xref: /freebsd/sys/vm/uma_core.c (revision c998f2d39d81550474170e121387aabe6fa7ac83)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
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
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice unmodified, this list of conditions, and the following
14  *    disclaimer.
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.
18  *
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.
29  */
30 
31 /*
32  * uma_core.c  Implementation of the Universal Memory allocator
33  *
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.
40  *
41  * The basic ideas stem from similar slab/zone based allocators whose algorithms
42  * are well known.
43  *
44  */
45 
46 /*
47  * TODO:
48  *	- Improve memory usage for large allocations
49  *	- Investigate cache size adjustments
50  */
51 
52 #include <sys/cdefs.h>
53 __FBSDID("$FreeBSD$");
54 
55 #include "opt_ddb.h"
56 #include "opt_param.h"
57 #include "opt_vm.h"
58 
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>
69 #include <sys/ktr.h>
70 #include <sys/lock.h>
71 #include <sys/sysctl.h>
72 #include <sys/mutex.h>
73 #include <sys/proc.h>
74 #include <sys/random.h>
75 #include <sys/rwlock.h>
76 #include <sys/sbuf.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smp.h>
80 #include <sys/smr.h>
81 #include <sys/taskqueue.h>
82 #include <sys/vmmeter.h>
83 
84 #include <vm/vm.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>
95 #include <vm/uma.h>
96 #include <vm/uma_int.h>
97 #include <vm/uma_dbg.h>
98 
99 #include <ddb/ddb.h>
100 
101 #ifdef DEBUG_MEMGUARD
102 #include <vm/memguard.h>
103 #endif
104 
105 #include <machine/md_var.h>
106 
107 #ifdef INVARIANTS
108 #define	UMA_ALWAYS_CTORDTOR	1
109 #else
110 #define	UMA_ALWAYS_CTORDTOR	0
111 #endif
112 
113 /*
114  * This is the zone and keg from which all zones are spawned.
115  */
116 static uma_zone_t kegs;
117 static uma_zone_t zones;
118 
119 /*
120  * These are the two zones from which all offpage uma_slab_ts are allocated.
121  *
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
125  * efficient.
126  */
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];
134 
135 /*
136  * The initial hash tables come out of this zone so they can be allocated
137  * prior to malloc coming up.
138  */
139 static uma_zone_t hashzone;
140 
141 /* The boot-time adjusted value for cache line alignment. */
142 int uma_align_cache = 64 - 1;
143 
144 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
145 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
146 
147 /*
148  * Are we allowed to allocate buckets?
149  */
150 static int bucketdisable = 1;
151 
152 /* Linked list of all kegs in the system */
153 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
154 
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);
158 
159 /* This RW lock protects the keg list */
160 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
161 
162 /*
163  * First available virual address for boot time allocations.
164  */
165 static vm_offset_t bootstart;
166 static vm_offset_t bootmem;
167 
168 static struct sx uma_reclaim_lock;
169 
170 /*
171  * kmem soft limit, initialized by uma_set_limit().  Ensure that early
172  * allocations don't trigger a wakeup of the reclaim thread.
173  */
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");
180 
181 /* Is the VM done starting up? */
182 static enum {
183 	BOOT_COLD,
184 	BOOT_KVA,
185 	BOOT_RUNNING,
186 	BOOT_SHUTDOWN,
187 } booted = BOOT_COLD;
188 
189 /*
190  * This is the handle used to schedule events that need to happen
191  * outside of the allocation fast path.
192  */
193 static struct callout uma_callout;
194 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
195 
196 /*
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.
199  */
200 struct uma_zctor_args {
201 	const char *name;
202 	size_t size;
203 	uma_ctor ctor;
204 	uma_dtor dtor;
205 	uma_init uminit;
206 	uma_fini fini;
207 	uma_import import;
208 	uma_release release;
209 	void *arg;
210 	uma_keg_t keg;
211 	int align;
212 	uint32_t flags;
213 };
214 
215 struct uma_kctor_args {
216 	uma_zone_t zone;
217 	size_t size;
218 	uma_init uminit;
219 	uma_fini fini;
220 	int align;
221 	uint32_t flags;
222 };
223 
224 struct uma_bucket_zone {
225 	uma_zone_t	ubz_zone;
226 	char		*ubz_name;
227 	int		ubz_entries;	/* Number of items it can hold. */
228 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
229 };
230 
231 /*
232  * Compute the actual number of bucket entries to pack them in power
233  * of two sizes for more efficient space utilization.
234  */
235 #define	BUCKET_SIZE(n)						\
236     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
237 
238 #define	BUCKET_MAX	BUCKET_SIZE(256)
239 #define	BUCKET_MIN	BUCKET_SIZE(4)
240 
241 struct uma_bucket_zone bucket_zones[] = {
242 #ifndef __ILP32__
243 	{ NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
244 #endif
245 	{ NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
246 	{ NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
247 	{ NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
248 	{ NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
249 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
250 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
251 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
252 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
253 	{ NULL, NULL, 0}
254 };
255 
256 /*
257  * Flags and enumerations to be passed to internal functions.
258  */
259 enum zfreeskip {
260 	SKIP_NONE =	0,
261 	SKIP_CNT =	0x00000001,
262 	SKIP_DTOR =	0x00010000,
263 	SKIP_FINI =	0x00020000,
264 };
265 
266 /* Prototypes.. */
267 
268 void	uma_startup1(vm_offset_t);
269 void	uma_startup2(void);
270 
271 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
272 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
273 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
274 static void *startup_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);
315 
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);
323 
324 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
325 
326 #ifdef INVARIANTS
327 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
328 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
329 
330 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
331 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
332 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
333 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
334 
335 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD, 0,
336     "Memory allocation debugging");
337 
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");
342 
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");
349 #endif
350 
351 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
352 
353 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW, 0, "Universal Memory Allocator");
354 
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");
357 
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");
360 
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");
364 
365 /*
366  * Select the slab zone for an offpage slab with the given maximum item count.
367  */
368 static inline uma_zone_t
369 slabzone(int ipers)
370 {
371 
372 	return (slabzones[ipers > SLABZONE0_SETSIZE]);
373 }
374 
375 /*
376  * This routine checks to see whether or not it's safe to enable buckets.
377  */
378 static void
379 bucket_enable(void)
380 {
381 
382 	KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
383 	bucketdisable = vm_page_count_min();
384 }
385 
386 /*
387  * Initialize bucket_zones, the array of zones of buckets of various sizes.
388  *
389  * For each zone, calculate the memory required for each bucket, consisting
390  * of the header and an array of pointers.
391  */
392 static void
393 bucket_init(void)
394 {
395 	struct uma_bucket_zone *ubz;
396 	int size;
397 
398 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
399 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
400 		size += sizeof(void *) * ubz->ubz_entries;
401 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
402 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
403 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
404 		    UMA_ZONE_FIRSTTOUCH);
405 	}
406 }
407 
408 /*
409  * Given a desired number of entries for a bucket, return the zone from which
410  * to allocate the bucket.
411  */
412 static struct uma_bucket_zone *
413 bucket_zone_lookup(int entries)
414 {
415 	struct uma_bucket_zone *ubz;
416 
417 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
418 		if (ubz->ubz_entries >= entries)
419 			return (ubz);
420 	ubz--;
421 	return (ubz);
422 }
423 
424 static struct uma_bucket_zone *
425 bucket_zone_max(uma_zone_t zone, int nitems)
426 {
427 	struct uma_bucket_zone *ubz;
428 	int bpcpu;
429 
430 	bpcpu = 2;
431 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
432 		/* Count the cross-domain bucket. */
433 		bpcpu++;
434 
435 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
436 		if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems)
437 			break;
438 	if (ubz == &bucket_zones[0])
439 		ubz = NULL;
440 	else
441 		ubz--;
442 	return (ubz);
443 }
444 
445 static int
446 bucket_select(int size)
447 {
448 	struct uma_bucket_zone *ubz;
449 
450 	ubz = &bucket_zones[0];
451 	if (size > ubz->ubz_maxsize)
452 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
453 
454 	for (; ubz->ubz_entries != 0; ubz++)
455 		if (ubz->ubz_maxsize < size)
456 			break;
457 	ubz--;
458 	return (ubz->ubz_entries);
459 }
460 
461 static uma_bucket_t
462 bucket_alloc(uma_zone_t zone, void *udata, int flags)
463 {
464 	struct uma_bucket_zone *ubz;
465 	uma_bucket_t bucket;
466 
467 	/*
468 	 * Don't allocate buckets early in boot.
469 	 */
470 	if (__predict_false(booted < BOOT_KVA))
471 		return (NULL);
472 
473 	/*
474 	 * To limit bucket recursion we store the original zone flags
475 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
476 	 * NOVM flag to persist even through deep recursions.  We also
477 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
478 	 * a bucket for a bucket zone so we do not allow infinite bucket
479 	 * recursion.  This cookie will even persist to frees of unused
480 	 * buckets via the allocation path or bucket allocations in the
481 	 * free path.
482 	 */
483 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
484 		udata = (void *)(uintptr_t)zone->uz_flags;
485 	else {
486 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
487 			return (NULL);
488 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
489 	}
490 	if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
491 		flags |= M_NOVM;
492 	ubz = bucket_zone_lookup(zone->uz_bucket_size);
493 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
494 		ubz++;
495 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
496 	if (bucket) {
497 #ifdef INVARIANTS
498 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
499 #endif
500 		bucket->ub_cnt = 0;
501 		bucket->ub_entries = ubz->ubz_entries;
502 		bucket->ub_seq = SMR_SEQ_INVALID;
503 		CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
504 		    zone->uz_name, zone, bucket);
505 	}
506 
507 	return (bucket);
508 }
509 
510 static void
511 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
512 {
513 	struct uma_bucket_zone *ubz;
514 
515 	KASSERT(bucket->ub_cnt == 0,
516 	    ("bucket_free: Freeing a non free bucket."));
517 	KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
518 	    ("bucket_free: Freeing an SMR bucket."));
519 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
520 		udata = (void *)(uintptr_t)zone->uz_flags;
521 	ubz = bucket_zone_lookup(bucket->ub_entries);
522 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
523 }
524 
525 static void
526 bucket_zone_drain(void)
527 {
528 	struct uma_bucket_zone *ubz;
529 
530 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
531 		uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN);
532 }
533 
534 /*
535  * Attempt to satisfy an allocation by retrieving a full bucket from one of the
536  * zone's caches.  If a bucket is found the zone is not locked on return.
537  */
538 static uma_bucket_t
539 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom)
540 {
541 	uma_bucket_t bucket;
542 	int i;
543 	bool dtor = false;
544 
545 	ZONE_LOCK_ASSERT(zone);
546 
547 	if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
548 		return (NULL);
549 
550 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
551 	    bucket->ub_seq != SMR_SEQ_INVALID) {
552 		if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
553 			return (NULL);
554 		bucket->ub_seq = SMR_SEQ_INVALID;
555 		dtor = (zone->uz_dtor != NULL) | UMA_ALWAYS_CTORDTOR;
556 	}
557 	MPASS(zdom->uzd_nitems >= bucket->ub_cnt);
558 	STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
559 	zdom->uzd_nitems -= bucket->ub_cnt;
560 	if (zdom->uzd_imin > zdom->uzd_nitems)
561 		zdom->uzd_imin = zdom->uzd_nitems;
562 	zone->uz_bkt_count -= bucket->ub_cnt;
563 	ZONE_UNLOCK(zone);
564 	if (dtor)
565 		for (i = 0; i < bucket->ub_cnt; i++)
566 			item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
567 			    NULL, SKIP_NONE);
568 
569 	return (bucket);
570 }
571 
572 /*
573  * Insert a full bucket into the specified cache.  The "ws" parameter indicates
574  * whether the bucket's contents should be counted as part of the zone's working
575  * set.
576  */
577 static void
578 zone_put_bucket(uma_zone_t zone, uma_zone_domain_t zdom, uma_bucket_t bucket,
579     const bool ws)
580 {
581 
582 	ZONE_LOCK_ASSERT(zone);
583 	KASSERT(!ws || zone->uz_bkt_count < zone->uz_bkt_max,
584 	    ("%s: zone %p overflow", __func__, zone));
585 
586 	STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
587 	zdom->uzd_nitems += bucket->ub_cnt;
588 	if (ws && zdom->uzd_imax < zdom->uzd_nitems)
589 		zdom->uzd_imax = zdom->uzd_nitems;
590 	zone->uz_bkt_count += bucket->ub_cnt;
591 }
592 
593 /* Pops an item out of a per-cpu cache bucket. */
594 static inline void *
595 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
596 {
597 	void *item;
598 
599 	CRITICAL_ASSERT(curthread);
600 
601 	bucket->ucb_cnt--;
602 	item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
603 #ifdef INVARIANTS
604 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
605 	KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
606 #endif
607 	cache->uc_allocs++;
608 
609 	return (item);
610 }
611 
612 /* Pushes an item into a per-cpu cache bucket. */
613 static inline void
614 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
615 {
616 
617 	CRITICAL_ASSERT(curthread);
618 	KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
619 	    ("uma_zfree: Freeing to non free bucket index."));
620 
621 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
622 	bucket->ucb_cnt++;
623 	cache->uc_frees++;
624 }
625 
626 /*
627  * Unload a UMA bucket from a per-cpu cache.
628  */
629 static inline uma_bucket_t
630 cache_bucket_unload(uma_cache_bucket_t bucket)
631 {
632 	uma_bucket_t b;
633 
634 	b = bucket->ucb_bucket;
635 	if (b != NULL) {
636 		MPASS(b->ub_entries == bucket->ucb_entries);
637 		b->ub_cnt = bucket->ucb_cnt;
638 		bucket->ucb_bucket = NULL;
639 		bucket->ucb_entries = bucket->ucb_cnt = 0;
640 	}
641 
642 	return (b);
643 }
644 
645 static inline uma_bucket_t
646 cache_bucket_unload_alloc(uma_cache_t cache)
647 {
648 
649 	return (cache_bucket_unload(&cache->uc_allocbucket));
650 }
651 
652 static inline uma_bucket_t
653 cache_bucket_unload_free(uma_cache_t cache)
654 {
655 
656 	return (cache_bucket_unload(&cache->uc_freebucket));
657 }
658 
659 static inline uma_bucket_t
660 cache_bucket_unload_cross(uma_cache_t cache)
661 {
662 
663 	return (cache_bucket_unload(&cache->uc_crossbucket));
664 }
665 
666 /*
667  * Load a bucket into a per-cpu cache bucket.
668  */
669 static inline void
670 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
671 {
672 
673 	CRITICAL_ASSERT(curthread);
674 	MPASS(bucket->ucb_bucket == NULL);
675 
676 	bucket->ucb_bucket = b;
677 	bucket->ucb_cnt = b->ub_cnt;
678 	bucket->ucb_entries = b->ub_entries;
679 }
680 
681 static inline void
682 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
683 {
684 
685 	cache_bucket_load(&cache->uc_allocbucket, b);
686 }
687 
688 static inline void
689 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
690 {
691 
692 	cache_bucket_load(&cache->uc_freebucket, b);
693 }
694 
695 #ifdef NUMA
696 static inline void
697 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
698 {
699 
700 	cache_bucket_load(&cache->uc_crossbucket, b);
701 }
702 #endif
703 
704 /*
705  * Copy and preserve ucb_spare.
706  */
707 static inline void
708 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
709 {
710 
711 	b1->ucb_bucket = b2->ucb_bucket;
712 	b1->ucb_entries = b2->ucb_entries;
713 	b1->ucb_cnt = b2->ucb_cnt;
714 }
715 
716 /*
717  * Swap two cache buckets.
718  */
719 static inline void
720 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
721 {
722 	struct uma_cache_bucket b3;
723 
724 	CRITICAL_ASSERT(curthread);
725 
726 	cache_bucket_copy(&b3, b1);
727 	cache_bucket_copy(b1, b2);
728 	cache_bucket_copy(b2, &b3);
729 }
730 
731 static void
732 zone_log_warning(uma_zone_t zone)
733 {
734 	static const struct timeval warninterval = { 300, 0 };
735 
736 	if (!zone_warnings || zone->uz_warning == NULL)
737 		return;
738 
739 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
740 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
741 }
742 
743 static inline void
744 zone_maxaction(uma_zone_t zone)
745 {
746 
747 	if (zone->uz_maxaction.ta_func != NULL)
748 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
749 }
750 
751 /*
752  * Routine called by timeout which is used to fire off some time interval
753  * based calculations.  (stats, hash size, etc.)
754  *
755  * Arguments:
756  *	arg   Unused
757  *
758  * Returns:
759  *	Nothing
760  */
761 static void
762 uma_timeout(void *unused)
763 {
764 	bucket_enable();
765 	zone_foreach(zone_timeout, NULL);
766 
767 	/* Reschedule this event */
768 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
769 }
770 
771 /*
772  * Update the working set size estimate for the zone's bucket cache.
773  * The constants chosen here are somewhat arbitrary.  With an update period of
774  * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
775  * last 100s.
776  */
777 static void
778 zone_domain_update_wss(uma_zone_domain_t zdom)
779 {
780 	long wss;
781 
782 	MPASS(zdom->uzd_imax >= zdom->uzd_imin);
783 	wss = zdom->uzd_imax - zdom->uzd_imin;
784 	zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
785 	zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
786 }
787 
788 /*
789  * Routine to perform timeout driven calculations.  This expands the
790  * hashes and does per cpu statistics aggregation.
791  *
792  *  Returns nothing.
793  */
794 static void
795 zone_timeout(uma_zone_t zone, void *unused)
796 {
797 	uma_keg_t keg;
798 	u_int slabs, pages;
799 
800 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
801 		goto update_wss;
802 
803 	keg = zone->uz_keg;
804 
805 	/*
806 	 * Hash zones are non-numa by definition so the first domain
807 	 * is the only one present.
808 	 */
809 	KEG_LOCK(keg, 0);
810 	pages = keg->uk_domain[0].ud_pages;
811 
812 	/*
813 	 * Expand the keg hash table.
814 	 *
815 	 * This is done if the number of slabs is larger than the hash size.
816 	 * What I'm trying to do here is completely reduce collisions.  This
817 	 * may be a little aggressive.  Should I allow for two collisions max?
818 	 */
819 	if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
820 		struct uma_hash newhash;
821 		struct uma_hash oldhash;
822 		int ret;
823 
824 		/*
825 		 * This is so involved because allocating and freeing
826 		 * while the keg lock is held will lead to deadlock.
827 		 * I have to do everything in stages and check for
828 		 * races.
829 		 */
830 		KEG_UNLOCK(keg, 0);
831 		ret = hash_alloc(&newhash, 1 << fls(slabs));
832 		KEG_LOCK(keg, 0);
833 		if (ret) {
834 			if (hash_expand(&keg->uk_hash, &newhash)) {
835 				oldhash = keg->uk_hash;
836 				keg->uk_hash = newhash;
837 			} else
838 				oldhash = newhash;
839 
840 			KEG_UNLOCK(keg, 0);
841 			hash_free(&oldhash);
842 			goto update_wss;
843 		}
844 	}
845 	KEG_UNLOCK(keg, 0);
846 
847 update_wss:
848 	ZONE_LOCK(zone);
849 	for (int i = 0; i < vm_ndomains; i++)
850 		zone_domain_update_wss(&zone->uz_domain[i]);
851 	ZONE_UNLOCK(zone);
852 }
853 
854 /*
855  * Allocate and zero fill the next sized hash table from the appropriate
856  * backing store.
857  *
858  * Arguments:
859  *	hash  A new hash structure with the old hash size in uh_hashsize
860  *
861  * Returns:
862  *	1 on success and 0 on failure.
863  */
864 static int
865 hash_alloc(struct uma_hash *hash, u_int size)
866 {
867 	size_t alloc;
868 
869 	KASSERT(powerof2(size), ("hash size must be power of 2"));
870 	if (size > UMA_HASH_SIZE_INIT)  {
871 		hash->uh_hashsize = size;
872 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
873 		hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
874 	} else {
875 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
876 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
877 		    UMA_ANYDOMAIN, M_WAITOK);
878 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
879 	}
880 	if (hash->uh_slab_hash) {
881 		bzero(hash->uh_slab_hash, alloc);
882 		hash->uh_hashmask = hash->uh_hashsize - 1;
883 		return (1);
884 	}
885 
886 	return (0);
887 }
888 
889 /*
890  * Expands the hash table for HASH zones.  This is done from zone_timeout
891  * to reduce collisions.  This must not be done in the regular allocation
892  * path, otherwise, we can recurse on the vm while allocating pages.
893  *
894  * Arguments:
895  *	oldhash  The hash you want to expand
896  *	newhash  The hash structure for the new table
897  *
898  * Returns:
899  *	Nothing
900  *
901  * Discussion:
902  */
903 static int
904 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
905 {
906 	uma_hash_slab_t slab;
907 	u_int hval;
908 	u_int idx;
909 
910 	if (!newhash->uh_slab_hash)
911 		return (0);
912 
913 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
914 		return (0);
915 
916 	/*
917 	 * I need to investigate hash algorithms for resizing without a
918 	 * full rehash.
919 	 */
920 
921 	for (idx = 0; idx < oldhash->uh_hashsize; idx++)
922 		while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
923 			slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
924 			LIST_REMOVE(slab, uhs_hlink);
925 			hval = UMA_HASH(newhash, slab->uhs_data);
926 			LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
927 			    slab, uhs_hlink);
928 		}
929 
930 	return (1);
931 }
932 
933 /*
934  * Free the hash bucket to the appropriate backing store.
935  *
936  * Arguments:
937  *	slab_hash  The hash bucket we're freeing
938  *	hashsize   The number of entries in that hash bucket
939  *
940  * Returns:
941  *	Nothing
942  */
943 static void
944 hash_free(struct uma_hash *hash)
945 {
946 	if (hash->uh_slab_hash == NULL)
947 		return;
948 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
949 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
950 	else
951 		free(hash->uh_slab_hash, M_UMAHASH);
952 }
953 
954 /*
955  * Frees all outstanding items in a bucket
956  *
957  * Arguments:
958  *	zone   The zone to free to, must be unlocked.
959  *	bucket The free/alloc bucket with items.
960  *
961  * Returns:
962  *	Nothing
963  */
964 
965 static void
966 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
967 {
968 	int i;
969 
970 	if (bucket == NULL || bucket->ub_cnt == 0)
971 		return;
972 
973 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
974 	    bucket->ub_seq != SMR_SEQ_INVALID) {
975 		smr_wait(zone->uz_smr, bucket->ub_seq);
976 		for (i = 0; i < bucket->ub_cnt; i++)
977 			item_dtor(zone, bucket->ub_bucket[i],
978 			    zone->uz_size, NULL, SKIP_NONE);
979 		bucket->ub_seq = SMR_SEQ_INVALID;
980 	}
981 	if (zone->uz_fini)
982 		for (i = 0; i < bucket->ub_cnt; i++)
983 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
984 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
985 	if (zone->uz_max_items > 0)
986 		zone_free_limit(zone, bucket->ub_cnt);
987 #ifdef INVARIANTS
988 	bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
989 #endif
990 	bucket->ub_cnt = 0;
991 }
992 
993 /*
994  * Drains the per cpu caches for a zone.
995  *
996  * NOTE: This may only be called while the zone is being torn down, and not
997  * during normal operation.  This is necessary in order that we do not have
998  * to migrate CPUs to drain the per-CPU caches.
999  *
1000  * Arguments:
1001  *	zone     The zone to drain, must be unlocked.
1002  *
1003  * Returns:
1004  *	Nothing
1005  */
1006 static void
1007 cache_drain(uma_zone_t zone)
1008 {
1009 	uma_cache_t cache;
1010 	uma_bucket_t bucket;
1011 	int cpu;
1012 
1013 	/*
1014 	 * XXX: It is safe to not lock the per-CPU caches, because we're
1015 	 * tearing down the zone anyway.  I.e., there will be no further use
1016 	 * of the caches at this point.
1017 	 *
1018 	 * XXX: It would good to be able to assert that the zone is being
1019 	 * torn down to prevent improper use of cache_drain().
1020 	 */
1021 	CPU_FOREACH(cpu) {
1022 		cache = &zone->uz_cpu[cpu];
1023 		bucket = cache_bucket_unload_alloc(cache);
1024 		if (bucket != NULL) {
1025 			bucket_drain(zone, bucket);
1026 			bucket_free(zone, bucket, NULL);
1027 		}
1028 		bucket = cache_bucket_unload_free(cache);
1029 		if (bucket != NULL) {
1030 			bucket_drain(zone, bucket);
1031 			bucket_free(zone, bucket, NULL);
1032 		}
1033 		bucket = cache_bucket_unload_cross(cache);
1034 		if (bucket != NULL) {
1035 			bucket_drain(zone, bucket);
1036 			bucket_free(zone, bucket, NULL);
1037 		}
1038 	}
1039 	bucket_cache_reclaim(zone, true);
1040 }
1041 
1042 static void
1043 cache_shrink(uma_zone_t zone, void *unused)
1044 {
1045 
1046 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1047 		return;
1048 
1049 	ZONE_LOCK(zone);
1050 	zone->uz_bucket_size =
1051 	    (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1052 	ZONE_UNLOCK(zone);
1053 }
1054 
1055 static void
1056 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1057 {
1058 	uma_cache_t cache;
1059 	uma_bucket_t b1, b2, b3;
1060 	int domain;
1061 
1062 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1063 		return;
1064 
1065 	b1 = b2 = b3 = NULL;
1066 	ZONE_LOCK(zone);
1067 	critical_enter();
1068 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
1069 		domain = PCPU_GET(domain);
1070 	else
1071 		domain = 0;
1072 	cache = &zone->uz_cpu[curcpu];
1073 	b1 = cache_bucket_unload_alloc(cache);
1074 	if (b1 != NULL && b1->ub_cnt != 0) {
1075 		zone_put_bucket(zone, &zone->uz_domain[domain], b1, false);
1076 		b1 = NULL;
1077 	}
1078 
1079 	/*
1080 	 * Don't flush SMR zone buckets.  This leaves the zone without a
1081 	 * bucket and forces every free to synchronize().
1082 	 */
1083 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1084 		goto out;
1085 	b2 = cache_bucket_unload_free(cache);
1086 	if (b2 != NULL && b2->ub_cnt != 0) {
1087 		zone_put_bucket(zone, &zone->uz_domain[domain], b2, false);
1088 		b2 = NULL;
1089 	}
1090 	b3 = cache_bucket_unload_cross(cache);
1091 
1092 out:
1093 	critical_exit();
1094 	ZONE_UNLOCK(zone);
1095 	if (b1)
1096 		bucket_free(zone, b1, NULL);
1097 	if (b2)
1098 		bucket_free(zone, b2, NULL);
1099 	if (b3) {
1100 		bucket_drain(zone, b3);
1101 		bucket_free(zone, b3, NULL);
1102 	}
1103 }
1104 
1105 /*
1106  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1107  * This is an expensive call because it needs to bind to all CPUs
1108  * one by one and enter a critical section on each of them in order
1109  * to safely access their cache buckets.
1110  * Zone lock must not be held on call this function.
1111  */
1112 static void
1113 pcpu_cache_drain_safe(uma_zone_t zone)
1114 {
1115 	int cpu;
1116 
1117 	/*
1118 	 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1119 	 */
1120 	if (zone)
1121 		cache_shrink(zone, NULL);
1122 	else
1123 		zone_foreach(cache_shrink, NULL);
1124 
1125 	CPU_FOREACH(cpu) {
1126 		thread_lock(curthread);
1127 		sched_bind(curthread, cpu);
1128 		thread_unlock(curthread);
1129 
1130 		if (zone)
1131 			cache_drain_safe_cpu(zone, NULL);
1132 		else
1133 			zone_foreach(cache_drain_safe_cpu, NULL);
1134 	}
1135 	thread_lock(curthread);
1136 	sched_unbind(curthread);
1137 	thread_unlock(curthread);
1138 }
1139 
1140 /*
1141  * Reclaim cached buckets from a zone.  All buckets are reclaimed if the caller
1142  * requested a drain, otherwise the per-domain caches are trimmed to either
1143  * estimated working set size.
1144  */
1145 static void
1146 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1147 {
1148 	uma_zone_domain_t zdom;
1149 	uma_bucket_t bucket;
1150 	long target, tofree;
1151 	int i;
1152 
1153 	for (i = 0; i < vm_ndomains; i++) {
1154 		/*
1155 		 * The cross bucket is partially filled and not part of
1156 		 * the item count.  Reclaim it individually here.
1157 		 */
1158 		zdom = &zone->uz_domain[i];
1159 		ZONE_CROSS_LOCK(zone);
1160 		bucket = zdom->uzd_cross;
1161 		zdom->uzd_cross = NULL;
1162 		ZONE_CROSS_UNLOCK(zone);
1163 		if (bucket != NULL) {
1164 			bucket_drain(zone, bucket);
1165 			bucket_free(zone, bucket, NULL);
1166 		}
1167 
1168 		/*
1169 		 * Shrink the zone bucket size to ensure that the per-CPU caches
1170 		 * don't grow too large.
1171 		 */
1172 		ZONE_LOCK(zone);
1173 		if (i == 0 && zone->uz_bucket_size > zone->uz_bucket_size_min)
1174 			zone->uz_bucket_size--;
1175 
1176 		/*
1177 		 * If we were asked to drain the zone, we are done only once
1178 		 * this bucket cache is empty.  Otherwise, we reclaim items in
1179 		 * excess of the zone's estimated working set size.  If the
1180 		 * difference nitems - imin is larger than the WSS estimate,
1181 		 * then the estimate will grow at the end of this interval and
1182 		 * we ignore the historical average.
1183 		 */
1184 		target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1185 		    zdom->uzd_imin);
1186 		while (zdom->uzd_nitems > target) {
1187 			bucket = STAILQ_FIRST(&zdom->uzd_buckets);
1188 			if (bucket == NULL)
1189 				break;
1190 			tofree = bucket->ub_cnt;
1191 			STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
1192 			zdom->uzd_nitems -= tofree;
1193 
1194 			/*
1195 			 * Shift the bounds of the current WSS interval to avoid
1196 			 * perturbing the estimate.
1197 			 */
1198 			zdom->uzd_imax -= lmin(zdom->uzd_imax, tofree);
1199 			zdom->uzd_imin -= lmin(zdom->uzd_imin, tofree);
1200 
1201 			ZONE_UNLOCK(zone);
1202 			bucket_drain(zone, bucket);
1203 			bucket_free(zone, bucket, NULL);
1204 			ZONE_LOCK(zone);
1205 		}
1206 		ZONE_UNLOCK(zone);
1207 	}
1208 }
1209 
1210 static void
1211 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1212 {
1213 	uint8_t *mem;
1214 	int i;
1215 	uint8_t flags;
1216 
1217 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1218 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1219 
1220 	mem = slab_data(slab, keg);
1221 	flags = slab->us_flags;
1222 	i = start;
1223 	if (keg->uk_fini != NULL) {
1224 		for (i--; i > -1; i--)
1225 #ifdef INVARIANTS
1226 		/*
1227 		 * trash_fini implies that dtor was trash_dtor. trash_fini
1228 		 * would check that memory hasn't been modified since free,
1229 		 * which executed trash_dtor.
1230 		 * That's why we need to run uma_dbg_kskip() check here,
1231 		 * albeit we don't make skip check for other init/fini
1232 		 * invocations.
1233 		 */
1234 		if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1235 		    keg->uk_fini != trash_fini)
1236 #endif
1237 			keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1238 	}
1239 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1240 		zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1241 		    NULL, SKIP_NONE);
1242 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1243 	uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1244 }
1245 
1246 /*
1247  * Frees pages from a keg back to the system.  This is done on demand from
1248  * the pageout daemon.
1249  *
1250  * Returns nothing.
1251  */
1252 static void
1253 keg_drain(uma_keg_t keg)
1254 {
1255 	struct slabhead freeslabs = { 0 };
1256 	uma_domain_t dom;
1257 	uma_slab_t slab, tmp;
1258 	int i, n;
1259 
1260 	/*
1261 	 * We don't want to take pages from statically allocated kegs at this
1262 	 * time
1263 	 */
1264 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1265 		return;
1266 
1267 	for (i = 0; i < vm_ndomains; i++) {
1268 		CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1269 		    keg->uk_name, keg, i, dom->ud_free);
1270 		n = 0;
1271 		dom = &keg->uk_domain[i];
1272 		KEG_LOCK(keg, i);
1273 		LIST_FOREACH_SAFE(slab, &dom->ud_free_slab, us_link, tmp) {
1274 			if (keg->uk_flags & UMA_ZFLAG_HASH)
1275 				UMA_HASH_REMOVE(&keg->uk_hash, slab);
1276 			n++;
1277 			LIST_REMOVE(slab, us_link);
1278 			LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1279 		}
1280 		dom->ud_pages -= n * keg->uk_ppera;
1281 		dom->ud_free -= n * keg->uk_ipers;
1282 		KEG_UNLOCK(keg, i);
1283 	}
1284 
1285 	while ((slab = LIST_FIRST(&freeslabs)) != NULL) {
1286 		LIST_REMOVE(slab, us_link);
1287 		keg_free_slab(keg, slab, keg->uk_ipers);
1288 	}
1289 }
1290 
1291 static void
1292 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1293 {
1294 
1295 	/*
1296 	 * Set draining to interlock with zone_dtor() so we can release our
1297 	 * locks as we go.  Only dtor() should do a WAITOK call since it
1298 	 * is the only call that knows the structure will still be available
1299 	 * when it wakes up.
1300 	 */
1301 	ZONE_LOCK(zone);
1302 	while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1303 		if (waitok == M_NOWAIT)
1304 			goto out;
1305 		msleep(zone, &zone->uz_lock, PVM, "zonedrain", 1);
1306 	}
1307 	zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1308 	ZONE_UNLOCK(zone);
1309 	bucket_cache_reclaim(zone, drain);
1310 
1311 	/*
1312 	 * The DRAINING flag protects us from being freed while
1313 	 * we're running.  Normally the uma_rwlock would protect us but we
1314 	 * must be able to release and acquire the right lock for each keg.
1315 	 */
1316 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1317 		keg_drain(zone->uz_keg);
1318 	ZONE_LOCK(zone);
1319 	zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1320 	wakeup(zone);
1321 out:
1322 	ZONE_UNLOCK(zone);
1323 }
1324 
1325 static void
1326 zone_drain(uma_zone_t zone, void *unused)
1327 {
1328 
1329 	zone_reclaim(zone, M_NOWAIT, true);
1330 }
1331 
1332 static void
1333 zone_trim(uma_zone_t zone, void *unused)
1334 {
1335 
1336 	zone_reclaim(zone, M_NOWAIT, false);
1337 }
1338 
1339 /*
1340  * Allocate a new slab for a keg and inserts it into the partial slab list.
1341  * The keg should be unlocked on entry.  If the allocation succeeds it will
1342  * be locked on return.
1343  *
1344  * Arguments:
1345  *	flags   Wait flags for the item initialization routine
1346  *	aflags  Wait flags for the slab allocation
1347  *
1348  * Returns:
1349  *	The slab that was allocated or NULL if there is no memory and the
1350  *	caller specified M_NOWAIT.
1351  */
1352 static uma_slab_t
1353 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1354     int aflags)
1355 {
1356 	uma_domain_t dom;
1357 	uma_alloc allocf;
1358 	uma_slab_t slab;
1359 	unsigned long size;
1360 	uint8_t *mem;
1361 	uint8_t sflags;
1362 	int i;
1363 
1364 	KASSERT(domain >= 0 && domain < vm_ndomains,
1365 	    ("keg_alloc_slab: domain %d out of range", domain));
1366 
1367 	allocf = keg->uk_allocf;
1368 	slab = NULL;
1369 	mem = NULL;
1370 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1371 		uma_hash_slab_t hslab;
1372 		hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1373 		    domain, aflags);
1374 		if (hslab == NULL)
1375 			goto fail;
1376 		slab = &hslab->uhs_slab;
1377 	}
1378 
1379 	/*
1380 	 * This reproduces the old vm_zone behavior of zero filling pages the
1381 	 * first time they are added to a zone.
1382 	 *
1383 	 * Malloced items are zeroed in uma_zalloc.
1384 	 */
1385 
1386 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1387 		aflags |= M_ZERO;
1388 	else
1389 		aflags &= ~M_ZERO;
1390 
1391 	if (keg->uk_flags & UMA_ZONE_NODUMP)
1392 		aflags |= M_NODUMP;
1393 
1394 	/* zone is passed for legacy reasons. */
1395 	size = keg->uk_ppera * PAGE_SIZE;
1396 	mem = allocf(zone, size, domain, &sflags, aflags);
1397 	if (mem == NULL) {
1398 		if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1399 			zone_free_item(slabzone(keg->uk_ipers),
1400 			    slab_tohashslab(slab), NULL, SKIP_NONE);
1401 		goto fail;
1402 	}
1403 	uma_total_inc(size);
1404 
1405 	/* For HASH zones all pages go to the same uma_domain. */
1406 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1407 		domain = 0;
1408 
1409 	/* Point the slab into the allocated memory */
1410 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1411 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
1412 	else
1413 		slab_tohashslab(slab)->uhs_data = mem;
1414 
1415 	if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1416 		for (i = 0; i < keg->uk_ppera; i++)
1417 			vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1418 			    zone, slab);
1419 
1420 	slab->us_freecount = keg->uk_ipers;
1421 	slab->us_flags = sflags;
1422 	slab->us_domain = domain;
1423 
1424 	BIT_FILL(keg->uk_ipers, &slab->us_free);
1425 #ifdef INVARIANTS
1426 	BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1427 #endif
1428 
1429 	if (keg->uk_init != NULL) {
1430 		for (i = 0; i < keg->uk_ipers; i++)
1431 			if (keg->uk_init(slab_item(slab, keg, i),
1432 			    keg->uk_size, flags) != 0)
1433 				break;
1434 		if (i != keg->uk_ipers) {
1435 			keg_free_slab(keg, slab, i);
1436 			goto fail;
1437 		}
1438 	}
1439 	KEG_LOCK(keg, domain);
1440 
1441 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1442 	    slab, keg->uk_name, keg);
1443 
1444 	if (keg->uk_flags & UMA_ZFLAG_HASH)
1445 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1446 
1447 	/*
1448 	 * If we got a slab here it's safe to mark it partially used
1449 	 * and return.  We assume that the caller is going to remove
1450 	 * at least one item.
1451 	 */
1452 	dom = &keg->uk_domain[domain];
1453 	LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1454 	dom->ud_pages += keg->uk_ppera;
1455 	dom->ud_free += keg->uk_ipers;
1456 
1457 	return (slab);
1458 
1459 fail:
1460 	return (NULL);
1461 }
1462 
1463 /*
1464  * This function is intended to be used early on in place of page_alloc() so
1465  * that we may use the boot time page cache to satisfy allocations before
1466  * the VM is ready.
1467  */
1468 static void *
1469 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1470     int wait)
1471 {
1472 	vm_paddr_t pa;
1473 	vm_page_t m;
1474 	void *mem;
1475 	int pages;
1476 	int i;
1477 
1478 	pages = howmany(bytes, PAGE_SIZE);
1479 	KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1480 
1481 	*pflag = UMA_SLAB_BOOT;
1482 	m = vm_page_alloc_contig_domain(NULL, 0, domain,
1483 	    malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1484 	    (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1485 	if (m == NULL)
1486 		return (NULL);
1487 
1488 	pa = VM_PAGE_TO_PHYS(m);
1489 	for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1490 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1491     defined(__riscv) || defined(__powerpc64__)
1492 		if ((wait & M_NODUMP) == 0)
1493 			dump_add_page(pa);
1494 #endif
1495 	}
1496 	/* Allocate KVA and indirectly advance bootmem. */
1497 	mem = (void *)pmap_map(&bootmem, m->phys_addr,
1498 	    m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1499         if ((wait & M_ZERO) != 0)
1500                 bzero(mem, pages * PAGE_SIZE);
1501 
1502         return (mem);
1503 }
1504 
1505 static void
1506 startup_free(void *mem, vm_size_t bytes)
1507 {
1508 	vm_offset_t va;
1509 	vm_page_t m;
1510 
1511 	va = (vm_offset_t)mem;
1512 	m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1513 	pmap_remove(kernel_pmap, va, va + bytes);
1514 	for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1515 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1516     defined(__riscv) || defined(__powerpc64__)
1517 		dump_drop_page(VM_PAGE_TO_PHYS(m));
1518 #endif
1519 		vm_page_unwire_noq(m);
1520 		vm_page_free(m);
1521 	}
1522 }
1523 
1524 /*
1525  * Allocates a number of pages from the system
1526  *
1527  * Arguments:
1528  *	bytes  The number of bytes requested
1529  *	wait  Shall we wait?
1530  *
1531  * Returns:
1532  *	A pointer to the alloced memory or possibly
1533  *	NULL if M_NOWAIT is set.
1534  */
1535 static void *
1536 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1537     int wait)
1538 {
1539 	void *p;	/* Returned page */
1540 
1541 	*pflag = UMA_SLAB_KERNEL;
1542 	p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1543 
1544 	return (p);
1545 }
1546 
1547 static void *
1548 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1549     int wait)
1550 {
1551 	struct pglist alloctail;
1552 	vm_offset_t addr, zkva;
1553 	int cpu, flags;
1554 	vm_page_t p, p_next;
1555 #ifdef NUMA
1556 	struct pcpu *pc;
1557 #endif
1558 
1559 	MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1560 
1561 	TAILQ_INIT(&alloctail);
1562 	flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1563 	    malloc2vm_flags(wait);
1564 	*pflag = UMA_SLAB_KERNEL;
1565 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1566 		if (CPU_ABSENT(cpu)) {
1567 			p = vm_page_alloc(NULL, 0, flags);
1568 		} else {
1569 #ifndef NUMA
1570 			p = vm_page_alloc(NULL, 0, flags);
1571 #else
1572 			pc = pcpu_find(cpu);
1573 			if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1574 				p = NULL;
1575 			else
1576 				p = vm_page_alloc_domain(NULL, 0,
1577 				    pc->pc_domain, flags);
1578 			if (__predict_false(p == NULL))
1579 				p = vm_page_alloc(NULL, 0, flags);
1580 #endif
1581 		}
1582 		if (__predict_false(p == NULL))
1583 			goto fail;
1584 		TAILQ_INSERT_TAIL(&alloctail, p, listq);
1585 	}
1586 	if ((addr = kva_alloc(bytes)) == 0)
1587 		goto fail;
1588 	zkva = addr;
1589 	TAILQ_FOREACH(p, &alloctail, listq) {
1590 		pmap_qenter(zkva, &p, 1);
1591 		zkva += PAGE_SIZE;
1592 	}
1593 	return ((void*)addr);
1594 fail:
1595 	TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1596 		vm_page_unwire_noq(p);
1597 		vm_page_free(p);
1598 	}
1599 	return (NULL);
1600 }
1601 
1602 /*
1603  * Allocates a number of pages from within an object
1604  *
1605  * Arguments:
1606  *	bytes  The number of bytes requested
1607  *	wait   Shall we wait?
1608  *
1609  * Returns:
1610  *	A pointer to the alloced memory or possibly
1611  *	NULL if M_NOWAIT is set.
1612  */
1613 static void *
1614 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1615     int wait)
1616 {
1617 	TAILQ_HEAD(, vm_page) alloctail;
1618 	u_long npages;
1619 	vm_offset_t retkva, zkva;
1620 	vm_page_t p, p_next;
1621 	uma_keg_t keg;
1622 
1623 	TAILQ_INIT(&alloctail);
1624 	keg = zone->uz_keg;
1625 
1626 	npages = howmany(bytes, PAGE_SIZE);
1627 	while (npages > 0) {
1628 		p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1629 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1630 		    ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1631 		    VM_ALLOC_NOWAIT));
1632 		if (p != NULL) {
1633 			/*
1634 			 * Since the page does not belong to an object, its
1635 			 * listq is unused.
1636 			 */
1637 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1638 			npages--;
1639 			continue;
1640 		}
1641 		/*
1642 		 * Page allocation failed, free intermediate pages and
1643 		 * exit.
1644 		 */
1645 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1646 			vm_page_unwire_noq(p);
1647 			vm_page_free(p);
1648 		}
1649 		return (NULL);
1650 	}
1651 	*flags = UMA_SLAB_PRIV;
1652 	zkva = keg->uk_kva +
1653 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1654 	retkva = zkva;
1655 	TAILQ_FOREACH(p, &alloctail, listq) {
1656 		pmap_qenter(zkva, &p, 1);
1657 		zkva += PAGE_SIZE;
1658 	}
1659 
1660 	return ((void *)retkva);
1661 }
1662 
1663 /*
1664  * Frees a number of pages to the system
1665  *
1666  * Arguments:
1667  *	mem   A pointer to the memory to be freed
1668  *	size  The size of the memory being freed
1669  *	flags The original p->us_flags field
1670  *
1671  * Returns:
1672  *	Nothing
1673  */
1674 static void
1675 page_free(void *mem, vm_size_t size, uint8_t flags)
1676 {
1677 
1678 	if ((flags & UMA_SLAB_BOOT) != 0) {
1679 		startup_free(mem, size);
1680 		return;
1681 	}
1682 
1683 	if ((flags & UMA_SLAB_KERNEL) == 0)
1684 		panic("UMA: page_free used with invalid flags %x", flags);
1685 
1686 	kmem_free((vm_offset_t)mem, size);
1687 }
1688 
1689 /*
1690  * Frees pcpu zone allocations
1691  *
1692  * Arguments:
1693  *	mem   A pointer to the memory to be freed
1694  *	size  The size of the memory being freed
1695  *	flags The original p->us_flags field
1696  *
1697  * Returns:
1698  *	Nothing
1699  */
1700 static void
1701 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1702 {
1703 	vm_offset_t sva, curva;
1704 	vm_paddr_t paddr;
1705 	vm_page_t m;
1706 
1707 	MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1708 	sva = (vm_offset_t)mem;
1709 	for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1710 		paddr = pmap_kextract(curva);
1711 		m = PHYS_TO_VM_PAGE(paddr);
1712 		vm_page_unwire_noq(m);
1713 		vm_page_free(m);
1714 	}
1715 	pmap_qremove(sva, size >> PAGE_SHIFT);
1716 	kva_free(sva, size);
1717 }
1718 
1719 
1720 /*
1721  * Zero fill initializer
1722  *
1723  * Arguments/Returns follow uma_init specifications
1724  */
1725 static int
1726 zero_init(void *mem, int size, int flags)
1727 {
1728 	bzero(mem, size);
1729 	return (0);
1730 }
1731 
1732 #ifdef INVARIANTS
1733 struct noslabbits *
1734 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1735 {
1736 
1737 	return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1738 }
1739 #endif
1740 
1741 /*
1742  * Actual size of embedded struct slab (!OFFPAGE).
1743  */
1744 size_t
1745 slab_sizeof(int nitems)
1746 {
1747 	size_t s;
1748 
1749 	s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1750 	return (roundup(s, UMA_ALIGN_PTR + 1));
1751 }
1752 
1753 /*
1754  * Size of memory for embedded slabs (!OFFPAGE).
1755  */
1756 size_t
1757 slab_space(int nitems)
1758 {
1759 	return (UMA_SLAB_SIZE - slab_sizeof(nitems));
1760 }
1761 
1762 #define	UMA_FIXPT_SHIFT	31
1763 #define	UMA_FRAC_FIXPT(n, d)						\
1764 	((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1765 #define	UMA_FIXPT_PCT(f)						\
1766 	((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1767 #define	UMA_PCT_FIXPT(pct)	UMA_FRAC_FIXPT((pct), 100)
1768 #define	UMA_MIN_EFF	UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1769 
1770 /*
1771  * Compute the number of items that will fit in a slab.  If hdr is true, the
1772  * item count may be limited to provide space in the slab for an inline slab
1773  * header.  Otherwise, all slab space will be provided for item storage.
1774  */
1775 static u_int
1776 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1777 {
1778 	u_int ipers;
1779 	u_int padpi;
1780 
1781 	/* The padding between items is not needed after the last item. */
1782 	padpi = rsize - size;
1783 
1784 	if (hdr) {
1785 		/*
1786 		 * Start with the maximum item count and remove items until
1787 		 * the slab header first alongside the allocatable memory.
1788 		 */
1789 		for (ipers = MIN(SLAB_MAX_SETSIZE,
1790 		    (slabsize + padpi - slab_sizeof(1)) / rsize);
1791 		    ipers > 0 &&
1792 		    ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1793 		    ipers--)
1794 			continue;
1795 	} else {
1796 		ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1797 	}
1798 
1799 	return (ipers);
1800 }
1801 
1802 /*
1803  * Compute the number of items that will fit in a slab for a startup zone.
1804  */
1805 int
1806 slab_ipers(size_t size, int align)
1807 {
1808 	int rsize;
1809 
1810 	rsize = roundup(size, align + 1); /* Assume no CACHESPREAD */
1811 	return (slab_ipers_hdr(size, rsize, UMA_SLAB_SIZE, true));
1812 }
1813 
1814 /*
1815  * Determine the format of a uma keg.  This determines where the slab header
1816  * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
1817  *
1818  * Arguments
1819  *	keg  The zone we should initialize
1820  *
1821  * Returns
1822  *	Nothing
1823  */
1824 static void
1825 keg_layout(uma_keg_t keg)
1826 {
1827 	u_int alignsize;
1828 	u_int eff;
1829 	u_int eff_offpage;
1830 	u_int format;
1831 	u_int ipers;
1832 	u_int ipers_offpage;
1833 	u_int pages;
1834 	u_int rsize;
1835 	u_int slabsize;
1836 
1837 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1838 	    (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
1839 	     (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
1840 	    ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
1841 	     __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
1842 	     PRINT_UMA_ZFLAGS));
1843 	KASSERT((keg->uk_flags &
1844 	    (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY)) == 0 ||
1845 	    (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
1846 	    ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
1847 	     PRINT_UMA_ZFLAGS));
1848 
1849 	alignsize = keg->uk_align + 1;
1850 	format = 0;
1851 	ipers = 0;
1852 
1853 	/*
1854 	 * Calculate the size of each allocation (rsize) according to
1855 	 * alignment.  If the requested size is smaller than we have
1856 	 * allocation bits for we round it up.
1857 	 */
1858 	rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
1859 	rsize = roundup2(rsize, alignsize);
1860 
1861 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0) {
1862 		slabsize = UMA_PCPU_ALLOC_SIZE;
1863 		pages = mp_maxid + 1;
1864 	} else if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
1865 		/*
1866 		 * We want one item to start on every align boundary in a page.
1867 		 * To do this we will span pages.  We will also extend the item
1868 		 * by the size of align if it is an even multiple of align.
1869 		 * Otherwise, it would fall on the same boundary every time.
1870 		 */
1871 		if ((rsize & alignsize) == 0)
1872 			rsize += alignsize;
1873 		slabsize = rsize * (PAGE_SIZE / alignsize);
1874 		slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
1875 		slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
1876 		pages = howmany(slabsize, PAGE_SIZE);
1877 		slabsize = ptoa(pages);
1878 	} else {
1879 		/*
1880 		 * Choose a slab size of as many pages as it takes to represent
1881 		 * a single item.  We will then try to fit as many additional
1882 		 * items into the slab as possible.  At some point, we may want
1883 		 * to increase the slab size for awkward item sizes in order to
1884 		 * increase efficiency.
1885 		 */
1886 		pages = howmany(keg->uk_size, PAGE_SIZE);
1887 		slabsize = ptoa(pages);
1888 	}
1889 
1890 	/* Evaluate an inline slab layout. */
1891 	if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
1892 		ipers = slab_ipers_hdr(keg->uk_size, rsize, slabsize, true);
1893 
1894 	/* TODO: vm_page-embedded slab. */
1895 
1896 	/*
1897 	 * We can't do OFFPAGE if we're internal or if we've been
1898 	 * asked to not go to the VM for buckets.  If we do this we
1899 	 * may end up going to the VM  for slabs which we do not
1900 	 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1901 	 * of UMA_ZONE_VM, which clearly forbids it.
1902 	 */
1903 	if ((keg->uk_flags &
1904 	    (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY)) != 0) {
1905 		if (ipers == 0) {
1906 			/* We need an extra page for the slab header. */
1907 			pages++;
1908 			slabsize = ptoa(pages);
1909 			ipers = slab_ipers_hdr(keg->uk_size, rsize, slabsize,
1910 			    true);
1911 		}
1912 		goto out;
1913 	}
1914 
1915 	/*
1916 	 * See if using an OFFPAGE slab will improve our efficiency.
1917 	 * Only do this if we are below our efficiency threshold.
1918 	 *
1919 	 * XXX We could try growing slabsize to limit max waste as well.
1920 	 * Historically this was not done because the VM could not
1921 	 * efficiently handle contiguous allocations.
1922 	 */
1923 	eff = UMA_FRAC_FIXPT(ipers * rsize, slabsize);
1924 	ipers_offpage = slab_ipers_hdr(keg->uk_size, rsize, slabsize, false);
1925 	eff_offpage = UMA_FRAC_FIXPT(ipers_offpage * rsize,
1926 	    slabsize + slabzone(ipers_offpage)->uz_keg->uk_rsize);
1927 	if (ipers == 0 || (eff < UMA_MIN_EFF && eff < eff_offpage)) {
1928 		CTR5(KTR_UMA, "UMA decided we need offpage slab headers for "
1929 		    "keg: %s(%p), minimum efficiency allowed = %u%%, "
1930 		    "old efficiency = %u%%, offpage efficiency = %u%%",
1931 		    keg->uk_name, keg, UMA_FIXPT_PCT(UMA_MIN_EFF),
1932 		    UMA_FIXPT_PCT(eff), UMA_FIXPT_PCT(eff_offpage));
1933 		format = UMA_ZFLAG_OFFPAGE;
1934 		ipers = ipers_offpage;
1935 	}
1936 
1937 out:
1938 	/*
1939 	 * How do we find the slab header if it is offpage or if not all item
1940 	 * start addresses are in the same page?  We could solve the latter
1941 	 * case with vaddr alignment, but we don't.
1942 	 */
1943 	if ((format & UMA_ZFLAG_OFFPAGE) != 0 ||
1944 	    (ipers - 1) * rsize >= PAGE_SIZE) {
1945 		if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
1946 			format |= UMA_ZFLAG_HASH;
1947 		else
1948 			format |= UMA_ZFLAG_VTOSLAB;
1949 	}
1950 	keg->uk_ipers = ipers;
1951 	keg->uk_rsize = rsize;
1952 	keg->uk_flags |= format;
1953 	keg->uk_ppera = pages;
1954 	CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
1955 	    __func__, keg->uk_name, keg->uk_flags, rsize, ipers, pages);
1956 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
1957 	    ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
1958 	     keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize, ipers,
1959 	     pages));
1960 }
1961 
1962 /*
1963  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1964  * the keg onto the global keg list.
1965  *
1966  * Arguments/Returns follow uma_ctor specifications
1967  *	udata  Actually uma_kctor_args
1968  */
1969 static int
1970 keg_ctor(void *mem, int size, void *udata, int flags)
1971 {
1972 	struct uma_kctor_args *arg = udata;
1973 	uma_keg_t keg = mem;
1974 	uma_zone_t zone;
1975 	int i;
1976 
1977 	bzero(keg, size);
1978 	keg->uk_size = arg->size;
1979 	keg->uk_init = arg->uminit;
1980 	keg->uk_fini = arg->fini;
1981 	keg->uk_align = arg->align;
1982 	keg->uk_reserve = 0;
1983 	keg->uk_flags = arg->flags;
1984 
1985 	/*
1986 	 * We use a global round-robin policy by default.  Zones with
1987 	 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
1988 	 * case the iterator is never run.
1989 	 */
1990 	keg->uk_dr.dr_policy = DOMAINSET_RR();
1991 	keg->uk_dr.dr_iter = 0;
1992 
1993 	/*
1994 	 * The master zone is passed to us at keg-creation time.
1995 	 */
1996 	zone = arg->zone;
1997 	keg->uk_name = zone->uz_name;
1998 
1999 	if (arg->flags & UMA_ZONE_VM)
2000 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
2001 
2002 	if (arg->flags & UMA_ZONE_ZINIT)
2003 		keg->uk_init = zero_init;
2004 
2005 	if (arg->flags & UMA_ZONE_MALLOC)
2006 		keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2007 
2008 #ifndef SMP
2009 	keg->uk_flags &= ~UMA_ZONE_PCPU;
2010 #endif
2011 
2012 	keg_layout(keg);
2013 
2014 	/*
2015 	 * Use a first-touch NUMA policy for all kegs that pmap_extract()
2016 	 * will work on with the exception of critical VM structures
2017 	 * necessary for paging.
2018 	 *
2019 	 * Zones may override the default by specifying either.
2020 	 */
2021 #ifdef NUMA
2022 	if ((keg->uk_flags &
2023 	    (UMA_ZFLAG_HASH | UMA_ZONE_VM | UMA_ZONE_ROUNDROBIN)) == 0)
2024 		keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2025 	else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2026 		keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2027 #endif
2028 
2029 	/*
2030 	 * If we haven't booted yet we need allocations to go through the
2031 	 * startup cache until the vm is ready.
2032 	 */
2033 #ifdef UMA_MD_SMALL_ALLOC
2034 	if (keg->uk_ppera == 1)
2035 		keg->uk_allocf = uma_small_alloc;
2036 	else
2037 #endif
2038 	if (booted < BOOT_KVA)
2039 		keg->uk_allocf = startup_alloc;
2040 	else if (keg->uk_flags & UMA_ZONE_PCPU)
2041 		keg->uk_allocf = pcpu_page_alloc;
2042 	else
2043 		keg->uk_allocf = page_alloc;
2044 #ifdef UMA_MD_SMALL_ALLOC
2045 	if (keg->uk_ppera == 1)
2046 		keg->uk_freef = uma_small_free;
2047 	else
2048 #endif
2049 	if (keg->uk_flags & UMA_ZONE_PCPU)
2050 		keg->uk_freef = pcpu_page_free;
2051 	else
2052 		keg->uk_freef = page_free;
2053 
2054 	/*
2055 	 * Initialize keg's locks.
2056 	 */
2057 	for (i = 0; i < vm_ndomains; i++)
2058 		KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2059 
2060 	/*
2061 	 * If we're putting the slab header in the actual page we need to
2062 	 * figure out where in each page it goes.  See slab_sizeof
2063 	 * definition.
2064 	 */
2065 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2066 		size_t shsize;
2067 
2068 		shsize = slab_sizeof(keg->uk_ipers);
2069 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2070 		/*
2071 		 * The only way the following is possible is if with our
2072 		 * UMA_ALIGN_PTR adjustments we are now bigger than
2073 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
2074 		 * mathematically possible for all cases, so we make
2075 		 * sure here anyway.
2076 		 */
2077 		KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2078 		    ("zone %s ipers %d rsize %d size %d slab won't fit",
2079 		    zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2080 	}
2081 
2082 	if (keg->uk_flags & UMA_ZFLAG_HASH)
2083 		hash_alloc(&keg->uk_hash, 0);
2084 
2085 	CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2086 
2087 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2088 
2089 	rw_wlock(&uma_rwlock);
2090 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2091 	rw_wunlock(&uma_rwlock);
2092 	return (0);
2093 }
2094 
2095 static void
2096 zone_kva_available(uma_zone_t zone, void *unused)
2097 {
2098 	uma_keg_t keg;
2099 
2100 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2101 		return;
2102 	KEG_GET(zone, keg);
2103 	if (keg->uk_flags & UMA_ZONE_PCPU)
2104 		keg->uk_allocf = pcpu_page_alloc;
2105 	else if (keg->uk_allocf == startup_alloc)
2106 		keg->uk_allocf = page_alloc;
2107 }
2108 
2109 static void
2110 zone_alloc_counters(uma_zone_t zone, void *unused)
2111 {
2112 
2113 	zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2114 	zone->uz_frees = counter_u64_alloc(M_WAITOK);
2115 	zone->uz_fails = counter_u64_alloc(M_WAITOK);
2116 }
2117 
2118 static void
2119 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2120 {
2121 	uma_zone_domain_t zdom;
2122 	uma_domain_t dom;
2123 	uma_keg_t keg;
2124 	struct sysctl_oid *oid, *domainoid;
2125 	int domains, i, cnt;
2126 	static const char *nokeg = "cache zone";
2127 	char *c;
2128 
2129 	/*
2130 	 * Make a sysctl safe copy of the zone name by removing
2131 	 * any special characters and handling dups by appending
2132 	 * an index.
2133 	 */
2134 	if (zone->uz_namecnt != 0) {
2135 		/* Count the number of decimal digits and '_' separator. */
2136 		for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2137 			cnt /= 10;
2138 		zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2139 		    M_UMA, M_WAITOK);
2140 		sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2141 		    zone->uz_namecnt);
2142 	} else
2143 		zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2144 	for (c = zone->uz_ctlname; *c != '\0'; c++)
2145 		if (strchr("./\\ -", *c) != NULL)
2146 			*c = '_';
2147 
2148 	/*
2149 	 * Basic parameters at the root.
2150 	 */
2151 	zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2152 	    OID_AUTO, zone->uz_ctlname, CTLFLAG_RD, NULL, "");
2153 	oid = zone->uz_oid;
2154 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2155 	    "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2156 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2157 	    "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2158 	    zone, 0, sysctl_handle_uma_zone_flags, "A",
2159 	    "Allocator configuration flags");
2160 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2161 	    "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2162 	    "Desired per-cpu cache size");
2163 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2164 	    "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2165 	    "Maximum allowed per-cpu cache size");
2166 
2167 	/*
2168 	 * keg if present.
2169 	 */
2170 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2171 		domains = vm_ndomains;
2172 	else
2173 		domains = 1;
2174 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2175 	    "keg", CTLFLAG_RD, NULL, "");
2176 	keg = zone->uz_keg;
2177 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2178 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2179 		    "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2180 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2181 		    "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2182 		    "Real object size with alignment");
2183 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2184 		    "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2185 		    "pages per-slab allocation");
2186 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2187 		    "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2188 		    "items available per-slab");
2189 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2190 		    "align", CTLFLAG_RD, &keg->uk_align, 0,
2191 		    "item alignment mask");
2192 		SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2193 		    "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2194 		    keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2195 		    "Slab utilization (100 - internal fragmentation %)");
2196 		domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2197 		    OID_AUTO, "domain", CTLFLAG_RD, NULL, "");
2198 		for (i = 0; i < domains; i++) {
2199 			dom = &keg->uk_domain[i];
2200 			oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2201 			    OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD,
2202 			    NULL, "");
2203 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2204 			    "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2205 			    "Total pages currently allocated from VM");
2206 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2207 			    "free", CTLFLAG_RD, &dom->ud_free, 0,
2208 			    "items free in the slab layer");
2209 		}
2210 	} else
2211 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2212 		    "name", CTLFLAG_RD, nokeg, "Keg name");
2213 
2214 	/*
2215 	 * Information about zone limits.
2216 	 */
2217 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2218 	    "limit", CTLFLAG_RD, NULL, "");
2219 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2220 	    "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2221 	    zone, 0, sysctl_handle_uma_zone_items, "QU",
2222 	    "current number of allocated items if limit is set");
2223 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2224 	    "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2225 	    "Maximum number of cached items");
2226 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2227 	    "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2228 	    "Number of threads sleeping at limit");
2229 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2230 	    "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2231 	    "Total zone limit sleeps");
2232 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2233 	    "bucket_max", CTLFLAG_RD, &zone->uz_bkt_max, 0,
2234 	    "Maximum number of items in the bucket cache");
2235 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2236 	    "bucket_cnt", CTLFLAG_RD, &zone->uz_bkt_count, 0,
2237 	    "Number of items in the bucket cache");
2238 
2239 	/*
2240 	 * Per-domain zone information.
2241 	 */
2242 	domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2243 	    OID_AUTO, "domain", CTLFLAG_RD, NULL, "");
2244 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2245 		domains = 1;
2246 	for (i = 0; i < domains; i++) {
2247 		zdom = &zone->uz_domain[i];
2248 		oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2249 		    OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD, NULL, "");
2250 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2251 		    "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2252 		    "number of items in this domain");
2253 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2254 		    "imax", CTLFLAG_RD, &zdom->uzd_imax,
2255 		    "maximum item count in this period");
2256 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2257 		    "imin", CTLFLAG_RD, &zdom->uzd_imin,
2258 		    "minimum item count in this period");
2259 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2260 		    "wss", CTLFLAG_RD, &zdom->uzd_wss,
2261 		    "Working set size");
2262 	}
2263 
2264 	/*
2265 	 * General statistics.
2266 	 */
2267 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2268 	    "stats", CTLFLAG_RD, NULL, "");
2269 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2270 	    "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2271 	    zone, 1, sysctl_handle_uma_zone_cur, "I",
2272 	    "Current number of allocated items");
2273 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2274 	    "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2275 	    zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2276 	    "Total allocation calls");
2277 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2278 	    "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2279 	    zone, 0, sysctl_handle_uma_zone_frees, "QU",
2280 	    "Total free calls");
2281 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2282 	    "fails", CTLFLAG_RD, &zone->uz_fails,
2283 	    "Number of allocation failures");
2284 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2285 	    "xdomain", CTLFLAG_RD, &zone->uz_xdomain, 0,
2286 	    "Free calls from the wrong domain");
2287 }
2288 
2289 struct uma_zone_count {
2290 	const char	*name;
2291 	int		count;
2292 };
2293 
2294 static void
2295 zone_count(uma_zone_t zone, void *arg)
2296 {
2297 	struct uma_zone_count *cnt;
2298 
2299 	cnt = arg;
2300 	/*
2301 	 * Some zones are rapidly created with identical names and
2302 	 * destroyed out of order.  This can lead to gaps in the count.
2303 	 * Use one greater than the maximum observed for this name.
2304 	 */
2305 	if (strcmp(zone->uz_name, cnt->name) == 0)
2306 		cnt->count = MAX(cnt->count,
2307 		    zone->uz_namecnt + 1);
2308 }
2309 
2310 static void
2311 zone_update_caches(uma_zone_t zone)
2312 {
2313 	int i;
2314 
2315 	for (i = 0; i <= mp_maxid; i++) {
2316 		cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2317 		cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2318 	}
2319 }
2320 
2321 /*
2322  * Zone header ctor.  This initializes all fields, locks, etc.
2323  *
2324  * Arguments/Returns follow uma_ctor specifications
2325  *	udata  Actually uma_zctor_args
2326  */
2327 static int
2328 zone_ctor(void *mem, int size, void *udata, int flags)
2329 {
2330 	struct uma_zone_count cnt;
2331 	struct uma_zctor_args *arg = udata;
2332 	uma_zone_t zone = mem;
2333 	uma_zone_t z;
2334 	uma_keg_t keg;
2335 	int i;
2336 
2337 	bzero(zone, size);
2338 	zone->uz_name = arg->name;
2339 	zone->uz_ctor = arg->ctor;
2340 	zone->uz_dtor = arg->dtor;
2341 	zone->uz_init = NULL;
2342 	zone->uz_fini = NULL;
2343 	zone->uz_sleeps = 0;
2344 	zone->uz_xdomain = 0;
2345 	zone->uz_bucket_size = 0;
2346 	zone->uz_bucket_size_min = 0;
2347 	zone->uz_bucket_size_max = BUCKET_MAX;
2348 	zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2349 	zone->uz_warning = NULL;
2350 	/* The domain structures follow the cpu structures. */
2351 	zone->uz_domain =
2352 	    (struct uma_zone_domain *)&zone->uz_cpu[mp_maxid + 1];
2353 	zone->uz_bkt_max = ULONG_MAX;
2354 	timevalclear(&zone->uz_ratecheck);
2355 
2356 	/* Count the number of duplicate names. */
2357 	cnt.name = arg->name;
2358 	cnt.count = 0;
2359 	zone_foreach(zone_count, &cnt);
2360 	zone->uz_namecnt = cnt.count;
2361 	ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
2362 	ZONE_CROSS_LOCK_INIT(zone);
2363 
2364 	for (i = 0; i < vm_ndomains; i++)
2365 		STAILQ_INIT(&zone->uz_domain[i].uzd_buckets);
2366 
2367 #ifdef INVARIANTS
2368 	if (arg->uminit == trash_init && arg->fini == trash_fini)
2369 		zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2370 #endif
2371 
2372 	/*
2373 	 * This is a pure cache zone, no kegs.
2374 	 */
2375 	if (arg->import) {
2376 		KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2377 		    ("zone_ctor: Import specified for non-cache zone."));
2378 		if (arg->flags & UMA_ZONE_VM)
2379 			arg->flags |= UMA_ZFLAG_CACHEONLY;
2380 		zone->uz_flags = arg->flags;
2381 		zone->uz_size = arg->size;
2382 		zone->uz_import = arg->import;
2383 		zone->uz_release = arg->release;
2384 		zone->uz_arg = arg->arg;
2385 		rw_wlock(&uma_rwlock);
2386 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2387 		rw_wunlock(&uma_rwlock);
2388 		goto out;
2389 	}
2390 
2391 	/*
2392 	 * Use the regular zone/keg/slab allocator.
2393 	 */
2394 	zone->uz_import = zone_import;
2395 	zone->uz_release = zone_release;
2396 	zone->uz_arg = zone;
2397 	keg = arg->keg;
2398 
2399 	if (arg->flags & UMA_ZONE_SECONDARY) {
2400 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2401 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2402 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2403 		zone->uz_init = arg->uminit;
2404 		zone->uz_fini = arg->fini;
2405 		zone->uz_flags |= UMA_ZONE_SECONDARY;
2406 		rw_wlock(&uma_rwlock);
2407 		ZONE_LOCK(zone);
2408 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2409 			if (LIST_NEXT(z, uz_link) == NULL) {
2410 				LIST_INSERT_AFTER(z, zone, uz_link);
2411 				break;
2412 			}
2413 		}
2414 		ZONE_UNLOCK(zone);
2415 		rw_wunlock(&uma_rwlock);
2416 	} else if (keg == NULL) {
2417 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2418 		    arg->align, arg->flags)) == NULL)
2419 			return (ENOMEM);
2420 	} else {
2421 		struct uma_kctor_args karg;
2422 		int error;
2423 
2424 		/* We should only be here from uma_startup() */
2425 		karg.size = arg->size;
2426 		karg.uminit = arg->uminit;
2427 		karg.fini = arg->fini;
2428 		karg.align = arg->align;
2429 		karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2430 		karg.zone = zone;
2431 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2432 		    flags);
2433 		if (error)
2434 			return (error);
2435 	}
2436 
2437 	/* Inherit properties from the keg. */
2438 	zone->uz_keg = keg;
2439 	zone->uz_size = keg->uk_size;
2440 	zone->uz_flags |= (keg->uk_flags &
2441 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2442 
2443 out:
2444 	if (__predict_true(booted >= BOOT_RUNNING)) {
2445 		zone_alloc_counters(zone, NULL);
2446 		zone_alloc_sysctl(zone, NULL);
2447 	} else {
2448 		zone->uz_allocs = EARLY_COUNTER;
2449 		zone->uz_frees = EARLY_COUNTER;
2450 		zone->uz_fails = EARLY_COUNTER;
2451 	}
2452 
2453 	/* Caller requests a private SMR context. */
2454 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2455 		zone->uz_smr = smr_create(zone->uz_name);
2456 
2457 	KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2458 	    (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2459 	    ("Invalid zone flag combination"));
2460 	if (arg->flags & UMA_ZFLAG_INTERNAL)
2461 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2462 	if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2463 		zone->uz_bucket_size = BUCKET_MAX;
2464 	else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2465 		zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2466 	else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2467 		zone->uz_bucket_size = 0;
2468 	else
2469 		zone->uz_bucket_size = bucket_select(zone->uz_size);
2470 	zone->uz_bucket_size_min = zone->uz_bucket_size;
2471 	if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2472 		zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2473 	zone_update_caches(zone);
2474 
2475 	return (0);
2476 }
2477 
2478 /*
2479  * Keg header dtor.  This frees all data, destroys locks, frees the hash
2480  * table and removes the keg from the global list.
2481  *
2482  * Arguments/Returns follow uma_dtor specifications
2483  *	udata  unused
2484  */
2485 static void
2486 keg_dtor(void *arg, int size, void *udata)
2487 {
2488 	uma_keg_t keg;
2489 	uint32_t free, pages;
2490 	int i;
2491 
2492 	keg = (uma_keg_t)arg;
2493 	free = pages = 0;
2494 	for (i = 0; i < vm_ndomains; i++) {
2495 		free += keg->uk_domain[i].ud_free;
2496 		pages += keg->uk_domain[i].ud_pages;
2497 		KEG_LOCK_FINI(keg, i);
2498 	}
2499 	if (pages != 0)
2500 		printf("Freed UMA keg (%s) was not empty (%u items). "
2501 		    " Lost %u pages of memory.\n",
2502 		    keg->uk_name ? keg->uk_name : "",
2503 		    pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2504 
2505 	hash_free(&keg->uk_hash);
2506 }
2507 
2508 /*
2509  * Zone header dtor.
2510  *
2511  * Arguments/Returns follow uma_dtor specifications
2512  *	udata  unused
2513  */
2514 static void
2515 zone_dtor(void *arg, int size, void *udata)
2516 {
2517 	uma_zone_t zone;
2518 	uma_keg_t keg;
2519 
2520 	zone = (uma_zone_t)arg;
2521 
2522 	sysctl_remove_oid(zone->uz_oid, 1, 1);
2523 
2524 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2525 		cache_drain(zone);
2526 
2527 	rw_wlock(&uma_rwlock);
2528 	LIST_REMOVE(zone, uz_link);
2529 	rw_wunlock(&uma_rwlock);
2530 	/*
2531 	 * XXX there are some races here where
2532 	 * the zone can be drained but zone lock
2533 	 * released and then refilled before we
2534 	 * remove it... we dont care for now
2535 	 */
2536 	zone_reclaim(zone, M_WAITOK, true);
2537 	/*
2538 	 * We only destroy kegs from non secondary/non cache zones.
2539 	 */
2540 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2541 		keg = zone->uz_keg;
2542 		rw_wlock(&uma_rwlock);
2543 		LIST_REMOVE(keg, uk_link);
2544 		rw_wunlock(&uma_rwlock);
2545 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
2546 	}
2547 	counter_u64_free(zone->uz_allocs);
2548 	counter_u64_free(zone->uz_frees);
2549 	counter_u64_free(zone->uz_fails);
2550 	free(zone->uz_ctlname, M_UMA);
2551 	ZONE_LOCK_FINI(zone);
2552 	ZONE_CROSS_LOCK_FINI(zone);
2553 }
2554 
2555 static void
2556 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2557 {
2558 	uma_keg_t keg;
2559 	uma_zone_t zone;
2560 
2561 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
2562 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2563 			zfunc(zone, arg);
2564 	}
2565 	LIST_FOREACH(zone, &uma_cachezones, uz_link)
2566 		zfunc(zone, arg);
2567 }
2568 
2569 /*
2570  * Traverses every zone in the system and calls a callback
2571  *
2572  * Arguments:
2573  *	zfunc  A pointer to a function which accepts a zone
2574  *		as an argument.
2575  *
2576  * Returns:
2577  *	Nothing
2578  */
2579 static void
2580 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2581 {
2582 
2583 	rw_rlock(&uma_rwlock);
2584 	zone_foreach_unlocked(zfunc, arg);
2585 	rw_runlock(&uma_rwlock);
2586 }
2587 
2588 /*
2589  * Initialize the kernel memory allocator.  This is done after pages can be
2590  * allocated but before general KVA is available.
2591  */
2592 void
2593 uma_startup1(vm_offset_t virtual_avail)
2594 {
2595 	struct uma_zctor_args args;
2596 	size_t ksize, zsize, size;
2597 	uma_keg_t masterkeg;
2598 	uintptr_t m;
2599 	uint8_t pflag;
2600 
2601 	bootstart = bootmem = virtual_avail;
2602 
2603 	rw_init(&uma_rwlock, "UMA lock");
2604 	sx_init(&uma_reclaim_lock, "umareclaim");
2605 
2606 	ksize = sizeof(struct uma_keg) +
2607 	    (sizeof(struct uma_domain) * vm_ndomains);
2608 	ksize = roundup(ksize, UMA_SUPER_ALIGN);
2609 	zsize = sizeof(struct uma_zone) +
2610 	    (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2611 	    (sizeof(struct uma_zone_domain) * vm_ndomains);
2612 	zsize = roundup(zsize, UMA_SUPER_ALIGN);
2613 
2614 	/* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2615 	size = (zsize * 2) + ksize;
2616 	m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO);
2617 	zones = (uma_zone_t)m;
2618 	m += zsize;
2619 	kegs = (uma_zone_t)m;
2620 	m += zsize;
2621 	masterkeg = (uma_keg_t)m;
2622 
2623 	/* "manually" create the initial zone */
2624 	memset(&args, 0, sizeof(args));
2625 	args.name = "UMA Kegs";
2626 	args.size = ksize;
2627 	args.ctor = keg_ctor;
2628 	args.dtor = keg_dtor;
2629 	args.uminit = zero_init;
2630 	args.fini = NULL;
2631 	args.keg = masterkeg;
2632 	args.align = UMA_SUPER_ALIGN - 1;
2633 	args.flags = UMA_ZFLAG_INTERNAL;
2634 	zone_ctor(kegs, zsize, &args, M_WAITOK);
2635 
2636 	args.name = "UMA Zones";
2637 	args.size = zsize;
2638 	args.ctor = zone_ctor;
2639 	args.dtor = zone_dtor;
2640 	args.uminit = zero_init;
2641 	args.fini = NULL;
2642 	args.keg = NULL;
2643 	args.align = UMA_SUPER_ALIGN - 1;
2644 	args.flags = UMA_ZFLAG_INTERNAL;
2645 	zone_ctor(zones, zsize, &args, M_WAITOK);
2646 
2647 	/* Now make zones for slab headers */
2648 	slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2649 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2650 	slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2651 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2652 
2653 	hashzone = uma_zcreate("UMA Hash",
2654 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2655 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2656 
2657 	bucket_init();
2658 	smr_init();
2659 }
2660 
2661 #ifndef UMA_MD_SMALL_ALLOC
2662 extern void vm_radix_reserve_kva(void);
2663 #endif
2664 
2665 /*
2666  * Advertise the availability of normal kva allocations and switch to
2667  * the default back-end allocator.  Marks the KVA we consumed on startup
2668  * as used in the map.
2669  */
2670 void
2671 uma_startup2(void)
2672 {
2673 
2674 	if (bootstart != bootmem) {
2675 		vm_map_lock(kernel_map);
2676 		(void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2677 		    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2678 		vm_map_unlock(kernel_map);
2679 	}
2680 
2681 #ifndef UMA_MD_SMALL_ALLOC
2682 	/* Set up radix zone to use noobj_alloc. */
2683 	vm_radix_reserve_kva();
2684 #endif
2685 
2686 	booted = BOOT_KVA;
2687 	zone_foreach_unlocked(zone_kva_available, NULL);
2688 	bucket_enable();
2689 }
2690 
2691 /*
2692  * Finish our initialization steps.
2693  */
2694 static void
2695 uma_startup3(void)
2696 {
2697 
2698 #ifdef INVARIANTS
2699 	TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2700 	uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2701 	uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2702 #endif
2703 	zone_foreach_unlocked(zone_alloc_counters, NULL);
2704 	zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2705 	callout_init(&uma_callout, 1);
2706 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2707 	booted = BOOT_RUNNING;
2708 
2709 	EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2710 	    EVENTHANDLER_PRI_FIRST);
2711 }
2712 
2713 static void
2714 uma_shutdown(void)
2715 {
2716 
2717 	booted = BOOT_SHUTDOWN;
2718 }
2719 
2720 static uma_keg_t
2721 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2722 		int align, uint32_t flags)
2723 {
2724 	struct uma_kctor_args args;
2725 
2726 	args.size = size;
2727 	args.uminit = uminit;
2728 	args.fini = fini;
2729 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2730 	args.flags = flags;
2731 	args.zone = zone;
2732 	return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2733 }
2734 
2735 /* Public functions */
2736 /* See uma.h */
2737 void
2738 uma_set_align(int align)
2739 {
2740 
2741 	if (align != UMA_ALIGN_CACHE)
2742 		uma_align_cache = align;
2743 }
2744 
2745 /* See uma.h */
2746 uma_zone_t
2747 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2748 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
2749 
2750 {
2751 	struct uma_zctor_args args;
2752 	uma_zone_t res;
2753 
2754 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2755 	    align, name));
2756 
2757 	/* This stuff is essential for the zone ctor */
2758 	memset(&args, 0, sizeof(args));
2759 	args.name = name;
2760 	args.size = size;
2761 	args.ctor = ctor;
2762 	args.dtor = dtor;
2763 	args.uminit = uminit;
2764 	args.fini = fini;
2765 #ifdef  INVARIANTS
2766 	/*
2767 	 * Inject procedures which check for memory use after free if we are
2768 	 * allowed to scramble the memory while it is not allocated.  This
2769 	 * requires that: UMA is actually able to access the memory, no init
2770 	 * or fini procedures, no dependency on the initial value of the
2771 	 * memory, and no (legitimate) use of the memory after free.  Note,
2772 	 * the ctor and dtor do not need to be empty.
2773 	 */
2774 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
2775 	    UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
2776 		args.uminit = trash_init;
2777 		args.fini = trash_fini;
2778 	}
2779 #endif
2780 	args.align = align;
2781 	args.flags = flags;
2782 	args.keg = NULL;
2783 
2784 	sx_slock(&uma_reclaim_lock);
2785 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
2786 	sx_sunlock(&uma_reclaim_lock);
2787 
2788 	return (res);
2789 }
2790 
2791 /* See uma.h */
2792 uma_zone_t
2793 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
2794 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
2795 {
2796 	struct uma_zctor_args args;
2797 	uma_keg_t keg;
2798 	uma_zone_t res;
2799 
2800 	keg = master->uz_keg;
2801 	memset(&args, 0, sizeof(args));
2802 	args.name = name;
2803 	args.size = keg->uk_size;
2804 	args.ctor = ctor;
2805 	args.dtor = dtor;
2806 	args.uminit = zinit;
2807 	args.fini = zfini;
2808 	args.align = keg->uk_align;
2809 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
2810 	args.keg = keg;
2811 
2812 	sx_slock(&uma_reclaim_lock);
2813 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
2814 	sx_sunlock(&uma_reclaim_lock);
2815 
2816 	return (res);
2817 }
2818 
2819 /* See uma.h */
2820 uma_zone_t
2821 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
2822 		    uma_init zinit, uma_fini zfini, uma_import zimport,
2823 		    uma_release zrelease, void *arg, int flags)
2824 {
2825 	struct uma_zctor_args args;
2826 
2827 	memset(&args, 0, sizeof(args));
2828 	args.name = name;
2829 	args.size = size;
2830 	args.ctor = ctor;
2831 	args.dtor = dtor;
2832 	args.uminit = zinit;
2833 	args.fini = zfini;
2834 	args.import = zimport;
2835 	args.release = zrelease;
2836 	args.arg = arg;
2837 	args.align = 0;
2838 	args.flags = flags | UMA_ZFLAG_CACHE;
2839 
2840 	return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
2841 }
2842 
2843 /* See uma.h */
2844 void
2845 uma_zdestroy(uma_zone_t zone)
2846 {
2847 
2848 	/*
2849 	 * Large slabs are expensive to reclaim, so don't bother doing
2850 	 * unnecessary work if we're shutting down.
2851 	 */
2852 	if (booted == BOOT_SHUTDOWN &&
2853 	    zone->uz_fini == NULL && zone->uz_release == zone_release)
2854 		return;
2855 	sx_slock(&uma_reclaim_lock);
2856 	zone_free_item(zones, zone, NULL, SKIP_NONE);
2857 	sx_sunlock(&uma_reclaim_lock);
2858 }
2859 
2860 void
2861 uma_zwait(uma_zone_t zone)
2862 {
2863 	void *item;
2864 
2865 	item = uma_zalloc_arg(zone, NULL, M_WAITOK);
2866 	uma_zfree(zone, item);
2867 }
2868 
2869 void *
2870 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
2871 {
2872 	void *item;
2873 #ifdef SMP
2874 	int i;
2875 
2876 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
2877 #endif
2878 	item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
2879 	if (item != NULL && (flags & M_ZERO)) {
2880 #ifdef SMP
2881 		for (i = 0; i <= mp_maxid; i++)
2882 			bzero(zpcpu_get_cpu(item, i), zone->uz_size);
2883 #else
2884 		bzero(item, zone->uz_size);
2885 #endif
2886 	}
2887 	return (item);
2888 }
2889 
2890 /*
2891  * A stub while both regular and pcpu cases are identical.
2892  */
2893 void
2894 uma_zfree_pcpu_arg(uma_zone_t zone, void *item, void *udata)
2895 {
2896 
2897 #ifdef SMP
2898 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
2899 #endif
2900 	uma_zfree_arg(zone, item, udata);
2901 }
2902 
2903 static inline void *
2904 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
2905     void *item)
2906 {
2907 #ifdef INVARIANTS
2908 	bool skipdbg;
2909 
2910 	skipdbg = uma_dbg_zskip(zone, item);
2911 	if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
2912 	    zone->uz_ctor != trash_ctor)
2913 		trash_ctor(item, size, udata, flags);
2914 #endif
2915 	/* Check flags before loading ctor pointer. */
2916 	if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
2917 	    __predict_false(zone->uz_ctor != NULL) &&
2918 	    zone->uz_ctor(item, size, udata, flags) != 0) {
2919 		counter_u64_add(zone->uz_fails, 1);
2920 		zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
2921 		return (NULL);
2922 	}
2923 #ifdef INVARIANTS
2924 	if (!skipdbg)
2925 		uma_dbg_alloc(zone, NULL, item);
2926 #endif
2927 	if (flags & M_ZERO)
2928 		bzero(item, size);
2929 
2930 	return (item);
2931 }
2932 
2933 static inline void
2934 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
2935     enum zfreeskip skip)
2936 {
2937 #ifdef INVARIANTS
2938 	bool skipdbg;
2939 
2940 	skipdbg = uma_dbg_zskip(zone, item);
2941 	if (skip == SKIP_NONE && !skipdbg) {
2942 		if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
2943 			uma_dbg_free(zone, udata, item);
2944 		else
2945 			uma_dbg_free(zone, NULL, item);
2946 	}
2947 #endif
2948 	if (__predict_true(skip < SKIP_DTOR)) {
2949 		if (zone->uz_dtor != NULL)
2950 			zone->uz_dtor(item, size, udata);
2951 #ifdef INVARIANTS
2952 		if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
2953 		    zone->uz_dtor != trash_dtor)
2954 			trash_dtor(item, size, udata);
2955 #endif
2956 	}
2957 }
2958 
2959 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
2960 #define	UMA_ZALLOC_DEBUG
2961 static int
2962 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
2963 {
2964 	int error;
2965 
2966 	error = 0;
2967 #ifdef WITNESS
2968 	if (flags & M_WAITOK) {
2969 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2970 		    "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
2971 	}
2972 #endif
2973 
2974 #ifdef INVARIANTS
2975 	KASSERT((flags & M_EXEC) == 0,
2976 	    ("uma_zalloc_debug: called with M_EXEC"));
2977 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2978 	    ("uma_zalloc_debug: called within spinlock or critical section"));
2979 	KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
2980 	    ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
2981 #endif
2982 
2983 #ifdef DEBUG_MEMGUARD
2984 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
2985 		void *item;
2986 		item = memguard_alloc(zone->uz_size, flags);
2987 		if (item != NULL) {
2988 			error = EJUSTRETURN;
2989 			if (zone->uz_init != NULL &&
2990 			    zone->uz_init(item, zone->uz_size, flags) != 0) {
2991 				*itemp = NULL;
2992 				return (error);
2993 			}
2994 			if (zone->uz_ctor != NULL &&
2995 			    zone->uz_ctor(item, zone->uz_size, udata,
2996 			    flags) != 0) {
2997 				counter_u64_add(zone->uz_fails, 1);
2998 			    	zone->uz_fini(item, zone->uz_size);
2999 				*itemp = NULL;
3000 				return (error);
3001 			}
3002 			*itemp = item;
3003 			return (error);
3004 		}
3005 		/* This is unfortunate but should not be fatal. */
3006 	}
3007 #endif
3008 	return (error);
3009 }
3010 
3011 static int
3012 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3013 {
3014 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3015 	    ("uma_zfree_debug: called with spinlock or critical section held"));
3016 
3017 #ifdef DEBUG_MEMGUARD
3018 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3019 		if (zone->uz_dtor != NULL)
3020 			zone->uz_dtor(item, zone->uz_size, udata);
3021 		if (zone->uz_fini != NULL)
3022 			zone->uz_fini(item, zone->uz_size);
3023 		memguard_free(item);
3024 		return (EJUSTRETURN);
3025 	}
3026 #endif
3027 	return (0);
3028 }
3029 #endif
3030 
3031 static __noinline void *
3032 uma_zalloc_single(uma_zone_t zone, void *udata, int flags)
3033 {
3034 	int domain;
3035 
3036 	/*
3037 	 * We can not get a bucket so try to return a single item.
3038 	 */
3039 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3040 		domain = PCPU_GET(domain);
3041 	else
3042 		domain = UMA_ANYDOMAIN;
3043 	return (zone_alloc_item(zone, udata, domain, flags));
3044 }
3045 
3046 /* See uma.h */
3047 void *
3048 uma_zalloc_smr(uma_zone_t zone, int flags)
3049 {
3050 	uma_cache_bucket_t bucket;
3051 	uma_cache_t cache;
3052 	void *item;
3053 	int size, uz_flags;
3054 
3055 #ifdef UMA_ZALLOC_DEBUG
3056 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3057 	    ("uma_zalloc_arg: called with non-SMR zone.\n"));
3058 	if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3059 		return (item);
3060 #endif
3061 
3062 	critical_enter();
3063 	do {
3064 		cache = &zone->uz_cpu[curcpu];
3065 		bucket = &cache->uc_allocbucket;
3066 		size = cache_uz_size(cache);
3067 		uz_flags = cache_uz_flags(cache);
3068 		if (__predict_true(bucket->ucb_cnt != 0)) {
3069 			item = cache_bucket_pop(cache, bucket);
3070 			critical_exit();
3071 			return (item_ctor(zone, uz_flags, size, NULL, flags,
3072 			    item));
3073 		}
3074 	} while (cache_alloc(zone, cache, NULL, flags));
3075 	critical_exit();
3076 
3077 	return (uma_zalloc_single(zone, NULL, flags));
3078 }
3079 
3080 /* See uma.h */
3081 void *
3082 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3083 {
3084 	uma_cache_bucket_t bucket;
3085 	uma_cache_t cache;
3086 	void *item;
3087 	int size, uz_flags;
3088 
3089 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3090 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3091 
3092 	/* This is the fast path allocation */
3093 	CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3094 	    zone, flags);
3095 
3096 #ifdef UMA_ZALLOC_DEBUG
3097 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3098 	    ("uma_zalloc_arg: called with SMR zone.\n"));
3099 	if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3100 		return (item);
3101 #endif
3102 
3103 	/*
3104 	 * If possible, allocate from the per-CPU cache.  There are two
3105 	 * requirements for safe access to the per-CPU cache: (1) the thread
3106 	 * accessing the cache must not be preempted or yield during access,
3107 	 * and (2) the thread must not migrate CPUs without switching which
3108 	 * cache it accesses.  We rely on a critical section to prevent
3109 	 * preemption and migration.  We release the critical section in
3110 	 * order to acquire the zone mutex if we are unable to allocate from
3111 	 * the current cache; when we re-acquire the critical section, we
3112 	 * must detect and handle migration if it has occurred.
3113 	 */
3114 	critical_enter();
3115 	do {
3116 		cache = &zone->uz_cpu[curcpu];
3117 		bucket = &cache->uc_allocbucket;
3118 		size = cache_uz_size(cache);
3119 		uz_flags = cache_uz_flags(cache);
3120 		if (__predict_true(bucket->ucb_cnt != 0)) {
3121 			item = cache_bucket_pop(cache, bucket);
3122 			critical_exit();
3123 			return (item_ctor(zone, uz_flags, size, udata, flags,
3124 			    item));
3125 		}
3126 	} while (cache_alloc(zone, cache, udata, flags));
3127 	critical_exit();
3128 
3129 	return (uma_zalloc_single(zone, udata, flags));
3130 }
3131 
3132 /*
3133  * Replenish an alloc bucket and possibly restore an old one.  Called in
3134  * a critical section.  Returns in a critical section.
3135  *
3136  * A false return value indicates an allocation failure.
3137  * A true return value indicates success and the caller should retry.
3138  */
3139 static __noinline bool
3140 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3141 {
3142 	uma_zone_domain_t zdom;
3143 	uma_bucket_t bucket;
3144 	int domain;
3145 	bool lockfail;
3146 
3147 	CRITICAL_ASSERT(curthread);
3148 
3149 	/*
3150 	 * If we have run out of items in our alloc bucket see
3151 	 * if we can switch with the free bucket.
3152 	 *
3153 	 * SMR Zones can't re-use the free bucket until the sequence has
3154 	 * expired.
3155 	 */
3156 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 &&
3157 	    cache->uc_freebucket.ucb_cnt != 0) {
3158 		cache_bucket_swap(&cache->uc_freebucket,
3159 		    &cache->uc_allocbucket);
3160 		return (true);
3161 	}
3162 
3163 	/*
3164 	 * Discard any empty allocation bucket while we hold no locks.
3165 	 */
3166 	bucket = cache_bucket_unload_alloc(cache);
3167 	critical_exit();
3168 	if (bucket != NULL)
3169 		bucket_free(zone, bucket, udata);
3170 
3171 	/* Short-circuit for zones without buckets and low memory. */
3172 	if (zone->uz_bucket_size == 0 || bucketdisable) {
3173 		critical_enter();
3174 		return (false);
3175 	}
3176 
3177 	/*
3178 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
3179 	 * we must go back to the zone.  This requires the zone lock, so we
3180 	 * must drop the critical section, then re-acquire it when we go back
3181 	 * to the cache.  Since the critical section is released, we may be
3182 	 * preempted or migrate.  As such, make sure not to maintain any
3183 	 * thread-local state specific to the cache from prior to releasing
3184 	 * the critical section.
3185 	 */
3186 	lockfail = 0;
3187 	if (ZONE_TRYLOCK(zone) == 0) {
3188 		/* Record contention to size the buckets. */
3189 		ZONE_LOCK(zone);
3190 		lockfail = 1;
3191 	}
3192 
3193 	/* See if we lost the race to fill the cache. */
3194 	critical_enter();
3195 	cache = &zone->uz_cpu[curcpu];
3196 	if (cache->uc_allocbucket.ucb_bucket != NULL) {
3197 		ZONE_UNLOCK(zone);
3198 		return (true);
3199 	}
3200 
3201 	/*
3202 	 * Check the zone's cache of buckets.
3203 	 */
3204 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) {
3205 		domain = PCPU_GET(domain);
3206 		zdom = &zone->uz_domain[domain];
3207 	} else {
3208 		domain = UMA_ANYDOMAIN;
3209 		zdom = &zone->uz_domain[0];
3210 	}
3211 
3212 	if ((bucket = zone_fetch_bucket(zone, zdom)) != NULL) {
3213 		KASSERT(bucket->ub_cnt != 0,
3214 		    ("uma_zalloc_arg: Returning an empty bucket."));
3215 		cache_bucket_load_alloc(cache, bucket);
3216 		return (true);
3217 	}
3218 	/* We are no longer associated with this CPU. */
3219 	critical_exit();
3220 
3221 	/*
3222 	 * We bump the uz count when the cache size is insufficient to
3223 	 * handle the working set.
3224 	 */
3225 	if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
3226 		zone->uz_bucket_size++;
3227 	ZONE_UNLOCK(zone);
3228 
3229 	/*
3230 	 * Fill a bucket and attempt to use it as the alloc bucket.
3231 	 */
3232 	bucket = zone_alloc_bucket(zone, udata, domain, flags);
3233 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3234 	    zone->uz_name, zone, bucket);
3235 	if (bucket == NULL) {
3236 		critical_enter();
3237 		return (false);
3238 	}
3239 
3240 	/*
3241 	 * See if we lost the race or were migrated.  Cache the
3242 	 * initialized bucket to make this less likely or claim
3243 	 * the memory directly.
3244 	 */
3245 	ZONE_LOCK(zone);
3246 	critical_enter();
3247 	cache = &zone->uz_cpu[curcpu];
3248 	if (cache->uc_allocbucket.ucb_bucket == NULL &&
3249 	    ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0 ||
3250 	    domain == PCPU_GET(domain))) {
3251 		cache_bucket_load_alloc(cache, bucket);
3252 		zdom->uzd_imax += bucket->ub_cnt;
3253 	} else if (zone->uz_bkt_count >= zone->uz_bkt_max) {
3254 		critical_exit();
3255 		ZONE_UNLOCK(zone);
3256 		bucket_drain(zone, bucket);
3257 		bucket_free(zone, bucket, udata);
3258 		critical_enter();
3259 		return (true);
3260 	} else
3261 		zone_put_bucket(zone, zdom, bucket, false);
3262 	ZONE_UNLOCK(zone);
3263 	return (true);
3264 }
3265 
3266 void *
3267 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3268 {
3269 
3270 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3271 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3272 
3273 	/* This is the fast path allocation */
3274 	CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3275 	    zone->uz_name, zone, domain, flags);
3276 
3277 	if (flags & M_WAITOK) {
3278 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3279 		    "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3280 	}
3281 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3282 	    ("uma_zalloc_domain: called with spinlock or critical section held"));
3283 
3284 	return (zone_alloc_item(zone, udata, domain, flags));
3285 }
3286 
3287 /*
3288  * Find a slab with some space.  Prefer slabs that are partially used over those
3289  * that are totally full.  This helps to reduce fragmentation.
3290  *
3291  * If 'rr' is 1, search all domains starting from 'domain'.  Otherwise check
3292  * only 'domain'.
3293  */
3294 static uma_slab_t
3295 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3296 {
3297 	uma_domain_t dom;
3298 	uma_slab_t slab;
3299 	int start;
3300 
3301 	KASSERT(domain >= 0 && domain < vm_ndomains,
3302 	    ("keg_first_slab: domain %d out of range", domain));
3303 	KEG_LOCK_ASSERT(keg, domain);
3304 
3305 	slab = NULL;
3306 	start = domain;
3307 	do {
3308 		dom = &keg->uk_domain[domain];
3309 		if (!LIST_EMPTY(&dom->ud_part_slab))
3310 			return (LIST_FIRST(&dom->ud_part_slab));
3311 		if (!LIST_EMPTY(&dom->ud_free_slab)) {
3312 			slab = LIST_FIRST(&dom->ud_free_slab);
3313 			LIST_REMOVE(slab, us_link);
3314 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3315 			return (slab);
3316 		}
3317 		if (rr)
3318 			domain = (domain + 1) % vm_ndomains;
3319 	} while (domain != start);
3320 
3321 	return (NULL);
3322 }
3323 
3324 /*
3325  * Fetch an existing slab from a free or partial list.  Returns with the
3326  * keg domain lock held if a slab was found or unlocked if not.
3327  */
3328 static uma_slab_t
3329 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3330 {
3331 	uma_slab_t slab;
3332 	uint32_t reserve;
3333 
3334 	/* HASH has a single free list. */
3335 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3336 		domain = 0;
3337 
3338 	KEG_LOCK(keg, domain);
3339 	reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3340 	if (keg->uk_domain[domain].ud_free <= reserve ||
3341 	    (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3342 		KEG_UNLOCK(keg, domain);
3343 		return (NULL);
3344 	}
3345 	return (slab);
3346 }
3347 
3348 static uma_slab_t
3349 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3350 {
3351 	struct vm_domainset_iter di;
3352 	uma_slab_t slab;
3353 	int aflags, domain;
3354 	bool rr;
3355 
3356 restart:
3357 	/*
3358 	 * Use the keg's policy if upper layers haven't already specified a
3359 	 * domain (as happens with first-touch zones).
3360 	 *
3361 	 * To avoid races we run the iterator with the keg lock held, but that
3362 	 * means that we cannot allow the vm_domainset layer to sleep.  Thus,
3363 	 * clear M_WAITOK and handle low memory conditions locally.
3364 	 */
3365 	rr = rdomain == UMA_ANYDOMAIN;
3366 	if (rr) {
3367 		aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3368 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3369 		    &aflags);
3370 	} else {
3371 		aflags = flags;
3372 		domain = rdomain;
3373 	}
3374 
3375 	for (;;) {
3376 		slab = keg_fetch_free_slab(keg, domain, rr, flags);
3377 		if (slab != NULL)
3378 			return (slab);
3379 
3380 		/*
3381 		 * M_NOVM means don't ask at all!
3382 		 */
3383 		if (flags & M_NOVM)
3384 			break;
3385 
3386 		slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3387 		if (slab != NULL)
3388 			return (slab);
3389 		if (!rr && (flags & M_WAITOK) == 0)
3390 			break;
3391 		if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3392 			if ((flags & M_WAITOK) != 0) {
3393 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
3394 				goto restart;
3395 			}
3396 			break;
3397 		}
3398 	}
3399 
3400 	/*
3401 	 * We might not have been able to get a slab but another cpu
3402 	 * could have while we were unlocked.  Check again before we
3403 	 * fail.
3404 	 */
3405 	if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3406 		return (slab);
3407 
3408 	return (NULL);
3409 }
3410 
3411 static void *
3412 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3413 {
3414 	uma_domain_t dom;
3415 	void *item;
3416 	int freei;
3417 
3418 	KEG_LOCK_ASSERT(keg, slab->us_domain);
3419 
3420 	dom = &keg->uk_domain[slab->us_domain];
3421 	freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3422 	BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3423 	item = slab_item(slab, keg, freei);
3424 	slab->us_freecount--;
3425 	dom->ud_free--;
3426 
3427 	/* Move this slab to the full list */
3428 	if (slab->us_freecount == 0) {
3429 		LIST_REMOVE(slab, us_link);
3430 		LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3431 	}
3432 
3433 	return (item);
3434 }
3435 
3436 static int
3437 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3438 {
3439 	uma_domain_t dom;
3440 	uma_zone_t zone;
3441 	uma_slab_t slab;
3442 	uma_keg_t keg;
3443 #ifdef NUMA
3444 	int stripe;
3445 #endif
3446 	int i;
3447 
3448 	zone = arg;
3449 	slab = NULL;
3450 	keg = zone->uz_keg;
3451 	/* Try to keep the buckets totally full */
3452 	for (i = 0; i < max; ) {
3453 		if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3454 			break;
3455 #ifdef NUMA
3456 		stripe = howmany(max, vm_ndomains);
3457 #endif
3458 		dom = &keg->uk_domain[slab->us_domain];
3459 		while (slab->us_freecount && i < max) {
3460 			bucket[i++] = slab_alloc_item(keg, slab);
3461 			if (dom->ud_free <= keg->uk_reserve)
3462 				break;
3463 #ifdef NUMA
3464 			/*
3465 			 * If the zone is striped we pick a new slab for every
3466 			 * N allocations.  Eliminating this conditional will
3467 			 * instead pick a new domain for each bucket rather
3468 			 * than stripe within each bucket.  The current option
3469 			 * produces more fragmentation and requires more cpu
3470 			 * time but yields better distribution.
3471 			 */
3472 			if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3473 			    vm_ndomains > 1 && --stripe == 0)
3474 				break;
3475 #endif
3476 		}
3477 		KEG_UNLOCK(keg, slab->us_domain);
3478 		/* Don't block if we allocated any successfully. */
3479 		flags &= ~M_WAITOK;
3480 		flags |= M_NOWAIT;
3481 	}
3482 
3483 	return i;
3484 }
3485 
3486 static int
3487 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3488 {
3489 	uint64_t old, new, total, max;
3490 
3491 	/*
3492 	 * The hard case.  We're going to sleep because there were existing
3493 	 * sleepers or because we ran out of items.  This routine enforces
3494 	 * fairness by keeping fifo order.
3495 	 *
3496 	 * First release our ill gotten gains and make some noise.
3497 	 */
3498 	for (;;) {
3499 		zone_free_limit(zone, count);
3500 		zone_log_warning(zone);
3501 		zone_maxaction(zone);
3502 		if (flags & M_NOWAIT)
3503 			return (0);
3504 
3505 		/*
3506 		 * We need to allocate an item or set ourself as a sleeper
3507 		 * while the sleepq lock is held to avoid wakeup races.  This
3508 		 * is essentially a home rolled semaphore.
3509 		 */
3510 		sleepq_lock(&zone->uz_max_items);
3511 		old = zone->uz_items;
3512 		do {
3513 			MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3514 			/* Cache the max since we will evaluate twice. */
3515 			max = zone->uz_max_items;
3516 			if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3517 			    UZ_ITEMS_COUNT(old) >= max)
3518 				new = old + UZ_ITEMS_SLEEPER;
3519 			else
3520 				new = old + MIN(count, max - old);
3521 		} while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3522 
3523 		/* We may have successfully allocated under the sleepq lock. */
3524 		if (UZ_ITEMS_SLEEPERS(new) == 0) {
3525 			sleepq_release(&zone->uz_max_items);
3526 			return (new - old);
3527 		}
3528 
3529 		/*
3530 		 * This is in a different cacheline from uz_items so that we
3531 		 * don't constantly invalidate the fastpath cacheline when we
3532 		 * adjust item counts.  This could be limited to toggling on
3533 		 * transitions.
3534 		 */
3535 		atomic_add_32(&zone->uz_sleepers, 1);
3536 		atomic_add_64(&zone->uz_sleeps, 1);
3537 
3538 		/*
3539 		 * We have added ourselves as a sleeper.  The sleepq lock
3540 		 * protects us from wakeup races.  Sleep now and then retry.
3541 		 */
3542 		sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3543 		sleepq_wait(&zone->uz_max_items, PVM);
3544 
3545 		/*
3546 		 * After wakeup, remove ourselves as a sleeper and try
3547 		 * again.  We no longer have the sleepq lock for protection.
3548 		 *
3549 		 * Subract ourselves as a sleeper while attempting to add
3550 		 * our count.
3551 		 */
3552 		atomic_subtract_32(&zone->uz_sleepers, 1);
3553 		old = atomic_fetchadd_64(&zone->uz_items,
3554 		    -(UZ_ITEMS_SLEEPER - count));
3555 		/* We're no longer a sleeper. */
3556 		old -= UZ_ITEMS_SLEEPER;
3557 
3558 		/*
3559 		 * If we're still at the limit, restart.  Notably do not
3560 		 * block on other sleepers.  Cache the max value to protect
3561 		 * against changes via sysctl.
3562 		 */
3563 		total = UZ_ITEMS_COUNT(old);
3564 		max = zone->uz_max_items;
3565 		if (total >= max)
3566 			continue;
3567 		/* Truncate if necessary, otherwise wake other sleepers. */
3568 		if (total + count > max) {
3569 			zone_free_limit(zone, total + count - max);
3570 			count = max - total;
3571 		} else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3572 			wakeup_one(&zone->uz_max_items);
3573 
3574 		return (count);
3575 	}
3576 }
3577 
3578 /*
3579  * Allocate 'count' items from our max_items limit.  Returns the number
3580  * available.  If M_NOWAIT is not specified it will sleep until at least
3581  * one item can be allocated.
3582  */
3583 static int
3584 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3585 {
3586 	uint64_t old;
3587 	uint64_t max;
3588 
3589 	max = zone->uz_max_items;
3590 	MPASS(max > 0);
3591 
3592 	/*
3593 	 * We expect normal allocations to succeed with a simple
3594 	 * fetchadd.
3595 	 */
3596 	old = atomic_fetchadd_64(&zone->uz_items, count);
3597 	if (__predict_true(old + count <= max))
3598 		return (count);
3599 
3600 	/*
3601 	 * If we had some items and no sleepers just return the
3602 	 * truncated value.  We have to release the excess space
3603 	 * though because that may wake sleepers who weren't woken
3604 	 * because we were temporarily over the limit.
3605 	 */
3606 	if (old < max) {
3607 		zone_free_limit(zone, (old + count) - max);
3608 		return (max - old);
3609 	}
3610 	return (zone_alloc_limit_hard(zone, count, flags));
3611 }
3612 
3613 /*
3614  * Free a number of items back to the limit.
3615  */
3616 static void
3617 zone_free_limit(uma_zone_t zone, int count)
3618 {
3619 	uint64_t old;
3620 
3621 	MPASS(count > 0);
3622 
3623 	/*
3624 	 * In the common case we either have no sleepers or
3625 	 * are still over the limit and can just return.
3626 	 */
3627 	old = atomic_fetchadd_64(&zone->uz_items, -count);
3628 	if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3629 	   UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3630 		return;
3631 
3632 	/*
3633 	 * Moderate the rate of wakeups.  Sleepers will continue
3634 	 * to generate wakeups if necessary.
3635 	 */
3636 	wakeup_one(&zone->uz_max_items);
3637 }
3638 
3639 static uma_bucket_t
3640 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3641 {
3642 	uma_bucket_t bucket;
3643 	int maxbucket, cnt;
3644 
3645 	CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3646 	    zone, domain);
3647 
3648 	/* Avoid allocs targeting empty domains. */
3649 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3650 		domain = UMA_ANYDOMAIN;
3651 
3652 	if (zone->uz_max_items > 0)
3653 		maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3654 		    M_NOWAIT);
3655 	else
3656 		maxbucket = zone->uz_bucket_size;
3657 	if (maxbucket == 0)
3658 		return (false);
3659 
3660 	/* Don't wait for buckets, preserve caller's NOVM setting. */
3661 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3662 	if (bucket == NULL) {
3663 		cnt = 0;
3664 		goto out;
3665 	}
3666 
3667 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3668 	    MIN(maxbucket, bucket->ub_entries), domain, flags);
3669 
3670 	/*
3671 	 * Initialize the memory if necessary.
3672 	 */
3673 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3674 		int i;
3675 
3676 		for (i = 0; i < bucket->ub_cnt; i++)
3677 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3678 			    flags) != 0)
3679 				break;
3680 		/*
3681 		 * If we couldn't initialize the whole bucket, put the
3682 		 * rest back onto the freelist.
3683 		 */
3684 		if (i != bucket->ub_cnt) {
3685 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3686 			    bucket->ub_cnt - i);
3687 #ifdef INVARIANTS
3688 			bzero(&bucket->ub_bucket[i],
3689 			    sizeof(void *) * (bucket->ub_cnt - i));
3690 #endif
3691 			bucket->ub_cnt = i;
3692 		}
3693 	}
3694 
3695 	cnt = bucket->ub_cnt;
3696 	if (bucket->ub_cnt == 0) {
3697 		bucket_free(zone, bucket, udata);
3698 		counter_u64_add(zone->uz_fails, 1);
3699 		bucket = NULL;
3700 	}
3701 out:
3702 	if (zone->uz_max_items > 0 && cnt < maxbucket)
3703 		zone_free_limit(zone, maxbucket - cnt);
3704 
3705 	return (bucket);
3706 }
3707 
3708 /*
3709  * Allocates a single item from a zone.
3710  *
3711  * Arguments
3712  *	zone   The zone to alloc for.
3713  *	udata  The data to be passed to the constructor.
3714  *	domain The domain to allocate from or UMA_ANYDOMAIN.
3715  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
3716  *
3717  * Returns
3718  *	NULL if there is no memory and M_NOWAIT is set
3719  *	An item if successful
3720  */
3721 
3722 static void *
3723 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
3724 {
3725 	void *item;
3726 
3727 	if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0)
3728 		return (NULL);
3729 
3730 	/* Avoid allocs targeting empty domains. */
3731 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3732 		domain = UMA_ANYDOMAIN;
3733 
3734 	if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
3735 		goto fail_cnt;
3736 
3737 	/*
3738 	 * We have to call both the zone's init (not the keg's init)
3739 	 * and the zone's ctor.  This is because the item is going from
3740 	 * a keg slab directly to the user, and the user is expecting it
3741 	 * to be both zone-init'd as well as zone-ctor'd.
3742 	 */
3743 	if (zone->uz_init != NULL) {
3744 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
3745 			zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
3746 			goto fail_cnt;
3747 		}
3748 	}
3749 	item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
3750 	    item);
3751 	if (item == NULL)
3752 		goto fail;
3753 
3754 	counter_u64_add(zone->uz_allocs, 1);
3755 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
3756 	    zone->uz_name, zone);
3757 
3758 	return (item);
3759 
3760 fail_cnt:
3761 	counter_u64_add(zone->uz_fails, 1);
3762 fail:
3763 	if (zone->uz_max_items > 0)
3764 		zone_free_limit(zone, 1);
3765 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
3766 	    zone->uz_name, zone);
3767 
3768 	return (NULL);
3769 }
3770 
3771 /* See uma.h */
3772 void
3773 uma_zfree_smr(uma_zone_t zone, void *item)
3774 {
3775 	uma_cache_t cache;
3776 	uma_cache_bucket_t bucket;
3777 	int domain, itemdomain, uz_flags;
3778 
3779 #ifdef UMA_ZALLOC_DEBUG
3780 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3781 	    ("uma_zfree_smr: called with non-SMR zone.\n"));
3782 	KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
3783 	if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
3784 		return;
3785 #endif
3786 	cache = &zone->uz_cpu[curcpu];
3787 	uz_flags = cache_uz_flags(cache);
3788 	domain = itemdomain = 0;
3789 #ifdef NUMA
3790 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
3791 		itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
3792 #endif
3793 	critical_enter();
3794 	do {
3795 		cache = &zone->uz_cpu[curcpu];
3796 		/* SMR Zones must free to the free bucket. */
3797 		bucket = &cache->uc_freebucket;
3798 #ifdef NUMA
3799 		domain = PCPU_GET(domain);
3800 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
3801 		    domain != itemdomain) {
3802 			bucket = &cache->uc_crossbucket;
3803 		}
3804 #endif
3805 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
3806 			cache_bucket_push(cache, bucket, item);
3807 			critical_exit();
3808 			return;
3809 		}
3810 	} while (cache_free(zone, cache, NULL, item, itemdomain));
3811 	critical_exit();
3812 
3813 	/*
3814 	 * If nothing else caught this, we'll just do an internal free.
3815 	 */
3816 	zone_free_item(zone, item, NULL, SKIP_NONE);
3817 }
3818 
3819 /* See uma.h */
3820 void
3821 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
3822 {
3823 	uma_cache_t cache;
3824 	uma_cache_bucket_t bucket;
3825 	int domain, itemdomain, uz_flags;
3826 
3827 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3828 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3829 
3830 	CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
3831 
3832 #ifdef UMA_ZALLOC_DEBUG
3833 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3834 	    ("uma_zfree_arg: called with SMR zone.\n"));
3835 	if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
3836 		return;
3837 #endif
3838         /* uma_zfree(..., NULL) does nothing, to match free(9). */
3839         if (item == NULL)
3840                 return;
3841 
3842 	/*
3843 	 * We are accessing the per-cpu cache without a critical section to
3844 	 * fetch size and flags.  This is acceptable, if we are preempted we
3845 	 * will simply read another cpu's line.
3846 	 */
3847 	cache = &zone->uz_cpu[curcpu];
3848 	uz_flags = cache_uz_flags(cache);
3849 	if (UMA_ALWAYS_CTORDTOR ||
3850 	    __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
3851 		item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
3852 
3853 	/*
3854 	 * The race here is acceptable.  If we miss it we'll just have to wait
3855 	 * a little longer for the limits to be reset.
3856 	 */
3857 	if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
3858 		if (zone->uz_sleepers > 0)
3859 			goto zfree_item;
3860 	}
3861 
3862 	/*
3863 	 * If possible, free to the per-CPU cache.  There are two
3864 	 * requirements for safe access to the per-CPU cache: (1) the thread
3865 	 * accessing the cache must not be preempted or yield during access,
3866 	 * and (2) the thread must not migrate CPUs without switching which
3867 	 * cache it accesses.  We rely on a critical section to prevent
3868 	 * preemption and migration.  We release the critical section in
3869 	 * order to acquire the zone mutex if we are unable to free to the
3870 	 * current cache; when we re-acquire the critical section, we must
3871 	 * detect and handle migration if it has occurred.
3872 	 */
3873 	domain = itemdomain = 0;
3874 #ifdef NUMA
3875 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
3876 		itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
3877 #endif
3878 	critical_enter();
3879 	do {
3880 		cache = &zone->uz_cpu[curcpu];
3881 		/*
3882 		 * Try to free into the allocbucket first to give LIFO
3883 		 * ordering for cache-hot datastructures.  Spill over
3884 		 * into the freebucket if necessary.  Alloc will swap
3885 		 * them if one runs dry.
3886 		 */
3887 		bucket = &cache->uc_allocbucket;
3888 #ifdef NUMA
3889 		domain = PCPU_GET(domain);
3890 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
3891 		    domain != itemdomain) {
3892 			bucket = &cache->uc_crossbucket;
3893 		} else
3894 #endif
3895 		if (bucket->ucb_cnt >= bucket->ucb_entries)
3896 			bucket = &cache->uc_freebucket;
3897 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
3898 			cache_bucket_push(cache, bucket, item);
3899 			critical_exit();
3900 			return;
3901 		}
3902 	} while (cache_free(zone, cache, udata, item, itemdomain));
3903 	critical_exit();
3904 
3905 	/*
3906 	 * If nothing else caught this, we'll just do an internal free.
3907 	 */
3908 zfree_item:
3909 	zone_free_item(zone, item, udata, SKIP_DTOR);
3910 }
3911 
3912 #ifdef NUMA
3913 /*
3914  * sort crossdomain free buckets to domain correct buckets and cache
3915  * them.
3916  */
3917 static void
3918 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
3919 {
3920 	struct uma_bucketlist fullbuckets;
3921 	uma_zone_domain_t zdom;
3922 	uma_bucket_t b;
3923 	void *item;
3924 	int domain;
3925 
3926 	CTR3(KTR_UMA,
3927 	    "uma_zfree: zone %s(%p) draining cross bucket %p",
3928 	    zone->uz_name, zone, bucket);
3929 
3930 	STAILQ_INIT(&fullbuckets);
3931 
3932 	/*
3933 	 * To avoid having ndomain * ndomain buckets for sorting we have a
3934 	 * lock on the current crossfree bucket.  A full matrix with
3935 	 * per-domain locking could be used if necessary.
3936 	 */
3937 	ZONE_CROSS_LOCK(zone);
3938 	while (bucket->ub_cnt > 0) {
3939 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
3940 		domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
3941 		zdom = &zone->uz_domain[domain];
3942 		if (zdom->uzd_cross == NULL) {
3943 			zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
3944 			if (zdom->uzd_cross == NULL)
3945 				break;
3946 		}
3947 		zdom->uzd_cross->ub_bucket[zdom->uzd_cross->ub_cnt++] = item;
3948 		if (zdom->uzd_cross->ub_cnt == zdom->uzd_cross->ub_entries) {
3949 			STAILQ_INSERT_HEAD(&fullbuckets, zdom->uzd_cross,
3950 			    ub_link);
3951 			zdom->uzd_cross = NULL;
3952 		}
3953 		bucket->ub_cnt--;
3954 	}
3955 	ZONE_CROSS_UNLOCK(zone);
3956 	if (!STAILQ_EMPTY(&fullbuckets)) {
3957 		ZONE_LOCK(zone);
3958 		while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
3959 			if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3960 				bucket->ub_seq = smr_current(zone->uz_smr);
3961 			STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
3962 			if (zone->uz_bkt_count >= zone->uz_bkt_max) {
3963 				ZONE_UNLOCK(zone);
3964 				bucket_drain(zone, b);
3965 				bucket_free(zone, b, udata);
3966 				ZONE_LOCK(zone);
3967 			} else {
3968 				domain = _vm_phys_domain(
3969 				    pmap_kextract(
3970 				    (vm_offset_t)b->ub_bucket[0]));
3971 				zdom = &zone->uz_domain[domain];
3972 				zone_put_bucket(zone, zdom, b, true);
3973 			}
3974 		}
3975 		ZONE_UNLOCK(zone);
3976 	}
3977 	if (bucket->ub_cnt != 0)
3978 		bucket_drain(zone, bucket);
3979 	bucket->ub_seq = SMR_SEQ_INVALID;
3980 	bucket_free(zone, bucket, udata);
3981 }
3982 #endif
3983 
3984 static void
3985 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
3986     int domain, int itemdomain)
3987 {
3988 	uma_zone_domain_t zdom;
3989 
3990 #ifdef NUMA
3991 	/*
3992 	 * Buckets coming from the wrong domain will be entirely for the
3993 	 * only other domain on two domain systems.  In this case we can
3994 	 * simply cache them.  Otherwise we need to sort them back to
3995 	 * correct domains.
3996 	 */
3997 	if (domain != itemdomain && vm_ndomains > 2) {
3998 		zone_free_cross(zone, bucket, udata);
3999 		return;
4000 	}
4001 #endif
4002 
4003 	/*
4004 	 * Attempt to save the bucket in the zone's domain bucket cache.
4005 	 *
4006 	 * We bump the uz count when the cache size is insufficient to
4007 	 * handle the working set.
4008 	 */
4009 	if (ZONE_TRYLOCK(zone) == 0) {
4010 		/* Record contention to size the buckets. */
4011 		ZONE_LOCK(zone);
4012 		if (zone->uz_bucket_size < zone->uz_bucket_size_max)
4013 			zone->uz_bucket_size++;
4014 	}
4015 
4016 	CTR3(KTR_UMA,
4017 	    "uma_zfree: zone %s(%p) putting bucket %p on free list",
4018 	    zone->uz_name, zone, bucket);
4019 	/* ub_cnt is pointing to the last free item */
4020 	KASSERT(bucket->ub_cnt == bucket->ub_entries,
4021 	    ("uma_zfree: Attempting to insert partial  bucket onto the full list.\n"));
4022 	if (zone->uz_bkt_count >= zone->uz_bkt_max) {
4023 		ZONE_UNLOCK(zone);
4024 		bucket_drain(zone, bucket);
4025 		bucket_free(zone, bucket, udata);
4026 	} else {
4027 		zdom = &zone->uz_domain[itemdomain];
4028 		zone_put_bucket(zone, zdom, bucket, true);
4029 		ZONE_UNLOCK(zone);
4030 	}
4031 }
4032 
4033 /*
4034  * Populate a free or cross bucket for the current cpu cache.  Free any
4035  * existing full bucket either to the zone cache or back to the slab layer.
4036  *
4037  * Enters and returns in a critical section.  false return indicates that
4038  * we can not satisfy this free in the cache layer.  true indicates that
4039  * the caller should retry.
4040  */
4041 static __noinline bool
4042 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4043     int itemdomain)
4044 {
4045 	uma_cache_bucket_t cbucket;
4046 	uma_bucket_t newbucket, bucket;
4047 	int domain;
4048 
4049 	CRITICAL_ASSERT(curthread);
4050 
4051 	if (zone->uz_bucket_size == 0)
4052 		return false;
4053 
4054 	cache = &zone->uz_cpu[curcpu];
4055 	newbucket = NULL;
4056 
4057 	/*
4058 	 * FIRSTTOUCH domains need to free to the correct zdom.  When
4059 	 * enabled this is the zdom of the item.   The bucket is the
4060 	 * cross bucket if the current domain and itemdomain do not match.
4061 	 */
4062 	cbucket = &cache->uc_freebucket;
4063 #ifdef NUMA
4064 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) {
4065 		domain = PCPU_GET(domain);
4066 		if (domain != itemdomain) {
4067 			cbucket = &cache->uc_crossbucket;
4068 			if (cbucket->ucb_cnt != 0)
4069 				atomic_add_64(&zone->uz_xdomain,
4070 				    cbucket->ucb_cnt);
4071 		}
4072 	} else
4073 #endif
4074 		itemdomain = domain = 0;
4075 	bucket = cache_bucket_unload(cbucket);
4076 
4077 	/* We are no longer associated with this CPU. */
4078 	critical_exit();
4079 
4080 	/*
4081 	 * Don't let SMR zones operate without a free bucket.  Force
4082 	 * a synchronize and re-use this one.  We will only degrade
4083 	 * to a synchronize every bucket_size items rather than every
4084 	 * item if we fail to allocate a bucket.
4085 	 */
4086 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4087 		if (bucket != NULL)
4088 			bucket->ub_seq = smr_advance(zone->uz_smr);
4089 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4090 		if (newbucket == NULL && bucket != NULL) {
4091 			bucket_drain(zone, bucket);
4092 			newbucket = bucket;
4093 			bucket = NULL;
4094 		}
4095 	} else if (!bucketdisable)
4096 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4097 
4098 	if (bucket != NULL)
4099 		zone_free_bucket(zone, bucket, udata, domain, itemdomain);
4100 
4101 	critical_enter();
4102 	if ((bucket = newbucket) == NULL)
4103 		return (false);
4104 	cache = &zone->uz_cpu[curcpu];
4105 #ifdef NUMA
4106 	/*
4107 	 * Check to see if we should be populating the cross bucket.  If it
4108 	 * is already populated we will fall through and attempt to populate
4109 	 * the free bucket.
4110 	 */
4111 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) {
4112 		domain = PCPU_GET(domain);
4113 		if (domain != itemdomain &&
4114 		    cache->uc_crossbucket.ucb_bucket == NULL) {
4115 			cache_bucket_load_cross(cache, bucket);
4116 			return (true);
4117 		}
4118 	}
4119 #endif
4120 	/*
4121 	 * We may have lost the race to fill the bucket or switched CPUs.
4122 	 */
4123 	if (cache->uc_freebucket.ucb_bucket != NULL) {
4124 		critical_exit();
4125 		bucket_free(zone, bucket, udata);
4126 		critical_enter();
4127 	} else
4128 		cache_bucket_load_free(cache, bucket);
4129 
4130 	return (true);
4131 }
4132 
4133 void
4134 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
4135 {
4136 
4137 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4138 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4139 
4140 	CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone);
4141 
4142 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
4143 	    ("uma_zfree_domain: called with spinlock or critical section held"));
4144 
4145         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4146         if (item == NULL)
4147                 return;
4148 	zone_free_item(zone, item, udata, SKIP_NONE);
4149 }
4150 
4151 static void
4152 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4153 {
4154 	uma_keg_t keg;
4155 	uma_domain_t dom;
4156 	int freei;
4157 
4158 	keg = zone->uz_keg;
4159 	KEG_LOCK_ASSERT(keg, slab->us_domain);
4160 
4161 	/* Do we need to remove from any lists? */
4162 	dom = &keg->uk_domain[slab->us_domain];
4163 	if (slab->us_freecount+1 == keg->uk_ipers) {
4164 		LIST_REMOVE(slab, us_link);
4165 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4166 	} else if (slab->us_freecount == 0) {
4167 		LIST_REMOVE(slab, us_link);
4168 		LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4169 	}
4170 
4171 	/* Slab management. */
4172 	freei = slab_item_index(slab, keg, item);
4173 	BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4174 	slab->us_freecount++;
4175 
4176 	/* Keg statistics. */
4177 	dom->ud_free++;
4178 }
4179 
4180 static void
4181 zone_release(void *arg, void **bucket, int cnt)
4182 {
4183 	struct mtx *lock;
4184 	uma_zone_t zone;
4185 	uma_slab_t slab;
4186 	uma_keg_t keg;
4187 	uint8_t *mem;
4188 	void *item;
4189 	int i;
4190 
4191 	zone = arg;
4192 	keg = zone->uz_keg;
4193 	lock = NULL;
4194 	if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4195 		lock = KEG_LOCK(keg, 0);
4196 	for (i = 0; i < cnt; i++) {
4197 		item = bucket[i];
4198 		if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4199 			slab = vtoslab((vm_offset_t)item);
4200 		} else {
4201 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4202 			if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4203 				slab = hash_sfind(&keg->uk_hash, mem);
4204 			else
4205 				slab = (uma_slab_t)(mem + keg->uk_pgoff);
4206 		}
4207 		if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4208 			if (lock != NULL)
4209 				mtx_unlock(lock);
4210 			lock = KEG_LOCK(keg, slab->us_domain);
4211 		}
4212 		slab_free_item(zone, slab, item);
4213 	}
4214 	if (lock != NULL)
4215 		mtx_unlock(lock);
4216 }
4217 
4218 /*
4219  * Frees a single item to any zone.
4220  *
4221  * Arguments:
4222  *	zone   The zone to free to
4223  *	item   The item we're freeing
4224  *	udata  User supplied data for the dtor
4225  *	skip   Skip dtors and finis
4226  */
4227 static void
4228 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4229 {
4230 
4231 	/*
4232 	 * If a free is sent directly to an SMR zone we have to
4233 	 * synchronize immediately because the item can instantly
4234 	 * be reallocated. This should only happen in degenerate
4235 	 * cases when no memory is available for per-cpu caches.
4236 	 */
4237 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4238 		smr_synchronize(zone->uz_smr);
4239 
4240 	item_dtor(zone, item, zone->uz_size, udata, skip);
4241 
4242 	if (skip < SKIP_FINI && zone->uz_fini)
4243 		zone->uz_fini(item, zone->uz_size);
4244 
4245 	zone->uz_release(zone->uz_arg, &item, 1);
4246 
4247 	if (skip & SKIP_CNT)
4248 		return;
4249 
4250 	counter_u64_add(zone->uz_frees, 1);
4251 
4252 	if (zone->uz_max_items > 0)
4253 		zone_free_limit(zone, 1);
4254 }
4255 
4256 /* See uma.h */
4257 int
4258 uma_zone_set_max(uma_zone_t zone, int nitems)
4259 {
4260 	struct uma_bucket_zone *ubz;
4261 	int count;
4262 
4263 	/*
4264 	 * XXX This can misbehave if the zone has any allocations with
4265 	 * no limit and a limit is imposed.  There is currently no
4266 	 * way to clear a limit.
4267 	 */
4268 	ZONE_LOCK(zone);
4269 	ubz = bucket_zone_max(zone, nitems);
4270 	count = ubz != NULL ? ubz->ubz_entries : 0;
4271 	zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4272 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4273 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4274 	zone->uz_max_items = nitems;
4275 	zone->uz_flags |= UMA_ZFLAG_LIMIT;
4276 	zone_update_caches(zone);
4277 	/* We may need to wake waiters. */
4278 	wakeup(&zone->uz_max_items);
4279 	ZONE_UNLOCK(zone);
4280 
4281 	return (nitems);
4282 }
4283 
4284 /* See uma.h */
4285 void
4286 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4287 {
4288 	struct uma_bucket_zone *ubz;
4289 	int bpcpu;
4290 
4291 	ZONE_LOCK(zone);
4292 	ubz = bucket_zone_max(zone, nitems);
4293 	if (ubz != NULL) {
4294 		bpcpu = 2;
4295 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4296 			/* Count the cross-domain bucket. */
4297 			bpcpu++;
4298 		nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4299 		zone->uz_bucket_size_max = ubz->ubz_entries;
4300 	} else {
4301 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4302 	}
4303 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4304 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4305 	zone->uz_bkt_max = nitems;
4306 	ZONE_UNLOCK(zone);
4307 }
4308 
4309 /* See uma.h */
4310 int
4311 uma_zone_get_max(uma_zone_t zone)
4312 {
4313 	int nitems;
4314 
4315 	nitems = atomic_load_64(&zone->uz_max_items);
4316 
4317 	return (nitems);
4318 }
4319 
4320 /* See uma.h */
4321 void
4322 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4323 {
4324 
4325 	ZONE_ASSERT_COLD(zone);
4326 	zone->uz_warning = warning;
4327 }
4328 
4329 /* See uma.h */
4330 void
4331 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4332 {
4333 
4334 	ZONE_ASSERT_COLD(zone);
4335 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4336 }
4337 
4338 /* See uma.h */
4339 int
4340 uma_zone_get_cur(uma_zone_t zone)
4341 {
4342 	int64_t nitems;
4343 	u_int i;
4344 
4345 	nitems = 0;
4346 	if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4347 		nitems = counter_u64_fetch(zone->uz_allocs) -
4348 		    counter_u64_fetch(zone->uz_frees);
4349 	CPU_FOREACH(i)
4350 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4351 		    atomic_load_64(&zone->uz_cpu[i].uc_frees);
4352 
4353 	return (nitems < 0 ? 0 : nitems);
4354 }
4355 
4356 static uint64_t
4357 uma_zone_get_allocs(uma_zone_t zone)
4358 {
4359 	uint64_t nitems;
4360 	u_int i;
4361 
4362 	nitems = 0;
4363 	if (zone->uz_allocs != EARLY_COUNTER)
4364 		nitems = counter_u64_fetch(zone->uz_allocs);
4365 	CPU_FOREACH(i)
4366 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4367 
4368 	return (nitems);
4369 }
4370 
4371 static uint64_t
4372 uma_zone_get_frees(uma_zone_t zone)
4373 {
4374 	uint64_t nitems;
4375 	u_int i;
4376 
4377 	nitems = 0;
4378 	if (zone->uz_frees != EARLY_COUNTER)
4379 		nitems = counter_u64_fetch(zone->uz_frees);
4380 	CPU_FOREACH(i)
4381 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4382 
4383 	return (nitems);
4384 }
4385 
4386 #ifdef INVARIANTS
4387 /* Used only for KEG_ASSERT_COLD(). */
4388 static uint64_t
4389 uma_keg_get_allocs(uma_keg_t keg)
4390 {
4391 	uma_zone_t z;
4392 	uint64_t nitems;
4393 
4394 	nitems = 0;
4395 	LIST_FOREACH(z, &keg->uk_zones, uz_link)
4396 		nitems += uma_zone_get_allocs(z);
4397 
4398 	return (nitems);
4399 }
4400 #endif
4401 
4402 /* See uma.h */
4403 void
4404 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4405 {
4406 	uma_keg_t keg;
4407 
4408 	KEG_GET(zone, keg);
4409 	KEG_ASSERT_COLD(keg);
4410 	keg->uk_init = uminit;
4411 }
4412 
4413 /* See uma.h */
4414 void
4415 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4416 {
4417 	uma_keg_t keg;
4418 
4419 	KEG_GET(zone, keg);
4420 	KEG_ASSERT_COLD(keg);
4421 	keg->uk_fini = fini;
4422 }
4423 
4424 /* See uma.h */
4425 void
4426 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4427 {
4428 
4429 	ZONE_ASSERT_COLD(zone);
4430 	zone->uz_init = zinit;
4431 }
4432 
4433 /* See uma.h */
4434 void
4435 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4436 {
4437 
4438 	ZONE_ASSERT_COLD(zone);
4439 	zone->uz_fini = zfini;
4440 }
4441 
4442 /* See uma.h */
4443 void
4444 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4445 {
4446 	uma_keg_t keg;
4447 
4448 	KEG_GET(zone, keg);
4449 	KEG_ASSERT_COLD(keg);
4450 	keg->uk_freef = freef;
4451 }
4452 
4453 /* See uma.h */
4454 void
4455 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4456 {
4457 	uma_keg_t keg;
4458 
4459 	KEG_GET(zone, keg);
4460 	KEG_ASSERT_COLD(keg);
4461 	keg->uk_allocf = allocf;
4462 }
4463 
4464 /* See uma.h */
4465 void
4466 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4467 {
4468 
4469 	ZONE_ASSERT_COLD(zone);
4470 
4471 	zone->uz_flags |= UMA_ZONE_SMR;
4472 	zone->uz_smr = smr;
4473 	zone_update_caches(zone);
4474 }
4475 
4476 smr_t
4477 uma_zone_get_smr(uma_zone_t zone)
4478 {
4479 
4480 	return (zone->uz_smr);
4481 }
4482 
4483 /* See uma.h */
4484 void
4485 uma_zone_reserve(uma_zone_t zone, int items)
4486 {
4487 	uma_keg_t keg;
4488 
4489 	KEG_GET(zone, keg);
4490 	KEG_ASSERT_COLD(keg);
4491 	keg->uk_reserve = items;
4492 }
4493 
4494 /* See uma.h */
4495 int
4496 uma_zone_reserve_kva(uma_zone_t zone, int count)
4497 {
4498 	uma_keg_t keg;
4499 	vm_offset_t kva;
4500 	u_int pages;
4501 
4502 	KEG_GET(zone, keg);
4503 	KEG_ASSERT_COLD(keg);
4504 	ZONE_ASSERT_COLD(zone);
4505 
4506 	pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4507 
4508 #ifdef UMA_MD_SMALL_ALLOC
4509 	if (keg->uk_ppera > 1) {
4510 #else
4511 	if (1) {
4512 #endif
4513 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4514 		if (kva == 0)
4515 			return (0);
4516 	} else
4517 		kva = 0;
4518 
4519 	ZONE_LOCK(zone);
4520 	MPASS(keg->uk_kva == 0);
4521 	keg->uk_kva = kva;
4522 	keg->uk_offset = 0;
4523 	zone->uz_max_items = pages * keg->uk_ipers;
4524 #ifdef UMA_MD_SMALL_ALLOC
4525 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4526 #else
4527 	keg->uk_allocf = noobj_alloc;
4528 #endif
4529 	keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4530 	zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4531 	zone_update_caches(zone);
4532 	ZONE_UNLOCK(zone);
4533 
4534 	return (1);
4535 }
4536 
4537 /* See uma.h */
4538 void
4539 uma_prealloc(uma_zone_t zone, int items)
4540 {
4541 	struct vm_domainset_iter di;
4542 	uma_domain_t dom;
4543 	uma_slab_t slab;
4544 	uma_keg_t keg;
4545 	int aflags, domain, slabs;
4546 
4547 	KEG_GET(zone, keg);
4548 	slabs = howmany(items, keg->uk_ipers);
4549 	while (slabs-- > 0) {
4550 		aflags = M_NOWAIT;
4551 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4552 		    &aflags);
4553 		for (;;) {
4554 			slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4555 			    aflags);
4556 			if (slab != NULL) {
4557 				dom = &keg->uk_domain[slab->us_domain];
4558 				LIST_REMOVE(slab, us_link);
4559 				LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4560 				    us_link);
4561 				KEG_UNLOCK(keg, slab->us_domain);
4562 				break;
4563 			}
4564 			if (vm_domainset_iter_policy(&di, &domain) != 0)
4565 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
4566 		}
4567 	}
4568 }
4569 
4570 /* See uma.h */
4571 void
4572 uma_reclaim(int req)
4573 {
4574 
4575 	CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4576 	sx_xlock(&uma_reclaim_lock);
4577 	bucket_enable();
4578 
4579 	switch (req) {
4580 	case UMA_RECLAIM_TRIM:
4581 		zone_foreach(zone_trim, NULL);
4582 		break;
4583 	case UMA_RECLAIM_DRAIN:
4584 	case UMA_RECLAIM_DRAIN_CPU:
4585 		zone_foreach(zone_drain, NULL);
4586 		if (req == UMA_RECLAIM_DRAIN_CPU) {
4587 			pcpu_cache_drain_safe(NULL);
4588 			zone_foreach(zone_drain, NULL);
4589 		}
4590 		break;
4591 	default:
4592 		panic("unhandled reclamation request %d", req);
4593 	}
4594 
4595 	/*
4596 	 * Some slabs may have been freed but this zone will be visited early
4597 	 * we visit again so that we can free pages that are empty once other
4598 	 * zones are drained.  We have to do the same for buckets.
4599 	 */
4600 	zone_drain(slabzones[0], NULL);
4601 	zone_drain(slabzones[1], NULL);
4602 	bucket_zone_drain();
4603 	sx_xunlock(&uma_reclaim_lock);
4604 }
4605 
4606 static volatile int uma_reclaim_needed;
4607 
4608 void
4609 uma_reclaim_wakeup(void)
4610 {
4611 
4612 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4613 		wakeup(uma_reclaim);
4614 }
4615 
4616 void
4617 uma_reclaim_worker(void *arg __unused)
4618 {
4619 
4620 	for (;;) {
4621 		sx_xlock(&uma_reclaim_lock);
4622 		while (atomic_load_int(&uma_reclaim_needed) == 0)
4623 			sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4624 			    hz);
4625 		sx_xunlock(&uma_reclaim_lock);
4626 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4627 		uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4628 		atomic_store_int(&uma_reclaim_needed, 0);
4629 		/* Don't fire more than once per-second. */
4630 		pause("umarclslp", hz);
4631 	}
4632 }
4633 
4634 /* See uma.h */
4635 void
4636 uma_zone_reclaim(uma_zone_t zone, int req)
4637 {
4638 
4639 	switch (req) {
4640 	case UMA_RECLAIM_TRIM:
4641 		zone_trim(zone, NULL);
4642 		break;
4643 	case UMA_RECLAIM_DRAIN:
4644 		zone_drain(zone, NULL);
4645 		break;
4646 	case UMA_RECLAIM_DRAIN_CPU:
4647 		pcpu_cache_drain_safe(zone);
4648 		zone_drain(zone, NULL);
4649 		break;
4650 	default:
4651 		panic("unhandled reclamation request %d", req);
4652 	}
4653 }
4654 
4655 /* See uma.h */
4656 int
4657 uma_zone_exhausted(uma_zone_t zone)
4658 {
4659 
4660 	return (atomic_load_32(&zone->uz_sleepers) > 0);
4661 }
4662 
4663 unsigned long
4664 uma_limit(void)
4665 {
4666 
4667 	return (uma_kmem_limit);
4668 }
4669 
4670 void
4671 uma_set_limit(unsigned long limit)
4672 {
4673 
4674 	uma_kmem_limit = limit;
4675 }
4676 
4677 unsigned long
4678 uma_size(void)
4679 {
4680 
4681 	return (atomic_load_long(&uma_kmem_total));
4682 }
4683 
4684 long
4685 uma_avail(void)
4686 {
4687 
4688 	return (uma_kmem_limit - uma_size());
4689 }
4690 
4691 #ifdef DDB
4692 /*
4693  * Generate statistics across both the zone and its per-cpu cache's.  Return
4694  * desired statistics if the pointer is non-NULL for that statistic.
4695  *
4696  * Note: does not update the zone statistics, as it can't safely clear the
4697  * per-CPU cache statistic.
4698  *
4699  */
4700 static void
4701 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4702     uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4703 {
4704 	uma_cache_t cache;
4705 	uint64_t allocs, frees, sleeps, xdomain;
4706 	int cachefree, cpu;
4707 
4708 	allocs = frees = sleeps = xdomain = 0;
4709 	cachefree = 0;
4710 	CPU_FOREACH(cpu) {
4711 		cache = &z->uz_cpu[cpu];
4712 		cachefree += cache->uc_allocbucket.ucb_cnt;
4713 		cachefree += cache->uc_freebucket.ucb_cnt;
4714 		xdomain += cache->uc_crossbucket.ucb_cnt;
4715 		cachefree += cache->uc_crossbucket.ucb_cnt;
4716 		allocs += cache->uc_allocs;
4717 		frees += cache->uc_frees;
4718 	}
4719 	allocs += counter_u64_fetch(z->uz_allocs);
4720 	frees += counter_u64_fetch(z->uz_frees);
4721 	sleeps += z->uz_sleeps;
4722 	xdomain += z->uz_xdomain;
4723 	if (cachefreep != NULL)
4724 		*cachefreep = cachefree;
4725 	if (allocsp != NULL)
4726 		*allocsp = allocs;
4727 	if (freesp != NULL)
4728 		*freesp = frees;
4729 	if (sleepsp != NULL)
4730 		*sleepsp = sleeps;
4731 	if (xdomainp != NULL)
4732 		*xdomainp = xdomain;
4733 }
4734 #endif /* DDB */
4735 
4736 static int
4737 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
4738 {
4739 	uma_keg_t kz;
4740 	uma_zone_t z;
4741 	int count;
4742 
4743 	count = 0;
4744 	rw_rlock(&uma_rwlock);
4745 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
4746 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
4747 			count++;
4748 	}
4749 	LIST_FOREACH(z, &uma_cachezones, uz_link)
4750 		count++;
4751 
4752 	rw_runlock(&uma_rwlock);
4753 	return (sysctl_handle_int(oidp, &count, 0, req));
4754 }
4755 
4756 static void
4757 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
4758     struct uma_percpu_stat *ups, bool internal)
4759 {
4760 	uma_zone_domain_t zdom;
4761 	uma_cache_t cache;
4762 	int i;
4763 
4764 
4765 	for (i = 0; i < vm_ndomains; i++) {
4766 		zdom = &z->uz_domain[i];
4767 		uth->uth_zone_free += zdom->uzd_nitems;
4768 	}
4769 	uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
4770 	uth->uth_frees = counter_u64_fetch(z->uz_frees);
4771 	uth->uth_fails = counter_u64_fetch(z->uz_fails);
4772 	uth->uth_sleeps = z->uz_sleeps;
4773 	uth->uth_xdomain = z->uz_xdomain;
4774 
4775 	/*
4776 	 * While it is not normally safe to access the cache bucket pointers
4777 	 * while not on the CPU that owns the cache, we only allow the pointers
4778 	 * to be exchanged without the zone lock held, not invalidated, so
4779 	 * accept the possible race associated with bucket exchange during
4780 	 * monitoring.  Use atomic_load_ptr() to ensure that the bucket pointers
4781 	 * are loaded only once.
4782 	 */
4783 	for (i = 0; i < mp_maxid + 1; i++) {
4784 		bzero(&ups[i], sizeof(*ups));
4785 		if (internal || CPU_ABSENT(i))
4786 			continue;
4787 		cache = &z->uz_cpu[i];
4788 		ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
4789 		ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
4790 		ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
4791 		ups[i].ups_allocs = cache->uc_allocs;
4792 		ups[i].ups_frees = cache->uc_frees;
4793 	}
4794 }
4795 
4796 static int
4797 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
4798 {
4799 	struct uma_stream_header ush;
4800 	struct uma_type_header uth;
4801 	struct uma_percpu_stat *ups;
4802 	struct sbuf sbuf;
4803 	uma_keg_t kz;
4804 	uma_zone_t z;
4805 	uint64_t items;
4806 	uint32_t kfree, pages;
4807 	int count, error, i;
4808 
4809 	error = sysctl_wire_old_buffer(req, 0);
4810 	if (error != 0)
4811 		return (error);
4812 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
4813 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
4814 	ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
4815 
4816 	count = 0;
4817 	rw_rlock(&uma_rwlock);
4818 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
4819 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
4820 			count++;
4821 	}
4822 
4823 	LIST_FOREACH(z, &uma_cachezones, uz_link)
4824 		count++;
4825 
4826 	/*
4827 	 * Insert stream header.
4828 	 */
4829 	bzero(&ush, sizeof(ush));
4830 	ush.ush_version = UMA_STREAM_VERSION;
4831 	ush.ush_maxcpus = (mp_maxid + 1);
4832 	ush.ush_count = count;
4833 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
4834 
4835 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
4836 		kfree = pages = 0;
4837 		for (i = 0; i < vm_ndomains; i++) {
4838 			kfree += kz->uk_domain[i].ud_free;
4839 			pages += kz->uk_domain[i].ud_pages;
4840 		}
4841 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
4842 			bzero(&uth, sizeof(uth));
4843 			ZONE_LOCK(z);
4844 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
4845 			uth.uth_align = kz->uk_align;
4846 			uth.uth_size = kz->uk_size;
4847 			uth.uth_rsize = kz->uk_rsize;
4848 			if (z->uz_max_items > 0) {
4849 				items = UZ_ITEMS_COUNT(z->uz_items);
4850 				uth.uth_pages = (items / kz->uk_ipers) *
4851 					kz->uk_ppera;
4852 			} else
4853 				uth.uth_pages = pages;
4854 			uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
4855 			    kz->uk_ppera;
4856 			uth.uth_limit = z->uz_max_items;
4857 			uth.uth_keg_free = kfree;
4858 
4859 			/*
4860 			 * A zone is secondary is it is not the first entry
4861 			 * on the keg's zone list.
4862 			 */
4863 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
4864 			    (LIST_FIRST(&kz->uk_zones) != z))
4865 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
4866 			uma_vm_zone_stats(&uth, z, &sbuf, ups,
4867 			    kz->uk_flags & UMA_ZFLAG_INTERNAL);
4868 			ZONE_UNLOCK(z);
4869 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
4870 			for (i = 0; i < mp_maxid + 1; i++)
4871 				(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
4872 		}
4873 	}
4874 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
4875 		bzero(&uth, sizeof(uth));
4876 		ZONE_LOCK(z);
4877 		strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
4878 		uth.uth_size = z->uz_size;
4879 		uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
4880 		ZONE_UNLOCK(z);
4881 		(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
4882 		for (i = 0; i < mp_maxid + 1; i++)
4883 			(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
4884 	}
4885 
4886 	rw_runlock(&uma_rwlock);
4887 	error = sbuf_finish(&sbuf);
4888 	sbuf_delete(&sbuf);
4889 	free(ups, M_TEMP);
4890 	return (error);
4891 }
4892 
4893 int
4894 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
4895 {
4896 	uma_zone_t zone = *(uma_zone_t *)arg1;
4897 	int error, max;
4898 
4899 	max = uma_zone_get_max(zone);
4900 	error = sysctl_handle_int(oidp, &max, 0, req);
4901 	if (error || !req->newptr)
4902 		return (error);
4903 
4904 	uma_zone_set_max(zone, max);
4905 
4906 	return (0);
4907 }
4908 
4909 int
4910 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
4911 {
4912 	uma_zone_t zone;
4913 	int cur;
4914 
4915 	/*
4916 	 * Some callers want to add sysctls for global zones that
4917 	 * may not yet exist so they pass a pointer to a pointer.
4918 	 */
4919 	if (arg2 == 0)
4920 		zone = *(uma_zone_t *)arg1;
4921 	else
4922 		zone = arg1;
4923 	cur = uma_zone_get_cur(zone);
4924 	return (sysctl_handle_int(oidp, &cur, 0, req));
4925 }
4926 
4927 static int
4928 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
4929 {
4930 	uma_zone_t zone = arg1;
4931 	uint64_t cur;
4932 
4933 	cur = uma_zone_get_allocs(zone);
4934 	return (sysctl_handle_64(oidp, &cur, 0, req));
4935 }
4936 
4937 static int
4938 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
4939 {
4940 	uma_zone_t zone = arg1;
4941 	uint64_t cur;
4942 
4943 	cur = uma_zone_get_frees(zone);
4944 	return (sysctl_handle_64(oidp, &cur, 0, req));
4945 }
4946 
4947 static int
4948 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
4949 {
4950 	struct sbuf sbuf;
4951 	uma_zone_t zone = arg1;
4952 	int error;
4953 
4954 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
4955 	if (zone->uz_flags != 0)
4956 		sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
4957 	else
4958 		sbuf_printf(&sbuf, "0");
4959 	error = sbuf_finish(&sbuf);
4960 	sbuf_delete(&sbuf);
4961 
4962 	return (error);
4963 }
4964 
4965 static int
4966 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
4967 {
4968 	uma_keg_t keg = arg1;
4969 	int avail, effpct, total;
4970 
4971 	total = keg->uk_ppera * PAGE_SIZE;
4972 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
4973 		total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
4974 	/*
4975 	 * We consider the client's requested size and alignment here, not the
4976 	 * real size determination uk_rsize, because we also adjust the real
4977 	 * size for internal implementation reasons (max bitset size).
4978 	 */
4979 	avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
4980 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
4981 		avail *= mp_maxid + 1;
4982 	effpct = 100 * avail / total;
4983 	return (sysctl_handle_int(oidp, &effpct, 0, req));
4984 }
4985 
4986 static int
4987 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
4988 {
4989 	uma_zone_t zone = arg1;
4990 	uint64_t cur;
4991 
4992 	cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
4993 	return (sysctl_handle_64(oidp, &cur, 0, req));
4994 }
4995 
4996 #ifdef INVARIANTS
4997 static uma_slab_t
4998 uma_dbg_getslab(uma_zone_t zone, void *item)
4999 {
5000 	uma_slab_t slab;
5001 	uma_keg_t keg;
5002 	uint8_t *mem;
5003 
5004 	/*
5005 	 * It is safe to return the slab here even though the
5006 	 * zone is unlocked because the item's allocation state
5007 	 * essentially holds a reference.
5008 	 */
5009 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5010 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5011 		return (NULL);
5012 	if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5013 		return (vtoslab((vm_offset_t)mem));
5014 	keg = zone->uz_keg;
5015 	if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5016 		return ((uma_slab_t)(mem + keg->uk_pgoff));
5017 	KEG_LOCK(keg, 0);
5018 	slab = hash_sfind(&keg->uk_hash, mem);
5019 	KEG_UNLOCK(keg, 0);
5020 
5021 	return (slab);
5022 }
5023 
5024 static bool
5025 uma_dbg_zskip(uma_zone_t zone, void *mem)
5026 {
5027 
5028 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5029 		return (true);
5030 
5031 	return (uma_dbg_kskip(zone->uz_keg, mem));
5032 }
5033 
5034 static bool
5035 uma_dbg_kskip(uma_keg_t keg, void *mem)
5036 {
5037 	uintptr_t idx;
5038 
5039 	if (dbg_divisor == 0)
5040 		return (true);
5041 
5042 	if (dbg_divisor == 1)
5043 		return (false);
5044 
5045 	idx = (uintptr_t)mem >> PAGE_SHIFT;
5046 	if (keg->uk_ipers > 1) {
5047 		idx *= keg->uk_ipers;
5048 		idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5049 	}
5050 
5051 	if ((idx / dbg_divisor) * dbg_divisor != idx) {
5052 		counter_u64_add(uma_skip_cnt, 1);
5053 		return (true);
5054 	}
5055 	counter_u64_add(uma_dbg_cnt, 1);
5056 
5057 	return (false);
5058 }
5059 
5060 /*
5061  * Set up the slab's freei data such that uma_dbg_free can function.
5062  *
5063  */
5064 static void
5065 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5066 {
5067 	uma_keg_t keg;
5068 	int freei;
5069 
5070 	if (slab == NULL) {
5071 		slab = uma_dbg_getslab(zone, item);
5072 		if (slab == NULL)
5073 			panic("uma: item %p did not belong to zone %s\n",
5074 			    item, zone->uz_name);
5075 	}
5076 	keg = zone->uz_keg;
5077 	freei = slab_item_index(slab, keg, item);
5078 
5079 	if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5080 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
5081 		    item, zone, zone->uz_name, slab, freei);
5082 	BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5083 }
5084 
5085 /*
5086  * Verifies freed addresses.  Checks for alignment, valid slab membership
5087  * and duplicate frees.
5088  *
5089  */
5090 static void
5091 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5092 {
5093 	uma_keg_t keg;
5094 	int freei;
5095 
5096 	if (slab == NULL) {
5097 		slab = uma_dbg_getslab(zone, item);
5098 		if (slab == NULL)
5099 			panic("uma: Freed item %p did not belong to zone %s\n",
5100 			    item, zone->uz_name);
5101 	}
5102 	keg = zone->uz_keg;
5103 	freei = slab_item_index(slab, keg, item);
5104 
5105 	if (freei >= keg->uk_ipers)
5106 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
5107 		    item, zone, zone->uz_name, slab, freei);
5108 
5109 	if (slab_item(slab, keg, freei) != item)
5110 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
5111 		    item, zone, zone->uz_name, slab, freei);
5112 
5113 	if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5114 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
5115 		    item, zone, zone->uz_name, slab, freei);
5116 
5117 	BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5118 }
5119 #endif /* INVARIANTS */
5120 
5121 #ifdef DDB
5122 static int64_t
5123 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5124     uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5125 {
5126 	uint64_t frees;
5127 	int i;
5128 
5129 	if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5130 		*allocs = counter_u64_fetch(z->uz_allocs);
5131 		frees = counter_u64_fetch(z->uz_frees);
5132 		*sleeps = z->uz_sleeps;
5133 		*cachefree = 0;
5134 		*xdomain = 0;
5135 	} else
5136 		uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5137 		    xdomain);
5138 	for (i = 0; i < vm_ndomains; i++) {
5139 		*cachefree += z->uz_domain[i].uzd_nitems;
5140 		if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5141 		    (LIST_FIRST(&kz->uk_zones) != z)))
5142 			*cachefree += kz->uk_domain[i].ud_free;
5143 	}
5144 	*used = *allocs - frees;
5145 	return (((int64_t)*used + *cachefree) * kz->uk_size);
5146 }
5147 
5148 DB_SHOW_COMMAND(uma, db_show_uma)
5149 {
5150 	const char *fmt_hdr, *fmt_entry;
5151 	uma_keg_t kz;
5152 	uma_zone_t z;
5153 	uint64_t allocs, used, sleeps, xdomain;
5154 	long cachefree;
5155 	/* variables for sorting */
5156 	uma_keg_t cur_keg;
5157 	uma_zone_t cur_zone, last_zone;
5158 	int64_t cur_size, last_size, size;
5159 	int ties;
5160 
5161 	/* /i option produces machine-parseable CSV output */
5162 	if (modif[0] == 'i') {
5163 		fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5164 		fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5165 	} else {
5166 		fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5167 		fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5168 	}
5169 
5170 	db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5171 	    "Sleeps", "Bucket", "Total Mem", "XFree");
5172 
5173 	/* Sort the zones with largest size first. */
5174 	last_zone = NULL;
5175 	last_size = INT64_MAX;
5176 	for (;;) {
5177 		cur_zone = NULL;
5178 		cur_size = -1;
5179 		ties = 0;
5180 		LIST_FOREACH(kz, &uma_kegs, uk_link) {
5181 			LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5182 				/*
5183 				 * In the case of size ties, print out zones
5184 				 * in the order they are encountered.  That is,
5185 				 * when we encounter the most recently output
5186 				 * zone, we have already printed all preceding
5187 				 * ties, and we must print all following ties.
5188 				 */
5189 				if (z == last_zone) {
5190 					ties = 1;
5191 					continue;
5192 				}
5193 				size = get_uma_stats(kz, z, &allocs, &used,
5194 				    &sleeps, &cachefree, &xdomain);
5195 				if (size > cur_size && size < last_size + ties)
5196 				{
5197 					cur_size = size;
5198 					cur_zone = z;
5199 					cur_keg = kz;
5200 				}
5201 			}
5202 		}
5203 		if (cur_zone == NULL)
5204 			break;
5205 
5206 		size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5207 		    &sleeps, &cachefree, &xdomain);
5208 		db_printf(fmt_entry, cur_zone->uz_name,
5209 		    (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5210 		    (uintmax_t)allocs, (uintmax_t)sleeps,
5211 		    (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5212 		    xdomain);
5213 
5214 		if (db_pager_quit)
5215 			return;
5216 		last_zone = cur_zone;
5217 		last_size = cur_size;
5218 	}
5219 }
5220 
5221 DB_SHOW_COMMAND(umacache, db_show_umacache)
5222 {
5223 	uma_zone_t z;
5224 	uint64_t allocs, frees;
5225 	long cachefree;
5226 	int i;
5227 
5228 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5229 	    "Requests", "Bucket");
5230 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5231 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5232 		for (i = 0; i < vm_ndomains; i++)
5233 			cachefree += z->uz_domain[i].uzd_nitems;
5234 		db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5235 		    z->uz_name, (uintmax_t)z->uz_size,
5236 		    (intmax_t)(allocs - frees), cachefree,
5237 		    (uintmax_t)allocs, z->uz_bucket_size);
5238 		if (db_pager_quit)
5239 			return;
5240 	}
5241 }
5242 #endif	/* DDB */
5243