = hash_alloc(&newhash, 1 << fls(slabs));
1154 if (hash_expand(&keg->uk_hash, &newhash)) {
1155 oldhash = keg->uk_hash;
1156 keg->uk_hash = newhash;
1161 hash_free(&oldhash);
1168 /* Trim caches not used for a long time. */
1169 for (int i = 0; i < vm_ndomains; i++) {
1170 if (bucket_cache_reclaim_domain(zone, false, false, i) &&
1171 (zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1172 keg_drain(zone->uz_keg, i);
1177 * Allocate and zero fill the next sized hash table from the appropriate
1181 * hash A new hash structure with the old hash size in uh_hashsize
1184 * 1 on success and 0 on failure.
1187 hash_alloc(struct uma_hash *hash, u_int size)
1191 KASSERT(powerof2(size), ("hash size must be power of 2"));
1192 if (size > UMA_HASH_SIZE_INIT) {
1193 hash->uh_hashsize = size;
1194 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1195 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1197 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1198 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1199 UMA_ANYDOMAIN, M_WAITOK);
1200 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1202 if (hash->uh_slab_hash) {
1203 bzero(hash->uh_slab_hash, alloc);
1204 hash->uh_hashmask = hash->uh_hashsize - 1;
1212 * Expands the hash table for HASH zones. This is done from zone_timeout
1213 * to reduce collisions. This must not be done in the regular allocation
1214 * path, otherwise, we can recurse on the vm while allocating pages.
1217 * oldhash The hash you want to expand
1218 * newhash The hash structure for the new table
1226 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1228 uma_hash_slab_t slab;
1232 if (!newhash->uh_slab_hash)
1235 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1239 * I need to investigate hash algorithms for resizing without a
1243 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1244 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1245 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1246 LIST_REMOVE(slab, uhs_hlink);
1247 hval = UMA_HASH(newhash, slab->uhs_data);
1248 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1256 * Free the hash bucket to the appropriate backing store.
1259 * slab_hash The hash bucket we're freeing
1260 * hashsize The number of entries in that hash bucket
1266 hash_free(struct uma_hash *hash)
1268 if (hash->uh_slab_hash == NULL)
1270 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1271 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1273 free(hash->uh_slab_hash, M_UMAHASH);
1277 * Frees all outstanding items in a bucket
1280 * zone The zone to free to, must be unlocked.
1281 * bucket The free/alloc bucket with items.
1287 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1291 if (bucket->ub_cnt == 0)
1294 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1295 bucket->ub_seq != SMR_SEQ_INVALID) {
1296 smr_wait(zone->uz_smr, bucket->ub_seq);
1297 bucket->ub_seq = SMR_SEQ_INVALID;
1298 for (i = 0; i < bucket->ub_cnt; i++)
1299 item_dtor(zone, bucket->ub_bucket[i],
1300 zone->uz_size, NULL, SKIP_NONE);
1303 for (i = 0; i < bucket->ub_cnt; i++) {
1304 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
1305 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1306 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
1308 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1309 if (zone->uz_max_items > 0)
1310 zone_free_limit(zone, bucket->ub_cnt);
1312 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1318 * Drains the per cpu caches for a zone.
1320 * NOTE: This may only be called while the zone is being torn down, and not
1321 * during normal operation. This is necessary in order that we do not have
1322 * to migrate CPUs to drain the per-CPU caches.
1325 * zone The zone to drain, must be unlocked.
1331 cache_drain(uma_zone_t zone)
1334 uma_bucket_t bucket;
1339 * XXX: It is safe to not lock the per-CPU caches, because we're
1340 * tearing down the zone anyway. I.e., there will be no further use
1341 * of the caches at this point.
1343 * XXX: It would good to be able to assert that the zone is being
1344 * torn down to prevent improper use of cache_drain().
1346 seq = SMR_SEQ_INVALID;
1347 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1348 seq = smr_advance(zone->uz_smr);
1350 cache = &zone->uz_cpu[cpu];
1351 bucket = cache_bucket_unload_alloc(cache);
1353 bucket_free(zone, bucket, NULL);
1354 bucket = cache_bucket_unload_free(cache);
1355 if (bucket != NULL) {
1356 bucket->ub_seq = seq;
1357 bucket_free(zone, bucket, NULL);
1359 bucket = cache_bucket_unload_cross(cache);
1360 if (bucket != NULL) {
1361 bucket->ub_seq = seq;
1362 bucket_free(zone, bucket, NULL);
1365 bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN);
1369 cache_shrink(uma_zone_t zone, void *unused)
1372 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1376 zone->uz_bucket_size =
1377 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1382 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1385 uma_bucket_t b1, b2, b3;
1388 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1391 b1 = b2 = b3 = NULL;
1393 cache = &zone->uz_cpu[curcpu];
1394 domain = PCPU_GET(domain);
1395 b1 = cache_bucket_unload_alloc(cache);
1398 * Don't flush SMR zone buckets. This leaves the zone without a
1399 * bucket and forces every free to synchronize().
1401 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1402 b2 = cache_bucket_unload_free(cache);
1403 b3 = cache_bucket_unload_cross(cache);
1408 zone_free_bucket(zone, b1, NULL, domain, false);
1410 zone_free_bucket(zone, b2, NULL, domain, false);
1412 /* Adjust the domain so it goes to zone_free_cross. */
1413 domain = (domain + 1) % vm_ndomains;
1414 zone_free_bucket(zone, b3, NULL, domain, false);
1419 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1420 * This is an expensive call because it needs to bind to all CPUs
1421 * one by one and enter a critical section on each of them in order
1422 * to safely access their cache buckets.
1423 * Zone lock must not be held on call this function.
1426 pcpu_cache_drain_safe(uma_zone_t zone)
1431 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1434 cache_shrink(zone, NULL);
1436 zone_foreach(cache_shrink, NULL);
1439 thread_lock(curthread);
1440 sched_bind(curthread, cpu);
1441 thread_unlock(curthread);
1444 cache_drain_safe_cpu(zone, NULL);
1446 zone_foreach(cache_drain_safe_cpu, NULL);
1448 thread_lock(curthread);
1449 sched_unbind(curthread);
1450 thread_unlock(curthread);
1454 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1455 * requested a drain, otherwise the per-domain caches are trimmed to either
1456 * estimated working set size.
1459 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain)
1461 uma_zone_domain_t zdom;
1462 uma_bucket_t bucket;
1467 * The cross bucket is partially filled and not part of
1468 * the item count. Reclaim it individually here.
1470 zdom = ZDOM_GET(zone, domain);
1471 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1472 ZONE_CROSS_LOCK(zone);
1473 bucket = zdom->uzd_cross;
1474 zdom->uzd_cross = NULL;
1475 ZONE_CROSS_UNLOCK(zone);
1477 bucket_free(zone, bucket, NULL);
1481 * If we were asked to drain the zone, we are done only once
1482 * this bucket cache is empty. If trim, we reclaim items in
1483 * excess of the zone's estimated working set size. Multiple
1484 * consecutive calls will shrink the WSS and so reclaim more.
1485 * If neither drain nor trim, then voluntarily reclaim 1/4
1486 * (to reduce first spike) of items not used for a long time.
1489 zone_domain_update_wss(zdom);
1493 target = zdom->uzd_wss;
1494 else if (zdom->uzd_timin > 900 / UMA_TIMEOUT)
1495 target = zdom->uzd_nitems - zdom->uzd_limin / 4;
1500 while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL &&
1501 zdom->uzd_nitems >= target + bucket->ub_cnt) {
1502 bucket = zone_fetch_bucket(zone, zdom, true);
1505 bucket_free(zone, bucket, NULL);
1514 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain)
1519 * Shrink the zone bucket size to ensure that the per-CPU caches
1520 * don't grow too large.
1522 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1523 zone->uz_bucket_size--;
1525 if (domain != UMA_ANYDOMAIN &&
1526 (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) {
1527 bucket_cache_reclaim_domain(zone, drain, true, domain);
1529 for (i = 0; i < vm_ndomains; i++)
1530 bucket_cache_reclaim_domain(zone, drain, true, i);
1535 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1542 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1543 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1545 mem = slab_data(slab, keg);
1546 size = PAGE_SIZE * keg->uk_ppera;
1548 kasan_mark_slab_valid(keg, mem);
1549 if (keg->uk_fini != NULL) {
1550 for (i = start - 1; i > -1; i--)
1553 * trash_fini implies that dtor was trash_dtor. trash_fini
1554 * would check that memory hasn't been modified since free,
1555 * which executed trash_dtor.
1556 * That's why we need to run uma_dbg_kskip() check here,
1557 * albeit we don't make skip check for other init/fini
1560 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1561 keg->uk_fini != trash_fini)
1563 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1565 flags = slab->us_flags;
1566 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1567 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1570 keg->uk_freef(mem, size, flags);
1571 uma_total_dec(size);
1575 keg_drain_domain(uma_keg_t keg, int domain)
1577 struct slabhead freeslabs;
1579 uma_slab_t slab, tmp;
1580 uint32_t i, stofree, stokeep, partial;
1582 dom = &keg->uk_domain[domain];
1583 LIST_INIT(&freeslabs);
1585 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1586 keg->uk_name, keg, domain, dom->ud_free_items);
1588 KEG_LOCK(keg, domain);
1591 * Are the free items in partially allocated slabs sufficient to meet
1592 * the reserve? If not, compute the number of fully free slabs that must
1595 partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1596 if (partial < keg->uk_reserve) {
1597 stokeep = min(dom->ud_free_slabs,
1598 howmany(keg->uk_reserve - partial, keg->uk_ipers));
1602 stofree = dom->ud_free_slabs - stokeep;
1605 * Partition the free slabs into two sets: those that must be kept in
1606 * order to maintain the reserve, and those that may be released back to
1607 * the system. Since one set may be much larger than the other,
1608 * populate the smaller of the two sets and swap them if necessary.
1610 for (i = min(stofree, stokeep); i > 0; i--) {
1611 slab = LIST_FIRST(&dom->ud_free_slab);
1612 LIST_REMOVE(slab, us_link);
1613 LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1615 if (stofree > stokeep)
1616 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1618 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1619 LIST_FOREACH(slab, &freeslabs, us_link)
1620 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1622 dom->ud_free_items -= stofree * keg->uk_ipers;
1623 dom->ud_free_slabs -= stofree;
1624 dom->ud_pages -= stofree * keg->uk_ppera;
1625 KEG_UNLOCK(keg, domain);
1627 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1628 keg_free_slab(keg, slab, keg->uk_ipers);
1632 * Frees pages from a keg back to the system. This is done on demand from
1633 * the pageout daemon.
1638 keg_drain(uma_keg_t keg, int domain)
1642 if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1644 if (domain != UMA_ANYDOMAIN) {
1645 keg_drain_domain(keg, domain);
1647 for (i = 0; i < vm_ndomains; i++)
1648 keg_drain_domain(keg, i);
1653 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain)
1656 * Count active reclaim operations in order to interlock with
1657 * zone_dtor(), which removes the zone from global lists before
1658 * attempting to reclaim items itself.
1660 * The zone may be destroyed while sleeping, so only zone_dtor() should
1664 if (waitok == M_WAITOK) {
1665 while (zone->uz_reclaimers > 0)
1666 msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1);
1668 zone->uz_reclaimers++;
1670 bucket_cache_reclaim(zone, drain, domain);
1672 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1673 keg_drain(zone->uz_keg, domain);
1675 zone->uz_reclaimers--;
1676 if (zone->uz_reclaimers == 0)
1682 zone_drain(uma_zone_t zone, void *arg)
1686 domain = (int)(uintptr_t)arg;
1687 zone_reclaim(zone, domain, M_NOWAIT, true);
1691 zone_trim(uma_zone_t zone, void *arg)
1695 domain = (int)(uintptr_t)arg;
1696 zone_reclaim(zone, domain, M_NOWAIT, false);
1700 * Allocate a new slab for a keg and inserts it into the partial slab list.
1701 * The keg should be unlocked on entry. If the allocation succeeds it will
1702 * be locked on return.
1705 * flags Wait flags for the item initialization routine
1706 * aflags Wait flags for the slab allocation
1709 * The slab that was allocated or NULL if there is no memory and the
1710 * caller specified M_NOWAIT.
1713 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1723 KASSERT(domain >= 0 && domain < vm_ndomains,
1724 ("keg_alloc_slab: domain %d out of range", domain));
1728 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1729 uma_hash_slab_t hslab;
1730 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1734 slab = &hslab->uhs_slab;
1738 * This reproduces the old vm_zone behavior of zero filling pages the
1739 * first time they are added to a zone.
1741 * Malloced items are zeroed in uma_zalloc.
1744 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1749 if (keg->uk_flags & UMA_ZONE_NODUMP)
1752 /* zone is passed for legacy reasons. */
1753 size = keg->uk_ppera * PAGE_SIZE;
1754 mem = keg->uk_allocf(zone, size, domain, &sflags, aflags);
1756 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1757 zone_free_item(slabzone(keg->uk_ipers),
1758 slab_tohashslab(slab), NULL, SKIP_NONE);
1761 uma_total_inc(size);
1763 /* For HASH zones all pages go to the same uma_domain. */
1764 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1767 /* Point the slab into the allocated memory */
1768 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1769 slab = (uma_slab_t)(mem + keg->uk_pgoff);
1771 slab_tohashslab(slab)->uhs_data = mem;
1773 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1774 for (i = 0; i < keg->uk_ppera; i++)
1775 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1778 slab->us_freecount = keg->uk_ipers;
1779 slab->us_flags = sflags;
1780 slab->us_domain = domain;
1782 BIT_FILL(keg->uk_ipers, &slab->us_free);
1784 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1787 if (keg->uk_init != NULL) {
1788 for (i = 0; i < keg->uk_ipers; i++)
1789 if (keg->uk_init(slab_item(slab, keg, i),
1790 keg->uk_size, flags) != 0)
1792 if (i != keg->uk_ipers) {
1793 keg_free_slab(keg, slab, i);
1797 kasan_mark_slab_invalid(keg, mem);
1798 KEG_LOCK(keg, domain);
1800 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1801 slab, keg->uk_name, keg);
1803 if (keg->uk_flags & UMA_ZFLAG_HASH)
1804 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1807 * If we got a slab here it's safe to mark it partially used
1808 * and return. We assume that the caller is going to remove
1809 * at least one item.
1811 dom = &keg->uk_domain[domain];
1812 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1813 dom->ud_pages += keg->uk_ppera;
1814 dom->ud_free_items += keg->uk_ipers;
1823 * This function is intended to be used early on in place of page_alloc(). It
1824 * performs contiguous physical memory allocations and uses a bump allocator for
1825 * KVA, so is usable before the kernel map is initialized.
1828 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1835 pages = howmany(bytes, PAGE_SIZE);
1836 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1838 *pflag = UMA_SLAB_BOOT;
1839 m = vm_page_alloc_noobj_contig_domain(domain, malloc2vm_flags(wait) |
1840 VM_ALLOC_WIRED, pages, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
1841 VM_MEMATTR_DEFAULT);
1845 pa = VM_PAGE_TO_PHYS(m);
1846 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1847 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1848 defined(__riscv) || defined(__powerpc64__)
1849 if ((wait & M_NODUMP) == 0)
1854 /* Allocate KVA and indirectly advance bootmem. */
1855 return ((void *)pmap_map(&bootmem, m->phys_addr,
1856 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE));
1860 startup_free(void *mem, vm_size_t bytes)
1865 va = (vm_offset_t)mem;
1866 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1869 * startup_alloc() returns direct-mapped slabs on some platforms. Avoid
1870 * unmapping ranges of the direct map.
1872 if (va >= bootstart && va + bytes <= bootmem)
1873 pmap_remove(kernel_pmap, va, va + bytes);
1874 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1875 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1876 defined(__riscv) || defined(__powerpc64__)
1877 dump_drop_page(VM_PAGE_TO_PHYS(m));
1879 vm_page_unwire_noq(m);
1885 * Allocates a number of pages from the system
1888 * bytes The number of bytes requested
1889 * wait Shall we wait?
1892 * A pointer to the alloced memory or possibly
1893 * NULL if M_NOWAIT is set.
1896 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1899 void *p; /* Returned page */
1901 *pflag = UMA_SLAB_KERNEL;
1902 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1908 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1911 struct pglist alloctail;
1912 vm_offset_t addr, zkva;
1914 vm_page_t p, p_next;
1919 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1921 TAILQ_INIT(&alloctail);
1922 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | malloc2vm_flags(wait);
1923 *pflag = UMA_SLAB_KERNEL;
1924 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1925 if (CPU_ABSENT(cpu)) {
1926 p = vm_page_alloc_noobj(flags);
1929 p = vm_page_alloc_noobj(flags);
1931 pc = pcpu_find(cpu);
1932 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1935 p = vm_page_alloc_noobj_domain(pc->pc_domain,
1937 if (__predict_false(p == NULL))
1938 p = vm_page_alloc_noobj(flags);
1941 if (__predict_false(p == NULL))
1943 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1945 if ((addr = kva_alloc(bytes)) == 0)
1948 TAILQ_FOREACH(p, &alloctail, listq) {
1949 pmap_qenter(zkva, &p, 1);
1952 return ((void*)addr);
1954 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1955 vm_page_unwire_noq(p);
1962 * Allocates a number of pages from within an object
1965 * bytes The number of bytes requested
1966 * wait Shall we wait?
1969 * A pointer to the alloced memory or possibly
1970 * NULL if M_NOWAIT is set.
1973 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1976 TAILQ_HEAD(, vm_page) alloctail;
1978 vm_offset_t retkva, zkva;
1979 vm_page_t p, p_next;
1983 TAILQ_INIT(&alloctail);
1985 req = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
1986 if ((wait & M_WAITOK) != 0)
1987 req |= VM_ALLOC_WAITOK;
1989 npages = howmany(bytes, PAGE_SIZE);
1990 while (npages > 0) {
1991 p = vm_page_alloc_noobj_domain(domain, req);
1994 * Since the page does not belong to an object, its
1997 TAILQ_INSERT_TAIL(&alloctail, p, listq);
2002 * Page allocation failed, free intermediate pages and
2005 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2006 vm_page_unwire_noq(p);
2011 *flags = UMA_SLAB_PRIV;
2012 zkva = keg->uk_kva +
2013 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
2015 TAILQ_FOREACH(p, &alloctail, listq) {
2016 pmap_qenter(zkva, &p, 1);
2020 return ((void *)retkva);
2024 * Allocate physically contiguous pages.
2027 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
2031 *pflag = UMA_SLAB_KERNEL;
2032 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
2033 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
2037 * Frees a number of pages to the system
2040 * mem A pointer to the memory to be freed
2041 * size The size of the memory being freed
2042 * flags The original p->us_flags field
2048 page_free(void *mem, vm_size_t size, uint8_t flags)
2051 if ((flags & UMA_SLAB_BOOT) != 0) {
2052 startup_free(mem, size);
2056 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
2057 ("UMA: page_free used with invalid flags %x", flags));
2059 kmem_free((vm_offset_t)mem, size);
2063 * Frees pcpu zone allocations
2066 * mem A pointer to the memory to be freed
2067 * size The size of the memory being freed
2068 * flags The original p->us_flags field
2074 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
2076 vm_offset_t sva, curva;
2080 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
2082 if ((flags & UMA_SLAB_BOOT) != 0) {
2083 startup_free(mem, size);
2087 sva = (vm_offset_t)mem;
2088 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
2089 paddr = pmap_kextract(curva);
2090 m = PHYS_TO_VM_PAGE(paddr);
2091 vm_page_unwire_noq(m);
2094 pmap_qremove(sva, size >> PAGE_SHIFT);
2095 kva_free(sva, size);
2099 * Zero fill initializer
2101 * Arguments/Returns follow uma_init specifications
2104 zero_init(void *mem, int size, int flags)
2111 static struct noslabbits *
2112 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2115 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2120 * Actual size of embedded struct slab (!OFFPAGE).
2123 slab_sizeof(int nitems)
2127 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2128 return (roundup(s, UMA_ALIGN_PTR + 1));
2131 #define UMA_FIXPT_SHIFT 31
2132 #define UMA_FRAC_FIXPT(n, d) \
2133 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2134 #define UMA_FIXPT_PCT(f) \
2135 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2136 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
2137 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2140 * Compute the number of items that will fit in a slab. If hdr is true, the
2141 * item count may be limited to provide space in the slab for an inline slab
2142 * header. Otherwise, all slab space will be provided for item storage.
2145 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2150 /* The padding between items is not needed after the last item. */
2151 padpi = rsize - size;
2155 * Start with the maximum item count and remove items until
2156 * the slab header first alongside the allocatable memory.
2158 for (ipers = MIN(SLAB_MAX_SETSIZE,
2159 (slabsize + padpi - slab_sizeof(1)) / rsize);
2161 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2165 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2171 struct keg_layout_result {
2179 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2180 struct keg_layout_result *kl)
2185 kl->slabsize = slabsize;
2187 /* Handle INTERNAL as inline with an extra page. */
2188 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2189 kl->format &= ~UMA_ZFLAG_INTERNAL;
2190 kl->slabsize += PAGE_SIZE;
2193 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2194 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2196 /* Account for memory used by an offpage slab header. */
2197 total = kl->slabsize;
2198 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2199 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2201 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2205 * Determine the format of a uma keg. This determines where the slab header
2206 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2209 * keg The zone we should initialize
2215 keg_layout(uma_keg_t keg)
2217 struct keg_layout_result kl = {}, kl_tmp;
2226 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2227 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2228 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2229 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2230 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2232 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2233 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2234 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2237 alignsize = keg->uk_align + 1;
2240 * ASAN requires that each allocation be aligned to the shadow map
2243 if (alignsize < KASAN_SHADOW_SCALE)
2244 alignsize = KASAN_SHADOW_SCALE;
2248 * Calculate the size of each allocation (rsize) according to
2249 * alignment. If the requested size is smaller than we have
2250 * allocation bits for we round it up.
2252 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2253 rsize = roundup2(rsize, alignsize);
2255 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2257 * We want one item to start on every align boundary in a page.
2258 * To do this we will span pages. We will also extend the item
2259 * by the size of align if it is an even multiple of align.
2260 * Otherwise, it would fall on the same boundary every time.
2262 if ((rsize & alignsize) == 0)
2264 slabsize = rsize * (PAGE_SIZE / alignsize);
2265 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2266 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2267 slabsize = round_page(slabsize);
2270 * Start with a slab size of as many pages as it takes to
2271 * represent a single item. We will try to fit as many
2272 * additional items into the slab as possible.
2274 slabsize = round_page(keg->uk_size);
2277 /* Build a list of all of the available formats for this keg. */
2280 /* Evaluate an inline slab layout. */
2281 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2284 /* TODO: vm_page-embedded slab. */
2287 * We can't do OFFPAGE if we're internal or if we've been
2288 * asked to not go to the VM for buckets. If we do this we
2289 * may end up going to the VM for slabs which we do not want
2290 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2291 * In those cases, evaluate a pseudo-format called INTERNAL
2292 * which has an inline slab header and one extra page to
2293 * guarantee that it fits.
2295 * Otherwise, see if using an OFFPAGE slab will improve our
2298 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2299 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2301 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2304 * Choose a slab size and format which satisfy the minimum efficiency.
2305 * Prefer the smallest slab size that meets the constraints.
2307 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2308 * for small items (up to PAGE_SIZE), the iteration increment is one
2309 * page; and for large items, the increment is one item.
2311 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2312 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2313 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2316 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2317 round_page(rsize * (i - 1) + keg->uk_size);
2319 for (j = 0; j < nfmt; j++) {
2320 /* Only if we have no viable format yet. */
2321 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2325 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2326 if (kl_tmp.eff <= kl.eff)
2331 CTR6(KTR_UMA, "keg %s layout: format %#x "
2332 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2333 keg->uk_name, kl.format, kl.ipers, rsize,
2334 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2336 /* Stop when we reach the minimum efficiency. */
2337 if (kl.eff >= UMA_MIN_EFF)
2341 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2342 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2343 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2347 pages = atop(kl.slabsize);
2348 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2349 pages *= mp_maxid + 1;
2351 keg->uk_rsize = rsize;
2352 keg->uk_ipers = kl.ipers;
2353 keg->uk_ppera = pages;
2354 keg->uk_flags |= kl.format;
2357 * How do we find the slab header if it is offpage or if not all item
2358 * start addresses are in the same page? We could solve the latter
2359 * case with vaddr alignment, but we don't.
2361 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2362 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2363 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2364 keg->uk_flags |= UMA_ZFLAG_HASH;
2366 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2369 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2370 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2372 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2373 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2374 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2375 keg->uk_ipers, pages));
2379 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2380 * the keg onto the global keg list.
2382 * Arguments/Returns follow uma_ctor specifications
2383 * udata Actually uma_kctor_args
2386 keg_ctor(void *mem, int size, void *udata, int flags)
2388 struct uma_kctor_args *arg = udata;
2389 uma_keg_t keg = mem;
2394 keg->uk_size = arg->size;
2395 keg->uk_init = arg->uminit;
2396 keg->uk_fini = arg->fini;
2397 keg->uk_align = arg->align;
2398 keg->uk_reserve = 0;
2399 keg->uk_flags = arg->flags;
2402 * We use a global round-robin policy by default. Zones with
2403 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2404 * case the iterator is never run.
2406 keg->uk_dr.dr_policy = DOMAINSET_RR();
2407 keg->uk_dr.dr_iter = 0;
2410 * The primary zone is passed to us at keg-creation time.
2413 keg->uk_name = zone->uz_name;
2415 if (arg->flags & UMA_ZONE_ZINIT)
2416 keg->uk_init = zero_init;
2418 if (arg->flags & UMA_ZONE_MALLOC)
2419 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2422 keg->uk_flags &= ~UMA_ZONE_PCPU;
2428 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2429 * work on. Use round-robin for everything else.
2431 * Zones may override the default by specifying either.
2434 if ((keg->uk_flags &
2435 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2436 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2437 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2438 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2442 * If we haven't booted yet we need allocations to go through the
2443 * startup cache until the vm is ready.
2445 #ifdef UMA_MD_SMALL_ALLOC
2446 if (keg->uk_ppera == 1)
2447 keg->uk_allocf = uma_small_alloc;
2450 if (booted < BOOT_KVA)
2451 keg->uk_allocf = startup_alloc;
2452 else if (keg->uk_flags & UMA_ZONE_PCPU)
2453 keg->uk_allocf = pcpu_page_alloc;
2454 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2455 keg->uk_allocf = contig_alloc;
2457 keg->uk_allocf = page_alloc;
2458 #ifdef UMA_MD_SMALL_ALLOC
2459 if (keg->uk_ppera == 1)
2460 keg->uk_freef = uma_small_free;
2463 if (keg->uk_flags & UMA_ZONE_PCPU)
2464 keg->uk_freef = pcpu_page_free;
2466 keg->uk_freef = page_free;
2469 * Initialize keg's locks.
2471 for (i = 0; i < vm_ndomains; i++)
2472 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2475 * If we're putting the slab header in the actual page we need to
2476 * figure out where in each page it goes. See slab_sizeof
2479 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2482 shsize = slab_sizeof(keg->uk_ipers);
2483 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2485 * The only way the following is possible is if with our
2486 * UMA_ALIGN_PTR adjustments we are now bigger than
2487 * UMA_SLAB_SIZE. I haven't checked whether this is
2488 * mathematically possible for all cases, so we make
2491 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2492 ("zone %s ipers %d rsize %d size %d slab won't fit",
2493 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2496 if (keg->uk_flags & UMA_ZFLAG_HASH)
2497 hash_alloc(&keg->uk_hash, 0);
2499 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2501 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2503 rw_wlock(&uma_rwlock);
2504 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2505 rw_wunlock(&uma_rwlock);
2510 zone_kva_available(uma_zone_t zone, void *unused)
2514 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2518 if (keg->uk_allocf == startup_alloc) {
2519 /* Switch to the real allocator. */
2520 if (keg->uk_flags & UMA_ZONE_PCPU)
2521 keg->uk_allocf = pcpu_page_alloc;
2522 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2524 keg->uk_allocf = contig_alloc;
2526 keg->uk_allocf = page_alloc;
2531 zone_alloc_counters(uma_zone_t zone, void *unused)
2534 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2535 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2536 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2537 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2541 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2543 uma_zone_domain_t zdom;
2546 struct sysctl_oid *oid, *domainoid;
2547 int domains, i, cnt;
2548 static const char *nokeg = "cache zone";
2552 * Make a sysctl safe copy of the zone name by removing
2553 * any special characters and handling dups by appending
2556 if (zone->uz_namecnt != 0) {
2557 /* Count the number of decimal digits and '_' separator. */
2558 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2560 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2562 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2565 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2566 for (c = zone->uz_ctlname; *c != '\0'; c++)
2567 if (strchr("./\\ -", *c) != NULL)
2571 * Basic parameters at the root.
2573 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2574 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2576 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2577 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2578 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2579 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2580 zone, 0, sysctl_handle_uma_zone_flags, "A",
2581 "Allocator configuration flags");
2582 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2583 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2584 "Desired per-cpu cache size");
2585 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2586 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2587 "Maximum allowed per-cpu cache size");
2592 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2593 domains = vm_ndomains;
2596 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2597 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2599 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2600 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2601 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2602 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2603 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2604 "Real object size with alignment");
2605 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2606 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2607 "pages per-slab allocation");
2608 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2609 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2610 "items available per-slab");
2611 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2612 "align", CTLFLAG_RD, &keg->uk_align, 0,
2613 "item alignment mask");
2614 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2615 "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2616 "number of reserved items");
2617 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2618 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2619 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2620 "Slab utilization (100 - internal fragmentation %)");
2621 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2622 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2623 for (i = 0; i < domains; i++) {
2624 dom = &keg->uk_domain[i];
2625 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2626 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2627 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2628 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2629 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2630 "Total pages currently allocated from VM");
2631 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2632 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2633 "Items free in the slab layer");
2634 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2635 "free_slabs", CTLFLAG_RD, &dom->ud_free_slabs, 0,
2639 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2640 "name", CTLFLAG_RD, nokeg, "Keg name");
2643 * Information about zone limits.
2645 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2646 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2647 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2648 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2649 zone, 0, sysctl_handle_uma_zone_items, "QU",
2650 "Current number of allocated items if limit is set");
2651 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2652 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2653 "Maximum number of allocated and cached items");
2654 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2655 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2656 "Number of threads sleeping at limit");
2657 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2658 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2659 "Total zone limit sleeps");
2660 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2661 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2662 "Maximum number of items in each domain's bucket cache");
2665 * Per-domain zone information.
2667 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2668 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2669 for (i = 0; i < domains; i++) {
2670 zdom = ZDOM_GET(zone, i);
2671 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2672 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2673 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2674 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2675 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2676 "number of items in this domain");
2677 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2678 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2679 "maximum item count in this period");
2680 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2681 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2682 "minimum item count in this period");
2683 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2684 "bimin", CTLFLAG_RD, &zdom->uzd_bimin,
2685 "Minimum item count in this batch");
2686 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2687 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2688 "Working set size");
2689 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2690 "limin", CTLFLAG_RD, &zdom->uzd_limin,
2691 "Long time minimum item count");
2692 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2693 "timin", CTLFLAG_RD, &zdom->uzd_timin, 0,
2694 "Time since zero long time minimum item count");
2698 * General statistics.
2700 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2701 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2702 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2703 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2704 zone, 1, sysctl_handle_uma_zone_cur, "I",
2705 "Current number of allocated items");
2706 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2707 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2708 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2709 "Total allocation calls");
2710 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2711 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2712 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2713 "Total free calls");
2714 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2715 "fails", CTLFLAG_RD, &zone->uz_fails,
2716 "Number of allocation failures");
2717 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2718 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2719 "Free calls from the wrong domain");
2722 struct uma_zone_count {
2728 zone_count(uma_zone_t zone, void *arg)
2730 struct uma_zone_count *cnt;
2734 * Some zones are rapidly created with identical names and
2735 * destroyed out of order. This can lead to gaps in the count.
2736 * Use one greater than the maximum observed for this name.
2738 if (strcmp(zone->uz_name, cnt->name) == 0)
2739 cnt->count = MAX(cnt->count,
2740 zone->uz_namecnt + 1);
2744 zone_update_caches(uma_zone_t zone)
2748 for (i = 0; i <= mp_maxid; i++) {
2749 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2750 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2755 * Zone header ctor. This initializes all fields, locks, etc.
2757 * Arguments/Returns follow uma_ctor specifications
2758 * udata Actually uma_zctor_args
2761 zone_ctor(void *mem, int size, void *udata, int flags)
2763 struct uma_zone_count cnt;
2764 struct uma_zctor_args *arg = udata;
2765 uma_zone_domain_t zdom;
2766 uma_zone_t zone = mem;
2772 zone->uz_name = arg->name;
2773 zone->uz_ctor = arg->ctor;
2774 zone->uz_dtor = arg->dtor;
2775 zone->uz_init = NULL;
2776 zone->uz_fini = NULL;
2777 zone->uz_sleeps = 0;
2778 zone->uz_bucket_size = 0;
2779 zone->uz_bucket_size_min = 0;
2780 zone->uz_bucket_size_max = BUCKET_MAX;
2781 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2782 zone->uz_warning = NULL;
2783 /* The domain structures follow the cpu structures. */
2784 zone->uz_bucket_max = ULONG_MAX;
2785 timevalclear(&zone->uz_ratecheck);
2787 /* Count the number of duplicate names. */
2788 cnt.name = arg->name;
2790 zone_foreach(zone_count, &cnt);
2791 zone->uz_namecnt = cnt.count;
2792 ZONE_CROSS_LOCK_INIT(zone);
2794 for (i = 0; i < vm_ndomains; i++) {
2795 zdom = ZDOM_GET(zone, i);
2796 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2797 STAILQ_INIT(&zdom->uzd_buckets);
2800 #if defined(INVARIANTS) && !defined(KASAN)
2801 if (arg->uminit == trash_init && arg->fini == trash_fini)
2802 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2803 #elif defined(KASAN)
2804 if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0)
2805 arg->flags |= UMA_ZONE_NOKASAN;
2809 * This is a pure cache zone, no kegs.
2812 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2813 ("zone_ctor: Import specified for non-cache zone."));
2814 zone->uz_flags = arg->flags;
2815 zone->uz_size = arg->size;
2816 zone->uz_import = arg->import;
2817 zone->uz_release = arg->release;
2818 zone->uz_arg = arg->arg;
2821 * Cache zones are round-robin unless a policy is
2822 * specified because they may have incompatible
2825 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2826 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2828 rw_wlock(&uma_rwlock);
2829 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2830 rw_wunlock(&uma_rwlock);
2835 * Use the regular zone/keg/slab allocator.
2837 zone->uz_import = zone_import;
2838 zone->uz_release = zone_release;
2839 zone->uz_arg = zone;
2842 if (arg->flags & UMA_ZONE_SECONDARY) {
2843 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2844 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2845 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2846 zone->uz_init = arg->uminit;
2847 zone->uz_fini = arg->fini;
2848 zone->uz_flags |= UMA_ZONE_SECONDARY;
2849 rw_wlock(&uma_rwlock);
2851 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2852 if (LIST_NEXT(z, uz_link) == NULL) {
2853 LIST_INSERT_AFTER(z, zone, uz_link);
2858 rw_wunlock(&uma_rwlock);
2859 } else if (keg == NULL) {
2860 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2861 arg->align, arg->flags)) == NULL)
2864 struct uma_kctor_args karg;
2867 /* We should only be here from uma_startup() */
2868 karg.size = arg->size;
2869 karg.uminit = arg->uminit;
2870 karg.fini = arg->fini;
2871 karg.align = arg->align;
2872 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2874 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2880 /* Inherit properties from the keg. */
2882 zone->uz_size = keg->uk_size;
2883 zone->uz_flags |= (keg->uk_flags &
2884 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2887 if (booted >= BOOT_PCPU) {
2888 zone_alloc_counters(zone, NULL);
2889 if (booted >= BOOT_RUNNING)
2890 zone_alloc_sysctl(zone, NULL);
2892 zone->uz_allocs = EARLY_COUNTER;
2893 zone->uz_frees = EARLY_COUNTER;
2894 zone->uz_fails = EARLY_COUNTER;
2897 /* Caller requests a private SMR context. */
2898 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2899 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2901 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2902 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2903 ("Invalid zone flag combination"));
2904 if (arg->flags & UMA_ZFLAG_INTERNAL)
2905 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2906 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2907 zone->uz_bucket_size = BUCKET_MAX;
2908 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2909 zone->uz_bucket_size = 0;
2911 zone->uz_bucket_size = bucket_select(zone->uz_size);
2912 zone->uz_bucket_size_min = zone->uz_bucket_size;
2913 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2914 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2915 zone_update_caches(zone);
2921 * Keg header dtor. This frees all data, destroys locks, frees the hash
2922 * table and removes the keg from the global list.
2924 * Arguments/Returns follow uma_dtor specifications
2928 keg_dtor(void *arg, int size, void *udata)
2931 uint32_t free, pages;
2934 keg = (uma_keg_t)arg;
2936 for (i = 0; i < vm_ndomains; i++) {
2937 free += keg->uk_domain[i].ud_free_items;
2938 pages += keg->uk_domain[i].ud_pages;
2939 KEG_LOCK_FINI(keg, i);
2942 printf("Freed UMA keg (%s) was not empty (%u items). "
2943 " Lost %u pages of memory.\n",
2944 keg->uk_name ? keg->uk_name : "",
2945 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2947 hash_free(&keg->uk_hash);
2953 * Arguments/Returns follow uma_dtor specifications
2957 zone_dtor(void *arg, int size, void *udata)
2963 zone = (uma_zone_t)arg;
2965 sysctl_remove_oid(zone->uz_oid, 1, 1);
2967 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2970 rw_wlock(&uma_rwlock);
2971 LIST_REMOVE(zone, uz_link);
2972 rw_wunlock(&uma_rwlock);
2973 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2975 keg->uk_reserve = 0;
2977 zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true);
2980 * We only destroy kegs from non secondary/non cache zones.
2982 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2984 rw_wlock(&uma_rwlock);
2985 LIST_REMOVE(keg, uk_link);
2986 rw_wunlock(&uma_rwlock);
2987 zone_free_item(kegs, keg, NULL, SKIP_NONE);
2989 counter_u64_free(zone->uz_allocs);
2990 counter_u64_free(zone->uz_frees);
2991 counter_u64_free(zone->uz_fails);
2992 counter_u64_free(zone->uz_xdomain);
2993 free(zone->uz_ctlname, M_UMA);
2994 for (i = 0; i < vm_ndomains; i++)
2995 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2996 ZONE_CROSS_LOCK_FINI(zone);
3000 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3005 LIST_FOREACH(keg, &uma_kegs, uk_link) {
3006 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
3009 LIST_FOREACH(zone, &uma_cachezones, uz_link)
3014 * Traverses every zone in the system and calls a callback
3017 * zfunc A pointer to a function which accepts a zone
3024 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3027 rw_rlock(&uma_rwlock);
3028 zone_foreach_unlocked(zfunc, arg);
3029 rw_runlock(&uma_rwlock);
3033 * Initialize the kernel memory allocator. This is done after pages can be
3034 * allocated but before general KVA is available.
3037 uma_startup1(vm_offset_t virtual_avail)
3039 struct uma_zctor_args args;
3040 size_t ksize, zsize, size;
3041 uma_keg_t primarykeg;
3046 bootstart = bootmem = virtual_avail;
3048 rw_init(&uma_rwlock, "UMA lock");
3049 sx_init(&uma_reclaim_lock, "umareclaim");
3051 ksize = sizeof(struct uma_keg) +
3052 (sizeof(struct uma_domain) * vm_ndomains);
3053 ksize = roundup(ksize, UMA_SUPER_ALIGN);
3054 zsize = sizeof(struct uma_zone) +
3055 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
3056 (sizeof(struct uma_zone_domain) * vm_ndomains);
3057 zsize = roundup(zsize, UMA_SUPER_ALIGN);
3059 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
3060 size = (zsize * 2) + ksize;
3061 for (domain = 0; domain < vm_ndomains; domain++) {
3062 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
3067 zones = (uma_zone_t)m;
3069 kegs = (uma_zone_t)m;
3071 primarykeg = (uma_keg_t)m;
3073 /* "manually" create the initial zone */
3074 memset(&args, 0, sizeof(args));
3075 args.name = "UMA Kegs";
3077 args.ctor = keg_ctor;
3078 args.dtor = keg_dtor;
3079 args.uminit = zero_init;
3081 args.keg = primarykeg;
3082 args.align = UMA_SUPER_ALIGN - 1;
3083 args.flags = UMA_ZFLAG_INTERNAL;
3084 zone_ctor(kegs, zsize, &args, M_WAITOK);
3086 args.name = "UMA Zones";
3088 args.ctor = zone_ctor;
3089 args.dtor = zone_dtor;
3090 args.uminit = zero_init;
3093 args.align = UMA_SUPER_ALIGN - 1;
3094 args.flags = UMA_ZFLAG_INTERNAL;
3095 zone_ctor(zones, zsize, &args, M_WAITOK);
3097 /* Now make zones for slab headers */
3098 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
3099 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3100 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
3101 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3103 hashzone = uma_zcreate("UMA Hash",
3104 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
3105 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3111 #ifndef UMA_MD_SMALL_ALLOC
3112 extern void vm_radix_reserve_kva(void);
3116 * Advertise the availability of normal kva allocations and switch to
3117 * the default back-end allocator. Marks the KVA we consumed on startup
3118 * as used in the map.
3124 if (bootstart != bootmem) {
3125 vm_map_lock(kernel_map);
3126 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
3127 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
3128 vm_map_unlock(kernel_map);
3131 #ifndef UMA_MD_SMALL_ALLOC
3132 /* Set up radix zone to use noobj_alloc. */
3133 vm_radix_reserve_kva();
3137 zone_foreach_unlocked(zone_kva_available, NULL);
3142 * Allocate counters as early as possible so that boot-time allocations are
3143 * accounted more precisely.
3146 uma_startup_pcpu(void *arg __unused)
3149 zone_foreach_unlocked(zone_alloc_counters, NULL);
3152 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
3155 * Finish our initialization steps.
3158 uma_startup3(void *arg __unused)
3162 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
3163 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
3164 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
3166 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
3167 callout_init(&uma_callout, 1);
3168 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
3169 booted = BOOT_RUNNING;
3171 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
3172 EVENTHANDLER_PRI_FIRST);
3174 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
3180 booted = BOOT_SHUTDOWN;
3184 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3185 int align, uint32_t flags)
3187 struct uma_kctor_args args;
3190 args.uminit = uminit;
3192 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
3195 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3198 /* Public functions */
3201 uma_set_align(int align)
3204 if (align != UMA_ALIGN_CACHE)
3205 uma_align_cache = align;
3210 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3211 uma_init uminit, uma_fini fini, int align, uint32_t flags)
3214 struct uma_zctor_args args;
3217 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
3220 /* This stuff is essential for the zone ctor */
3221 memset(&args, 0, sizeof(args));
3226 args.uminit = uminit;
3228 #if defined(INVARIANTS) && !defined(KASAN)
3230 * Inject procedures which check for memory use after free if we are
3231 * allowed to scramble the memory while it is not allocated. This
3232 * requires that: UMA is actually able to access the memory, no init
3233 * or fini procedures, no dependency on the initial value of the
3234 * memory, and no (legitimate) use of the memory after free. Note,
3235 * the ctor and dtor do not need to be empty.
3237 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3238 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3239 args.uminit = trash_init;
3240 args.fini = trash_fini;
3247 sx_xlock(&uma_reclaim_lock);
3248 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3249 sx_xunlock(&uma_reclaim_lock);
3256 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3257 uma_init zinit, uma_fini zfini, uma_zone_t primary)
3259 struct uma_zctor_args args;
3263 keg = primary->uz_keg;
3264 memset(&args, 0, sizeof(args));
3266 args.size = keg->uk_size;
3269 args.uminit = zinit;
3271 args.align = keg->uk_align;
3272 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3275 sx_xlock(&uma_reclaim_lock);
3276 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3277 sx_xunlock(&uma_reclaim_lock);
3284 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3285 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3286 void *arg, int flags)
3288 struct uma_zctor_args args;
3290 memset(&args, 0, sizeof(args));
3295 args.uminit = zinit;
3297 args.import = zimport;
3298 args.release = zrelease;
3301 args.flags = flags | UMA_ZFLAG_CACHE;
3303 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3308 uma_zdestroy(uma_zone_t zone)
3312 * Large slabs are expensive to reclaim, so don't bother doing
3313 * unnecessary work if we're shutting down.
3315 if (booted == BOOT_SHUTDOWN &&
3316 zone->uz_fini == NULL && zone->uz_release == zone_release)
3318 sx_xlock(&uma_reclaim_lock);
3319 zone_free_item(zones, zone, NULL, SKIP_NONE);
3320 sx_xunlock(&uma_reclaim_lock);
3324 uma_zwait(uma_zone_t zone)
3327 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3328 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3329 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3330 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3332 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3336 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3338 void *item, *pcpu_item;
3342 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3344 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3347 pcpu_item = zpcpu_base_to_offset(item);
3348 if (flags & M_ZERO) {
3350 for (i = 0; i <= mp_maxid; i++)
3351 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3353 bzero(item, zone->uz_size);
3360 * A stub while both regular and pcpu cases are identical.
3363 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3368 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3371 /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */
3372 if (pcpu_item == NULL)
3375 item = zpcpu_offset_to_base(pcpu_item);
3376 uma_zfree_arg(zone, item, udata);
3379 static inline void *
3380 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3387 kasan_mark_item_valid(zone, item);
3390 skipdbg = uma_dbg_zskip(zone, item);
3391 if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3392 zone->uz_ctor != trash_ctor)
3393 trash_ctor(item, size, udata, flags);
3396 /* Check flags before loading ctor pointer. */
3397 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3398 __predict_false(zone->uz_ctor != NULL) &&
3399 zone->uz_ctor(item, size, udata, flags) != 0) {
3400 counter_u64_add(zone->uz_fails, 1);
3401 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3406 uma_dbg_alloc(zone, NULL, item);
3408 if (__predict_false(flags & M_ZERO))
3409 return (memset(item, 0, size));
3415 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3416 enum zfreeskip skip)
3421 skipdbg = uma_dbg_zskip(zone, item);
3422 if (skip == SKIP_NONE && !skipdbg) {
3423 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3424 uma_dbg_free(zone, udata, item);
3426 uma_dbg_free(zone, NULL, item);
3429 if (__predict_true(skip < SKIP_DTOR)) {
3430 if (zone->uz_dtor != NULL)
3431 zone->uz_dtor(item, size, udata);
3433 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3434 zone->uz_dtor != trash_dtor)
3435 trash_dtor(item, size, udata);
3438 kasan_mark_item_invalid(zone, item);
3443 item_domain(void *item)
3447 domain = vm_phys_domain(vtophys(item));
3448 KASSERT(domain >= 0 && domain < vm_ndomains,
3449 ("%s: unknown domain for item %p", __func__, item));
3454 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3455 #define UMA_ZALLOC_DEBUG
3457 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3463 if (flags & M_WAITOK) {
3464 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3465 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3470 KASSERT((flags & M_EXEC) == 0,
3471 ("uma_zalloc_debug: called with M_EXEC"));
3472 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3473 ("uma_zalloc_debug: called within spinlock or critical section"));
3474 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3475 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3478 #ifdef DEBUG_MEMGUARD
3479 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3481 item = memguard_alloc(zone->uz_size, flags);
3483 error = EJUSTRETURN;
3484 if (zone->uz_init != NULL &&
3485 zone->uz_init(item, zone->uz_size, flags) != 0) {
3489 if (zone->uz_ctor != NULL &&
3490 zone->uz_ctor(item, zone->uz_size, udata,
3492 counter_u64_add(zone->uz_fails, 1);
3493 zone->uz_fini(item, zone->uz_size);
3500 /* This is unfortunate but should not be fatal. */
3507 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3509 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3510 ("uma_zfree_debug: called with spinlock or critical section held"));
3512 #ifdef DEBUG_MEMGUARD
3513 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3514 if (zone->uz_dtor != NULL)
3515 zone->uz_dtor(item, zone->uz_size, udata);
3516 if (zone->uz_fini != NULL)
3517 zone->uz_fini(item, zone->uz_size);
3518 memguard_free(item);
3519 return (EJUSTRETURN);
3526 static inline void *
3527 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3528 void *udata, int flags)
3533 item = cache_bucket_pop(cache, bucket);
3534 size = cache_uz_size(cache);
3535 uz_flags = cache_uz_flags(cache);
3537 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3540 static __noinline void *
3541 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3543 uma_cache_bucket_t bucket;
3546 while (cache_alloc(zone, cache, udata, flags)) {
3547 cache = &zone->uz_cpu[curcpu];
3548 bucket = &cache->uc_allocbucket;
3549 if (__predict_false(bucket->ucb_cnt == 0))
3551 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3556 * We can not get a bucket so try to return a single item.
3558 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3559 domain = PCPU_GET(domain);
3561 domain = UMA_ANYDOMAIN;
3562 return (zone_alloc_item(zone, udata, domain, flags));
3567 uma_zalloc_smr(uma_zone_t zone, int flags)
3569 uma_cache_bucket_t bucket;
3572 #ifdef UMA_ZALLOC_DEBUG
3575 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3576 ("uma_zalloc_arg: called with non-SMR zone."));
3577 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3582 cache = &zone->uz_cpu[curcpu];
3583 bucket = &cache->uc_allocbucket;
3584 if (__predict_false(bucket->ucb_cnt == 0))
3585 return (cache_alloc_retry(zone, cache, NULL, flags));
3586 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3591 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3593 uma_cache_bucket_t bucket;
3596 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3597 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3599 /* This is the fast path allocation */
3600 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3603 #ifdef UMA_ZALLOC_DEBUG
3606 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3607 ("uma_zalloc_arg: called with SMR zone."));
3608 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3613 * If possible, allocate from the per-CPU cache. There are two
3614 * requirements for safe access to the per-CPU cache: (1) the thread
3615 * accessing the cache must not be preempted or yield during access,
3616 * and (2) the thread must not migrate CPUs without switching which
3617 * cache it accesses. We rely on a critical section to prevent
3618 * preemption and migration. We release the critical section in
3619 * order to acquire the zone mutex if we are unable to allocate from
3620 * the current cache; when we re-acquire the critical section, we
3621 * must detect and handle migration if it has occurred.
3624 cache = &zone->uz_cpu[curcpu];
3625 bucket = &cache->uc_allocbucket;
3626 if (__predict_false(bucket->ucb_cnt == 0))
3627 return (cache_alloc_retry(zone, cache, udata, flags));
3628 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3632 * Replenish an alloc bucket and possibly restore an old one. Called in
3633 * a critical section. Returns in a critical section.
3635 * A false return value indicates an allocation failure.
3636 * A true return value indicates success and the caller should retry.
3638 static __noinline bool
3639 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3641 uma_bucket_t bucket;
3642 int curdomain, domain;
3645 CRITICAL_ASSERT(curthread);
3648 * If we have run out of items in our alloc bucket see
3649 * if we can switch with the free bucket.
3651 * SMR Zones can't re-use the free bucket until the sequence has
3654 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3655 cache->uc_freebucket.ucb_cnt != 0) {
3656 cache_bucket_swap(&cache->uc_freebucket,
3657 &cache->uc_allocbucket);
3662 * Discard any empty allocation bucket while we hold no locks.
3664 bucket = cache_bucket_unload_alloc(cache);
3667 if (bucket != NULL) {
3668 KASSERT(bucket->ub_cnt == 0,
3669 ("cache_alloc: Entered with non-empty alloc bucket."));
3670 bucket_free(zone, bucket, udata);
3674 * Attempt to retrieve the item from the per-CPU cache has failed, so
3675 * we must go back to the zone. This requires the zdom lock, so we
3676 * must drop the critical section, then re-acquire it when we go back
3677 * to the cache. Since the critical section is released, we may be
3678 * preempted or migrate. As such, make sure not to maintain any
3679 * thread-local state specific to the cache from prior to releasing
3680 * the critical section.
3682 domain = PCPU_GET(domain);
3683 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3684 VM_DOMAIN_EMPTY(domain))
3685 domain = zone_domain_highest(zone, domain);
3686 bucket = cache_fetch_bucket(zone, cache, domain);
3687 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3688 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3694 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3695 zone->uz_name, zone, bucket);
3696 if (bucket == NULL) {
3702 * See if we lost the race or were migrated. Cache the
3703 * initialized bucket to make this less likely or claim
3704 * the memory directly.
3707 cache = &zone->uz_cpu[curcpu];
3708 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3709 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3710 (curdomain = PCPU_GET(domain)) == domain ||
3711 VM_DOMAIN_EMPTY(curdomain))) {
3713 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3715 cache_bucket_load_alloc(cache, bucket);
3720 * We lost the race, release this bucket and start over.
3723 zone_put_bucket(zone, domain, bucket, udata, !new);
3730 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3733 uma_bucket_t bucket;
3734 uma_zone_domain_t zdom;
3738 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3739 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3741 /* This is the fast path allocation */
3742 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3743 zone->uz_name, zone, domain, flags);
3745 if (flags & M_WAITOK) {
3746 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3747 "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3749 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3750 ("uma_zalloc_domain: called with spinlock or critical section held"));
3751 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3752 ("uma_zalloc_domain: called with SMR zone."));
3754 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3755 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3757 if (vm_ndomains == 1)
3758 return (uma_zalloc_arg(zone, udata, flags));
3761 * Try to allocate from the bucket cache before falling back to the keg.
3762 * We could try harder and attempt to allocate from per-CPU caches or
3763 * the per-domain cross-domain buckets, but the complexity is probably
3764 * not worth it. It is more important that frees of previous
3765 * cross-domain allocations do not blow up the cache.
3767 zdom = zone_domain_lock(zone, domain);
3768 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3769 item = bucket->ub_bucket[bucket->ub_cnt - 1];
3771 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3774 zone_put_bucket(zone, domain, bucket, udata, true);
3775 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3778 KASSERT(item_domain(item) == domain,
3779 ("%s: bucket cache item %p from wrong domain",
3781 counter_u64_add(zone->uz_allocs, 1);
3786 return (zone_alloc_item(zone, udata, domain, flags));
3788 return (uma_zalloc_arg(zone, udata, flags));
3793 * Find a slab with some space. Prefer slabs that are partially used over those
3794 * that are totally full. This helps to reduce fragmentation.
3796 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3800 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3806 KASSERT(domain >= 0 && domain < vm_ndomains,
3807 ("keg_first_slab: domain %d out of range", domain));
3808 KEG_LOCK_ASSERT(keg, domain);
3813 dom = &keg->uk_domain[domain];
3814 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3816 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3817 LIST_REMOVE(slab, us_link);
3818 dom->ud_free_slabs--;
3819 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3823 domain = (domain + 1) % vm_ndomains;
3824 } while (domain != start);
3830 * Fetch an existing slab from a free or partial list. Returns with the
3831 * keg domain lock held if a slab was found or unlocked if not.
3834 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3839 /* HASH has a single free list. */
3840 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3843 KEG_LOCK(keg, domain);
3844 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3845 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3846 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3847 KEG_UNLOCK(keg, domain);
3854 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3856 struct vm_domainset_iter di;
3861 KASSERT((flags & (M_WAITOK | M_NOVM)) != (M_WAITOK | M_NOVM),
3862 ("%s: invalid flags %#x", __func__, flags));
3866 * Use the keg's policy if upper layers haven't already specified a
3867 * domain (as happens with first-touch zones).
3869 * To avoid races we run the iterator with the keg lock held, but that
3870 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3871 * clear M_WAITOK and handle low memory conditions locally.
3873 rr = rdomain == UMA_ANYDOMAIN;
3875 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3876 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3884 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3889 * M_NOVM is used to break the recursion that can otherwise
3890 * occur if low-level memory management routines use UMA.
3892 if ((flags & M_NOVM) == 0) {
3893 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3899 if ((flags & M_USE_RESERVE) != 0) {
3901 * Drain reserves from other domains before
3902 * giving up or sleeping. It may be useful to
3903 * support per-domain reserves eventually.
3905 rdomain = UMA_ANYDOMAIN;
3908 if ((flags & M_WAITOK) == 0)
3910 vm_wait_domain(domain);
3911 } else if (vm_domainset_iter_policy(&di, &domain) != 0) {
3912 if ((flags & M_WAITOK) != 0) {
3913 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
3921 * We might not have been able to get a slab but another cpu
3922 * could have while we were unlocked. Check again before we
3925 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3932 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3938 KEG_LOCK_ASSERT(keg, slab->us_domain);
3940 dom = &keg->uk_domain[slab->us_domain];
3941 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3942 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3943 item = slab_item(slab, keg, freei);
3944 slab->us_freecount--;
3945 dom->ud_free_items--;
3948 * Move this slab to the full list. It must be on the partial list, so
3949 * we do not need to update the free slab count. In particular,
3950 * keg_fetch_slab() always returns slabs on the partial list.
3952 if (slab->us_freecount == 0) {
3953 LIST_REMOVE(slab, us_link);
3954 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3961 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3975 /* Try to keep the buckets totally full */
3976 for (i = 0; i < max; ) {
3977 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3980 stripe = howmany(max, vm_ndomains);
3982 dom = &keg->uk_domain[slab->us_domain];
3984 bucket[i++] = slab_alloc_item(keg, slab);
3985 if (keg->uk_reserve > 0 &&
3986 dom->ud_free_items <= keg->uk_reserve) {
3988 * Avoid depleting the reserve after a
3989 * successful item allocation, even if
3990 * M_USE_RESERVE is specified.
3992 KEG_UNLOCK(keg, slab->us_domain);
3997 * If the zone is striped we pick a new slab for every
3998 * N allocations. Eliminating this conditional will
3999 * instead pick a new domain for each bucket rather
4000 * than stripe within each bucket. The current option
4001 * produces more fragmentation and requires more cpu
4002 * time but yields better distribution.
4004 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
4005 vm_ndomains > 1 && --stripe == 0)
4008 } while (slab->us_freecount != 0 && i < max);
4009 KEG_UNLOCK(keg, slab->us_domain);
4011 /* Don't block if we allocated any successfully. */
4020 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
4022 uint64_t old, new, total, max;
4025 * The hard case. We're going to sleep because there were existing
4026 * sleepers or because we ran out of items. This routine enforces
4027 * fairness by keeping fifo order.
4029 * First release our ill gotten gains and make some noise.
4032 zone_free_limit(zone, count);
4033 zone_log_warning(zone);
4034 zone_maxaction(zone);
4035 if (flags & M_NOWAIT)
4039 * We need to allocate an item or set ourself as a sleeper
4040 * while the sleepq lock is held to avoid wakeup races. This
4041 * is essentially a home rolled semaphore.
4043 sleepq_lock(&zone->uz_max_items);
4044 old = zone->uz_items;
4046 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
4047 /* Cache the max since we will evaluate twice. */
4048 max = zone->uz_max_items;
4049 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
4050 UZ_ITEMS_COUNT(old) >= max)
4051 new = old + UZ_ITEMS_SLEEPER;
4053 new = old + MIN(count, max - old);
4054 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
4056 /* We may have successfully allocated under the sleepq lock. */
4057 if (UZ_ITEMS_SLEEPERS(new) == 0) {
4058 sleepq_release(&zone->uz_max_items);
4063 * This is in a different cacheline from uz_items so that we
4064 * don't constantly invalidate the fastpath cacheline when we
4065 * adjust item counts. This could be limited to toggling on
4068 atomic_add_32(&zone->uz_sleepers, 1);
4069 atomic_add_64(&zone->uz_sleeps, 1);
4072 * We have added ourselves as a sleeper. The sleepq lock
4073 * protects us from wakeup races. Sleep now and then retry.
4075 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
4076 sleepq_wait(&zone->uz_max_items, PVM);
4079 * After wakeup, remove ourselves as a sleeper and try
4080 * again. We no longer have the sleepq lock for protection.
4082 * Subract ourselves as a sleeper while attempting to add
4085 atomic_subtract_32(&zone->uz_sleepers, 1);
4086 old = atomic_fetchadd_64(&zone->uz_items,
4087 -(UZ_ITEMS_SLEEPER - count));
4088 /* We're no longer a sleeper. */
4089 old -= UZ_ITEMS_SLEEPER;
4092 * If we're still at the limit, restart. Notably do not
4093 * block on other sleepers. Cache the max value to protect
4094 * against changes via sysctl.
4096 total = UZ_ITEMS_COUNT(old);
4097 max = zone->uz_max_items;
4100 /* Truncate if necessary, otherwise wake other sleepers. */
4101 if (total + count > max) {
4102 zone_free_limit(zone, total + count - max);
4103 count = max - total;
4104 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
4105 wakeup_one(&zone->uz_max_items);
4112 * Allocate 'count' items from our max_items limit. Returns the number
4113 * available. If M_NOWAIT is not specified it will sleep until at least
4114 * one item can be allocated.
4117 zone_alloc_limit(uma_zone_t zone, int count, int flags)
4122 max = zone->uz_max_items;
4126 * We expect normal allocations to succeed with a simple
4129 old = atomic_fetchadd_64(&zone->uz_items, count);
4130 if (__predict_true(old + count <= max))
4134 * If we had some items and no sleepers just return the
4135 * truncated value. We have to release the excess space
4136 * though because that may wake sleepers who weren't woken
4137 * because we were temporarily over the limit.
4140 zone_free_limit(zone, (old + count) - max);
4143 return (zone_alloc_limit_hard(zone, count, flags));
4147 * Free a number of items back to the limit.
4150 zone_free_limit(uma_zone_t zone, int count)
4157 * In the common case we either have no sleepers or
4158 * are still over the limit and can just return.
4160 old = atomic_fetchadd_64(&zone->uz_items, -count);
4161 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4162 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4166 * Moderate the rate of wakeups. Sleepers will continue
4167 * to generate wakeups if necessary.
4169 wakeup_one(&zone->uz_max_items);
4173 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4175 uma_bucket_t bucket;
4176 int error, maxbucket, cnt;
4178 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4181 /* Avoid allocs targeting empty domains. */
4182 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4183 domain = UMA_ANYDOMAIN;
4184 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4185 domain = UMA_ANYDOMAIN;
4187 if (zone->uz_max_items > 0)
4188 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4191 maxbucket = zone->uz_bucket_size;
4195 /* Don't wait for buckets, preserve caller's NOVM setting. */
4196 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4197 if (bucket == NULL) {
4202 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4203 MIN(maxbucket, bucket->ub_entries), domain, flags);
4206 * Initialize the memory if necessary.
4208 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4211 for (i = 0; i < bucket->ub_cnt; i++) {
4212 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
4213 error = zone->uz_init(bucket->ub_bucket[i],
4214 zone->uz_size, flags);
4215 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
4221 * If we couldn't initialize the whole bucket, put the
4222 * rest back onto the freelist.
4224 if (i != bucket->ub_cnt) {
4225 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4226 bucket->ub_cnt - i);
4228 bzero(&bucket->ub_bucket[i],
4229 sizeof(void *) * (bucket->ub_cnt - i));
4235 cnt = bucket->ub_cnt;
4236 if (bucket->ub_cnt == 0) {
4237 bucket_free(zone, bucket, udata);
4238 counter_u64_add(zone->uz_fails, 1);
4242 if (zone->uz_max_items > 0 && cnt < maxbucket)
4243 zone_free_limit(zone, maxbucket - cnt);
4249 * Allocates a single item from a zone.
4252 * zone The zone to alloc for.
4253 * udata The data to be passed to the constructor.
4254 * domain The domain to allocate from or UMA_ANYDOMAIN.
4255 * flags M_WAITOK, M_NOWAIT, M_ZERO.
4258 * NULL if there is no memory and M_NOWAIT is set
4259 * An item if successful
4263 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4267 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4268 counter_u64_add(zone->uz_fails, 1);
4272 /* Avoid allocs targeting empty domains. */
4273 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4274 domain = UMA_ANYDOMAIN;
4276 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4280 * We have to call both the zone's init (not the keg's init)
4281 * and the zone's ctor. This is because the item is going from
4282 * a keg slab directly to the user, and the user is expecting it
4283 * to be both zone-init'd as well as zone-ctor'd.
4285 if (zone->uz_init != NULL) {
4288 kasan_mark_item_valid(zone, item);
4289 error = zone->uz_init(item, zone->uz_size, flags);
4290 kasan_mark_item_invalid(zone, item);
4292 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4296 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4301 counter_u64_add(zone->uz_allocs, 1);
4302 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4303 zone->uz_name, zone);
4308 counter_u64_add(zone->uz_fails, 1);
4310 if (zone->uz_max_items > 0)
4311 zone_free_limit(zone, 1);
4312 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4313 zone->uz_name, zone);
4320 uma_zfree_smr(uma_zone_t zone, void *item)
4323 uma_cache_bucket_t bucket;
4324 int itemdomain, uz_flags;
4326 #ifdef UMA_ZALLOC_DEBUG
4327 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4328 ("uma_zfree_smr: called with non-SMR zone."));
4329 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4330 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4331 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4334 cache = &zone->uz_cpu[curcpu];
4335 uz_flags = cache_uz_flags(cache);
4338 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4339 itemdomain = item_domain(item);
4343 cache = &zone->uz_cpu[curcpu];
4344 /* SMR Zones must free to the free bucket. */
4345 bucket = &cache->uc_freebucket;
4347 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4348 PCPU_GET(domain) != itemdomain) {
4349 bucket = &cache->uc_crossbucket;
4352 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4353 cache_bucket_push(cache, bucket, item);
4357 } while (cache_free(zone, cache, NULL, item, itemdomain));
4361 * If nothing else caught this, we'll just do an internal free.
4363 zone_free_item(zone, item, NULL, SKIP_NONE);
4368 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4371 uma_cache_bucket_t bucket;
4372 int itemdomain, uz_flags;
4374 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4375 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4377 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4379 #ifdef UMA_ZALLOC_DEBUG
4380 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4381 ("uma_zfree_arg: called with SMR zone."));
4382 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4385 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4390 * We are accessing the per-cpu cache without a critical section to
4391 * fetch size and flags. This is acceptable, if we are preempted we
4392 * will simply read another cpu's line.
4394 cache = &zone->uz_cpu[curcpu];
4395 uz_flags = cache_uz_flags(cache);
4396 if (UMA_ALWAYS_CTORDTOR ||
4397 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4398 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4401 * The race here is acceptable. If we miss it we'll just have to wait
4402 * a little longer for the limits to be reset.
4404 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4405 if (atomic_load_32(&zone->uz_sleepers) > 0)
4410 * If possible, free to the per-CPU cache. There are two
4411 * requirements for safe access to the per-CPU cache: (1) the thread
4412 * accessing the cache must not be preempted or yield during access,
4413 * and (2) the thread must not migrate CPUs without switching which
4414 * cache it accesses. We rely on a critical section to prevent
4415 * preemption and migration. We release the critical section in
4416 * order to acquire the zone mutex if we are unable to free to the
4417 * current cache; when we re-acquire the critical section, we must
4418 * detect and handle migration if it has occurred.
4422 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4423 itemdomain = item_domain(item);
4427 cache = &zone->uz_cpu[curcpu];
4429 * Try to free into the allocbucket first to give LIFO
4430 * ordering for cache-hot datastructures. Spill over
4431 * into the freebucket if necessary. Alloc will swap
4432 * them if one runs dry.
4434 bucket = &cache->uc_allocbucket;
4436 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4437 PCPU_GET(domain) != itemdomain) {
4438 bucket = &cache->uc_crossbucket;
4441 if (bucket->ucb_cnt == bucket->ucb_entries &&
4442 cache->uc_freebucket.ucb_cnt <
4443 cache->uc_freebucket.ucb_entries)
4444 cache_bucket_swap(&cache->uc_freebucket,
4445 &cache->uc_allocbucket);
4446 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4447 cache_bucket_push(cache, bucket, item);
4451 } while (cache_free(zone, cache, udata, item, itemdomain));
4455 * If nothing else caught this, we'll just do an internal free.
4458 zone_free_item(zone, item, udata, SKIP_DTOR);
4463 * sort crossdomain free buckets to domain correct buckets and cache
4467 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4469 struct uma_bucketlist emptybuckets, fullbuckets;
4470 uma_zone_domain_t zdom;
4477 "uma_zfree: zone %s(%p) draining cross bucket %p",
4478 zone->uz_name, zone, bucket);
4481 * It is possible for buckets to arrive here out of order so we fetch
4482 * the current smr seq rather than accepting the bucket's.
4484 seq = SMR_SEQ_INVALID;
4485 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4486 seq = smr_advance(zone->uz_smr);
4489 * To avoid having ndomain * ndomain buckets for sorting we have a
4490 * lock on the current crossfree bucket. A full matrix with
4491 * per-domain locking could be used if necessary.
4493 STAILQ_INIT(&emptybuckets);
4494 STAILQ_INIT(&fullbuckets);
4495 ZONE_CROSS_LOCK(zone);
4496 for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4497 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4498 domain = item_domain(item);
4499 zdom = ZDOM_GET(zone, domain);
4500 if (zdom->uzd_cross == NULL) {
4501 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4502 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4503 zdom->uzd_cross = b;
4506 * Avoid allocating a bucket with the cross lock
4507 * held, since allocation can trigger a
4508 * cross-domain free and bucket zones may
4509 * allocate from each other.
4511 ZONE_CROSS_UNLOCK(zone);
4512 b = bucket_alloc(zone, udata, M_NOWAIT);
4515 ZONE_CROSS_LOCK(zone);
4516 if (zdom->uzd_cross != NULL) {
4517 STAILQ_INSERT_HEAD(&emptybuckets, b,
4520 zdom->uzd_cross = b;
4524 b = zdom->uzd_cross;
4525 b->ub_bucket[b->ub_cnt++] = item;
4527 if (b->ub_cnt == b->ub_entries) {
4528 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4529 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4530 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4531 zdom->uzd_cross = b;
4534 ZONE_CROSS_UNLOCK(zone);
4536 if (bucket->ub_cnt == 0)
4537 bucket->ub_seq = SMR_SEQ_INVALID;
4538 bucket_free(zone, bucket, udata);
4540 while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4541 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4542 bucket_free(zone, b, udata);
4544 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4545 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4546 domain = item_domain(b->ub_bucket[0]);
4547 zone_put_bucket(zone, domain, b, udata, true);
4553 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4554 int itemdomain, bool ws)
4559 * Buckets coming from the wrong domain will be entirely for the
4560 * only other domain on two domain systems. In this case we can
4561 * simply cache them. Otherwise we need to sort them back to
4564 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4565 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4566 zone_free_cross(zone, bucket, udata);
4572 * Attempt to save the bucket in the zone's domain bucket cache.
4575 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4576 zone->uz_name, zone, bucket);
4577 /* ub_cnt is pointing to the last free item */
4578 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4579 itemdomain = zone_domain_lowest(zone, itemdomain);
4580 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4584 * Populate a free or cross bucket for the current cpu cache. Free any
4585 * existing full bucket either to the zone cache or back to the slab layer.
4587 * Enters and returns in a critical section. false return indicates that
4588 * we can not satisfy this free in the cache layer. true indicates that
4589 * the caller should retry.
4591 static __noinline bool
4592 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4595 uma_cache_bucket_t cbucket;
4596 uma_bucket_t newbucket, bucket;
4598 CRITICAL_ASSERT(curthread);
4600 if (zone->uz_bucket_size == 0)
4603 cache = &zone->uz_cpu[curcpu];
4607 * FIRSTTOUCH domains need to free to the correct zdom. When
4608 * enabled this is the zdom of the item. The bucket is the
4609 * cross bucket if the current domain and itemdomain do not match.
4611 cbucket = &cache->uc_freebucket;
4613 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4614 if (PCPU_GET(domain) != itemdomain) {
4615 cbucket = &cache->uc_crossbucket;
4616 if (cbucket->ucb_cnt != 0)
4617 counter_u64_add(zone->uz_xdomain,
4622 bucket = cache_bucket_unload(cbucket);
4623 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4624 ("cache_free: Entered with non-full free bucket."));
4626 /* We are no longer associated with this CPU. */
4630 * Don't let SMR zones operate without a free bucket. Force
4631 * a synchronize and re-use this one. We will only degrade
4632 * to a synchronize every bucket_size items rather than every
4633 * item if we fail to allocate a bucket.
4635 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4637 bucket->ub_seq = smr_advance(zone->uz_smr);
4638 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4639 if (newbucket == NULL && bucket != NULL) {
4640 bucket_drain(zone, bucket);
4644 } else if (!bucketdisable)
4645 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4648 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4651 if ((bucket = newbucket) == NULL)
4653 cache = &zone->uz_cpu[curcpu];
4656 * Check to see if we should be populating the cross bucket. If it
4657 * is already populated we will fall through and attempt to populate
4660 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4661 if (PCPU_GET(domain) != itemdomain &&
4662 cache->uc_crossbucket.ucb_bucket == NULL) {
4663 cache_bucket_load_cross(cache, bucket);
4669 * We may have lost the race to fill the bucket or switched CPUs.
4671 if (cache->uc_freebucket.ucb_bucket != NULL) {
4673 bucket_free(zone, bucket, udata);
4676 cache_bucket_load_free(cache, bucket);
4682 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4689 KEG_LOCK_ASSERT(keg, slab->us_domain);
4691 /* Do we need to remove from any lists? */
4692 dom = &keg->uk_domain[slab->us_domain];
4693 if (slab->us_freecount + 1 == keg->uk_ipers) {
4694 LIST_REMOVE(slab, us_link);
4695 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4696 dom->ud_free_slabs++;
4697 } else if (slab->us_freecount == 0) {
4698 LIST_REMOVE(slab, us_link);
4699 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4702 /* Slab management. */
4703 freei = slab_item_index(slab, keg, item);
4704 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4705 slab->us_freecount++;
4707 /* Keg statistics. */
4708 dom->ud_free_items++;
4712 zone_release(void *arg, void **bucket, int cnt)
4725 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4726 lock = KEG_LOCK(keg, 0);
4727 for (i = 0; i < cnt; i++) {
4729 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4730 slab = vtoslab((vm_offset_t)item);
4732 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4733 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4734 slab = hash_sfind(&keg->uk_hash, mem);
4736 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4738 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4741 lock = KEG_LOCK(keg, slab->us_domain);
4743 slab_free_item(zone, slab, item);
4750 * Frees a single item to any zone.
4753 * zone The zone to free to
4754 * item The item we're freeing
4755 * udata User supplied data for the dtor
4756 * skip Skip dtors and finis
4758 static __noinline void
4759 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4763 * If a free is sent directly to an SMR zone we have to
4764 * synchronize immediately because the item can instantly
4765 * be reallocated. This should only happen in degenerate
4766 * cases when no memory is available for per-cpu caches.
4768 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4769 smr_synchronize(zone->uz_smr);
4771 item_dtor(zone, item, zone->uz_size, udata, skip);
4773 if (skip < SKIP_FINI && zone->uz_fini) {
4774 kasan_mark_item_valid(zone, item);
4775 zone->uz_fini(item, zone->uz_size);
4776 kasan_mark_item_invalid(zone, item);
4779 zone->uz_release(zone->uz_arg, &item, 1);
4781 if (skip & SKIP_CNT)
4784 counter_u64_add(zone->uz_frees, 1);
4786 if (zone->uz_max_items > 0)
4787 zone_free_limit(zone, 1);
4792 uma_zone_set_max(uma_zone_t zone, int nitems)
4796 * If the limit is small, we may need to constrain the maximum per-CPU
4797 * cache size, or disable caching entirely.
4799 uma_zone_set_maxcache(zone, nitems);
4802 * XXX This can misbehave if the zone has any allocations with
4803 * no limit and a limit is imposed. There is currently no
4804 * way to clear a limit.
4807 zone->uz_max_items = nitems;
4808 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4809 zone_update_caches(zone);
4810 /* We may need to wake waiters. */
4811 wakeup(&zone->uz_max_items);
4819 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4821 int bpcpu, bpdom, bsize, nb;
4826 * Compute a lower bound on the number of items that may be cached in
4827 * the zone. Each CPU gets at least two buckets, and for cross-domain
4828 * frees we use an additional bucket per CPU and per domain. Select the
4829 * largest bucket size that does not exceed half of the requested limit,
4830 * with the left over space given to the full bucket cache.
4835 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4840 nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4841 bsize = nitems / nb / 2;
4842 if (bsize > BUCKET_MAX)
4844 else if (bsize == 0 && nitems / nb > 0)
4846 zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
4847 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4848 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4849 zone->uz_bucket_max = nitems - nb * bsize;
4855 uma_zone_get_max(uma_zone_t zone)
4859 nitems = atomic_load_64(&zone->uz_max_items);
4866 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4869 ZONE_ASSERT_COLD(zone);
4870 zone->uz_warning = warning;
4875 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4878 ZONE_ASSERT_COLD(zone);
4879 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4884 uma_zone_get_cur(uma_zone_t zone)
4890 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4891 nitems = counter_u64_fetch(zone->uz_allocs) -
4892 counter_u64_fetch(zone->uz_frees);
4894 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4895 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4897 return (nitems < 0 ? 0 : nitems);
4901 uma_zone_get_allocs(uma_zone_t zone)
4907 if (zone->uz_allocs != EARLY_COUNTER)
4908 nitems = counter_u64_fetch(zone->uz_allocs);
4910 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4916 uma_zone_get_frees(uma_zone_t zone)
4922 if (zone->uz_frees != EARLY_COUNTER)
4923 nitems = counter_u64_fetch(zone->uz_frees);
4925 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4931 /* Used only for KEG_ASSERT_COLD(). */
4933 uma_keg_get_allocs(uma_keg_t keg)
4939 LIST_FOREACH(z, &keg->uk_zones, uz_link)
4940 nitems += uma_zone_get_allocs(z);
4948 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4953 KEG_ASSERT_COLD(keg);
4954 keg->uk_init = uminit;
4959 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4964 KEG_ASSERT_COLD(keg);
4965 keg->uk_fini = fini;
4970 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4973 ZONE_ASSERT_COLD(zone);
4974 zone->uz_init = zinit;
4979 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4982 ZONE_ASSERT_COLD(zone);
4983 zone->uz_fini = zfini;
4988 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4993 KEG_ASSERT_COLD(keg);
4994 keg->uk_freef = freef;
4999 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
5004 KEG_ASSERT_COLD(keg);
5005 keg->uk_allocf = allocf;
5010 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
5013 ZONE_ASSERT_COLD(zone);
5015 KASSERT(smr != NULL, ("Got NULL smr"));
5016 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
5017 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
5018 zone->uz_flags |= UMA_ZONE_SMR;
5020 zone_update_caches(zone);
5024 uma_zone_get_smr(uma_zone_t zone)
5027 return (zone->uz_smr);
5032 uma_zone_reserve(uma_zone_t zone, int items)
5037 KEG_ASSERT_COLD(keg);
5038 keg->uk_reserve = items;
5043 uma_zone_reserve_kva(uma_zone_t zone, int count)
5050 KEG_ASSERT_COLD(keg);
5051 ZONE_ASSERT_COLD(zone);
5053 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
5055 #ifdef UMA_MD_SMALL_ALLOC
5056 if (keg->uk_ppera > 1) {
5060 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
5066 MPASS(keg->uk_kva == 0);
5069 zone->uz_max_items = pages * keg->uk_ipers;
5070 #ifdef UMA_MD_SMALL_ALLOC
5071 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
5073 keg->uk_allocf = noobj_alloc;
5075 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5076 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5077 zone_update_caches(zone);
5084 uma_prealloc(uma_zone_t zone, int items)
5086 struct vm_domainset_iter di;
5090 int aflags, domain, slabs;
5093 slabs = howmany(items, keg->uk_ipers);
5094 while (slabs-- > 0) {
5096 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
5099 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
5102 dom = &keg->uk_domain[slab->us_domain];
5104 * keg_alloc_slab() always returns a slab on the
5107 LIST_REMOVE(slab, us_link);
5108 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
5110 dom->ud_free_slabs++;
5111 KEG_UNLOCK(keg, slab->us_domain);
5114 if (vm_domainset_iter_policy(&di, &domain) != 0)
5115 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
5121 * Returns a snapshot of memory consumption in bytes.
5124 uma_zone_memory(uma_zone_t zone)
5130 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
5131 for (i = 0; i < vm_ndomains; i++)
5132 sz += ZDOM_GET(zone, i)->uzd_nitems;
5133 return (sz * zone->uz_size);
5135 for (i = 0; i < vm_ndomains; i++)
5136 sz += zone->uz_keg->uk_domain[i].ud_pages;
5138 return (sz * PAGE_SIZE);
5143 uma_reclaim(int req)
5145 uma_reclaim_domain(req, UMA_ANYDOMAIN);
5149 uma_reclaim_domain(int req, int domain)
5155 arg = (void *)(uintptr_t)domain;
5156 sx_slock(&uma_reclaim_lock);
5158 case UMA_RECLAIM_TRIM:
5159 zone_foreach(zone_trim, arg);
5161 case UMA_RECLAIM_DRAIN:
5162 zone_foreach(zone_drain, arg);
5164 case UMA_RECLAIM_DRAIN_CPU:
5165 zone_foreach(zone_drain, arg);
5166 pcpu_cache_drain_safe(NULL);
5167 zone_foreach(zone_drain, arg);
5170 panic("unhandled reclamation request %d", req);
5174 * Some slabs may have been freed but this zone will be visited early
5175 * we visit again so that we can free pages that are empty once other
5176 * zones are drained. We have to do the same for buckets.
5178 zone_drain(slabzones[0], arg);
5179 zone_drain(slabzones[1], arg);
5180 bucket_zone_drain(domain);
5181 sx_sunlock(&uma_reclaim_lock);
5184 static volatile int uma_reclaim_needed;
5187 uma_reclaim_wakeup(void)
5190 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5191 wakeup(uma_reclaim);
5195 uma_reclaim_worker(void *arg __unused)
5199 sx_xlock(&uma_reclaim_lock);
5200 while (atomic_load_int(&uma_reclaim_needed) == 0)
5201 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5203 sx_xunlock(&uma_reclaim_lock);
5204 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5205 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5206 atomic_store_int(&uma_reclaim_needed, 0);
5207 /* Don't fire more than once per-second. */
5208 pause("umarclslp", hz);
5214 uma_zone_reclaim(uma_zone_t zone, int req)
5216 uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5220 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5224 arg = (void *)(uintptr_t)domain;
5226 case UMA_RECLAIM_TRIM:
5227 zone_trim(zone, arg);
5229 case UMA_RECLAIM_DRAIN:
5230 zone_drain(zone, arg);
5232 case UMA_RECLAIM_DRAIN_CPU:
5233 pcpu_cache_drain_safe(zone);
5234 zone_drain(zone, arg);
5237 panic("unhandled reclamation request %d", req);
5243 uma_zone_exhausted(uma_zone_t zone)
5246 return (atomic_load_32(&zone->uz_sleepers) > 0);
5253 return (uma_kmem_limit);
5257 uma_set_limit(unsigned long limit)
5260 uma_kmem_limit = limit;
5267 return (atomic_load_long(&uma_kmem_total));
5274 return (uma_kmem_limit - uma_size());
5279 * Generate statistics across both the zone and its per-cpu cache's. Return
5280 * desired statistics if the pointer is non-NULL for that statistic.
5282 * Note: does not update the zone statistics, as it can't safely clear the
5283 * per-CPU cache statistic.
5287 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5288 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5291 uint64_t allocs, frees, sleeps, xdomain;
5294 allocs = frees = sleeps = xdomain = 0;
5297 cache = &z->uz_cpu[cpu];
5298 cachefree += cache->uc_allocbucket.ucb_cnt;
5299 cachefree += cache->uc_freebucket.ucb_cnt;
5300 xdomain += cache->uc_crossbucket.ucb_cnt;
5301 cachefree += cache->uc_crossbucket.ucb_cnt;
5302 allocs += cache->uc_allocs;
5303 frees += cache->uc_frees;
5305 allocs += counter_u64_fetch(z->uz_allocs);
5306 frees += counter_u64_fetch(z->uz_frees);
5307 xdomain += counter_u64_fetch(z->uz_xdomain);
5308 sleeps += z->uz_sleeps;
5309 if (cachefreep != NULL)
5310 *cachefreep = cachefree;
5311 if (allocsp != NULL)
5315 if (sleepsp != NULL)
5317 if (xdomainp != NULL)
5318 *xdomainp = xdomain;
5323 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5330 rw_rlock(&uma_rwlock);
5331 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5332 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5335 LIST_FOREACH(z, &uma_cachezones, uz_link)
5338 rw_runlock(&uma_rwlock);
5339 return (sysctl_handle_int(oidp, &count, 0, req));
5343 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5344 struct uma_percpu_stat *ups, bool internal)
5346 uma_zone_domain_t zdom;
5350 for (i = 0; i < vm_ndomains; i++) {
5351 zdom = ZDOM_GET(z, i);
5352 uth->uth_zone_free += zdom->uzd_nitems;
5354 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5355 uth->uth_frees = counter_u64_fetch(z->uz_frees);
5356 uth->uth_fails = counter_u64_fetch(z->uz_fails);
5357 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5358 uth->uth_sleeps = z->uz_sleeps;
5360 for (i = 0; i < mp_maxid + 1; i++) {
5361 bzero(&ups[i], sizeof(*ups));
5362 if (internal || CPU_ABSENT(i))
5364 cache = &z->uz_cpu[i];
5365 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5366 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5367 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5368 ups[i].ups_allocs = cache->uc_allocs;
5369 ups[i].ups_frees = cache->uc_frees;
5374 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5376 struct uma_stream_header ush;
5377 struct uma_type_header uth;
5378 struct uma_percpu_stat *ups;
5383 uint32_t kfree, pages;
5384 int count, error, i;
5386 error = sysctl_wire_old_buffer(req, 0);
5389 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5390 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5391 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5394 rw_rlock(&uma_rwlock);
5395 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5396 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5400 LIST_FOREACH(z, &uma_cachezones, uz_link)
5404 * Insert stream header.
5406 bzero(&ush, sizeof(ush));
5407 ush.ush_version = UMA_STREAM_VERSION;
5408 ush.ush_maxcpus = (mp_maxid + 1);
5409 ush.ush_count = count;
5410 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5412 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5414 for (i = 0; i < vm_ndomains; i++) {
5415 kfree += kz->uk_domain[i].ud_free_items;
5416 pages += kz->uk_domain[i].ud_pages;
5418 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5419 bzero(&uth, sizeof(uth));
5420 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5421 uth.uth_align = kz->uk_align;
5422 uth.uth_size = kz->uk_size;
5423 uth.uth_rsize = kz->uk_rsize;
5424 if (z->uz_max_items > 0) {
5425 items = UZ_ITEMS_COUNT(z->uz_items);
5426 uth.uth_pages = (items / kz->uk_ipers) *
5429 uth.uth_pages = pages;
5430 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5432 uth.uth_limit = z->uz_max_items;
5433 uth.uth_keg_free = kfree;
5436 * A zone is secondary is it is not the first entry
5437 * on the keg's zone list.
5439 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5440 (LIST_FIRST(&kz->uk_zones) != z))
5441 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5442 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5443 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5444 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5445 for (i = 0; i < mp_maxid + 1; i++)
5446 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5449 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5450 bzero(&uth, sizeof(uth));
5451 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5452 uth.uth_size = z->uz_size;
5453 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5454 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5455 for (i = 0; i < mp_maxid + 1; i++)
5456 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5459 rw_runlock(&uma_rwlock);
5460 error = sbuf_finish(&sbuf);
5467 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5469 uma_zone_t zone = *(uma_zone_t *)arg1;
5472 max = uma_zone_get_max(zone);
5473 error = sysctl_handle_int(oidp, &max, 0, req);
5474 if (error || !req->newptr)
5477 uma_zone_set_max(zone, max);
5483 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5489 * Some callers want to add sysctls for global zones that
5490 * may not yet exist so they pass a pointer to a pointer.
5493 zone = *(uma_zone_t *)arg1;
5496 cur = uma_zone_get_cur(zone);
5497 return (sysctl_handle_int(oidp, &cur, 0, req));
5501 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5503 uma_zone_t zone = arg1;
5506 cur = uma_zone_get_allocs(zone);
5507 return (sysctl_handle_64(oidp, &cur, 0, req));
5511 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5513 uma_zone_t zone = arg1;
5516 cur = uma_zone_get_frees(zone);
5517 return (sysctl_handle_64(oidp, &cur, 0, req));
5521 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5524 uma_zone_t zone = arg1;
5527 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5528 if (zone->uz_flags != 0)
5529 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5531 sbuf_printf(&sbuf, "0");
5532 error = sbuf_finish(&sbuf);
5539 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5541 uma_keg_t keg = arg1;
5542 int avail, effpct, total;
5544 total = keg->uk_ppera * PAGE_SIZE;
5545 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5546 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5548 * We consider the client's requested size and alignment here, not the
5549 * real size determination uk_rsize, because we also adjust the real
5550 * size for internal implementation reasons (max bitset size).
5552 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5553 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5554 avail *= mp_maxid + 1;
5555 effpct = 100 * avail / total;
5556 return (sysctl_handle_int(oidp, &effpct, 0, req));
5560 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5562 uma_zone_t zone = arg1;
5565 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5566 return (sysctl_handle_64(oidp, &cur, 0, req));
5571 uma_dbg_getslab(uma_zone_t zone, void *item)
5578 * It is safe to return the slab here even though the
5579 * zone is unlocked because the item's allocation state
5580 * essentially holds a reference.
5582 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5583 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5585 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5586 return (vtoslab((vm_offset_t)mem));
5588 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5589 return ((uma_slab_t)(mem + keg->uk_pgoff));
5591 slab = hash_sfind(&keg->uk_hash, mem);
5598 uma_dbg_zskip(uma_zone_t zone, void *mem)
5601 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5604 return (uma_dbg_kskip(zone->uz_keg, mem));
5608 uma_dbg_kskip(uma_keg_t keg, void *mem)
5612 if (dbg_divisor == 0)
5615 if (dbg_divisor == 1)
5618 idx = (uintptr_t)mem >> PAGE_SHIFT;
5619 if (keg->uk_ipers > 1) {
5620 idx *= keg->uk_ipers;
5621 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5624 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5625 counter_u64_add(uma_skip_cnt, 1);
5628 counter_u64_add(uma_dbg_cnt, 1);
5634 * Set up the slab's freei data such that uma_dbg_free can function.
5638 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5644 slab = uma_dbg_getslab(zone, item);
5646 panic("uma: item %p did not belong to zone %s",
5647 item, zone->uz_name);
5650 freei = slab_item_index(slab, keg, item);
5652 if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5653 slab_dbg_bits(slab, keg)))
5654 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5655 item, zone, zone->uz_name, slab, freei);
5659 * Verifies freed addresses. Checks for alignment, valid slab membership
5660 * and duplicate frees.
5664 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5670 slab = uma_dbg_getslab(zone, item);
5672 panic("uma: Freed item %p did not belong to zone %s",
5673 item, zone->uz_name);
5676 freei = slab_item_index(slab, keg, item);
5678 if (freei >= keg->uk_ipers)
5679 panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5680 item, zone, zone->uz_name, slab, freei);
5682 if (slab_item(slab, keg, freei) != item)
5683 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5684 item, zone, zone->uz_name, slab, freei);
5686 if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5687 slab_dbg_bits(slab, keg)))
5688 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5689 item, zone, zone->uz_name, slab, freei);
5691 #endif /* INVARIANTS */
5695 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5696 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5701 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5702 *allocs = counter_u64_fetch(z->uz_allocs);
5703 frees = counter_u64_fetch(z->uz_frees);
5704 *sleeps = z->uz_sleeps;
5708 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5710 for (i = 0; i < vm_ndomains; i++) {
5711 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5712 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5713 (LIST_FIRST(&kz->uk_zones) != z)))
5714 *cachefree += kz->uk_domain[i].ud_free_items;
5716 *used = *allocs - frees;
5717 return (((int64_t)*used + *cachefree) * kz->uk_size);
5720 DB_SHOW_COMMAND(uma, db_show_uma)
5722 const char *fmt_hdr, *fmt_entry;
5725 uint64_t allocs, used, sleeps, xdomain;
5727 /* variables for sorting */
5729 uma_zone_t cur_zone, last_zone;
5730 int64_t cur_size, last_size, size;
5733 /* /i option produces machine-parseable CSV output */
5734 if (modif[0] == 'i') {
5735 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5736 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5738 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5739 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5742 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5743 "Sleeps", "Bucket", "Total Mem", "XFree");
5745 /* Sort the zones with largest size first. */
5747 last_size = INT64_MAX;
5752 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5753 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5755 * In the case of size ties, print out zones
5756 * in the order they are encountered. That is,
5757 * when we encounter the most recently output
5758 * zone, we have already printed all preceding
5759 * ties, and we must print all following ties.
5761 if (z == last_zone) {
5765 size = get_uma_stats(kz, z, &allocs, &used,
5766 &sleeps, &cachefree, &xdomain);
5767 if (size > cur_size && size < last_size + ties)
5775 if (cur_zone == NULL)
5778 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5779 &sleeps, &cachefree, &xdomain);
5780 db_printf(fmt_entry, cur_zone->uz_name,
5781 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5782 (uintmax_t)allocs, (uintmax_t)sleeps,
5783 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5788 last_zone = cur_zone;
5789 last_size = cur_size;
5793 DB_SHOW_COMMAND(umacache, db_show_umacache)
5796 uint64_t allocs, frees;
5800 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5801 "Requests", "Bucket");
5802 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5803 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5804 for (i = 0; i < vm_ndomains; i++)
5805 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5806 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5807 z->uz_name, (uintmax_t)z->uz_size,
5808 (intmax_t)(allocs - frees), cachefree,
5809 (uintmax_t)allocs, z->uz_bucket_size);