2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2002-2019 Jeffrey Roberson <jeff@FreeBSD.org>
5 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
6 * Copyright (c) 2004-2006 Robert N. M. Watson
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice unmodified, this list of conditions, and the following
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
20 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
21 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
22 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
24 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
28 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32 * uma_core.c Implementation of the Universal Memory allocator
34 * This allocator is intended to replace the multitude of similar object caches
35 * in the standard FreeBSD kernel. The intent is to be flexible as well as
36 * efficient. A primary design goal is to return unused memory to the rest of
37 * the system. This will make the system as a whole more flexible due to the
38 * ability to move memory to subsystems which most need it instead of leaving
39 * pools of reserved memory unused.
41 * The basic ideas stem from similar slab/zone based allocators whose algorithms
48 * - Improve memory usage for large allocations
49 * - Investigate cache size adjustments
52 #include <sys/cdefs.h>
53 __FBSDID("$FreeBSD$");
56 #include "opt_param.h"
59 #include <sys/param.h>
60 #include <sys/systm.h>
61 #include <sys/bitset.h>
62 #include <sys/domainset.h>
63 #include <sys/eventhandler.h>
64 #include <sys/kernel.h>
65 #include <sys/types.h>
66 #include <sys/limits.h>
67 #include <sys/queue.h>
68 #include <sys/malloc.h>
71 #include <sys/sysctl.h>
72 #include <sys/mutex.h>
74 #include <sys/random.h>
75 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
81 #include <sys/taskqueue.h>
82 #include <sys/vmmeter.h>
85 #include <vm/vm_domainset.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_param.h>
90 #include <vm/vm_phys.h>
91 #include <vm/vm_pagequeue.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
96 #include <vm/uma_int.h>
97 #include <vm/uma_dbg.h>
101 #ifdef DEBUG_MEMGUARD
102 #include <vm/memguard.h>
105 #include <machine/md_var.h>
108 #define UMA_ALWAYS_CTORDTOR 1
110 #define UMA_ALWAYS_CTORDTOR 0
114 * This is the zone and keg from which all zones are spawned.
116 static uma_zone_t kegs;
117 static uma_zone_t zones;
120 * These are the two zones from which all offpage uma_slab_ts are allocated.
122 * One zone is for slab headers that can represent a larger number of items,
123 * making the slabs themselves more efficient, and the other zone is for
124 * headers that are smaller and represent fewer items, making the headers more
127 #define SLABZONE_SIZE(setsize) \
128 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
129 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16)
130 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE
131 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE)
132 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE)
133 static uma_zone_t slabzones[2];
136 * The initial hash tables come out of this zone so they can be allocated
137 * prior to malloc coming up.
139 static uma_zone_t hashzone;
141 /* The boot-time adjusted value for cache line alignment. */
142 int uma_align_cache = 64 - 1;
144 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
145 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
148 * Are we allowed to allocate buckets?
150 static int bucketdisable = 1;
152 /* Linked list of all kegs in the system */
153 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
155 /* Linked list of all cache-only zones in the system */
156 static LIST_HEAD(,uma_zone) uma_cachezones =
157 LIST_HEAD_INITIALIZER(uma_cachezones);
159 /* This RW lock protects the keg list */
160 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
163 * First available virual address for boot time allocations.
165 static vm_offset_t bootstart;
166 static vm_offset_t bootmem;
168 static struct sx uma_reclaim_lock;
171 * kmem soft limit, initialized by uma_set_limit(). Ensure that early
172 * allocations don't trigger a wakeup of the reclaim thread.
174 unsigned long uma_kmem_limit = LONG_MAX;
175 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
176 "UMA kernel memory soft limit");
177 unsigned long uma_kmem_total;
178 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
179 "UMA kernel memory usage");
181 /* Is the VM done starting up? */
187 } booted = BOOT_COLD;
190 * This is the handle used to schedule events that need to happen
191 * outside of the allocation fast path.
193 static struct callout uma_callout;
194 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
197 * This structure is passed as the zone ctor arg so that I don't have to create
198 * a special allocation function just for zones.
200 struct uma_zctor_args {
215 struct uma_kctor_args {
224 struct uma_bucket_zone {
227 int ubz_entries; /* Number of items it can hold. */
228 int ubz_maxsize; /* Maximum allocation size per-item. */
232 * Compute the actual number of bucket entries to pack them in power
233 * of two sizes for more efficient space utilization.
235 #define BUCKET_SIZE(n) \
236 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
238 #define BUCKET_MAX BUCKET_SIZE(256)
241 struct uma_bucket_zone bucket_zones[] = {
242 /* Literal bucket sizes. */
243 { NULL, "2 Bucket", 2, 4096 },
244 { NULL, "4 Bucket", 4, 3072 },
245 { NULL, "8 Bucket", 8, 2048 },
246 { NULL, "16 Bucket", 16, 1024 },
247 /* Rounded down power of 2 sizes for efficiency. */
248 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
249 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
250 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
251 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
256 * Flags and enumerations to be passed to internal functions.
260 SKIP_CNT = 0x00000001,
261 SKIP_DTOR = 0x00010000,
262 SKIP_FINI = 0x00020000,
267 void uma_startup1(vm_offset_t);
268 void uma_startup2(void);
270 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
271 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
272 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
273 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
274 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
275 static void page_free(void *, vm_size_t, uint8_t);
276 static void pcpu_page_free(void *, vm_size_t, uint8_t);
277 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
278 static void cache_drain(uma_zone_t);
279 static void bucket_drain(uma_zone_t, uma_bucket_t);
280 static void bucket_cache_reclaim(uma_zone_t zone, bool);
281 static int keg_ctor(void *, int, void *, int);
282 static void keg_dtor(void *, int, void *);
283 static int zone_ctor(void *, int, void *, int);
284 static void zone_dtor(void *, int, void *);
285 static inline void item_dtor(uma_zone_t zone, void *item, int size,
286 void *udata, enum zfreeskip skip);
287 static int zero_init(void *, int, int);
288 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
289 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
290 static void zone_timeout(uma_zone_t zone, void *);
291 static int hash_alloc(struct uma_hash *, u_int);
292 static int hash_expand(struct uma_hash *, struct uma_hash *);
293 static void hash_free(struct uma_hash *hash);
294 static void uma_timeout(void *);
295 static void uma_startup3(void);
296 static void uma_shutdown(void);
297 static void *zone_alloc_item(uma_zone_t, void *, int, int);
298 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
299 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
300 static void zone_free_limit(uma_zone_t zone, int count);
301 static void bucket_enable(void);
302 static void bucket_init(void);
303 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
304 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
305 static void bucket_zone_drain(void);
306 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
307 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
308 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
309 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
310 uma_fini fini, int align, uint32_t flags);
311 static int zone_import(void *, void **, int, int, int);
312 static void zone_release(void *, void **, int);
313 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
314 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
316 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
317 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
318 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
319 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
320 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
321 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
322 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
324 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
326 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD, 0,
327 "Memory allocation debugging");
330 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
331 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
333 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
334 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
335 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
336 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
338 static u_int dbg_divisor = 1;
339 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
340 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
341 "Debug & thrash every this item in memory allocator");
343 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
344 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
345 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
346 &uma_dbg_cnt, "memory items debugged");
347 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
348 &uma_skip_cnt, "memory items skipped, not debugged");
351 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
353 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW, 0, "Universal Memory Allocator");
355 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
356 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
358 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
359 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
361 static int zone_warnings = 1;
362 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
363 "Warn when UMA zones becomes full");
365 static int multipage_slabs = 1;
366 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
367 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
368 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
369 "UMA may choose larger slab sizes for better efficiency");
372 * Select the slab zone for an offpage slab with the given maximum item count.
374 static inline uma_zone_t
378 return (slabzones[ipers > SLABZONE0_SETSIZE]);
382 * This routine checks to see whether or not it's safe to enable buckets.
388 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
389 bucketdisable = vm_page_count_min();
393 * Initialize bucket_zones, the array of zones of buckets of various sizes.
395 * For each zone, calculate the memory required for each bucket, consisting
396 * of the header and an array of pointers.
401 struct uma_bucket_zone *ubz;
404 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
405 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
406 size += sizeof(void *) * ubz->ubz_entries;
407 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
408 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
409 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
410 UMA_ZONE_FIRSTTOUCH);
415 * Given a desired number of entries for a bucket, return the zone from which
416 * to allocate the bucket.
418 static struct uma_bucket_zone *
419 bucket_zone_lookup(int entries)
421 struct uma_bucket_zone *ubz;
423 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
424 if (ubz->ubz_entries >= entries)
430 static struct uma_bucket_zone *
431 bucket_zone_max(uma_zone_t zone, int nitems)
433 struct uma_bucket_zone *ubz;
437 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
438 /* Count the cross-domain bucket. */
441 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
442 if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems)
444 if (ubz == &bucket_zones[0])
452 bucket_select(int size)
454 struct uma_bucket_zone *ubz;
456 ubz = &bucket_zones[0];
457 if (size > ubz->ubz_maxsize)
458 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
460 for (; ubz->ubz_entries != 0; ubz++)
461 if (ubz->ubz_maxsize < size)
464 return (ubz->ubz_entries);
468 bucket_alloc(uma_zone_t zone, void *udata, int flags)
470 struct uma_bucket_zone *ubz;
474 * Don't allocate buckets early in boot.
476 if (__predict_false(booted < BOOT_KVA))
480 * To limit bucket recursion we store the original zone flags
481 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
482 * NOVM flag to persist even through deep recursions. We also
483 * store ZFLAG_BUCKET once we have recursed attempting to allocate
484 * a bucket for a bucket zone so we do not allow infinite bucket
485 * recursion. This cookie will even persist to frees of unused
486 * buckets via the allocation path or bucket allocations in the
489 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
490 udata = (void *)(uintptr_t)zone->uz_flags;
492 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
494 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
496 if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
498 ubz = bucket_zone_lookup(zone->uz_bucket_size);
499 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
501 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
504 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
507 bucket->ub_entries = ubz->ubz_entries;
508 bucket->ub_seq = SMR_SEQ_INVALID;
509 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
510 zone->uz_name, zone, bucket);
517 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
519 struct uma_bucket_zone *ubz;
521 KASSERT(bucket->ub_cnt == 0,
522 ("bucket_free: Freeing a non free bucket."));
523 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
524 ("bucket_free: Freeing an SMR bucket."));
525 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
526 udata = (void *)(uintptr_t)zone->uz_flags;
527 ubz = bucket_zone_lookup(bucket->ub_entries);
528 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
532 bucket_zone_drain(void)
534 struct uma_bucket_zone *ubz;
536 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
537 uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN);
541 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
542 * zone's caches. If a bucket is found the zone is not locked on return.
545 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom)
551 ZONE_LOCK_ASSERT(zone);
553 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
556 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
557 bucket->ub_seq != SMR_SEQ_INVALID) {
558 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
560 bucket->ub_seq = SMR_SEQ_INVALID;
561 dtor = (zone->uz_dtor != NULL) | UMA_ALWAYS_CTORDTOR;
563 MPASS(zdom->uzd_nitems >= bucket->ub_cnt);
564 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
565 zdom->uzd_nitems -= bucket->ub_cnt;
566 if (zdom->uzd_imin > zdom->uzd_nitems)
567 zdom->uzd_imin = zdom->uzd_nitems;
568 zone->uz_bkt_count -= bucket->ub_cnt;
571 for (i = 0; i < bucket->ub_cnt; i++)
572 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
579 * Insert a full bucket into the specified cache. The "ws" parameter indicates
580 * whether the bucket's contents should be counted as part of the zone's working
584 zone_put_bucket(uma_zone_t zone, uma_zone_domain_t zdom, uma_bucket_t bucket,
588 ZONE_LOCK_ASSERT(zone);
589 KASSERT(!ws || zone->uz_bkt_count < zone->uz_bkt_max,
590 ("%s: zone %p overflow", __func__, zone));
592 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
593 zdom->uzd_nitems += bucket->ub_cnt;
594 if (ws && zdom->uzd_imax < zdom->uzd_nitems)
595 zdom->uzd_imax = zdom->uzd_nitems;
596 zone->uz_bkt_count += bucket->ub_cnt;
599 /* Pops an item out of a per-cpu cache bucket. */
601 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
605 CRITICAL_ASSERT(curthread);
608 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
610 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
611 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
618 /* Pushes an item into a per-cpu cache bucket. */
620 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
623 CRITICAL_ASSERT(curthread);
624 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
625 ("uma_zfree: Freeing to non free bucket index."));
627 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
633 * Unload a UMA bucket from a per-cpu cache.
635 static inline uma_bucket_t
636 cache_bucket_unload(uma_cache_bucket_t bucket)
640 b = bucket->ucb_bucket;
642 MPASS(b->ub_entries == bucket->ucb_entries);
643 b->ub_cnt = bucket->ucb_cnt;
644 bucket->ucb_bucket = NULL;
645 bucket->ucb_entries = bucket->ucb_cnt = 0;
651 static inline uma_bucket_t
652 cache_bucket_unload_alloc(uma_cache_t cache)
655 return (cache_bucket_unload(&cache->uc_allocbucket));
658 static inline uma_bucket_t
659 cache_bucket_unload_free(uma_cache_t cache)
662 return (cache_bucket_unload(&cache->uc_freebucket));
665 static inline uma_bucket_t
666 cache_bucket_unload_cross(uma_cache_t cache)
669 return (cache_bucket_unload(&cache->uc_crossbucket));
673 * Load a bucket into a per-cpu cache bucket.
676 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
679 CRITICAL_ASSERT(curthread);
680 MPASS(bucket->ucb_bucket == NULL);
682 bucket->ucb_bucket = b;
683 bucket->ucb_cnt = b->ub_cnt;
684 bucket->ucb_entries = b->ub_entries;
688 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
691 cache_bucket_load(&cache->uc_allocbucket, b);
695 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
698 cache_bucket_load(&cache->uc_freebucket, b);
703 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
706 cache_bucket_load(&cache->uc_crossbucket, b);
711 * Copy and preserve ucb_spare.
714 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
717 b1->ucb_bucket = b2->ucb_bucket;
718 b1->ucb_entries = b2->ucb_entries;
719 b1->ucb_cnt = b2->ucb_cnt;
723 * Swap two cache buckets.
726 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
728 struct uma_cache_bucket b3;
730 CRITICAL_ASSERT(curthread);
732 cache_bucket_copy(&b3, b1);
733 cache_bucket_copy(b1, b2);
734 cache_bucket_copy(b2, &b3);
738 zone_log_warning(uma_zone_t zone)
740 static const struct timeval warninterval = { 300, 0 };
742 if (!zone_warnings || zone->uz_warning == NULL)
745 if (ratecheck(&zone->uz_ratecheck, &warninterval))
746 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
750 zone_maxaction(uma_zone_t zone)
753 if (zone->uz_maxaction.ta_func != NULL)
754 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
758 * Routine called by timeout which is used to fire off some time interval
759 * based calculations. (stats, hash size, etc.)
768 uma_timeout(void *unused)
771 zone_foreach(zone_timeout, NULL);
773 /* Reschedule this event */
774 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
778 * Update the working set size estimate for the zone's bucket cache.
779 * The constants chosen here are somewhat arbitrary. With an update period of
780 * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
784 zone_domain_update_wss(uma_zone_domain_t zdom)
788 MPASS(zdom->uzd_imax >= zdom->uzd_imin);
789 wss = zdom->uzd_imax - zdom->uzd_imin;
790 zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
791 zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
795 * Routine to perform timeout driven calculations. This expands the
796 * hashes and does per cpu statistics aggregation.
801 zone_timeout(uma_zone_t zone, void *unused)
806 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
812 * Hash zones are non-numa by definition so the first domain
813 * is the only one present.
816 pages = keg->uk_domain[0].ud_pages;
819 * Expand the keg hash table.
821 * This is done if the number of slabs is larger than the hash size.
822 * What I'm trying to do here is completely reduce collisions. This
823 * may be a little aggressive. Should I allow for two collisions max?
825 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
826 struct uma_hash newhash;
827 struct uma_hash oldhash;
831 * This is so involved because allocating and freeing
832 * while the keg lock is held will lead to deadlock.
833 * I have to do everything in stages and check for
837 ret = hash_alloc(&newhash, 1 << fls(slabs));
840 if (hash_expand(&keg->uk_hash, &newhash)) {
841 oldhash = keg->uk_hash;
842 keg->uk_hash = newhash;
855 for (int i = 0; i < vm_ndomains; i++)
856 zone_domain_update_wss(&zone->uz_domain[i]);
861 * Allocate and zero fill the next sized hash table from the appropriate
865 * hash A new hash structure with the old hash size in uh_hashsize
868 * 1 on success and 0 on failure.
871 hash_alloc(struct uma_hash *hash, u_int size)
875 KASSERT(powerof2(size), ("hash size must be power of 2"));
876 if (size > UMA_HASH_SIZE_INIT) {
877 hash->uh_hashsize = size;
878 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
879 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
881 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
882 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
883 UMA_ANYDOMAIN, M_WAITOK);
884 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
886 if (hash->uh_slab_hash) {
887 bzero(hash->uh_slab_hash, alloc);
888 hash->uh_hashmask = hash->uh_hashsize - 1;
896 * Expands the hash table for HASH zones. This is done from zone_timeout
897 * to reduce collisions. This must not be done in the regular allocation
898 * path, otherwise, we can recurse on the vm while allocating pages.
901 * oldhash The hash you want to expand
902 * newhash The hash structure for the new table
910 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
912 uma_hash_slab_t slab;
916 if (!newhash->uh_slab_hash)
919 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
923 * I need to investigate hash algorithms for resizing without a
927 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
928 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
929 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
930 LIST_REMOVE(slab, uhs_hlink);
931 hval = UMA_HASH(newhash, slab->uhs_data);
932 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
940 * Free the hash bucket to the appropriate backing store.
943 * slab_hash The hash bucket we're freeing
944 * hashsize The number of entries in that hash bucket
950 hash_free(struct uma_hash *hash)
952 if (hash->uh_slab_hash == NULL)
954 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
955 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
957 free(hash->uh_slab_hash, M_UMAHASH);
961 * Frees all outstanding items in a bucket
964 * zone The zone to free to, must be unlocked.
965 * bucket The free/alloc bucket with items.
972 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
976 if (bucket == NULL || bucket->ub_cnt == 0)
979 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
980 bucket->ub_seq != SMR_SEQ_INVALID) {
981 smr_wait(zone->uz_smr, bucket->ub_seq);
982 for (i = 0; i < bucket->ub_cnt; i++)
983 item_dtor(zone, bucket->ub_bucket[i],
984 zone->uz_size, NULL, SKIP_NONE);
985 bucket->ub_seq = SMR_SEQ_INVALID;
988 for (i = 0; i < bucket->ub_cnt; i++)
989 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
990 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
991 if (zone->uz_max_items > 0)
992 zone_free_limit(zone, bucket->ub_cnt);
994 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1000 * Drains the per cpu caches for a zone.
1002 * NOTE: This may only be called while the zone is being torn down, and not
1003 * during normal operation. This is necessary in order that we do not have
1004 * to migrate CPUs to drain the per-CPU caches.
1007 * zone The zone to drain, must be unlocked.
1013 cache_drain(uma_zone_t zone)
1016 uma_bucket_t bucket;
1020 * XXX: It is safe to not lock the per-CPU caches, because we're
1021 * tearing down the zone anyway. I.e., there will be no further use
1022 * of the caches at this point.
1024 * XXX: It would good to be able to assert that the zone is being
1025 * torn down to prevent improper use of cache_drain().
1028 cache = &zone->uz_cpu[cpu];
1029 bucket = cache_bucket_unload_alloc(cache);
1030 if (bucket != NULL) {
1031 bucket_drain(zone, bucket);
1032 bucket_free(zone, bucket, NULL);
1034 bucket = cache_bucket_unload_free(cache);
1035 if (bucket != NULL) {
1036 bucket_drain(zone, bucket);
1037 bucket_free(zone, bucket, NULL);
1039 bucket = cache_bucket_unload_cross(cache);
1040 if (bucket != NULL) {
1041 bucket_drain(zone, bucket);
1042 bucket_free(zone, bucket, NULL);
1045 bucket_cache_reclaim(zone, true);
1049 cache_shrink(uma_zone_t zone, void *unused)
1052 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1056 zone->uz_bucket_size =
1057 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1062 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1065 uma_bucket_t b1, b2, b3;
1068 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1071 b1 = b2 = b3 = NULL;
1074 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
1075 domain = PCPU_GET(domain);
1078 cache = &zone->uz_cpu[curcpu];
1079 b1 = cache_bucket_unload_alloc(cache);
1080 if (b1 != NULL && b1->ub_cnt != 0) {
1081 zone_put_bucket(zone, &zone->uz_domain[domain], b1, false);
1086 * Don't flush SMR zone buckets. This leaves the zone without a
1087 * bucket and forces every free to synchronize().
1089 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1091 b2 = cache_bucket_unload_free(cache);
1092 if (b2 != NULL && b2->ub_cnt != 0) {
1093 zone_put_bucket(zone, &zone->uz_domain[domain], b2, false);
1096 b3 = cache_bucket_unload_cross(cache);
1102 bucket_free(zone, b1, NULL);
1104 bucket_free(zone, b2, NULL);
1106 bucket_drain(zone, b3);
1107 bucket_free(zone, b3, NULL);
1112 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1113 * This is an expensive call because it needs to bind to all CPUs
1114 * one by one and enter a critical section on each of them in order
1115 * to safely access their cache buckets.
1116 * Zone lock must not be held on call this function.
1119 pcpu_cache_drain_safe(uma_zone_t zone)
1124 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1127 cache_shrink(zone, NULL);
1129 zone_foreach(cache_shrink, NULL);
1132 thread_lock(curthread);
1133 sched_bind(curthread, cpu);
1134 thread_unlock(curthread);
1137 cache_drain_safe_cpu(zone, NULL);
1139 zone_foreach(cache_drain_safe_cpu, NULL);
1141 thread_lock(curthread);
1142 sched_unbind(curthread);
1143 thread_unlock(curthread);
1147 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1148 * requested a drain, otherwise the per-domain caches are trimmed to either
1149 * estimated working set size.
1152 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1154 uma_zone_domain_t zdom;
1155 uma_bucket_t bucket;
1156 long target, tofree;
1159 for (i = 0; i < vm_ndomains; i++) {
1161 * The cross bucket is partially filled and not part of
1162 * the item count. Reclaim it individually here.
1164 zdom = &zone->uz_domain[i];
1165 ZONE_CROSS_LOCK(zone);
1166 bucket = zdom->uzd_cross;
1167 zdom->uzd_cross = NULL;
1168 ZONE_CROSS_UNLOCK(zone);
1169 if (bucket != NULL) {
1170 bucket_drain(zone, bucket);
1171 bucket_free(zone, bucket, NULL);
1175 * Shrink the zone bucket size to ensure that the per-CPU caches
1176 * don't grow too large.
1179 if (i == 0 && zone->uz_bucket_size > zone->uz_bucket_size_min)
1180 zone->uz_bucket_size--;
1183 * If we were asked to drain the zone, we are done only once
1184 * this bucket cache is empty. Otherwise, we reclaim items in
1185 * excess of the zone's estimated working set size. If the
1186 * difference nitems - imin is larger than the WSS estimate,
1187 * then the estimate will grow at the end of this interval and
1188 * we ignore the historical average.
1190 target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1192 while (zdom->uzd_nitems > target) {
1193 bucket = STAILQ_FIRST(&zdom->uzd_buckets);
1196 tofree = bucket->ub_cnt;
1197 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
1198 zdom->uzd_nitems -= tofree;
1201 * Shift the bounds of the current WSS interval to avoid
1202 * perturbing the estimate.
1204 zdom->uzd_imax -= lmin(zdom->uzd_imax, tofree);
1205 zdom->uzd_imin -= lmin(zdom->uzd_imin, tofree);
1208 bucket_drain(zone, bucket);
1209 bucket_free(zone, bucket, NULL);
1217 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1223 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1224 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1226 mem = slab_data(slab, keg);
1227 flags = slab->us_flags;
1229 if (keg->uk_fini != NULL) {
1230 for (i--; i > -1; i--)
1233 * trash_fini implies that dtor was trash_dtor. trash_fini
1234 * would check that memory hasn't been modified since free,
1235 * which executed trash_dtor.
1236 * That's why we need to run uma_dbg_kskip() check here,
1237 * albeit we don't make skip check for other init/fini
1240 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1241 keg->uk_fini != trash_fini)
1243 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1245 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1246 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1248 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1249 uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1253 * Frees pages from a keg back to the system. This is done on demand from
1254 * the pageout daemon.
1259 keg_drain(uma_keg_t keg)
1261 struct slabhead freeslabs;
1263 uma_slab_t slab, tmp;
1266 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1269 for (i = 0; i < vm_ndomains; i++) {
1270 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1271 keg->uk_name, keg, i, dom->ud_free_items);
1272 dom = &keg->uk_domain[i];
1273 LIST_INIT(&freeslabs);
1276 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1277 LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
1278 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1280 n = dom->ud_free_slabs;
1281 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1282 dom->ud_free_slabs = 0;
1283 dom->ud_free_items -= n * keg->uk_ipers;
1284 dom->ud_pages -= n * keg->uk_ppera;
1287 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1288 keg_free_slab(keg, slab, keg->uk_ipers);
1293 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1297 * Set draining to interlock with zone_dtor() so we can release our
1298 * locks as we go. Only dtor() should do a WAITOK call since it
1299 * is the only call that knows the structure will still be available
1303 while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1304 if (waitok == M_NOWAIT)
1306 msleep(zone, &zone->uz_lock, PVM, "zonedrain", 1);
1308 zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1310 bucket_cache_reclaim(zone, drain);
1313 * The DRAINING flag protects us from being freed while
1314 * we're running. Normally the uma_rwlock would protect us but we
1315 * must be able to release and acquire the right lock for each keg.
1317 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1318 keg_drain(zone->uz_keg);
1320 zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1327 zone_drain(uma_zone_t zone, void *unused)
1330 zone_reclaim(zone, M_NOWAIT, true);
1334 zone_trim(uma_zone_t zone, void *unused)
1337 zone_reclaim(zone, M_NOWAIT, false);
1341 * Allocate a new slab for a keg and inserts it into the partial slab list.
1342 * The keg should be unlocked on entry. If the allocation succeeds it will
1343 * be locked on return.
1346 * flags Wait flags for the item initialization routine
1347 * aflags Wait flags for the slab allocation
1350 * The slab that was allocated or NULL if there is no memory and the
1351 * caller specified M_NOWAIT.
1354 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1365 KASSERT(domain >= 0 && domain < vm_ndomains,
1366 ("keg_alloc_slab: domain %d out of range", domain));
1368 allocf = keg->uk_allocf;
1371 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1372 uma_hash_slab_t hslab;
1373 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1377 slab = &hslab->uhs_slab;
1381 * This reproduces the old vm_zone behavior of zero filling pages the
1382 * first time they are added to a zone.
1384 * Malloced items are zeroed in uma_zalloc.
1387 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1392 if (keg->uk_flags & UMA_ZONE_NODUMP)
1395 /* zone is passed for legacy reasons. */
1396 size = keg->uk_ppera * PAGE_SIZE;
1397 mem = allocf(zone, size, domain, &sflags, aflags);
1399 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1400 zone_free_item(slabzone(keg->uk_ipers),
1401 slab_tohashslab(slab), NULL, SKIP_NONE);
1404 uma_total_inc(size);
1406 /* For HASH zones all pages go to the same uma_domain. */
1407 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1410 /* Point the slab into the allocated memory */
1411 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1412 slab = (uma_slab_t )(mem + keg->uk_pgoff);
1414 slab_tohashslab(slab)->uhs_data = mem;
1416 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1417 for (i = 0; i < keg->uk_ppera; i++)
1418 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1421 slab->us_freecount = keg->uk_ipers;
1422 slab->us_flags = sflags;
1423 slab->us_domain = domain;
1425 BIT_FILL(keg->uk_ipers, &slab->us_free);
1427 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1430 if (keg->uk_init != NULL) {
1431 for (i = 0; i < keg->uk_ipers; i++)
1432 if (keg->uk_init(slab_item(slab, keg, i),
1433 keg->uk_size, flags) != 0)
1435 if (i != keg->uk_ipers) {
1436 keg_free_slab(keg, slab, i);
1440 KEG_LOCK(keg, domain);
1442 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1443 slab, keg->uk_name, keg);
1445 if (keg->uk_flags & UMA_ZFLAG_HASH)
1446 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1449 * If we got a slab here it's safe to mark it partially used
1450 * and return. We assume that the caller is going to remove
1451 * at least one item.
1453 dom = &keg->uk_domain[domain];
1454 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1455 dom->ud_pages += keg->uk_ppera;
1456 dom->ud_free_items += keg->uk_ipers;
1465 * This function is intended to be used early on in place of page_alloc() so
1466 * that we may use the boot time page cache to satisfy allocations before
1470 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1479 pages = howmany(bytes, PAGE_SIZE);
1480 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1482 *pflag = UMA_SLAB_BOOT;
1483 m = vm_page_alloc_contig_domain(NULL, 0, domain,
1484 malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1485 (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1489 pa = VM_PAGE_TO_PHYS(m);
1490 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1491 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1492 defined(__riscv) || defined(__powerpc64__)
1493 if ((wait & M_NODUMP) == 0)
1497 /* Allocate KVA and indirectly advance bootmem. */
1498 mem = (void *)pmap_map(&bootmem, m->phys_addr,
1499 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1500 if ((wait & M_ZERO) != 0)
1501 bzero(mem, pages * PAGE_SIZE);
1507 startup_free(void *mem, vm_size_t bytes)
1512 va = (vm_offset_t)mem;
1513 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1514 pmap_remove(kernel_pmap, va, va + bytes);
1515 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1516 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1517 defined(__riscv) || defined(__powerpc64__)
1518 dump_drop_page(VM_PAGE_TO_PHYS(m));
1520 vm_page_unwire_noq(m);
1526 * Allocates a number of pages from the system
1529 * bytes The number of bytes requested
1530 * wait Shall we wait?
1533 * A pointer to the alloced memory or possibly
1534 * NULL if M_NOWAIT is set.
1537 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1540 void *p; /* Returned page */
1542 *pflag = UMA_SLAB_KERNEL;
1543 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1549 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1552 struct pglist alloctail;
1553 vm_offset_t addr, zkva;
1555 vm_page_t p, p_next;
1560 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1562 TAILQ_INIT(&alloctail);
1563 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1564 malloc2vm_flags(wait);
1565 *pflag = UMA_SLAB_KERNEL;
1566 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1567 if (CPU_ABSENT(cpu)) {
1568 p = vm_page_alloc(NULL, 0, flags);
1571 p = vm_page_alloc(NULL, 0, flags);
1573 pc = pcpu_find(cpu);
1574 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1577 p = vm_page_alloc_domain(NULL, 0,
1578 pc->pc_domain, flags);
1579 if (__predict_false(p == NULL))
1580 p = vm_page_alloc(NULL, 0, flags);
1583 if (__predict_false(p == NULL))
1585 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1587 if ((addr = kva_alloc(bytes)) == 0)
1590 TAILQ_FOREACH(p, &alloctail, listq) {
1591 pmap_qenter(zkva, &p, 1);
1594 return ((void*)addr);
1596 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1597 vm_page_unwire_noq(p);
1604 * Allocates a number of pages from within an object
1607 * bytes The number of bytes requested
1608 * wait Shall we wait?
1611 * A pointer to the alloced memory or possibly
1612 * NULL if M_NOWAIT is set.
1615 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1618 TAILQ_HEAD(, vm_page) alloctail;
1620 vm_offset_t retkva, zkva;
1621 vm_page_t p, p_next;
1624 TAILQ_INIT(&alloctail);
1627 npages = howmany(bytes, PAGE_SIZE);
1628 while (npages > 0) {
1629 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1630 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1631 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1635 * Since the page does not belong to an object, its
1638 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1643 * Page allocation failed, free intermediate pages and
1646 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1647 vm_page_unwire_noq(p);
1652 *flags = UMA_SLAB_PRIV;
1653 zkva = keg->uk_kva +
1654 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1656 TAILQ_FOREACH(p, &alloctail, listq) {
1657 pmap_qenter(zkva, &p, 1);
1661 return ((void *)retkva);
1665 * Allocate physically contiguous pages.
1668 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1672 *pflag = UMA_SLAB_KERNEL;
1673 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1674 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1678 * Frees a number of pages to the system
1681 * mem A pointer to the memory to be freed
1682 * size The size of the memory being freed
1683 * flags The original p->us_flags field
1689 page_free(void *mem, vm_size_t size, uint8_t flags)
1692 if ((flags & UMA_SLAB_BOOT) != 0) {
1693 startup_free(mem, size);
1697 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1698 ("UMA: page_free used with invalid flags %x", flags));
1700 kmem_free((vm_offset_t)mem, size);
1704 * Frees pcpu zone allocations
1707 * mem A pointer to the memory to be freed
1708 * size The size of the memory being freed
1709 * flags The original p->us_flags field
1715 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1717 vm_offset_t sva, curva;
1721 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1723 if ((flags & UMA_SLAB_BOOT) != 0) {
1724 startup_free(mem, size);
1728 sva = (vm_offset_t)mem;
1729 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1730 paddr = pmap_kextract(curva);
1731 m = PHYS_TO_VM_PAGE(paddr);
1732 vm_page_unwire_noq(m);
1735 pmap_qremove(sva, size >> PAGE_SHIFT);
1736 kva_free(sva, size);
1741 * Zero fill initializer
1743 * Arguments/Returns follow uma_init specifications
1746 zero_init(void *mem, int size, int flags)
1754 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1757 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1762 * Actual size of embedded struct slab (!OFFPAGE).
1765 slab_sizeof(int nitems)
1769 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1770 return (roundup(s, UMA_ALIGN_PTR + 1));
1774 * Size of memory for embedded slabs (!OFFPAGE).
1777 slab_space(int nitems)
1779 return (UMA_SLAB_SIZE - slab_sizeof(nitems));
1782 #define UMA_FIXPT_SHIFT 31
1783 #define UMA_FRAC_FIXPT(n, d) \
1784 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1785 #define UMA_FIXPT_PCT(f) \
1786 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1787 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
1788 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1791 * Compute the number of items that will fit in a slab. If hdr is true, the
1792 * item count may be limited to provide space in the slab for an inline slab
1793 * header. Otherwise, all slab space will be provided for item storage.
1796 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1801 /* The padding between items is not needed after the last item. */
1802 padpi = rsize - size;
1806 * Start with the maximum item count and remove items until
1807 * the slab header first alongside the allocatable memory.
1809 for (ipers = MIN(SLAB_MAX_SETSIZE,
1810 (slabsize + padpi - slab_sizeof(1)) / rsize);
1812 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1816 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1823 * Compute the number of items that will fit in a slab for a startup zone.
1826 slab_ipers(size_t size, int align)
1830 rsize = roundup(size, align + 1); /* Assume no CACHESPREAD */
1831 return (slab_ipers_hdr(size, rsize, UMA_SLAB_SIZE, true));
1834 struct keg_layout_result {
1842 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
1843 struct keg_layout_result *kl)
1848 kl->slabsize = slabsize;
1850 /* Handle INTERNAL as inline with an extra page. */
1851 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
1852 kl->format &= ~UMA_ZFLAG_INTERNAL;
1853 kl->slabsize += PAGE_SIZE;
1856 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
1857 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
1859 /* Account for memory used by an offpage slab header. */
1860 total = kl->slabsize;
1861 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
1862 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
1864 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
1868 * Determine the format of a uma keg. This determines where the slab header
1869 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
1872 * keg The zone we should initialize
1878 keg_layout(uma_keg_t keg)
1880 struct keg_layout_result kl = {}, kl_tmp;
1889 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1890 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
1891 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
1892 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
1893 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
1895 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
1896 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
1897 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
1900 alignsize = keg->uk_align + 1;
1903 * Calculate the size of each allocation (rsize) according to
1904 * alignment. If the requested size is smaller than we have
1905 * allocation bits for we round it up.
1907 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
1908 rsize = roundup2(rsize, alignsize);
1910 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
1912 * We want one item to start on every align boundary in a page.
1913 * To do this we will span pages. We will also extend the item
1914 * by the size of align if it is an even multiple of align.
1915 * Otherwise, it would fall on the same boundary every time.
1917 if ((rsize & alignsize) == 0)
1919 slabsize = rsize * (PAGE_SIZE / alignsize);
1920 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
1921 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
1922 slabsize = round_page(slabsize);
1925 * Start with a slab size of as many pages as it takes to
1926 * represent a single item. We will try to fit as many
1927 * additional items into the slab as possible.
1929 slabsize = round_page(keg->uk_size);
1932 /* Build a list of all of the available formats for this keg. */
1935 /* Evaluate an inline slab layout. */
1936 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
1939 /* TODO: vm_page-embedded slab. */
1942 * We can't do OFFPAGE if we're internal or if we've been
1943 * asked to not go to the VM for buckets. If we do this we
1944 * may end up going to the VM for slabs which we do not want
1945 * to do if we're UMA_ZONE_VM, which clearly forbids it.
1946 * In those cases, evaluate a pseudo-format called INTERNAL
1947 * which has an inline slab header and one extra page to
1948 * guarantee that it fits.
1950 * Otherwise, see if using an OFFPAGE slab will improve our
1953 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
1954 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
1956 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
1959 * Choose a slab size and format which satisfy the minimum efficiency.
1960 * Prefer the smallest slab size that meets the constraints.
1962 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
1963 * for small items (up to PAGE_SIZE), the iteration increment is one
1964 * page; and for large items, the increment is one item.
1966 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
1967 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
1968 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
1971 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
1972 round_page(rsize * (i - 1) + keg->uk_size);
1974 for (j = 0; j < nfmt; j++) {
1975 /* Only if we have no viable format yet. */
1976 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
1980 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
1981 if (kl_tmp.eff <= kl.eff)
1986 CTR6(KTR_UMA, "keg %s layout: format %#x "
1987 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
1988 keg->uk_name, kl.format, kl.ipers, rsize,
1989 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
1991 /* Stop when we reach the minimum efficiency. */
1992 if (kl.eff >= UMA_MIN_EFF)
1996 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
1997 slabsize >= SLAB_MAX_SETSIZE * rsize ||
1998 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2002 pages = atop(kl.slabsize);
2003 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2004 pages *= mp_maxid + 1;
2006 keg->uk_rsize = rsize;
2007 keg->uk_ipers = kl.ipers;
2008 keg->uk_ppera = pages;
2009 keg->uk_flags |= kl.format;
2012 * How do we find the slab header if it is offpage or if not all item
2013 * start addresses are in the same page? We could solve the latter
2014 * case with vaddr alignment, but we don't.
2016 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2017 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2018 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2019 keg->uk_flags |= UMA_ZFLAG_HASH;
2021 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2024 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2025 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2027 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2028 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2029 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2030 keg->uk_ipers, pages));
2034 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2035 * the keg onto the global keg list.
2037 * Arguments/Returns follow uma_ctor specifications
2038 * udata Actually uma_kctor_args
2041 keg_ctor(void *mem, int size, void *udata, int flags)
2043 struct uma_kctor_args *arg = udata;
2044 uma_keg_t keg = mem;
2049 keg->uk_size = arg->size;
2050 keg->uk_init = arg->uminit;
2051 keg->uk_fini = arg->fini;
2052 keg->uk_align = arg->align;
2053 keg->uk_reserve = 0;
2054 keg->uk_flags = arg->flags;
2057 * We use a global round-robin policy by default. Zones with
2058 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2059 * case the iterator is never run.
2061 keg->uk_dr.dr_policy = DOMAINSET_RR();
2062 keg->uk_dr.dr_iter = 0;
2065 * The master zone is passed to us at keg-creation time.
2068 keg->uk_name = zone->uz_name;
2070 if (arg->flags & UMA_ZONE_ZINIT)
2071 keg->uk_init = zero_init;
2073 if (arg->flags & UMA_ZONE_MALLOC)
2074 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2077 keg->uk_flags &= ~UMA_ZONE_PCPU;
2083 * Use a first-touch NUMA policy for all kegs that pmap_extract()
2084 * will work on with the exception of critical VM structures
2085 * necessary for paging.
2087 * Zones may override the default by specifying either.
2090 if ((keg->uk_flags &
2091 (UMA_ZFLAG_HASH | UMA_ZONE_VM | UMA_ZONE_ROUNDROBIN)) == 0)
2092 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2093 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2094 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2098 * If we haven't booted yet we need allocations to go through the
2099 * startup cache until the vm is ready.
2101 #ifdef UMA_MD_SMALL_ALLOC
2102 if (keg->uk_ppera == 1)
2103 keg->uk_allocf = uma_small_alloc;
2106 if (booted < BOOT_KVA)
2107 keg->uk_allocf = startup_alloc;
2108 else if (keg->uk_flags & UMA_ZONE_PCPU)
2109 keg->uk_allocf = pcpu_page_alloc;
2110 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2111 keg->uk_allocf = contig_alloc;
2113 keg->uk_allocf = page_alloc;
2114 #ifdef UMA_MD_SMALL_ALLOC
2115 if (keg->uk_ppera == 1)
2116 keg->uk_freef = uma_small_free;
2119 if (keg->uk_flags & UMA_ZONE_PCPU)
2120 keg->uk_freef = pcpu_page_free;
2122 keg->uk_freef = page_free;
2125 * Initialize keg's locks.
2127 for (i = 0; i < vm_ndomains; i++)
2128 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2131 * If we're putting the slab header in the actual page we need to
2132 * figure out where in each page it goes. See slab_sizeof
2135 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2138 shsize = slab_sizeof(keg->uk_ipers);
2139 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2141 * The only way the following is possible is if with our
2142 * UMA_ALIGN_PTR adjustments we are now bigger than
2143 * UMA_SLAB_SIZE. I haven't checked whether this is
2144 * mathematically possible for all cases, so we make
2147 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2148 ("zone %s ipers %d rsize %d size %d slab won't fit",
2149 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2152 if (keg->uk_flags & UMA_ZFLAG_HASH)
2153 hash_alloc(&keg->uk_hash, 0);
2155 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2157 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2159 rw_wlock(&uma_rwlock);
2160 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2161 rw_wunlock(&uma_rwlock);
2166 zone_kva_available(uma_zone_t zone, void *unused)
2170 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2174 if (keg->uk_allocf == startup_alloc) {
2175 /* Switch to the real allocator. */
2176 if (keg->uk_flags & UMA_ZONE_PCPU)
2177 keg->uk_allocf = pcpu_page_alloc;
2178 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2180 keg->uk_allocf = contig_alloc;
2182 keg->uk_allocf = page_alloc;
2187 zone_alloc_counters(uma_zone_t zone, void *unused)
2190 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2191 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2192 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2196 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2198 uma_zone_domain_t zdom;
2201 struct sysctl_oid *oid, *domainoid;
2202 int domains, i, cnt;
2203 static const char *nokeg = "cache zone";
2207 * Make a sysctl safe copy of the zone name by removing
2208 * any special characters and handling dups by appending
2211 if (zone->uz_namecnt != 0) {
2212 /* Count the number of decimal digits and '_' separator. */
2213 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2215 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2217 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2220 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2221 for (c = zone->uz_ctlname; *c != '\0'; c++)
2222 if (strchr("./\\ -", *c) != NULL)
2226 * Basic parameters at the root.
2228 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2229 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD, NULL, "");
2231 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2232 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2233 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2234 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2235 zone, 0, sysctl_handle_uma_zone_flags, "A",
2236 "Allocator configuration flags");
2237 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2238 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2239 "Desired per-cpu cache size");
2240 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2241 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2242 "Maximum allowed per-cpu cache size");
2247 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2248 domains = vm_ndomains;
2251 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2252 "keg", CTLFLAG_RD, NULL, "");
2254 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2255 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2256 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2257 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2258 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2259 "Real object size with alignment");
2260 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2261 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2262 "pages per-slab allocation");
2263 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2264 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2265 "items available per-slab");
2266 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2267 "align", CTLFLAG_RD, &keg->uk_align, 0,
2268 "item alignment mask");
2269 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2270 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2271 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2272 "Slab utilization (100 - internal fragmentation %)");
2273 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2274 OID_AUTO, "domain", CTLFLAG_RD, NULL, "");
2275 for (i = 0; i < domains; i++) {
2276 dom = &keg->uk_domain[i];
2277 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2278 OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD,
2280 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2281 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2282 "Total pages currently allocated from VM");
2283 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2284 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2285 "items free in the slab layer");
2288 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2289 "name", CTLFLAG_RD, nokeg, "Keg name");
2292 * Information about zone limits.
2294 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2295 "limit", CTLFLAG_RD, NULL, "");
2296 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2297 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2298 zone, 0, sysctl_handle_uma_zone_items, "QU",
2299 "current number of allocated items if limit is set");
2300 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2301 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2302 "Maximum number of cached items");
2303 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2304 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2305 "Number of threads sleeping at limit");
2306 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2307 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2308 "Total zone limit sleeps");
2309 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2310 "bucket_max", CTLFLAG_RD, &zone->uz_bkt_max, 0,
2311 "Maximum number of items in the bucket cache");
2312 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2313 "bucket_cnt", CTLFLAG_RD, &zone->uz_bkt_count, 0,
2314 "Number of items in the bucket cache");
2317 * Per-domain zone information.
2319 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2320 OID_AUTO, "domain", CTLFLAG_RD, NULL, "");
2321 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2323 for (i = 0; i < domains; i++) {
2324 zdom = &zone->uz_domain[i];
2325 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2326 OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD, NULL, "");
2327 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2328 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2329 "number of items in this domain");
2330 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2331 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2332 "maximum item count in this period");
2333 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2334 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2335 "minimum item count in this period");
2336 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2337 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2338 "Working set size");
2342 * General statistics.
2344 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2345 "stats", CTLFLAG_RD, NULL, "");
2346 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2347 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2348 zone, 1, sysctl_handle_uma_zone_cur, "I",
2349 "Current number of allocated items");
2350 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2351 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2352 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2353 "Total allocation calls");
2354 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2355 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2356 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2357 "Total free calls");
2358 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2359 "fails", CTLFLAG_RD, &zone->uz_fails,
2360 "Number of allocation failures");
2361 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2362 "xdomain", CTLFLAG_RD, &zone->uz_xdomain, 0,
2363 "Free calls from the wrong domain");
2366 struct uma_zone_count {
2372 zone_count(uma_zone_t zone, void *arg)
2374 struct uma_zone_count *cnt;
2378 * Some zones are rapidly created with identical names and
2379 * destroyed out of order. This can lead to gaps in the count.
2380 * Use one greater than the maximum observed for this name.
2382 if (strcmp(zone->uz_name, cnt->name) == 0)
2383 cnt->count = MAX(cnt->count,
2384 zone->uz_namecnt + 1);
2388 zone_update_caches(uma_zone_t zone)
2392 for (i = 0; i <= mp_maxid; i++) {
2393 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2394 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2399 * Zone header ctor. This initializes all fields, locks, etc.
2401 * Arguments/Returns follow uma_ctor specifications
2402 * udata Actually uma_zctor_args
2405 zone_ctor(void *mem, int size, void *udata, int flags)
2407 struct uma_zone_count cnt;
2408 struct uma_zctor_args *arg = udata;
2409 uma_zone_t zone = mem;
2415 zone->uz_name = arg->name;
2416 zone->uz_ctor = arg->ctor;
2417 zone->uz_dtor = arg->dtor;
2418 zone->uz_init = NULL;
2419 zone->uz_fini = NULL;
2420 zone->uz_sleeps = 0;
2421 zone->uz_xdomain = 0;
2422 zone->uz_bucket_size = 0;
2423 zone->uz_bucket_size_min = 0;
2424 zone->uz_bucket_size_max = BUCKET_MAX;
2425 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2426 zone->uz_warning = NULL;
2427 /* The domain structures follow the cpu structures. */
2429 (struct uma_zone_domain *)&zone->uz_cpu[mp_maxid + 1];
2430 zone->uz_bkt_max = ULONG_MAX;
2431 timevalclear(&zone->uz_ratecheck);
2433 /* Count the number of duplicate names. */
2434 cnt.name = arg->name;
2436 zone_foreach(zone_count, &cnt);
2437 zone->uz_namecnt = cnt.count;
2438 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
2439 ZONE_CROSS_LOCK_INIT(zone);
2441 for (i = 0; i < vm_ndomains; i++)
2442 STAILQ_INIT(&zone->uz_domain[i].uzd_buckets);
2445 if (arg->uminit == trash_init && arg->fini == trash_fini)
2446 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2450 * This is a pure cache zone, no kegs.
2453 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2454 ("zone_ctor: Import specified for non-cache zone."));
2455 zone->uz_flags = arg->flags;
2456 zone->uz_size = arg->size;
2457 zone->uz_import = arg->import;
2458 zone->uz_release = arg->release;
2459 zone->uz_arg = arg->arg;
2460 rw_wlock(&uma_rwlock);
2461 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2462 rw_wunlock(&uma_rwlock);
2467 * Use the regular zone/keg/slab allocator.
2469 zone->uz_import = zone_import;
2470 zone->uz_release = zone_release;
2471 zone->uz_arg = zone;
2474 if (arg->flags & UMA_ZONE_SECONDARY) {
2475 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2476 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2477 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2478 zone->uz_init = arg->uminit;
2479 zone->uz_fini = arg->fini;
2480 zone->uz_flags |= UMA_ZONE_SECONDARY;
2481 rw_wlock(&uma_rwlock);
2483 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2484 if (LIST_NEXT(z, uz_link) == NULL) {
2485 LIST_INSERT_AFTER(z, zone, uz_link);
2490 rw_wunlock(&uma_rwlock);
2491 } else if (keg == NULL) {
2492 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2493 arg->align, arg->flags)) == NULL)
2496 struct uma_kctor_args karg;
2499 /* We should only be here from uma_startup() */
2500 karg.size = arg->size;
2501 karg.uminit = arg->uminit;
2502 karg.fini = arg->fini;
2503 karg.align = arg->align;
2504 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2506 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2512 /* Inherit properties from the keg. */
2514 zone->uz_size = keg->uk_size;
2515 zone->uz_flags |= (keg->uk_flags &
2516 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2519 if (__predict_true(booted >= BOOT_RUNNING)) {
2520 zone_alloc_counters(zone, NULL);
2521 zone_alloc_sysctl(zone, NULL);
2523 zone->uz_allocs = EARLY_COUNTER;
2524 zone->uz_frees = EARLY_COUNTER;
2525 zone->uz_fails = EARLY_COUNTER;
2528 /* Caller requests a private SMR context. */
2529 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2530 zone->uz_smr = smr_create(zone->uz_name);
2532 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2533 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2534 ("Invalid zone flag combination"));
2535 if (arg->flags & UMA_ZFLAG_INTERNAL)
2536 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2537 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2538 zone->uz_bucket_size = BUCKET_MAX;
2539 else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2540 zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2541 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2542 zone->uz_bucket_size = 0;
2544 zone->uz_bucket_size = bucket_select(zone->uz_size);
2545 zone->uz_bucket_size_min = zone->uz_bucket_size;
2546 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2547 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2548 zone_update_caches(zone);
2554 * Keg header dtor. This frees all data, destroys locks, frees the hash
2555 * table and removes the keg from the global list.
2557 * Arguments/Returns follow uma_dtor specifications
2561 keg_dtor(void *arg, int size, void *udata)
2564 uint32_t free, pages;
2567 keg = (uma_keg_t)arg;
2569 for (i = 0; i < vm_ndomains; i++) {
2570 free += keg->uk_domain[i].ud_free_items;
2571 pages += keg->uk_domain[i].ud_pages;
2572 KEG_LOCK_FINI(keg, i);
2575 printf("Freed UMA keg (%s) was not empty (%u items). "
2576 " Lost %u pages of memory.\n",
2577 keg->uk_name ? keg->uk_name : "",
2578 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2580 hash_free(&keg->uk_hash);
2586 * Arguments/Returns follow uma_dtor specifications
2590 zone_dtor(void *arg, int size, void *udata)
2595 zone = (uma_zone_t)arg;
2597 sysctl_remove_oid(zone->uz_oid, 1, 1);
2599 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2602 rw_wlock(&uma_rwlock);
2603 LIST_REMOVE(zone, uz_link);
2604 rw_wunlock(&uma_rwlock);
2606 * XXX there are some races here where
2607 * the zone can be drained but zone lock
2608 * released and then refilled before we
2609 * remove it... we dont care for now
2611 zone_reclaim(zone, M_WAITOK, true);
2613 * We only destroy kegs from non secondary/non cache zones.
2615 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2617 rw_wlock(&uma_rwlock);
2618 LIST_REMOVE(keg, uk_link);
2619 rw_wunlock(&uma_rwlock);
2620 zone_free_item(kegs, keg, NULL, SKIP_NONE);
2622 counter_u64_free(zone->uz_allocs);
2623 counter_u64_free(zone->uz_frees);
2624 counter_u64_free(zone->uz_fails);
2625 free(zone->uz_ctlname, M_UMA);
2626 ZONE_LOCK_FINI(zone);
2627 ZONE_CROSS_LOCK_FINI(zone);
2631 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2636 LIST_FOREACH(keg, &uma_kegs, uk_link) {
2637 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2640 LIST_FOREACH(zone, &uma_cachezones, uz_link)
2645 * Traverses every zone in the system and calls a callback
2648 * zfunc A pointer to a function which accepts a zone
2655 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2658 rw_rlock(&uma_rwlock);
2659 zone_foreach_unlocked(zfunc, arg);
2660 rw_runlock(&uma_rwlock);
2664 * Initialize the kernel memory allocator. This is done after pages can be
2665 * allocated but before general KVA is available.
2668 uma_startup1(vm_offset_t virtual_avail)
2670 struct uma_zctor_args args;
2671 size_t ksize, zsize, size;
2672 uma_keg_t masterkeg;
2676 bootstart = bootmem = virtual_avail;
2678 rw_init(&uma_rwlock, "UMA lock");
2679 sx_init(&uma_reclaim_lock, "umareclaim");
2681 ksize = sizeof(struct uma_keg) +
2682 (sizeof(struct uma_domain) * vm_ndomains);
2683 ksize = roundup(ksize, UMA_SUPER_ALIGN);
2684 zsize = sizeof(struct uma_zone) +
2685 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2686 (sizeof(struct uma_zone_domain) * vm_ndomains);
2687 zsize = roundup(zsize, UMA_SUPER_ALIGN);
2689 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2690 size = (zsize * 2) + ksize;
2691 m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO);
2692 zones = (uma_zone_t)m;
2694 kegs = (uma_zone_t)m;
2696 masterkeg = (uma_keg_t)m;
2698 /* "manually" create the initial zone */
2699 memset(&args, 0, sizeof(args));
2700 args.name = "UMA Kegs";
2702 args.ctor = keg_ctor;
2703 args.dtor = keg_dtor;
2704 args.uminit = zero_init;
2706 args.keg = masterkeg;
2707 args.align = UMA_SUPER_ALIGN - 1;
2708 args.flags = UMA_ZFLAG_INTERNAL;
2709 zone_ctor(kegs, zsize, &args, M_WAITOK);
2711 args.name = "UMA Zones";
2713 args.ctor = zone_ctor;
2714 args.dtor = zone_dtor;
2715 args.uminit = zero_init;
2718 args.align = UMA_SUPER_ALIGN - 1;
2719 args.flags = UMA_ZFLAG_INTERNAL;
2720 zone_ctor(zones, zsize, &args, M_WAITOK);
2722 /* Now make zones for slab headers */
2723 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2724 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2725 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2726 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2728 hashzone = uma_zcreate("UMA Hash",
2729 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2730 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2736 #ifndef UMA_MD_SMALL_ALLOC
2737 extern void vm_radix_reserve_kva(void);
2741 * Advertise the availability of normal kva allocations and switch to
2742 * the default back-end allocator. Marks the KVA we consumed on startup
2743 * as used in the map.
2749 if (bootstart != bootmem) {
2750 vm_map_lock(kernel_map);
2751 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2752 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2753 vm_map_unlock(kernel_map);
2756 #ifndef UMA_MD_SMALL_ALLOC
2757 /* Set up radix zone to use noobj_alloc. */
2758 vm_radix_reserve_kva();
2762 zone_foreach_unlocked(zone_kva_available, NULL);
2767 * Finish our initialization steps.
2774 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2775 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2776 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2778 zone_foreach_unlocked(zone_alloc_counters, NULL);
2779 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2780 callout_init(&uma_callout, 1);
2781 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2782 booted = BOOT_RUNNING;
2784 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2785 EVENTHANDLER_PRI_FIRST);
2792 booted = BOOT_SHUTDOWN;
2796 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2797 int align, uint32_t flags)
2799 struct uma_kctor_args args;
2802 args.uminit = uminit;
2804 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2807 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2810 /* Public functions */
2813 uma_set_align(int align)
2816 if (align != UMA_ALIGN_CACHE)
2817 uma_align_cache = align;
2822 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2823 uma_init uminit, uma_fini fini, int align, uint32_t flags)
2826 struct uma_zctor_args args;
2829 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2832 /* This stuff is essential for the zone ctor */
2833 memset(&args, 0, sizeof(args));
2838 args.uminit = uminit;
2842 * Inject procedures which check for memory use after free if we are
2843 * allowed to scramble the memory while it is not allocated. This
2844 * requires that: UMA is actually able to access the memory, no init
2845 * or fini procedures, no dependency on the initial value of the
2846 * memory, and no (legitimate) use of the memory after free. Note,
2847 * the ctor and dtor do not need to be empty.
2849 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
2850 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
2851 args.uminit = trash_init;
2852 args.fini = trash_fini;
2859 sx_slock(&uma_reclaim_lock);
2860 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
2861 sx_sunlock(&uma_reclaim_lock);
2868 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
2869 uma_init zinit, uma_fini zfini, uma_zone_t master)
2871 struct uma_zctor_args args;
2875 keg = master->uz_keg;
2876 memset(&args, 0, sizeof(args));
2878 args.size = keg->uk_size;
2881 args.uminit = zinit;
2883 args.align = keg->uk_align;
2884 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
2887 sx_slock(&uma_reclaim_lock);
2888 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
2889 sx_sunlock(&uma_reclaim_lock);
2896 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
2897 uma_init zinit, uma_fini zfini, uma_import zimport,
2898 uma_release zrelease, void *arg, int flags)
2900 struct uma_zctor_args args;
2902 memset(&args, 0, sizeof(args));
2907 args.uminit = zinit;
2909 args.import = zimport;
2910 args.release = zrelease;
2913 args.flags = flags | UMA_ZFLAG_CACHE;
2915 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
2920 uma_zdestroy(uma_zone_t zone)
2924 * Large slabs are expensive to reclaim, so don't bother doing
2925 * unnecessary work if we're shutting down.
2927 if (booted == BOOT_SHUTDOWN &&
2928 zone->uz_fini == NULL && zone->uz_release == zone_release)
2930 sx_slock(&uma_reclaim_lock);
2931 zone_free_item(zones, zone, NULL, SKIP_NONE);
2932 sx_sunlock(&uma_reclaim_lock);
2936 uma_zwait(uma_zone_t zone)
2940 item = uma_zalloc_arg(zone, NULL, M_WAITOK);
2941 uma_zfree(zone, item);
2945 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
2947 void *item, *pcpu_item;
2951 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
2953 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
2956 pcpu_item = zpcpu_base_to_offset(item);
2957 if (flags & M_ZERO) {
2959 for (i = 0; i <= mp_maxid; i++)
2960 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
2962 bzero(item, zone->uz_size);
2969 * A stub while both regular and pcpu cases are identical.
2972 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
2977 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
2979 item = zpcpu_offset_to_base(pcpu_item);
2980 uma_zfree_arg(zone, item, udata);
2983 static inline void *
2984 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
2990 skipdbg = uma_dbg_zskip(zone, item);
2991 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
2992 zone->uz_ctor != trash_ctor)
2993 trash_ctor(item, size, udata, flags);
2995 /* Check flags before loading ctor pointer. */
2996 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
2997 __predict_false(zone->uz_ctor != NULL) &&
2998 zone->uz_ctor(item, size, udata, flags) != 0) {
2999 counter_u64_add(zone->uz_fails, 1);
3000 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3005 uma_dbg_alloc(zone, NULL, item);
3014 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3015 enum zfreeskip skip)
3020 skipdbg = uma_dbg_zskip(zone, item);
3021 if (skip == SKIP_NONE && !skipdbg) {
3022 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3023 uma_dbg_free(zone, udata, item);
3025 uma_dbg_free(zone, NULL, item);
3028 if (__predict_true(skip < SKIP_DTOR)) {
3029 if (zone->uz_dtor != NULL)
3030 zone->uz_dtor(item, size, udata);
3032 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3033 zone->uz_dtor != trash_dtor)
3034 trash_dtor(item, size, udata);
3039 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3040 #define UMA_ZALLOC_DEBUG
3042 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3048 if (flags & M_WAITOK) {
3049 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3050 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3055 KASSERT((flags & M_EXEC) == 0,
3056 ("uma_zalloc_debug: called with M_EXEC"));
3057 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3058 ("uma_zalloc_debug: called within spinlock or critical section"));
3059 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3060 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3063 #ifdef DEBUG_MEMGUARD
3064 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3066 item = memguard_alloc(zone->uz_size, flags);
3068 error = EJUSTRETURN;
3069 if (zone->uz_init != NULL &&
3070 zone->uz_init(item, zone->uz_size, flags) != 0) {
3074 if (zone->uz_ctor != NULL &&
3075 zone->uz_ctor(item, zone->uz_size, udata,
3077 counter_u64_add(zone->uz_fails, 1);
3078 zone->uz_fini(item, zone->uz_size);
3085 /* This is unfortunate but should not be fatal. */
3092 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3094 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3095 ("uma_zfree_debug: called with spinlock or critical section held"));
3097 #ifdef DEBUG_MEMGUARD
3098 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3099 if (zone->uz_dtor != NULL)
3100 zone->uz_dtor(item, zone->uz_size, udata);
3101 if (zone->uz_fini != NULL)
3102 zone->uz_fini(item, zone->uz_size);
3103 memguard_free(item);
3104 return (EJUSTRETURN);
3111 static __noinline void *
3112 uma_zalloc_single(uma_zone_t zone, void *udata, int flags)
3117 * We can not get a bucket so try to return a single item.
3119 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3120 domain = PCPU_GET(domain);
3122 domain = UMA_ANYDOMAIN;
3123 return (zone_alloc_item(zone, udata, domain, flags));
3128 uma_zalloc_smr(uma_zone_t zone, int flags)
3130 uma_cache_bucket_t bucket;
3135 #ifdef UMA_ZALLOC_DEBUG
3136 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3137 ("uma_zalloc_arg: called with non-SMR zone.\n"));
3138 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3144 cache = &zone->uz_cpu[curcpu];
3145 bucket = &cache->uc_allocbucket;
3146 size = cache_uz_size(cache);
3147 uz_flags = cache_uz_flags(cache);
3148 if (__predict_true(bucket->ucb_cnt != 0)) {
3149 item = cache_bucket_pop(cache, bucket);
3151 return (item_ctor(zone, uz_flags, size, NULL, flags,
3154 } while (cache_alloc(zone, cache, NULL, flags));
3157 return (uma_zalloc_single(zone, NULL, flags));
3162 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3164 uma_cache_bucket_t bucket;
3169 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3170 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3172 /* This is the fast path allocation */
3173 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3176 #ifdef UMA_ZALLOC_DEBUG
3177 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3178 ("uma_zalloc_arg: called with SMR zone.\n"));
3179 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3184 * If possible, allocate from the per-CPU cache. There are two
3185 * requirements for safe access to the per-CPU cache: (1) the thread
3186 * accessing the cache must not be preempted or yield during access,
3187 * and (2) the thread must not migrate CPUs without switching which
3188 * cache it accesses. We rely on a critical section to prevent
3189 * preemption and migration. We release the critical section in
3190 * order to acquire the zone mutex if we are unable to allocate from
3191 * the current cache; when we re-acquire the critical section, we
3192 * must detect and handle migration if it has occurred.
3196 cache = &zone->uz_cpu[curcpu];
3197 bucket = &cache->uc_allocbucket;
3198 size = cache_uz_size(cache);
3199 uz_flags = cache_uz_flags(cache);
3200 if (__predict_true(bucket->ucb_cnt != 0)) {
3201 item = cache_bucket_pop(cache, bucket);
3203 return (item_ctor(zone, uz_flags, size, udata, flags,
3206 } while (cache_alloc(zone, cache, udata, flags));
3209 return (uma_zalloc_single(zone, udata, flags));
3213 * Replenish an alloc bucket and possibly restore an old one. Called in
3214 * a critical section. Returns in a critical section.
3216 * A false return value indicates an allocation failure.
3217 * A true return value indicates success and the caller should retry.
3219 static __noinline bool
3220 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3222 uma_zone_domain_t zdom;
3223 uma_bucket_t bucket;
3227 CRITICAL_ASSERT(curthread);
3230 * If we have run out of items in our alloc bucket see
3231 * if we can switch with the free bucket.
3233 * SMR Zones can't re-use the free bucket until the sequence has
3236 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 &&
3237 cache->uc_freebucket.ucb_cnt != 0) {
3238 cache_bucket_swap(&cache->uc_freebucket,
3239 &cache->uc_allocbucket);
3244 * Discard any empty allocation bucket while we hold no locks.
3246 bucket = cache_bucket_unload_alloc(cache);
3249 bucket_free(zone, bucket, udata);
3251 /* Short-circuit for zones without buckets and low memory. */
3252 if (zone->uz_bucket_size == 0 || bucketdisable) {
3258 * Attempt to retrieve the item from the per-CPU cache has failed, so
3259 * we must go back to the zone. This requires the zone lock, so we
3260 * must drop the critical section, then re-acquire it when we go back
3261 * to the cache. Since the critical section is released, we may be
3262 * preempted or migrate. As such, make sure not to maintain any
3263 * thread-local state specific to the cache from prior to releasing
3264 * the critical section.
3267 if (ZONE_TRYLOCK(zone) == 0) {
3268 /* Record contention to size the buckets. */
3273 /* See if we lost the race to fill the cache. */
3275 cache = &zone->uz_cpu[curcpu];
3276 if (cache->uc_allocbucket.ucb_bucket != NULL) {
3282 * Check the zone's cache of buckets.
3284 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) {
3285 domain = PCPU_GET(domain);
3286 zdom = &zone->uz_domain[domain];
3288 domain = UMA_ANYDOMAIN;
3289 zdom = &zone->uz_domain[0];
3292 if ((bucket = zone_fetch_bucket(zone, zdom)) != NULL) {
3293 KASSERT(bucket->ub_cnt != 0,
3294 ("uma_zalloc_arg: Returning an empty bucket."));
3295 cache_bucket_load_alloc(cache, bucket);
3298 /* We are no longer associated with this CPU. */
3302 * We bump the uz count when the cache size is insufficient to
3303 * handle the working set.
3305 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
3306 zone->uz_bucket_size++;
3310 * Fill a bucket and attempt to use it as the alloc bucket.
3312 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3313 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3314 zone->uz_name, zone, bucket);
3315 if (bucket == NULL) {
3321 * See if we lost the race or were migrated. Cache the
3322 * initialized bucket to make this less likely or claim
3323 * the memory directly.
3327 cache = &zone->uz_cpu[curcpu];
3328 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3329 ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0 ||
3330 domain == PCPU_GET(domain))) {
3331 cache_bucket_load_alloc(cache, bucket);
3332 zdom->uzd_imax += bucket->ub_cnt;
3333 } else if (zone->uz_bkt_count >= zone->uz_bkt_max) {
3336 bucket_drain(zone, bucket);
3337 bucket_free(zone, bucket, udata);
3341 zone_put_bucket(zone, zdom, bucket, false);
3347 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3350 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3351 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3353 /* This is the fast path allocation */
3354 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3355 zone->uz_name, zone, domain, flags);
3357 if (flags & M_WAITOK) {
3358 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3359 "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3361 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3362 ("uma_zalloc_domain: called with spinlock or critical section held"));
3364 return (zone_alloc_item(zone, udata, domain, flags));
3368 * Find a slab with some space. Prefer slabs that are partially used over those
3369 * that are totally full. This helps to reduce fragmentation.
3371 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3375 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3381 KASSERT(domain >= 0 && domain < vm_ndomains,
3382 ("keg_first_slab: domain %d out of range", domain));
3383 KEG_LOCK_ASSERT(keg, domain);
3388 dom = &keg->uk_domain[domain];
3389 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3391 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3392 LIST_REMOVE(slab, us_link);
3393 dom->ud_free_slabs--;
3394 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3398 domain = (domain + 1) % vm_ndomains;
3399 } while (domain != start);
3405 * Fetch an existing slab from a free or partial list. Returns with the
3406 * keg domain lock held if a slab was found or unlocked if not.
3409 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3414 /* HASH has a single free list. */
3415 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3418 KEG_LOCK(keg, domain);
3419 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3420 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3421 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3422 KEG_UNLOCK(keg, domain);
3429 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3431 struct vm_domainset_iter di;
3438 * Use the keg's policy if upper layers haven't already specified a
3439 * domain (as happens with first-touch zones).
3441 * To avoid races we run the iterator with the keg lock held, but that
3442 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3443 * clear M_WAITOK and handle low memory conditions locally.
3445 rr = rdomain == UMA_ANYDOMAIN;
3447 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3448 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3456 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3461 * M_NOVM means don't ask at all!
3466 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3469 if (!rr && (flags & M_WAITOK) == 0)
3471 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3472 if ((flags & M_WAITOK) != 0) {
3473 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
3481 * We might not have been able to get a slab but another cpu
3482 * could have while we were unlocked. Check again before we
3485 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3492 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3498 KEG_LOCK_ASSERT(keg, slab->us_domain);
3500 dom = &keg->uk_domain[slab->us_domain];
3501 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3502 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3503 item = slab_item(slab, keg, freei);
3504 slab->us_freecount--;
3505 dom->ud_free_items--;
3508 * Move this slab to the full list. It must be on the partial list, so
3509 * we do not need to update the free slab count. In particular,
3510 * keg_fetch_slab() always returns slabs on the partial list.
3512 if (slab->us_freecount == 0) {
3513 LIST_REMOVE(slab, us_link);
3514 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3521 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3535 /* Try to keep the buckets totally full */
3536 for (i = 0; i < max; ) {
3537 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3540 stripe = howmany(max, vm_ndomains);
3542 dom = &keg->uk_domain[slab->us_domain];
3543 while (slab->us_freecount && i < max) {
3544 bucket[i++] = slab_alloc_item(keg, slab);
3545 if (dom->ud_free_items <= keg->uk_reserve)
3549 * If the zone is striped we pick a new slab for every
3550 * N allocations. Eliminating this conditional will
3551 * instead pick a new domain for each bucket rather
3552 * than stripe within each bucket. The current option
3553 * produces more fragmentation and requires more cpu
3554 * time but yields better distribution.
3556 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3557 vm_ndomains > 1 && --stripe == 0)
3561 KEG_UNLOCK(keg, slab->us_domain);
3562 /* Don't block if we allocated any successfully. */
3571 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3573 uint64_t old, new, total, max;
3576 * The hard case. We're going to sleep because there were existing
3577 * sleepers or because we ran out of items. This routine enforces
3578 * fairness by keeping fifo order.
3580 * First release our ill gotten gains and make some noise.
3583 zone_free_limit(zone, count);
3584 zone_log_warning(zone);
3585 zone_maxaction(zone);
3586 if (flags & M_NOWAIT)
3590 * We need to allocate an item or set ourself as a sleeper
3591 * while the sleepq lock is held to avoid wakeup races. This
3592 * is essentially a home rolled semaphore.
3594 sleepq_lock(&zone->uz_max_items);
3595 old = zone->uz_items;
3597 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3598 /* Cache the max since we will evaluate twice. */
3599 max = zone->uz_max_items;
3600 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3601 UZ_ITEMS_COUNT(old) >= max)
3602 new = old + UZ_ITEMS_SLEEPER;
3604 new = old + MIN(count, max - old);
3605 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3607 /* We may have successfully allocated under the sleepq lock. */
3608 if (UZ_ITEMS_SLEEPERS(new) == 0) {
3609 sleepq_release(&zone->uz_max_items);
3614 * This is in a different cacheline from uz_items so that we
3615 * don't constantly invalidate the fastpath cacheline when we
3616 * adjust item counts. This could be limited to toggling on
3619 atomic_add_32(&zone->uz_sleepers, 1);
3620 atomic_add_64(&zone->uz_sleeps, 1);
3623 * We have added ourselves as a sleeper. The sleepq lock
3624 * protects us from wakeup races. Sleep now and then retry.
3626 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3627 sleepq_wait(&zone->uz_max_items, PVM);
3630 * After wakeup, remove ourselves as a sleeper and try
3631 * again. We no longer have the sleepq lock for protection.
3633 * Subract ourselves as a sleeper while attempting to add
3636 atomic_subtract_32(&zone->uz_sleepers, 1);
3637 old = atomic_fetchadd_64(&zone->uz_items,
3638 -(UZ_ITEMS_SLEEPER - count));
3639 /* We're no longer a sleeper. */
3640 old -= UZ_ITEMS_SLEEPER;
3643 * If we're still at the limit, restart. Notably do not
3644 * block on other sleepers. Cache the max value to protect
3645 * against changes via sysctl.
3647 total = UZ_ITEMS_COUNT(old);
3648 max = zone->uz_max_items;
3651 /* Truncate if necessary, otherwise wake other sleepers. */
3652 if (total + count > max) {
3653 zone_free_limit(zone, total + count - max);
3654 count = max - total;
3655 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3656 wakeup_one(&zone->uz_max_items);
3663 * Allocate 'count' items from our max_items limit. Returns the number
3664 * available. If M_NOWAIT is not specified it will sleep until at least
3665 * one item can be allocated.
3668 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3673 max = zone->uz_max_items;
3677 * We expect normal allocations to succeed with a simple
3680 old = atomic_fetchadd_64(&zone->uz_items, count);
3681 if (__predict_true(old + count <= max))
3685 * If we had some items and no sleepers just return the
3686 * truncated value. We have to release the excess space
3687 * though because that may wake sleepers who weren't woken
3688 * because we were temporarily over the limit.
3691 zone_free_limit(zone, (old + count) - max);
3694 return (zone_alloc_limit_hard(zone, count, flags));
3698 * Free a number of items back to the limit.
3701 zone_free_limit(uma_zone_t zone, int count)
3708 * In the common case we either have no sleepers or
3709 * are still over the limit and can just return.
3711 old = atomic_fetchadd_64(&zone->uz_items, -count);
3712 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3713 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3717 * Moderate the rate of wakeups. Sleepers will continue
3718 * to generate wakeups if necessary.
3720 wakeup_one(&zone->uz_max_items);
3724 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3726 uma_bucket_t bucket;
3729 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3732 /* Avoid allocs targeting empty domains. */
3733 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3734 domain = UMA_ANYDOMAIN;
3736 if (zone->uz_max_items > 0)
3737 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3740 maxbucket = zone->uz_bucket_size;
3744 /* Don't wait for buckets, preserve caller's NOVM setting. */
3745 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3746 if (bucket == NULL) {
3751 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3752 MIN(maxbucket, bucket->ub_entries), domain, flags);
3755 * Initialize the memory if necessary.
3757 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3760 for (i = 0; i < bucket->ub_cnt; i++)
3761 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3765 * If we couldn't initialize the whole bucket, put the
3766 * rest back onto the freelist.
3768 if (i != bucket->ub_cnt) {
3769 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3770 bucket->ub_cnt - i);
3772 bzero(&bucket->ub_bucket[i],
3773 sizeof(void *) * (bucket->ub_cnt - i));
3779 cnt = bucket->ub_cnt;
3780 if (bucket->ub_cnt == 0) {
3781 bucket_free(zone, bucket, udata);
3782 counter_u64_add(zone->uz_fails, 1);
3786 if (zone->uz_max_items > 0 && cnt < maxbucket)
3787 zone_free_limit(zone, maxbucket - cnt);
3793 * Allocates a single item from a zone.
3796 * zone The zone to alloc for.
3797 * udata The data to be passed to the constructor.
3798 * domain The domain to allocate from or UMA_ANYDOMAIN.
3799 * flags M_WAITOK, M_NOWAIT, M_ZERO.
3802 * NULL if there is no memory and M_NOWAIT is set
3803 * An item if successful
3807 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
3811 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0)
3814 /* Avoid allocs targeting empty domains. */
3815 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3816 domain = UMA_ANYDOMAIN;
3818 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
3822 * We have to call both the zone's init (not the keg's init)
3823 * and the zone's ctor. This is because the item is going from
3824 * a keg slab directly to the user, and the user is expecting it
3825 * to be both zone-init'd as well as zone-ctor'd.
3827 if (zone->uz_init != NULL) {
3828 if (zone->uz_init(item, zone->uz_size, flags) != 0) {
3829 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
3833 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
3838 counter_u64_add(zone->uz_allocs, 1);
3839 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
3840 zone->uz_name, zone);
3845 counter_u64_add(zone->uz_fails, 1);
3847 if (zone->uz_max_items > 0)
3848 zone_free_limit(zone, 1);
3849 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
3850 zone->uz_name, zone);
3857 uma_zfree_smr(uma_zone_t zone, void *item)
3860 uma_cache_bucket_t bucket;
3861 int domain, itemdomain, uz_flags;
3863 #ifdef UMA_ZALLOC_DEBUG
3864 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3865 ("uma_zfree_smr: called with non-SMR zone.\n"));
3866 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
3867 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
3870 cache = &zone->uz_cpu[curcpu];
3871 uz_flags = cache_uz_flags(cache);
3872 domain = itemdomain = 0;
3874 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
3875 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
3879 cache = &zone->uz_cpu[curcpu];
3880 /* SMR Zones must free to the free bucket. */
3881 bucket = &cache->uc_freebucket;
3883 domain = PCPU_GET(domain);
3884 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
3885 domain != itemdomain) {
3886 bucket = &cache->uc_crossbucket;
3889 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
3890 cache_bucket_push(cache, bucket, item);
3894 } while (cache_free(zone, cache, NULL, item, itemdomain));
3898 * If nothing else caught this, we'll just do an internal free.
3900 zone_free_item(zone, item, NULL, SKIP_NONE);
3905 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
3908 uma_cache_bucket_t bucket;
3909 int domain, itemdomain, uz_flags;
3911 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3912 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3914 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
3916 #ifdef UMA_ZALLOC_DEBUG
3917 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3918 ("uma_zfree_arg: called with SMR zone.\n"));
3919 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
3922 /* uma_zfree(..., NULL) does nothing, to match free(9). */
3927 * We are accessing the per-cpu cache without a critical section to
3928 * fetch size and flags. This is acceptable, if we are preempted we
3929 * will simply read another cpu's line.
3931 cache = &zone->uz_cpu[curcpu];
3932 uz_flags = cache_uz_flags(cache);
3933 if (UMA_ALWAYS_CTORDTOR ||
3934 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
3935 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
3938 * The race here is acceptable. If we miss it we'll just have to wait
3939 * a little longer for the limits to be reset.
3941 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
3942 if (zone->uz_sleepers > 0)
3947 * If possible, free to the per-CPU cache. There are two
3948 * requirements for safe access to the per-CPU cache: (1) the thread
3949 * accessing the cache must not be preempted or yield during access,
3950 * and (2) the thread must not migrate CPUs without switching which
3951 * cache it accesses. We rely on a critical section to prevent
3952 * preemption and migration. We release the critical section in
3953 * order to acquire the zone mutex if we are unable to free to the
3954 * current cache; when we re-acquire the critical section, we must
3955 * detect and handle migration if it has occurred.
3957 domain = itemdomain = 0;
3959 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
3960 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
3964 cache = &zone->uz_cpu[curcpu];
3966 * Try to free into the allocbucket first to give LIFO
3967 * ordering for cache-hot datastructures. Spill over
3968 * into the freebucket if necessary. Alloc will swap
3969 * them if one runs dry.
3971 bucket = &cache->uc_allocbucket;
3973 domain = PCPU_GET(domain);
3974 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
3975 domain != itemdomain) {
3976 bucket = &cache->uc_crossbucket;
3979 if (bucket->ucb_cnt >= bucket->ucb_entries)
3980 bucket = &cache->uc_freebucket;
3981 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
3982 cache_bucket_push(cache, bucket, item);
3986 } while (cache_free(zone, cache, udata, item, itemdomain));
3990 * If nothing else caught this, we'll just do an internal free.
3993 zone_free_item(zone, item, udata, SKIP_DTOR);
3998 * sort crossdomain free buckets to domain correct buckets and cache
4002 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4004 struct uma_bucketlist fullbuckets;
4005 uma_zone_domain_t zdom;
4011 "uma_zfree: zone %s(%p) draining cross bucket %p",
4012 zone->uz_name, zone, bucket);
4014 STAILQ_INIT(&fullbuckets);
4017 * To avoid having ndomain * ndomain buckets for sorting we have a
4018 * lock on the current crossfree bucket. A full matrix with
4019 * per-domain locking could be used if necessary.
4021 ZONE_CROSS_LOCK(zone);
4022 while (bucket->ub_cnt > 0) {
4023 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4024 domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4025 zdom = &zone->uz_domain[domain];
4026 if (zdom->uzd_cross == NULL) {
4027 zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
4028 if (zdom->uzd_cross == NULL)
4031 zdom->uzd_cross->ub_bucket[zdom->uzd_cross->ub_cnt++] = item;
4032 if (zdom->uzd_cross->ub_cnt == zdom->uzd_cross->ub_entries) {
4033 STAILQ_INSERT_HEAD(&fullbuckets, zdom->uzd_cross,
4035 zdom->uzd_cross = NULL;
4039 ZONE_CROSS_UNLOCK(zone);
4040 if (!STAILQ_EMPTY(&fullbuckets)) {
4042 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4043 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4044 bucket->ub_seq = smr_current(zone->uz_smr);
4045 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4046 if (zone->uz_bkt_count >= zone->uz_bkt_max) {
4048 bucket_drain(zone, b);
4049 bucket_free(zone, b, udata);
4052 domain = _vm_phys_domain(
4054 (vm_offset_t)b->ub_bucket[0]));
4055 zdom = &zone->uz_domain[domain];
4056 zone_put_bucket(zone, zdom, b, true);
4061 if (bucket->ub_cnt != 0)
4062 bucket_drain(zone, bucket);
4063 bucket->ub_seq = SMR_SEQ_INVALID;
4064 bucket_free(zone, bucket, udata);
4069 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4070 int domain, int itemdomain)
4072 uma_zone_domain_t zdom;
4076 * Buckets coming from the wrong domain will be entirely for the
4077 * only other domain on two domain systems. In this case we can
4078 * simply cache them. Otherwise we need to sort them back to
4081 if (domain != itemdomain && vm_ndomains > 2) {
4082 zone_free_cross(zone, bucket, udata);
4088 * Attempt to save the bucket in the zone's domain bucket cache.
4090 * We bump the uz count when the cache size is insufficient to
4091 * handle the working set.
4093 if (ZONE_TRYLOCK(zone) == 0) {
4094 /* Record contention to size the buckets. */
4096 if (zone->uz_bucket_size < zone->uz_bucket_size_max)
4097 zone->uz_bucket_size++;
4101 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4102 zone->uz_name, zone, bucket);
4103 /* ub_cnt is pointing to the last free item */
4104 KASSERT(bucket->ub_cnt == bucket->ub_entries,
4105 ("uma_zfree: Attempting to insert partial bucket onto the full list.\n"));
4106 if (zone->uz_bkt_count >= zone->uz_bkt_max) {
4108 bucket_drain(zone, bucket);
4109 bucket_free(zone, bucket, udata);
4111 zdom = &zone->uz_domain[itemdomain];
4112 zone_put_bucket(zone, zdom, bucket, true);
4118 * Populate a free or cross bucket for the current cpu cache. Free any
4119 * existing full bucket either to the zone cache or back to the slab layer.
4121 * Enters and returns in a critical section. false return indicates that
4122 * we can not satisfy this free in the cache layer. true indicates that
4123 * the caller should retry.
4125 static __noinline bool
4126 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4129 uma_cache_bucket_t cbucket;
4130 uma_bucket_t newbucket, bucket;
4133 CRITICAL_ASSERT(curthread);
4135 if (zone->uz_bucket_size == 0)
4138 cache = &zone->uz_cpu[curcpu];
4142 * FIRSTTOUCH domains need to free to the correct zdom. When
4143 * enabled this is the zdom of the item. The bucket is the
4144 * cross bucket if the current domain and itemdomain do not match.
4146 cbucket = &cache->uc_freebucket;
4148 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) {
4149 domain = PCPU_GET(domain);
4150 if (domain != itemdomain) {
4151 cbucket = &cache->uc_crossbucket;
4152 if (cbucket->ucb_cnt != 0)
4153 atomic_add_64(&zone->uz_xdomain,
4158 itemdomain = domain = 0;
4159 bucket = cache_bucket_unload(cbucket);
4161 /* We are no longer associated with this CPU. */
4165 * Don't let SMR zones operate without a free bucket. Force
4166 * a synchronize and re-use this one. We will only degrade
4167 * to a synchronize every bucket_size items rather than every
4168 * item if we fail to allocate a bucket.
4170 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4172 bucket->ub_seq = smr_advance(zone->uz_smr);
4173 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4174 if (newbucket == NULL && bucket != NULL) {
4175 bucket_drain(zone, bucket);
4179 } else if (!bucketdisable)
4180 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4183 zone_free_bucket(zone, bucket, udata, domain, itemdomain);
4186 if ((bucket = newbucket) == NULL)
4188 cache = &zone->uz_cpu[curcpu];
4191 * Check to see if we should be populating the cross bucket. If it
4192 * is already populated we will fall through and attempt to populate
4195 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) {
4196 domain = PCPU_GET(domain);
4197 if (domain != itemdomain &&
4198 cache->uc_crossbucket.ucb_bucket == NULL) {
4199 cache_bucket_load_cross(cache, bucket);
4205 * We may have lost the race to fill the bucket or switched CPUs.
4207 if (cache->uc_freebucket.ucb_bucket != NULL) {
4209 bucket_free(zone, bucket, udata);
4212 cache_bucket_load_free(cache, bucket);
4218 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
4221 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4222 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4224 CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone);
4226 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
4227 ("uma_zfree_domain: called with spinlock or critical section held"));
4229 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4232 zone_free_item(zone, item, udata, SKIP_NONE);
4236 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4243 KEG_LOCK_ASSERT(keg, slab->us_domain);
4245 /* Do we need to remove from any lists? */
4246 dom = &keg->uk_domain[slab->us_domain];
4247 if (slab->us_freecount + 1 == keg->uk_ipers) {
4248 LIST_REMOVE(slab, us_link);
4249 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4250 dom->ud_free_slabs++;
4251 } else if (slab->us_freecount == 0) {
4252 LIST_REMOVE(slab, us_link);
4253 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4256 /* Slab management. */
4257 freei = slab_item_index(slab, keg, item);
4258 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4259 slab->us_freecount++;
4261 /* Keg statistics. */
4262 dom->ud_free_items++;
4266 zone_release(void *arg, void **bucket, int cnt)
4279 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4280 lock = KEG_LOCK(keg, 0);
4281 for (i = 0; i < cnt; i++) {
4283 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4284 slab = vtoslab((vm_offset_t)item);
4286 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4287 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4288 slab = hash_sfind(&keg->uk_hash, mem);
4290 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4292 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4295 lock = KEG_LOCK(keg, slab->us_domain);
4297 slab_free_item(zone, slab, item);
4304 * Frees a single item to any zone.
4307 * zone The zone to free to
4308 * item The item we're freeing
4309 * udata User supplied data for the dtor
4310 * skip Skip dtors and finis
4313 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4317 * If a free is sent directly to an SMR zone we have to
4318 * synchronize immediately because the item can instantly
4319 * be reallocated. This should only happen in degenerate
4320 * cases when no memory is available for per-cpu caches.
4322 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4323 smr_synchronize(zone->uz_smr);
4325 item_dtor(zone, item, zone->uz_size, udata, skip);
4327 if (skip < SKIP_FINI && zone->uz_fini)
4328 zone->uz_fini(item, zone->uz_size);
4330 zone->uz_release(zone->uz_arg, &item, 1);
4332 if (skip & SKIP_CNT)
4335 counter_u64_add(zone->uz_frees, 1);
4337 if (zone->uz_max_items > 0)
4338 zone_free_limit(zone, 1);
4343 uma_zone_set_max(uma_zone_t zone, int nitems)
4345 struct uma_bucket_zone *ubz;
4349 * XXX This can misbehave if the zone has any allocations with
4350 * no limit and a limit is imposed. There is currently no
4351 * way to clear a limit.
4354 ubz = bucket_zone_max(zone, nitems);
4355 count = ubz != NULL ? ubz->ubz_entries : 0;
4356 zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4357 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4358 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4359 zone->uz_max_items = nitems;
4360 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4361 zone_update_caches(zone);
4362 /* We may need to wake waiters. */
4363 wakeup(&zone->uz_max_items);
4371 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4373 struct uma_bucket_zone *ubz;
4377 ubz = bucket_zone_max(zone, nitems);
4380 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4381 /* Count the cross-domain bucket. */
4383 nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4384 zone->uz_bucket_size_max = ubz->ubz_entries;
4386 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4388 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4389 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4390 zone->uz_bkt_max = nitems;
4396 uma_zone_get_max(uma_zone_t zone)
4400 nitems = atomic_load_64(&zone->uz_max_items);
4407 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4410 ZONE_ASSERT_COLD(zone);
4411 zone->uz_warning = warning;
4416 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4419 ZONE_ASSERT_COLD(zone);
4420 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4425 uma_zone_get_cur(uma_zone_t zone)
4431 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4432 nitems = counter_u64_fetch(zone->uz_allocs) -
4433 counter_u64_fetch(zone->uz_frees);
4435 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4436 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4438 return (nitems < 0 ? 0 : nitems);
4442 uma_zone_get_allocs(uma_zone_t zone)
4448 if (zone->uz_allocs != EARLY_COUNTER)
4449 nitems = counter_u64_fetch(zone->uz_allocs);
4451 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4457 uma_zone_get_frees(uma_zone_t zone)
4463 if (zone->uz_frees != EARLY_COUNTER)
4464 nitems = counter_u64_fetch(zone->uz_frees);
4466 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4472 /* Used only for KEG_ASSERT_COLD(). */
4474 uma_keg_get_allocs(uma_keg_t keg)
4480 LIST_FOREACH(z, &keg->uk_zones, uz_link)
4481 nitems += uma_zone_get_allocs(z);
4489 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4494 KEG_ASSERT_COLD(keg);
4495 keg->uk_init = uminit;
4500 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4505 KEG_ASSERT_COLD(keg);
4506 keg->uk_fini = fini;
4511 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4514 ZONE_ASSERT_COLD(zone);
4515 zone->uz_init = zinit;
4520 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4523 ZONE_ASSERT_COLD(zone);
4524 zone->uz_fini = zfini;
4529 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4534 KEG_ASSERT_COLD(keg);
4535 keg->uk_freef = freef;
4540 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4545 KEG_ASSERT_COLD(keg);
4546 keg->uk_allocf = allocf;
4551 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4554 ZONE_ASSERT_COLD(zone);
4556 zone->uz_flags |= UMA_ZONE_SMR;
4558 zone_update_caches(zone);
4562 uma_zone_get_smr(uma_zone_t zone)
4565 return (zone->uz_smr);
4570 uma_zone_reserve(uma_zone_t zone, int items)
4575 KEG_ASSERT_COLD(keg);
4576 keg->uk_reserve = items;
4581 uma_zone_reserve_kva(uma_zone_t zone, int count)
4588 KEG_ASSERT_COLD(keg);
4589 ZONE_ASSERT_COLD(zone);
4591 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4593 #ifdef UMA_MD_SMALL_ALLOC
4594 if (keg->uk_ppera > 1) {
4598 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4605 MPASS(keg->uk_kva == 0);
4608 zone->uz_max_items = pages * keg->uk_ipers;
4609 #ifdef UMA_MD_SMALL_ALLOC
4610 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4612 keg->uk_allocf = noobj_alloc;
4614 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4615 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4616 zone_update_caches(zone);
4624 uma_prealloc(uma_zone_t zone, int items)
4626 struct vm_domainset_iter di;
4630 int aflags, domain, slabs;
4633 slabs = howmany(items, keg->uk_ipers);
4634 while (slabs-- > 0) {
4636 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4639 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4642 dom = &keg->uk_domain[slab->us_domain];
4644 * keg_alloc_slab() always returns a slab on the
4647 LIST_REMOVE(slab, us_link);
4648 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4650 dom->ud_free_slabs++;
4651 KEG_UNLOCK(keg, slab->us_domain);
4654 if (vm_domainset_iter_policy(&di, &domain) != 0)
4655 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
4662 uma_reclaim(int req)
4665 CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4666 sx_xlock(&uma_reclaim_lock);
4670 case UMA_RECLAIM_TRIM:
4671 zone_foreach(zone_trim, NULL);
4673 case UMA_RECLAIM_DRAIN:
4674 case UMA_RECLAIM_DRAIN_CPU:
4675 zone_foreach(zone_drain, NULL);
4676 if (req == UMA_RECLAIM_DRAIN_CPU) {
4677 pcpu_cache_drain_safe(NULL);
4678 zone_foreach(zone_drain, NULL);
4682 panic("unhandled reclamation request %d", req);
4686 * Some slabs may have been freed but this zone will be visited early
4687 * we visit again so that we can free pages that are empty once other
4688 * zones are drained. We have to do the same for buckets.
4690 zone_drain(slabzones[0], NULL);
4691 zone_drain(slabzones[1], NULL);
4692 bucket_zone_drain();
4693 sx_xunlock(&uma_reclaim_lock);
4696 static volatile int uma_reclaim_needed;
4699 uma_reclaim_wakeup(void)
4702 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4703 wakeup(uma_reclaim);
4707 uma_reclaim_worker(void *arg __unused)
4711 sx_xlock(&uma_reclaim_lock);
4712 while (atomic_load_int(&uma_reclaim_needed) == 0)
4713 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4715 sx_xunlock(&uma_reclaim_lock);
4716 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4717 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4718 atomic_store_int(&uma_reclaim_needed, 0);
4719 /* Don't fire more than once per-second. */
4720 pause("umarclslp", hz);
4726 uma_zone_reclaim(uma_zone_t zone, int req)
4730 case UMA_RECLAIM_TRIM:
4731 zone_trim(zone, NULL);
4733 case UMA_RECLAIM_DRAIN:
4734 zone_drain(zone, NULL);
4736 case UMA_RECLAIM_DRAIN_CPU:
4737 pcpu_cache_drain_safe(zone);
4738 zone_drain(zone, NULL);
4741 panic("unhandled reclamation request %d", req);
4747 uma_zone_exhausted(uma_zone_t zone)
4750 return (atomic_load_32(&zone->uz_sleepers) > 0);
4757 return (uma_kmem_limit);
4761 uma_set_limit(unsigned long limit)
4764 uma_kmem_limit = limit;
4771 return (atomic_load_long(&uma_kmem_total));
4778 return (uma_kmem_limit - uma_size());
4783 * Generate statistics across both the zone and its per-cpu cache's. Return
4784 * desired statistics if the pointer is non-NULL for that statistic.
4786 * Note: does not update the zone statistics, as it can't safely clear the
4787 * per-CPU cache statistic.
4791 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4792 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4795 uint64_t allocs, frees, sleeps, xdomain;
4798 allocs = frees = sleeps = xdomain = 0;
4801 cache = &z->uz_cpu[cpu];
4802 cachefree += cache->uc_allocbucket.ucb_cnt;
4803 cachefree += cache->uc_freebucket.ucb_cnt;
4804 xdomain += cache->uc_crossbucket.ucb_cnt;
4805 cachefree += cache->uc_crossbucket.ucb_cnt;
4806 allocs += cache->uc_allocs;
4807 frees += cache->uc_frees;
4809 allocs += counter_u64_fetch(z->uz_allocs);
4810 frees += counter_u64_fetch(z->uz_frees);
4811 sleeps += z->uz_sleeps;
4812 xdomain += z->uz_xdomain;
4813 if (cachefreep != NULL)
4814 *cachefreep = cachefree;
4815 if (allocsp != NULL)
4819 if (sleepsp != NULL)
4821 if (xdomainp != NULL)
4822 *xdomainp = xdomain;
4827 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
4834 rw_rlock(&uma_rwlock);
4835 LIST_FOREACH(kz, &uma_kegs, uk_link) {
4836 LIST_FOREACH(z, &kz->uk_zones, uz_link)
4839 LIST_FOREACH(z, &uma_cachezones, uz_link)
4842 rw_runlock(&uma_rwlock);
4843 return (sysctl_handle_int(oidp, &count, 0, req));
4847 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
4848 struct uma_percpu_stat *ups, bool internal)
4850 uma_zone_domain_t zdom;
4855 for (i = 0; i < vm_ndomains; i++) {
4856 zdom = &z->uz_domain[i];
4857 uth->uth_zone_free += zdom->uzd_nitems;
4859 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
4860 uth->uth_frees = counter_u64_fetch(z->uz_frees);
4861 uth->uth_fails = counter_u64_fetch(z->uz_fails);
4862 uth->uth_sleeps = z->uz_sleeps;
4863 uth->uth_xdomain = z->uz_xdomain;
4866 * While it is not normally safe to access the cache bucket pointers
4867 * while not on the CPU that owns the cache, we only allow the pointers
4868 * to be exchanged without the zone lock held, not invalidated, so
4869 * accept the possible race associated with bucket exchange during
4870 * monitoring. Use atomic_load_ptr() to ensure that the bucket pointers
4871 * are loaded only once.
4873 for (i = 0; i < mp_maxid + 1; i++) {
4874 bzero(&ups[i], sizeof(*ups));
4875 if (internal || CPU_ABSENT(i))
4877 cache = &z->uz_cpu[i];
4878 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
4879 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
4880 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
4881 ups[i].ups_allocs = cache->uc_allocs;
4882 ups[i].ups_frees = cache->uc_frees;
4887 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
4889 struct uma_stream_header ush;
4890 struct uma_type_header uth;
4891 struct uma_percpu_stat *ups;
4896 uint32_t kfree, pages;
4897 int count, error, i;
4899 error = sysctl_wire_old_buffer(req, 0);
4902 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
4903 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
4904 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
4907 rw_rlock(&uma_rwlock);
4908 LIST_FOREACH(kz, &uma_kegs, uk_link) {
4909 LIST_FOREACH(z, &kz->uk_zones, uz_link)
4913 LIST_FOREACH(z, &uma_cachezones, uz_link)
4917 * Insert stream header.
4919 bzero(&ush, sizeof(ush));
4920 ush.ush_version = UMA_STREAM_VERSION;
4921 ush.ush_maxcpus = (mp_maxid + 1);
4922 ush.ush_count = count;
4923 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
4925 LIST_FOREACH(kz, &uma_kegs, uk_link) {
4927 for (i = 0; i < vm_ndomains; i++) {
4928 kfree += kz->uk_domain[i].ud_free_items;
4929 pages += kz->uk_domain[i].ud_pages;
4931 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
4932 bzero(&uth, sizeof(uth));
4934 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
4935 uth.uth_align = kz->uk_align;
4936 uth.uth_size = kz->uk_size;
4937 uth.uth_rsize = kz->uk_rsize;
4938 if (z->uz_max_items > 0) {
4939 items = UZ_ITEMS_COUNT(z->uz_items);
4940 uth.uth_pages = (items / kz->uk_ipers) *
4943 uth.uth_pages = pages;
4944 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
4946 uth.uth_limit = z->uz_max_items;
4947 uth.uth_keg_free = kfree;
4950 * A zone is secondary is it is not the first entry
4951 * on the keg's zone list.
4953 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
4954 (LIST_FIRST(&kz->uk_zones) != z))
4955 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
4956 uma_vm_zone_stats(&uth, z, &sbuf, ups,
4957 kz->uk_flags & UMA_ZFLAG_INTERNAL);
4959 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
4960 for (i = 0; i < mp_maxid + 1; i++)
4961 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
4964 LIST_FOREACH(z, &uma_cachezones, uz_link) {
4965 bzero(&uth, sizeof(uth));
4967 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
4968 uth.uth_size = z->uz_size;
4969 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
4971 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
4972 for (i = 0; i < mp_maxid + 1; i++)
4973 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
4976 rw_runlock(&uma_rwlock);
4977 error = sbuf_finish(&sbuf);
4984 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
4986 uma_zone_t zone = *(uma_zone_t *)arg1;
4989 max = uma_zone_get_max(zone);
4990 error = sysctl_handle_int(oidp, &max, 0, req);
4991 if (error || !req->newptr)
4994 uma_zone_set_max(zone, max);
5000 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5006 * Some callers want to add sysctls for global zones that
5007 * may not yet exist so they pass a pointer to a pointer.
5010 zone = *(uma_zone_t *)arg1;
5013 cur = uma_zone_get_cur(zone);
5014 return (sysctl_handle_int(oidp, &cur, 0, req));
5018 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5020 uma_zone_t zone = arg1;
5023 cur = uma_zone_get_allocs(zone);
5024 return (sysctl_handle_64(oidp, &cur, 0, req));
5028 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5030 uma_zone_t zone = arg1;
5033 cur = uma_zone_get_frees(zone);
5034 return (sysctl_handle_64(oidp, &cur, 0, req));
5038 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5041 uma_zone_t zone = arg1;
5044 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5045 if (zone->uz_flags != 0)
5046 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5048 sbuf_printf(&sbuf, "0");
5049 error = sbuf_finish(&sbuf);
5056 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5058 uma_keg_t keg = arg1;
5059 int avail, effpct, total;
5061 total = keg->uk_ppera * PAGE_SIZE;
5062 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5063 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5065 * We consider the client's requested size and alignment here, not the
5066 * real size determination uk_rsize, because we also adjust the real
5067 * size for internal implementation reasons (max bitset size).
5069 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5070 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5071 avail *= mp_maxid + 1;
5072 effpct = 100 * avail / total;
5073 return (sysctl_handle_int(oidp, &effpct, 0, req));
5077 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5079 uma_zone_t zone = arg1;
5082 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5083 return (sysctl_handle_64(oidp, &cur, 0, req));
5088 uma_dbg_getslab(uma_zone_t zone, void *item)
5095 * It is safe to return the slab here even though the
5096 * zone is unlocked because the item's allocation state
5097 * essentially holds a reference.
5099 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5100 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5102 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5103 return (vtoslab((vm_offset_t)mem));
5105 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5106 return ((uma_slab_t)(mem + keg->uk_pgoff));
5108 slab = hash_sfind(&keg->uk_hash, mem);
5115 uma_dbg_zskip(uma_zone_t zone, void *mem)
5118 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5121 return (uma_dbg_kskip(zone->uz_keg, mem));
5125 uma_dbg_kskip(uma_keg_t keg, void *mem)
5129 if (dbg_divisor == 0)
5132 if (dbg_divisor == 1)
5135 idx = (uintptr_t)mem >> PAGE_SHIFT;
5136 if (keg->uk_ipers > 1) {
5137 idx *= keg->uk_ipers;
5138 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5141 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5142 counter_u64_add(uma_skip_cnt, 1);
5145 counter_u64_add(uma_dbg_cnt, 1);
5151 * Set up the slab's freei data such that uma_dbg_free can function.
5155 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5161 slab = uma_dbg_getslab(zone, item);
5163 panic("uma: item %p did not belong to zone %s\n",
5164 item, zone->uz_name);
5167 freei = slab_item_index(slab, keg, item);
5169 if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5170 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
5171 item, zone, zone->uz_name, slab, freei);
5172 BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5176 * Verifies freed addresses. Checks for alignment, valid slab membership
5177 * and duplicate frees.
5181 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5187 slab = uma_dbg_getslab(zone, item);
5189 panic("uma: Freed item %p did not belong to zone %s\n",
5190 item, zone->uz_name);
5193 freei = slab_item_index(slab, keg, item);
5195 if (freei >= keg->uk_ipers)
5196 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
5197 item, zone, zone->uz_name, slab, freei);
5199 if (slab_item(slab, keg, freei) != item)
5200 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
5201 item, zone, zone->uz_name, slab, freei);
5203 if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5204 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
5205 item, zone, zone->uz_name, slab, freei);
5207 BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5209 #endif /* INVARIANTS */
5213 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5214 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5219 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5220 *allocs = counter_u64_fetch(z->uz_allocs);
5221 frees = counter_u64_fetch(z->uz_frees);
5222 *sleeps = z->uz_sleeps;
5226 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5228 for (i = 0; i < vm_ndomains; i++) {
5229 *cachefree += z->uz_domain[i].uzd_nitems;
5230 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5231 (LIST_FIRST(&kz->uk_zones) != z)))
5232 *cachefree += kz->uk_domain[i].ud_free_items;
5234 *used = *allocs - frees;
5235 return (((int64_t)*used + *cachefree) * kz->uk_size);
5238 DB_SHOW_COMMAND(uma, db_show_uma)
5240 const char *fmt_hdr, *fmt_entry;
5243 uint64_t allocs, used, sleeps, xdomain;
5245 /* variables for sorting */
5247 uma_zone_t cur_zone, last_zone;
5248 int64_t cur_size, last_size, size;
5251 /* /i option produces machine-parseable CSV output */
5252 if (modif[0] == 'i') {
5253 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5254 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5256 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5257 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5260 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5261 "Sleeps", "Bucket", "Total Mem", "XFree");
5263 /* Sort the zones with largest size first. */
5265 last_size = INT64_MAX;
5270 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5271 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5273 * In the case of size ties, print out zones
5274 * in the order they are encountered. That is,
5275 * when we encounter the most recently output
5276 * zone, we have already printed all preceding
5277 * ties, and we must print all following ties.
5279 if (z == last_zone) {
5283 size = get_uma_stats(kz, z, &allocs, &used,
5284 &sleeps, &cachefree, &xdomain);
5285 if (size > cur_size && size < last_size + ties)
5293 if (cur_zone == NULL)
5296 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5297 &sleeps, &cachefree, &xdomain);
5298 db_printf(fmt_entry, cur_zone->uz_name,
5299 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5300 (uintmax_t)allocs, (uintmax_t)sleeps,
5301 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5306 last_zone = cur_zone;
5307 last_size = cur_size;
5311 DB_SHOW_COMMAND(umacache, db_show_umacache)
5314 uint64_t allocs, frees;
5318 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5319 "Requests", "Bucket");
5320 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5321 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5322 for (i = 0; i < vm_ndomains; i++)
5323 cachefree += z->uz_domain[i].uzd_nitems;
5324 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5325 z->uz_name, (uintmax_t)z->uz_size,
5326 (intmax_t)(allocs - frees), cachefree,
5327 (uintmax_t)allocs, z->uz_bucket_size);