xref: /freebsd/sys/vm/uma_core.c (revision c66ec88fed842fbaad62c30d510644ceb7bd2d71)
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_param.h>
86 #include <vm/vm_domainset.h>
87 #include <vm/vm_object.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_pageout.h>
90 #include <vm/vm_phys.h>
91 #include <vm/vm_pagequeue.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
95 #include <vm/vm_dumpset.h>
96 #include <vm/uma.h>
97 #include <vm/uma_int.h>
98 #include <vm/uma_dbg.h>
99 
100 #include <ddb/ddb.h>
101 
102 #ifdef DEBUG_MEMGUARD
103 #include <vm/memguard.h>
104 #endif
105 
106 #include <machine/md_var.h>
107 
108 #ifdef INVARIANTS
109 #define	UMA_ALWAYS_CTORDTOR	1
110 #else
111 #define	UMA_ALWAYS_CTORDTOR	0
112 #endif
113 
114 /*
115  * This is the zone and keg from which all zones are spawned.
116  */
117 static uma_zone_t kegs;
118 static uma_zone_t zones;
119 
120 /*
121  * On INVARIANTS builds, the slab contains a second bitset of the same size,
122  * "dbg_bits", which is laid out immediately after us_free.
123  */
124 #ifdef INVARIANTS
125 #define	SLAB_BITSETS	2
126 #else
127 #define	SLAB_BITSETS	1
128 #endif
129 
130 /*
131  * These are the two zones from which all offpage uma_slab_ts are allocated.
132  *
133  * One zone is for slab headers that can represent a larger number of items,
134  * making the slabs themselves more efficient, and the other zone is for
135  * headers that are smaller and represent fewer items, making the headers more
136  * efficient.
137  */
138 #define	SLABZONE_SIZE(setsize)					\
139     (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
140 #define	SLABZONE0_SETSIZE	(PAGE_SIZE / 16)
141 #define	SLABZONE1_SETSIZE	SLAB_MAX_SETSIZE
142 #define	SLABZONE0_SIZE	SLABZONE_SIZE(SLABZONE0_SETSIZE)
143 #define	SLABZONE1_SIZE	SLABZONE_SIZE(SLABZONE1_SETSIZE)
144 static uma_zone_t slabzones[2];
145 
146 /*
147  * The initial hash tables come out of this zone so they can be allocated
148  * prior to malloc coming up.
149  */
150 static uma_zone_t hashzone;
151 
152 /* The boot-time adjusted value for cache line alignment. */
153 int uma_align_cache = 64 - 1;
154 
155 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
156 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
157 
158 /*
159  * Are we allowed to allocate buckets?
160  */
161 static int bucketdisable = 1;
162 
163 /* Linked list of all kegs in the system */
164 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
165 
166 /* Linked list of all cache-only zones in the system */
167 static LIST_HEAD(,uma_zone) uma_cachezones =
168     LIST_HEAD_INITIALIZER(uma_cachezones);
169 
170 /* This RW lock protects the keg list */
171 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
172 
173 /*
174  * First available virual address for boot time allocations.
175  */
176 static vm_offset_t bootstart;
177 static vm_offset_t bootmem;
178 
179 static struct sx uma_reclaim_lock;
180 
181 /*
182  * kmem soft limit, initialized by uma_set_limit().  Ensure that early
183  * allocations don't trigger a wakeup of the reclaim thread.
184  */
185 unsigned long uma_kmem_limit = LONG_MAX;
186 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
187     "UMA kernel memory soft limit");
188 unsigned long uma_kmem_total;
189 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
190     "UMA kernel memory usage");
191 
192 /* Is the VM done starting up? */
193 static enum {
194 	BOOT_COLD,
195 	BOOT_KVA,
196 	BOOT_PCPU,
197 	BOOT_RUNNING,
198 	BOOT_SHUTDOWN,
199 } booted = BOOT_COLD;
200 
201 /*
202  * This is the handle used to schedule events that need to happen
203  * outside of the allocation fast path.
204  */
205 static struct callout uma_callout;
206 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
207 
208 /*
209  * This structure is passed as the zone ctor arg so that I don't have to create
210  * a special allocation function just for zones.
211  */
212 struct uma_zctor_args {
213 	const char *name;
214 	size_t size;
215 	uma_ctor ctor;
216 	uma_dtor dtor;
217 	uma_init uminit;
218 	uma_fini fini;
219 	uma_import import;
220 	uma_release release;
221 	void *arg;
222 	uma_keg_t keg;
223 	int align;
224 	uint32_t flags;
225 };
226 
227 struct uma_kctor_args {
228 	uma_zone_t zone;
229 	size_t size;
230 	uma_init uminit;
231 	uma_fini fini;
232 	int align;
233 	uint32_t flags;
234 };
235 
236 struct uma_bucket_zone {
237 	uma_zone_t	ubz_zone;
238 	const char	*ubz_name;
239 	int		ubz_entries;	/* Number of items it can hold. */
240 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
241 };
242 
243 /*
244  * Compute the actual number of bucket entries to pack them in power
245  * of two sizes for more efficient space utilization.
246  */
247 #define	BUCKET_SIZE(n)						\
248     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
249 
250 #define	BUCKET_MAX	BUCKET_SIZE(256)
251 
252 struct uma_bucket_zone bucket_zones[] = {
253 	/* Literal bucket sizes. */
254 	{ NULL, "2 Bucket", 2, 4096 },
255 	{ NULL, "4 Bucket", 4, 3072 },
256 	{ NULL, "8 Bucket", 8, 2048 },
257 	{ NULL, "16 Bucket", 16, 1024 },
258 	/* Rounded down power of 2 sizes for efficiency. */
259 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
260 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
261 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
262 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
263 	{ NULL, NULL, 0}
264 };
265 
266 /*
267  * Flags and enumerations to be passed to internal functions.
268  */
269 enum zfreeskip {
270 	SKIP_NONE =	0,
271 	SKIP_CNT =	0x00000001,
272 	SKIP_DTOR =	0x00010000,
273 	SKIP_FINI =	0x00020000,
274 };
275 
276 /* Prototypes.. */
277 
278 void	uma_startup1(vm_offset_t);
279 void	uma_startup2(void);
280 
281 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
282 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
283 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
284 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
285 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
286 static void page_free(void *, vm_size_t, uint8_t);
287 static void pcpu_page_free(void *, vm_size_t, uint8_t);
288 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
289 static void cache_drain(uma_zone_t);
290 static void bucket_drain(uma_zone_t, uma_bucket_t);
291 static void bucket_cache_reclaim(uma_zone_t zone, bool);
292 static int keg_ctor(void *, int, void *, int);
293 static void keg_dtor(void *, int, void *);
294 static int zone_ctor(void *, int, void *, int);
295 static void zone_dtor(void *, int, void *);
296 static inline void item_dtor(uma_zone_t zone, void *item, int size,
297     void *udata, enum zfreeskip skip);
298 static int zero_init(void *, int, int);
299 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
300     int itemdomain, bool ws);
301 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
302 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
303 static void zone_timeout(uma_zone_t zone, void *);
304 static int hash_alloc(struct uma_hash *, u_int);
305 static int hash_expand(struct uma_hash *, struct uma_hash *);
306 static void hash_free(struct uma_hash *hash);
307 static void uma_timeout(void *);
308 static void uma_shutdown(void);
309 static void *zone_alloc_item(uma_zone_t, void *, int, int);
310 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
311 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
312 static void zone_free_limit(uma_zone_t zone, int count);
313 static void bucket_enable(void);
314 static void bucket_init(void);
315 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
316 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
317 static void bucket_zone_drain(void);
318 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
319 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
320 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
321 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
322     uma_fini fini, int align, uint32_t flags);
323 static int zone_import(void *, void **, int, int, int);
324 static void zone_release(void *, void **, int);
325 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
326 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
327 
328 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
329 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
330 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
331 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
332 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
333 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
334 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
335 
336 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
337 
338 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
339     "Memory allocation debugging");
340 
341 #ifdef INVARIANTS
342 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
343 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
344 
345 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
346 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
347 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
348 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
349 
350 static u_int dbg_divisor = 1;
351 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
352     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
353     "Debug & thrash every this item in memory allocator");
354 
355 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
356 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
357 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
358     &uma_dbg_cnt, "memory items debugged");
359 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
360     &uma_skip_cnt, "memory items skipped, not debugged");
361 #endif
362 
363 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
364     "Universal Memory Allocator");
365 
366 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
367     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
368 
369 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
370     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
371 
372 static int zone_warnings = 1;
373 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
374     "Warn when UMA zones becomes full");
375 
376 static int multipage_slabs = 1;
377 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
378 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
379     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
380     "UMA may choose larger slab sizes for better efficiency");
381 
382 /*
383  * Select the slab zone for an offpage slab with the given maximum item count.
384  */
385 static inline uma_zone_t
386 slabzone(int ipers)
387 {
388 
389 	return (slabzones[ipers > SLABZONE0_SETSIZE]);
390 }
391 
392 /*
393  * This routine checks to see whether or not it's safe to enable buckets.
394  */
395 static void
396 bucket_enable(void)
397 {
398 
399 	KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
400 	bucketdisable = vm_page_count_min();
401 }
402 
403 /*
404  * Initialize bucket_zones, the array of zones of buckets of various sizes.
405  *
406  * For each zone, calculate the memory required for each bucket, consisting
407  * of the header and an array of pointers.
408  */
409 static void
410 bucket_init(void)
411 {
412 	struct uma_bucket_zone *ubz;
413 	int size;
414 
415 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
416 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
417 		size += sizeof(void *) * ubz->ubz_entries;
418 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
419 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
420 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
421 		    UMA_ZONE_FIRSTTOUCH);
422 	}
423 }
424 
425 /*
426  * Given a desired number of entries for a bucket, return the zone from which
427  * to allocate the bucket.
428  */
429 static struct uma_bucket_zone *
430 bucket_zone_lookup(int entries)
431 {
432 	struct uma_bucket_zone *ubz;
433 
434 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
435 		if (ubz->ubz_entries >= entries)
436 			return (ubz);
437 	ubz--;
438 	return (ubz);
439 }
440 
441 static int
442 bucket_select(int size)
443 {
444 	struct uma_bucket_zone *ubz;
445 
446 	ubz = &bucket_zones[0];
447 	if (size > ubz->ubz_maxsize)
448 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
449 
450 	for (; ubz->ubz_entries != 0; ubz++)
451 		if (ubz->ubz_maxsize < size)
452 			break;
453 	ubz--;
454 	return (ubz->ubz_entries);
455 }
456 
457 static uma_bucket_t
458 bucket_alloc(uma_zone_t zone, void *udata, int flags)
459 {
460 	struct uma_bucket_zone *ubz;
461 	uma_bucket_t bucket;
462 
463 	/*
464 	 * Don't allocate buckets early in boot.
465 	 */
466 	if (__predict_false(booted < BOOT_KVA))
467 		return (NULL);
468 
469 	/*
470 	 * To limit bucket recursion we store the original zone flags
471 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
472 	 * NOVM flag to persist even through deep recursions.  We also
473 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
474 	 * a bucket for a bucket zone so we do not allow infinite bucket
475 	 * recursion.  This cookie will even persist to frees of unused
476 	 * buckets via the allocation path or bucket allocations in the
477 	 * free path.
478 	 */
479 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
480 		udata = (void *)(uintptr_t)zone->uz_flags;
481 	else {
482 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
483 			return (NULL);
484 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
485 	}
486 	if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
487 		flags |= M_NOVM;
488 	ubz = bucket_zone_lookup(atomic_load_16(&zone->uz_bucket_size));
489 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
490 		ubz++;
491 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
492 	if (bucket) {
493 #ifdef INVARIANTS
494 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
495 #endif
496 		bucket->ub_cnt = 0;
497 		bucket->ub_entries = min(ubz->ubz_entries,
498 		    zone->uz_bucket_size_max);
499 		bucket->ub_seq = SMR_SEQ_INVALID;
500 		CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
501 		    zone->uz_name, zone, bucket);
502 	}
503 
504 	return (bucket);
505 }
506 
507 static void
508 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
509 {
510 	struct uma_bucket_zone *ubz;
511 
512 	if (bucket->ub_cnt != 0)
513 		bucket_drain(zone, bucket);
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  * Acquire the domain lock and record contention.
536  */
537 static uma_zone_domain_t
538 zone_domain_lock(uma_zone_t zone, int domain)
539 {
540 	uma_zone_domain_t zdom;
541 	bool lockfail;
542 
543 	zdom = ZDOM_GET(zone, domain);
544 	lockfail = false;
545 	if (ZDOM_OWNED(zdom))
546 		lockfail = true;
547 	ZDOM_LOCK(zdom);
548 	/* This is unsynchronized.  The counter does not need to be precise. */
549 	if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
550 		zone->uz_bucket_size++;
551 	return (zdom);
552 }
553 
554 /*
555  * Search for the domain with the least cached items and return it if it
556  * is out of balance with the preferred domain.
557  */
558 static __noinline int
559 zone_domain_lowest(uma_zone_t zone, int pref)
560 {
561 	long least, nitems, prefitems;
562 	int domain;
563 	int i;
564 
565 	prefitems = least = LONG_MAX;
566 	domain = 0;
567 	for (i = 0; i < vm_ndomains; i++) {
568 		nitems = ZDOM_GET(zone, i)->uzd_nitems;
569 		if (nitems < least) {
570 			domain = i;
571 			least = nitems;
572 		}
573 		if (domain == pref)
574 			prefitems = nitems;
575 	}
576 	if (prefitems < least * 2)
577 		return (pref);
578 
579 	return (domain);
580 }
581 
582 /*
583  * Search for the domain with the most cached items and return it or the
584  * preferred domain if it has enough to proceed.
585  */
586 static __noinline int
587 zone_domain_highest(uma_zone_t zone, int pref)
588 {
589 	long most, nitems;
590 	int domain;
591 	int i;
592 
593 	if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
594 		return (pref);
595 
596 	most = 0;
597 	domain = 0;
598 	for (i = 0; i < vm_ndomains; i++) {
599 		nitems = ZDOM_GET(zone, i)->uzd_nitems;
600 		if (nitems > most) {
601 			domain = i;
602 			most = nitems;
603 		}
604 	}
605 
606 	return (domain);
607 }
608 
609 /*
610  * Safely subtract cnt from imax.
611  */
612 static void
613 zone_domain_imax_sub(uma_zone_domain_t zdom, int cnt)
614 {
615 	long new;
616 	long old;
617 
618 	old = zdom->uzd_imax;
619 	do {
620 		if (old <= cnt)
621 			new = 0;
622 		else
623 			new = old - cnt;
624 	} while (atomic_fcmpset_long(&zdom->uzd_imax, &old, new) == 0);
625 }
626 
627 /*
628  * Set the maximum imax value.
629  */
630 static void
631 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
632 {
633 	long old;
634 
635 	old = zdom->uzd_imax;
636 	do {
637 		if (old >= nitems)
638 			break;
639 	} while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
640 }
641 
642 /*
643  * Attempt to satisfy an allocation by retrieving a full bucket from one of the
644  * zone's caches.  If a bucket is found the zone is not locked on return.
645  */
646 static uma_bucket_t
647 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
648 {
649 	uma_bucket_t bucket;
650 	int i;
651 	bool dtor = false;
652 
653 	ZDOM_LOCK_ASSERT(zdom);
654 
655 	if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
656 		return (NULL);
657 
658 	/* SMR Buckets can not be re-used until readers expire. */
659 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
660 	    bucket->ub_seq != SMR_SEQ_INVALID) {
661 		if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
662 			return (NULL);
663 		bucket->ub_seq = SMR_SEQ_INVALID;
664 		dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
665 		if (STAILQ_NEXT(bucket, ub_link) != NULL)
666 			zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
667 	}
668 	STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
669 
670 	KASSERT(zdom->uzd_nitems >= bucket->ub_cnt,
671 	    ("%s: item count underflow (%ld, %d)",
672 	    __func__, zdom->uzd_nitems, bucket->ub_cnt));
673 	KASSERT(bucket->ub_cnt > 0,
674 	    ("%s: empty bucket in bucket cache", __func__));
675 	zdom->uzd_nitems -= bucket->ub_cnt;
676 
677 	/*
678 	 * Shift the bounds of the current WSS interval to avoid
679 	 * perturbing the estimate.
680 	 */
681 	if (reclaim) {
682 		zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
683 		zone_domain_imax_sub(zdom, bucket->ub_cnt);
684 	} else if (zdom->uzd_imin > zdom->uzd_nitems)
685 		zdom->uzd_imin = zdom->uzd_nitems;
686 
687 	ZDOM_UNLOCK(zdom);
688 	if (dtor)
689 		for (i = 0; i < bucket->ub_cnt; i++)
690 			item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
691 			    NULL, SKIP_NONE);
692 
693 	return (bucket);
694 }
695 
696 /*
697  * Insert a full bucket into the specified cache.  The "ws" parameter indicates
698  * whether the bucket's contents should be counted as part of the zone's working
699  * set.  The bucket may be freed if it exceeds the bucket limit.
700  */
701 static void
702 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
703     const bool ws)
704 {
705 	uma_zone_domain_t zdom;
706 
707 	/* We don't cache empty buckets.  This can happen after a reclaim. */
708 	if (bucket->ub_cnt == 0)
709 		goto out;
710 	zdom = zone_domain_lock(zone, domain);
711 
712 	/*
713 	 * Conditionally set the maximum number of items.
714 	 */
715 	zdom->uzd_nitems += bucket->ub_cnt;
716 	if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
717 		if (ws)
718 			zone_domain_imax_set(zdom, zdom->uzd_nitems);
719 		if (STAILQ_EMPTY(&zdom->uzd_buckets))
720 			zdom->uzd_seq = bucket->ub_seq;
721 
722 		/*
723 		 * Try to promote reuse of recently used items.  For items
724 		 * protected by SMR, try to defer reuse to minimize polling.
725 		 */
726 		if (bucket->ub_seq == SMR_SEQ_INVALID)
727 			STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
728 		else
729 			STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
730 		ZDOM_UNLOCK(zdom);
731 		return;
732 	}
733 	zdom->uzd_nitems -= bucket->ub_cnt;
734 	ZDOM_UNLOCK(zdom);
735 out:
736 	bucket_free(zone, bucket, udata);
737 }
738 
739 /* Pops an item out of a per-cpu cache bucket. */
740 static inline void *
741 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
742 {
743 	void *item;
744 
745 	CRITICAL_ASSERT(curthread);
746 
747 	bucket->ucb_cnt--;
748 	item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
749 #ifdef INVARIANTS
750 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
751 	KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
752 #endif
753 	cache->uc_allocs++;
754 
755 	return (item);
756 }
757 
758 /* Pushes an item into a per-cpu cache bucket. */
759 static inline void
760 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
761 {
762 
763 	CRITICAL_ASSERT(curthread);
764 	KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
765 	    ("uma_zfree: Freeing to non free bucket index."));
766 
767 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
768 	bucket->ucb_cnt++;
769 	cache->uc_frees++;
770 }
771 
772 /*
773  * Unload a UMA bucket from a per-cpu cache.
774  */
775 static inline uma_bucket_t
776 cache_bucket_unload(uma_cache_bucket_t bucket)
777 {
778 	uma_bucket_t b;
779 
780 	b = bucket->ucb_bucket;
781 	if (b != NULL) {
782 		MPASS(b->ub_entries == bucket->ucb_entries);
783 		b->ub_cnt = bucket->ucb_cnt;
784 		bucket->ucb_bucket = NULL;
785 		bucket->ucb_entries = bucket->ucb_cnt = 0;
786 	}
787 
788 	return (b);
789 }
790 
791 static inline uma_bucket_t
792 cache_bucket_unload_alloc(uma_cache_t cache)
793 {
794 
795 	return (cache_bucket_unload(&cache->uc_allocbucket));
796 }
797 
798 static inline uma_bucket_t
799 cache_bucket_unload_free(uma_cache_t cache)
800 {
801 
802 	return (cache_bucket_unload(&cache->uc_freebucket));
803 }
804 
805 static inline uma_bucket_t
806 cache_bucket_unload_cross(uma_cache_t cache)
807 {
808 
809 	return (cache_bucket_unload(&cache->uc_crossbucket));
810 }
811 
812 /*
813  * Load a bucket into a per-cpu cache bucket.
814  */
815 static inline void
816 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
817 {
818 
819 	CRITICAL_ASSERT(curthread);
820 	MPASS(bucket->ucb_bucket == NULL);
821 	MPASS(b->ub_seq == SMR_SEQ_INVALID);
822 
823 	bucket->ucb_bucket = b;
824 	bucket->ucb_cnt = b->ub_cnt;
825 	bucket->ucb_entries = b->ub_entries;
826 }
827 
828 static inline void
829 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
830 {
831 
832 	cache_bucket_load(&cache->uc_allocbucket, b);
833 }
834 
835 static inline void
836 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
837 {
838 
839 	cache_bucket_load(&cache->uc_freebucket, b);
840 }
841 
842 #ifdef NUMA
843 static inline void
844 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
845 {
846 
847 	cache_bucket_load(&cache->uc_crossbucket, b);
848 }
849 #endif
850 
851 /*
852  * Copy and preserve ucb_spare.
853  */
854 static inline void
855 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
856 {
857 
858 	b1->ucb_bucket = b2->ucb_bucket;
859 	b1->ucb_entries = b2->ucb_entries;
860 	b1->ucb_cnt = b2->ucb_cnt;
861 }
862 
863 /*
864  * Swap two cache buckets.
865  */
866 static inline void
867 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
868 {
869 	struct uma_cache_bucket b3;
870 
871 	CRITICAL_ASSERT(curthread);
872 
873 	cache_bucket_copy(&b3, b1);
874 	cache_bucket_copy(b1, b2);
875 	cache_bucket_copy(b2, &b3);
876 }
877 
878 /*
879  * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
880  */
881 static uma_bucket_t
882 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
883 {
884 	uma_zone_domain_t zdom;
885 	uma_bucket_t bucket;
886 
887 	/*
888 	 * Avoid the lock if possible.
889 	 */
890 	zdom = ZDOM_GET(zone, domain);
891 	if (zdom->uzd_nitems == 0)
892 		return (NULL);
893 
894 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
895 	    !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
896 		return (NULL);
897 
898 	/*
899 	 * Check the zone's cache of buckets.
900 	 */
901 	zdom = zone_domain_lock(zone, domain);
902 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL)
903 		return (bucket);
904 	ZDOM_UNLOCK(zdom);
905 
906 	return (NULL);
907 }
908 
909 static void
910 zone_log_warning(uma_zone_t zone)
911 {
912 	static const struct timeval warninterval = { 300, 0 };
913 
914 	if (!zone_warnings || zone->uz_warning == NULL)
915 		return;
916 
917 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
918 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
919 }
920 
921 static inline void
922 zone_maxaction(uma_zone_t zone)
923 {
924 
925 	if (zone->uz_maxaction.ta_func != NULL)
926 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
927 }
928 
929 /*
930  * Routine called by timeout which is used to fire off some time interval
931  * based calculations.  (stats, hash size, etc.)
932  *
933  * Arguments:
934  *	arg   Unused
935  *
936  * Returns:
937  *	Nothing
938  */
939 static void
940 uma_timeout(void *unused)
941 {
942 	bucket_enable();
943 	zone_foreach(zone_timeout, NULL);
944 
945 	/* Reschedule this event */
946 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
947 }
948 
949 /*
950  * Update the working set size estimate for the zone's bucket cache.
951  * The constants chosen here are somewhat arbitrary.  With an update period of
952  * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
953  * last 100s.
954  */
955 static void
956 zone_domain_update_wss(uma_zone_domain_t zdom)
957 {
958 	long wss;
959 
960 	ZDOM_LOCK(zdom);
961 	MPASS(zdom->uzd_imax >= zdom->uzd_imin);
962 	wss = zdom->uzd_imax - zdom->uzd_imin;
963 	zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
964 	zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
965 	ZDOM_UNLOCK(zdom);
966 }
967 
968 /*
969  * Routine to perform timeout driven calculations.  This expands the
970  * hashes and does per cpu statistics aggregation.
971  *
972  *  Returns nothing.
973  */
974 static void
975 zone_timeout(uma_zone_t zone, void *unused)
976 {
977 	uma_keg_t keg;
978 	u_int slabs, pages;
979 
980 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
981 		goto update_wss;
982 
983 	keg = zone->uz_keg;
984 
985 	/*
986 	 * Hash zones are non-numa by definition so the first domain
987 	 * is the only one present.
988 	 */
989 	KEG_LOCK(keg, 0);
990 	pages = keg->uk_domain[0].ud_pages;
991 
992 	/*
993 	 * Expand the keg hash table.
994 	 *
995 	 * This is done if the number of slabs is larger than the hash size.
996 	 * What I'm trying to do here is completely reduce collisions.  This
997 	 * may be a little aggressive.  Should I allow for two collisions max?
998 	 */
999 	if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1000 		struct uma_hash newhash;
1001 		struct uma_hash oldhash;
1002 		int ret;
1003 
1004 		/*
1005 		 * This is so involved because allocating and freeing
1006 		 * while the keg lock is held will lead to deadlock.
1007 		 * I have to do everything in stages and check for
1008 		 * races.
1009 		 */
1010 		KEG_UNLOCK(keg, 0);
1011 		ret = hash_alloc(&newhash, 1 << fls(slabs));
1012 		KEG_LOCK(keg, 0);
1013 		if (ret) {
1014 			if (hash_expand(&keg->uk_hash, &newhash)) {
1015 				oldhash = keg->uk_hash;
1016 				keg->uk_hash = newhash;
1017 			} else
1018 				oldhash = newhash;
1019 
1020 			KEG_UNLOCK(keg, 0);
1021 			hash_free(&oldhash);
1022 			goto update_wss;
1023 		}
1024 	}
1025 	KEG_UNLOCK(keg, 0);
1026 
1027 update_wss:
1028 	for (int i = 0; i < vm_ndomains; i++)
1029 		zone_domain_update_wss(ZDOM_GET(zone, i));
1030 }
1031 
1032 /*
1033  * Allocate and zero fill the next sized hash table from the appropriate
1034  * backing store.
1035  *
1036  * Arguments:
1037  *	hash  A new hash structure with the old hash size in uh_hashsize
1038  *
1039  * Returns:
1040  *	1 on success and 0 on failure.
1041  */
1042 static int
1043 hash_alloc(struct uma_hash *hash, u_int size)
1044 {
1045 	size_t alloc;
1046 
1047 	KASSERT(powerof2(size), ("hash size must be power of 2"));
1048 	if (size > UMA_HASH_SIZE_INIT)  {
1049 		hash->uh_hashsize = size;
1050 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1051 		hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1052 	} else {
1053 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1054 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1055 		    UMA_ANYDOMAIN, M_WAITOK);
1056 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1057 	}
1058 	if (hash->uh_slab_hash) {
1059 		bzero(hash->uh_slab_hash, alloc);
1060 		hash->uh_hashmask = hash->uh_hashsize - 1;
1061 		return (1);
1062 	}
1063 
1064 	return (0);
1065 }
1066 
1067 /*
1068  * Expands the hash table for HASH zones.  This is done from zone_timeout
1069  * to reduce collisions.  This must not be done in the regular allocation
1070  * path, otherwise, we can recurse on the vm while allocating pages.
1071  *
1072  * Arguments:
1073  *	oldhash  The hash you want to expand
1074  *	newhash  The hash structure for the new table
1075  *
1076  * Returns:
1077  *	Nothing
1078  *
1079  * Discussion:
1080  */
1081 static int
1082 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1083 {
1084 	uma_hash_slab_t slab;
1085 	u_int hval;
1086 	u_int idx;
1087 
1088 	if (!newhash->uh_slab_hash)
1089 		return (0);
1090 
1091 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1092 		return (0);
1093 
1094 	/*
1095 	 * I need to investigate hash algorithms for resizing without a
1096 	 * full rehash.
1097 	 */
1098 
1099 	for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1100 		while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1101 			slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1102 			LIST_REMOVE(slab, uhs_hlink);
1103 			hval = UMA_HASH(newhash, slab->uhs_data);
1104 			LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1105 			    slab, uhs_hlink);
1106 		}
1107 
1108 	return (1);
1109 }
1110 
1111 /*
1112  * Free the hash bucket to the appropriate backing store.
1113  *
1114  * Arguments:
1115  *	slab_hash  The hash bucket we're freeing
1116  *	hashsize   The number of entries in that hash bucket
1117  *
1118  * Returns:
1119  *	Nothing
1120  */
1121 static void
1122 hash_free(struct uma_hash *hash)
1123 {
1124 	if (hash->uh_slab_hash == NULL)
1125 		return;
1126 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1127 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1128 	else
1129 		free(hash->uh_slab_hash, M_UMAHASH);
1130 }
1131 
1132 /*
1133  * Frees all outstanding items in a bucket
1134  *
1135  * Arguments:
1136  *	zone   The zone to free to, must be unlocked.
1137  *	bucket The free/alloc bucket with items.
1138  *
1139  * Returns:
1140  *	Nothing
1141  */
1142 static void
1143 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1144 {
1145 	int i;
1146 
1147 	if (bucket->ub_cnt == 0)
1148 		return;
1149 
1150 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1151 	    bucket->ub_seq != SMR_SEQ_INVALID) {
1152 		smr_wait(zone->uz_smr, bucket->ub_seq);
1153 		bucket->ub_seq = SMR_SEQ_INVALID;
1154 		for (i = 0; i < bucket->ub_cnt; i++)
1155 			item_dtor(zone, bucket->ub_bucket[i],
1156 			    zone->uz_size, NULL, SKIP_NONE);
1157 	}
1158 	if (zone->uz_fini)
1159 		for (i = 0; i < bucket->ub_cnt; i++)
1160 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1161 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1162 	if (zone->uz_max_items > 0)
1163 		zone_free_limit(zone, bucket->ub_cnt);
1164 #ifdef INVARIANTS
1165 	bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1166 #endif
1167 	bucket->ub_cnt = 0;
1168 }
1169 
1170 /*
1171  * Drains the per cpu caches for a zone.
1172  *
1173  * NOTE: This may only be called while the zone is being torn down, and not
1174  * during normal operation.  This is necessary in order that we do not have
1175  * to migrate CPUs to drain the per-CPU caches.
1176  *
1177  * Arguments:
1178  *	zone     The zone to drain, must be unlocked.
1179  *
1180  * Returns:
1181  *	Nothing
1182  */
1183 static void
1184 cache_drain(uma_zone_t zone)
1185 {
1186 	uma_cache_t cache;
1187 	uma_bucket_t bucket;
1188 	smr_seq_t seq;
1189 	int cpu;
1190 
1191 	/*
1192 	 * XXX: It is safe to not lock the per-CPU caches, because we're
1193 	 * tearing down the zone anyway.  I.e., there will be no further use
1194 	 * of the caches at this point.
1195 	 *
1196 	 * XXX: It would good to be able to assert that the zone is being
1197 	 * torn down to prevent improper use of cache_drain().
1198 	 */
1199 	seq = SMR_SEQ_INVALID;
1200 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1201 		seq = smr_advance(zone->uz_smr);
1202 	CPU_FOREACH(cpu) {
1203 		cache = &zone->uz_cpu[cpu];
1204 		bucket = cache_bucket_unload_alloc(cache);
1205 		if (bucket != NULL)
1206 			bucket_free(zone, bucket, NULL);
1207 		bucket = cache_bucket_unload_free(cache);
1208 		if (bucket != NULL) {
1209 			bucket->ub_seq = seq;
1210 			bucket_free(zone, bucket, NULL);
1211 		}
1212 		bucket = cache_bucket_unload_cross(cache);
1213 		if (bucket != NULL) {
1214 			bucket->ub_seq = seq;
1215 			bucket_free(zone, bucket, NULL);
1216 		}
1217 	}
1218 	bucket_cache_reclaim(zone, true);
1219 }
1220 
1221 static void
1222 cache_shrink(uma_zone_t zone, void *unused)
1223 {
1224 
1225 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1226 		return;
1227 
1228 	zone->uz_bucket_size =
1229 	    (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1230 }
1231 
1232 static void
1233 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1234 {
1235 	uma_cache_t cache;
1236 	uma_bucket_t b1, b2, b3;
1237 	int domain;
1238 
1239 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1240 		return;
1241 
1242 	b1 = b2 = b3 = NULL;
1243 	critical_enter();
1244 	cache = &zone->uz_cpu[curcpu];
1245 	domain = PCPU_GET(domain);
1246 	b1 = cache_bucket_unload_alloc(cache);
1247 
1248 	/*
1249 	 * Don't flush SMR zone buckets.  This leaves the zone without a
1250 	 * bucket and forces every free to synchronize().
1251 	 */
1252 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1253 		b2 = cache_bucket_unload_free(cache);
1254 		b3 = cache_bucket_unload_cross(cache);
1255 	}
1256 	critical_exit();
1257 
1258 	if (b1 != NULL)
1259 		zone_free_bucket(zone, b1, NULL, domain, false);
1260 	if (b2 != NULL)
1261 		zone_free_bucket(zone, b2, NULL, domain, false);
1262 	if (b3 != NULL) {
1263 		/* Adjust the domain so it goes to zone_free_cross. */
1264 		domain = (domain + 1) % vm_ndomains;
1265 		zone_free_bucket(zone, b3, NULL, domain, false);
1266 	}
1267 }
1268 
1269 /*
1270  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1271  * This is an expensive call because it needs to bind to all CPUs
1272  * one by one and enter a critical section on each of them in order
1273  * to safely access their cache buckets.
1274  * Zone lock must not be held on call this function.
1275  */
1276 static void
1277 pcpu_cache_drain_safe(uma_zone_t zone)
1278 {
1279 	int cpu;
1280 
1281 	/*
1282 	 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1283 	 */
1284 	if (zone)
1285 		cache_shrink(zone, NULL);
1286 	else
1287 		zone_foreach(cache_shrink, NULL);
1288 
1289 	CPU_FOREACH(cpu) {
1290 		thread_lock(curthread);
1291 		sched_bind(curthread, cpu);
1292 		thread_unlock(curthread);
1293 
1294 		if (zone)
1295 			cache_drain_safe_cpu(zone, NULL);
1296 		else
1297 			zone_foreach(cache_drain_safe_cpu, NULL);
1298 	}
1299 	thread_lock(curthread);
1300 	sched_unbind(curthread);
1301 	thread_unlock(curthread);
1302 }
1303 
1304 /*
1305  * Reclaim cached buckets from a zone.  All buckets are reclaimed if the caller
1306  * requested a drain, otherwise the per-domain caches are trimmed to either
1307  * estimated working set size.
1308  */
1309 static void
1310 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1311 {
1312 	uma_zone_domain_t zdom;
1313 	uma_bucket_t bucket;
1314 	long target;
1315 	int i;
1316 
1317 	/*
1318 	 * Shrink the zone bucket size to ensure that the per-CPU caches
1319 	 * don't grow too large.
1320 	 */
1321 	if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1322 		zone->uz_bucket_size--;
1323 
1324 	for (i = 0; i < vm_ndomains; i++) {
1325 		/*
1326 		 * The cross bucket is partially filled and not part of
1327 		 * the item count.  Reclaim it individually here.
1328 		 */
1329 		zdom = ZDOM_GET(zone, i);
1330 		if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1331 			ZONE_CROSS_LOCK(zone);
1332 			bucket = zdom->uzd_cross;
1333 			zdom->uzd_cross = NULL;
1334 			ZONE_CROSS_UNLOCK(zone);
1335 			if (bucket != NULL)
1336 				bucket_free(zone, bucket, NULL);
1337 		}
1338 
1339 		/*
1340 		 * If we were asked to drain the zone, we are done only once
1341 		 * this bucket cache is empty.  Otherwise, we reclaim items in
1342 		 * excess of the zone's estimated working set size.  If the
1343 		 * difference nitems - imin is larger than the WSS estimate,
1344 		 * then the estimate will grow at the end of this interval and
1345 		 * we ignore the historical average.
1346 		 */
1347 		ZDOM_LOCK(zdom);
1348 		target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1349 		    zdom->uzd_imin);
1350 		while (zdom->uzd_nitems > target) {
1351 			bucket = zone_fetch_bucket(zone, zdom, true);
1352 			if (bucket == NULL)
1353 				break;
1354 			bucket_free(zone, bucket, NULL);
1355 			ZDOM_LOCK(zdom);
1356 		}
1357 		ZDOM_UNLOCK(zdom);
1358 	}
1359 }
1360 
1361 static void
1362 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1363 {
1364 	uint8_t *mem;
1365 	int i;
1366 	uint8_t flags;
1367 
1368 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1369 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1370 
1371 	mem = slab_data(slab, keg);
1372 	flags = slab->us_flags;
1373 	i = start;
1374 	if (keg->uk_fini != NULL) {
1375 		for (i--; i > -1; i--)
1376 #ifdef INVARIANTS
1377 		/*
1378 		 * trash_fini implies that dtor was trash_dtor. trash_fini
1379 		 * would check that memory hasn't been modified since free,
1380 		 * which executed trash_dtor.
1381 		 * That's why we need to run uma_dbg_kskip() check here,
1382 		 * albeit we don't make skip check for other init/fini
1383 		 * invocations.
1384 		 */
1385 		if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1386 		    keg->uk_fini != trash_fini)
1387 #endif
1388 			keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1389 	}
1390 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1391 		zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1392 		    NULL, SKIP_NONE);
1393 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1394 	uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1395 }
1396 
1397 static void
1398 keg_drain_domain(uma_keg_t keg, int domain)
1399 {
1400 	struct slabhead freeslabs;
1401 	uma_domain_t dom;
1402 	uma_slab_t slab, tmp;
1403 	uint32_t i, stofree, stokeep, partial;
1404 
1405 	dom = &keg->uk_domain[domain];
1406 	LIST_INIT(&freeslabs);
1407 
1408 	CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1409 	    keg->uk_name, keg, domain, dom->ud_free_items);
1410 
1411 	KEG_LOCK(keg, domain);
1412 
1413 	/*
1414 	 * Are the free items in partially allocated slabs sufficient to meet
1415 	 * the reserve? If not, compute the number of fully free slabs that must
1416 	 * be kept.
1417 	 */
1418 	partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1419 	if (partial < keg->uk_reserve) {
1420 		stokeep = min(dom->ud_free_slabs,
1421 		    howmany(keg->uk_reserve - partial, keg->uk_ipers));
1422 	} else {
1423 		stokeep = 0;
1424 	}
1425 	stofree = dom->ud_free_slabs - stokeep;
1426 
1427 	/*
1428 	 * Partition the free slabs into two sets: those that must be kept in
1429 	 * order to maintain the reserve, and those that may be released back to
1430 	 * the system.  Since one set may be much larger than the other,
1431 	 * populate the smaller of the two sets and swap them if necessary.
1432 	 */
1433 	for (i = min(stofree, stokeep); i > 0; i--) {
1434 		slab = LIST_FIRST(&dom->ud_free_slab);
1435 		LIST_REMOVE(slab, us_link);
1436 		LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1437 	}
1438 	if (stofree > stokeep)
1439 		LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1440 
1441 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1442 		LIST_FOREACH(slab, &freeslabs, us_link)
1443 			UMA_HASH_REMOVE(&keg->uk_hash, slab);
1444 	}
1445 	dom->ud_free_items -= stofree * keg->uk_ipers;
1446 	dom->ud_free_slabs -= stofree;
1447 	dom->ud_pages -= stofree * keg->uk_ppera;
1448 	KEG_UNLOCK(keg, domain);
1449 
1450 	LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1451 		keg_free_slab(keg, slab, keg->uk_ipers);
1452 }
1453 
1454 /*
1455  * Frees pages from a keg back to the system.  This is done on demand from
1456  * the pageout daemon.
1457  *
1458  * Returns nothing.
1459  */
1460 static void
1461 keg_drain(uma_keg_t keg)
1462 {
1463 	int i;
1464 
1465 	if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1466 		return;
1467 	for (i = 0; i < vm_ndomains; i++)
1468 		keg_drain_domain(keg, i);
1469 }
1470 
1471 static void
1472 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1473 {
1474 
1475 	/*
1476 	 * Set draining to interlock with zone_dtor() so we can release our
1477 	 * locks as we go.  Only dtor() should do a WAITOK call since it
1478 	 * is the only call that knows the structure will still be available
1479 	 * when it wakes up.
1480 	 */
1481 	ZONE_LOCK(zone);
1482 	while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1483 		if (waitok == M_NOWAIT)
1484 			goto out;
1485 		msleep(zone, &ZDOM_GET(zone, 0)->uzd_lock, PVM, "zonedrain",
1486 		    1);
1487 	}
1488 	zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1489 	ZONE_UNLOCK(zone);
1490 	bucket_cache_reclaim(zone, drain);
1491 
1492 	/*
1493 	 * The DRAINING flag protects us from being freed while
1494 	 * we're running.  Normally the uma_rwlock would protect us but we
1495 	 * must be able to release and acquire the right lock for each keg.
1496 	 */
1497 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1498 		keg_drain(zone->uz_keg);
1499 	ZONE_LOCK(zone);
1500 	zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1501 	wakeup(zone);
1502 out:
1503 	ZONE_UNLOCK(zone);
1504 }
1505 
1506 static void
1507 zone_drain(uma_zone_t zone, void *unused)
1508 {
1509 
1510 	zone_reclaim(zone, M_NOWAIT, true);
1511 }
1512 
1513 static void
1514 zone_trim(uma_zone_t zone, void *unused)
1515 {
1516 
1517 	zone_reclaim(zone, M_NOWAIT, false);
1518 }
1519 
1520 /*
1521  * Allocate a new slab for a keg and inserts it into the partial slab list.
1522  * The keg should be unlocked on entry.  If the allocation succeeds it will
1523  * be locked on return.
1524  *
1525  * Arguments:
1526  *	flags   Wait flags for the item initialization routine
1527  *	aflags  Wait flags for the slab allocation
1528  *
1529  * Returns:
1530  *	The slab that was allocated or NULL if there is no memory and the
1531  *	caller specified M_NOWAIT.
1532  */
1533 static uma_slab_t
1534 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1535     int aflags)
1536 {
1537 	uma_domain_t dom;
1538 	uma_alloc allocf;
1539 	uma_slab_t slab;
1540 	unsigned long size;
1541 	uint8_t *mem;
1542 	uint8_t sflags;
1543 	int i;
1544 
1545 	KASSERT(domain >= 0 && domain < vm_ndomains,
1546 	    ("keg_alloc_slab: domain %d out of range", domain));
1547 
1548 	allocf = keg->uk_allocf;
1549 	slab = NULL;
1550 	mem = NULL;
1551 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1552 		uma_hash_slab_t hslab;
1553 		hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1554 		    domain, aflags);
1555 		if (hslab == NULL)
1556 			goto fail;
1557 		slab = &hslab->uhs_slab;
1558 	}
1559 
1560 	/*
1561 	 * This reproduces the old vm_zone behavior of zero filling pages the
1562 	 * first time they are added to a zone.
1563 	 *
1564 	 * Malloced items are zeroed in uma_zalloc.
1565 	 */
1566 
1567 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1568 		aflags |= M_ZERO;
1569 	else
1570 		aflags &= ~M_ZERO;
1571 
1572 	if (keg->uk_flags & UMA_ZONE_NODUMP)
1573 		aflags |= M_NODUMP;
1574 
1575 	/* zone is passed for legacy reasons. */
1576 	size = keg->uk_ppera * PAGE_SIZE;
1577 	mem = allocf(zone, size, domain, &sflags, aflags);
1578 	if (mem == NULL) {
1579 		if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1580 			zone_free_item(slabzone(keg->uk_ipers),
1581 			    slab_tohashslab(slab), NULL, SKIP_NONE);
1582 		goto fail;
1583 	}
1584 	uma_total_inc(size);
1585 
1586 	/* For HASH zones all pages go to the same uma_domain. */
1587 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1588 		domain = 0;
1589 
1590 	/* Point the slab into the allocated memory */
1591 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1592 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
1593 	else
1594 		slab_tohashslab(slab)->uhs_data = mem;
1595 
1596 	if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1597 		for (i = 0; i < keg->uk_ppera; i++)
1598 			vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1599 			    zone, slab);
1600 
1601 	slab->us_freecount = keg->uk_ipers;
1602 	slab->us_flags = sflags;
1603 	slab->us_domain = domain;
1604 
1605 	BIT_FILL(keg->uk_ipers, &slab->us_free);
1606 #ifdef INVARIANTS
1607 	BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1608 #endif
1609 
1610 	if (keg->uk_init != NULL) {
1611 		for (i = 0; i < keg->uk_ipers; i++)
1612 			if (keg->uk_init(slab_item(slab, keg, i),
1613 			    keg->uk_size, flags) != 0)
1614 				break;
1615 		if (i != keg->uk_ipers) {
1616 			keg_free_slab(keg, slab, i);
1617 			goto fail;
1618 		}
1619 	}
1620 	KEG_LOCK(keg, domain);
1621 
1622 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1623 	    slab, keg->uk_name, keg);
1624 
1625 	if (keg->uk_flags & UMA_ZFLAG_HASH)
1626 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1627 
1628 	/*
1629 	 * If we got a slab here it's safe to mark it partially used
1630 	 * and return.  We assume that the caller is going to remove
1631 	 * at least one item.
1632 	 */
1633 	dom = &keg->uk_domain[domain];
1634 	LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1635 	dom->ud_pages += keg->uk_ppera;
1636 	dom->ud_free_items += keg->uk_ipers;
1637 
1638 	return (slab);
1639 
1640 fail:
1641 	return (NULL);
1642 }
1643 
1644 /*
1645  * This function is intended to be used early on in place of page_alloc() so
1646  * that we may use the boot time page cache to satisfy allocations before
1647  * the VM is ready.
1648  */
1649 static void *
1650 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1651     int wait)
1652 {
1653 	vm_paddr_t pa;
1654 	vm_page_t m;
1655 	void *mem;
1656 	int pages;
1657 	int i;
1658 
1659 	pages = howmany(bytes, PAGE_SIZE);
1660 	KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1661 
1662 	*pflag = UMA_SLAB_BOOT;
1663 	m = vm_page_alloc_contig_domain(NULL, 0, domain,
1664 	    malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1665 	    (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1666 	if (m == NULL)
1667 		return (NULL);
1668 
1669 	pa = VM_PAGE_TO_PHYS(m);
1670 	for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1671 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1672     defined(__riscv) || defined(__powerpc64__)
1673 		if ((wait & M_NODUMP) == 0)
1674 			dump_add_page(pa);
1675 #endif
1676 	}
1677 	/* Allocate KVA and indirectly advance bootmem. */
1678 	mem = (void *)pmap_map(&bootmem, m->phys_addr,
1679 	    m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1680         if ((wait & M_ZERO) != 0)
1681                 bzero(mem, pages * PAGE_SIZE);
1682 
1683         return (mem);
1684 }
1685 
1686 static void
1687 startup_free(void *mem, vm_size_t bytes)
1688 {
1689 	vm_offset_t va;
1690 	vm_page_t m;
1691 
1692 	va = (vm_offset_t)mem;
1693 	m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1694 
1695 	/*
1696 	 * startup_alloc() returns direct-mapped slabs on some platforms.  Avoid
1697 	 * unmapping ranges of the direct map.
1698 	 */
1699 	if (va >= bootstart && va + bytes <= bootmem)
1700 		pmap_remove(kernel_pmap, va, va + bytes);
1701 	for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1702 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1703     defined(__riscv) || defined(__powerpc64__)
1704 		dump_drop_page(VM_PAGE_TO_PHYS(m));
1705 #endif
1706 		vm_page_unwire_noq(m);
1707 		vm_page_free(m);
1708 	}
1709 }
1710 
1711 /*
1712  * Allocates a number of pages from the system
1713  *
1714  * Arguments:
1715  *	bytes  The number of bytes requested
1716  *	wait  Shall we wait?
1717  *
1718  * Returns:
1719  *	A pointer to the alloced memory or possibly
1720  *	NULL if M_NOWAIT is set.
1721  */
1722 static void *
1723 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1724     int wait)
1725 {
1726 	void *p;	/* Returned page */
1727 
1728 	*pflag = UMA_SLAB_KERNEL;
1729 	p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1730 
1731 	return (p);
1732 }
1733 
1734 static void *
1735 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1736     int wait)
1737 {
1738 	struct pglist alloctail;
1739 	vm_offset_t addr, zkva;
1740 	int cpu, flags;
1741 	vm_page_t p, p_next;
1742 #ifdef NUMA
1743 	struct pcpu *pc;
1744 #endif
1745 
1746 	MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1747 
1748 	TAILQ_INIT(&alloctail);
1749 	flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1750 	    malloc2vm_flags(wait);
1751 	*pflag = UMA_SLAB_KERNEL;
1752 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1753 		if (CPU_ABSENT(cpu)) {
1754 			p = vm_page_alloc(NULL, 0, flags);
1755 		} else {
1756 #ifndef NUMA
1757 			p = vm_page_alloc(NULL, 0, flags);
1758 #else
1759 			pc = pcpu_find(cpu);
1760 			if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1761 				p = NULL;
1762 			else
1763 				p = vm_page_alloc_domain(NULL, 0,
1764 				    pc->pc_domain, flags);
1765 			if (__predict_false(p == NULL))
1766 				p = vm_page_alloc(NULL, 0, flags);
1767 #endif
1768 		}
1769 		if (__predict_false(p == NULL))
1770 			goto fail;
1771 		TAILQ_INSERT_TAIL(&alloctail, p, listq);
1772 	}
1773 	if ((addr = kva_alloc(bytes)) == 0)
1774 		goto fail;
1775 	zkva = addr;
1776 	TAILQ_FOREACH(p, &alloctail, listq) {
1777 		pmap_qenter(zkva, &p, 1);
1778 		zkva += PAGE_SIZE;
1779 	}
1780 	return ((void*)addr);
1781 fail:
1782 	TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1783 		vm_page_unwire_noq(p);
1784 		vm_page_free(p);
1785 	}
1786 	return (NULL);
1787 }
1788 
1789 /*
1790  * Allocates a number of pages from within an object
1791  *
1792  * Arguments:
1793  *	bytes  The number of bytes requested
1794  *	wait   Shall we wait?
1795  *
1796  * Returns:
1797  *	A pointer to the alloced memory or possibly
1798  *	NULL if M_NOWAIT is set.
1799  */
1800 static void *
1801 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1802     int wait)
1803 {
1804 	TAILQ_HEAD(, vm_page) alloctail;
1805 	u_long npages;
1806 	vm_offset_t retkva, zkva;
1807 	vm_page_t p, p_next;
1808 	uma_keg_t keg;
1809 
1810 	TAILQ_INIT(&alloctail);
1811 	keg = zone->uz_keg;
1812 
1813 	npages = howmany(bytes, PAGE_SIZE);
1814 	while (npages > 0) {
1815 		p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1816 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1817 		    ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1818 		    VM_ALLOC_NOWAIT));
1819 		if (p != NULL) {
1820 			/*
1821 			 * Since the page does not belong to an object, its
1822 			 * listq is unused.
1823 			 */
1824 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1825 			npages--;
1826 			continue;
1827 		}
1828 		/*
1829 		 * Page allocation failed, free intermediate pages and
1830 		 * exit.
1831 		 */
1832 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1833 			vm_page_unwire_noq(p);
1834 			vm_page_free(p);
1835 		}
1836 		return (NULL);
1837 	}
1838 	*flags = UMA_SLAB_PRIV;
1839 	zkva = keg->uk_kva +
1840 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1841 	retkva = zkva;
1842 	TAILQ_FOREACH(p, &alloctail, listq) {
1843 		pmap_qenter(zkva, &p, 1);
1844 		zkva += PAGE_SIZE;
1845 	}
1846 
1847 	return ((void *)retkva);
1848 }
1849 
1850 /*
1851  * Allocate physically contiguous pages.
1852  */
1853 static void *
1854 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1855     int wait)
1856 {
1857 
1858 	*pflag = UMA_SLAB_KERNEL;
1859 	return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1860 	    bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1861 }
1862 
1863 /*
1864  * Frees a number of pages to the system
1865  *
1866  * Arguments:
1867  *	mem   A pointer to the memory to be freed
1868  *	size  The size of the memory being freed
1869  *	flags The original p->us_flags field
1870  *
1871  * Returns:
1872  *	Nothing
1873  */
1874 static void
1875 page_free(void *mem, vm_size_t size, uint8_t flags)
1876 {
1877 
1878 	if ((flags & UMA_SLAB_BOOT) != 0) {
1879 		startup_free(mem, size);
1880 		return;
1881 	}
1882 
1883 	KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1884 	    ("UMA: page_free used with invalid flags %x", flags));
1885 
1886 	kmem_free((vm_offset_t)mem, size);
1887 }
1888 
1889 /*
1890  * Frees pcpu zone allocations
1891  *
1892  * Arguments:
1893  *	mem   A pointer to the memory to be freed
1894  *	size  The size of the memory being freed
1895  *	flags The original p->us_flags field
1896  *
1897  * Returns:
1898  *	Nothing
1899  */
1900 static void
1901 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1902 {
1903 	vm_offset_t sva, curva;
1904 	vm_paddr_t paddr;
1905 	vm_page_t m;
1906 
1907 	MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1908 
1909 	if ((flags & UMA_SLAB_BOOT) != 0) {
1910 		startup_free(mem, size);
1911 		return;
1912 	}
1913 
1914 	sva = (vm_offset_t)mem;
1915 	for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1916 		paddr = pmap_kextract(curva);
1917 		m = PHYS_TO_VM_PAGE(paddr);
1918 		vm_page_unwire_noq(m);
1919 		vm_page_free(m);
1920 	}
1921 	pmap_qremove(sva, size >> PAGE_SHIFT);
1922 	kva_free(sva, size);
1923 }
1924 
1925 /*
1926  * Zero fill initializer
1927  *
1928  * Arguments/Returns follow uma_init specifications
1929  */
1930 static int
1931 zero_init(void *mem, int size, int flags)
1932 {
1933 	bzero(mem, size);
1934 	return (0);
1935 }
1936 
1937 #ifdef INVARIANTS
1938 static struct noslabbits *
1939 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1940 {
1941 
1942 	return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1943 }
1944 #endif
1945 
1946 /*
1947  * Actual size of embedded struct slab (!OFFPAGE).
1948  */
1949 static size_t
1950 slab_sizeof(int nitems)
1951 {
1952 	size_t s;
1953 
1954 	s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1955 	return (roundup(s, UMA_ALIGN_PTR + 1));
1956 }
1957 
1958 #define	UMA_FIXPT_SHIFT	31
1959 #define	UMA_FRAC_FIXPT(n, d)						\
1960 	((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1961 #define	UMA_FIXPT_PCT(f)						\
1962 	((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1963 #define	UMA_PCT_FIXPT(pct)	UMA_FRAC_FIXPT((pct), 100)
1964 #define	UMA_MIN_EFF	UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1965 
1966 /*
1967  * Compute the number of items that will fit in a slab.  If hdr is true, the
1968  * item count may be limited to provide space in the slab for an inline slab
1969  * header.  Otherwise, all slab space will be provided for item storage.
1970  */
1971 static u_int
1972 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1973 {
1974 	u_int ipers;
1975 	u_int padpi;
1976 
1977 	/* The padding between items is not needed after the last item. */
1978 	padpi = rsize - size;
1979 
1980 	if (hdr) {
1981 		/*
1982 		 * Start with the maximum item count and remove items until
1983 		 * the slab header first alongside the allocatable memory.
1984 		 */
1985 		for (ipers = MIN(SLAB_MAX_SETSIZE,
1986 		    (slabsize + padpi - slab_sizeof(1)) / rsize);
1987 		    ipers > 0 &&
1988 		    ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1989 		    ipers--)
1990 			continue;
1991 	} else {
1992 		ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1993 	}
1994 
1995 	return (ipers);
1996 }
1997 
1998 struct keg_layout_result {
1999 	u_int format;
2000 	u_int slabsize;
2001 	u_int ipers;
2002 	u_int eff;
2003 };
2004 
2005 static void
2006 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2007     struct keg_layout_result *kl)
2008 {
2009 	u_int total;
2010 
2011 	kl->format = fmt;
2012 	kl->slabsize = slabsize;
2013 
2014 	/* Handle INTERNAL as inline with an extra page. */
2015 	if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2016 		kl->format &= ~UMA_ZFLAG_INTERNAL;
2017 		kl->slabsize += PAGE_SIZE;
2018 	}
2019 
2020 	kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2021 	    (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2022 
2023 	/* Account for memory used by an offpage slab header. */
2024 	total = kl->slabsize;
2025 	if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2026 		total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2027 
2028 	kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2029 }
2030 
2031 /*
2032  * Determine the format of a uma keg.  This determines where the slab header
2033  * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2034  *
2035  * Arguments
2036  *	keg  The zone we should initialize
2037  *
2038  * Returns
2039  *	Nothing
2040  */
2041 static void
2042 keg_layout(uma_keg_t keg)
2043 {
2044 	struct keg_layout_result kl = {}, kl_tmp;
2045 	u_int fmts[2];
2046 	u_int alignsize;
2047 	u_int nfmt;
2048 	u_int pages;
2049 	u_int rsize;
2050 	u_int slabsize;
2051 	u_int i, j;
2052 
2053 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2054 	    (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2055 	     (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2056 	    ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2057 	     __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2058 	     PRINT_UMA_ZFLAGS));
2059 	KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2060 	    (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2061 	    ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2062 	     PRINT_UMA_ZFLAGS));
2063 
2064 	alignsize = keg->uk_align + 1;
2065 
2066 	/*
2067 	 * Calculate the size of each allocation (rsize) according to
2068 	 * alignment.  If the requested size is smaller than we have
2069 	 * allocation bits for we round it up.
2070 	 */
2071 	rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2072 	rsize = roundup2(rsize, alignsize);
2073 
2074 	if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2075 		/*
2076 		 * We want one item to start on every align boundary in a page.
2077 		 * To do this we will span pages.  We will also extend the item
2078 		 * by the size of align if it is an even multiple of align.
2079 		 * Otherwise, it would fall on the same boundary every time.
2080 		 */
2081 		if ((rsize & alignsize) == 0)
2082 			rsize += alignsize;
2083 		slabsize = rsize * (PAGE_SIZE / alignsize);
2084 		slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2085 		slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2086 		slabsize = round_page(slabsize);
2087 	} else {
2088 		/*
2089 		 * Start with a slab size of as many pages as it takes to
2090 		 * represent a single item.  We will try to fit as many
2091 		 * additional items into the slab as possible.
2092 		 */
2093 		slabsize = round_page(keg->uk_size);
2094 	}
2095 
2096 	/* Build a list of all of the available formats for this keg. */
2097 	nfmt = 0;
2098 
2099 	/* Evaluate an inline slab layout. */
2100 	if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2101 		fmts[nfmt++] = 0;
2102 
2103 	/* TODO: vm_page-embedded slab. */
2104 
2105 	/*
2106 	 * We can't do OFFPAGE if we're internal or if we've been
2107 	 * asked to not go to the VM for buckets.  If we do this we
2108 	 * may end up going to the VM for slabs which we do not want
2109 	 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2110 	 * In those cases, evaluate a pseudo-format called INTERNAL
2111 	 * which has an inline slab header and one extra page to
2112 	 * guarantee that it fits.
2113 	 *
2114 	 * Otherwise, see if using an OFFPAGE slab will improve our
2115 	 * efficiency.
2116 	 */
2117 	if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2118 		fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2119 	else
2120 		fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2121 
2122 	/*
2123 	 * Choose a slab size and format which satisfy the minimum efficiency.
2124 	 * Prefer the smallest slab size that meets the constraints.
2125 	 *
2126 	 * Start with a minimum slab size, to accommodate CACHESPREAD.  Then,
2127 	 * for small items (up to PAGE_SIZE), the iteration increment is one
2128 	 * page; and for large items, the increment is one item.
2129 	 */
2130 	i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2131 	KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2132 	    keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2133 	    rsize, i));
2134 	for ( ; ; i++) {
2135 		slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2136 		    round_page(rsize * (i - 1) + keg->uk_size);
2137 
2138 		for (j = 0; j < nfmt; j++) {
2139 			/* Only if we have no viable format yet. */
2140 			if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2141 			    kl.ipers > 0)
2142 				continue;
2143 
2144 			keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2145 			if (kl_tmp.eff <= kl.eff)
2146 				continue;
2147 
2148 			kl = kl_tmp;
2149 
2150 			CTR6(KTR_UMA, "keg %s layout: format %#x "
2151 			    "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2152 			    keg->uk_name, kl.format, kl.ipers, rsize,
2153 			    kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2154 
2155 			/* Stop when we reach the minimum efficiency. */
2156 			if (kl.eff >= UMA_MIN_EFF)
2157 				break;
2158 		}
2159 
2160 		if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2161 		    slabsize >= SLAB_MAX_SETSIZE * rsize ||
2162 		    (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2163 			break;
2164 	}
2165 
2166 	pages = atop(kl.slabsize);
2167 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2168 		pages *= mp_maxid + 1;
2169 
2170 	keg->uk_rsize = rsize;
2171 	keg->uk_ipers = kl.ipers;
2172 	keg->uk_ppera = pages;
2173 	keg->uk_flags |= kl.format;
2174 
2175 	/*
2176 	 * How do we find the slab header if it is offpage or if not all item
2177 	 * start addresses are in the same page?  We could solve the latter
2178 	 * case with vaddr alignment, but we don't.
2179 	 */
2180 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2181 	    (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2182 		if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2183 			keg->uk_flags |= UMA_ZFLAG_HASH;
2184 		else
2185 			keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2186 	}
2187 
2188 	CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2189 	    __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2190 	    pages);
2191 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2192 	    ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2193 	     keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2194 	     keg->uk_ipers, pages));
2195 }
2196 
2197 /*
2198  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
2199  * the keg onto the global keg list.
2200  *
2201  * Arguments/Returns follow uma_ctor specifications
2202  *	udata  Actually uma_kctor_args
2203  */
2204 static int
2205 keg_ctor(void *mem, int size, void *udata, int flags)
2206 {
2207 	struct uma_kctor_args *arg = udata;
2208 	uma_keg_t keg = mem;
2209 	uma_zone_t zone;
2210 	int i;
2211 
2212 	bzero(keg, size);
2213 	keg->uk_size = arg->size;
2214 	keg->uk_init = arg->uminit;
2215 	keg->uk_fini = arg->fini;
2216 	keg->uk_align = arg->align;
2217 	keg->uk_reserve = 0;
2218 	keg->uk_flags = arg->flags;
2219 
2220 	/*
2221 	 * We use a global round-robin policy by default.  Zones with
2222 	 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2223 	 * case the iterator is never run.
2224 	 */
2225 	keg->uk_dr.dr_policy = DOMAINSET_RR();
2226 	keg->uk_dr.dr_iter = 0;
2227 
2228 	/*
2229 	 * The primary zone is passed to us at keg-creation time.
2230 	 */
2231 	zone = arg->zone;
2232 	keg->uk_name = zone->uz_name;
2233 
2234 	if (arg->flags & UMA_ZONE_ZINIT)
2235 		keg->uk_init = zero_init;
2236 
2237 	if (arg->flags & UMA_ZONE_MALLOC)
2238 		keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2239 
2240 #ifndef SMP
2241 	keg->uk_flags &= ~UMA_ZONE_PCPU;
2242 #endif
2243 
2244 	keg_layout(keg);
2245 
2246 	/*
2247 	 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2248 	 * work on.  Use round-robin for everything else.
2249 	 *
2250 	 * Zones may override the default by specifying either.
2251 	 */
2252 #ifdef NUMA
2253 	if ((keg->uk_flags &
2254 	    (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2255 		keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2256 	else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2257 		keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2258 #endif
2259 
2260 	/*
2261 	 * If we haven't booted yet we need allocations to go through the
2262 	 * startup cache until the vm is ready.
2263 	 */
2264 #ifdef UMA_MD_SMALL_ALLOC
2265 	if (keg->uk_ppera == 1)
2266 		keg->uk_allocf = uma_small_alloc;
2267 	else
2268 #endif
2269 	if (booted < BOOT_KVA)
2270 		keg->uk_allocf = startup_alloc;
2271 	else if (keg->uk_flags & UMA_ZONE_PCPU)
2272 		keg->uk_allocf = pcpu_page_alloc;
2273 	else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2274 		keg->uk_allocf = contig_alloc;
2275 	else
2276 		keg->uk_allocf = page_alloc;
2277 #ifdef UMA_MD_SMALL_ALLOC
2278 	if (keg->uk_ppera == 1)
2279 		keg->uk_freef = uma_small_free;
2280 	else
2281 #endif
2282 	if (keg->uk_flags & UMA_ZONE_PCPU)
2283 		keg->uk_freef = pcpu_page_free;
2284 	else
2285 		keg->uk_freef = page_free;
2286 
2287 	/*
2288 	 * Initialize keg's locks.
2289 	 */
2290 	for (i = 0; i < vm_ndomains; i++)
2291 		KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2292 
2293 	/*
2294 	 * If we're putting the slab header in the actual page we need to
2295 	 * figure out where in each page it goes.  See slab_sizeof
2296 	 * definition.
2297 	 */
2298 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2299 		size_t shsize;
2300 
2301 		shsize = slab_sizeof(keg->uk_ipers);
2302 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2303 		/*
2304 		 * The only way the following is possible is if with our
2305 		 * UMA_ALIGN_PTR adjustments we are now bigger than
2306 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
2307 		 * mathematically possible for all cases, so we make
2308 		 * sure here anyway.
2309 		 */
2310 		KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2311 		    ("zone %s ipers %d rsize %d size %d slab won't fit",
2312 		    zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2313 	}
2314 
2315 	if (keg->uk_flags & UMA_ZFLAG_HASH)
2316 		hash_alloc(&keg->uk_hash, 0);
2317 
2318 	CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2319 
2320 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2321 
2322 	rw_wlock(&uma_rwlock);
2323 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2324 	rw_wunlock(&uma_rwlock);
2325 	return (0);
2326 }
2327 
2328 static void
2329 zone_kva_available(uma_zone_t zone, void *unused)
2330 {
2331 	uma_keg_t keg;
2332 
2333 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2334 		return;
2335 	KEG_GET(zone, keg);
2336 
2337 	if (keg->uk_allocf == startup_alloc) {
2338 		/* Switch to the real allocator. */
2339 		if (keg->uk_flags & UMA_ZONE_PCPU)
2340 			keg->uk_allocf = pcpu_page_alloc;
2341 		else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2342 		    keg->uk_ppera > 1)
2343 			keg->uk_allocf = contig_alloc;
2344 		else
2345 			keg->uk_allocf = page_alloc;
2346 	}
2347 }
2348 
2349 static void
2350 zone_alloc_counters(uma_zone_t zone, void *unused)
2351 {
2352 
2353 	zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2354 	zone->uz_frees = counter_u64_alloc(M_WAITOK);
2355 	zone->uz_fails = counter_u64_alloc(M_WAITOK);
2356 	zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2357 }
2358 
2359 static void
2360 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2361 {
2362 	uma_zone_domain_t zdom;
2363 	uma_domain_t dom;
2364 	uma_keg_t keg;
2365 	struct sysctl_oid *oid, *domainoid;
2366 	int domains, i, cnt;
2367 	static const char *nokeg = "cache zone";
2368 	char *c;
2369 
2370 	/*
2371 	 * Make a sysctl safe copy of the zone name by removing
2372 	 * any special characters and handling dups by appending
2373 	 * an index.
2374 	 */
2375 	if (zone->uz_namecnt != 0) {
2376 		/* Count the number of decimal digits and '_' separator. */
2377 		for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2378 			cnt /= 10;
2379 		zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2380 		    M_UMA, M_WAITOK);
2381 		sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2382 		    zone->uz_namecnt);
2383 	} else
2384 		zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2385 	for (c = zone->uz_ctlname; *c != '\0'; c++)
2386 		if (strchr("./\\ -", *c) != NULL)
2387 			*c = '_';
2388 
2389 	/*
2390 	 * Basic parameters at the root.
2391 	 */
2392 	zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2393 	    OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2394 	oid = zone->uz_oid;
2395 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2396 	    "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2397 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2398 	    "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2399 	    zone, 0, sysctl_handle_uma_zone_flags, "A",
2400 	    "Allocator configuration flags");
2401 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2402 	    "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2403 	    "Desired per-cpu cache size");
2404 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2405 	    "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2406 	    "Maximum allowed per-cpu cache size");
2407 
2408 	/*
2409 	 * keg if present.
2410 	 */
2411 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2412 		domains = vm_ndomains;
2413 	else
2414 		domains = 1;
2415 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2416 	    "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2417 	keg = zone->uz_keg;
2418 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2419 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2420 		    "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2421 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2422 		    "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2423 		    "Real object size with alignment");
2424 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2425 		    "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2426 		    "pages per-slab allocation");
2427 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2428 		    "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2429 		    "items available per-slab");
2430 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2431 		    "align", CTLFLAG_RD, &keg->uk_align, 0,
2432 		    "item alignment mask");
2433 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2434 		    "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2435 		    "number of reserved items");
2436 		SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2437 		    "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2438 		    keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2439 		    "Slab utilization (100 - internal fragmentation %)");
2440 		domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2441 		    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2442 		for (i = 0; i < domains; i++) {
2443 			dom = &keg->uk_domain[i];
2444 			oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2445 			    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2446 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2447 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2448 			    "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2449 			    "Total pages currently allocated from VM");
2450 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2451 			    "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2452 			    "items free in the slab layer");
2453 		}
2454 	} else
2455 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2456 		    "name", CTLFLAG_RD, nokeg, "Keg name");
2457 
2458 	/*
2459 	 * Information about zone limits.
2460 	 */
2461 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2462 	    "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2463 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2464 	    "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2465 	    zone, 0, sysctl_handle_uma_zone_items, "QU",
2466 	    "Current number of allocated items if limit is set");
2467 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2468 	    "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2469 	    "Maximum number of allocated and cached items");
2470 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2471 	    "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2472 	    "Number of threads sleeping at limit");
2473 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2474 	    "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2475 	    "Total zone limit sleeps");
2476 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2477 	    "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2478 	    "Maximum number of items in each domain's bucket cache");
2479 
2480 	/*
2481 	 * Per-domain zone information.
2482 	 */
2483 	domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2484 	    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2485 	for (i = 0; i < domains; i++) {
2486 		zdom = ZDOM_GET(zone, i);
2487 		oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2488 		    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2489 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2490 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2491 		    "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2492 		    "number of items in this domain");
2493 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2494 		    "imax", CTLFLAG_RD, &zdom->uzd_imax,
2495 		    "maximum item count in this period");
2496 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2497 		    "imin", CTLFLAG_RD, &zdom->uzd_imin,
2498 		    "minimum item count in this period");
2499 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2500 		    "wss", CTLFLAG_RD, &zdom->uzd_wss,
2501 		    "Working set size");
2502 	}
2503 
2504 	/*
2505 	 * General statistics.
2506 	 */
2507 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2508 	    "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2509 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2510 	    "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2511 	    zone, 1, sysctl_handle_uma_zone_cur, "I",
2512 	    "Current number of allocated items");
2513 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2514 	    "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2515 	    zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2516 	    "Total allocation calls");
2517 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2518 	    "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2519 	    zone, 0, sysctl_handle_uma_zone_frees, "QU",
2520 	    "Total free calls");
2521 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2522 	    "fails", CTLFLAG_RD, &zone->uz_fails,
2523 	    "Number of allocation failures");
2524 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2525 	    "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2526 	    "Free calls from the wrong domain");
2527 }
2528 
2529 struct uma_zone_count {
2530 	const char	*name;
2531 	int		count;
2532 };
2533 
2534 static void
2535 zone_count(uma_zone_t zone, void *arg)
2536 {
2537 	struct uma_zone_count *cnt;
2538 
2539 	cnt = arg;
2540 	/*
2541 	 * Some zones are rapidly created with identical names and
2542 	 * destroyed out of order.  This can lead to gaps in the count.
2543 	 * Use one greater than the maximum observed for this name.
2544 	 */
2545 	if (strcmp(zone->uz_name, cnt->name) == 0)
2546 		cnt->count = MAX(cnt->count,
2547 		    zone->uz_namecnt + 1);
2548 }
2549 
2550 static void
2551 zone_update_caches(uma_zone_t zone)
2552 {
2553 	int i;
2554 
2555 	for (i = 0; i <= mp_maxid; i++) {
2556 		cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2557 		cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2558 	}
2559 }
2560 
2561 /*
2562  * Zone header ctor.  This initializes all fields, locks, etc.
2563  *
2564  * Arguments/Returns follow uma_ctor specifications
2565  *	udata  Actually uma_zctor_args
2566  */
2567 static int
2568 zone_ctor(void *mem, int size, void *udata, int flags)
2569 {
2570 	struct uma_zone_count cnt;
2571 	struct uma_zctor_args *arg = udata;
2572 	uma_zone_domain_t zdom;
2573 	uma_zone_t zone = mem;
2574 	uma_zone_t z;
2575 	uma_keg_t keg;
2576 	int i;
2577 
2578 	bzero(zone, size);
2579 	zone->uz_name = arg->name;
2580 	zone->uz_ctor = arg->ctor;
2581 	zone->uz_dtor = arg->dtor;
2582 	zone->uz_init = NULL;
2583 	zone->uz_fini = NULL;
2584 	zone->uz_sleeps = 0;
2585 	zone->uz_bucket_size = 0;
2586 	zone->uz_bucket_size_min = 0;
2587 	zone->uz_bucket_size_max = BUCKET_MAX;
2588 	zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2589 	zone->uz_warning = NULL;
2590 	/* The domain structures follow the cpu structures. */
2591 	zone->uz_bucket_max = ULONG_MAX;
2592 	timevalclear(&zone->uz_ratecheck);
2593 
2594 	/* Count the number of duplicate names. */
2595 	cnt.name = arg->name;
2596 	cnt.count = 0;
2597 	zone_foreach(zone_count, &cnt);
2598 	zone->uz_namecnt = cnt.count;
2599 	ZONE_CROSS_LOCK_INIT(zone);
2600 
2601 	for (i = 0; i < vm_ndomains; i++) {
2602 		zdom = ZDOM_GET(zone, i);
2603 		ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2604 		STAILQ_INIT(&zdom->uzd_buckets);
2605 	}
2606 
2607 #ifdef INVARIANTS
2608 	if (arg->uminit == trash_init && arg->fini == trash_fini)
2609 		zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2610 #endif
2611 
2612 	/*
2613 	 * This is a pure cache zone, no kegs.
2614 	 */
2615 	if (arg->import) {
2616 		KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2617 		    ("zone_ctor: Import specified for non-cache zone."));
2618 		zone->uz_flags = arg->flags;
2619 		zone->uz_size = arg->size;
2620 		zone->uz_import = arg->import;
2621 		zone->uz_release = arg->release;
2622 		zone->uz_arg = arg->arg;
2623 #ifdef NUMA
2624 		/*
2625 		 * Cache zones are round-robin unless a policy is
2626 		 * specified because they may have incompatible
2627 		 * constraints.
2628 		 */
2629 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2630 			zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2631 #endif
2632 		rw_wlock(&uma_rwlock);
2633 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2634 		rw_wunlock(&uma_rwlock);
2635 		goto out;
2636 	}
2637 
2638 	/*
2639 	 * Use the regular zone/keg/slab allocator.
2640 	 */
2641 	zone->uz_import = zone_import;
2642 	zone->uz_release = zone_release;
2643 	zone->uz_arg = zone;
2644 	keg = arg->keg;
2645 
2646 	if (arg->flags & UMA_ZONE_SECONDARY) {
2647 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2648 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2649 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2650 		zone->uz_init = arg->uminit;
2651 		zone->uz_fini = arg->fini;
2652 		zone->uz_flags |= UMA_ZONE_SECONDARY;
2653 		rw_wlock(&uma_rwlock);
2654 		ZONE_LOCK(zone);
2655 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2656 			if (LIST_NEXT(z, uz_link) == NULL) {
2657 				LIST_INSERT_AFTER(z, zone, uz_link);
2658 				break;
2659 			}
2660 		}
2661 		ZONE_UNLOCK(zone);
2662 		rw_wunlock(&uma_rwlock);
2663 	} else if (keg == NULL) {
2664 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2665 		    arg->align, arg->flags)) == NULL)
2666 			return (ENOMEM);
2667 	} else {
2668 		struct uma_kctor_args karg;
2669 		int error;
2670 
2671 		/* We should only be here from uma_startup() */
2672 		karg.size = arg->size;
2673 		karg.uminit = arg->uminit;
2674 		karg.fini = arg->fini;
2675 		karg.align = arg->align;
2676 		karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2677 		karg.zone = zone;
2678 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2679 		    flags);
2680 		if (error)
2681 			return (error);
2682 	}
2683 
2684 	/* Inherit properties from the keg. */
2685 	zone->uz_keg = keg;
2686 	zone->uz_size = keg->uk_size;
2687 	zone->uz_flags |= (keg->uk_flags &
2688 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2689 
2690 out:
2691 	if (booted >= BOOT_PCPU) {
2692 		zone_alloc_counters(zone, NULL);
2693 		if (booted >= BOOT_RUNNING)
2694 			zone_alloc_sysctl(zone, NULL);
2695 	} else {
2696 		zone->uz_allocs = EARLY_COUNTER;
2697 		zone->uz_frees = EARLY_COUNTER;
2698 		zone->uz_fails = EARLY_COUNTER;
2699 	}
2700 
2701 	/* Caller requests a private SMR context. */
2702 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2703 		zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2704 
2705 	KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2706 	    (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2707 	    ("Invalid zone flag combination"));
2708 	if (arg->flags & UMA_ZFLAG_INTERNAL)
2709 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2710 	if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2711 		zone->uz_bucket_size = BUCKET_MAX;
2712 	else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2713 		zone->uz_bucket_size = 0;
2714 	else
2715 		zone->uz_bucket_size = bucket_select(zone->uz_size);
2716 	zone->uz_bucket_size_min = zone->uz_bucket_size;
2717 	if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2718 		zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2719 	zone_update_caches(zone);
2720 
2721 	return (0);
2722 }
2723 
2724 /*
2725  * Keg header dtor.  This frees all data, destroys locks, frees the hash
2726  * table and removes the keg from the global list.
2727  *
2728  * Arguments/Returns follow uma_dtor specifications
2729  *	udata  unused
2730  */
2731 static void
2732 keg_dtor(void *arg, int size, void *udata)
2733 {
2734 	uma_keg_t keg;
2735 	uint32_t free, pages;
2736 	int i;
2737 
2738 	keg = (uma_keg_t)arg;
2739 	free = pages = 0;
2740 	for (i = 0; i < vm_ndomains; i++) {
2741 		free += keg->uk_domain[i].ud_free_items;
2742 		pages += keg->uk_domain[i].ud_pages;
2743 		KEG_LOCK_FINI(keg, i);
2744 	}
2745 	if (pages != 0)
2746 		printf("Freed UMA keg (%s) was not empty (%u items). "
2747 		    " Lost %u pages of memory.\n",
2748 		    keg->uk_name ? keg->uk_name : "",
2749 		    pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2750 
2751 	hash_free(&keg->uk_hash);
2752 }
2753 
2754 /*
2755  * Zone header dtor.
2756  *
2757  * Arguments/Returns follow uma_dtor specifications
2758  *	udata  unused
2759  */
2760 static void
2761 zone_dtor(void *arg, int size, void *udata)
2762 {
2763 	uma_zone_t zone;
2764 	uma_keg_t keg;
2765 	int i;
2766 
2767 	zone = (uma_zone_t)arg;
2768 
2769 	sysctl_remove_oid(zone->uz_oid, 1, 1);
2770 
2771 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2772 		cache_drain(zone);
2773 
2774 	rw_wlock(&uma_rwlock);
2775 	LIST_REMOVE(zone, uz_link);
2776 	rw_wunlock(&uma_rwlock);
2777 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2778 		keg = zone->uz_keg;
2779 		keg->uk_reserve = 0;
2780 	}
2781 	zone_reclaim(zone, M_WAITOK, true);
2782 
2783 	/*
2784 	 * We only destroy kegs from non secondary/non cache zones.
2785 	 */
2786 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2787 		keg = zone->uz_keg;
2788 		rw_wlock(&uma_rwlock);
2789 		LIST_REMOVE(keg, uk_link);
2790 		rw_wunlock(&uma_rwlock);
2791 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
2792 	}
2793 	counter_u64_free(zone->uz_allocs);
2794 	counter_u64_free(zone->uz_frees);
2795 	counter_u64_free(zone->uz_fails);
2796 	counter_u64_free(zone->uz_xdomain);
2797 	free(zone->uz_ctlname, M_UMA);
2798 	for (i = 0; i < vm_ndomains; i++)
2799 		ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2800 	ZONE_CROSS_LOCK_FINI(zone);
2801 }
2802 
2803 static void
2804 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2805 {
2806 	uma_keg_t keg;
2807 	uma_zone_t zone;
2808 
2809 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
2810 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2811 			zfunc(zone, arg);
2812 	}
2813 	LIST_FOREACH(zone, &uma_cachezones, uz_link)
2814 		zfunc(zone, arg);
2815 }
2816 
2817 /*
2818  * Traverses every zone in the system and calls a callback
2819  *
2820  * Arguments:
2821  *	zfunc  A pointer to a function which accepts a zone
2822  *		as an argument.
2823  *
2824  * Returns:
2825  *	Nothing
2826  */
2827 static void
2828 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2829 {
2830 
2831 	rw_rlock(&uma_rwlock);
2832 	zone_foreach_unlocked(zfunc, arg);
2833 	rw_runlock(&uma_rwlock);
2834 }
2835 
2836 /*
2837  * Initialize the kernel memory allocator.  This is done after pages can be
2838  * allocated but before general KVA is available.
2839  */
2840 void
2841 uma_startup1(vm_offset_t virtual_avail)
2842 {
2843 	struct uma_zctor_args args;
2844 	size_t ksize, zsize, size;
2845 	uma_keg_t primarykeg;
2846 	uintptr_t m;
2847 	int domain;
2848 	uint8_t pflag;
2849 
2850 	bootstart = bootmem = virtual_avail;
2851 
2852 	rw_init(&uma_rwlock, "UMA lock");
2853 	sx_init(&uma_reclaim_lock, "umareclaim");
2854 
2855 	ksize = sizeof(struct uma_keg) +
2856 	    (sizeof(struct uma_domain) * vm_ndomains);
2857 	ksize = roundup(ksize, UMA_SUPER_ALIGN);
2858 	zsize = sizeof(struct uma_zone) +
2859 	    (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2860 	    (sizeof(struct uma_zone_domain) * vm_ndomains);
2861 	zsize = roundup(zsize, UMA_SUPER_ALIGN);
2862 
2863 	/* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2864 	size = (zsize * 2) + ksize;
2865 	for (domain = 0; domain < vm_ndomains; domain++) {
2866 		m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
2867 		    M_NOWAIT | M_ZERO);
2868 		if (m != 0)
2869 			break;
2870 	}
2871 	zones = (uma_zone_t)m;
2872 	m += zsize;
2873 	kegs = (uma_zone_t)m;
2874 	m += zsize;
2875 	primarykeg = (uma_keg_t)m;
2876 
2877 	/* "manually" create the initial zone */
2878 	memset(&args, 0, sizeof(args));
2879 	args.name = "UMA Kegs";
2880 	args.size = ksize;
2881 	args.ctor = keg_ctor;
2882 	args.dtor = keg_dtor;
2883 	args.uminit = zero_init;
2884 	args.fini = NULL;
2885 	args.keg = primarykeg;
2886 	args.align = UMA_SUPER_ALIGN - 1;
2887 	args.flags = UMA_ZFLAG_INTERNAL;
2888 	zone_ctor(kegs, zsize, &args, M_WAITOK);
2889 
2890 	args.name = "UMA Zones";
2891 	args.size = zsize;
2892 	args.ctor = zone_ctor;
2893 	args.dtor = zone_dtor;
2894 	args.uminit = zero_init;
2895 	args.fini = NULL;
2896 	args.keg = NULL;
2897 	args.align = UMA_SUPER_ALIGN - 1;
2898 	args.flags = UMA_ZFLAG_INTERNAL;
2899 	zone_ctor(zones, zsize, &args, M_WAITOK);
2900 
2901 	/* Now make zones for slab headers */
2902 	slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2903 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2904 	slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2905 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2906 
2907 	hashzone = uma_zcreate("UMA Hash",
2908 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2909 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2910 
2911 	bucket_init();
2912 	smr_init();
2913 }
2914 
2915 #ifndef UMA_MD_SMALL_ALLOC
2916 extern void vm_radix_reserve_kva(void);
2917 #endif
2918 
2919 /*
2920  * Advertise the availability of normal kva allocations and switch to
2921  * the default back-end allocator.  Marks the KVA we consumed on startup
2922  * as used in the map.
2923  */
2924 void
2925 uma_startup2(void)
2926 {
2927 
2928 	if (bootstart != bootmem) {
2929 		vm_map_lock(kernel_map);
2930 		(void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2931 		    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2932 		vm_map_unlock(kernel_map);
2933 	}
2934 
2935 #ifndef UMA_MD_SMALL_ALLOC
2936 	/* Set up radix zone to use noobj_alloc. */
2937 	vm_radix_reserve_kva();
2938 #endif
2939 
2940 	booted = BOOT_KVA;
2941 	zone_foreach_unlocked(zone_kva_available, NULL);
2942 	bucket_enable();
2943 }
2944 
2945 /*
2946  * Allocate counters as early as possible so that boot-time allocations are
2947  * accounted more precisely.
2948  */
2949 static void
2950 uma_startup_pcpu(void *arg __unused)
2951 {
2952 
2953 	zone_foreach_unlocked(zone_alloc_counters, NULL);
2954 	booted = BOOT_PCPU;
2955 }
2956 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
2957 
2958 /*
2959  * Finish our initialization steps.
2960  */
2961 static void
2962 uma_startup3(void *arg __unused)
2963 {
2964 
2965 #ifdef INVARIANTS
2966 	TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2967 	uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2968 	uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2969 #endif
2970 	zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2971 	callout_init(&uma_callout, 1);
2972 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2973 	booted = BOOT_RUNNING;
2974 
2975 	EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2976 	    EVENTHANDLER_PRI_FIRST);
2977 }
2978 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
2979 
2980 static void
2981 uma_shutdown(void)
2982 {
2983 
2984 	booted = BOOT_SHUTDOWN;
2985 }
2986 
2987 static uma_keg_t
2988 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2989 		int align, uint32_t flags)
2990 {
2991 	struct uma_kctor_args args;
2992 
2993 	args.size = size;
2994 	args.uminit = uminit;
2995 	args.fini = fini;
2996 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2997 	args.flags = flags;
2998 	args.zone = zone;
2999 	return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3000 }
3001 
3002 /* Public functions */
3003 /* See uma.h */
3004 void
3005 uma_set_align(int align)
3006 {
3007 
3008 	if (align != UMA_ALIGN_CACHE)
3009 		uma_align_cache = align;
3010 }
3011 
3012 /* See uma.h */
3013 uma_zone_t
3014 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3015 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
3016 
3017 {
3018 	struct uma_zctor_args args;
3019 	uma_zone_t res;
3020 
3021 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
3022 	    align, name));
3023 
3024 	/* This stuff is essential for the zone ctor */
3025 	memset(&args, 0, sizeof(args));
3026 	args.name = name;
3027 	args.size = size;
3028 	args.ctor = ctor;
3029 	args.dtor = dtor;
3030 	args.uminit = uminit;
3031 	args.fini = fini;
3032 #ifdef  INVARIANTS
3033 	/*
3034 	 * Inject procedures which check for memory use after free if we are
3035 	 * allowed to scramble the memory while it is not allocated.  This
3036 	 * requires that: UMA is actually able to access the memory, no init
3037 	 * or fini procedures, no dependency on the initial value of the
3038 	 * memory, and no (legitimate) use of the memory after free.  Note,
3039 	 * the ctor and dtor do not need to be empty.
3040 	 */
3041 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3042 	    UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3043 		args.uminit = trash_init;
3044 		args.fini = trash_fini;
3045 	}
3046 #endif
3047 	args.align = align;
3048 	args.flags = flags;
3049 	args.keg = NULL;
3050 
3051 	sx_slock(&uma_reclaim_lock);
3052 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3053 	sx_sunlock(&uma_reclaim_lock);
3054 
3055 	return (res);
3056 }
3057 
3058 /* See uma.h */
3059 uma_zone_t
3060 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3061     uma_init zinit, uma_fini zfini, uma_zone_t primary)
3062 {
3063 	struct uma_zctor_args args;
3064 	uma_keg_t keg;
3065 	uma_zone_t res;
3066 
3067 	keg = primary->uz_keg;
3068 	memset(&args, 0, sizeof(args));
3069 	args.name = name;
3070 	args.size = keg->uk_size;
3071 	args.ctor = ctor;
3072 	args.dtor = dtor;
3073 	args.uminit = zinit;
3074 	args.fini = zfini;
3075 	args.align = keg->uk_align;
3076 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3077 	args.keg = keg;
3078 
3079 	sx_slock(&uma_reclaim_lock);
3080 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3081 	sx_sunlock(&uma_reclaim_lock);
3082 
3083 	return (res);
3084 }
3085 
3086 /* See uma.h */
3087 uma_zone_t
3088 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3089     uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3090     void *arg, int flags)
3091 {
3092 	struct uma_zctor_args args;
3093 
3094 	memset(&args, 0, sizeof(args));
3095 	args.name = name;
3096 	args.size = size;
3097 	args.ctor = ctor;
3098 	args.dtor = dtor;
3099 	args.uminit = zinit;
3100 	args.fini = zfini;
3101 	args.import = zimport;
3102 	args.release = zrelease;
3103 	args.arg = arg;
3104 	args.align = 0;
3105 	args.flags = flags | UMA_ZFLAG_CACHE;
3106 
3107 	return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3108 }
3109 
3110 /* See uma.h */
3111 void
3112 uma_zdestroy(uma_zone_t zone)
3113 {
3114 
3115 	/*
3116 	 * Large slabs are expensive to reclaim, so don't bother doing
3117 	 * unnecessary work if we're shutting down.
3118 	 */
3119 	if (booted == BOOT_SHUTDOWN &&
3120 	    zone->uz_fini == NULL && zone->uz_release == zone_release)
3121 		return;
3122 	sx_slock(&uma_reclaim_lock);
3123 	zone_free_item(zones, zone, NULL, SKIP_NONE);
3124 	sx_sunlock(&uma_reclaim_lock);
3125 }
3126 
3127 void
3128 uma_zwait(uma_zone_t zone)
3129 {
3130 
3131 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3132 		uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3133 	else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3134 		uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3135 	else
3136 		uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3137 }
3138 
3139 void *
3140 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3141 {
3142 	void *item, *pcpu_item;
3143 #ifdef SMP
3144 	int i;
3145 
3146 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3147 #endif
3148 	item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3149 	if (item == NULL)
3150 		return (NULL);
3151 	pcpu_item = zpcpu_base_to_offset(item);
3152 	if (flags & M_ZERO) {
3153 #ifdef SMP
3154 		for (i = 0; i <= mp_maxid; i++)
3155 			bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3156 #else
3157 		bzero(item, zone->uz_size);
3158 #endif
3159 	}
3160 	return (pcpu_item);
3161 }
3162 
3163 /*
3164  * A stub while both regular and pcpu cases are identical.
3165  */
3166 void
3167 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3168 {
3169 	void *item;
3170 
3171 #ifdef SMP
3172 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3173 #endif
3174 	item = zpcpu_offset_to_base(pcpu_item);
3175 	uma_zfree_arg(zone, item, udata);
3176 }
3177 
3178 static inline void *
3179 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3180     void *item)
3181 {
3182 #ifdef INVARIANTS
3183 	bool skipdbg;
3184 
3185 	skipdbg = uma_dbg_zskip(zone, item);
3186 	if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3187 	    zone->uz_ctor != trash_ctor)
3188 		trash_ctor(item, size, udata, flags);
3189 #endif
3190 	/* Check flags before loading ctor pointer. */
3191 	if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3192 	    __predict_false(zone->uz_ctor != NULL) &&
3193 	    zone->uz_ctor(item, size, udata, flags) != 0) {
3194 		counter_u64_add(zone->uz_fails, 1);
3195 		zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3196 		return (NULL);
3197 	}
3198 #ifdef INVARIANTS
3199 	if (!skipdbg)
3200 		uma_dbg_alloc(zone, NULL, item);
3201 #endif
3202 	if (__predict_false(flags & M_ZERO))
3203 		return (memset(item, 0, size));
3204 
3205 	return (item);
3206 }
3207 
3208 static inline void
3209 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3210     enum zfreeskip skip)
3211 {
3212 #ifdef INVARIANTS
3213 	bool skipdbg;
3214 
3215 	skipdbg = uma_dbg_zskip(zone, item);
3216 	if (skip == SKIP_NONE && !skipdbg) {
3217 		if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3218 			uma_dbg_free(zone, udata, item);
3219 		else
3220 			uma_dbg_free(zone, NULL, item);
3221 	}
3222 #endif
3223 	if (__predict_true(skip < SKIP_DTOR)) {
3224 		if (zone->uz_dtor != NULL)
3225 			zone->uz_dtor(item, size, udata);
3226 #ifdef INVARIANTS
3227 		if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3228 		    zone->uz_dtor != trash_dtor)
3229 			trash_dtor(item, size, udata);
3230 #endif
3231 	}
3232 }
3233 
3234 #ifdef NUMA
3235 static int
3236 item_domain(void *item)
3237 {
3238 	int domain;
3239 
3240 	domain = vm_phys_domain(vtophys(item));
3241 	KASSERT(domain >= 0 && domain < vm_ndomains,
3242 	    ("%s: unknown domain for item %p", __func__, item));
3243 	return (domain);
3244 }
3245 #endif
3246 
3247 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3248 #define	UMA_ZALLOC_DEBUG
3249 static int
3250 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3251 {
3252 	int error;
3253 
3254 	error = 0;
3255 #ifdef WITNESS
3256 	if (flags & M_WAITOK) {
3257 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3258 		    "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3259 	}
3260 #endif
3261 
3262 #ifdef INVARIANTS
3263 	KASSERT((flags & M_EXEC) == 0,
3264 	    ("uma_zalloc_debug: called with M_EXEC"));
3265 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3266 	    ("uma_zalloc_debug: called within spinlock or critical section"));
3267 	KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3268 	    ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3269 #endif
3270 
3271 #ifdef DEBUG_MEMGUARD
3272 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3273 		void *item;
3274 		item = memguard_alloc(zone->uz_size, flags);
3275 		if (item != NULL) {
3276 			error = EJUSTRETURN;
3277 			if (zone->uz_init != NULL &&
3278 			    zone->uz_init(item, zone->uz_size, flags) != 0) {
3279 				*itemp = NULL;
3280 				return (error);
3281 			}
3282 			if (zone->uz_ctor != NULL &&
3283 			    zone->uz_ctor(item, zone->uz_size, udata,
3284 			    flags) != 0) {
3285 				counter_u64_add(zone->uz_fails, 1);
3286 			    	zone->uz_fini(item, zone->uz_size);
3287 				*itemp = NULL;
3288 				return (error);
3289 			}
3290 			*itemp = item;
3291 			return (error);
3292 		}
3293 		/* This is unfortunate but should not be fatal. */
3294 	}
3295 #endif
3296 	return (error);
3297 }
3298 
3299 static int
3300 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3301 {
3302 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3303 	    ("uma_zfree_debug: called with spinlock or critical section held"));
3304 
3305 #ifdef DEBUG_MEMGUARD
3306 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3307 		if (zone->uz_dtor != NULL)
3308 			zone->uz_dtor(item, zone->uz_size, udata);
3309 		if (zone->uz_fini != NULL)
3310 			zone->uz_fini(item, zone->uz_size);
3311 		memguard_free(item);
3312 		return (EJUSTRETURN);
3313 	}
3314 #endif
3315 	return (0);
3316 }
3317 #endif
3318 
3319 static inline void *
3320 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3321     void *udata, int flags)
3322 {
3323 	void *item;
3324 	int size, uz_flags;
3325 
3326 	item = cache_bucket_pop(cache, bucket);
3327 	size = cache_uz_size(cache);
3328 	uz_flags = cache_uz_flags(cache);
3329 	critical_exit();
3330 	return (item_ctor(zone, uz_flags, size, udata, flags, item));
3331 }
3332 
3333 static __noinline void *
3334 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3335 {
3336 	uma_cache_bucket_t bucket;
3337 	int domain;
3338 
3339 	while (cache_alloc(zone, cache, udata, flags)) {
3340 		cache = &zone->uz_cpu[curcpu];
3341 		bucket = &cache->uc_allocbucket;
3342 		if (__predict_false(bucket->ucb_cnt == 0))
3343 			continue;
3344 		return (cache_alloc_item(zone, cache, bucket, udata, flags));
3345 	}
3346 	critical_exit();
3347 
3348 	/*
3349 	 * We can not get a bucket so try to return a single item.
3350 	 */
3351 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3352 		domain = PCPU_GET(domain);
3353 	else
3354 		domain = UMA_ANYDOMAIN;
3355 	return (zone_alloc_item(zone, udata, domain, flags));
3356 }
3357 
3358 /* See uma.h */
3359 void *
3360 uma_zalloc_smr(uma_zone_t zone, int flags)
3361 {
3362 	uma_cache_bucket_t bucket;
3363 	uma_cache_t cache;
3364 
3365 #ifdef UMA_ZALLOC_DEBUG
3366 	void *item;
3367 
3368 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3369 	    ("uma_zalloc_arg: called with non-SMR zone."));
3370 	if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3371 		return (item);
3372 #endif
3373 
3374 	critical_enter();
3375 	cache = &zone->uz_cpu[curcpu];
3376 	bucket = &cache->uc_allocbucket;
3377 	if (__predict_false(bucket->ucb_cnt == 0))
3378 		return (cache_alloc_retry(zone, cache, NULL, flags));
3379 	return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3380 }
3381 
3382 /* See uma.h */
3383 void *
3384 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3385 {
3386 	uma_cache_bucket_t bucket;
3387 	uma_cache_t cache;
3388 
3389 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3390 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3391 
3392 	/* This is the fast path allocation */
3393 	CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3394 	    zone, flags);
3395 
3396 #ifdef UMA_ZALLOC_DEBUG
3397 	void *item;
3398 
3399 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3400 	    ("uma_zalloc_arg: called with SMR zone."));
3401 	if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3402 		return (item);
3403 #endif
3404 
3405 	/*
3406 	 * If possible, allocate from the per-CPU cache.  There are two
3407 	 * requirements for safe access to the per-CPU cache: (1) the thread
3408 	 * accessing the cache must not be preempted or yield during access,
3409 	 * and (2) the thread must not migrate CPUs without switching which
3410 	 * cache it accesses.  We rely on a critical section to prevent
3411 	 * preemption and migration.  We release the critical section in
3412 	 * order to acquire the zone mutex if we are unable to allocate from
3413 	 * the current cache; when we re-acquire the critical section, we
3414 	 * must detect and handle migration if it has occurred.
3415 	 */
3416 	critical_enter();
3417 	cache = &zone->uz_cpu[curcpu];
3418 	bucket = &cache->uc_allocbucket;
3419 	if (__predict_false(bucket->ucb_cnt == 0))
3420 		return (cache_alloc_retry(zone, cache, udata, flags));
3421 	return (cache_alloc_item(zone, cache, bucket, udata, flags));
3422 }
3423 
3424 /*
3425  * Replenish an alloc bucket and possibly restore an old one.  Called in
3426  * a critical section.  Returns in a critical section.
3427  *
3428  * A false return value indicates an allocation failure.
3429  * A true return value indicates success and the caller should retry.
3430  */
3431 static __noinline bool
3432 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3433 {
3434 	uma_bucket_t bucket;
3435 	int curdomain, domain;
3436 	bool new;
3437 
3438 	CRITICAL_ASSERT(curthread);
3439 
3440 	/*
3441 	 * If we have run out of items in our alloc bucket see
3442 	 * if we can switch with the free bucket.
3443 	 *
3444 	 * SMR Zones can't re-use the free bucket until the sequence has
3445 	 * expired.
3446 	 */
3447 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3448 	    cache->uc_freebucket.ucb_cnt != 0) {
3449 		cache_bucket_swap(&cache->uc_freebucket,
3450 		    &cache->uc_allocbucket);
3451 		return (true);
3452 	}
3453 
3454 	/*
3455 	 * Discard any empty allocation bucket while we hold no locks.
3456 	 */
3457 	bucket = cache_bucket_unload_alloc(cache);
3458 	critical_exit();
3459 
3460 	if (bucket != NULL) {
3461 		KASSERT(bucket->ub_cnt == 0,
3462 		    ("cache_alloc: Entered with non-empty alloc bucket."));
3463 		bucket_free(zone, bucket, udata);
3464 	}
3465 
3466 	/*
3467 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
3468 	 * we must go back to the zone.  This requires the zdom lock, so we
3469 	 * must drop the critical section, then re-acquire it when we go back
3470 	 * to the cache.  Since the critical section is released, we may be
3471 	 * preempted or migrate.  As such, make sure not to maintain any
3472 	 * thread-local state specific to the cache from prior to releasing
3473 	 * the critical section.
3474 	 */
3475 	domain = PCPU_GET(domain);
3476 	if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3477 	    VM_DOMAIN_EMPTY(domain))
3478 		domain = zone_domain_highest(zone, domain);
3479 	bucket = cache_fetch_bucket(zone, cache, domain);
3480 	if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3481 		bucket = zone_alloc_bucket(zone, udata, domain, flags);
3482 		new = true;
3483 	} else {
3484 		new = false;
3485 	}
3486 
3487 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3488 	    zone->uz_name, zone, bucket);
3489 	if (bucket == NULL) {
3490 		critical_enter();
3491 		return (false);
3492 	}
3493 
3494 	/*
3495 	 * See if we lost the race or were migrated.  Cache the
3496 	 * initialized bucket to make this less likely or claim
3497 	 * the memory directly.
3498 	 */
3499 	critical_enter();
3500 	cache = &zone->uz_cpu[curcpu];
3501 	if (cache->uc_allocbucket.ucb_bucket == NULL &&
3502 	    ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3503 	    (curdomain = PCPU_GET(domain)) == domain ||
3504 	    VM_DOMAIN_EMPTY(curdomain))) {
3505 		if (new)
3506 			atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3507 			    bucket->ub_cnt);
3508 		cache_bucket_load_alloc(cache, bucket);
3509 		return (true);
3510 	}
3511 
3512 	/*
3513 	 * We lost the race, release this bucket and start over.
3514 	 */
3515 	critical_exit();
3516 	zone_put_bucket(zone, domain, bucket, udata, false);
3517 	critical_enter();
3518 
3519 	return (true);
3520 }
3521 
3522 void *
3523 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3524 {
3525 #ifdef NUMA
3526 	uma_bucket_t bucket;
3527 	uma_zone_domain_t zdom;
3528 	void *item;
3529 #endif
3530 
3531 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3532 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3533 
3534 	/* This is the fast path allocation */
3535 	CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3536 	    zone->uz_name, zone, domain, flags);
3537 
3538 	if (flags & M_WAITOK) {
3539 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3540 		    "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3541 	}
3542 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3543 	    ("uma_zalloc_domain: called with spinlock or critical section held"));
3544 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3545 	    ("uma_zalloc_domain: called with SMR zone."));
3546 #ifdef NUMA
3547 	KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3548 	    ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3549 
3550 	if (vm_ndomains == 1)
3551 		return (uma_zalloc_arg(zone, udata, flags));
3552 
3553 	/*
3554 	 * Try to allocate from the bucket cache before falling back to the keg.
3555 	 * We could try harder and attempt to allocate from per-CPU caches or
3556 	 * the per-domain cross-domain buckets, but the complexity is probably
3557 	 * not worth it.  It is more important that frees of previous
3558 	 * cross-domain allocations do not blow up the cache.
3559 	 */
3560 	zdom = zone_domain_lock(zone, domain);
3561 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3562 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
3563 #ifdef INVARIANTS
3564 		bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3565 #endif
3566 		bucket->ub_cnt--;
3567 		zone_put_bucket(zone, domain, bucket, udata, true);
3568 		item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3569 		    flags, item);
3570 		if (item != NULL) {
3571 			KASSERT(item_domain(item) == domain,
3572 			    ("%s: bucket cache item %p from wrong domain",
3573 			    __func__, item));
3574 			counter_u64_add(zone->uz_allocs, 1);
3575 		}
3576 		return (item);
3577 	}
3578 	ZDOM_UNLOCK(zdom);
3579 	return (zone_alloc_item(zone, udata, domain, flags));
3580 #else
3581 	return (uma_zalloc_arg(zone, udata, flags));
3582 #endif
3583 }
3584 
3585 /*
3586  * Find a slab with some space.  Prefer slabs that are partially used over those
3587  * that are totally full.  This helps to reduce fragmentation.
3588  *
3589  * If 'rr' is 1, search all domains starting from 'domain'.  Otherwise check
3590  * only 'domain'.
3591  */
3592 static uma_slab_t
3593 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3594 {
3595 	uma_domain_t dom;
3596 	uma_slab_t slab;
3597 	int start;
3598 
3599 	KASSERT(domain >= 0 && domain < vm_ndomains,
3600 	    ("keg_first_slab: domain %d out of range", domain));
3601 	KEG_LOCK_ASSERT(keg, domain);
3602 
3603 	slab = NULL;
3604 	start = domain;
3605 	do {
3606 		dom = &keg->uk_domain[domain];
3607 		if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3608 			return (slab);
3609 		if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3610 			LIST_REMOVE(slab, us_link);
3611 			dom->ud_free_slabs--;
3612 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3613 			return (slab);
3614 		}
3615 		if (rr)
3616 			domain = (domain + 1) % vm_ndomains;
3617 	} while (domain != start);
3618 
3619 	return (NULL);
3620 }
3621 
3622 /*
3623  * Fetch an existing slab from a free or partial list.  Returns with the
3624  * keg domain lock held if a slab was found or unlocked if not.
3625  */
3626 static uma_slab_t
3627 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3628 {
3629 	uma_slab_t slab;
3630 	uint32_t reserve;
3631 
3632 	/* HASH has a single free list. */
3633 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3634 		domain = 0;
3635 
3636 	KEG_LOCK(keg, domain);
3637 	reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3638 	if (keg->uk_domain[domain].ud_free_items <= reserve ||
3639 	    (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3640 		KEG_UNLOCK(keg, domain);
3641 		return (NULL);
3642 	}
3643 	return (slab);
3644 }
3645 
3646 static uma_slab_t
3647 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3648 {
3649 	struct vm_domainset_iter di;
3650 	uma_slab_t slab;
3651 	int aflags, domain;
3652 	bool rr;
3653 
3654 restart:
3655 	/*
3656 	 * Use the keg's policy if upper layers haven't already specified a
3657 	 * domain (as happens with first-touch zones).
3658 	 *
3659 	 * To avoid races we run the iterator with the keg lock held, but that
3660 	 * means that we cannot allow the vm_domainset layer to sleep.  Thus,
3661 	 * clear M_WAITOK and handle low memory conditions locally.
3662 	 */
3663 	rr = rdomain == UMA_ANYDOMAIN;
3664 	if (rr) {
3665 		aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3666 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3667 		    &aflags);
3668 	} else {
3669 		aflags = flags;
3670 		domain = rdomain;
3671 	}
3672 
3673 	for (;;) {
3674 		slab = keg_fetch_free_slab(keg, domain, rr, flags);
3675 		if (slab != NULL)
3676 			return (slab);
3677 
3678 		/*
3679 		 * M_NOVM means don't ask at all!
3680 		 */
3681 		if (flags & M_NOVM)
3682 			break;
3683 
3684 		slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3685 		if (slab != NULL)
3686 			return (slab);
3687 		if (!rr && (flags & M_WAITOK) == 0)
3688 			break;
3689 		if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3690 			if ((flags & M_WAITOK) != 0) {
3691 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
3692 				goto restart;
3693 			}
3694 			break;
3695 		}
3696 	}
3697 
3698 	/*
3699 	 * We might not have been able to get a slab but another cpu
3700 	 * could have while we were unlocked.  Check again before we
3701 	 * fail.
3702 	 */
3703 	if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3704 		return (slab);
3705 
3706 	return (NULL);
3707 }
3708 
3709 static void *
3710 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3711 {
3712 	uma_domain_t dom;
3713 	void *item;
3714 	int freei;
3715 
3716 	KEG_LOCK_ASSERT(keg, slab->us_domain);
3717 
3718 	dom = &keg->uk_domain[slab->us_domain];
3719 	freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3720 	BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3721 	item = slab_item(slab, keg, freei);
3722 	slab->us_freecount--;
3723 	dom->ud_free_items--;
3724 
3725 	/*
3726 	 * Move this slab to the full list.  It must be on the partial list, so
3727 	 * we do not need to update the free slab count.  In particular,
3728 	 * keg_fetch_slab() always returns slabs on the partial list.
3729 	 */
3730 	if (slab->us_freecount == 0) {
3731 		LIST_REMOVE(slab, us_link);
3732 		LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3733 	}
3734 
3735 	return (item);
3736 }
3737 
3738 static int
3739 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3740 {
3741 	uma_domain_t dom;
3742 	uma_zone_t zone;
3743 	uma_slab_t slab;
3744 	uma_keg_t keg;
3745 #ifdef NUMA
3746 	int stripe;
3747 #endif
3748 	int i;
3749 
3750 	zone = arg;
3751 	slab = NULL;
3752 	keg = zone->uz_keg;
3753 	/* Try to keep the buckets totally full */
3754 	for (i = 0; i < max; ) {
3755 		if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3756 			break;
3757 #ifdef NUMA
3758 		stripe = howmany(max, vm_ndomains);
3759 #endif
3760 		dom = &keg->uk_domain[slab->us_domain];
3761 		do {
3762 			bucket[i++] = slab_alloc_item(keg, slab);
3763 			if (dom->ud_free_items <= keg->uk_reserve) {
3764 				/*
3765 				 * Avoid depleting the reserve after a
3766 				 * successful item allocation, even if
3767 				 * M_USE_RESERVE is specified.
3768 				 */
3769 				KEG_UNLOCK(keg, slab->us_domain);
3770 				goto out;
3771 			}
3772 #ifdef NUMA
3773 			/*
3774 			 * If the zone is striped we pick a new slab for every
3775 			 * N allocations.  Eliminating this conditional will
3776 			 * instead pick a new domain for each bucket rather
3777 			 * than stripe within each bucket.  The current option
3778 			 * produces more fragmentation and requires more cpu
3779 			 * time but yields better distribution.
3780 			 */
3781 			if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3782 			    vm_ndomains > 1 && --stripe == 0)
3783 				break;
3784 #endif
3785 		} while (slab->us_freecount != 0 && i < max);
3786 		KEG_UNLOCK(keg, slab->us_domain);
3787 
3788 		/* Don't block if we allocated any successfully. */
3789 		flags &= ~M_WAITOK;
3790 		flags |= M_NOWAIT;
3791 	}
3792 out:
3793 	return i;
3794 }
3795 
3796 static int
3797 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3798 {
3799 	uint64_t old, new, total, max;
3800 
3801 	/*
3802 	 * The hard case.  We're going to sleep because there were existing
3803 	 * sleepers or because we ran out of items.  This routine enforces
3804 	 * fairness by keeping fifo order.
3805 	 *
3806 	 * First release our ill gotten gains and make some noise.
3807 	 */
3808 	for (;;) {
3809 		zone_free_limit(zone, count);
3810 		zone_log_warning(zone);
3811 		zone_maxaction(zone);
3812 		if (flags & M_NOWAIT)
3813 			return (0);
3814 
3815 		/*
3816 		 * We need to allocate an item or set ourself as a sleeper
3817 		 * while the sleepq lock is held to avoid wakeup races.  This
3818 		 * is essentially a home rolled semaphore.
3819 		 */
3820 		sleepq_lock(&zone->uz_max_items);
3821 		old = zone->uz_items;
3822 		do {
3823 			MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3824 			/* Cache the max since we will evaluate twice. */
3825 			max = zone->uz_max_items;
3826 			if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3827 			    UZ_ITEMS_COUNT(old) >= max)
3828 				new = old + UZ_ITEMS_SLEEPER;
3829 			else
3830 				new = old + MIN(count, max - old);
3831 		} while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3832 
3833 		/* We may have successfully allocated under the sleepq lock. */
3834 		if (UZ_ITEMS_SLEEPERS(new) == 0) {
3835 			sleepq_release(&zone->uz_max_items);
3836 			return (new - old);
3837 		}
3838 
3839 		/*
3840 		 * This is in a different cacheline from uz_items so that we
3841 		 * don't constantly invalidate the fastpath cacheline when we
3842 		 * adjust item counts.  This could be limited to toggling on
3843 		 * transitions.
3844 		 */
3845 		atomic_add_32(&zone->uz_sleepers, 1);
3846 		atomic_add_64(&zone->uz_sleeps, 1);
3847 
3848 		/*
3849 		 * We have added ourselves as a sleeper.  The sleepq lock
3850 		 * protects us from wakeup races.  Sleep now and then retry.
3851 		 */
3852 		sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3853 		sleepq_wait(&zone->uz_max_items, PVM);
3854 
3855 		/*
3856 		 * After wakeup, remove ourselves as a sleeper and try
3857 		 * again.  We no longer have the sleepq lock for protection.
3858 		 *
3859 		 * Subract ourselves as a sleeper while attempting to add
3860 		 * our count.
3861 		 */
3862 		atomic_subtract_32(&zone->uz_sleepers, 1);
3863 		old = atomic_fetchadd_64(&zone->uz_items,
3864 		    -(UZ_ITEMS_SLEEPER - count));
3865 		/* We're no longer a sleeper. */
3866 		old -= UZ_ITEMS_SLEEPER;
3867 
3868 		/*
3869 		 * If we're still at the limit, restart.  Notably do not
3870 		 * block on other sleepers.  Cache the max value to protect
3871 		 * against changes via sysctl.
3872 		 */
3873 		total = UZ_ITEMS_COUNT(old);
3874 		max = zone->uz_max_items;
3875 		if (total >= max)
3876 			continue;
3877 		/* Truncate if necessary, otherwise wake other sleepers. */
3878 		if (total + count > max) {
3879 			zone_free_limit(zone, total + count - max);
3880 			count = max - total;
3881 		} else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3882 			wakeup_one(&zone->uz_max_items);
3883 
3884 		return (count);
3885 	}
3886 }
3887 
3888 /*
3889  * Allocate 'count' items from our max_items limit.  Returns the number
3890  * available.  If M_NOWAIT is not specified it will sleep until at least
3891  * one item can be allocated.
3892  */
3893 static int
3894 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3895 {
3896 	uint64_t old;
3897 	uint64_t max;
3898 
3899 	max = zone->uz_max_items;
3900 	MPASS(max > 0);
3901 
3902 	/*
3903 	 * We expect normal allocations to succeed with a simple
3904 	 * fetchadd.
3905 	 */
3906 	old = atomic_fetchadd_64(&zone->uz_items, count);
3907 	if (__predict_true(old + count <= max))
3908 		return (count);
3909 
3910 	/*
3911 	 * If we had some items and no sleepers just return the
3912 	 * truncated value.  We have to release the excess space
3913 	 * though because that may wake sleepers who weren't woken
3914 	 * because we were temporarily over the limit.
3915 	 */
3916 	if (old < max) {
3917 		zone_free_limit(zone, (old + count) - max);
3918 		return (max - old);
3919 	}
3920 	return (zone_alloc_limit_hard(zone, count, flags));
3921 }
3922 
3923 /*
3924  * Free a number of items back to the limit.
3925  */
3926 static void
3927 zone_free_limit(uma_zone_t zone, int count)
3928 {
3929 	uint64_t old;
3930 
3931 	MPASS(count > 0);
3932 
3933 	/*
3934 	 * In the common case we either have no sleepers or
3935 	 * are still over the limit and can just return.
3936 	 */
3937 	old = atomic_fetchadd_64(&zone->uz_items, -count);
3938 	if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3939 	   UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3940 		return;
3941 
3942 	/*
3943 	 * Moderate the rate of wakeups.  Sleepers will continue
3944 	 * to generate wakeups if necessary.
3945 	 */
3946 	wakeup_one(&zone->uz_max_items);
3947 }
3948 
3949 static uma_bucket_t
3950 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3951 {
3952 	uma_bucket_t bucket;
3953 	int maxbucket, cnt;
3954 
3955 	CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3956 	    zone, domain);
3957 
3958 	/* Avoid allocs targeting empty domains. */
3959 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3960 		domain = UMA_ANYDOMAIN;
3961 	else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
3962 		domain = UMA_ANYDOMAIN;
3963 
3964 	if (zone->uz_max_items > 0)
3965 		maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3966 		    M_NOWAIT);
3967 	else
3968 		maxbucket = zone->uz_bucket_size;
3969 	if (maxbucket == 0)
3970 		return (false);
3971 
3972 	/* Don't wait for buckets, preserve caller's NOVM setting. */
3973 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3974 	if (bucket == NULL) {
3975 		cnt = 0;
3976 		goto out;
3977 	}
3978 
3979 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3980 	    MIN(maxbucket, bucket->ub_entries), domain, flags);
3981 
3982 	/*
3983 	 * Initialize the memory if necessary.
3984 	 */
3985 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3986 		int i;
3987 
3988 		for (i = 0; i < bucket->ub_cnt; i++)
3989 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3990 			    flags) != 0)
3991 				break;
3992 		/*
3993 		 * If we couldn't initialize the whole bucket, put the
3994 		 * rest back onto the freelist.
3995 		 */
3996 		if (i != bucket->ub_cnt) {
3997 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3998 			    bucket->ub_cnt - i);
3999 #ifdef INVARIANTS
4000 			bzero(&bucket->ub_bucket[i],
4001 			    sizeof(void *) * (bucket->ub_cnt - i));
4002 #endif
4003 			bucket->ub_cnt = i;
4004 		}
4005 	}
4006 
4007 	cnt = bucket->ub_cnt;
4008 	if (bucket->ub_cnt == 0) {
4009 		bucket_free(zone, bucket, udata);
4010 		counter_u64_add(zone->uz_fails, 1);
4011 		bucket = NULL;
4012 	}
4013 out:
4014 	if (zone->uz_max_items > 0 && cnt < maxbucket)
4015 		zone_free_limit(zone, maxbucket - cnt);
4016 
4017 	return (bucket);
4018 }
4019 
4020 /*
4021  * Allocates a single item from a zone.
4022  *
4023  * Arguments
4024  *	zone   The zone to alloc for.
4025  *	udata  The data to be passed to the constructor.
4026  *	domain The domain to allocate from or UMA_ANYDOMAIN.
4027  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
4028  *
4029  * Returns
4030  *	NULL if there is no memory and M_NOWAIT is set
4031  *	An item if successful
4032  */
4033 
4034 static void *
4035 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4036 {
4037 	void *item;
4038 
4039 	if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4040 		counter_u64_add(zone->uz_fails, 1);
4041 		return (NULL);
4042 	}
4043 
4044 	/* Avoid allocs targeting empty domains. */
4045 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4046 		domain = UMA_ANYDOMAIN;
4047 
4048 	if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4049 		goto fail_cnt;
4050 
4051 	/*
4052 	 * We have to call both the zone's init (not the keg's init)
4053 	 * and the zone's ctor.  This is because the item is going from
4054 	 * a keg slab directly to the user, and the user is expecting it
4055 	 * to be both zone-init'd as well as zone-ctor'd.
4056 	 */
4057 	if (zone->uz_init != NULL) {
4058 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
4059 			zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4060 			goto fail_cnt;
4061 		}
4062 	}
4063 	item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4064 	    item);
4065 	if (item == NULL)
4066 		goto fail;
4067 
4068 	counter_u64_add(zone->uz_allocs, 1);
4069 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4070 	    zone->uz_name, zone);
4071 
4072 	return (item);
4073 
4074 fail_cnt:
4075 	counter_u64_add(zone->uz_fails, 1);
4076 fail:
4077 	if (zone->uz_max_items > 0)
4078 		zone_free_limit(zone, 1);
4079 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4080 	    zone->uz_name, zone);
4081 
4082 	return (NULL);
4083 }
4084 
4085 /* See uma.h */
4086 void
4087 uma_zfree_smr(uma_zone_t zone, void *item)
4088 {
4089 	uma_cache_t cache;
4090 	uma_cache_bucket_t bucket;
4091 	int itemdomain, uz_flags;
4092 
4093 #ifdef UMA_ZALLOC_DEBUG
4094 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4095 	    ("uma_zfree_smr: called with non-SMR zone."));
4096 	KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4097 	SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4098 	if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4099 		return;
4100 #endif
4101 	cache = &zone->uz_cpu[curcpu];
4102 	uz_flags = cache_uz_flags(cache);
4103 	itemdomain = 0;
4104 #ifdef NUMA
4105 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4106 		itemdomain = item_domain(item);
4107 #endif
4108 	critical_enter();
4109 	do {
4110 		cache = &zone->uz_cpu[curcpu];
4111 		/* SMR Zones must free to the free bucket. */
4112 		bucket = &cache->uc_freebucket;
4113 #ifdef NUMA
4114 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4115 		    PCPU_GET(domain) != itemdomain) {
4116 			bucket = &cache->uc_crossbucket;
4117 		}
4118 #endif
4119 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4120 			cache_bucket_push(cache, bucket, item);
4121 			critical_exit();
4122 			return;
4123 		}
4124 	} while (cache_free(zone, cache, NULL, item, itemdomain));
4125 	critical_exit();
4126 
4127 	/*
4128 	 * If nothing else caught this, we'll just do an internal free.
4129 	 */
4130 	zone_free_item(zone, item, NULL, SKIP_NONE);
4131 }
4132 
4133 /* See uma.h */
4134 void
4135 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4136 {
4137 	uma_cache_t cache;
4138 	uma_cache_bucket_t bucket;
4139 	int itemdomain, uz_flags;
4140 
4141 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4142 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4143 
4144 	CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4145 
4146 #ifdef UMA_ZALLOC_DEBUG
4147 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4148 	    ("uma_zfree_arg: called with SMR zone."));
4149 	if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4150 		return;
4151 #endif
4152         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4153         if (item == NULL)
4154                 return;
4155 
4156 	/*
4157 	 * We are accessing the per-cpu cache without a critical section to
4158 	 * fetch size and flags.  This is acceptable, if we are preempted we
4159 	 * will simply read another cpu's line.
4160 	 */
4161 	cache = &zone->uz_cpu[curcpu];
4162 	uz_flags = cache_uz_flags(cache);
4163 	if (UMA_ALWAYS_CTORDTOR ||
4164 	    __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4165 		item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4166 
4167 	/*
4168 	 * The race here is acceptable.  If we miss it we'll just have to wait
4169 	 * a little longer for the limits to be reset.
4170 	 */
4171 	if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4172 		if (atomic_load_32(&zone->uz_sleepers) > 0)
4173 			goto zfree_item;
4174 	}
4175 
4176 	/*
4177 	 * If possible, free to the per-CPU cache.  There are two
4178 	 * requirements for safe access to the per-CPU cache: (1) the thread
4179 	 * accessing the cache must not be preempted or yield during access,
4180 	 * and (2) the thread must not migrate CPUs without switching which
4181 	 * cache it accesses.  We rely on a critical section to prevent
4182 	 * preemption and migration.  We release the critical section in
4183 	 * order to acquire the zone mutex if we are unable to free to the
4184 	 * current cache; when we re-acquire the critical section, we must
4185 	 * detect and handle migration if it has occurred.
4186 	 */
4187 	itemdomain = 0;
4188 #ifdef NUMA
4189 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4190 		itemdomain = item_domain(item);
4191 #endif
4192 	critical_enter();
4193 	do {
4194 		cache = &zone->uz_cpu[curcpu];
4195 		/*
4196 		 * Try to free into the allocbucket first to give LIFO
4197 		 * ordering for cache-hot datastructures.  Spill over
4198 		 * into the freebucket if necessary.  Alloc will swap
4199 		 * them if one runs dry.
4200 		 */
4201 		bucket = &cache->uc_allocbucket;
4202 #ifdef NUMA
4203 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4204 		    PCPU_GET(domain) != itemdomain) {
4205 			bucket = &cache->uc_crossbucket;
4206 		} else
4207 #endif
4208 		if (bucket->ucb_cnt == bucket->ucb_entries &&
4209 		   cache->uc_freebucket.ucb_cnt <
4210 		   cache->uc_freebucket.ucb_entries)
4211 			cache_bucket_swap(&cache->uc_freebucket,
4212 			    &cache->uc_allocbucket);
4213 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4214 			cache_bucket_push(cache, bucket, item);
4215 			critical_exit();
4216 			return;
4217 		}
4218 	} while (cache_free(zone, cache, udata, item, itemdomain));
4219 	critical_exit();
4220 
4221 	/*
4222 	 * If nothing else caught this, we'll just do an internal free.
4223 	 */
4224 zfree_item:
4225 	zone_free_item(zone, item, udata, SKIP_DTOR);
4226 }
4227 
4228 #ifdef NUMA
4229 /*
4230  * sort crossdomain free buckets to domain correct buckets and cache
4231  * them.
4232  */
4233 static void
4234 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4235 {
4236 	struct uma_bucketlist emptybuckets, fullbuckets;
4237 	uma_zone_domain_t zdom;
4238 	uma_bucket_t b;
4239 	smr_seq_t seq;
4240 	void *item;
4241 	int domain;
4242 
4243 	CTR3(KTR_UMA,
4244 	    "uma_zfree: zone %s(%p) draining cross bucket %p",
4245 	    zone->uz_name, zone, bucket);
4246 
4247 	/*
4248 	 * It is possible for buckets to arrive here out of order so we fetch
4249 	 * the current smr seq rather than accepting the bucket's.
4250 	 */
4251 	seq = SMR_SEQ_INVALID;
4252 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4253 		seq = smr_advance(zone->uz_smr);
4254 
4255 	/*
4256 	 * To avoid having ndomain * ndomain buckets for sorting we have a
4257 	 * lock on the current crossfree bucket.  A full matrix with
4258 	 * per-domain locking could be used if necessary.
4259 	 */
4260 	STAILQ_INIT(&emptybuckets);
4261 	STAILQ_INIT(&fullbuckets);
4262 	ZONE_CROSS_LOCK(zone);
4263 	for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4264 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
4265 		domain = item_domain(item);
4266 		zdom = ZDOM_GET(zone, domain);
4267 		if (zdom->uzd_cross == NULL) {
4268 			if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4269 				STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4270 				zdom->uzd_cross = b;
4271 			} else {
4272 				/*
4273 				 * Avoid allocating a bucket with the cross lock
4274 				 * held, since allocation can trigger a
4275 				 * cross-domain free and bucket zones may
4276 				 * allocate from each other.
4277 				 */
4278 				ZONE_CROSS_UNLOCK(zone);
4279 				b = bucket_alloc(zone, udata, M_NOWAIT);
4280 				if (b == NULL)
4281 					goto out;
4282 				ZONE_CROSS_LOCK(zone);
4283 				if (zdom->uzd_cross != NULL) {
4284 					STAILQ_INSERT_HEAD(&emptybuckets, b,
4285 					    ub_link);
4286 				} else {
4287 					zdom->uzd_cross = b;
4288 				}
4289 			}
4290 		}
4291 		b = zdom->uzd_cross;
4292 		b->ub_bucket[b->ub_cnt++] = item;
4293 		b->ub_seq = seq;
4294 		if (b->ub_cnt == b->ub_entries) {
4295 			STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4296 			if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4297 				STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4298 			zdom->uzd_cross = b;
4299 		}
4300 	}
4301 	ZONE_CROSS_UNLOCK(zone);
4302 out:
4303 	if (bucket->ub_cnt == 0)
4304 		bucket->ub_seq = SMR_SEQ_INVALID;
4305 	bucket_free(zone, bucket, udata);
4306 
4307 	while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4308 		STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4309 		bucket_free(zone, b, udata);
4310 	}
4311 	while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4312 		STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4313 		domain = item_domain(b->ub_bucket[0]);
4314 		zone_put_bucket(zone, domain, b, udata, true);
4315 	}
4316 }
4317 #endif
4318 
4319 static void
4320 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4321     int itemdomain, bool ws)
4322 {
4323 
4324 #ifdef NUMA
4325 	/*
4326 	 * Buckets coming from the wrong domain will be entirely for the
4327 	 * only other domain on two domain systems.  In this case we can
4328 	 * simply cache them.  Otherwise we need to sort them back to
4329 	 * correct domains.
4330 	 */
4331 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4332 	    vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4333 		zone_free_cross(zone, bucket, udata);
4334 		return;
4335 	}
4336 #endif
4337 
4338 	/*
4339 	 * Attempt to save the bucket in the zone's domain bucket cache.
4340 	 */
4341 	CTR3(KTR_UMA,
4342 	    "uma_zfree: zone %s(%p) putting bucket %p on free list",
4343 	    zone->uz_name, zone, bucket);
4344 	/* ub_cnt is pointing to the last free item */
4345 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4346 		itemdomain = zone_domain_lowest(zone, itemdomain);
4347 	zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4348 }
4349 
4350 /*
4351  * Populate a free or cross bucket for the current cpu cache.  Free any
4352  * existing full bucket either to the zone cache or back to the slab layer.
4353  *
4354  * Enters and returns in a critical section.  false return indicates that
4355  * we can not satisfy this free in the cache layer.  true indicates that
4356  * the caller should retry.
4357  */
4358 static __noinline bool
4359 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4360     int itemdomain)
4361 {
4362 	uma_cache_bucket_t cbucket;
4363 	uma_bucket_t newbucket, bucket;
4364 
4365 	CRITICAL_ASSERT(curthread);
4366 
4367 	if (zone->uz_bucket_size == 0)
4368 		return false;
4369 
4370 	cache = &zone->uz_cpu[curcpu];
4371 	newbucket = NULL;
4372 
4373 	/*
4374 	 * FIRSTTOUCH domains need to free to the correct zdom.  When
4375 	 * enabled this is the zdom of the item.   The bucket is the
4376 	 * cross bucket if the current domain and itemdomain do not match.
4377 	 */
4378 	cbucket = &cache->uc_freebucket;
4379 #ifdef NUMA
4380 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4381 		if (PCPU_GET(domain) != itemdomain) {
4382 			cbucket = &cache->uc_crossbucket;
4383 			if (cbucket->ucb_cnt != 0)
4384 				counter_u64_add(zone->uz_xdomain,
4385 				    cbucket->ucb_cnt);
4386 		}
4387 	}
4388 #endif
4389 	bucket = cache_bucket_unload(cbucket);
4390 	KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4391 	    ("cache_free: Entered with non-full free bucket."));
4392 
4393 	/* We are no longer associated with this CPU. */
4394 	critical_exit();
4395 
4396 	/*
4397 	 * Don't let SMR zones operate without a free bucket.  Force
4398 	 * a synchronize and re-use this one.  We will only degrade
4399 	 * to a synchronize every bucket_size items rather than every
4400 	 * item if we fail to allocate a bucket.
4401 	 */
4402 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4403 		if (bucket != NULL)
4404 			bucket->ub_seq = smr_advance(zone->uz_smr);
4405 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4406 		if (newbucket == NULL && bucket != NULL) {
4407 			bucket_drain(zone, bucket);
4408 			newbucket = bucket;
4409 			bucket = NULL;
4410 		}
4411 	} else if (!bucketdisable)
4412 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4413 
4414 	if (bucket != NULL)
4415 		zone_free_bucket(zone, bucket, udata, itemdomain, true);
4416 
4417 	critical_enter();
4418 	if ((bucket = newbucket) == NULL)
4419 		return (false);
4420 	cache = &zone->uz_cpu[curcpu];
4421 #ifdef NUMA
4422 	/*
4423 	 * Check to see if we should be populating the cross bucket.  If it
4424 	 * is already populated we will fall through and attempt to populate
4425 	 * the free bucket.
4426 	 */
4427 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4428 		if (PCPU_GET(domain) != itemdomain &&
4429 		    cache->uc_crossbucket.ucb_bucket == NULL) {
4430 			cache_bucket_load_cross(cache, bucket);
4431 			return (true);
4432 		}
4433 	}
4434 #endif
4435 	/*
4436 	 * We may have lost the race to fill the bucket or switched CPUs.
4437 	 */
4438 	if (cache->uc_freebucket.ucb_bucket != NULL) {
4439 		critical_exit();
4440 		bucket_free(zone, bucket, udata);
4441 		critical_enter();
4442 	} else
4443 		cache_bucket_load_free(cache, bucket);
4444 
4445 	return (true);
4446 }
4447 
4448 static void
4449 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4450 {
4451 	uma_keg_t keg;
4452 	uma_domain_t dom;
4453 	int freei;
4454 
4455 	keg = zone->uz_keg;
4456 	KEG_LOCK_ASSERT(keg, slab->us_domain);
4457 
4458 	/* Do we need to remove from any lists? */
4459 	dom = &keg->uk_domain[slab->us_domain];
4460 	if (slab->us_freecount + 1 == keg->uk_ipers) {
4461 		LIST_REMOVE(slab, us_link);
4462 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4463 		dom->ud_free_slabs++;
4464 	} else if (slab->us_freecount == 0) {
4465 		LIST_REMOVE(slab, us_link);
4466 		LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4467 	}
4468 
4469 	/* Slab management. */
4470 	freei = slab_item_index(slab, keg, item);
4471 	BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4472 	slab->us_freecount++;
4473 
4474 	/* Keg statistics. */
4475 	dom->ud_free_items++;
4476 }
4477 
4478 static void
4479 zone_release(void *arg, void **bucket, int cnt)
4480 {
4481 	struct mtx *lock;
4482 	uma_zone_t zone;
4483 	uma_slab_t slab;
4484 	uma_keg_t keg;
4485 	uint8_t *mem;
4486 	void *item;
4487 	int i;
4488 
4489 	zone = arg;
4490 	keg = zone->uz_keg;
4491 	lock = NULL;
4492 	if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4493 		lock = KEG_LOCK(keg, 0);
4494 	for (i = 0; i < cnt; i++) {
4495 		item = bucket[i];
4496 		if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4497 			slab = vtoslab((vm_offset_t)item);
4498 		} else {
4499 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4500 			if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4501 				slab = hash_sfind(&keg->uk_hash, mem);
4502 			else
4503 				slab = (uma_slab_t)(mem + keg->uk_pgoff);
4504 		}
4505 		if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4506 			if (lock != NULL)
4507 				mtx_unlock(lock);
4508 			lock = KEG_LOCK(keg, slab->us_domain);
4509 		}
4510 		slab_free_item(zone, slab, item);
4511 	}
4512 	if (lock != NULL)
4513 		mtx_unlock(lock);
4514 }
4515 
4516 /*
4517  * Frees a single item to any zone.
4518  *
4519  * Arguments:
4520  *	zone   The zone to free to
4521  *	item   The item we're freeing
4522  *	udata  User supplied data for the dtor
4523  *	skip   Skip dtors and finis
4524  */
4525 static __noinline void
4526 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4527 {
4528 
4529 	/*
4530 	 * If a free is sent directly to an SMR zone we have to
4531 	 * synchronize immediately because the item can instantly
4532 	 * be reallocated. This should only happen in degenerate
4533 	 * cases when no memory is available for per-cpu caches.
4534 	 */
4535 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4536 		smr_synchronize(zone->uz_smr);
4537 
4538 	item_dtor(zone, item, zone->uz_size, udata, skip);
4539 
4540 	if (skip < SKIP_FINI && zone->uz_fini)
4541 		zone->uz_fini(item, zone->uz_size);
4542 
4543 	zone->uz_release(zone->uz_arg, &item, 1);
4544 
4545 	if (skip & SKIP_CNT)
4546 		return;
4547 
4548 	counter_u64_add(zone->uz_frees, 1);
4549 
4550 	if (zone->uz_max_items > 0)
4551 		zone_free_limit(zone, 1);
4552 }
4553 
4554 /* See uma.h */
4555 int
4556 uma_zone_set_max(uma_zone_t zone, int nitems)
4557 {
4558 
4559 	/*
4560 	 * If the limit is small, we may need to constrain the maximum per-CPU
4561 	 * cache size, or disable caching entirely.
4562 	 */
4563 	uma_zone_set_maxcache(zone, nitems);
4564 
4565 	/*
4566 	 * XXX This can misbehave if the zone has any allocations with
4567 	 * no limit and a limit is imposed.  There is currently no
4568 	 * way to clear a limit.
4569 	 */
4570 	ZONE_LOCK(zone);
4571 	zone->uz_max_items = nitems;
4572 	zone->uz_flags |= UMA_ZFLAG_LIMIT;
4573 	zone_update_caches(zone);
4574 	/* We may need to wake waiters. */
4575 	wakeup(&zone->uz_max_items);
4576 	ZONE_UNLOCK(zone);
4577 
4578 	return (nitems);
4579 }
4580 
4581 /* See uma.h */
4582 void
4583 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4584 {
4585 	int bpcpu, bpdom, bsize, nb;
4586 
4587 	ZONE_LOCK(zone);
4588 
4589 	/*
4590 	 * Compute a lower bound on the number of items that may be cached in
4591 	 * the zone.  Each CPU gets at least two buckets, and for cross-domain
4592 	 * frees we use an additional bucket per CPU and per domain.  Select the
4593 	 * largest bucket size that does not exceed half of the requested limit,
4594 	 * with the left over space given to the full bucket cache.
4595 	 */
4596 	bpdom = 0;
4597 	bpcpu = 2;
4598 #ifdef NUMA
4599 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4600 		bpcpu++;
4601 		bpdom++;
4602 	}
4603 #endif
4604 	nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4605 	bsize = nitems / nb / 2;
4606 	if (bsize > BUCKET_MAX)
4607 		bsize = BUCKET_MAX;
4608 	else if (bsize == 0 && nitems / nb > 0)
4609 		bsize = 1;
4610 	zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
4611 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4612 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4613 	zone->uz_bucket_max = nitems - nb * bsize;
4614 	ZONE_UNLOCK(zone);
4615 }
4616 
4617 /* See uma.h */
4618 int
4619 uma_zone_get_max(uma_zone_t zone)
4620 {
4621 	int nitems;
4622 
4623 	nitems = atomic_load_64(&zone->uz_max_items);
4624 
4625 	return (nitems);
4626 }
4627 
4628 /* See uma.h */
4629 void
4630 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4631 {
4632 
4633 	ZONE_ASSERT_COLD(zone);
4634 	zone->uz_warning = warning;
4635 }
4636 
4637 /* See uma.h */
4638 void
4639 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4640 {
4641 
4642 	ZONE_ASSERT_COLD(zone);
4643 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4644 }
4645 
4646 /* See uma.h */
4647 int
4648 uma_zone_get_cur(uma_zone_t zone)
4649 {
4650 	int64_t nitems;
4651 	u_int i;
4652 
4653 	nitems = 0;
4654 	if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4655 		nitems = counter_u64_fetch(zone->uz_allocs) -
4656 		    counter_u64_fetch(zone->uz_frees);
4657 	CPU_FOREACH(i)
4658 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4659 		    atomic_load_64(&zone->uz_cpu[i].uc_frees);
4660 
4661 	return (nitems < 0 ? 0 : nitems);
4662 }
4663 
4664 static uint64_t
4665 uma_zone_get_allocs(uma_zone_t zone)
4666 {
4667 	uint64_t nitems;
4668 	u_int i;
4669 
4670 	nitems = 0;
4671 	if (zone->uz_allocs != EARLY_COUNTER)
4672 		nitems = counter_u64_fetch(zone->uz_allocs);
4673 	CPU_FOREACH(i)
4674 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4675 
4676 	return (nitems);
4677 }
4678 
4679 static uint64_t
4680 uma_zone_get_frees(uma_zone_t zone)
4681 {
4682 	uint64_t nitems;
4683 	u_int i;
4684 
4685 	nitems = 0;
4686 	if (zone->uz_frees != EARLY_COUNTER)
4687 		nitems = counter_u64_fetch(zone->uz_frees);
4688 	CPU_FOREACH(i)
4689 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4690 
4691 	return (nitems);
4692 }
4693 
4694 #ifdef INVARIANTS
4695 /* Used only for KEG_ASSERT_COLD(). */
4696 static uint64_t
4697 uma_keg_get_allocs(uma_keg_t keg)
4698 {
4699 	uma_zone_t z;
4700 	uint64_t nitems;
4701 
4702 	nitems = 0;
4703 	LIST_FOREACH(z, &keg->uk_zones, uz_link)
4704 		nitems += uma_zone_get_allocs(z);
4705 
4706 	return (nitems);
4707 }
4708 #endif
4709 
4710 /* See uma.h */
4711 void
4712 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4713 {
4714 	uma_keg_t keg;
4715 
4716 	KEG_GET(zone, keg);
4717 	KEG_ASSERT_COLD(keg);
4718 	keg->uk_init = uminit;
4719 }
4720 
4721 /* See uma.h */
4722 void
4723 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4724 {
4725 	uma_keg_t keg;
4726 
4727 	KEG_GET(zone, keg);
4728 	KEG_ASSERT_COLD(keg);
4729 	keg->uk_fini = fini;
4730 }
4731 
4732 /* See uma.h */
4733 void
4734 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4735 {
4736 
4737 	ZONE_ASSERT_COLD(zone);
4738 	zone->uz_init = zinit;
4739 }
4740 
4741 /* See uma.h */
4742 void
4743 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4744 {
4745 
4746 	ZONE_ASSERT_COLD(zone);
4747 	zone->uz_fini = zfini;
4748 }
4749 
4750 /* See uma.h */
4751 void
4752 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4753 {
4754 	uma_keg_t keg;
4755 
4756 	KEG_GET(zone, keg);
4757 	KEG_ASSERT_COLD(keg);
4758 	keg->uk_freef = freef;
4759 }
4760 
4761 /* See uma.h */
4762 void
4763 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4764 {
4765 	uma_keg_t keg;
4766 
4767 	KEG_GET(zone, keg);
4768 	KEG_ASSERT_COLD(keg);
4769 	keg->uk_allocf = allocf;
4770 }
4771 
4772 /* See uma.h */
4773 void
4774 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4775 {
4776 
4777 	ZONE_ASSERT_COLD(zone);
4778 
4779 	KASSERT(smr != NULL, ("Got NULL smr"));
4780 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4781 	    ("zone %p (%s) already uses SMR", zone, zone->uz_name));
4782 	zone->uz_flags |= UMA_ZONE_SMR;
4783 	zone->uz_smr = smr;
4784 	zone_update_caches(zone);
4785 }
4786 
4787 smr_t
4788 uma_zone_get_smr(uma_zone_t zone)
4789 {
4790 
4791 	return (zone->uz_smr);
4792 }
4793 
4794 /* See uma.h */
4795 void
4796 uma_zone_reserve(uma_zone_t zone, int items)
4797 {
4798 	uma_keg_t keg;
4799 
4800 	KEG_GET(zone, keg);
4801 	KEG_ASSERT_COLD(keg);
4802 	keg->uk_reserve = items;
4803 }
4804 
4805 /* See uma.h */
4806 int
4807 uma_zone_reserve_kva(uma_zone_t zone, int count)
4808 {
4809 	uma_keg_t keg;
4810 	vm_offset_t kva;
4811 	u_int pages;
4812 
4813 	KEG_GET(zone, keg);
4814 	KEG_ASSERT_COLD(keg);
4815 	ZONE_ASSERT_COLD(zone);
4816 
4817 	pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4818 
4819 #ifdef UMA_MD_SMALL_ALLOC
4820 	if (keg->uk_ppera > 1) {
4821 #else
4822 	if (1) {
4823 #endif
4824 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4825 		if (kva == 0)
4826 			return (0);
4827 	} else
4828 		kva = 0;
4829 
4830 	MPASS(keg->uk_kva == 0);
4831 	keg->uk_kva = kva;
4832 	keg->uk_offset = 0;
4833 	zone->uz_max_items = pages * keg->uk_ipers;
4834 #ifdef UMA_MD_SMALL_ALLOC
4835 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4836 #else
4837 	keg->uk_allocf = noobj_alloc;
4838 #endif
4839 	keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4840 	zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4841 	zone_update_caches(zone);
4842 
4843 	return (1);
4844 }
4845 
4846 /* See uma.h */
4847 void
4848 uma_prealloc(uma_zone_t zone, int items)
4849 {
4850 	struct vm_domainset_iter di;
4851 	uma_domain_t dom;
4852 	uma_slab_t slab;
4853 	uma_keg_t keg;
4854 	int aflags, domain, slabs;
4855 
4856 	KEG_GET(zone, keg);
4857 	slabs = howmany(items, keg->uk_ipers);
4858 	while (slabs-- > 0) {
4859 		aflags = M_NOWAIT;
4860 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4861 		    &aflags);
4862 		for (;;) {
4863 			slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4864 			    aflags);
4865 			if (slab != NULL) {
4866 				dom = &keg->uk_domain[slab->us_domain];
4867 				/*
4868 				 * keg_alloc_slab() always returns a slab on the
4869 				 * partial list.
4870 				 */
4871 				LIST_REMOVE(slab, us_link);
4872 				LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4873 				    us_link);
4874 				dom->ud_free_slabs++;
4875 				KEG_UNLOCK(keg, slab->us_domain);
4876 				break;
4877 			}
4878 			if (vm_domainset_iter_policy(&di, &domain) != 0)
4879 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
4880 		}
4881 	}
4882 }
4883 
4884 /*
4885  * Returns a snapshot of memory consumption in bytes.
4886  */
4887 size_t
4888 uma_zone_memory(uma_zone_t zone)
4889 {
4890 	size_t sz;
4891 	int i;
4892 
4893 	sz = 0;
4894 	if (zone->uz_flags & UMA_ZFLAG_CACHE) {
4895 		for (i = 0; i < vm_ndomains; i++)
4896 			sz += ZDOM_GET(zone, i)->uzd_nitems;
4897 		return (sz * zone->uz_size);
4898 	}
4899 	for (i = 0; i < vm_ndomains; i++)
4900 		sz += zone->uz_keg->uk_domain[i].ud_pages;
4901 
4902 	return (sz * PAGE_SIZE);
4903 }
4904 
4905 /* See uma.h */
4906 void
4907 uma_reclaim(int req)
4908 {
4909 
4910 	CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4911 	sx_xlock(&uma_reclaim_lock);
4912 	bucket_enable();
4913 
4914 	switch (req) {
4915 	case UMA_RECLAIM_TRIM:
4916 		zone_foreach(zone_trim, NULL);
4917 		break;
4918 	case UMA_RECLAIM_DRAIN:
4919 	case UMA_RECLAIM_DRAIN_CPU:
4920 		zone_foreach(zone_drain, NULL);
4921 		if (req == UMA_RECLAIM_DRAIN_CPU) {
4922 			pcpu_cache_drain_safe(NULL);
4923 			zone_foreach(zone_drain, NULL);
4924 		}
4925 		break;
4926 	default:
4927 		panic("unhandled reclamation request %d", req);
4928 	}
4929 
4930 	/*
4931 	 * Some slabs may have been freed but this zone will be visited early
4932 	 * we visit again so that we can free pages that are empty once other
4933 	 * zones are drained.  We have to do the same for buckets.
4934 	 */
4935 	zone_drain(slabzones[0], NULL);
4936 	zone_drain(slabzones[1], NULL);
4937 	bucket_zone_drain();
4938 	sx_xunlock(&uma_reclaim_lock);
4939 }
4940 
4941 static volatile int uma_reclaim_needed;
4942 
4943 void
4944 uma_reclaim_wakeup(void)
4945 {
4946 
4947 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4948 		wakeup(uma_reclaim);
4949 }
4950 
4951 void
4952 uma_reclaim_worker(void *arg __unused)
4953 {
4954 
4955 	for (;;) {
4956 		sx_xlock(&uma_reclaim_lock);
4957 		while (atomic_load_int(&uma_reclaim_needed) == 0)
4958 			sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4959 			    hz);
4960 		sx_xunlock(&uma_reclaim_lock);
4961 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4962 		uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4963 		atomic_store_int(&uma_reclaim_needed, 0);
4964 		/* Don't fire more than once per-second. */
4965 		pause("umarclslp", hz);
4966 	}
4967 }
4968 
4969 /* See uma.h */
4970 void
4971 uma_zone_reclaim(uma_zone_t zone, int req)
4972 {
4973 
4974 	switch (req) {
4975 	case UMA_RECLAIM_TRIM:
4976 		zone_trim(zone, NULL);
4977 		break;
4978 	case UMA_RECLAIM_DRAIN:
4979 		zone_drain(zone, NULL);
4980 		break;
4981 	case UMA_RECLAIM_DRAIN_CPU:
4982 		pcpu_cache_drain_safe(zone);
4983 		zone_drain(zone, NULL);
4984 		break;
4985 	default:
4986 		panic("unhandled reclamation request %d", req);
4987 	}
4988 }
4989 
4990 /* See uma.h */
4991 int
4992 uma_zone_exhausted(uma_zone_t zone)
4993 {
4994 
4995 	return (atomic_load_32(&zone->uz_sleepers) > 0);
4996 }
4997 
4998 unsigned long
4999 uma_limit(void)
5000 {
5001 
5002 	return (uma_kmem_limit);
5003 }
5004 
5005 void
5006 uma_set_limit(unsigned long limit)
5007 {
5008 
5009 	uma_kmem_limit = limit;
5010 }
5011 
5012 unsigned long
5013 uma_size(void)
5014 {
5015 
5016 	return (atomic_load_long(&uma_kmem_total));
5017 }
5018 
5019 long
5020 uma_avail(void)
5021 {
5022 
5023 	return (uma_kmem_limit - uma_size());
5024 }
5025 
5026 #ifdef DDB
5027 /*
5028  * Generate statistics across both the zone and its per-cpu cache's.  Return
5029  * desired statistics if the pointer is non-NULL for that statistic.
5030  *
5031  * Note: does not update the zone statistics, as it can't safely clear the
5032  * per-CPU cache statistic.
5033  *
5034  */
5035 static void
5036 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5037     uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5038 {
5039 	uma_cache_t cache;
5040 	uint64_t allocs, frees, sleeps, xdomain;
5041 	int cachefree, cpu;
5042 
5043 	allocs = frees = sleeps = xdomain = 0;
5044 	cachefree = 0;
5045 	CPU_FOREACH(cpu) {
5046 		cache = &z->uz_cpu[cpu];
5047 		cachefree += cache->uc_allocbucket.ucb_cnt;
5048 		cachefree += cache->uc_freebucket.ucb_cnt;
5049 		xdomain += cache->uc_crossbucket.ucb_cnt;
5050 		cachefree += cache->uc_crossbucket.ucb_cnt;
5051 		allocs += cache->uc_allocs;
5052 		frees += cache->uc_frees;
5053 	}
5054 	allocs += counter_u64_fetch(z->uz_allocs);
5055 	frees += counter_u64_fetch(z->uz_frees);
5056 	xdomain += counter_u64_fetch(z->uz_xdomain);
5057 	sleeps += z->uz_sleeps;
5058 	if (cachefreep != NULL)
5059 		*cachefreep = cachefree;
5060 	if (allocsp != NULL)
5061 		*allocsp = allocs;
5062 	if (freesp != NULL)
5063 		*freesp = frees;
5064 	if (sleepsp != NULL)
5065 		*sleepsp = sleeps;
5066 	if (xdomainp != NULL)
5067 		*xdomainp = xdomain;
5068 }
5069 #endif /* DDB */
5070 
5071 static int
5072 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5073 {
5074 	uma_keg_t kz;
5075 	uma_zone_t z;
5076 	int count;
5077 
5078 	count = 0;
5079 	rw_rlock(&uma_rwlock);
5080 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5081 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5082 			count++;
5083 	}
5084 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5085 		count++;
5086 
5087 	rw_runlock(&uma_rwlock);
5088 	return (sysctl_handle_int(oidp, &count, 0, req));
5089 }
5090 
5091 static void
5092 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5093     struct uma_percpu_stat *ups, bool internal)
5094 {
5095 	uma_zone_domain_t zdom;
5096 	uma_cache_t cache;
5097 	int i;
5098 
5099 	for (i = 0; i < vm_ndomains; i++) {
5100 		zdom = ZDOM_GET(z, i);
5101 		uth->uth_zone_free += zdom->uzd_nitems;
5102 	}
5103 	uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5104 	uth->uth_frees = counter_u64_fetch(z->uz_frees);
5105 	uth->uth_fails = counter_u64_fetch(z->uz_fails);
5106 	uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5107 	uth->uth_sleeps = z->uz_sleeps;
5108 
5109 	for (i = 0; i < mp_maxid + 1; i++) {
5110 		bzero(&ups[i], sizeof(*ups));
5111 		if (internal || CPU_ABSENT(i))
5112 			continue;
5113 		cache = &z->uz_cpu[i];
5114 		ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5115 		ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5116 		ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5117 		ups[i].ups_allocs = cache->uc_allocs;
5118 		ups[i].ups_frees = cache->uc_frees;
5119 	}
5120 }
5121 
5122 static int
5123 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5124 {
5125 	struct uma_stream_header ush;
5126 	struct uma_type_header uth;
5127 	struct uma_percpu_stat *ups;
5128 	struct sbuf sbuf;
5129 	uma_keg_t kz;
5130 	uma_zone_t z;
5131 	uint64_t items;
5132 	uint32_t kfree, pages;
5133 	int count, error, i;
5134 
5135 	error = sysctl_wire_old_buffer(req, 0);
5136 	if (error != 0)
5137 		return (error);
5138 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5139 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5140 	ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5141 
5142 	count = 0;
5143 	rw_rlock(&uma_rwlock);
5144 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5145 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5146 			count++;
5147 	}
5148 
5149 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5150 		count++;
5151 
5152 	/*
5153 	 * Insert stream header.
5154 	 */
5155 	bzero(&ush, sizeof(ush));
5156 	ush.ush_version = UMA_STREAM_VERSION;
5157 	ush.ush_maxcpus = (mp_maxid + 1);
5158 	ush.ush_count = count;
5159 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5160 
5161 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5162 		kfree = pages = 0;
5163 		for (i = 0; i < vm_ndomains; i++) {
5164 			kfree += kz->uk_domain[i].ud_free_items;
5165 			pages += kz->uk_domain[i].ud_pages;
5166 		}
5167 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5168 			bzero(&uth, sizeof(uth));
5169 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5170 			uth.uth_align = kz->uk_align;
5171 			uth.uth_size = kz->uk_size;
5172 			uth.uth_rsize = kz->uk_rsize;
5173 			if (z->uz_max_items > 0) {
5174 				items = UZ_ITEMS_COUNT(z->uz_items);
5175 				uth.uth_pages = (items / kz->uk_ipers) *
5176 					kz->uk_ppera;
5177 			} else
5178 				uth.uth_pages = pages;
5179 			uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5180 			    kz->uk_ppera;
5181 			uth.uth_limit = z->uz_max_items;
5182 			uth.uth_keg_free = kfree;
5183 
5184 			/*
5185 			 * A zone is secondary is it is not the first entry
5186 			 * on the keg's zone list.
5187 			 */
5188 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5189 			    (LIST_FIRST(&kz->uk_zones) != z))
5190 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5191 			uma_vm_zone_stats(&uth, z, &sbuf, ups,
5192 			    kz->uk_flags & UMA_ZFLAG_INTERNAL);
5193 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5194 			for (i = 0; i < mp_maxid + 1; i++)
5195 				(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5196 		}
5197 	}
5198 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5199 		bzero(&uth, sizeof(uth));
5200 		strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5201 		uth.uth_size = z->uz_size;
5202 		uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5203 		(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5204 		for (i = 0; i < mp_maxid + 1; i++)
5205 			(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5206 	}
5207 
5208 	rw_runlock(&uma_rwlock);
5209 	error = sbuf_finish(&sbuf);
5210 	sbuf_delete(&sbuf);
5211 	free(ups, M_TEMP);
5212 	return (error);
5213 }
5214 
5215 int
5216 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5217 {
5218 	uma_zone_t zone = *(uma_zone_t *)arg1;
5219 	int error, max;
5220 
5221 	max = uma_zone_get_max(zone);
5222 	error = sysctl_handle_int(oidp, &max, 0, req);
5223 	if (error || !req->newptr)
5224 		return (error);
5225 
5226 	uma_zone_set_max(zone, max);
5227 
5228 	return (0);
5229 }
5230 
5231 int
5232 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5233 {
5234 	uma_zone_t zone;
5235 	int cur;
5236 
5237 	/*
5238 	 * Some callers want to add sysctls for global zones that
5239 	 * may not yet exist so they pass a pointer to a pointer.
5240 	 */
5241 	if (arg2 == 0)
5242 		zone = *(uma_zone_t *)arg1;
5243 	else
5244 		zone = arg1;
5245 	cur = uma_zone_get_cur(zone);
5246 	return (sysctl_handle_int(oidp, &cur, 0, req));
5247 }
5248 
5249 static int
5250 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5251 {
5252 	uma_zone_t zone = arg1;
5253 	uint64_t cur;
5254 
5255 	cur = uma_zone_get_allocs(zone);
5256 	return (sysctl_handle_64(oidp, &cur, 0, req));
5257 }
5258 
5259 static int
5260 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5261 {
5262 	uma_zone_t zone = arg1;
5263 	uint64_t cur;
5264 
5265 	cur = uma_zone_get_frees(zone);
5266 	return (sysctl_handle_64(oidp, &cur, 0, req));
5267 }
5268 
5269 static int
5270 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5271 {
5272 	struct sbuf sbuf;
5273 	uma_zone_t zone = arg1;
5274 	int error;
5275 
5276 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5277 	if (zone->uz_flags != 0)
5278 		sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5279 	else
5280 		sbuf_printf(&sbuf, "0");
5281 	error = sbuf_finish(&sbuf);
5282 	sbuf_delete(&sbuf);
5283 
5284 	return (error);
5285 }
5286 
5287 static int
5288 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5289 {
5290 	uma_keg_t keg = arg1;
5291 	int avail, effpct, total;
5292 
5293 	total = keg->uk_ppera * PAGE_SIZE;
5294 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5295 		total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5296 	/*
5297 	 * We consider the client's requested size and alignment here, not the
5298 	 * real size determination uk_rsize, because we also adjust the real
5299 	 * size for internal implementation reasons (max bitset size).
5300 	 */
5301 	avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5302 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5303 		avail *= mp_maxid + 1;
5304 	effpct = 100 * avail / total;
5305 	return (sysctl_handle_int(oidp, &effpct, 0, req));
5306 }
5307 
5308 static int
5309 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5310 {
5311 	uma_zone_t zone = arg1;
5312 	uint64_t cur;
5313 
5314 	cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5315 	return (sysctl_handle_64(oidp, &cur, 0, req));
5316 }
5317 
5318 #ifdef INVARIANTS
5319 static uma_slab_t
5320 uma_dbg_getslab(uma_zone_t zone, void *item)
5321 {
5322 	uma_slab_t slab;
5323 	uma_keg_t keg;
5324 	uint8_t *mem;
5325 
5326 	/*
5327 	 * It is safe to return the slab here even though the
5328 	 * zone is unlocked because the item's allocation state
5329 	 * essentially holds a reference.
5330 	 */
5331 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5332 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5333 		return (NULL);
5334 	if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5335 		return (vtoslab((vm_offset_t)mem));
5336 	keg = zone->uz_keg;
5337 	if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5338 		return ((uma_slab_t)(mem + keg->uk_pgoff));
5339 	KEG_LOCK(keg, 0);
5340 	slab = hash_sfind(&keg->uk_hash, mem);
5341 	KEG_UNLOCK(keg, 0);
5342 
5343 	return (slab);
5344 }
5345 
5346 static bool
5347 uma_dbg_zskip(uma_zone_t zone, void *mem)
5348 {
5349 
5350 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5351 		return (true);
5352 
5353 	return (uma_dbg_kskip(zone->uz_keg, mem));
5354 }
5355 
5356 static bool
5357 uma_dbg_kskip(uma_keg_t keg, void *mem)
5358 {
5359 	uintptr_t idx;
5360 
5361 	if (dbg_divisor == 0)
5362 		return (true);
5363 
5364 	if (dbg_divisor == 1)
5365 		return (false);
5366 
5367 	idx = (uintptr_t)mem >> PAGE_SHIFT;
5368 	if (keg->uk_ipers > 1) {
5369 		idx *= keg->uk_ipers;
5370 		idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5371 	}
5372 
5373 	if ((idx / dbg_divisor) * dbg_divisor != idx) {
5374 		counter_u64_add(uma_skip_cnt, 1);
5375 		return (true);
5376 	}
5377 	counter_u64_add(uma_dbg_cnt, 1);
5378 
5379 	return (false);
5380 }
5381 
5382 /*
5383  * Set up the slab's freei data such that uma_dbg_free can function.
5384  *
5385  */
5386 static void
5387 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5388 {
5389 	uma_keg_t keg;
5390 	int freei;
5391 
5392 	if (slab == NULL) {
5393 		slab = uma_dbg_getslab(zone, item);
5394 		if (slab == NULL)
5395 			panic("uma: item %p did not belong to zone %s",
5396 			    item, zone->uz_name);
5397 	}
5398 	keg = zone->uz_keg;
5399 	freei = slab_item_index(slab, keg, item);
5400 
5401 	if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5402 	    slab_dbg_bits(slab, keg)))
5403 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5404 		    item, zone, zone->uz_name, slab, freei);
5405 }
5406 
5407 /*
5408  * Verifies freed addresses.  Checks for alignment, valid slab membership
5409  * and duplicate frees.
5410  *
5411  */
5412 static void
5413 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5414 {
5415 	uma_keg_t keg;
5416 	int freei;
5417 
5418 	if (slab == NULL) {
5419 		slab = uma_dbg_getslab(zone, item);
5420 		if (slab == NULL)
5421 			panic("uma: Freed item %p did not belong to zone %s",
5422 			    item, zone->uz_name);
5423 	}
5424 	keg = zone->uz_keg;
5425 	freei = slab_item_index(slab, keg, item);
5426 
5427 	if (freei >= keg->uk_ipers)
5428 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5429 		    item, zone, zone->uz_name, slab, freei);
5430 
5431 	if (slab_item(slab, keg, freei) != item)
5432 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5433 		    item, zone, zone->uz_name, slab, freei);
5434 
5435 	if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5436 	    slab_dbg_bits(slab, keg)))
5437 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5438 		    item, zone, zone->uz_name, slab, freei);
5439 }
5440 #endif /* INVARIANTS */
5441 
5442 #ifdef DDB
5443 static int64_t
5444 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5445     uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5446 {
5447 	uint64_t frees;
5448 	int i;
5449 
5450 	if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5451 		*allocs = counter_u64_fetch(z->uz_allocs);
5452 		frees = counter_u64_fetch(z->uz_frees);
5453 		*sleeps = z->uz_sleeps;
5454 		*cachefree = 0;
5455 		*xdomain = 0;
5456 	} else
5457 		uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5458 		    xdomain);
5459 	for (i = 0; i < vm_ndomains; i++) {
5460 		*cachefree += ZDOM_GET(z, i)->uzd_nitems;
5461 		if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5462 		    (LIST_FIRST(&kz->uk_zones) != z)))
5463 			*cachefree += kz->uk_domain[i].ud_free_items;
5464 	}
5465 	*used = *allocs - frees;
5466 	return (((int64_t)*used + *cachefree) * kz->uk_size);
5467 }
5468 
5469 DB_SHOW_COMMAND(uma, db_show_uma)
5470 {
5471 	const char *fmt_hdr, *fmt_entry;
5472 	uma_keg_t kz;
5473 	uma_zone_t z;
5474 	uint64_t allocs, used, sleeps, xdomain;
5475 	long cachefree;
5476 	/* variables for sorting */
5477 	uma_keg_t cur_keg;
5478 	uma_zone_t cur_zone, last_zone;
5479 	int64_t cur_size, last_size, size;
5480 	int ties;
5481 
5482 	/* /i option produces machine-parseable CSV output */
5483 	if (modif[0] == 'i') {
5484 		fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5485 		fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5486 	} else {
5487 		fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5488 		fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5489 	}
5490 
5491 	db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5492 	    "Sleeps", "Bucket", "Total Mem", "XFree");
5493 
5494 	/* Sort the zones with largest size first. */
5495 	last_zone = NULL;
5496 	last_size = INT64_MAX;
5497 	for (;;) {
5498 		cur_zone = NULL;
5499 		cur_size = -1;
5500 		ties = 0;
5501 		LIST_FOREACH(kz, &uma_kegs, uk_link) {
5502 			LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5503 				/*
5504 				 * In the case of size ties, print out zones
5505 				 * in the order they are encountered.  That is,
5506 				 * when we encounter the most recently output
5507 				 * zone, we have already printed all preceding
5508 				 * ties, and we must print all following ties.
5509 				 */
5510 				if (z == last_zone) {
5511 					ties = 1;
5512 					continue;
5513 				}
5514 				size = get_uma_stats(kz, z, &allocs, &used,
5515 				    &sleeps, &cachefree, &xdomain);
5516 				if (size > cur_size && size < last_size + ties)
5517 				{
5518 					cur_size = size;
5519 					cur_zone = z;
5520 					cur_keg = kz;
5521 				}
5522 			}
5523 		}
5524 		if (cur_zone == NULL)
5525 			break;
5526 
5527 		size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5528 		    &sleeps, &cachefree, &xdomain);
5529 		db_printf(fmt_entry, cur_zone->uz_name,
5530 		    (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5531 		    (uintmax_t)allocs, (uintmax_t)sleeps,
5532 		    (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5533 		    xdomain);
5534 
5535 		if (db_pager_quit)
5536 			return;
5537 		last_zone = cur_zone;
5538 		last_size = cur_size;
5539 	}
5540 }
5541 
5542 DB_SHOW_COMMAND(umacache, db_show_umacache)
5543 {
5544 	uma_zone_t z;
5545 	uint64_t allocs, frees;
5546 	long cachefree;
5547 	int i;
5548 
5549 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5550 	    "Requests", "Bucket");
5551 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5552 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5553 		for (i = 0; i < vm_ndomains; i++)
5554 			cachefree += ZDOM_GET(z, i)->uzd_nitems;
5555 		db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5556 		    z->uz_name, (uintmax_t)z->uz_size,
5557 		    (intmax_t)(allocs - frees), cachefree,
5558 		    (uintmax_t)allocs, z->uz_bucket_size);
5559 		if (db_pager_quit)
5560 			return;
5561 	}
5562 }
5563 #endif	/* DDB */
5564